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.1007373 | Glycolysis regulates pollen tube polarity via Rho GTPase signaling | As a universal energy generation pathway utilizing carbon metabolism, glycolysis plays an important housekeeping role in all organisms. Pollen tubes expand rapidly via a mechanism of polarized growth, known as tip growth, to deliver sperm for fertilization. Here, we report a novel and surprising role of glycolysis in the regulation of growth polarity in Arabidopsis pollen tubes via impingement of Rho GTPase-dependent signaling. We identified a cytosolic phosphoglycerate kinase (pgkc-1) mutant with accelerated pollen germination and compromised pollen tube growth polarity. pgkc-1 mutation greatly diminished apical exocytic vesicular distribution of REN1 RopGAP (Rop GTPase activating protein), leading to ROP1 hyper-activation at the apical plasma membrane. Consequently, pgkc-1 pollen tubes contained higher amounts of exocytic vesicles and actin microfilaments in the apical region, and showed reduced sensitivity to Brefeldin A and Latrunculin B, respectively. While inhibition of mitochondrial respiration could not explain the pgkc-1 phenotype, the glycolytic activity is indeed required for PGKc function in pollen tubes. Moreover, the pgkc-1 pollen tube phenotype was mimicked by the inhibition of another glycolytic enzyme. These findings highlight an unconventional regulatory function for a housekeeping metabolic pathway in the spatial control of a fundamental cellular process.
| Glycolysis, which breaks down glucose to produce energy, has long been considered a “housekeeping” pathway in living cells, i.e., it helps maintain basic cellular functions. Here, we found that the glycolysis pathway plays an unconventional regulatory role in cell polarity, i.e., the intrinsic asymmetry in the shape, structure, and organization of cellular components. Mutation in the gene encoding the glycolytic enzyme cytosolic phosphoglycerate kinase (PGKc) leads to swollen and shorter pollen tubes in Arabidopsis thaliana, which is associated with the over-activation of Rho GTPase—a master regulator of cell polarity. Our results suggest that this phenomenon is caused by a specific regulatory role of cytosolic glycolysis rather than the global energy supply or moonlighting functions of glycolytic enzymes that modulate pollen tube growth polarity. Our findings shed light on the diverse biological roles of glycolysis in plants beyond simple “housekeeping” functions.
| Glycolysis, which generates two ATP from each glucose molecule and produces two pyruvate molecules to fuel the mitochondrial tricarboxylic acid cycle, is a central enzymatic process in carbon metabolism. In addition, glycolysis also produces metabolic intermediates and reduced cofactors for secondary metabolism, as well as amino acid and fatty acid biosynthesis [1, 2]. Recent studies have hinted at a role for energy in the regulation of cellular processes independent of the housekeeping function. For instance, aldolase, a glycolytic enzyme, acts as a sensor of glucose availability in mammalian cells, and represses the energy sensing AMP-dependent kinase (AMPK) pathway, which is known to coordinate cell growth, metabolism, and cell polarity [3–6]. Therefore, glycolysis may play a regulatory role in determining cell polarity regulation, although direct evidence for this role is lacking thus far.
Polarized cell growth is a conserved cellular process shared by many diverse systems in eukaryotic species, such as the mating tubes in budding yeast, cell growth and morphogenesis in fission yeast and filamentous fungi, axon outgrowth in animals, and root hair and pollen tube formation in plants. Pollen tubes are a well-established and favorite model system for studying cell polarity formation and polar cell growth [7]. Pollen tubes are among the most rapidly extending polarized cells, growing at rates of up to 250 nm per second [8]. The rapid tip growth exhibited by pollen tubes is supported by cytoskeletal organization/dynamics and vesicular trafficking coordinated by a conserved signaling network dependent upon a plant Rho GTPase (ROP1) [7–12]. ROP1 is activated in the apical region, where it orchestrates F-actin dynamics and calcium homeostasis to dynamically maintain apical growth in the pollen tube [11, 13]. REN1, a RhoGAP, acts as a global inhibitor to spatially restrict ROP1 activity to the apical plasma membrane at the pollen-tube tip region [14]. This self-organizing ROP signaling network is comprised of multiple coordinated pathways and feedback loops, providing a robust molecular linkage between the cytoskeleton, vesicular trafficking, and polarity formation [11–13, 15–19].
It is conceivable that the rapid tip growth exhibited by pollen tubes is extremely energy-demanding. Overall elevations in energy metabolism in pollen tubes appears to rely on plastid-localized glycolysis and mitochondrial-localized respiration pathways [20–23]. As a result, respiration rates in pollen tubes are up to ten times greater than those in vegetative tissues [24]. In addition, the ethanol fermentation pathway is also active in support of pollen tube growth [25]. Apart from an overall increase in energy metabolism, rapid tip growth may also require a tight spatiotemporal regulation of energy production, given the tightly regulated spatiotemporal dynamic of the aforementioned processes.
Phosphoglycerate kinase (PGK) is a key enzyme in the glycolytic pathway, responsible for catalyzing the reversible conversion of 1,3-biphosphoglycerate (1,3BPG) to 3-phosphoglycerate (3PG). Here, we report that the Arabidopsis cytosolic phosphoglycerate kinase (PGKc) plays a regulatory role in the regulation of pollen tube polarity by modulating the apical distribution of the REN1 RhoGAP, and thus the activity of the apical ROP1 RhoGTPase as well. This action of PGKc is specific for the cytosolic glycolysis pathway and is independent of mitochondrial respiration. Our findings provide the first conclusive evidence that glycolysis plays an important and specific role in the regulation of cell polarity.
To discover new genes regulating pollen tube polarity, we performed a genetic screen for Arabidopsis thaliana mutants from the SALK collection of individually indexed homozygous T-DNA insertion lines presenting altered growth polarity. Among over 8000 individual lines screened, pollen tubes from SALK_066422C were identified to present defective polarized growth in in vitro germination medium. According to the annotation, SALK_066422C contains a T-DNA inserted into the 5th exon of AT1G79550, which encodes a cytosolic phosphoglycerate kinase (designated hereafter as PGKc) (S1A and S1B Fig). We therefore designated the mutant as pgkc-1. The pgkc-1 mutant pollen germinated at a much faster rate than wild type (WT) plants (Fig 1D). However, pgkc-1 mutant pollen tubes were significantly shorter than WT ones after 9 h (Fig 1A, 1B and 1E). Moreover, the majority of mutant pollen tubes were swollen relative to WT, exhibiting irregular morphology and wider tube width (Fig 1A, 1B and 1F).
We also obtained an independent allele mutant with a T-DNA insertion in the 3rd intron of PGKc (SALK_062377, designated pgkc-2), which showed similar pollen tube phenotypes (S1A, S1B and S1G Fig). Quantitative reverse transcription polymerase chain reaction (Q-RT-PCR) showed that both pgkc-1 and pgkc-2 are knockout mutants for PGKc (S1C Fig). We also performed a backcross of pgkc-1 with WT plants, where F2 progeny pgkc-1 homozygous plants showed defects in pollen tube polarity while WT progeny remained normal (S1E and S1F Fig). This indicates the co-segregation of the pgkc-1 locus with the mutant phenotype. Finally, the pgkc-1 mutant was rescued by introducing PGKc genomic sequences, including the native promoter and terminator (Fig 1C to 1F and S1D Fig). Taken together, our results confirm that loss of PGKc is indeed responsible for the pollen tube polarity phenotype. The vegetative growth and flowering of pgkc-1 plants were slightly delayed relative to WT, but mutant plant morphology was normal otherwise (S2A–S2C Fig).
The Arabidopsis genome contains three PGK genes, AT1G79550 (PGKc), AT3G12780, and AT1G56190. Recent reports have shown that AT1G79550 encodes the sole cytosolic PGK, while AT1G79550 and AT3G12780 are plastid localized [26]. We also performed subcellular localization analysis using a GFP fusion protein. Consistent with the results of a previous study, we found PGKc to be localized to the cytoplasm and nuclei while the other 2 PGKs were localized to the chloroplasts (plastids) (S3A–S3F Fig). Finally, both a previous study and publicly available microarray expression data showed that PGKc is expressed ubiquitously in most plant tissues, including pollen (https://genevestigator.com/) [26].
Our surprising findings regarding PGKc knockouts prompted us to assess how a housekeeping glycolytic enzyme can regulate cell polarity. We first performed a series of assays to assess pgkc-1 mutant phenotype cellular mechanisms with known links to cell polarity defects. The spatiotemporal dynamics of apical actin microfilaments (F-actin) and vesicle trafficking is crucial for generation of cell polarity and pollen tube tip growth [7, 13]. We observed F-actin organization in pgkc-1 pollen tubes by introducing a Lifeact-mEGFP marker via crossing [2, 27]. In WT pollen tubes, highly dynamic fine F-actin structures were observed in the apical region, dense F-actin fringe structures were present in sub-apical regions, and parallel longitudinal F-actin bundles were found in shank regions (Fig 2A). Dynamic apical F-actin has been shown to be disrupted by treatment with 1.5 nM Latrunculin B (LatB), a chemical promoting actin depolymerization [9] [28] (Fig 2A and 2B). In pgkc-1 pollen tubes, no significant difference was detected in the shank and sub-apical regions. However, fine F-actin filaments were significantly over-accumulated towards the apex of the apical tip region in pgkc-1 pollen tubes, even after LatB treatment (Fig 2A and 2B). Indeed, treatment with 1.5 nM LatB had no significant effect on the germination, length, and morphology of pgkc-1 mutant pollen tubes, but greatly inhibited similar mechanisms in WT pollen tubes (Fig 2C to 2G).Taken together, these results indicate that pgkc-1 mutation promotes the accumulation of F-actin in the apical tip region of the pollen tube.
A previous study has shown that an increased level of apical F-actin leads to greater apical accumulation of exocytic vesicles [11]. Thus, we examined the distribution of exocytic vesicles in pgkc-1 mutant pollen tubes. The Rab GTPase RABA4D is a pollen-specific Arabidopsis homolog of animal Rab11 known to localize to post-Golgi compartments, including exocytic vesicles in pollen tube tips [14, 29]. We introduced an EYFP-RABA4D marker into pgkc-1 pollen tubes via crossing. In WT pollen tubes, EYFP-RABA4D-labeled membrane compartments were punctuated and enriched in the tip region (Fig 3A). In pgkc-1 pollen tubes, the apical distribution pattern of RABA4D was similar to that of WT pollen tubes (Fig 3A), but quantification of EYFP-RABA4D signal showed that apical EYFP-RABA4D compartments were much more enriched in pgkc-1 pollen tube compared to WT despite lower signal intensity in the shank region (Fig 3B), a similar pattern to that observed in pollen tubes with ROP1 over-activation [11].
We next examined whether pgkc-1 pollen tubes respond differently to Brefeldin A (BFA), an inhibitor which interrupts vesicle trafficking by inhibiting vesicle formation from TGN and recycling endosomes [30–33]. Application of 0.4 μM BFA abolished the apical enrichment of RABA4D signal observed in WT pollen tubes, but had a markedly reduced effect on RABA4D localization in pgkc-1 pollen tubes (Fig 3A and 3B). Moreover, BFA greatly inhibited WT pollen germination but only moderately affected pgkc-1 pollen germination (Fig 3C to 3G). Interestingly, pgkc-1 pollen tubes exhibited enhanced growth depolarization when treated with BFA (Fig 3C to 3G). Germinated pgkc-1 pollen tubes were shorter and wider, and multiple tips occasionally formed from a single pollen grain (Fig 3D). These results suggested that the pgkc-1 mutation appears to enhance the production or accumulation of exocytic vesicles in pollen grains and tubes. This altered vesicular trafficking behavior in pollen tube tips is consistent with the aforementioned observed over-accumulation of apical F-actin [11].
Given the role of ROP1 GTPase signaling in regulating F-actin dynamics and vesicle trafficking [7, 11, 13], we speculated that the F-actin dynamics and vesicle trafficking phenotype in the pgkc-1 mutant may be linked to altered ROP1 signaling. RIC4 binds active ROP1 via its CRIB4 domain. CRIB4-GFP localization to the plasma membrane indicates ROP1 activity in pollen tubes [14]. To evaluate whether ROP1 activity was altered in pgkc-1 mutants, we introduced CRIB4-GFP into the pgkc-1 mutant background. CRIB4-GFP localization to the apical plasma membrane was significantly broader in pgkc-1 mutant pollen tubes than in WT (Fig 4A). Quantification revealed stronger CRIB4-GFP signal in pgkc-1 pollen tubes than in WT counterparts (Fig 4B). This result suggested that active ROP1 levels were indeed excessive in pgkc-1 pollen tubes.
A previous study has shown that the RhoGAP REN1 is an important regulator of ROP1 negative feedback loops. REN1 is localized to exocytic vesicles in the pollen tube tip [14]. A mutation in REN1 causes swollen pollen tubes and is correlated with hyper-activation of ROP1 [14]. We therefore introduced a GFP-REN1 reporter into pgkc-1 plants to observe the subcellular distribution of this negative feedback regulator of ROP signaling. Consistent with the previous study, GFP-REN1 was enriched in the apical region in an inverted-cone pattern, reminiscent of the distribution of RABA4D-labeled vesicles (Figs 4C and 3A). Strikingly, in pgkc-1 pollen tubes, this apical localization of GFP-REN1 was abolished in stark contrast to the enhanced apical accumulation of RABA4D-labeled vesicles observed (Fig 4C and 4D). These results indicate that pgkc-1 mutation disrupted apical localization of REN1, which may be associated with ROP1 hyper-activation.
To examine the functional interaction between PGKc and REN1, we generated double mutants using pgkc-1 and ren1-3 mutant plants. ren1-3 plants contain a weak mutation consisting of a C terminus truncation which confers a mild polarization defect [14] (S4 Fig). If PGKc functionally interacts with REN1, the tip-targeting defect of REN1 present in pgkc-1 plants would have a synergistic effect with the phenotype observed in the ren1-3 mutant. We found that in standard medium, ren1-3 pollen tubes displayed near normal growth and morphology, while pgkc-1 pollen tubes exhibited reduced growth and moderate polarity defects (Fig 4E and 4F). However, the pollen tubes of the pgkc-1/ren1-3 double mutant plants were much shorter and dramatically more swollen compared to either single mutant (Fig 4G to 4I). These results indicate that a moderate REN1 defect in ren1-3 was synergistically enhanced by pgkc-1, demonstrating the genetic interaction between PGKc and REN1.
We reasoned that PGKc, as a glycolytic enzyme, regulated pollen tube polarity through one or more of the following possible mechanisms: (1) pollen tube polarity may be linked to overall cellular ATP level, which is dependent on both glycolysis and mitochondrial respiration; (2) glycolysis may play a regulatory role in determining pollen tube polarity; and (3) PGKc may have evolved a new, so-called “moonlighting” function distinct from its role in glycolysis. We performed a series of assays to examine these possibilities.
To assess a possible relationship between cellular ATP level and pollen tube polarity, we determined whether mitochondrial respiration, the downstream pathway of glycolysis and the main source of cellular ATP production, was involved in pollen tube polarity. The potent inhibitor oligomycin has been used to block mitochondrial respiration in pollen germination medium [34]. In our assay, 40 nM oligomycin significantly inhibited WT pollen tube growth (Fig 5A and 5B). However, oligomycin-treated pollen tubes were uniformly short and thin, exhibiting a distinctly different phenotype than pgkc-1 pollen tubes (Fig 5A to 5D). Therefore, we concluded that the pgkc-1 pollen tube phenotype was likely not caused by inhibition of respiration.
To check if the role of PGKc in pollen tube polarity could be attributed to its glycolytic enzymatic activity, we generated a mutant version of PGKc termed mPGKc where an evolutionally conserved residue Glutamate179 was changed to Glutamine. This mutation has been shown to impair PGK catalytic activity but not binding kinetics in yeast [35, 36] (Fig 6A). We introduced native promoter-driven PGKc or mPGKc cDNA into pgkc-1 mutants, and found that WT PGKc cDNA transgene expression, while lower than native PGKc expression, was still able to complement the mutant phenotype (Fig 6B to 6E). In contrast, mPGKc could not rescue the mutant phenotype despite similar levels of gene expression. These results indicated that glycolytic activity was required for PGKc function in pollen tube polarity (Fig 6B to 6E, S5 Fig).
If PGKc regulates pollen tube polarity through its glycolytic activity, we would anticipate that other glycolytic enzymes are also involved in this process. GAPDH is an enzyme which catalyzes the conversion of glyceraldehyde-3-phosphate to 1,3BPG (Fig 7A). When we applied 40 μM of CGP 3466B maleate, a specific inhibitor of GAPDH [37], WT pollen tubes exhibited a pgkc-1-like phenotype with depolarized morphology [38](Fig 7B, 7C and 7E). Furthermore, when either pgkc-1 or ren1-3 single mutants were treated with CGP, cell polarity defect magnitude was greatly enhanced, exhibiting significantly ballooned pollen tubes (Fig 7B to 7I). To validate that the cellular mechanism underlying the GAPDH inhibition phenotype was similar to that underlying the pgkc-1 mutation phenotype, we observed the distribution of GFP-REN1, CRIB4-GFP, EYFP-RABA4D and Lifeact-mEGFP in WT pollen tubes treated with 40 μM CGP 3466B. Similar to in pgkc-1 mutants, GFP-REN1 signal was diminished while CRIB4-GFP, EYFP-RABA4D and Lifeact-mEGFP signals were enhanced in the apical region after CGP 3466B treatment (Fig 7J to 7M). Similarly, a double mutant of cytosolic GAPDHs, gapc1-1/gapc2-1 [30],and the application of another GAPDH inhibitor, iodoacetate, also resulted in pollen tube phenotypes resembling that of pgkc-1 (Fig 8A to 8D and S6 Fig). These results indicate that GAPDH activity is also involved in the regulation of pollen tube polarity. Taken together, we conclude that glycolysis plays an important role in the regulation of pollen tube polarity by affecting the association of the REN1 RopGAP with exocytic vesicles.
Our findings here clearly demonstrate that cytosolic glycolysis has a novel function in the regulation of cellular signaling, distinct from its conventional housekeeping role in carbon and energy metabolism. The global energy level is important for pollen development and pollen tube elongation [20–23, 31, 39]. In this study, we found that inhibition of mitochondrial respiration using oligomycin resulted in reduced pollen tube length and width. This phenotype is consistent with previous reports, while distinct from the reduced growth polarity induced by the pgkc-1 mutation or GAPDH inhibition (Fig 5A to 5D).
The ethanol fermentation pathway serves, concomitantly with oxidative respiration metabolism, as a bypass route to help maintain metabolic flux and energy supply in pollen tubes [25, 40]. This pathway is also downstream of glycolysis and consists of two key enzymes, pyruvate decarboxylase (PDC) and alcohol dehydrogenase (ADH) [25]. In petunia, the mutation of a pollen-specific PDC2 gene was shown to cause reduced elongation of pollen tubes in the style, leading to a competitive disadvantage relative to WT pollen [41]. However, pollen tube polarity in pdc2 mutants appeared to be normal [41]. Therefore, we believe that pollen tube growth polarity is modulated by a specific regulatory aspect of cytosolic glycolysis rather than glycolysis-dependent respiration or fermentation (Fig 9).
In animal cells, many glycolytic enzymes participate in moonlighting functions, including RNA binding, membrane fusion, cytoskeletal dynamics, autophagy, and cell death [42–46]. Similarly, cytosolic GAPDHs in plants demonstrate nuclear uracil-DNA-glycosylase activity and participate in plant immunity [47]. Here, we demonstrated that glycolytic activity is required for PGKc function in pollen tubes. Moreover, GAPDH, another enzyme in the cytosolic glycolysis pathway, plays a similar role as PGKc in pollen tube polarity. Based on these results, it is more likely that it is the glycolysis pathway which regulates pollen tube polarity, rather than a moonlighting function of a glycolytic enzyme (Fig 9).
The pollen tube polarity defects present in the pgkc mutant have been associated with the over-activation of ROP1, as well as the over-accumulation of F-actin and exocytic vesicles in the tip region. Previous findings have suggested that REN1-based negative feedback globally inhibited ROP1, as ROP1 activity is dependent upon the association of REN1 with exocytic vesicles at the apical plasma membrane [14]. Both ren1 mutation and constitutively active ROP1 (CA-ROP1) expression has been shown to cause ROP1 hyper-activation, leading to F-actin stabilization, apical cortex vesicle accumulation, and pollen tube depolarization [11, 14]. In the pgkc mutant or during treatment with a GAPDH inhibitor, the association between REN1 and the exocytic vesicles is abolished, thus accounting for the observed over-activation of ROP1. Accordingly, tip region over-accumulation of F-actin and exocytic vesicles appears to be attributed to ROP activation (Fig 9).
The oscillation of apical ROP1 activity is regulated by positive and negative feedback via F-actin-mediated exocytosis [7]. Could the aberrant REN1 localization be the consequence of disrupted F-actin in the pgkc-1 mutant, rather than the cause? According to previous studies, if the loss of PGKc activity simply enhance F-actin accumulation, then one may expect overall alteration of pollen tube elongation, rather than polarity. Mutations of F-actin severing factors RIC1 or MAP18, also caused aberrant F-actin overaccumulation in the apical tip of pollen tubes [27, 48]. However, ric1 mutant exhibited enhanced elongation and map18 is defective in growth direction of pollen tubes, while the pollen tube polarity was normal in both cases [27, 48]. Therefore, we interpret disrupted RhoGTPase signaling in the pgkc pollen tubes as a reason rather than consequence of the aberrant cellular activities (Fig 9). Nevertheless, our study does not exclude the possibility that pgkc mutation might directly interrupt other unelucidated cellular processes, which simultaneously affect multiple steps in the feedback loops of RhoGTPase signaling, including REN1 distribution, F-actin dynamics, and exocytic vesicle trafficking.
Several possible underlying mechanisms may link the glycolysis pathway with pollen tube polarity. Mitochondria provide most of the energy required by the cell. However, mitochondria are not evenly distributed in polarized cells, and may not meet the needs of all organelles [49, 50]. In contrast, although net energy gain is low, glycolysis could produce ATP close to energy sinks, thus complementing mitochondrial function. For instance, in neurons, the vesicles in fast axonal transport are energized by on-board ATP provided by specifically localized glycolytic machinery rather than mitochondrial respiration [51]. Mitochondria are absent from the apical tip of pollen tubes, where PGKc is present [50, 52]. Therefore, it is possible that cytosolic glycolysis may provide an ATP source in close proximity to some unclear vesicle activities, similar to the fast axonal transport, which are required for the targeting and/or trafficking of REN1 protein in pollen tube tips (Fig 9).
Glycolysis is a fundamental energy metabolism pathway, but glycolytic enzymes and intermediates may also play important signaling roles in growth and development. One of the most important signaling hubs is the enzyme hexokinase (HXK). As the first enzyme in glycolysis, HXK is able to phosphorylate glucose, producing glucose-6-phosphate [53]. Independent of its catalytic activity, plant HXK has also been proven as a glucose sensor for the regulation of sugar metabolism and signaling pathways [54, 55]. Since PGK and GAPDH are downstream of HXK and aldolase, there is a possibility that PGKc or GAPDH inhibition might cause accumulation of glucose in pollen tube, resulting in a hyperactivation of HXK signaling in pollen tubes (Fig 9). It would be helpful to examine this possibility in the future by overexpressing HXK in pollen tubes.
It is less likely but still possible that glycolysis may regulate pollen tube polarity through signaling by downstream intermediate metabolites, such as 3-phosphoglyceric acid (3PG), a product of PGKs. However, given that the plastidial glycolysis pathway remains intact in pgkc mutant pollen tubes, metabolic intermediates are unlikely to be deficient. Consistent with this, adding 3PG or pyruvate did not affect the pollen tube polarity phenotype of the pgkc mutant, even at concentrations inhibitory to WT pollen tubes (S7 Fig and S8 Fig). Although we could not clarify whether exogenous metabolites could substitute for intracellular metabolic intermediates under our experimental conditions, this result indicates that these metabolites have no effect on pollen tube polarity. Nonetheless, future studies are needed to elucidate the mechanisms by which cytosolic glycolysis regulates the association of REN1 with apical vesicles and subsequent cell polarity modulation in pollen tubes.
Arabidopsis (Columbia ecotype) were used as WT specimens. All plants were grown under a 16 h photoperiod at 22°C. SALK collections of individually indexed homozygous T-DNA insertion lines were obtained from the ABRC (http://signal.salk.edu/cgi-bin/homozygotes.cgi). For in vitro pollen germination screening, 5–10 seeds from each SALK line were grown in individual pots. Pollen grains from three plants for each line were collected and germinated in in vitro germination medium as previously described [14, 56]. Lines with the pollen tube polarity phenotype were selected as mutant candidates for further verification. SALK_066422C (pgkc-1) was identified during screening. Another allele, SALK_062377 (pgkc-2), was obtained from the ABRC. Genotyping was performed based on the protocol provided on the SALK website. gapc1-1/gapc2-1 double mutant seeds were gifts from Dr. Xueming Wang, and genotype was confirmed using primers as described [30]. All primers used are listed in Supplemental S1 Table
Total RNA was extracted from indicated tissues using the E.Z.N.A. RNA extraction kit (Omega) according to manufacturer’s instructions. Oligo dT-primed cDNA was synthesized from 500 mg of total RNA using the PrimeScript RT reagent Kit with gDNA Eraser (Takara). Quantitative PCR analysis was performed with the SYBR Premix Ex Taq II ROX plus kit (Takara) using a Mx3005 device (Agilent). Relative levels of each transcript were calculated after being normalized to UBC21 endogenous control.
All constructs were generated using Gateway technology (Invitrogen). Primers used are listed in Supplemental S1 Table. All entry vectors were generated from the pDONR-zeo vector (Invitrogen). LR reactions were conducted using LR Clonase II (Invitrogen) with corresponding entry vectors and destination vectors.
To construct catalytic inactive mPGKc complementation vector, PGKc cDNA was cloned first. Then DpnI-mediated site-directed mutagenesis was performed to change the G535 to C [35, 57]. pGWB604 vector was modified by inserting a 2.1 kb PGKc promoter with HindIII and SbfI. LR cloning were then performed to generate the proPGKc::PGKc-GFP or proPGKc::mPGKc-GFP constructs, respectively.
Open flowers were collected and pollen grains were dusted onto standard agar-germination medium with 18% sucrose, 0.01% boric acid, 1 mM CaCl2, 1 mM Ca(NO3)2, 1 mM MgSO4, pH 6.0, and 0.5% Difc Noble agar (BD Biosciences). Incubation times ranged from 2 to 9 h at 23°C, and pollen tubes were observed under an Imager M2 inverted microscope (Olympus). Tube length and width were measured using ImageJ software. Since the morphology of pgkc-1 pollen tubes was non-uniform, pollen tube width was measured at the widest point.50-100 pollen grains or pollen tubes were measured.
For pollen tube chemical treatment, LatB (Invitrogen), BFA(Invitrogen), iodoacetate (Sigma), oligomycin (Sigma), CGP 3466B (Tocris Bioscience) at indicated concentrations were added to the above solid pollen germination medium. WT and mutant pollen grains were germinated using the same medium and compared side-by-side.
For Lifeact-mEGFP, EYFP-RABA4D, REN1-GFP marker observation, previously reported marker lines were used for crossing with pgkc-1 mutants. Double homozygote F2 progeny were identified by genotyping for the presence of mutant pgkc-1 and observation of pollen GFP/YFP fluorescent signal, respectively. For the CRIB4-GFP marker, as pgkc-1 and CRIB4-GFP are linked on the same chromosome, a vector containing CRIB4-GFP was used to transform pgkc-1. The chosen line was then backcrossed with WT, and CRIB4-GFP on a WT background was obtained as part of F2 progeny. Fluorescent microscopy was performed with a Spinning Disk Confocal Microscope Andor Revolution WD.
To quantitatively measure GFP signal intensity, the ImageJ line profile tool was used according to user guidelines. Briefly, a five pixel block was drawn from the background toward the tip along the axis of a pollen tube. The signal intensity along the line was measured by the line profile tool. The apical tip was defined as the position where signal intensity was two-fold greater than the black background, and this position was designated as 0 μm. Fifteen to twenty pollen tubes were measured for each sample, with the data from each pollen tube aligned by tip position, and average intensities were calculated.
|
10.1371/journal.pgen.1005542 | EP4 Receptor–Associated Protein in Macrophages Ameliorates Colitis and Colitis-Associated Tumorigenesis | Prostaglandin E2 plays important roles in the maintenance of colonic homeostasis. The recently identified prostaglandin E receptor (EP) 4–associated protein (EPRAP) is essential for an anti-inflammatory function of EP4 signaling in macrophages in vitro. To investigate the in vivo roles of EPRAP, we examined the effects of EPRAP on colitis and colitis-associated tumorigenesis. In mice, EPRAP deficiency exacerbated colitis induced by dextran sodium sulfate (DSS) treatment. Wild-type (WT) or EPRAP-deficient recipients transplanted with EPRAP-deficient bone marrow developed more severe DSS-induced colitis than WT or EPRAP-deficient recipients of WT bone marrow. In the context of colitis-associated tumorigenesis, both systemic EPRAP null mutation and EPRAP-deficiency in the bone marrow enhanced intestinal polyp formation induced by azoxymethane (AOM)/DSS treatment. Administration of an EP4-selective agonist, ONO-AE1-329, ameliorated DSS-induced colitis in WT, but not in EPRAP-deficient mice. EPRAP deficiency increased the levels of the phosphorylated forms of p105, MEK, and ERK, resulting in activation of stromal macrophages in DSS-induced colitis. Macrophages of DSS-treated EPRAP-deficient mice exhibited a marked increase in the expression of pro-inflammatory genes, relative to WT mice. By contrast, forced expression of EPRAP in macrophages ameliorated DSS-induced colitis and AOM/DSS-induced intestinal polyp formation. These data suggest that EPRAP in macrophages functions crucially in suppressing colonic inflammation. Consistently, EPRAP-positive macrophages were also accumulated in the colonic stroma of ulcerative colitis patients. Thus, EPRAP may be a potential therapeutic target for inflammatory bowel disease and associated intestinal tumorigenesis.
| Inflammatory bowel disease (IBD) is one of the most prevalent and serious gastrointestinal diseases in Western countries and associates with cancer development. EP4 receptor signaling can suppress intestinal inflammation and shows promise as a target for the development of novel therapies for IBD. To date, however, the lack of detailed molecular targets has hampered the development of effective drugs. This study focused on EPRAP, a novel EP4 receptor–associated protein, implicated in its signaling pathway. The generation of EPRAP-gene mutated mice permitted exploration of EPRAP functions in vivo. In addition, EPRAP was localized in stromal macrophages of ulcerative colitis patients. This study revealed that EPRAP in macrophage participates critically in EP4 receptor signaling-mediated inhibition of intestinal inflammation. The macrophage EP4–EPRAP axis thus comprises a novel therapeutic target for IBD.
| The incidence of inflammatory bowel disease (IBD) is increasing, and current concepts attribute IBD to inappropriate chronic inflammatory responses to commensal microbes in genetically susceptible patients [1]. Prostaglandin E2 (PGE2) plays a pivotal role in maintaining local homeostasis in a variety of pathophysiological settings. PGE2 receptors (EPs) mediate the effects of this molecule and include four subtypes: EP1–4 [2]. PGE2 participates decisively in the defense of the colonic mucosa. For example, misoprostol, a synthetic PGE1 analog, potently protects human colonic mucosa against mucosal insults [3]. In addition, in mice, PGE2 and EP4-selective agonists significantly improved colitis induced by dextran sodium sulfate (DSS) treatment [4,5]. Yet, misoprostol frequently induces severe diarrhea in humans [6], and treatment with EP receptor agonists can induce vasodilatation, causing hypotension [7]. Indeed, in a phase 2 clinical trial of an EP4-selective agonist for ulcerative colitis patients, diarrhea or hypotension occurred in patients with EP4 agonist treatment [8]. These undesired effects of PGE2 and EP4 agonists likely result from increased cAMP production in colonic or vascular endothelial cells. Hence the need is urgent to develop molecules that stimulate EP4 receptors but have fewer side effects to treat IBD, and potentially other inflammatory diseases.
EPRAP, a cytoplasmic EP4–interacting molecule, emerged from yeast two-hybrid screening, using the C-terminus of EP4 receptor as a bait [9]. EPRAP contains multiple ankyrin repeat motifs, but has no predicted enzymatic or catalytic domain. EPRAP transcripts abound in the heart, skeletal muscle, and the kidney, and localize at lower amounts in several other human tissues [10]. The counterpart of EPRAP in mice is Fem1a, a mammalian ortholog of Caenorhabditis elegans FEM-1. FEM-1 participates in nematode sex determination [11], although the functions of mouse EPRAP/Fem1a and of human EPRAP remain uncertain.
In vitro, EPRAP in macrophages mediates an anti-inflammatory function of PGE2–EP4 signaling [12]; however, the pathophysiological roles of EPRAP in colonic inflammation in vivo remain unknown. In this study, we evaluated the role of EPRAP in the development of colitis and colitis-associated tumorigenesis using EPRAP-deficient mice and macrophage-specific EPRAP-overexpressing mice.
The generation of mice lacking the gene encoding EPRAP enabled investigation of the role of EPRAP in colonic inflammation (S1 Fig). EPRAP-deficient mice were fertile and grew normally without apparent malformation. DSS-induced colitis is a widely used rodent model of human IBD [13]. After the administration of 2.5% DSS for 5 days and regular water for the following 16 days, EPRAP-deficient mice exhibited significantly higher mortality (Fig 1A) and markedly lower body weight (Fig 1B) relative to WT mice. In addition, EPRAP-deficient mice had reduced colon length (Fig 1C and 1D). Histopathologically, EPRAP-deficient mice treated with DSS exhibited more prominent crypt loss, as well as elevated infiltration by inflammatory cells: DSS-treated EPRAP-deficient mice had significantly higher histological damage scores [14] than did WT mice (Fig 1E and 1F). The colons of DSS-treated EPRAP-deficient mice showed significantly elevated accumulation of macrophages, neutrophils, B cells, CD4+ T cells, or CD8+ T cells (Figs 1G and S2A). Furthermore, the colons of DSS-treated EPRAP-deficient mice contained markedly higher levels of pro-inflammatory cytokines and chemokines such as TNF-α, IL-1β, IL-6, CXCL1, and MCP-1 compared to controls (Fig 1H). Markers for classically activated macrophages such as iNOS and CXCL10 also increased in EPRAP-deficient mice (S2B Fig). Thus, the colonic mucosa of EPRAP-deficient mice showed accentuated inflammatory damage caused by DSS treatment.
Colitis-associated tumorigenesis associates closely with the duration and severity of colonic inflammation [15]. Because EPRAP-deficient mice had more prominent colitis induced by DSS, we tested the hypothesis that EPRAP reduces colitis-associated tumorigenesis. WT and EPRAP-deficient mice were intraperitoneally injected with AOM, followed by three cycles of 5-day administration of 2% DSS. After this treatment, EPRAP-deficient mice exhibited a significantly higher mortality rate than WT mice (Fig 1I), with more prominent crypt loss and elevated histological damage scores (S3A and S3C Fig). The colons of AOM/DSS-treated EPRAP-deficient mice showed significantly increased accumulation of inflammatory cells including macrophages, neutrophils, B cells, CD4+ T cells, or CD8+ T cells (S3B and S3D Fig). Concomitantly, EPRAP-deficient mice developed a significantly greater number of colonic polyps at day 63 (Fig 1J and 1K). Furthermore, tumors of EPRAP-deficient mice had more Ki-67–positive (Fig 1L), but fewer TUNEL-positive cells, than those of WT mice (Fig 1M), indicating that EPRAP, either directly or indirectly, suppressed cell proliferation and promoted apoptosis during colitis-associated tumor development.
To determine which EPRAP-deficient cells contribute to the suppression of colitis and associated tumorigenesis, we used four kinds of chimeric mice generated by bone marrow transplantation. Following DSS treatment, WT and EPRAP-deficient recipients of EPRAP-deficient bone marrow exhibited greater colon shortening, crypt loss, content of inflammatory cells, and a higher histological damage score than WT or EPRAP-deficient recipients of WT bone marrow (Fig 2A–2D). Similarly, after AOM/DSS treatment, WT or EPRAP-deficient recipients of EPRAP-deficient bone marrow developed significantly more polyps (Fig 2E and 2F) than recipients of WT bone marrow. These observations indicated that EPRAP-expressing bone marrow cells predominated over epithelial cells in suppression of DSS-induced colitis and of AOM/DSS-induced tumorigenesis. In agreement with these in vivo findings in mice, human colon cancer cell lines with gain or loss of function of human EPRAP (S4 Fig) also showed that EPRAP expression did not affect cell proliferation (Fig 2G).
Normal colonic tissues of WT mice showed negligible immunoreactive EPRAP; however, not epithelial cells but mononuclear cells in lamina propria and submucosal regions in the colonic tissues of DSS- or AOM/DSS-treated WT mice displayed abundant EPRAP (S5A and S5G Fig). Double-color immunofluorescence analyses revealed predominent EPRAP protein in F4/80-positive macrophages (S5B Fig), but not in CD4+ T cells, CD8+ T cells, dendritic cells, B cells, or NK cells (S5C Fig). Neutrophils also contained EPRAP (S5D Fig); however, EPRAP deficiency did not alter myeloperoxidase activity in the colonic tissues of DSS-treated mice (S5E Fig), suggesting that EPRAP deficiency in neutrophils does not play a major role in enhancing colitis and colitis-associated tumorigenesis. We further isolated lamina propria macrophages from DSS-treated mice by flow cytometry, and measured mRNAs that encode pro-inflammatory cytokines and chemokines by quantitative PCR. Lamina propria macrophages of EPRAP-deficient mice exhibited marked increases in the mRNAs corresponding to TNF-α, IL-1β, IL-6, CXCL1, and MCP-1 relative to those of WT mice (S5F Fig), indicating that EPRAP contributes critically to inhibiting macrophage activation and colonic inflammation.
EPRAP contributes to PGE2/EP4-mediated inhibition of inflammation of macrophages in vitro [12]. To determine whether EPRAP alters anti-inflammatory functions of EP4 signaling in vivo, we examined the pharmacological effects of ONO-AE1-329, a selective EP4 agonist, in DSS-induced colitis of WT and EPRAP-deficient mice. Administration of ONO-AE1-329 significantly decreased colon shortening, crypt loss, infiltration of inflammatory cells, and histological damage score in WT mice with DSS-induced colitis (Fig 3A–3D). In contrast, the EP4 agonist did not protect EPRAP-deficient mice from DSS-induced colitis, indicating that EP4 signaling improved colitis through EPRAP. Consistent with the results of previous in vitro experiments [12], administration of EP4 agonist decreased the phosphorylation of MEK and ERK in the stromal macrophages of DSS-treated WT but not EPRAP-deficient mice (S6 Fig). EP4 activation increases intracellular levels of cAMP, a major downstream effector of EP4 signaling. To test the involvement of EPRAP in the cAMP increase resulting from EP4 activation, we measured the levels of cAMP in peritoneal macrophages and colonic epithelial cells obtained from WT and EPRAP-deficient mice, with or without ONO-AE1-329 treatment. Cyclic AMP production induced by the EP4 agonist in both peritoneal macrophages (Fig 3E) and colonic epithelial cells (Fig 3F) did not differ significantly between WT and EPRAP-deficient mice. Furthermore, neither cell type showed a significant difference in the level of EP4 mRNA (Fig 3G and 3H). These data suggested that the anti-inflammatory function of EP4–EPRAP pathway does not involve cAMP increase.
Via a direct interaction, EPRAP suppresses stimulus-induced phosphorylation and subsequent degradation of NF-κB1 p105, resulting in the inhibition of MEK and ERK activation in macrophages [12]. Study of WT and EPRAP-deficient mice affirmed that lacking EPRAP enhanced LPS-induced phosphorylation of p105, MEK and ERK (S7A and S7B Fig), with elevated mRNA production of TNF-α, IL-1β, IL-6, and CXCL1 (S7C Fig) in peritoneal macrophages in vitro. Pretreatment with the MEK inhibitor cancelled the increased expression of those pro-inflammatory molecules, indicating that EPRAP deficiency in macrophages causes unopposed MEK activation and subsequent inflammatory responses induced by LPS (S7D Fig). Indeed, phosphorylation levels of MEK increased significantly in inflamed colons isolated from EPRAP-deficient mice compared to those of WT mice following DSS treatment (Fig 4A). To corroborate EPRAP participation in this pathway in the context of DSS-induced colitis, we assessed the levels of the phosphorylated forms of p105, MEK, and ERK in stromal macrophages by immunohistochemistry and flow cytometry. Double-color immunofluorescence analyses demonstrated more prominent phosphorylated forms of p105, MEK, and ERK in the stromal macrophages of EPRAP-deficient mice than in those of WT mice (Fig 4B and 4C). Flow-cytometric analyses with lamina propria macrophages isolated from colitis lesions yielded similar results (Fig 4D): mean fluorescence intensity (MFI) indicated that the intracellular levels of the phosphorylated forms of each molecule varied inversely with the levels of EPRAP expression in macrophages. These observations suggested that EPRAP suppressed colitis through its interaction with NF-κB1 p105, thereby limiting the activation of the MEK–ERK MAPK pathway in stromal macrophages.
To verify the pivotal role of EPRAP in macrophages, we generated transgenic mice in which the murine CD68 promoter directed murine EPRAP expression (CD68–mEPRAP transgenic mice), leading to overexpression of EPRAP selectively in macrophages (S8 Fig). CD68–mEPRAP transgenic mice were fertile, accumulated body weight normally, and did not develop spontaneous diarrhea. During the course of DSS-induced colitis, WT and CD68–mEPRAP transgenic mice had no significant difference in mortality or body weight loss (S9A and S9B Fig); however, CD68–mEPRAP transgenic mice showed markedly reduced colon shortening (Fig 5A and 5B). Histological examination of the colon obtained from DSS-treated CD68–mEPRAP transgenic mice revealed reduced crypt loss and fewer inflammatory cells than in colons of WT mice (Fig 5C and 5D). Indeed, DSS-treated CD68–mEPRAP transgenic mice, had fewer neutrophils, B cells, CD4+ T cells, and CD8+ T cells as well as macrophages (Figs 5E and S9C), and reduced concentrations of TNF-α, IL-1β, IL-6, CXCL1, or MCP-1 (Fig 5F). As well, after 3 cycles of DSS treatment following intraperitoneal AOM injection, CD68–mEPRAP transgenic mice also exhibited less crypt loss, decreased infiltration of inflammatory cells with lower histological damage scores than WT mice (S10 Fig). In colitis-associated tumorigenesis, in contrast to WT mice, all CD68–mEPRAP transgenic mice survived the AOM/DSS treatment (Fig 5G), and CD68–mEPRAP transgenic mice developed fewer polyps (Fig 5H and 5I). In addition, polyps of CD68–mEPRAP transgenic mice contained significantly fewer Ki-67–positive cells (Fig 5J), but more TUNEL-positive cells, than those of WT mice (Fig 5K). These results indicated that forced expression of EPRAP in macrophages suppressed colitis and colitis-associated tumorigenesis.
Regarding the impact of EPRAP overexpression in macrophages on p105 phosphorylation and MEK–ERK activation, double-color immunofluorescence analyses demonstrated less phosphorylation of these molecules in stromal macrophages of CD68–mEPRAP transgenic mice with DSS treatment than in those of WT mice (Fig 4B and 4C). Flow-cytometric analyses with lamina propria macrophages isolated from colitis lesions yielded similar results (Fig 4D).
To examine the role of EPRAP on human IBD pathogenesis, we immunostained EPRAP in resected colonic specimens of ulcerative colitis (UC) patients. Normal colonic specimens obtained from another cohort showed few EPRAP-positive mononuclear cells; however, mononuclear cells in lamina propria in the colonic tissues of UC patients displayed a large number of EPRAP-positive cells (Fig 6A and 6B). Double-color immunofluorescence analyses suggested those EPRAP-positive cells are stromal macrophages (Fig 6B). Notably, there was an inverse relationship between the percentage of EPRAP-positive cells in stromal macrophages and the disease severity of the UC patients (Fig 6C), suggesting that inflammatory stimuli could induce EPRAP-positive macrophages to mitigate excess local immune responses; and less EPRAP expression in macrophages may cause more severe intestinal inflammation. Accordingly, phosphorylations or activations of p105, MEK and ERK were more prominent in EPRAP-negative mononuclear cells in lamina propria of human UC samples (S11 Fig).These data are consistent with DSS- or AOM/DSS-treated mouse colon, and suggest that EPRAP plays roles in human IBD.
Our previous studies showed that EPRAP, a novel EP4 receptor interactor, mediates the anti-inflammatory actions of EP4 in macrophages [9,12]. This study demonstrated that EPRAP expression is elevated in the stromal macrophages of both mouse and human colitis, and that EPRAP participates critically in suppressing colonic inflammation. Systemic or bone marrow–specific deficiency of EPRAP exacerbated DSS-induced colitis and increased AOM/DSS-induced polyp formation. In contrast, forced expression of EPRAP in macrophages or transplantation of EPRAP-sufficient bone marrow into EPRAP-deficient mice ameliorated these endpoints. Improvement of DSS-induced colitis by EP4 agonist administration required EPRAP, whereas cAMP, a downstream mediator of EP4, did not appear to contribute to the anti-inflammatory effect of EPRAP in the colon.
In mice treated with either DSS or AOM/DSS, EPRAP participated essentially in the regulation of pro-inflammatory gene expression and accumulation of inflammatory cells during mucosal inflammation and tumorigenesis. IBD associates with the development of colonic cancer [16], and EPRAP appeared to suppress colonic tumorigenesis indirectly by attenuating mucosal inflammation.
To determine which cell types are responsible for the suppression of colonic inflammation and tumorigenesis, we first examined cell proliferation assays using human colon cancer cell lines; however, EPRAP expression levels did not affect cell proliferation, suggesting that EPRAP does not play a role in epithelial or tumor cells. Consistent with this, EPRAP was not detectable by immunohistochemistry in colonic epithelial cells or lamina propria cells in the absence of inflammation; once these tissues were inflamed, however, lamina propria and submucosal inflammatory cells such as macrophages expressed EPRAP abundantly. Bone marrow transplantation experiments determined that EPRAP-expressing cells from this source mediated the reduction in bowel inflammation and a pivotal role for EPRAP-expressing cells in the stroma. Systemic or bone marrow-specific deficiency of EPRAP exacerbated colitis and colonic tumorigenesis; in contrast, forced expression of EPRAP in macrophages or restoration of EPRAP expression in bone marrow cells reversed these findings. Thus, EPRAP expression in macrophages mediates the suppression of inflammation and tumorigenesis in the colon.
In DSS-treated WT mice, neutrophils as well as macrophages in the colon expressed EPRAP. In human IBD, activated neutrophils accumulated within epithelial crypts produce reactive oxygen species, which induces oxidative stress and mucosal injury [17]; however, in this study, EPRAP deficiency did not affect myeloperoxidase activity, a marker for neutrophil oxidative stress. These data support the conclusion that EPRAP in macrophages, rather than in neutrophils, is critical for the suppression of colonic inflammation and tumorigenesis.
EP4 signaling attenuates colonic inflammation in mouse models of human IBD [4,5]. EP4 participates not only in reducing accumulation of macrophages but also in suppressing proliferation and activation of CD4+ T cells [5]. This study showed that the protective effect of an EP4 agonist against DSS-induced colitis depended largely on EPRAP. After DSS treatment, inflammatory cells as well as in colonic epithelial cells express EP4 [18]. Downstream of EP4, intracellular cAMP mediates many of the biological effects of EP4 signaling. For example, elevation of intracellular cAMP after EP receptor activation contributes to the defense of the gastrointestinal mucosa by stimulating mucin secretion from epithelial cells and regulating local immune responses [19,20]. Meanwhile, excessive secretion of mucin induced by the cAMP pathway promotes diarrhea. Furthermore, augmentation of intracellular cAMP by EP4 stimulation in vascular endothelial cells induces vasodilatation, evoking hypotension [7]. In this regard, the deletion of EPRAP affected neither the reduction in intracellular cAMP production mediated by EP4 activation nor EP4 expression. These results agree with our previous in vitro studies that showed that the anti-inflammatory effect of EP4–EPRAP in macrophages did not depend on intracellular cAMP [9,12]. Taken together, these data identify EPRAP signaling in macrophages as a target for IBD therapies that would avoid known unwanted actions of EP4 agonism in inflammatory diseases.
This study dissected the molecular mechanisms by which EPRAP suppresses colonic inflammation. Via a direct interaction, EPRAP in cultured macrophages inhibits stimulus-induced phosphorylation and subsequent degradation of NF-κB1 p105, thereby suppressing MAPK activation [12]. Indeed, NF-κB1 p105 is an important scaffold protein in MAPK signaling: p105 binds to tumor progression locus-2 (TPL2), a mitogen-activated protein kinase, and inhibits kinase activity [21]. In response to inflammatory stimulation, IκB kinase (IKK) complex phosphorylates p105, leading to proteolytic degradation of p105 and subsequent release of TPL2 [22]. TPL2 then directly phosphorylates and activates MEK, followed by phosphorylation and activation of ERK. ERK, a MAPK, phosphorylates a number of target proteins, and critically regulates pro-inflammatory activation in colitis and colitis-associated tumorigenesis [23]. Indeed, mice lacking the p105/p50 subunit have heightened susceptibility to colitis induced by Helicobacter hepaticus infection [24]. In addition, previous kinetic analyses using cell lysates from intestinal mucosa of IBD patients indicated marked acceleration of ATP-dependent degradation of p105 occurs in the presence of proteasomes from IBD patients [25]. This study demonstrated that EPRAP deficiency increased the proportions of the phosphorylated forms of p105, MEK, and ERK in stromal macrophages in DSS-induced colitis, whereas forced EPRAP expression had the opposite effect. In particular, EPRAP impaired the activation of MEK–ERK MAPK pathway, likely through p105, thereby decreasing the production of pro-inflammatory cytokines and chemokines in stromal macrophages and ultimately inhibiting colonic inflammation.
In summary, EPRAP in macrophages mediates attenuation of colonic inflammation by PGE2–EP4, and these functions do not depend on cAMP production. EPRAP in macrophages also suppressed tumorigenesis. These data identify EPRAP as a promising target for the treatment of IBD patients.
All animal care and experiments were conducted following the guidelines for the Japan’s Act on Welfare and Management of Animals. The study protocol was approved by the Institutional Animal Care and Use Committees (IACUC)/ethics committee of Kyoto University. All surgery was performed when mice were anesthetized by 40 mg/kg of pentobarbital sodium (Kyoritsu Seiyaku, Tokyo, Japan), and all efforts were made to minimize suffering.
For immunohistochemistry of human samples, surgically resected specimens were obtained from ulcerative colitis (for colitis) or colorectal cancer patients (for surrounding normal colon) who had been admitted to Hyogo College of Medicine or Kyoto University Hospital, respectively. Written informed consents were obtained from all patients with the protocol approved by the Ethics Committee of Hyogo College of Medicine or Kyoto University Graduate School of Medicine in accordance with the ethical guidelines for epidemiological research by the Japanese Ministry of Education, Culture, Sports, Science and Technology and the Japanese Ministry of Health, Labour and Welfare as well as the principles expressed in the Declaration of Helsinki.
Gene targeting in HK3i mouse ES cells derived from C57BL/6 embryos (Acc. No. CDB0852K: http://www.cdb.riken.jp/arg/protocol.html) generated EPRAP/Fem1a-deficient (KO) mice. The targeting vector was based on a modified pBluescript plasmid containing Mc1 DT-A-pA and a floxed Pr-NEO-pA cassette, and was designed to replace the mouse Eprap/Fem1a gene with a selection cassette (S1A Fig). PCR and Southern blot analysis identified homologous-recombinant ES clones. Chimeric mice were generated following aggregation of the targeted recombinant ES cells and transfer to recipient female mice. Chimeric males and C57BL/6 females were mated, and Southern blog analysis verified germline transmission and correct gene targeting. To produce EPRAP-deficient (KO) mice, we intercrossed heterozygous offspring. The primers used to identify WT and deleted alleles in PCR are listed in S1 Table. WT and KO littermates were housed individually until DSS or AOM/DSS treatment.
Mice overexpressing murine EPRAP/Fem1a under the control of the macrophage-specific promoter CD68 [CD68–mEPRAP transgenic (TG) mice] were generated as follows. The CD68–EPRAP/Fem1a transgene consists of the murine EPRAP/Fem1a cDNA downstream of the mouse CD68 promoter, which was derived from pDRIVE-mCD68 (InvivoGen, San Diego, CA, USA). We microinjected the transgene into fertilized C57BL/6 mouse eggs, and three lines were established from six founders. PCR, using a set of two primers (S1 Table), validated transgenic offspring. CD68-positive bone marrow cells were sorted by staining with FITC-labeled anti-CD68 antibody (BioLegend, San Diego, CA, USA), and quantitative PCR measured mouse EPRAP mRNA concentrations in each line. Mice with the highest level of EPRAP expression were bred to C57BL/6 mice to obtain heterozygous TG mice and WT littermates and used for the experiments described in this article.
For DSS-induced colitis, 8–10-week-old male WT, KO, and TG mice were given 2.5% (w/v) DSS (mol wt, 36,000–50,000; MP Biomedicals, Irvine, CA, USA) in their drinking water for 5 days, followed by a recovery period with regular drinking water through the end of the experiment (day 21). Control mice received DSS-free drinking water. For AOM/DSS-induced inflammatory tumorigenesis, 8–10-week-old male WT, KO, and TG mice received intraperitoneal injection of 12 mg/kg AOM (Sigma-Aldrich, St. Louis, MO, USA) before three cycles of DSS administration as described above. All mice were maintained on food and water ad libitum, and were age-matched as well as co-housed for all experiments.
Eight- to ten-week-old male WT and EPRAP-deficient (KO) mice were each divided into two groups, EP4 agonist–treated and vehicle-treated. All four groups of mice (EP4 agonist–treated WT, vehicle-treated WT, EP4 agonist–treated KO, and vehicle-treated KO) received 2.5% (w/v) DSS in their drinking water ad libitum for 7 days [5]. Mice consumed either 100 μg/kg of ONO-AE1-329 (Ono Pharmaceutical Co. Ltd., Osaka, Japan) or vehicle daily via the transanal route for 8 consecutive days, starting the day before DSS administration.
Four- to eight-week-old male C57BL/6 wild type (WT) or EPRAP-deficient (KO) recipient mice were irradiated with X-rays (10 Gy) and injected intravenously with 6–14×106 bone marrow cells derived from femurs and tibias of adult WT or KO mice. Quantitative PCR verified peripheral blood chimerism: circulating blood cells were isolated and EPRAP mRNA expression was determined (see RNA isolation and quantitative PCR).
Rabbit polyclonal antibody against mouse EPRAP was raised by immunization with keyhole limpet hemocyanin–conjugated synthetic peptides corresponding to amino-acid residues 244–260, 330–346, and 629–645 of murine EPRAP. For immunohistochemistry of mouse samples, paraffin-embedded sections were routinely stained with one of the following antibodies: rat anti-F4/80 (Abcam, Cambridge, MA, USA), rat anti–Gr-1 (eBioscience, San Diego, CA, USA), rat anti-B220 (eBioscience), rat anti-CD4 (eBioscience), rat anti-CD8 (Abcam), hamster anti-CD11c (eBioscience), mouse anti-NCAM (Abcam), rabbit anti–phospho-NF-κB p105 (Ser933) (Cell Signaling, Boston, MA, USA), rabbit anti–phospho-MEK1/2 (Ser221) (Cell Signaling), or rabbit anti–phospho-p44/42 MAPK (Thr202/Tyr204) (Cell Signaling), or rabbit anti-EPRAP.
For immunohistochemistry of human samples, paraffin-embedded sections were stained with rat anti-CD68 (Abcam), goat anti-FEM1A (Abcam), and rabbit anti–phospho-NF-κB p105 (Ser933), rabbit anti–phospho-MEK1/2 (Ser221), or rabbit anti–phospho-p44/42 MAPK (Thr202/Tyr204) antibodies. Staining with non-immune rat or rabbit IgG served as a negative control for each experiment.
A validated scoring scheme assessed colitis. In brief, histological scoring was based on three parameters. Inflammation severity was scored as follows: 0, none; 1, mild; 2, moderate; 3, severe. Inflammation extent was scored as follows: 0, none; 1, mucosa; 2, mucosa and submucosa; 3, transmural. Crypt damage was scored as follows: 0, none; 1, basal 1/3 damaged; 2, basal 2/3 damaged; 3, crypts lost, surface epithelium present; 4, crypts and surface epithelium lost. Inflammation severity score, inflammation extent score, and crypt damage score were each multiplied by percent involvement (0, 0%; 1, 1–25%; 2, 26–50%; 3, 51–75%; 4, 76–100%), and the resultant products were summed to yield a histological score ranging from 0 to 40.
Disease severity was determined based on the guidelines for the management of ulcerative colitis in Japan (modification of Truelove and Witts’ criteria) proposed by the Research Committee of Inflammatory Bowel Disease [26]. The severity criteria is listed in S1 Table.
To detect apoptotic cells, paraffin-embedded sections were stained using the ApopTag in situ apoptosis detection kit (Millipore, Billerica, MA, USA). Five different areas per tissue section were analyzed using BZ-H2C (Keyence, Osaka, Japan).
Peritoneal macrophages or colonic epithelial cells were harvested from 8-week-old WT, KO, and CD68–mEPRAP transgenic (TG) mice, and 3×105 cells were seeded into individual wells of a 96-well dish. After 10-minute treatment with different concentration of ONO-AE1-329 or vehicle, intracellular cAMP levels were determined using the Cyclic AMP EIA Kit (Cayman Chemical).
Colonic tissues were removed and homogenized in 1% NP-40 supplemented with protease inhibitor cocktail and PhosSTOP (Roche Diagnostics, Indianapolis, IN, USA). The BD CBA assay (BD Bioscience, San Jose, CA, USA) measured quantitative determinations of TNF-α, IL-6, and MCP-1 concentrations in the supernatants of tissue homogenates. Quantikine ELISA Immunoassay (R&D Systems, Minneapolis, MN, USA) measured IL-1β and CXCL1 concentrations.
The Myeloperoxidase (MPO) Activity Colorimetric Assay Kit (BioVision, Inc., Milpitas, CA, USA) measured MPO activity. MPO activity was compensated by the protein concentration of lysates.
Lysates of colonic tissue extracts from DSS-treated WT and KO mice were assayed at a protein of 0.5 mg/ml or 0.05 mg/ml with the PathScan Signaling Nodes Multi-Target Sandwich ELISA Kit (Cell Signaling) according to the manufacture’s protocol.
Peritoneal macrophages were harvested from 10–12-week-old WT and KO mice and 1.5×106 cells were seeded into individual wells of 48-well dishes. After 1 hour of treatment with 1 μg/ml of lipopolysaccharides (E. Coli 055:B5; LPS) (Calbiochem, San Diego, CA, USA) or vehicle, endogenous levels of phospho-MEK and phospho-ERK were determined at a protein of 0.3 mg/ml using the PathScan Phospho-MEK1 (Ser217/221) Chemiluminescent Sandwich ELISA Kit and the PathScan Phospho-p44/42 MAPK (Thr202/Tyr204) Sandwich ELISA Kit (Cell Signaling) according to the manufacture’s protocol. Immunoblot analyses were also performed using the whole cell lysates of WT and KO peritoneal macrophages. Antibodies against phospho-NF-κB p105 (Ser933), phospho-MEK1/2 (Ser217/221), MEK1/2, phospho-p44/42 MAPK (Thr202/Tyr204), p44/42 MAPK (Erk1/2) were from Cell Signaling Technology: anti-mouse p105/p50 antibody was from Abcom. For quantitative PCR, peritoneal macrophages were treated with or without LPS, followed by total RNA isolation as described in the Materials and Methods. To test the effects of pharmacologic inhibition of MEK, cells were pretreated with U0126 (Cell Signaling) at 10 μM for 1 hour prior to LPS treatment.
Lamina propria mononuclear cells and colonic epithelial cells were isolated as previously described [27]. Briefly, colonic tissues were incubated in HBSS containing 5 mM EDTA and 1 mM DTT at 37°C with shaking, and then filtered through a 100-μm cell strainer (BD Biosciences). Colonic epithelial cells were harvested from the sediment of the flow-through at this step. To isolate lamina propria mononuclear cells, the remaining pre-digested colonic tissues were minced and incubated with PBS containing 500 μg/ml of collagenase D (Roche), 500 μg/ml of DNase I (Sigma-Aldrich), and 3 mg/ml of dispase II (Eidia, Ibaraki, Japan) at 37°C with shaking. After removing all remaining cell clumps by passing the suspension through a 40-μm cell strainer (BD Biosciences), cells were resuspended with 3% (v/v) FCS in PBS. CD45+CD11b+ macrophages were isolated using a FACSAria II flow cytometer (BD Biosciences), following staining of cell suspensions with FITC-labeled anti-CD45 (eBioscience) and PerCP–Cy5.5-labeled anti-CD11b (eBioscience). For analyses of the levels of phosphorylated forms of p105, MEK and ERK, freshly isolated macrophages were fixed in 4% paraformaldehyde/PBS, permeabilized in 0.5% Triton X-100/PBS, and stained with rabbit anti–phospho-p105, anti–phospho-MEK1/2, or anti–phospho-p44/42 MAPK antibody, followed by incubation with PE-conjugated donkey anti-rabbit IgG (eBioscience). Flow cytometry was performed using FACSAria II and analyzed using the FlowJo software (Tree Star Inc., Ashland, OR, USA).
Trizol Reagent (Invitrogen) was used to extract total RNA, Superscript III (Invitrogen) was used to synthesize single-stranded cDNA. The mRNA level for each target gene was determined by SYBR Green–based quantitative PCR using a LightCycler 480 system (Roche). Primer sequences are shown in S2 Table. Data were normalized using GAPDH as a reference gene. All reactions were performed in triplicate.
To knockdown endogenous EPRAP expression in colon cancer cells (DLD-1, ATCC No.: CCL-221), we used the BLOCK–iTPol II miR RNAi Expression Vector Kit with EmGFP (Invitrogen). miRNA-expressing plasmids were constructed according to the manufacturer’s protocol. Target sequences were as follows: miR–EPRAP #1, 5′-ACCAACCGAAAGCTATGCAAG-3′; miR–EPRAP #2, 5′-AAACCAACCGAAAGCTATGCA-3′. We used the pcDNA6.2/EmGFP–miR-negative control vector (miR–NC from Invitrogen) as a negative control. Plasmid transfection experiments were performed using Lipofectamine 2000 reagent (Invitrogen).
Following over expression [12] or knockdown of EPRAP, cell proliferation was measured with the CellTiter 96 AQueous One Solution Cell Proliferation Assay (MTS; Promega, Madison, WI). Briefly, 5 × 103 cells were seeded in 96–well plates and cultured in the growth medium containing 10% fetal bovine serum for 0 h, 48 h, or 72 h. Then, absorbance at 492 nm was measured according to the manufacture’s protocol. All experiments were performed in octuplicate, and the results are shown as mean ± SEM of values.
Data are presented as means ± SEM. Statistical comparisons between groups were made using Student’s t test or one-way ANOVA followed by Tukey-Kramer analysis. P values less than 0.05 were considered statistically significant.
|
10.1371/journal.pcbi.1000770 | A Bayesian Framework to Account for Complex Non-Genetic Factors in Gene Expression Levels Greatly Increases Power in eQTL Studies | Gene expression measurements are influenced by a wide range of factors, such as the state of the cell, experimental conditions and variants in the sequence of regulatory regions. To understand the effect of a variable of interest, such as the genotype of a locus, it is important to account for variation that is due to confounding causes. Here, we present VBQTL, a probabilistic approach for mapping expression quantitative trait loci (eQTLs) that jointly models contributions from genotype as well as known and hidden confounding factors. VBQTL is implemented within an efficient and flexible inference framework, making it fast and tractable on large-scale problems. We compare the performance of VBQTL with alternative methods for dealing with confounding variability on eQTL mapping datasets from simulations, yeast, mouse, and human. Employing Bayesian complexity control and joint modelling is shown to result in more precise estimates of the contribution of different confounding factors resulting in additional associations to measured transcript levels compared to alternative approaches. We present a threefold larger collection of cis eQTLs than previously found in a whole-genome eQTL scan of an outbred human population. Altogether, 27% of the tested probes show a significant genetic association in cis, and we validate that the additional eQTLs are likely to be real by replicating them in different sets of individuals. Our method is the next step in the analysis of high-dimensional phenotype data, and its application has revealed insights into genetic regulation of gene expression by demonstrating more abundant cis-acting eQTLs in human than previously shown. Our software is freely available online at http://www.sanger.ac.uk/resources/software/peer/.
| Gene expression is a complex phenotype. The measured expression level in an experiment can be affected by a wide range of factors—state of the cell, experimental conditions, variants in the sequence of regulatory regions, and others. To understand genotype-to-phenotype relationships, we need to be able to distinguish the variation that is due to the genetic state from all the confounding causes. We present VBQTL, a probabilistic method for dissecting gene expression variation by jointly modelling the underlying global causes of variability and the genetic effect. Our method is implemented in a flexible framework that allows for quick model adaptation and comparison with alternative models. The probabilistic approach yields more accurate estimates of the contributions from different sources of variation. Applying VBQTL, we find that common genetic variation controlling gene expression levels in human is more abundant than previously shown, which has implications for a wide range of studies relating genotype to phenotype.
| DNA microarray technologies allow for quantification of expression levels of thousands of loci in the genome. These measurements enable exploring how a variable, such as clinical phenotype, tissue type, or genetic background, affects the transcriptional state of the sample. Recently, gene expression levels have been studied as quantitative genetic traits, investigating the effect of genotype as the primary variable. Studies have found and characterised large numbers of expression quantitative trait loci (eQTLs) [1]–[3], exploring their complexity [2], population genetics [4], [5] and associations with disease [6], [7].
An important issue in such studies is additional variation in expression data that is not due to the genetic state, as illustrated in Figure 1. Intracellular fluctuations, environmental conditions, and experimental procedures are factors that all can have a strong effect on the measured transcript levels [2], [8]–[10] and thereby obscure the association signal. When measured, correct estimation of the additional variation due to these known factors allows for a more sensitive analysis of the genetic effect. For example, it has been reported that additional human eQTLs can be found when including the known factors of age, and blood cell counts in the model [7]. It is also standard procedure to correct for batch effects, such as image artefacts or sample preparation differences [11].
In practise it is not possible to measure or even be aware of all potential sources of variation, but nevertheless it is important to account for them. Unobserved, hidden factors, such as cell culture conditions [12] often have an influence on large numbers of genes. We and others have proposed methods to detect and correct for such effects [9], [13], [14]. These studies demonstrated the importance of accounting for hidden factors, yielding a stronger statistical discrimination signal.
The challenge in modelling several confounding sources of variation (Figure 1) is to correctly estimate the contribution that is due to each one of them. There are open questions how to ensure that only spurious signal is eliminated by methods that account for hidden factors (see for instance discussion in [14]), and how to deal with situations when both known and hidden factors are present. The problem of identifying the correct causes of the signal is even harder in the presence of additional sources of variability. For example, when searching for epistatic or genotype-environment interactions, the primary effects of other known factors and hidden factors also need to be accounted for.
The key for correctly attributing expression variability is controlling the complexity of the statistical models for each source of variation. For example, the number of genotypes considered in an association scan can be enormous, and not all of them affect the expression level of every probe. Threshold values, obtained from likelihood ratio statistics or empirical p-value distributions, can be used to determine the significance of individual associations, thereby avoiding overfitting by controlling the model complexity [4], [15]. Similar measures are necessary for models of other sources of variability such as hidden factors.
In this work we present VBQTL (Variational Bayesian QTL mapper), a joint Bayesian framework for gene expression variability that accounts for the signal from genotype, known factors, and hidden factors. VBQTL is implemented within a general framework that provides commonly used models for sources of phenotypic variation, which can be combined as needed. While previous attempts have been specific to a narrow set of underlying sources, our approach is flexible and can be adapted to a particular study design. The probabilistic treatment allows uncertainty to be propagated between models, and yields a posterior distribution over model parameters. Complexity control is tackled at the level of individual models, where parameters are regularised in a Bayesian manner.
We compare the performance of VBQTL with existing approaches for detecting expression QTLs. A simulation experiment contrasts VBQTL with common approaches that use non-Bayesian techniques for distinguishing global hidden factor effects from genetic effects. This study highlights differences in the methodology to control model complexity with implications to eQTL detection power. The necessity and difficulty to account for variability that confounds the genetic signal is demonstrated. Results on datasets from a human outbred population and crosses of inbred yeast and mouse strains show that VBQTL identifies more significant associations than alternative methods. Finally, we apply VBQTL to perform a whole-genome eQTL scan on the HapMap phase 2 expression and genotype data, demonstrating the scalability of our framework to large numbers of samples and probes. We find three times more cis eQTLs than a standard association mapping method, suggesting more extensive genetic control of gene expression by common variants than previously shown.
Here, we present VBQTL, a configuration of a general framework for modelling diverse sources of gene expression variability. The model underlying this framework assumes that gene expression levels are influenced by additive effects from independent sources, e.g. in the case of VBQTL these are contributions from genotype, known factors, and hidden factors (Figures 1, 2a). We cast the full model in a probabilistic setting, treating its parameters as random variables.
We perform Bayesian inference in the joint model, which is appealing for several reasons. First, it allows possible dependencies between the different sources of variation to be captured. The effects of the genotype, known and hidden factors are learned jointly, taking other parts of the model into account. Propagation of uncertainty leads to more accurate parameter estimates [16], and avoids possible pathologies, for instance of maximum likelihood methods [17]. Second, Bayesian inference allows different models to be flexibly combined according to the needs of a particular study. Many existing approaches can be cast as special cases of this general framework, with some examples given in Figure 1. Finally, the Bayesian approach leads itself to efficient approximate inference schemes such as variational methods [18], rendering the resulting algorithms applicable to large-scale and high-dimensional datasets. Also, variational learning allows an inference schedule to be specified by the user, leading to distinct algorithms with different computational complexity and properties (see Inference).
In the following, we present the mathematical model of VBQTL, and an outline of the inference procedure. We then describe alternative non-Bayesian models for expression QTL studies used in the experiments. An in-depth treatment of the framework including full details about the parameter estimation is provided in Text S1.
The observed gene expression matrix for genes and individuals is modelled by the sum of contributions from the genotype, known and hidden factor models and Gaussian noise with precisions for each gene (1)with a gamma prior on the noise precisions (Figure 2a). The comprise the contribution of individual sources to the variability in the observed expression levels, and are themselves treated as random variables with different underlying models.
1) Genotype effect model represents the probabilistic variant of the standard genetic association model, where some of the SNP genotypes have a linear effect on gene expression levels. The genetic component of the expression level of the th gene probe in the th individual is explained by linear effects of the genotypes of SNPs (Figure 2a, green plate):(2)(3)(4)The weights control the magnitude of the effect of the SNP on the expression levels of genes . The binary variables determine whether the SNP effect is significant () or not (). The prior probability of an individual association controls the complexity of the model by influencing the a priori expected number of significant associations; this parameter corresponds to a significance threshold in a classical setting (Text S1).
To reduce the computational cost, inference in the association model is approximated, only considering a single most relevant SNP-regulator per gene, with the other forced to . This bottleneck approximation ensures tractability of the joint association model for large-scale studies (Text S1), avoiding the need to track the covariance between effects from multiple SNPs.
2) Known factor model accounts for the effect of known covariates of individual samples, such as environmental conditions, gender, or a population indicator. The linear effects of measured covariates in the th individual, , is taken into account using a variant of Bayesian regression (Figure 2a, blue plate):(5)(6)(7)Here, is the corresponding weight vector for each gene . The gamma prior on the inverse variance for weights of each factor introduces automatic relevance detection (ARD) [19], [20], driving the weights of unused factors to and thereby switching them off. This provides complexity control of the model by regularising the effective number of covariates.
3) Hidden factor model accounts for the effect of hidden factors (such as unmeasured covariates and global effects) on the gene expression levels. We use a probabilistic variant of the classical factor analysis model for this task. We have previously shown that this model captures hidden factors better than alternative linear models, such as probabilistic principal component analysis or independent component analysis [13]. Similarly to known factors, the expression level of gene in individual is modelled by linear effects from a chosen number of hidden factors (Figure 2a, red plate).(8)(9)(10)(11)Note that in contrast to the known factor model, the factor activations are unobserved random variables that need to be inferred from the expression profiles. Again, the ARD prior switches unused factors off, thereby providing probabilistic complexity control ([13], Results).
Parameter inference in VBQTL is implemented using variational Bayesian learning [18], a generalisation of the expectation maximisation algorithm. An approximate -distribution over model parameters is iteratively refined until convergence. In each iteration, approximate distributions of individual parameters are updated according to a specified schedule, taking the current state of all other parameter distributions into account (Figure 2b–e). Choosing an approximation that factorises over individual models, the variational update equations have an intuitive interpretation:
This iterative procedure, performing updates of local parameter distributions in turn, can be interpreted as a message passing algorithm, where sufficient statistics of parameter and data distributions are propagated across the graphical model [21].
The initial values of parameters are determined from maximum likelihood solutions. A random initialisation via sampling from the prior is possible as well; we have not explored the implications of this alternative here. Details on inference and the individual parameter update equations are given in Text S1.
In experiments, we compare two alternative inference schedules of VBQTL. In iterative VBQTL (iVBQTL), the model parameters are learned using several iterations through all model components, first updating the genetic model, then known and hidden factors (Text S1). An important property of iVBQTL is that hidden factors are estimated jointly with the genetic state and known factors. This choice of schedule and the iterative learning help to ensure that variability that is due to genetic associations is not explained away by other parts of the model (Results).
In cases where neither known nor hidden factors are correlated with the genetic state, their effect can be learned independently without running the risk of explaining away meaningful association signal. This motivates fast VBQTL (fVBQTL), which performs a single update iteration of the full model, first inferring the contribution from the known and hidden factors, and then from the genetic state. This simpler schedule can save significant computation time, since the factor effects can be precalculated, and only a single iteration of the computationally more expensive genetic association model is needed. In cases where the genetic state is approximately orthogonal to the known and hidden factors, this cheaper approximation performs equally with iVBQTL for finding genetic associations (Results).
We compared VBQTL with previous methods that account for confounding variance in the context of expression QTL mapping. Similarly to VBQTL, they model known and hidden factors in the expression levels. The differences between the alternative methods are in the hidden factor model used, which in turn vary in the complexity control approach employed as highlighted below. Thus these alternative models are named after the hidden factor estimation method.
For a quantitative evaluation of the performance of each method, we considered the resulting residuals of the estimated effects from known and hidden factors. To detect eQTLs we applied standard statistical tests employing a linear model on the SNP genotype on these residual datasets (Text S1). For iVBQTL and fVBQTL, we inferred the posterior parameter distributions, and subtracted off the estimated effect of known and hidden factors. For other methods, we first subtracted off the standard linear regression fit of the known factors, and then learned and subtracted off the hidden factor effects on the residuals. All these alternative methods are also implemented in the general framework; for details see Text S1.
While VBQTL shares basic assumptions with these alternatives, there are a number of differences. First, it is a probabilistic model that operates with uncertainties in the parameter estimates as explained above. Second, the hidden factor model allows for non-orthogonal components, and provides probabilistic complexity control based on ARD. Third, the iVBQTL schedule takes the genetic signal into account when estimating the hidden factor effect. Finally, the VBQTL model estimates a global gene-specific noise level, while the non-Bayesian models either estimate noise levels implicitly (SVA) or assume noise-free observations (PCA, PCAsig).
We employed a simulated dataset to highlight the differences between alternative approaches to account for global factors in eQTL finding. Our synthetic expression data combines linear effects from genetic associations (eQTLs), known, hidden, and genetic global factors, and gene-specific noise (Text S1). We used three known and seven unknown global factors whose influence varies significantly to simulate effects with a range of magnitudes. These factors are meant to represent sources of confounding variation that are encountered in the study of the real datasets. We also introduced three global genetic factors giving rise to trans eQTL hotspots, mimicking the action of a genetic variant in a transcriptional regulator (e.g. transcription factor or pathway component). Such loci have been observed in several eQTL mapping studies [1], [3]. We designated three genes with a simulated eQTL as such regulators, and simulated correlated expression levels for 15% of the genes for each. While the specific simulation scenario may be biased in the comparative performance of different methods, its underlying linear model is shared by all the considered approaches, and it gives intuition for the results on real datasets discussed later.
Next, we compared the same methods for expression QTL finding on yeast [2], mouse [3] and human [4] datasets. These represent common study designs of an outbred population (human), and a population of crosses between inbred strains (yeast, mouse). We considered 5, 15, 30, and 60 hidden factors for PCA and VBQTL, and , and as significance cutoffs for SVA and PCAsig. Expression QTLs were detected using a two-sided t test analogously to the simulation scenario. Again, results for alternative genetic association tests were similar (Figures S2, S3, S4).
Motivated by the results of the initial study of a single human chromosome, we applied fVBQTL, learning 30 hidden factors, to the 10,000 most variable expression probes of the HapMap 2 dataset. We searched for cis eQTLs in the original expression data (standard eQTLs) as well as the residuals of fVBQTL (VBeQTLs), using a 2-tailed t test with Bonferroni-corrected per-gene FPR to assess the significance of association.
On the CEU population, we found 1051 genes with a VBeQTL at false discovery rate (FDR) of , and 382 genes with a standard eQTL at FDR of (Figure 5). This result corresponds to nearly a threefold increase in the number of genes with an association, and is consistent across chromosomes. A similar increase in the number of associations was found for other populations (Table S1).
We repeated this genome-wide experiment on pooled populations. Due to the increased sample size, it was possible to detect additional associations. We found 2696 genes with a VBeQTL compared to 1045 genes with a standard eQTL at the 0.1% FPR (Figure 6a). The VBeQTLs in the pooled sample cover of all the considered probes, suggesting that the number of human genes whose expression levels are affected by common cis-acting genetic variation may be significantly higher than previously shown [24], [25]. This additional abundance of associations suggests that detection of cis eQTLs has not been saturated and larger sample sizes may lead to evidence of even more extensive cis regulation by common polymorphisms.
Exploratory results indicate additional power to find trans eQTLs without explaining away eQTL hotspots (Text S2). These should be interpreted with caution due to very stringent multiple testing corrections, however.
It is important to demonstrate that the additional associations found after removing the learned non-genetic factors are biologically meaningful. We provide evidence that the additional associations found in HapMap phase 2 data are real in three ways.
First, we investigated how many of the genes with a VBeQTL in each of the three populations individually were replicated using the standard method on a pooled data set containing all populations. Note that this will only validate weak associations that occur in multiple populations – we would not expect weak population-specific associations to be replicated in the pooled data set. However, we expect many of the associations to be replicated in multiple populations [24]. A total of of all and of the additional associations found in the CEU population were recovered using the standard method in the pooled population (Figure 6b). The remaining additional associations may be explained by even weaker signals that were recovered by applying fVBQTL, or as population-specific effects that do not stand out in the pooled sample. Analogous overlaps were found when excluding the CEU population from the pooled analysis (Table S3).
Second, we evaluated to what extent the additional genes with a VBeQTL in a single population were replicated in other populations. For instance, of genes with a CEU VBeQTL were replicated on the YRI population (Figure 6d), and on the CHB+JPT population (Figure 6e). These overlaps are consistent with overlaps of standard eQTLs, and are similar for other populations (Table S2), and alternative methods accounting for hidden factors.
Finally, we validated that the locations of the novel associations are distributed similarly to the original ones. We analysed the distribution of the position of additional cis associations around the gene start along with the association LOD scores. The additional VBeQTLs have very similar characteristics to the standard eQTLs, being concentrated around the gene start (Figure 6c, 6f), in line with previous results [24].
The hidden factor models hypothesise a set of unobserved non-genetic factors that influence the measured gene expression levels. To gain insights into their interpretation we considered correlations to known effects such as gender, population or environment, and the sets of genes most influenced.
We applied fVBQTL to expression data from individuals of all three HapMap populations, and tested for correlation between the inferred hidden factors and the population and gender indicator variables. The resulting correlation coefficients (Table S4) indicate that many of the learned latent causes are correlated with population and that one is strongly correlated with gender. This implies that the hidden factor model can recapture variance in the gene expression levels due to true underlying properties of individuals. However, none of the global factors learned in one population was correlated with a single SNP genotype.
A recent study in yeast looked for changes in eQTLs when segregating strains were grown in different media [26]. We applied fVBQTL to the expression data of this study (GEO accession GSE9376), without including any information about the growth condition. The first hidden factor learned was highly correlated with the indicator variable for the growth condition (), demonstrating that the VBQTL model can successfully recover an environmental effect if it is present.
The global factors identified can be further analysed for biological signals, looking for GO term over-representation in the genes that they affect. We used the ordered GO profiling method [27] to find significantly enriched GO categories for 30 genes most affected by each factor. Recent results [28] show that related linear Gaussian models find biologically relevant factors in the yeast expression dataset. We replicated these findings with our model, yielding factors enriched in biological functions, including sugar, alcohol and amino acid metabolic processes. Similar analysis in human and mouse did not show significant over-representation of GO categories, providing no evidence that the main axes of variation in the expression levels for these experiments are due to common biological function. This could be due to poor annotation of the genes, gene features not related to biological function, or more technical sources of global variation, such as cell culture conditions [12].
We have presented VBQTL, a probabilistic model to dissect gene expression variation in the context of genetic association studies. The model is implemented in a Bayesian inference framework that allows uncertainty to be propagated between different parts of the model, and yields posterior distributions over parameter estimates for more sensitive analysis. In comparative eQTL mapping experiments, VBQTL outperformed alternative methods for eQTL finding on simulated and real data. In the most striking example, VBQTL found up to three times more eQTLs than a standard method, and 45% more compared to the best alternative in the HapMap 2 expression dataset.
Our approach advances the methodology for understanding phenotypic variation. The implementation of a flexible framework allows models for explaining the observed variability to be straightforwardly combined. Notably, non-Bayesian models can also be included, as we demonstrated with PCA, SVA, and linear regression models. VBQTL controls the model complexity at the level of all individual components of expression variability, thereby preventing from over- and underfitting. Our experimental results on simulation and real data showed how explaining away too much variability removes some signal of interest from the data, and failing to account for all sources of confounding variation decreases power to detect the relevant signal. When the variable of interest is correlated with many gene expression levels, its effect can be falsely explained away by the hidden factor model. We showed that in such settings the choice of an iterative schedule helps to ensure that variability is explained by the appropriate part of the model. There can be no silver bullet solution that provides perfect results in any scenario with no supervision. Instead, modelling assumptions must be made explicit, and incorporated in the analysis, as is elegantly done in the Bayesian setting.
VBQTL and other methods that account for hidden factors all found additional expression QTLs in the datasets studied compared to the standard method. It is remarkable that, with only 270 samples, and looking in one tissue type, we can find significant genetic associations to of the expressed genes. While similar results have been reported before, we have shown a threefold increase in the number of associations for the HapMap dataset, and analysed their repeatability and location distribution. The replication of the additional associations in different populations suggests that they are genuine. The increase in power is due to the hidden factor model, which explains away unwanted non-genetic variability, thereby allowing the genetic effects to stand out to a greater extent. The high number of additional associations suggests that association finding studies in human have not saturated, and we expect the fraction of genes with an eQTL will increase further as the number of samples grows. It may be that the expression of majority of human genes varies as a result of segregating genetic variation. While previous studies have reported only 12% of heritable variation to be due to cis variants [29], this does not contradict the presence of weak cis eQTLs for a large fraction of the genes.
In conclusion, we believe that VBQTL provides a principled and accurate way to study gene expression and other high-dimensional data. Increasingly complex models combining genetic and other effects can explain significantly more of the variance in observed phenotypes, as suggested by this study and others. Our general framework provides the flexibility to facilitate these richer models, for example, we have already started exploring interaction effects as an additional model of the framework. It will be interesting to see how these approaches can contribute to our understanding of human disease genetics, potentially involving intermediate phenotypes such as gene expression and other factors.
The software used in this study is freely available online at http://www.sanger.ac.uk/resources/software/peer/.
|
10.1371/journal.pcbi.1002488 | Uncovering MicroRNA and Transcription Factor Mediated Regulatory Networks in Glioblastoma | Glioblastoma multiforme (GBM) is the most common and lethal brain tumor in humans. Recent studies revealed that patterns of microRNA (miRNA) expression in GBM tissue samples are different from those in normal brain tissues, suggesting that a number of miRNAs play critical roles in the pathogenesis of GBM. However, little is yet known about which miRNAs play central roles in the pathology of GBM and their regulatory mechanisms of action. To address this issue, in this study, we systematically explored the main regulation format (feed-forward loops, FFLs) consisting of miRNAs, transcription factors (TFs) and their impacting GBM-related genes, and developed a computational approach to construct a miRNA-TF regulatory network. First, we compiled GBM-related miRNAs, GBM-related genes, and known human TFs. We then identified 1,128 3-node FFLs and 805 4-node FFLs with statistical significance. By merging these FFLs together, we constructed a comprehensive GBM-specific miRNA-TF mediated regulatory network. Then, from the network, we extracted a composite GBM-specific regulatory network. To illustrate the GBM-specific regulatory network is promising for identification of critical miRNA components, we specifically examined a Notch signaling pathway subnetwork. Our follow up topological and functional analyses of the subnetwork revealed that six miRNAs (miR-124, miR-137, miR-219-5p, miR-34a, miR-9, and miR-92b) might play important roles in GBM, including some results that are supported by previous studies. In this study, we have developed a computational framework to construct a miRNA-TF regulatory network and generated the first miRNA-TF regulatory network for GBM, providing a valuable resource for further understanding the complex regulatory mechanisms in GBM. The observation of critical miRNAs in the Notch signaling pathway, with partial verification from previous studies, demonstrates that our network-based approach is promising for the identification of new and important miRNAs in GBM and, potentially, other cancers.
| Several recent studies have implicated the critical role of microRNAs (miRNAs) in the pathogenesis of glioblastoma (GBM), the most common and lethal brain tumor in humans, suggesting that miRNAs may be clinically useful as biomarkers for brain tumors and other cancers. However, to date, the regulatory mechanisms of miRNAs in GBM are unclear. In this study, we have systematically constructed miRNA and transcription factor (TF) mediated regulatory networks specific to GBM. To demonstrate that the GBM-specific regulatory network contains functional modules that may composite of critical miRNA components, we extracted a subnetwork including GBM-related genes involved in the Notch signaling pathway. Through network topological and functional analyses of the Notch signaling pathway subnetwork, several critical miRNAs have been identified, some of which have been reinforced by previous studies. This study not only provides novel miRNAs for further experimental design but also develops a novel computational framework to construct a miRNA-TF combinatory regulatory network for a specific disease.
| Glioblastoma multiforme (GBM) is the most common and lethal primary brain tumor in humans and is classified as a grade IV astrocytoma by the World Health Organization (WHO) [1]. The tumor is characterized by rapid growth, a high degree of invasiveness, and strong resistance to radiation and chemotherapy [2]. To illuminate its complex characteristics, an understanding of the underlying genetics is critical. During the last decade, numerous genetic studies, including microRNA (miRNA) and mRNA expression profiling, somatic mutation, copy number variation and methylation studies performed by the Cancer Genome Atlas (TCGA) project, and genome-wide association studies (GWAS) by other groups, have substantially contributed to the comprehensive profiling of GBM [3]–[6]. In addition to confirming previous findings, such as TP53 mutation, NF1 deletion or mutation, and EGFR amplification, these results included several new genetic discoveries such as frequent mutations of the IDH1 and IDH2 genes in secondary GBM [3]. Most importantly, these studies support the idea that many of the current risk factors are likely coordinated at the biological pathway or network level rather than at an individual molecular level [6]. Several studies have interrogated networks in the context of gene expression profiles and/or protein interactions to identify novel critical genes and core pathways for GBM, which provides us with new insights into the mechanisms of the disease pathology [7]–[10]. Another important type of biological network, a miRNA-transcription factor (TF) regulatory network, acts as a functional unit in the regulation of cell fate in many cell types and systems, including cancer [11], [12], but this type of network has not yet been systematically investigated in GBM.
In recent years, an increasing number of miRNAs have been identified and linked to cancer [13], [14]. miRNAs are small (∼22 nucleotides) non-coding RNAs that mainly regulate gene expression at the post-transcriptional level in animals [14]. They are involved in cellular development, differentiation, proliferation, apoptosis and tumorigenesis [15], [16]. Similar to other types of cancer, patterns of differential miRNA expression versus normal tissues have been identified for GBM [17]–[19]. For example, several studies consistently confirmed the overexpression of miR-21 in GBM [20]–[24], and several miRNAs are weakly expressed compared with the normal brain, including miR-124, miR-7, and miR-128 [18], [24].
In addition to traditional low-throughput studies, the TCGA project assessed the expression of 534 miRNAs in 240 tumor tissue samples and 10 normal tissue samples. The results have been used to establish GBM subclasses [25], identify miRNA expression signatures to predict GBM patient survival [17], and identify important miRNAs in GBM [26]. These and other studies have made it clear that miRNAs play important roles in GBM, and it appears increasingly likely that miRNAs will be clinically useful as biomarkers and/or therapeutic targets for brain tumors and other cancers [19]. Despite a number of miRNAs reported to be dysregulated in GBM, little is known about which miRNAs play critical roles in the pathology of GBM and their relevant targets [27]. To address these questions, we hypothesized that an investigation of miRNAs in the context of the regulatory transcriptional and post-transcriptional networks will provide a far more comprehensive view of their functional roles in GBM.
TFs regulate gene expression by translating cis-regulatory codes into specific gene-regulatory events [28]. Since TFs and miRNAs are both categorized as gene-regulatory molecules and share a common regulatory logic [29], they are capable of cooperatively regulating the same gene: TFs regulate a gene's transcription in the gene's promoter region, while miRNAs regulate a gene's post-transcription in the gene's 3′ untranslated region (UTR). At the network level, it has been demonstrated that the regulation of transcription by TFs and post-transcriptional regulation by miRNAs are tightly coupled [30], [31]. Moreover, the examination of regulatory networks showed that TFs, miRNAs and genes form a combination of transcriptional/post-transcriptional feed-forward loops (FFLs), which comprise over-represented motifs in the mammalian regulatory network [30], [31]. Therefore, the analysis of mixed FFLs in a cellular system has emerged as a powerful tool to understand specific biological events, such as the control of cell fate in many cell types and systems [11].
In a regulatory network, a typical mixed FFL motif contains three components: TF, miRNA and gene. This mixed FFL motif is defined as a 3-node FFL. Considering co-expressed genes may have similar regulation patterns [32], [33], i.e., genes regulated by the same TF and the same miRNA, we hypothesized that inclusion of co-expressed genes in FFL analysis would have more power to detect disease-specific regulatory modules. Accordingly, we extended the 3-node FFL model to a 4-node FFL model, which might complement to the former.
Here, we pursued a regulatory network-based approach for a comprehensive investigation of gene regulation patterns in GBM. This method can be used to identify network modules containing known GBM-related miRNAs and genes. It can also be used to reveal new components for core pathways. Among GBM candidate genes, we identified the potential targets of TFs and GBM-related miRNAs. These datasets and their regulations were used to construct a comprehensive GBM-specific miRNA-TF mediated regulatory network. Furthermore, we constructed the subnetwork from one well-known core pathway in GBM, the Notch signaling pathway, and identified miRNA components involved in it. Based on the network topological analysis and functional analysis, we identified six functionally critical miRNAs in this pathway. Among them, four have been implicated in GBM by previous work. These results demonstrated that the comprehensive GBM-specific miRNA-TF mediated regulatory network contains valuable information for GBM investigators to identify critical miRNAs and their targets for further experimental design, providing further understanding of the regulatory mechanisms of GBM.
One major purpose of this study was to develop an integrative framework for the construction of a comprehensive regulatory network for GBM. This network consisted of feed-forward regulation among three components: GBM-related genes, GBM-related miRNAs and known human TFs. GBM-related genes and miRNAs with evidence of involvement in the pathology of GBM were collected and curated from public databases and literature. For GBM-related genes, we restricted our analyses to the 415 genes with mutation evidence in previous studies (Table S1 and Text S1). For GBM-related miRNAs, we collected 124 mature miRNAs that were reported to be dysregulated in studies assessing miRNA expression only in GBM tissue samples or cell lines. Human TFs were extracted from TRANSFAC Professional (release 2011.4) [34], a manually curated database of eukaryotic TFs, their genomic binding sites and DNA binding profiles. There are five types of regulatory relationships: TF regulation of gene expression (TF-gene) or miRNA expression (TF-miRNA), miRNA repression of gene expression (miRNA-gene) or TF expression (miRNA-TF), and gene-gene coexpression (gene-gene). Each of these regulatory relationships was predicted using computational approaches (Table 1). Considering the disadvantage of these reverse engineering methods, we applied stringent parameters in prediction to obtain high confidence regulations.
To integrate these regulations into a miRNA-TF regulatory network, we only included FFLs with significant miRNA-TF pairs, pinpointed by the hypergeometric test, that potentially cooperate in regulating the same targets. Based on the combinatory regulatory network, we performed further analyses of the network topological properties and functional associations to identify critical miRNAs (see Figure 1 for the framework and the Materials and Methods for details). It is necessary to point out that, in this computational framework, a novel FFL model (4-node model) was developed for the construction of the regulatory network. To illustrate that the framework has a promising application in cancer investigation, in this study, we focused on the GBM regulatory network and identified the miRNA components for the Notch signaling pathway. The analyses illustrated the framework is promising for further identification of critical miRNAs in the pathology of cancer.
Table 1 summarizes the five types of potential regulatory relationships mentioned above and their related methods. We provide more details below.
FFLs have been demonstrated as one of the most common types of transcriptional network motifs [42]. Typically, a FFL consists of three components: a miRNA, a TF, and a joint target, which is defined as a 3-node FFL. In this study, we expanded the 3-node FFL model to a 4-node FFL model to explore more regulatory modules. Figure 2A shows the detailed relationships in these FFLs. According to the regulatory relationship between two regulators (TF and miRNA) in each FFL, we classified FFLs into 3 types: TF-FFL, miRNA-FFL and composite FFL (Figure 2). Specific to the 3-node FFLs, the TF-FFL model includes TF regulation of a miRNA and a gene, and it also includes miRNA repression of a target gene. The miRNA-FFL model includes miRNA repression of both a target gene and a targeted TF, as well as TF regulation of a target gene. The composite-FFL model includes TF regulation of both a miRNA and a target gene, as well as miRNA repression of the TF gene and the target gene. The three types of FFLs are exclusive to each other. For 4-node FFLs, the design is similar to the 3-node FFL model, but each TF or miRNA may regulate both co-expressed genes.
Furthermore, we merged those FFLs with the same TF-miRNA regulation. Thus, the merged FFLs composed of a known TF, a mature miRNA, and a list of GBM-related genes or a list of GBM co-regulated gene pairs (Figure S1). Table 2 summarizes the number of nodes and links in the 3-node and 4-node FFLs.
After converging the significant 3-node and 4-node FFLs identified in the previous subsection, we constructed a miRNA-TF mediated regulatory network for GBM, the major biological output of our computational analysis. The resultant network contained a total of 4,354 edges and 408 unique nodes (Table S7). Among the 4,354 edges, 1,033 belonged to miRNA-gene pairs, 550 to miRNA-TF pairs, 1,863 to TF-gene pairs, 804 to TF-miRNA pairs, and 104 to gene-gene pairs. Among the 408 nodes, 176 belonged to GBM-related genes, 99 to GBM-related miRNAs and 142 to human TFs. Among GBM-related genes and TFs in this regulatory network, 9 genes overlapped (ARNT, FLI1, FOXO3, FOXO4, GATA3, SMAD4, STAT3, TCF12, and ZEB1). Although the network only recruited 176 (43.46%) of the 415 GBM-related genes and 99 (79.84%) of the 124 GBM-related miRNAs, given the uncertainty of associations between candidate genes and the disease, we regarded it as a representation of the regulatory network in GBM.
To provide a general view of this regulatory network, we calculated degrees (connectivity) and their distribution, which are basic topological network measures [47]. In this complicated network, degree values of genes, miRNAs and TFs ranged from 2 to 66, 2 to 77, and 2 to 123, respectively. The average degrees of genes, miRNAs and TFs were 18.70, 24.11, and 23.80, respectively. The degree distribution for genes, miRNAs and TFs were strongly right-skewed, indicating that most nodes had a low degree, while only a small portion of nodes had a high degree (Figure S5). Therefore, we observed only a few miRNAs, GBM-related genes and TFs exhibited a high degree in the network. In the context of this regulatory network, these molecules act as hubs that might play important roles in GBM.
Hubs are highly connected nodes in a network, suggesting critical roles in maintaining the overall connectivity of the network [47]. Consistently, hubs in the PPI network are more likely to be essential genes [48], [49]. Using the hub definition method proposed by Yu et al. [50], we determined the degree cutoff values 38, 49, and 71 for genes, miRNAs and TF hubs, respectively. Accordingly, we identified 15 hub genes (FOXO3, SMAD4, TCF12, BCL11A, PDGFRA, KLF4, NRAS, SOX11, CACNA1E, ELAVL2, PIK3R1, RPS6KA3, SLC9A2, CYLD, and PTCH1), 4 hub miRNAs (miR-9, let-7i, miR-495 and miR-130a) and 6 hub TFs (TEAD1, SP1, MZF1, NEUROD1, GATA1, and TCF7). Among them, genes PIK3R1 and PDGFRA had been reported to have high mutation frequencies in 91 GBM samples (9% and 13%, respectively), and are involved in the RTK/PI3K signaling pathway, a core GBM pathway [6].
In the above FFL analyses, we noticed that composite 3-node and 4-node FFLs recruited the most GBM-related genes in each category (49.67% and 72.73%, respectively), which indicated that composite-FFLs could play important roles in regulating GBM candidate genes. Therefore, we converged these composite-FFLs and generated a regulatory subnetwork that only included composite-FFLs. The resulting subnetwork included 457 edges and 101 GBM-related genes, which accounted for 57.38% of GBM-related genes (176) in the GBM-specific miRNA-TF mediated regulatory network and were regulated by only 26 GBM-related miRNAs (24.24%) and 24 TFs (16.90%). We defined this subnetwork as the composite miRNA-TF regulatory network in GBM; it could provide a main framework for the regulatory systems involved in GBM (Figure 3A). In this regulatory network, the distribution of all nodes was again strongly right-skewed; that is, only a few nodes had high degree in the network (Figure 3B). Using the same method to define hubs, we identified four hub genes (NRP1, FOXO3, SMAD4, and TNFRSF1B), six hub miRNAs (miR-495, miR-9, miR-137, miR-30d, miR-181c, and miR-30e), and three hub TFs (TEAD1, SP1, and ZBTB7A). Previously, Zhang et al. [51] proposed that a higher-order network structure is a frequently observed motif in integrated mRNA-protein networks. In our regulatory network, we also found several miRNAs and TFs involved in higher-order subnetworks. For instance, we identified three higher-order composite subnetworks. The first one (Figure 3C) included one hub TF (SP1) and one hub miRNA (miR-137), which together regulated 10 genes. The second composite subnetwork included one TF, one hub miRNA, and 6 genes (Figure 3D). The third one included one hub TF, two hub miRNAs, and 12 genes (Figure 3E).
We further examined enriched pathways in these 101 GBM-related genes involved in the GBM composite regulatory network. This further examination was important, as biological pathways that are statistically enriched in a set of disease genes may provide important cellular process information for our understanding of the molecular pathology of the disease. For the 101 genes, we identified 39 pathways that were significantly enriched (adjusted P-value<0.01) (Table 3). Among these 39 pathways, 10 (25.6%) were directly related to cancer, including glioma and GBM. Several are well-known core pathways involved in GBM, such as PTEN signaling, PI3K/AKT signaling and Notch signaling.
To demonstrate that the GBM-specific miRNA-TF mediated regulatory network is useful to identify miRNA components for core pathways, we took a convergent strategy to narrow down the candidate list. We first generated subnetworks for core pathways in GBM and then performed network characteristic analyses, including degree and degree distribution, hub, network modularity, to identify key components. Aside from degree of the node and degree distribution and hub definition mentioned before, the most frequently used approach for biological network analysis is to cluster or partition the whole network into subcomponents, i.e., modularity. Previous studies have revealed that highly connected groups of proteins tend to participate in the same biological process or complex [52]. In this study, we selected the Notch signaling pathway as an example to illustrate that the network is a useful resource for hypothesis generation and that our computational framework is promising.
The Notch signaling pathway strongly influences stem cell maintenance, development and cell fate [53]. Growing evidence indicates it plays a key role in cancer, including gliomas [54], [55]. According to pathway information recorded in the KEGG database [56] and Ingenuity Canonical Pathways (http://www.ingenuity.com/), there were five genes in the GBM miRNA-TF mediated regulatory network that belonged to the Notch pathway: EP300, NOTCH1, NOTCH2, FURIN, and JAG1. We generated a subnetwork for these 5 genes by merging the FFLs that included at least one of these five genes (Figure 4A). We defined it as the GBM Notch-specific miRNA-TF regulatory network, which included 222 edges, 17 GBM-related genes, 32 GBM-related miRNAs and 31 TFs. These 32 miRNAs might be involved in the Notch signaling pathway, providing a potential pool for further experimental determination of miRNAs involved in this pathway (Table S8). We noticed that there was no 4-node FFL involved in the GBM Notch-specific regulatory network.
To identify the critical candidates from the above 32 miRNAs, we further evaluated their importance based on network topological and functional analyses. The degree distribution of all nodes in this subnetwork was also strongly right-skewed. Using the same method to identify the hubs above, we identified four GBM hub genes (NOTCH1, FURIN, NOTCH2, and EP300), four hub miRNAs (miR-9, miR-92b, miR-137 and miR-295-5p) and four hub TFs (EP300, SP1, TEAD1, and TBX5). Thus, the network global property analysis indicated that these four hub miRNAs might play important roles in the Notch signaling pathway.
To investigate other miRNAs in the GBM Notch-related miRNA-TF regulatory network, we used the software CFinder [57] to identify tightly connected subnetworks. CFinder is a popular network analysis tool for examination of nodes' distributions in networks and communities. We obtained four communities in the Notch regulatory network. The first one (Figure S6A) included 15 GBM-related genes, 14 GBM-related miRNAs and 18 TFs. Since the subnetwork included the most GBM-related genes (88.2%) involved in the GBM Notch related regulatory network, we called this subnetwork the gene-centered subnetwork. The second community (Figure S6B) includes two GBM-related genes, 17 GBM-related miRNAs and 15 TFs. Since most of the nodes in this subnetwork are regulators, we defined it as the regulator-centered subnetwork. The third one includes one GBM-related gene, two miRNAs, and three TFs (Figure S6C); the last one includes one GBM-related gene, one miRNA and one TF (Figure S6D). Considering that the last two subnetworks had one common GBM-related gene, JAG1, and both were located in the center of the Notch-specific network, we merged these subnetworks together and defined it as a centered subnetwork (Figure 4). Consequently, three Notch-specific subnetworks were identified (Figure 4B, 4C, and 4D).
The centered subnetwork included 8 nodes, none of which belonged to the hubs we identified above. When the centered subnetwork was removed, the connection between the other two subnetworks was lost (Figure S7). To further examine this feature, we removed the nodes directly linked to the centered subnetwork; most parts of the Notch regulatory network were loosely connected except among GBM-related genes (Figure S8). These local network analyses showed that the centered subnetwork could serve as a bridge subnetwork and play an important role in the development of GBM. To further examine the role of the centered subnetwork, we used a GO enrichment analysis to identify biological processes associated with the three subnetworks. The gene-centered subnetwork mainly corresponded to the development processes. The centered subnetwork corresponded to regulation of biological processes and developmental processes. The regulator-centered subnetwork corresponded to regulation of biological processes and metabolic processes. These functional association analyses revealed that the centered subnetwork could play the central role in this subnetwork. Based on the important role of this centered subnetwork in the Notch-specific pathway, and two miRNAs, miR-124 and miR-34a, which have direct connections with two other subnetworks, we proposed that these two miRNAs might play important roles in the Notch signaling pathway involved in GBM.
In summary, based on the network topological analysis of the GBM Notch regulatory network and its subnetworks, we identified 32 human miRNAs that might be involved in the Notch signaling pathway, and six of them (miR-124, miR-137, miR-219-5p, miR-34a, miR-9, and miR-92b) might play important roles in this pathway.
In this study, we explored the combinatory regulation of miRNAs and TFs that have an impact on genes involved in the pathology of GBM. We developed a computational framework to construct and analyze a regulatory network for complex diseases. Our framework started with a compilation of numerous data sources to identify disease candidate genes and miRNAs and then inferred regulatory relationships using a large panel of computational tools. Based on these relationships, we focused on 3-node FFLs and 4-node FFLs to generate a GBM-specific regulatory network. This unique computational framework illustrated that it is indeed possible to process multiple types of data (e.g., mutation data, gene expression data, and knowledgebase) by combining a large collection of methods to identify potential miRNAs in complex diseases.
A significant concern regarding the computational approaches used in this study is controlling false positives from both public databases and prediction results caused by computational tools. In our framework, to minimize the effect of these false positives, we first performed a comprehensive compilation from multiple data sources to identify genes and miRNAs relevant to GBM. Next, we chose the most popular databases and software to conduct the prediction. Finally, we applied stringent parameters in the prediction of TF-gene/miRNA, miRNA-gene/TF, and gene-gene relationships. For TF-gene/miRNA and miRNA-gene/TF, we further required conservation among multiple mammalian genomes. Thus, our framework could potentially detect the most important regulatory relationships and might be applied to other complex diseases for the purpose of deciphering their regulatory systems and identifying critical miRNAs.
Compared to high-throughput and low-throughput experimental methods that have been used to discover and profile miRNAs, our computational framework could complement them and facilitate the discovery of critical miRNAs in the pathology of disorders. As much more regulatory data is expected to be released in the near future, such as ChIP-Seq (chromatin immunoprecipitation sequencing), RNA-Seq (transcriptome sequencing) and GRO-Seq (global run-on sequencing), this framework could be improved with the integration of high-throughput data by filtering out interactions in low confidence.
One important output of this comprehensive study is the GBM-specific miRNA-TF combinatory regulatory network. The regulatory network was massive and complex, presenting us with another challenging task: finding the tactic to decipher this huge network to mine the important regulatory components. Recently, pathway analysis has been reported as a useful approach to investigate the pathology of complex diseases [6], [58]. Specifically in our work, our strategy was to apportion the large regulatory network and extract relatively small but functionally critical subnetworks for pathways that have been previously implicated in the corresponding disease. We then performed network topology analyses and investigated modularity to identify critical miRNAs in these small subnetworks.
To demonstrate this strategy, we used the Notch signaling pathway as an example and found six critical miRNAs in the pathway in GBM (Figure 4). Among them, miR-34a has already been shown in an independent study led by one of the authors in 2009 (B.P.) to be down-regulated in GBM, target Notch family members, and cause differentiation in GBM stem-like cells [59], [60]. Additional studies have shown that this miRNA has been involved in the Notch pathway in other cancers such as medulloblastoma [61], pancreatic cancer [62] and carcinoma [63]. Moreover, miR-124 and miR-137 have functioned in a tumor-suppressive fashion in GBM and caused differentiation when re-expressed in GBM cells [24]. miR-9 has also been strongly linked to GBM subtypes in a recent analysis [25]. Interestingly, miR-124 has been reported to be involved in the Notch signaling pathway during Ciona intestinalis neuronal development [64]. The evidence from these studies suggests the effectiveness of our approach. Further experimental validation of these miRNAs is warranted.
Among the six miRNAs, the most noteworthy one is miR-34a. It regulates a number of target proteins that are involved in cell cycle, apoptosis, differentiation and cellular development [65]. In the independent study mentioned above, led by one of the authors (B.P.), the effects of miR-34a on MET, NOTCH1, NOTCH2, CDK6, and PDGFRA expression in brain tumor cells and stem cells were tested. The results showed that miR-34a suppressed brain tumor growth by targeting MET and Notch [66]. To check if these results exist in our predicted regulatory network, we further extracted miR-34a FFLs and merged them to form a miR-34a-specific regulatory network (Figure S9). Among 15 miR-34a targets, 8 (NOTCH2, MET, PDGFRA, JAG1, MYCN, BCL2, DCX, and CACNA1E) belonged to GMB-related genes and 7 (FOSB, FOSL1, NFE2L1, NR4A2, SMAD4, TCF12, and YY1) belonged to human TFs. Among the 8 GBM-related genes, NOTCH2 and MET have been reported in our previous study to be targeted by miR-34a, while PDGFRA was not [66]. JAG1 has been reported to be targeted by the miRNA in the regulation of human monocyte-derived dendritic cell differentiation [67]. MYCN has been reported to be targeted by miR-34a in neuroblastoma cells [68], [69] and somatic cell reprogramming [70]. BCL2 has been reported to be targeted by the same miRNA in neuroblastoma cells [69]. All 7 targeted TFs were significantly involved in the transcription of DNA according to Biology Function Analysis in IPA (Ingenuity Pathway Analysis) (Fisher's exact test, P-value = 8.75×10−9) as expected. Among them, YY1 has been reported to be directly targeted by miR-34a in neuroblastoma cells [71]. Taken together, miR-34a is likely not only regulates GBM-related genes directly but also regulates the TFs for gene expression through transcriptional mechanism. This assertion needs further experimental confirmation. While our analyses, especially of miR-34a and its targets, support the utility of our regulatory network framework, it still needs to be improved. Most GBM-related genes have not been confirmed to be causal, the human TF and miRNA binding profiles are neither complete nor error- or bias-free, and reverse engineering software has its own weaknesses.
This work represents the first application of a 4-node FFL as a regulatory motif in complex disease. Although there have been several genome-wide studies applying integrative regulation of TFs and miRNAs [30], [72], [73], none have considered gene coexpression profiles in an FFL model. The 4-node FFL model contains four components: one miRNA, one TF, and two co-expressed genes related to GBM (Figure 2). There are four types of possible regulations between the co-expressed genes and the TF and miRNA, making the regulatory network more informative and tolerant (Figure S10). Compared with 3-node FFLs, the main impact of 4-node FFLs is the recruitment of more GBM-related genes and regulatory relationships into the regulatory network (Table 2, Figure S3, and Figure S4). We found that 4-node FFLs tended to regulate the genes that might belong to the same biological processes, the same protein family, or be located in the same cellular components (Figure S2). Additionally, among the 20 GBM-related genes involved in the miR-34a-specific regulatory network, 3 were in the 3-node FFLs and 4-node FFLs, 11 from 4-node FFLs, and 6 from 3-node FFLs. This observation indicated that the recruitment of GBM-related genes in miR-34a network was greatly improved by applying the 4-node FFLs. In summary, our comparison of the 4-node and 3-node FFLs and the performance in the recruitment of GBM-related genes by the 3-node FFLs and 4-node FFLs to the miR-34a-specific regulatory network indicate that both are useful models, and they may complement each other in a regulatory network analysis.
Another interesting observation in this study is composite-FFLs, in which TF and miRNA regulate each other. The regulation between a TF and a miRNA has been defined as a TF↔miRNA feedback loop [38]. In our study, we observed 40 TF↔miRNA feedback loops in 3-node FFLs and 24 TF↔miRNA feedback loops in 4-node FFLs. Among the two sets of feedback loops, there were 19 loops in common between two sets, resulting in 45 unique feedback loops in the whole regulatory network for GBM. Compared to the 759 unique TF-miRNA regulatory relationships and the 505 miRNA-TF regulatory relationships in the regulatory network, the TF↔miRNA regulatory relationships were rarely observed. This low frequency is consistent with previous reports involving a pure transcriptional regulatory network [42]. However, interestingly, these TF↔miRNA feedback loops regulate 101 GBM-related genes, accounting for 57.38% of the GBM-related genes (176) in the GBM miRNA-TF mediated regulatory network. This observation indicated that composite-FFLs are more effective in unveiling the regulatory systems underlying the complex disease.
To collect genes involved in the pathology of GBM, we compiled GBM-related genes from six sources, which included multiple types of variations with experimental evidence, such as point mutation, gene fusion, structure rearrangement, and copy number variation. These sources included the Catalogue Of Somatic Mutations In Cancer (COSMIC, version 51) [74], the Online Mendelian Inheritance in Man (OMIM) [75], The Cancer Genome Atlas (TCGA) [6], and the Genetic Association Database (GAD) [76], as well as one recently published integrative genomic analysis of GBM [39] and two genome-wide association studies [4], [5] (Text S1). We mapped these genes to Entrez gene symbols and ultimately obtained 415 unique genes.
To collect a set of dysregulated miRNAs in GBM, we conducted a comprehensive literature search to identify studies that directly assess miRNA dysregulation in GBM patients' cell lines or tissues. We first searched the miR2Disease [77], PhenomiR [78] and HMDD [79] databases for relevant articles using the keyword “glioblastoma” and PubMed using the keywords “glioblastoma AND microRNA.” Then, we manually checked each title and abstract for relevance and reviewed the full text if the abstract indicated that the article reported associations between miRNA expression and GBM. As a result, we included 24 papers that directly assessed miRNA expression in GBM samples or cell lines. From these papers, we retrieved 134 miRNAs with up/down-regulated information, which were mapped to 124 unique mature miRNAs based on human miRNAs from miRBase [80].
Currently, several online databases that predict binding sites and target genes of individual miRNAs are available, such as PicTar [81], TargetScan [35], [82], and miRanda [83]. Among them, TargetScan has demonstrated the best performance compared to other miRNA target prediction software [84], [85]. Therefore, we extracted the miRNA-gene pairs between GBM-related miRNAs and GBM-related genes from the TargetScan server (version 5.2, February 2011) [35]. We required that miRNA-target interactions be evolutionarily conserved in four species (human, mouse, rat and dog) and have a total context score higher than −0.30 [86]. The score quantitatively measures the overall target efficacy [84], [85]. To obtain the posttranscriptional repression of miRNAs on TFs, we first retrieved 428 TFs that have human genes as targets from the TRANSFAC Professional database (release 2011.4) [34] and used the same procedure to obtain the relationships between miRNAs and TFs.
To predict the regulatory relationship between TF and gene/miRNA, we first downloaded the defined promoter region (−1500/+500 around TSS) of 415 GBM-related genes or 134 GBM-related miRNAs from the UCSC Table Browser [87]. Then, we performed a binding sites search using the Match™ software that is integrated in TRANSFAC Professional (release 2011.4) [36]. For the purpose of this study, we used pre-calculated cut-offs to minimize false positive matches (minFP) and create a high-quality matrix. To restrict the search, we required a core score of 1.00, a matrix score of 0.95, and TFs that only belong to the human genome. To further reduce false positive prediction, we required the predicted pairs to be conserved among humans, mice and rats [73].
Recently, Verhaak et al. [39] integrated the gene expression data from 200 GBM and two normal brain samples examined by three gene expression microarray platforms (Affymetrix HuEx array, Affymetrix U133A array, and Agilent 244 K array) into a single, unified data set of 11,861 genes using a factor analysis model. Then, they filtered the unified genes down to 1,740 genes with consistent but highly variable expression across the platforms using several filters to eliminate unreliably measured genes. We directly applied the resulting data to identify co-regulated genes. Among the 415 GBM-related genes we collected, 120 were included in the 1,740 genes. We estimated co-regulated relationships among these genes via the ARACNE software, which implemented the mutual information (MI) theory to identify transcriptional interactions between genes [40]. We used a high significance threshold for MI values with a P-value of 1.0×10−7 to sort out possible false positive and true negative data. To remove indirect regulatory relationship, we employed a data process inequality (DPI) tolerance of 0.15 according to the recommendation by Margolin et al. [88].
To identify TF and miRNA pairs that cooperatively regulate the same target genes, we calculated a P-value using a cumulative hypergeometric test [43] based on the common targets of any pair of miRNAs and TFs as in the following function:where is the number of genes targeted by a given miRNA, is the number of genes regulated by a given TF, and Total is the number of common genes between all human genes targeted by human miRNAs and all human genes regulated by all human TFs. We further used the false discovery rate (FDR) to adjust for multiple testing [89], and only those pairs with a corrected P-value less than 0.05 were chosen as significant pairs of regulators.
To quantify functional similarity, we calculated GO semantic similarity scores for the GO terms for each pair of the co-regulated genes using the R GOSemSim package [44]. For each of the three GO categories (BP: biological process, MF: molecular function, and CC: cellular component), the semantic similarity scores were computed for all gene pairs in the 3-node and 4-node FFLs. A gene pair was compiled from any two genes targeted by the same miRNA-TF pairs. To evaluate the statistical significance of the functional similarity of co-targeted genes in FFLs, we randomly selected the same number of genes in 3-node or 4-node FFLs from the 20,441 Entrez protein-coding genes with GO annotations, and calculated their GO similarities. We repeated this process 1,000 times. We performed a Kolmogorov-Smirnov test (KS-test) to examine whether the GO similarity of all the gene pairs from the FFLs is significantly greater than that of randomly selected pairs.
In this work, we constructed three major networks. The first network was the GBM-specific miRNA-TF mediated gene regulatory network, which was generated by converging all significant 3-node and 4-node FFLs. The second one was the GBM composite regulatory network generated by merging only those significant 3-node and 4-node composite-FFLs. The third one was the subnetwork for the Notch signaling pathway. We first collected the genes belonging to the Notch pathway from the KEGG and Ingenuity systems and merged those FFLs that included at least one Notch pathway gene to generate a Notch-specific regulatory network in GBM.
Considering the complexity of regulatory networks and our goal of distilling critical elements, we simplified the network analysis by disregarding the direction of the edges. We computed nodes' degrees and their distributions in order to assess network characteristics. The degree of a node, the network's most elementary characteristic, is measured by the number of links of the node in the network. If the degree distribution of one network follows a power law, the network would have only a small portion of nodes with a large number of links (i.e., hubs) [47]. To determine the hubs in our network, we applied the method proposed by Yu et al. [50] to draw a degree distribution for each node in the network. For local network analysis, we used the software CFinder (version 2.0.5) [57] to generate tightly connected sub-networks from the pathway network, and we then visualized them using Cytoscape (version 2.8) [90].
To identify pathways overrepresented in GBM-related genes from the GBM composite regulatory network, we performed a pathway enrichment analysis using the Core Analysis Tool in Ingenuity Pathway Analyses (IPA) from Ingenuity Systems [68]. Given a list of genes, a right-tailed Fisher's exact test was performed for the enrichment of these genes based on its hand-curated canonical pathway database. To control the error rate in the analysis results, IPA also provided a corrected P-value based on the Benjamini-Hochberg method [89]. GO and KEGG enrichment of the subnetworks was analyzed using WebGestalt [46].
|
10.1371/journal.pgen.0030151 | Evidence for Transgenerational Transmission of Epigenetic Tumor Susceptibility in Drosophila | Transgenerational epigenetic inheritance results from incomplete erasure of parental epigenetic marks during epigenetic reprogramming at fertilization. The significance of this phenomenon, and the mechanism by which it occurs, remains obscure. Here, we show that genetic mutations in Drosophila may cause epigenetic alterations that, when inherited, influence tumor susceptibility of the offspring. We found that many of the mutations that affected tumorigenesis induced by a hyperactive JAK kinase, HopTum-l, also modified the tumor phenotype epigenetically, such that the modification persisted even in the offspring that did not inherit the modifier mutation. We analyzed mutations of the transcription repressor Krüppel (Kr), which is one of the hopTum-l enhancers known to affect ftz transcription. We demonstrate that the Kr mutation causes increased DNA methylation in the ftz promoter region, and that the aberrant ftz transcription and promoter methylation are both transgenerationally heritable if HopTum-l is present in the oocyte. These results suggest that genetic mutations may alter epigenetic markings in the form of DNA methylation, which are normally erased early in the next generation, and that JAK overactivation disrupts epigenetic reprogramming and allows inheritance of epimutations that influence tumorigenesis in future generations.
| It is well known that many genetic mutations in oncogenes or tumor suppressors can cause or greatly increase a person's susceptibility to cancer. It is generally assumed that persons should feel relieved if they have not inherited the particular “cancer-causing” mutation carried by their parents. However, we found that, under certain circumstances, fruit flies carrying tumor suppressor gene mutations can pass the increased tumor risk to all offspring, even those that have not inherited the particular mutation. A likely scenario is that many genetic mutations can lead to epigenetic alterations, that is, changes in the chemical modifications of DNA or the proteins that bind to DNA in the chromosomes, and these changes can have global effects on cell function. Normally, these epigenetic alterations are wiped out and reset in the early embryo, but under certain circumstances such alterations can be inherited. Interestingly, we found evidence that a particular oncoprotein, an overactivated form of a cell-signaling molecule called JAK kinase, can counteract the epigenetic resetting program that normally operates in the early embryo. Thus, the failure of epigenetic reprogramming allows the inheritance of parental epigenetic alterations that affect susceptibility to tumors.
| Epigenetic regulation of gene expression refers to repression or activation of gene expression via covalent modifications of DNA or histones, such as methylation or acetylation, without changing the DNA sequence of the gene [1–3]. Epigenetic modifications are usually stably heritable through subsequent cell divisions, resulting in permanent changes in gene expression profiles, such as those associated with terminal differentiation. However, at critical stages in normal development or disease situations, cells undergo genome-wide epigenetic reprogramming, erasing preexisting epigenetic marks and establishing a new set of marks. For instance, major epigenetic reprogramming occurs at fertilization prior to zygotic development, at dedifferentiation that leads to cancer development, and during somatic cell nuclear transfer, a procedure used for cloning or obtaining embryonic stem cells [4–7].
However, epigenetic marks are not always completely erased from one generation to the next. For instance, genomic imprinting, where clusters of genes or whole chromosomes are preferentially inactivated depending on their parental origin [8,9], can be considered an exception to epigenetic reprogramming, because in this case parental epigenetic markings are retained in the zygote. Loss of imprinting has been shown to increase the likelihood that cancer will develop [10–12]. Furthermore, human diseases, such as Prader-Willi and Angelman syndromes [13] and hereditary nonpolyposis colorectal cancer [14], are associated with germline inheritance of epimutations. Though transgenerational epigenetic inheritance has been documented for a variety of eukaryotic organisms ranging from plants to humans [15], the precise mechanisms that regulate epigenetic marking and erasure, as well as those that protect certain epigenetic marks from being reset, are not clear.
We have previously undertaken a genetic approach in order to identify genes that are important for hopTum-l-induced tumorigenesis in Drosophila, and in the process, have found that JAK signaling globally counteracts heterochromatin formation [16]. Further analyses of the identified mutations indicated that a number of those mutations that genetically modify hopTum-l tumorigenicity also do so epigenetically. In fact, hopTum-l itself plays an essential role in the maintenance of parental origin epigenetic alterations that subsequently affect tumorigenesis in a transgenerational manner. These results indicate a novel function for the hopTum-l oncogene: it interferes with the epigenetic reprogramming process.
We previously conducted a genetic screen for modifiers of the hopTum-l hematopoietic tumorigenic phenotype and identified 37 modifier mutations [M(Tum-l)] that dominantly enhanced or suppressed hopTum-l tumorigenesis in hopTum-l/+; M(Tum-l)/+ transheterozygotes [16]. Hematopoietic tumors in hopTum-l-containing flies were quantified by tumor index (TI) (see Materials and Methods and also [16]). Interestingly, many of the M(Tum-l) mutations (24/37) exhibited paternal-effect modification of hopTum-l tumorigenicity, such that when hopTum-l/+ females were mated to males heterozygous for the modifier mutation (M[Tum-l]/+), tumorigenesis associated with hopTum-l was modified (enhanced or suppressed) in the F1 generation regardless of the inheritance of M(Tum-l) (Table 1). The transgenerational effects were confirmed with rebalanced stocks, indicating that they are unlikely to be due to different genetic backgrounds. Since little or no paternal cytoplasmic proteins are carried in the sperm, the observed paternal effects on the zygote suggest an epigenetic mechanism. Possibly, the M(Tum-l) mutations caused epigenetic alterations in the paternal chromosomes and these epigenetic changes were maintained through male meiosis and transmitted to the F1 generation, thereby influencing hopTum-l tumorigenicity.
To understand the nature of the transgenerational epigenetic modification of hopTum-l tumorigenicity by the M(Tum-l) mutations, we conducted a detailed analysis of Kr, which is one of the first zygotically transcribed “gap” genes whose activity is required for the correct segmentation of the embryo [17]. First, we tested two loss-of-function alleles of Kr (Kr1 and Kr2) and found that they both enhanced hopTum-l genetically and epigenetically (Figure 1A; unpublished data), confirming Kr as an E(Tum-l) with epigenetic effects.
To rule out any genetic background effects, we extensively outcrossed a Kr1 allele, and isogenized and rebalanced it over a CyO balancer chromosome that in previous testing showed no enhancement of hopTum-l (see Materials and Methods). The new iso-Kr1/CyO stock again enhanced hopTum-l tumorigenicity both genetically and epigenetically, such that when hopTum-l/+ females were crossed to iso-Kr1/CyO males both hopTum-l/+; Kr1/+ and hopTum-l/+; +/CyO progeny exhibited significantly higher TI (Figure 1B, columns 2 and 3). Interestingly, when F1 males of +/Y; +/CyO, which did not inherit Kr1, were backcrossed to hopTum-l/+ females, we found that the enhancement persisted in the F2 generation in the absence of Kr1, but diminished in the F3 (Figure 1B, columns 4 and 5). Since half and a quarter of the P0 paternal DNA contents (originally exposed Kr1) are inherited in the F2 and F3 generation, respectively, the diluting effect of the enhancement in the absence of the original mutation (Kr1) is consistent with the idea that the modification is epigenetic in nature and is distributed genome wide at multiple loci. To rule out the possibility that Kr1 induced genome-wide genetic mutations, we conducted the reciprocal cross, mating iso-Kr1/CyO females with rare escaper hopTum-l/Y males. We found that Kr1 enhanced hopTum-l only genetically but not epigenetically, such that the TI increased in hopTum-l/+; Kr1/+ but not in hopTum-l/+; +/CyO female progeny flies (Figure 1C). The result of the reciprocal cross confirms that the modification is epigenetic in nature, as genetic mutations (changes in DNA sequence) would not be reversible under normal circumstances. However, such a result could also suggest a parent-specific effect of Kr on the hopTum-l mutation.
To test whether the epigenetic effects of Kr1 are specific for the male genome, we mated hopTum-l/+; Kr1/CyO recombinant females to wild-type males. In this cross, the tumor phenotype associated with hopTum-l was enhanced in both hopTum-l/+; Kr1/+ and hopTum-l/+; CyO/+ progeny flies (Figure 1D), indicating that the presence of Kr1 in the female parent can also have epigenetic effects on hopTum-l tumorigenicity in the F1 generation. Thus, it appeared that Kr1 was capable of epigenetically altering both male and female genomes, and these alterations could be transmitted through both male and female meioses to the F1. However, the inheritance and/or ability of these parental origin alterations to modify hopTum-l tumorigenicity epigenetically in the F1 progeny appeared to depend on the presence of hopTum-l as a maternal mutation.
To further test the ability of maternal hopTum-l to maintain parental origin epigenetic changes, we examined the effects of histone deacetylase (HDAC) inhibitors on hopTum-l tumorigenicity. Since Rpd3, encoding an HDAC, was identified as one of the genes which, when mutated, exhibited both genetic and epigenetic enhancement of hopTum-l tumorigenicity (Table 1), we reasoned that the epigenetic effect of an Rpd3 mutation on hopTum-l tumorigenicity might be mimicked by HDAC inhibitors such as tricostatin A (TSA) and sodium butyrate. Indeed, TSA treatment caused increased levels of acetylated histone H3 (Figure 1E), and increased the tumor index of hopTum-l/+ flies from 0.38 to 0.96 ± 0.06 (p < 0.01). Consistent with a transgenerational epigenetic effect, when wild-type flies that had been treated with TSA were mated with untreated hopTum-l/+ females and the progeny were raised in the absence of the drug, the TI of hopTum-l/+ F1 progeny was also significantly increased (Figure 1F). As with Kr1, no epigenetic effect was found in the reciprocal cross (Figure 1F), suggesting that the presence of hopTum-l in the early embryo is important for TSA treatment to have a transgenerational epigenetic effect on hopTum-l tumorigenicity. A similar transgenerational epigenetic effect on hopTum-l tumorigenicity was also found with another HDAC inhibitor, sodium butyrate (unpublished data).
To investigate the maternal hopTum-l-dependent transgenerational inheritance of epigenetic changes at the level of gene expression, we examined the effects of hopTum-l on Kr-dependent expression of the pair-rule gene fushi-tarazu (ftz), which encodes a homeodomain protein required for embryonic patterning [18]. It has been shown that Kr heterozygous embryos exhibit defects in ftz expression [19]. In wild-type embryos, ftz is expressed in seven stripes at the onset of gastrulation (Figure 2A). In Kr1/+ embryos, however, ftz stripe 3 is narrow or weak (Figure 2B; also see [19]). The same ftz stripe 3 phenotype was found in Kr2/+ embryos (unpublished data). We wondered whether the defects in ftz expression might involve epigenetic alterations, and whether these defects could be passed to the next generation in the presence of maternal hopTum-l mutation. Indeed, we found that the ftz promoter region is differentially methylated in Kr heterozygotes (see below).
We reasoned that if hopTum-l promotes transmission of parental origin epigenetic alterations to the next generation, then the ftz stripe 3 defect caused by Kr1 could be retained in embryos from hopTum-l/+ mothers and Kr1/+ fathers that did not inherit Kr1. To test this, we examined ftz expression from a ftz-lacZ transgene carried on the CyO balancer chromosome, which contains the Kr+ allele and segregates from Kr1 in the F1 when Kr1/CyO ftz-lacZ flies are used as a parent. In embryos from male and female Kr1/CyO ftz-lacZ flies, 70% (n = 61/87) of the β-gal+ embryos exhibited the typical Kr1 heterozygous defects, characterized by weakened or narrowed stripe 3 expression (Figure 2B), suggesting that all embryos that are genotypically Kr1/CyO ftz-lacZ exhibit the stripe 3 defect. When Kr1/CyO ftz-lacZ flies were crossed to wild-type flies, in the F1 embryos, ftz-lacZ was expressed in seven stripes identical to those in the wild-type background, such that these stripes were more or less evenly spaced and similar in intensity (Figure 2A; n = 54). When hopTum-l females were mated to +/CyO ftz-lacZ males, we found wild-type ftz-lacZ pattern and no stripe 3 defects similar to those in Kr1 heterozygotes in the F1 embryos (Figure 2C; n = 78). Notably, although the JAK/STAT pathway is involved in regulating even-skipped stripe 3 expression [20,21], ftz expression seemed not affected in hopTum-l mutants. This is consistent with a lack of STAT-binding sites in the ftz promoter region (unpublished data). However, when Kr1/CyO ftz-lacZ males were mated to hopTum-l females, 94% of the F1 β-gal+ embryos retained the stripe 3 defect characteristic of Kr1 heterozygotes (Figure 2D; n = 48/51). Since in this mating scheme ftz-lacZ segregated from Kr1, embryos that expressed the ftz-lacZ trangene would not inherit Kr1 and were genotypically +/+ for the Kr locus. Thus, the presence of hopTum-l caused retention of the Kr1-specific defective ftz expression pattern in embryos that did not inherit the Kr1 mutation. These results demonstrate that hopTum-l can cause transgenerational inheritance of epigenetic changes at a transcriptional level.
To identify the epigenetic alterations caused by Kr mutations, we examined the DNA methylation status of the 620-bp minimal ftz enhancer in the ftz-lacZ transgene, as the expression of this ftz-lacZ is epigenetically modified by Kr1. DNA methylation is the predominant epigenetic modification, and methylation of CpG islands is responsible for gene silencing and genomic imprinting in mammals [5–7]. There is evidence for the presence of DNA methylation in Drosophila [22,23]. Drosophila has a Dnmt2-like DNA methyltransferase that mediates methylation of cytosine residues in vivo [24], although the biochemical activity of Drosophila Dnmt2 as a DNA methyltransferase is still to be shown. Methylated cytosines in both CG and CT dinucleotides have been found in many transposons and repetitive sequences in Drosophila genomic DNA [25], and increased promoter DNA methylation is associated with gene silencing [26].
We first assessed the methylation status of the ftz minimal enhancer (Figure 3A) by digesting total genomic DNA with a methylation-sensitive restriction enzyme BstUI, which cuts unmethylated but not methylated CGCG sequences, followed by quantification of the undigested DNA by PCR. By comparing the time courses of BstUI digestion of genomic DNA samples isolated from Kr+/− versus wild-type control flies, we concluded that the former is more resistant to BstUI digestion (Figure 3B and 3C, top panels). Digestion of the same DNA samples with a methylation-insensitive restriction enzyme HaeIII produced no differences between the two samples (Figure 3B and 3C, bottom panels). These results suggest that the minimal enhancer of ftz-lacZ in Kr+/− flies is more methylated than in wild-type flies.
We next investigated whether the Kr-dependent differential methylation of the ftz minimal enhancer can be passed to the next generation. We crossed Kr1/CyO ftz-lacZ flies to hopTum-l/+ and wild type females, respectively, and isolated genomic DNA from the F1 flies that inherited the ftz-lacZ transgene. We analyzed the methylation status of the 620-bp minimal ftz enhancer using methylation-sensitive and -insensitive restriction digests as described above. Indeed, we found the ftz enhancer in F1 flies of hopTum-l/+ females and Kr1/CyO ftz-lacZ males was more resistant to a methylation-sensitive restriction enzyme than the ftz enhancer in F1 flies of +/+ females and Kr1/CyO ftz-lacZ males (Figure 3D and 3E), consistent with the idea that hopTum-l promotes transgenerational inheritance of epigenetic changes.
We employed a second method to confirm that the promoter of the ftz-lacZ transgene has increased DNA methylation in Kr mutants and that this methylation status is transgenerationally inheritable in the presence of hopTum-l maternal mutation. We isolated total genomic DNA from embryos of different parental genotypes, digested with restriction enzymes, and incubated with antibodies against methylated cytosine. Quantification of immunoprecipitated DNA indicates that that the ftz-lacZ fragment was more methylated in embryos of Kr1/CyO ftz-lacZ flies (Figure 3F) and the higher levels of methylation was maintained in embryos from Kr1/CyO ftz-lacZ fathers and hopTum-l mothers (Figure 3G).
Finally, to further demonstrate the differential methylation of the ftz minimal enhancer in different genetic backgrounds or pedigrees, we treated the genomic DNA samples with sodium bisulfite, which converts cytosines (C) to thymidines (T), and then cloned and sequenced independent clones for each sample. Sequencing results indicated the presence of two CG (or CT)-rich “islands” in the ftz minimal enhancer that are preferentially methylated in Kr+/− samples or in embryos of Kr1/CyO ftz-lacZ father and hopTum-l mothers (Figure 4). Thus, Kr mutations indeed induce epigenetic alterations, as exemplified by increased DNA methylation in the ftz minimal enhancer, and such alterations are normally erased, but are transmitted to the next generation if an overactivated JAK kinase is present in the early embryo.
Since the epigenetic effects of Kr mutations involve DNA methylation, we investigated the effects of inhibiting DNA methylation on the ability of Kr mutations in promoting hopTum-l tumorigenesis. We raised flies in food containing the DNA methyltransferase inhibitor 5-aza-2′-deoxycytidine (5-aza-dC) and determined the effects of drug treatment on hopTum-l–dependent blood tumor formation. When raised at 100 μM 5-aza-dC (a nonlethal dose), hopTum-l/+ flies exhibited dramatically increased tumors compared with untreated hopTum-l/+ flies, with TI increased from 0.41 ± 0.05 (untreated; n = 116) to 1.27 ± 0.15 (treated; n = 68; p < 0.001). Such results are in line with TSA treatment (see above). Similar to the effects of TSA treatment, when wild-type male flies raised in 5-aza-dC were crossed to hopTum-l/+ females and allowed to produce eggs in the absence of the drug, the F1 flies exhibited increased TIs (Figure 5), but no TI increase was detected in the reciprocal cross (unpublished data), suggesting a maternal hopTum-l-dependent transgenerational inheritance. Interestingly, we found that treatment with 5-aza-dC, although by itself promotes hopTum-l tumorigenesis, abolished the ability of Kr mutations to epigenetically enhance tumors, such that when Kr1/CyO male flies raised on 5-aza-dC food were crossed to hopTum-l/+ females, the epigenetic effects (associated with CyO), but not the genetic effects of Kr, were abolished (Figure 5). Thus, the DNA methylation methyltransferase inhibitor 5-aza-dC both promotes hopTum-l tumorigenesis and inhibits Kr epigenetic effects. These results suggest that hopTum-l-induced blood tumors can be both enhanced by a general loss of genomic DNA methylation and suppressed by preventing Kr mutation–induced methylation in specific promoters.
We have investigated the effects of genetic and epigenetic mutations on the tumorigenicity of the Drosophila hematopoietic oncogene hopTum-l, and found that hopTum-l and its modifier mutations mutually influence each other, both genetically and epigenetically. In particular, we have shown that a Kr mutation that enhances hopTum-l tumorigenicity induces ftz promoter methylation, which is associated with repression of ftz stripe 3, and that Tum-l promotes transgenerational inheritance of ftz stripe 3 silencing in the F1 generation in the absence of the Kr mutation.
Taken together, these results suggest that the oncogenic JAK kinase encoded by hopTum-l is able to antagonize a cellular program that erases epigenetic markings of parental origin, allowing such epigenetic alterations to be maintained in the F1 even in the absence of the original genetic mutation. The epigenetic alterations in turn influence the risk of hopTum-l-induced tumorigenesis in the F1 generation.
Many of the M(Tum-l) genes that exhibited paternal-effect modifications encode products with known chromatin remodeling functions. These include HP1, Rpd3, and several Suppressor of variegation (Su[var]) mutations. It is conceivable that flies heterozygous for these mutations have altered chromatin states that could directly influence the epigenetic state of the zygote, leading to paternal effects as shown recently in mice [27]. However, the M(Tum-l) genes that exhibited epigenetic effects on Tum-l tumorigenicity also include those whose functions in chromatin modification are not obvious. These include transcription factors such as Kr and signaling molecules such as the Notch ligand Serrate (Ser). This observation suggests that genetic mutations in genes other than those encoding chromatin remodeling proteins may also cause epigenetic alterations.
Although Kr is expressed only in 20% of the early embryo, lacking Kr causes profound patterning defects, resulting in deletion or defects in over 70% of embryonic segments [28]. As a first zygotically expressed “gap” gene, Kr is in the top tier of the regulatory hierarchy that controls pattern formation of the whole organism [28]. Thus, Kr mutations can affect expression of genes that are not directly regulated by Kr. A Kr neomorphic allele (Krif) has been shown to affect eye development by an epigenetic mechanism [29]. Our results indicate that the Kr mutation, which likely acts early on, results in the establishment of an epigenetic signature in the genome in the form of methylation of particular promoters, such as the ftz promoter. Repression of certain “tumor suppressor genes” may explain the enhancement of the hopTum-l tumorigenic phenotype by Kr mutations. As an epigenetic modification, DNA methylation is believed to be mitotically stable. In support of this notion, we detected similar methylation patterns in the ftz-lacZ promoter in embryos and adult flies of Kr heterozygotes (Figure 4). Although we have not directly examined germ cells, the transgenerational phenomenon suggests that the Kr-dependent epigenetic signature extends to germ cells, which give rise to sperm and eggs. We envision the possibility that the epigenetic signature of germ cells is established early together with somatic cells, and can be affected by mutations in Kr, which might have a global reach in the early embryo. Alternatively, there is constant communication between germ cells and somatic cells during animal development, such that their epigenetic states will stay in “sync.” The precise mechanisms by which germ cells acquire the epigenetic states of somatic cells remain to be investigated.
When hopTum-l is inherited from the mother, its product, a hyperactive JAK kinase, is present in the embryo from the very beginning as a maternal contribution. In contrast, when inherited from the father, the hopTum-l gene product is not present in the early embryo but is expressed as a zygotic gene. Zygotic genes are not transcribed until the midblastula transition or later. The parent-of-origin effect of hopTum-l on the ability of Kr1 to modify its tumorigenicity suggests the following scenario. The M(Tum-l) mutations are capable of altering the state of the chromatin, resulting in epigenetic changes in the genome. These “epigenetic marks” can be maintained through mitosis and meiosis and transmitted to the F1 progeny, where they are normally erased in the zygote during early embryogenesis. However, the hopTum-l mutation, if present in the early embryo as a maternal-effect mutation, is able to preserve certain epigenetic alterations of parental origin. In other words, hopTum-l may play a role in counteracting a mechanism that erases epigenetic marks of parental origin during early embryogenesis.
All crosses were carried out at 25 °C on standard cornmeal/agar medium. All fly stocks, including hopTum-l, Kr alleles, CyO [ftz-lacZ], and the Bloomington Deficiency Kit Stocks, are from the Bloomington Drosophila Stock Center (http://flystocks.bio.indiana.edu/). Accession numbers for mutations used in this study are list in Table S1.
Hematopoietic tumors induced by hopTum-l were scored in adult flies, which manifest as melanotic masses most frequently found in the abdomen (see Figure 1A), but were also found in other parts of the body. Tumors of all sizes and locations were scored. Typically more than 200 progeny flies were scored for each cross. More than two independent crosses were scored and the results averaged. Tumorigenicity was quantified by TI, which is defined as the sum of tumor size times occurrence, and divided by the total number of flies of a particular genotype (TI = ∑[tumor size × n]/N, where n is the number of occurrences for a particular tumor size and N is the total number of flies counted for a particular genotype). Tumor size 1 is defined as a tumor with a diameter equal to the width of an average abdominal segment (see Figure 1A). TI 1.0 is equivalent to all flies of a category each having a 1.0 size tumor.
To eliminate genetic background effects, hopTum-l and Kr1 heterozygotes were outcrossed to a y1 w1 stock for ten generations. hopTum-l was monitored by the presence of melanotic tumors in the females in each generation. To recover Kr1 from the outcrossed progeny, ten y w virgin females were selected after five generations of outcrossing and individually crossed to a y1 w1; Sco/CyO ftz-lacZ stock (in y1 w1 background). Three males from the F1 of each cross were individually backcrossed to y1 w1; Sco/CyO ftz-lacZ flies (to maintain a stock) and the same male was testcrossed to Kr2/CyO flies. The presence of Kr1 was inferred by noncomplementation in the testcross, and a y1 w1; Kr1/CyO ftz-lacZ male was used to repeat the same outcrossing procedure one more round to establish an outcrossed y1 w1; Kr1/CyO ftz-lacZ stock.
Anti-H3Ac and anti-H3 (both from Upstate, http://www.upstate.com/) were used as 1:1,000 dilutions in Western blots, sheep anti-5-meCytidine (Abcam, http://www.abcam.com/) was used for precipitating methylated DNA. For treatment with HDAC or methyltransferase inhibitors, flies were cultured in food containing TSA (4.5 μM; Sigma, http://www.sigmaaldrich.com/), sodium butyrate (10 mM, Sigma), or 5-aza-dC (100 μM; MP Biomedicals, http://www.mpbio.com) at 25 °C. To detect β-gal expression from the ftz-lacZ transgene, mouse anti-β-gal (1:1,000; Promega, http://www.promega.com/) and a biotinylated secondary antibody and the ABC Elite Kit (Vector Laboratories, http://www.vectorlabs.com/) were used for whole-mount immunostaining of embryos. Signals were detected with DAB solution according to the manufacturer's recommendations. Stained embryos were dehydrated with ethanol, mounted with Euparal, and photographed with an Axiophot microscope using DIC optics.
Gemonic DNA was isolated using the DNeasy Tissue kit (Qiagen) according to the manufacturer's instructions with minor modifications. Thirty 1–2-d-old adult flies or 100 μl of 0–12-h embryos of desired genotypes were homogenized in 180 μl of PBS and 20 μl of proteinase K (1 mg/ml) per manufacturer's protocol. The samples were treated with DNase-free RNase A (Sigma) for 2 h at 37 °C prior to column purification. For restriction digests, 3 μg of genomic DNA was incubated with 10 units of BstUI (New England Biolabs, http://www.neb.com/) or 10 units of HaeIII (New England Biolabs) in 60 μl of total volume at 37 °C. At different time points, an aliquot of the digests was removed and heated at 80 °C to inactivate the restriction enzyme. One microliter of each sample was used for PCR amplification with primers specific to ftz-lacZ (forward: 5′-CCCAGGGATCGGACGTAATGTTAT-3′; reverse: 5′-GGATGTGCTGCAAGGCGATTAAGT-3′). Bisulfite treatment was carried out with the EpiTect Bisulfite Kit (Qiagen, http://www1.qiagen.com/) according to the manufacturer's instructions. Genomic DNA (2 μg) was treated in Bisulfite Mix. Treated genomic DNA was amplified with the following strand-specific primers (forward: 5′-TTTAGGGATTGGATGTAATGTTAT-3′; reverse: 5′-AAATATACTACAAAACAATTAAAT-3′). The PCR fragments were cloned into pGEM-T vectors (Promega) and independent plasmid DNA isolates were sequenced. Sequencing was carried out by Gene Gateway (http://www.genegateway.com/). For immunoprecipitation, genomic DNA was first digested to completion with EcoRI and BamHI (New England Biolabs). Digested genomic DNA (2 μg) in 200 μl was used for immunoprecipitation with 5 μl of anti-5-meC (Abcam) or control antibody at 4 °C overnight, together with protein-G beads that had been preabsorbed with sonicated single-stranded salmon sperm DNA. The antibody complex was centrifuged and washed and eluted. The presence of ftz-lacZ promoter sequence was quantified by PCR with the above primers. |
10.1371/journal.pgen.1004911 | A Re-examination of the Selection of the Sensory Organ Precursor of the Bristle Sensilla of Drosophila melanogaster | The bristle sensillum of the imago of Drosophila is made of four cells that arise from a sensory organ precursor cell (SOP). This SOP is selected within proneural clusters (PNC) through a mechanism that involves Notch signalling. PNCs are defined through the expression domains of the proneural genes, whose activities enables cells to become SOPs. They encode tissue specific bHLH proteins that form functional heterodimers with the bHLH protein Daughterless (Da). In the prevailing lateral inhibition model for SOP selection, a transcriptional feedback loop that involves the Notch pathway amplifies small differences of proneural activity between cells of the PNC. As a result only one or two cells accumulate sufficient proneural activity to adopt the SOP fate. Most of the experiments that sustained the prevailing lateral inhibition model were performed a decade ago. We here re-examined the selection process using recently available reagents. Our data suggest a different picture of SOP selection. They indicate that a band-like region of proneural activity exists. In this proneural band the activity of the Notch pathway is required in combination with Emc to define the PNCs. We found a sub-group in the PNCs from which a pre-selected SOP arises. Our data indicate that most imaginal disc cells are able to adopt a proneural state from which they can progress to become SOPs. They further show that bristle formation can occur in the absence of the proneural genes if the function of emc is abolished. These results suggest that the tissue specific proneural proteins of Drosophila have a similar function as in the vertebrates, which is to determine the time of emergence and position of the SOP and to stabilise the proneural state.
| The sensory organ precursor cell (SOP) that forms the mechanosensory bristles of the adult PNS of Drosophila is a paradigm to study neural precursor determination. The current model states that the SOP is selected in proneural clusters (PNCs) defined through the expression of the proneural genes. The selection occurs through lateral inhibition mediated by the Notch signalling pathway. The SOP is pre-selected by differential expression of Extramacrochaetae (Emc), the only member of the Id proteins in Drosophila, which inactivates the proneural factors. We have re-examined the selection process using novel markers and mutants. Our data suggest a different picture of SOP selection. We discovered a band–like region of varying proneural activity where the peaks constitute the proneural clusters. Within the PNC, a subgroup exists from which the SOP arises. The Notch pathway has two distinct functions in the subgroup and in the rest of the band. We show that so far one unappreciated essential role of the proneural genes is the neutralisation of the activity of Emc. Our data suggest that the selection of the SOP is more similar to neural selection in vertebrates than previously anticipated.
| The body of the imago of Drosophila melanogaster is covered with mechanosensory bristles, called macrochaetae (MCs) and microchaetae (mcs). In the notum, mcs cover the central regions, whereas the larger MCs arise at precise positions in peripheral regions and form a stereotypic pattern. Both sensilla consist of only four cells, which are the progenies of a single neural precursor cell, termed sensory organ precursor cell (SOP). The SOPs of MCs develop in the wing imaginal disc during the second half of the third larval instar stage in a precise temporal sequence [1]. Its development is a paradigm to study fundamental aspects of the determination of a neural precursor cell (reviewed in [2]).
The SOP is selected within proneural clusters (PNC), which are defined through the expression of tissue-specific proneural genes. In the notum these are achaete (ac) and scute (sc), two members of the achaete-scute complex. Their activity conveys cells into a proneural state from which they can proceed to become SOPs if they reach a threshold level of proneural activity. Concomitant loss of their function results in the loss of all bristles of the notum. They encode transcription factors of the class II bHLH family, have identical expression patterns and function redundantly (bHLH factors and their classification are reviewed in [3]). Class II proteins possess a basic DNA binding domain and a HLH domain that mediates dimerization with the ubiquitously expressed Daughterless (Da), the only class I bHLH protein in Drosophila. Class V HLH proteins are antagonists of bHLH factors. They lack a basic DNA binding domain and form non-functional heterodimers with class I and II bHLH factors. Thus, they are negative posttranslational regulators of bHLH transcription factors. The only Drosophila class V member is Extramacrochaetae (Emc), which forms inactivating heterodimers with Ac, Sc and Da (reviewed in [4]). Weak alleles of emc cause formation of additional MCs in homozygousity. Analysis of the emc null alleles in the eye imaginal disc revealed a regulatory loop between Da and Emc, where Da activates expression of Emc and itself and Emc in turn inactivates Da [5]. This loop assures that both factors are expressed at correct levels. Loss of emc function causes up-regulation of Da expression. The consequences of this up-regulation for bristle development have not been investigated.
Proneural genes play a similar, but not identical role in mammals (reviewed in [6]): In Drosophila the activity of the proneural genes appears to confer a proneural state onto cells and promote neural differentiation, while their mammalian counterpart only promote neural differentiation of neural plate cells, which have adopted a proneural state through other mechanisms.
The SOP is selected among the cells of the PNC by a mechanism that is called lateral inhibition mediated by the Notch signalling pathway (reviewed in [7]). Upon loss of Notch activity, all cells in a PNC adopt the SOP fate, indicating that it prevents them progressing from the proneural to the SOP state. Most genes contributing to Notch signalling produce this neurogenic phenotype upon loss of function and are therefore classified as neurogenic genes. In Drosophila Notch is activated by two ligands of the DSL family, Delta (Dl) and Serrate (Ser) (reviewed in [8]). Dl is the main ligand for SOP selection, but Ser has a redundant role [9]. Binding of the ligands to Notch initiates the release of its intracellular domain (NICD) into the cytosol. NICD is transported into the nucleus where it associates with the CSL transcription factor Suppressor of Hairless (Su(H)) to activate the target genes. The release of NICD occurs through two proteolytic cleavages of Notch. A ligand-induced first cleavage (S2) by Kuzbanian (Kuz) creates a membrane inserted intermediate (Notch EXtracellular Truncation (NEXT)), which is cleaved by γ-secretase (S3-cleavage) to release NICD. γ-secretase is a complex consisting of Presenilin (Psn), Anterior pharynx defective 1 (Aph-1), Nicastrin (Nic) and Presenilin enhancer 2.
The membrane-associated E3-ubiquitin ligases Neuralized (Neur) and Mindbomb1 (Mib1) are important for the activity of both DSL ligands [10]. They probably mediate the ubiquitylation of the intracellular domains (ICDs) of Dl and Ser on lysines, which in turn initiates their endocytosis. In Drosophila the two ligases have similar functions [9], [11]. In imaginal discs of Drosophila, expression of Neur is restricted to the SOP, while Mib1 is ubiquitously expressed, indicating that most DSL signalling probably depends on Mib1 [12]. During SOP selection the main target of the Notch pathway are the members of the Enhancer of split complex (E(spl)-C), which encode class IV bHLH proteins. These proteins antagonise the activity of the proneural factors to suppress SOP development (reviewed in [13]).
According to the lateral inhibition model, activation of the Notch pathway antagonises a cells ambition to adopt the SOP fate, because it initiates the expression of the E(spl)-C genes, which suppress the activity of Ac and Sc (reviewed in [7]). While Notch is expressed ubiquitously, the activity of the pathway is linked to that of the proneural proteins through transcriptional regulation of Dl. Consequently, cells with high proneural activity express high levels of Dl and can send a strong inhibitory signal to their immediate neighbours, which prevents them from adopting the SOP fate. The resulting regulatory feedback-loop automatically selects the SOP: All cells of a PNC initially express similar levels of proneural activity and therefore mutually inhibit each other from adopting the SOP fate through Dl/Notch signalling (S1A Fig.). A small difference in activity of the proneural factors results in a small difference in the expression of Dl among cells of a cluster. The regulatory feedback loop between proneural activity and Dl expression amplifies the initially small difference and transforms it in an all or nothing situation: The result is a cell with high proneural activity and high Dl expression that becomes the SOP and neighbours with no proneural activity that switch fate to become epidermoblasts. Emc appears to contribute to the initial bias of proneural activity, through its differential expression among cells of the PNC [14]. Thus, while all cells initially express similar levels of Ac and Sc, they have different proneural activity due to the differential expression of Emc. The SOP arises at positions of the lowest Emc expression and, hence, highest proneural activity. Note, that the lateral inhibition model predicts changes in the expression of Dl and the activity of the Notch pathway during the selection of the SOP.
One problem of the lateral inhibition model is to explain how the nascent SOP inhibits cells in the PNC that are located more than one cell diameter away, since Dl, as a transmembrane protein, reaches only the next cell. Recent reports provided an explanation, which is based on the observation that SOPs directly contact remote cells through filopodia [15], [16]. These contacts are thought to transfer the Dl signal to these remote cells. However, experimental proof that shows that these contacts mediate inhibition is scarce. Moreover, the selection of SOPs in Dl Ser double mutant PNCs where Dl is expressed in all cells at uniform levels occurs normally [9]. A similar observation has been made during the selection of neuroblasts in the embryo [17]. These data suggest that differential expression of Dl, which is a hallmark of the lateral inhibition model, is dispensable for the selection process. A recent update of the lateral inhibition model suggests that not transcription, but the activity of Dl is regulated in a feedback loop among the cells of the PNC [18]. In this model the activity of Notch results in activation of expression of the Bearded proteins, which in turn suppress the activity of Neur and therefore that of Dl. In this case, a differential activity of the Notch pathway should be observed. However this model is not in agreement with the finding that Neur is expressed only in the emerging SOP, but not in its neighbours, which are inhibited by it [12].
Most experiments that addressed the mechanism of SOP formation were performed more than a decade ago. In the meantime novel markers and mutants became available that allow a more precise look. We therefore re-examined the process. Our data suggest that the inhibitory Notch signal is restricted in range to the next cell. We found that a band-like region with changing proneural activity exists in the notum, whereby the peaks constituting the PNCs. The SOP appears to be chosen by an unknown mechanism among cells of a subgroup within the PNC that is defined through the requirement of the activity of Neur. We failed to find evidence for the described feedback loop of the lateral inhibition model. Rather, the Notch pathway has two functions: it provides a baseline of activity in the proneural band that defines the PNCs and the subgroup. It also mediates a strong inhibitory signal emitted by emerging SOP to inhibit the same fate in its neighbours. We found that Emc is required in all cells of the imaginal discs to suppress the proneural state. In the absence of emc function, the cells are in a proneural state from which several proceed to become SOPs, even in the absence of the function of ac and sc. These findings suggest that one important function of the proneural proteins is to neutralise Emc. Moreover, the Notch mediated selection of the SOP is independent of the proneural activity of Ac and Sc. The presented results indicate that the selection of the neural precursor in Drosophila is more similar to that in vertebrates than anticipated.
In order to determine the activity of the Notch pathway during SOP selection, we performed several experiments. First, we monitored expression of the Notch activity reporter Gbe+Su(H) together with the SOP marker Hindsight (Hnt) in the notum of wing discs at the late third larval instar stage. Hnt is a marker for mature, already determined SOPs [19]. Expression of Gbe+Su(H) in the notum is complex and comprises many domains, which report the activation of Notch in several parallel running processes (Fig. 1A, B, [20]). Nevertheless, we observed an elevation of expression of Gbe+Su(H) (halo) in cells around several Hnt positive SOPs (Fig. 1A, B, arrowheads). We counted that this halo contained between 6 to 9 immediate neighbours (Fig. 1C, D). We found the same halos around mature SOPs using a recently available GFP variant of Gbe+Su(H), termed NRE-pGR [21]. In order to test whether the halos are related to SOP development, we monitored the expression of Gbe+Su(H) in sc10.1 mutant imaginal discs, which lack the function of the proneural genes ac and sc, which are responsible for bristle formation in the notum. In these discs, the halos were absent, while all other domains of expression were unaffected (Fig. 1E, arrowheads, compare with A). Thus, the halos are caused by Ac/Sc dependent Notch signalling in PNCs. The observed halos suggest that the selected SOP sends a strong signal that activates the Notch pathway in its direct neighbours.
The lateral inhibition model predicts differential activity of Notch among cells of the PNC during SOP selection. This should be reflected in differential expression of Gbe+Su(H), with decreasing expression in the nascent SOP and increasing expression in its neighbours. To test this prediction, we looked at SOP formation at different times and at different positions. We used E(spl)m8-SM-GFP. In contrast to E(spl)m8, this construct is not responsive to Notch activity due to deletion of its Su(H) binding sites in its promoter [22]. It is the earliest marker for nascent SOPs and expressed well before Hnt [22], [23]. We found that this marker is initially expressed in a small group of cells in the PNC from which the SOP arises (Fig. 1 K-M, O-R, yellow arrows) and is then restricted to the emerging SOP before it expresses Hnt. We failed to observe any differential expression of Gbe+Su(H) around E(spl)m8-SM-GFP positive cells, before the onset of expression of Hnt. Expression was uniform and weak during the time the nascent SOP expresses only E(spl)m8-SM-GFP (Fig. 1F-R, arrow and arrowhead for the SOPs arising at the pDC and aPA positions). The halos formed at the time the SOP enlarged and initiates Hnt expression. These results suggest that activation of Notch appears to be uniform during the selection of the SOP. Only after its selection and maturation to the Hnt positive state, it sends an inhibitory signal to its neighbours via the Notch pathway.
Expression of Gbe+Su(H) in the notum is complex and includes four stripe-like domains (1-4 Fig. 1A), the halos and a diffuse weak expression between the domains (Fig. 1A). This complex pattern complicated the analysis at several SOP positions. To get rid of the irrelevant domains of expression of Gbe+Su(H), we performed the further analysis in mib1 mutants. Loss of mib1 function, which encodes the ubiquitously expressed E3-ligase required for DSL ligand activity, affects only a subset of Notch-dependent processes and MC formation is only mildly affected, resulting in the formation of a few supernumerary bristles [24], [25]. We found that SOPs emerged in the normal sequence, but earlier in mib1 mutants. Ectopic SOPs emerged later. Moreover, the stripes of Gbe+Su(H) were lost or dramatically reduced with the exception of stripe 3 (see Fig. 1A, S, U). The halos around the SOPs remained intact and became recognisable at new positions (Fig. 1S, U, arrows). As in wildtype nota, the halos of Gbe+Su(H) expression were present only around mature, Hnt-positive SOPs. Also in this background, differential expression of Gbe+Su(H) during SOP formation of the aDC position was not observed (arrowhead in Fig. 1S-X, arrowhead). Note, expression of Gbe+Su(H) was observed also in the nascent SOPs at least at some positions, such as the PNP and pDC (arrows in Fig. 1W, X). We observed this also in wildtype discs, although not in that clarity because of the additional expression domains. This observation suggests that Notch is still active in the nascent SOP or the stable ß-galactosidase is not degraded at this time of SOP development. Expression of Gbe+Su(H) disappeared in fully determined SOPs, indicating that the Notch pathway is switched off and that differences of Notch activity among cells of the PNC can be detected (white arrow and arrowhead in Fig. 1M, asterisk in W). Altogether, the observations support the notion that a differential expression of Gbe+Su(H) and therefore differential activation among the cells of a PNC occurs only after selection of the SOP. Hence, lateral inhibition appears to occur after the SOP is selected, but not during its selection.
The observed halo of Gbe+Su(H) in cells around the SOP suggests that the inhibitory Notch signal emitted by the SOP reaches only adjacent cells. In order to experimentally confirm the short range of the signal, we generated kuz or aph-1 mutant PNC cells by clonal analysis and tested their abilities to activate the Notch pathway in adjacent non-mutant cells (Fig. 2A-D). kuz encodes the ADAM protease that mediates the S2 cleavage of Notch to create the NEXT intermediate and aph-1 encodes a subunit of the γ-secretase complex, which performs the subsequent S3 cleavage required to release NICD. Loss of function of each gene results in the inactivation of the Notch pathway at the signal receiving side. The lateral inhibition model predicts that these mutant cells should accumulate high proneural activity and become SOPs, because they cannot receive the inhibitory signal. Because of the high levels of proneural activity, they should express high levels of Dl and emit a strong Notch signal to their neighbours (S1A–C Fig.). To test this prediction, we monitored the expression of Gbe+Su(H) around mutant SOPs generated by clonal analysis. For simplification, we induced these clones in mib1 mutant discs where several irrelevant domains of Gbe+Su(H) expression are absent. The function of mib1 is not essential for SOP selection, since this process is largely mediated by the other E3 ligase Neur. We found that only mib1 cells that were direct neighbours of the mib1 kuz or mib1 aph-1 double mutant SOPs up-regulated Gbe+Su(H) expression (Fig. 2A–H, arrows). This observation suggests that the range of the Notch signal emitted by the mib1 kuz or mib1 aph-1 mutant SOPs reaches only to their direct neighbours.
Next we tested over what distance the inhibitory Notch signal emitted by a SOP is effective in a functional assay. We analysed the ability of ectopic SOPs mutant for several neurogenic genes to influence the development of a wildtype SOP nearby. The lateral inhibition model predicts that:
1. cells of a PNC that are mutant for both ligands are unable to emit an inhibitory signal, but able to receive it. Therefore, these cells should never adopt the SOP fate if located adjacent to a strong signalling wildtype SOP (case 1, S1B, b Fig.).
2. ectopic SOPs mutant for genes that are required for signal-reception cannot be inhibited and should sent a potent inhibitory signal that should prevent all wildtype cells in its neighbourhood from adopting the SOP fate. Hence, no wildtype SOP should develop adjacent to a mutant SOP (case 2, S1C, c Fig.).
If the inhibitory Notch signal reaches a longer distance, e.g. through filopodia, the described effects should extend also to cells located farer away from the mutant cells (see also S1 b, c Fig.). We found that Dl/Ser double mutant cells (Fig. 2I, J, arrowhead), which are not able to inhibit their neighbours (case 2), adopted the SOP fate despite the presence of a wildtype SOP one cell diameter away (Fig. 2I, J, arrow). This suggests that the wildtype SOP can only suppress the SOP fate in cells that are in direct contact to it.
SOPs mutant for kuz and Psn (Fig. 2K–T, arrow) could not prevent wildtype cells (Fig. 2K–T, arrowhead) from adopting the SOP fate at a distance of one or more cell diameters away (case 2, Fig. 2K–M). Moreover, we found several cases where these mutant SOPs are unable to inhibit a wildtype SOP in direct contact to them (Fig. 2N–T, arrowhead). This suggests that cells at the positions where the SOPs normally arise, can adopt the SOP fate, even next to strongly signalling cells. Cells at these positions appear to be pre-selected to become SOPs and immune to the inhibitory signal. Altogether, the results strongly support the conclusion that the effective inhibitory Notch signal emitted by the nascent SOP reaches only its direct neighbours and that the SOP in the PNC is pre-selected. We have previously arrived to the same conclusions upon clonal analysis of Su(H) mutants [19]. A problem with this analysis was that Su(H) mutant cells fail to properly express Neur and several target genes of the Notch pathway are de-repressed in them [19], [26]. This might cause an abnormal behaviour of the mutant cells during SOP selection. All these effects are absent in the kuz and Psn mutant cells analysed here.
We wondered how each ligand contributes to the activation of the Notch pathway in the notum. To answer this question, we compared the expression of Ser and Dl relative to that of Gbe+Su(H). We found that the combined expression of both ligands together covers most of the notum (S2A–G Fig.). The domains of high Dl expression are shifted with respect to that of Gbe+Su(H) expression (S2A, B, E Fig.). Thus, strong expression of Dl does not correlate with high Notch activation. To investigate the expression of Dl and Ser in the PNC, we used sca-lacZ, a faithful marker for proneural activity. In regions of the PNCs, Dl appears to be uniformly expressed (S2H, I Fig., arrow and arrowhead). In the region of the aPA/Tr1 cluster expression of Dl was significantly lower (S2B, E Fig., arrow). The domains of high expression of Ser correlated better with that of high Notch activity, although it was not a perfect match (S2A, C, F Fig.). Consequently, the domains of high expression of Dl and Ser are overlapping (S2D–F Fig.).
Optical sections showed that all cells of a PNC contain Dl positive vesicles (S2H, I Fig.). Most of these vesicles (89%, n = 276) were Rab7 positive and therefore maturing endosomes ([27], S2H, I Fig.). We did not observed differences in Dl expression or distribution among cells of the PNCs (S2H, I Fig.).
Expression of Dl was unchanged in sc10.1 mutants (S2J Fig., compare with E), suggesting that the proneural factors do not regulate the expression of Dl during the determination of the SOP of the MCs. This is in agreement with the analysis of expression of Dl during mcs development and our observation that the expression of Gbe+Su(H)-lacZ is unaffected, with the exception of the lost halos ([28], Fig. 1E).
We previously noticed that sca-lacZ, which faithfully reports proneural activity, is weakly expressed also in regions between PNCs [23]. This is highlighted by the comparison of sca with the expression of the general enhancer that regulates the expression of ac and sc in the DC cluster (DC–E) ([29], [30], Fig. 3A–D). The observation suggests the existence of a corresponding band-shaped curved region of changing proneural activity (proneural band), with the PNCs being only peaks of this region. We found that expression of sca is elevated in regions between PNCs in mib1 mutant nota (Fig. 3A, E, arrows). Moreover, more cells expressed the earliest SOP marker E(spl)m8-SM-GFP (Fig. 3F–H). We quantified this for the region of the aPA1/tr and pSA clusters and found that in wt 6 ± 2 cells are positive for E(spl)m8-SM-GFP (n = 6), while 18 ± 4 are positive in mib1 mutants (n = 8). Thus, more cells accumulate high proneural capacity, which is actively suppressed during normal development through the activity of the Notch pathway. We were curious whether the cells of the proneural band could be coaxed to adopt the SOP fate if the activity of Notch is abolished. Indeed, this was the case for most of the cells of the band: In Psn or nic mutant nota, where the γ-secretase complex is inactive, the PNCs increased in size over time and eventually fused at 2h after puparium formation (apf) to form a continuous band of strong sca expression (Fig. 3I–N). Moreover, most cells eventually expressed Hnt, indicating that they adopted the SOP fate (Fig. 3I–N). There are two exceptions: cells located between the DC and SC clusters never became SOPs, although they expressed sca (yellow arrow in Fig. 3I–N). Cells in the region between the PNP and ASA clusters expressed high levels of sca, but only few had adopted the SOP fate (red arrow in Fig. 3I–N).
Altogether, these results reveal the existence of a proneural band in the notal region of the wing disc, which is divided in regions of high and low proneural activity. They uncover a novel role of the Notch pathway, which is the definition of the PNCs through suppression of the proneural activity in cells located between the clusters. This activity is only partly generated through Mib1, since the cells between the PNCs do not adopt the SOP fate in its absence, although they have increased levels of proneural activity.
In the imaginal discs Neur function is restricted to neural development and the neurogenic phenotype of neur mutants is milder than that of mutants of other neurogenic genes [9]. We therefore determined how many cells of a PNC adopt the SOP fate in the absence of neur function. To do so, we compared expression of sca and Hnt in neur clones, induced by ptcGal4 UAS Flp. It revealed that only a subgroup of the PNC (neur-group) adopted the SOP fate (Fig. 4A–D). The neur-group comprised up to eight SOPs, which is approximately the sum of the SOP plus its neighbours (see Fig. 1C) and is in agreement with the results of a previous study [9]. Importantly, we observed many cases where one or a few cells in the neur-group were wildtype. In these cases, only one SOP formed (Fig. 4E–H, U-X). This SOP was always a wildtype cell, indicating that a Neur positive cell can efficiently inhibit the other cells of the subgroup (Fig. 4E–H, arrowhead, U–X, white arrow). A nice example including both described situations in the DC cluster is depicted in (Fig. 4I–L, arrowhead and arrow). In cases where the SOP was wildtype, we observed a halo of Gbe+Su(H) expression in the adjacent cells, confirming that it can send a potent inhibitory Notch signal to its neighbours in the presence of Neur (Fig. 4M–O, arrowheads). If all cells of the subgroup were mutant, a clear halo was absent (Fig. 4P, Q, arrowhead). Altogether, the results indicate that a subgroup exists within the PNC (neur-group), which requires the activity of Neur in at least one cell to prevent the formation of supernumerary SOPs through lateral inhibition.
The SOPs in neur mutant territories appear to emerge in a sequence. This is indicated that only one SOP can be detected with Hnt or elevated sca-lacZ expression during early phases of mutant PNCs development (see Fig. 4R-X, arrowhead for the early arising pDC, yellow arrows for the later arising PNCs). These observations support the notion that one cell of the neur-group is ahead in its development to become a SOP. During further development more cells become Hnt positive SOPs in mutant PNCs indicating that the advanced cell is unable to inhibit its neighbours in the absence of Neur (Fig. 4U–X, arrowhead, DC cluster).
In order to investigate how the neur group might be defined, we monitored the expression of DC-E. This enhancer controls the general expression of Ac and Sc in the DC cluster, but not their elevation in the nascent SOP (S3 Fig., [30]). To identify the position of the future SOP, we used the elevation of sca-lacZ expression in one cell of the PNC, which is the first known sign of SOP formation [31]. We found that the expression of the DC-E is not uniform and a group of cells at the position where the SOP arises expressed higher levels. This is exemplified for the two SOPs of the DC cluster (S3 Fig., arrowheads). Thus, differential expression of Ac and Sc from the beginning of the PNC is likely to contribute to the determination of a subgroup of cells from which the SOP arises. This subgroup is probably the neur group. If the cells of the neur group possess more proneural activity, they should adopt the SOP fate before the rest of the PNC in the absence of Notch activity. Hence, one should initially observe small groups of SOPs in Psn mutant discs. The groups should increase in size during the third instar. This is what we observed (S4 Fig.). Note, that during the third instar new SOPs are added at the edges of the existing group. Thus, the cells of the PNC have different proneural activity with the subgroup possessing the highest proneural activity, allowing them to adopt the SOP fate faster than the rest of the PNC. This group probably requires higher Notch activity to suppress the SOP fate provided by Neur. Outside the subgroup the proneural activity decreases towards the edges of the PNC.
In contrast to complete loss of function of Notch, cells of the PNCs outside the neur group and of regions between the PNCs never became SOPs upon loss of neur function. Moreover, no elevation of sca-lacZ expression in neur mutant cells of the proneural band outside the neur subgroup was observed (Fig. 4U,V, compare with Fig. 3A). This observation suggests that the requirement for function of Neur is restricted to the subgroup.
In wing imaginal discs homozygous for the weak allele emcpel, ectopic SOPs arise in regions outside the proneural band and expression of sca-lacZ is expanded ([14]; Fig. 5A–D). These observations indicate that Emc is involved in defining the proneural band. We went on to test the effects of two null alleles, emc1 and emcAP6, which have been recently used for clonal analysis in combination with the Minute technique [5]. We were able to recover emc clones with both alleles in imaginal discs even without the Minute technique (Fig. 5E–J). We obtained the largest clones using a ptcGal4 driven UAS Flp construct (Fig. 5K–M). Independently of the technique, we observed that emc mutant cells activated ectopic sca-lacZ expression (Fig. 5E–M). Moreover, several mutant cells developed further to become ectopic SOPs (Fig. 5E–M). These results indicate that emc mutant cells adopt a proneural state from which they can progress to become SOPs. Note, that loss of emc function caused these effects also in cells of the posterior compartment where ac and sc are not expressed (Fig. 5H–J). Interestingly, we found that mutant SOPs tended to arise away from the clone boundary.
We observed elevation of sca expression already in clones of early third instar discs (Fig. 5N–P). Thus, adopting a proneural state is an immediate response of the cells to loss of emc function. The same response to loss of emc function was found in the leg disc. Hence, it appears to be a general response to loss of emc activity in imaginal disc cells. Emc inhibits only the activity of proneural bHLH proteins, but not their expression. Therefore, the cells of the imaginal discs must posses a low level of proneural proteins, whose activity become elevated upon loss of Emc.
Only a fraction of cells in the emc mutant clones developed to ectopic SOPs, although all cells elevated their proneural activity (4, 4 ± 1,8% in emc null mutant clones containing at least 100 cells, n = 6). These SOPs were well separated from each other, suggesting that the Notch mediated selection process operates in emc mutant territories. To test this notion, we first monitored the expression of Gbe+Su(H) in emc cell clones. We found that it was not affected (S5A–D Fig.). We could detect halos Gbe+Su(H) expression around ectopic Hnt positive SOPs, indicating that they send a strong inhibitory signal to their neighbours (S5A–D Fig., arrows). Expression of Dl was unaffected (S5E–K Fig., arrow and arrowhead).
We went on to investigate the consequences of loss of Notch function in emc mutant wing discs in two ways: We analysed flies homozygous for the hypomorphic allele emcpel and found that, in contrast to emcpel single mutants, all sca-lacZ positive cells became Hnt positive SOPs in emcpel Psnc1 double mutants, with exception of very few cells at the edges, which expressed sca only weakly (Fig. 6A, C, arrows). We further analysed the consequences of inactivation of the Notch pathway in emc null mutant cells (emcAP6 Psnc1 clones). We found that most of the emcAP6 Psnc1 mutant cells that express sca-lacZ adopted the SOP fate in the absence of Notch activity (80,5 ± 16% in clones containing at least 100 cells, instead of 4,4 ± 1,8% in emc null mutant clones, n = 6), with the exception of cells in the central part of the hinge region, which did not progress beyond the sca expressing proneural state until the stage of analysis (highlighted by the arrow in Fig. 6D–G). In addition we found that an anterior small stripe of the notum fails to elevate sca-lacZ and Hnt expression (Fig. 6D–G, arrowhead). This indicates that these cells are not capable to become SOPs. In summary, the analysis showed that the activity of the Notch pathway is required for the selection of SOPs in emc mutant territories.
Loss of emc function causes SOP development in the posterior compartment of the wing disc, although Ac and Sc are not expressed there in wildtype discs. This raised the possibility that they are dispensable for SOP formation in emc mutant territories. To test this possibility, the consequence of loss of emc function in sc10.1 flies was analysed [32]. sc10.1 flies lack the function of ac and sc and consequently lack all bristles in the head and notum (Fig. 7A, B). We found that induction of emc null mutant clones in sc10.1 flies resulted in the re-appearance of MCs as well as mcs in both regions (Fig. 7C). Thus, the activity of Ac and Sc is dispensable for bristle development in the absence of emc function. This result also indicates that Emc regulates the formation of mcs. This involvement is only revealed in the sc10.1 background, since most ectopic bristles induced by loss of emc function in a wildtype background are MCs (Fig. 7D). Note, that the bristles present in sc10.1 emc double mutant territories are well separated from each other, suggesting that the Notch mediated selection process operates despite the lack of the proneural genes.
We found that the emcAP6 sc10.1 mutant bristles are associated with a 22C10 positive neuron and possess normal looking socket and bristle cells (Fig. 7E, F). Thus, the corresponding SOPs initiated the normal lineage. Although we observed bristles at ectopic positions in the sc10.1 background, the distribution of MCs and mcs was not random. The MCs tended to develop in lateral and scutellar regions of the notum and the mcs developed in central regions, like it is observed in the wildtype situation. This indicates that positional cues exist in the absence of Ac, Sc and Emc, which contribute to the regional specification of bristles. However, the positioning of individual MCs was lost. For example the MCs highlighted with white arrows in (Fig. 7C and P) arose in a region of the scutellum normally devoid of bristles.
Analysis of corresponding double mutant wing imaginal discs revealed that expression of sca is elevated to similar levels as seen in emc clones (Fig. 7H–K). This indicates that emcAP6 sc10.1 cells are in a proneural state despite the absence of the tissue specific proneural factors. Moreover, single cells progress to the SOP stage (Fig. 7G–K, arrows). Note, that the regions of clones that covered central areas of the notum did not form SOPs at this stage (see Fig. 7G, asterisk). Since we observe that these central areas are covered with the later developing mcs in the adult flies, we believe that the corresponding SOPs arise later. The double mutant cells also dramatically increased expression of Da (Fig. 7L–N, arrow). Thus, the previously observed elevation of Da expression in emc clones is independent of Ac and Sc [5]. In order to confirm that the loss of emc function is responsible for the re-appearance of the bristles and not a second mutation on the used chromosome, we depleted sc10.1 mutant nota of emc function by expressing an UAS emc-RNAi constructs with ciGal4, which drives expression throughout the anterior compartment. We observed the re-appearance of many mcs (Fig. 7O, arrowheads). To achieve maximal efficiency of depletion, we next co-expressed UAS emc-RNAi with UAS Dcr2. This resulted in the re-appearance of a higher number of mcs and also MCs (Fig. 7P, arrowhead and arrows respectively). These results confirm that the loss of emc function causes the re-appearance of bristles in sc10.1 mutant nota.
Recently a protein trap, Emc-YFP, which encodes a fully functional Emc-YFP fusion protein became available [33]. This was used to re-examine the expression of Emc during bristle development. The comparison revealed that the expression patterns of the previously available emc-lacZ and Emc-YFP are similar in the wing imaginal disc. However, emc-lacZ expression was elevated in determined SOPs of late third instar discs, while Emc-YFP was decreased and eventually switched off (S6A Fig.). Moreover, the “valleys” of expression of Emc-YFP were broader. This difference can be explained by the known stability of ß-galactosidase in emc-lacZ. The strong perdurance of ß-galactosidase in progenies of the expressing cells is likely to cause the observed gradual difference among cells around the SOP not seen with emc-YFP.
We compared the emergence of SOPs relative to Emc-YFP and sca (S6B-h Fig.). In early third instar discs the expression of sca-lacZ is initiated weakly throughout most of the notum (S6C, D Fig., asterisk). Within this domain of uniform expression, the PNCs arise in the previously described sequence and initially comprise few cells (S6D–F Fig.). Note, that at this time one cell eventually expresses higher levels of sca indicating that is has been pre-selected to become the SOP (arrowhead in S6F′′ Fig.). No Hnt expression was observed at this stage. Thus, the selection of at least the early arising SOPs occurs at a stage where the clusters comprise few cells.
As previously reported, Emc is expressed in all cells of the imaginal discs. In early third instar discs two domains of higher expression can be observed: one large central anterior located domain (domain 1) and a smaller posterior distal one (domain 2; S6C Fig.). An even smaller third domain (domain 3) follows at more anterior distal position (S6D Fig.). In late instar discs 5 domains of higher expression can be observed (S6H Fig.).
In general we found that the expression of Emc is low in the regions of the PNCs (high sca expression). For example, the Tr1 arises between domain 1, 2, and 3 (S6E-e Fig., arrow and arrowhead). Moreover, the early clusters arise at the edge of the large domain 1 in a band of low Emc expression. This band probably defines the proneural band.
We failed to find differential expression suggestive for a role of Emc to pre-select a cell within the PNC. The earliest known sign of SOP determination in a cell is the elevation of sca-lacZ expression. Upon close examination of Emc-YFP expression at this time at several positions, we failed to observe a clear reverse correlation between Emc-YFP and sca-lacZ expression (S6e, g Fig.). However, when the SOP is determined and expresses Hnt, the expression of Emc disappears (S6H, h Fig.).
We confirm the previous finding that the expression of Emc is unaffected in sc10.1 mutants, indicating that it is independent of the proneural factors ([14]; S7A Fig.) However, we found very weak residual expression of sca-lacZ in the notum (S7A Fig., arrowheads). This hints to the presence of residual proneural activity. In order to test this possibility, we abolished the function of the Notch pathway in sc10.1 discs (sc10.1 PsnC1 mutants). The concomitant loss of Notch activity should lead to an increase of the proneural activity and, thus, in an increase of the residual expression of sca-lacZ. In agreement with this conclusion the sca-expression was increased in the remaining ac and sc independent PNCs of the radius of sc10.1 PsnC1 wing discs that give rise to non-bristle type sensilla (arrows in S7B Fig.). The cells of these remaining clusters were also Hnt positive, indicating that they have adopted the SOP fate. However, we failed to observe an increase of the residual weak expression of sca-lacZ in the notum where the ac sc dependent PNCs are located (S7B Fig., arrow). This strongly suggests that the very weak residual expression of sca-lacZ in the notum is independent of proneural activity and that proneural activity is abolished in sc10.1 nota.
In contrast to proneural genes, depletion of da function results in the loss of expression of Emc ([5], S7C Fig., arrow). This confirms previous findings that Da acts independently of the proneural genes to regulate the expression of Emc [5].
The analysis of expression of Emc-YFP in Psn mutants revealed that the regions where the cells of the proneural band are most resistant to become SOPs upon loss of Notch activity, e. g. between the aSA and pSA or the DC and SC clusters, are regions of high Emc expression (S7D Fig., red and yellow arrows). The overall expression of Emc appears not to be affected upon loss of Psn function. Altogether, these results suggest that Emc is involved in definition and subdivision of the proneural band, but does not pre-select the SOP in the PNCs.
In this study, we re-examined the development of the SOP of the MC using recently available reagents. We found evidence that strongly suggests that the range of the Notch signal is restricted to the next cell: The elevated expression of Gbe+Su(H) around the SOP is observed only in adjacent cells. In addition, cells of PNCs that are not able to receive the Notch signal, but can send a strong signal to adjacent wildtype cells, cannot prevent a wildtype cell from adopting the SOP fate at a distance of two cell diameters away. Likewise, cells that are not able to send a signal cannot be prevented by wildtype SOPs from adopting the SOP fate more than one cell diameter away. These results suggest that the discovered filopodia of the SOP, which contact more remotely located cells do not extend the range of the inhibitory signal to these cells.
Our study reveals the existence of a band of proneural activity. The PNCs are regions of elevated proneural activity in this band, rather than discrete clusters. In the band, the Notch pathway exerts an additional novel function, which defines the extent of the PNCs. In the absence of Notch function, most cells in the proneural band accumulate high levels of proneural activity that allows them to become SOPs. Thus, the pathway suppresses the proneural activity and the SOP fate in cells located between the PNCs in the proneural band. The short range of the Notch signal indicates that it is probably local mutual signalling among direct neighbours that generates the necessary Notch activity (mutual inhibition). The expression of Dl and Ser and the overall activity of Gbe+Su(H) (with exception of the halos) is unchanged in the absence of Ac and Sc. This suggests that the widespread activity of Notch in the notum that prevents most cells in the proneural band to become SOPs is not influenced by the proneural factors. It provides a baseline activity of Notch that suppresses the proneural activity in the band to prevent the formation of ectopic SOPs.
The presented results indicate that a subgroup within the PNCs exists, which is operationally defined via the requirement of the activity of Neur. The existence of a subgroup has previously been suggested on basis of experiments with a temperature sensitive allele of Notch [34]. These data and the ones presented here, suggest that the cells of the subgroup require Notch activity that is stronger than the baseline activity to be inhibited from adopting the SOP fate. This increase in activity is generated by the nascent SOP through a Neur enhanced Dl signal: We here found that if only one cell in the subgroup is neur positive, it can prevent all other neur mutant members to adopt the SOP fate. Thus, initiating the expression of Neur first, is a critical step for a cell to adopt the SOP fate, since it allows a cell to strongly inhibit its neighbours. The inhibitory signal prevents the accumulation of sufficient proneural activity to also activate Neur in the neighbours. This inhibition is probably reflected in the observed halo of Gbe+Su(H) expression around SOPs. The findings are in good agreement with a previous study that showed that the level of Neur in a cell is a critical factor for the formation of the SOP of the mc [18].
Loss of Notch activity results in expression of Neur and a dramatic increase in proneural activity in all cells of the PNC (e. g. see [19]). Moreover, the nascent SOP, which contains the highest proneural activity, is the only cell that initiates Neur expression during normal development and expression of neur is abolished in ac sc mutant discs [23]. These data indicate, that high proneural activity is required for the expression of Neur. Thus, the cell in the subgroup with the highest proneural activity is the cell that will express Neur first. The expression of Neur enables it to inhibit its neighbours from adopting the SOP fate by suppressing their proneural activity.
Our data therefore indicate that two activities of Notch are present during SOP formation. One generated through mutual signalling, which is not regulated by Ac and Sc and is sufficient to inhibit all cells in the proneural band outside the neur subgroup to become SOPs. This signalling requires the ubiquitously expressed Mib1 and antagonises the activity of Ac, Sc and Da. However, there is residual activity of Notch in mib1 mutants sufficient to prevent most cells from adopting the SOP fate. This residual activity is generated either independently of E3 ligases or by another unknown E3-ligase. In any case this component contributes to the baseline activity of the Notch pathway in addition to Mib1. The second activity on top of the baseline activity in the neur subgroup is generated by a Neur mediated strong signal from the nascent SOP. This signal suppresses the proneural activity of the other members of the neur subgroup. It is dependent on proneural activity, which initiates the expression of Neur. Thus, lateral inhibition is probably operating after the emerging SOP reaches a threshold of proneural activity. It serves to prevent the formation of supernumerary SOPs in the neur group and assures that other cells can generate the necessary SOP in case the selected one is lost.
How is the neur subgroup defined? We found that the PNCs are small in their beginning, comprising the number of cells typical for the subgroup. These cells probably also constitute the small groups of SOPs observed in early third instar discs mutant for Psn. It is likely that E(spl)m8-SM expression defines this subgroup since we show here that it is expressed in a small group of cells from which the SOP arises. This construct contains only one E box, the binding sites for Ac and Sc, and response to high proneural activity [35]. We therefore believe that the cells of the early PNC are the neur group and possess the highest proneural activity.
During normal development, a cell with more proneural activity is already recognisable at the early phase of the PNCs. This suggests the existence of a pre-selection mechanism that assures that one cell in the neur-subgroup is advanced in its development. Evidence for such a mechanism has been also previously found during rescue experiments studying the function of the proneural genes Ac and Sc [36]. We have here obtained additional experimental evidence for this pre-selecting mechanism: In neur clones one of the cells is advanced in its development towards the SOP fate. Moreover, clonal analysis of kuz and Psn mutants revealed that wildtype cells at positions in the PNC where the SOP arises cannot be prevented from adopting the SOP fate, even if a mutant SOP that cannot be inhibited (e. g. kuz mutant), is its neighbour. We have shown that the mutant cells can generate a strong inhibitory Notch signal. This indicates that the pre-selecting mechanism renders the wildtype SOP immune to the signal. The nature of this mechanism is not clear, nor whether it is always the same cell in a cluster that is selected.
Recent work demonstrated that in the eye disc a regulatory loop between Da and Emc assures correct expression of both factors and results in their complementary expression [5]. Consequently, loss of emc function results in an increase of expression of Da. The consequences of this up-regulation for the proneural state of the mutant cells have not been investigated in detail. The work focused on the eye imaginal disc and revealed that a few of the mutant cells in clones could adopt the neural fate. The neural cells do not express Runt, a marker expressed in the normal neural cells. Thus, the loss of emc does not result in the complete determination of the neural fate. The state of the vast majority of the cells in clones remained unknown. We observed up-regulation of proneural activity in emc clones already in early third instar wing imaginal discs, indicating that it is an immediate reaction to the loss of emc function. Some of these cells progress to become SOPs. The increase in proneural activity was also observed in emc clones of the leg disc. Thus, the cells of imaginal discs must be permanently inhibited from adopting a proneural state through the activity of Emc. It has to be pointed out that this situation is remarkably similar to that in the early vertebrate embryo, where all cells of the blastula adopt the proneural state unless they are inhibited through BMP signalling. The cells of the neural plate maintain the proneural state due to the presence of BMP antagonists (reviewed in [37]).
In the eye disc and during oogenesis expression of Emc is regulated by the Notch pathway [38], [39]. We failed to find evidence that supports a regulatory relationship between Emc and the pathway in the notum during SOP development, since the loss of Psn function did not affect the expression of EMC. However, it has been previously shown that the expression of Emc along the dorso-ventral boundary in the wing primordium depends on the activity of the Notch pathway [40]. This correlates well with our finding that this domain is independent of the activity of Da. However, the genetic network of the wing is significantly different from that in the notum. For example Notch signalling induces the expression of Wg along the D/V boundary. However, its expression in the proximal wing and in the notum is independent of the activity of the Notch pathway. This appears to be true also for the different domains of expression of Emc.
We here found that the function of ac and sc is dispensable for bristle development in the absence of emc function. How is the SOP fate initiated in these emc ac sc triple mutant cells? We believe that the activity of Da is sufficient for SOP development in this situation for the following reasons: 1. Da is expressed ubiquitously and is required for the formation of all external sense organs [41]. 2. Strong over-expression of Da induces bristle formation in cells that lack the whole AS-C [42], [43]. In contrast, over-expression of Sc fails to induce SOP formation in the absence of Da [43]. 3. Da can form homodimers that bind to the same DNA target sequences as Ac/Da and Sc/Da heterodimers in bend-shift assays [44]. 4. Loss of emc activity increases the activity of Da [5]. We here show that this increase is independent of the activity of Ac and Sc. 5. Our results show that Da regulates the expression of sca independently of Ac and Sc. 6. It has been shown that the mammalian homologue of Da, E2A, acts without its class II partners during B-cell development (reviewed in [45]). Thus, it is likely that in the absence of function of emc, ac and sc, Da forms active homo-dimers that initiate the required neural program.
While it is clear that the activity of Ac and Sc is required during normal development, the formation of normal bristles in their absence after concomitant loss of emc function raises the question about their function. Our data suggest that an important function is the neutralisation of Emc through formation of heterodimers with it or with Da. This releases Da from inactive heterodimers with Emc. The neutralisation of Emc by Ac and Sc, which are expressed in precise spatial and temporal regulated patterns, allows the differentiation of neural precursors at the correct position and time. The recent finding that a Sc variant without its transactivation domain is fully active fits well to this view of the function of Ac and Sc [43]. Thus, through their intricate and dynamic expression, Ac and Sc and other tissue specific proneural factors determine when and where a neural precursor cell develops. In this view the function of the tissue-specific proneural genes of Drosophila, is similar to that in mammals where their orthologs also promote differentiation of neural precursors in a proneural field, the neural plate, at correct positions and time.
Based on our results, we suggest a working model for the selection of the SOP of the MC (Fig. 8A–C): The differential expression of Emc defines a proneural band in the notum with changing proneural activity. The PNCs in this band are determined and positioned through the cluster-like expression of Ac and Sc, which increases the proneural activity at these positions. A baseline of activity of the Notch pathway generated by mutual inhibition prevents cells between the PNCs to accumulate high levels of proneural activity. In addition, it prevents cells located in the PNC, but outside the neur group, to accumulate high proneural activity required for adopting the SOP fate.
In the PNCs, expression of Ac and Sc neutralise Emc. Consequently, the proneural activity increases dramatically, since the released Da can form homodimers and/or heterodimers with Ac or Sc. The cells of the initial small PNCs later constitute the neur subgroup. The cells of this subgroup have the highest level of proneural activity and experience this activity also for the longest time. Within this subgroup a cell is pre-selected to become the SOP by a so far unidentified mechanism. Hence, it is the first to reach the threshold level of proneural activity required to initiate the expression of Neur. The expression of Neur enables it to efficiently inhibit the other cells of the subgroup through lateral inhibition. As a consequence these cells never accumulate sufficient proneural activity to activate Neur expression and to become a SOP. The strong signal also further activates the expression of Brd proteins that inhibit the activation of Neur, which might be accidentally activated weakly in one of the neighbours [46]. This activation contributes to the precision of determination process. Thus, a combination of mutual and lateral inhibition mediated by the Notch pathway operates in the PNC during the determination of the SOP. Only the lateral inhibition component depends on proneural activity through transcriptional activation of expression of Neur. For further information and how the model can explain the phenotypes of neur and mib1 mutants, see (S8D-F Fig.) and the corresponding figure legend.
The model differs from the lateral inhibition model in the following points: No feedback loop between expression of Dl and proneural activity and, hence, no differential Dl expression is required. Instead the future SOP is pre-selected and advanced in its development. Subgroups within a proneural band defined through its requirement of Neur exist. In this subgroup the activation of the expression of Neur is critical for SOP development since it enables a cell to potently inhibit its neighbours. The pre-selection mechanism favours a cell at the right position to initiate the expression of Neur before the others of the Neur group and therefore secures its development as SOP. Moreover, the existence of mutual signalling explains the inhibition of cells in the proneural band outside the subgroup without the necessity of signalling of Dl over longer distances.
UAS lines: UAS Flp (Bloomington stock collection). Gal4 lines: ptcGal4 [47], ciGal4 [48]. Other lines: Gbe+Su(H)-lacZ [20], NRE-pGr [21], Gbe+Su(H)-GFP [49], E(spl)m8-SM-GFP [22], YFP-Emc ([33]), sca-lacZ (Bloomimgton stock collection), neurA101-lacZ (Bloomington stock collection), DCE-GFP [29], UAS emc-RNAi (VDRC, line #100587), tub. rab7-YFP [50].
Mutants: PsnC1 FRT2A (null allele; [51]), emcAP6 FRT80 [5] and emc1 FRT80 (Bloomington stock collection #5532), aph-1D35 FRT40A [52], nicA7 [53], sc10.1 (Bloomington stock collection), mib12 and mib13 [24], mib1EY09780 [25], neur1 FRT 82B (Bloomington stock collection, a gift from C. Delidakis), Dlrev10 SerRX22 FRT82B [54], kuzES24 FRT40A [55].
Clones were generated with the FLP/FRT or MARCM system [56] and induced at the first larval instar (24-48h after egg laying) by applying a 1h heat shock (37°C). Alternatively, the clones were induced using a UAS Flp construct driven by ptcGal4 in some of the cases of the neur and emc clones as indicated in the figures or figure legends. Flies were raised at 25C.
Antibody staining was performed according to standard protocols. Primary antibodies used: mouse anti-Wg (4D4), mouse Dl antibody against the extracellular domain (C594.9B), anti Hnt (1G9), anti NICD (C17.9C6), anti NECD (C458.2H), 22C10/futsch antibody and anti Cut (2B10). All antibodies were purchased from the Developmental Studies Hybridoma Bank (DSHB). Anti Ser antibody (gift of Elisabeth Knust, [57]), anti-Da [41], Fluoro-chrome conjugated secondary antibodies were purchased from Invitrogen/Molecular Probes (Dianova, anti-gp). Images were obtained with a Zeiss AxioImager Z1 Microscope equipped with a Zeiss Apotome.
For staging of discs during pupal stages white pupae (0–1 h apf) were selected and staged accordingly. For larval discs the emerged Hnt positive SOPs were used in combination with the description of the emergence of SOPs by Huang et al. (1991).
|
10.1371/journal.pmed.1002484 | Association between intake of less-healthy foods defined by the United Kingdom's nutrient profile model and cardiovascular disease: A population-based cohort study | In the United Kingdom, the Food Standards Agency-Ofcom nutrient profiling model (FSA-Ofcom model) is used to define less-healthy foods that cannot be advertised to children. However, there has been limited investigation of whether less-healthy foods defined by this model are associated with prospective health outcomes. The objective of this study was to test whether consumption of less-healthy food as defined by the FSA-Ofcom model is associated with cardiovascular disease (CVD).
We used data from the European Prospective Investigation of Cancer (EPIC)-Norfolk cohort study in adults (n = 25,639) aged 40–79 years who completed a 7-day diet diary between 1993 and 1997. Incident CVD (primary outcome), cardiovascular mortality, and all-cause mortality (secondary outcomes) were identified using record linkage to hospital admissions data and death certificates up to 31 March 2015. Each food and beverage item reported was coded and given a continuous score, using the FSA-Ofcom model, based on the consumption of energy; saturated fat; total sugar; sodium; nonsoluble fibre; protein; and fruits, vegetables, and nuts. Items were classified as less-healthy using Ofcom regulation thresholds. We used Cox proportional hazards regression to test for an association between consumption of less-healthy food and incident CVD. Sensitivity analyses explored whether the results differed based on the definition of the exposure. Analyses were adjusted for age, sex, behavioural risk factors, clinical risk factors, and socioeconomic status. Participants were followed up for a mean of 16.4 years. During follow-up, there were 4,965 incident cases of CVD (1,524 fatal within 30 days). In the unadjusted analyses, we observed an association between consumption of less-healthy food and incident CVD (test for linear trend over quintile groups, p < 0.01). After adjustment for covariates (sociodemographic, behavioural, and indices of cardiovascular risk), we found no association between consumption of less-healthy food and incident CVD (p = 0.84) or cardiovascular mortality (p = 0.90), but there was an association between consumption of less-healthy food and all-cause mortality (test for linear trend, p = 0.006; quintile group 5, highest consumption of less-healthy food, versus quintile group 1, HR = 1.11, 95% CI 1.02–1.20). Sensitivity analyses produced similar results. The study is observational and relies on self-report of dietary consumption. Despite adjustment for known and reported confounders, residual confounding is possible.
After adjustment for potential confounding factors, no significant association between consumption of less-healthy food (as classified by the FSA-Ofcom model) and CVD was observed in this study. This suggests, in the UK setting, that the FSA-Ofcom model is not consistently discriminating among foods with respect to their association with CVD. More studies are needed to understand better the relationship between consumption of less-healthy food, defined by the FSA-Ofcom model, and indices of health.
| The Food Standards Agency (FSA)-Ofcom model is used in the UK to identify ‘less-healthy’ foods in order to restrict their advertising to children.
Variants of the FSA-Ofcom model, as well as other nutrient profiling models, are increasingly being used to regulate food retailing or marketing for the purposes of improving health; yet, very few of these models have been validated.
The FSA-Ofcom model has been shown to classify foods in a way that is consistent with professional opinion, but there has been limited assessment of its association with health outcomes.
We used the European Prospective Investigation of Cancer (EPIC)-Norfolk study to test the prospective association of less-healthy food consumption with incident cardiovascular disease, cardiovascular mortality, and all-cause mortality.
Each item of food or drink reported in a participant’s 7-day diet diary was given a score based on its nutrient composition and then categorised as either ‘less-healthy’ or ‘healthy’.
Participants (n = 22,292) were allocated to 1 of 5 groups based on their consumption of less-healthy food (as a proportion of total dietary energy).
After adjustment for confounding factors, we found no association between consumption of less-healthy food and incident cardiovascular disease (n = 4,965) or cardiovascular mortality (n = 2,555)
The findings were robust to a variety of sensitivity analyses, including adjustment for exclusion based on different cardiovascular risk factors.
Whilst no single study is definitive and our findings are in contrast to similar work in a French cohort, these findings suggest that the FSA-Ofcom model is not consistently discriminating among foods with respect to their associations with cardiovascular disease in the UK context.
Public health officials and scientists may want to review whether and how the FSA-Ofcom scoring system could be improved for use in the UK and elsewhere.
There is a robust evidence base concerning the health risks associated with the consumption of many foods that are often labelled ‘unhealthy’ (e.g., red meat, sugar-sweetened beverages, and takeaway food), and it would be inappropriate to use this study to undermine present dietary advice for the public.
| Nutrient profiling is the science of classifying or ranking foods according to their nutritional composition for reasons related to preventing disease [1,2]. Over 100 nutrient profile models exist globally (around 60 of which are publicly available). One of the most prominent is a model originally devised by the Food Standards Agency (FSA) and used in the UK by the communications regulator (Ofcom) to restrict the advertising of unhealthy foods to children [3–5]. Variations on this model have been used in other countries (e.g., in Australia, New Zealand, France, and South Africa) [1,6,7].
The FSA-Ofcom model has 2 parts: a scoring system that assigns each food item a numerical score based on its nutrient composition and a classification system that then categorises each food or beverage item that exceeds a prespecified score as ‘less-healthy’. Ranking foods by the FSA-Ofcom model has been shown to correlate with the views of nutritional professionals, and classifications compare favourably with UK food-based dietary guidelines [8,9]. In two French cohorts, prospective associations between a diet consisting of foods with a higher mean score and weight gain, development of metabolic syndrome, cardiovascular risk, and cancer risk have been reported [10–14]. There are no similar studies in a UK population. The French studies do not reflect how the model is used in the UK presently. The French scoring system is similar to that of the FSA-Ofcom model but scores fats, cheeses, and beverages differently [13,15], and the French studies have tested the scoring system rather than the classification system. There may also be important differences between French and British diets [16,17], which could result in different associations.
Our objective was to test whether consumption of less-healthy food, as identified by the FSA-Ofcom model, was associated with incident cardiovascular disease (ischaemic heart disease and stroke). We chose to focus on cardiovascular disease (CVD) because the components of the scoring system (e.g., saturated fat, salt, sugar, fruits, and vegetables) suggest that it should identify foods that would be associated with a higher risk of CVD.
The EPIC-Norfolk study protocol was approved by the Norwich District Health Authority Ethics Committee, and all participants gave written informed consent.
The EPIC-Norfolk study is part of the European Prospective Investigation of Cancer (EPIC) study that spans 10 European countries. It has been described in detail elsewhere [18]. In brief, participants aged 40–79 years were recruited from the general population through general practices in the east of England between 1993 and 1997. Participants (n = 25,639) completed a baseline questionnaire covering sociodemographic factors, medical history, medication use and health behaviours, completed a 7-day diet diary [19], and attended a clinical research facility (for measurement of blood pressure, height, and weight). Health outcomes were ascertained by linkage to hospital admissions data and death certificates.
We excluded participants who did not complete at least 1 day of the 7-day diet diary and those who were in the top or bottom 0.5% of the distribution of the ratio of reported energy intake to basal metabolic rate (calculated using sex-specific Schofield equations) [20]. For analysis of incident CVD, we further excluded participants with prevalent disease (self-reported angina, heart attack, or stroke) as well as those with missing covariates. For analysis of mortality, we included participants with prevalent disease and excluded participants with missing covariates. Because missing covariate data were limited to a small proportion of the total sample (1.08%, 250/23,242, for analysis of incident CVD; 1.29%, 322/24,880, for analysis of mortality outcomes), we chose to exclude these participants rather than impute missing data.
Participants reported their food intake for 1 week using a 7-day diet diary. A trained nurse, during the visit to the clinical research facility, obtained a 24-hour-diet recall that formed the first day of the diet diary and served as a general instruction regarding the detail required for the diary. Participants were additionally provided with written instructions, and the diet diary contained colour photographs to aid portion size estimation [19,21]. The 7-day diet diaries were entered using the in-house developed DINER data-entry system and checked and calculated using the DINERMO processing programmes [22,23]. For each food item, we also ascertained the proportion (by weight) that was fruit, vegetables, pulses/lentils, or nuts, which we have previously described as ‘disaggregated food groups’ [23]. This resulted in nutrient quantities and (disaggregated) food weight intake for every food item consumed. The majority of included participants (90.8%; 20,885/22,992) completed all 7 days of the diary.
The FSA-Ofcom model assigns an overall numeric score for any given item of food, based on the following components: energy; saturated fat; total sugar; sodium; nonsoluble fibre; protein; and fruit, vegetable, and nut content. In summary, each component is scored based on the quantity per 100 g edible weight [24]. Scores for energy, saturated fat, total sugar, and sodium are positive (i.e., adverse score), graded on a 10-point scale. Scores for nonsoluble fibre and protein as well as fruits, vegetables, and nuts are negative (i.e., beneficial or healthy score), graded on a 5-point scale. A copy of the full algorithm is available for download [24] and outlines how the scores for the different components are added together to give the overall score. If a food scores 4 points or more, it is categorised as less-healthy, and a beverage is categorised as less-healthy if it scores 1 point or more. Reflecting the operational use of the FSA-Ofcom model, any beverage that contained alcohol was not scored [10,25,26].
For each participant, we summed the energy consumed from all foods and beverages (referred to as ‘food items’) that were classified as less-healthy. Energy from alcoholic beverages formed a separate group, since alcohol is not part of the score guidelines. For each participant, we estimated the proportion of energy consumed from food items that were classified as less-healthy by the FSA-Ofcom model:
(Energyfromless‑healthyfood+Energyfromless‑healthybeverages)(Totalenergyintake−Energyalcoholicbeverages)
We then divided the study sample into quintile groups (fifths) based on this proportion. Thus, our primary exposure measure was quintile groups of proportion of energy intake consumed from food items categorised as less healthy.
Our primary outcome measure was incident CVD. Secondary outcome measures were cardiovascular mortality and total (all-cause) mortality.
We defined incident cases of CVD as any primary fatal or nonfatal event of ischaemic heart disease (International Classification of Disease [ICD]-10 codes I20–I25) or cerebrovascular disease (stroke) (ICD-10 codes I60–I69). Incident cases were ascertained by record linkage to hospital admissions data and death certificates coded for CVD using the ICD-10 criteria. Death from any cause, including cardiovascular death, was ascertained by record linkage to mortality data confirmed via death certificates with ICD codes held at the UK Office for National Statistics. Record linkage for deaths and hospital admissions was complete to 31 March 2015.
We used Cox proportional hazards regression to estimate the hazard ratio and 95% confidence interval for the association between exposure and outcome. Whilst aspects of the analytic plan (e.g., classification of exposure, choice of outcomes, and use of Cox proportional hazards) were agreed prior to beginning the analysis (S1 Text), there was no preagreed study protocol specifying the choice of covariates and sensitivity analyses.
We adjusted analyses for two sets of potential confounders. Information on other covariates was obtained from the baseline questionnaire. Model 1 was adjusted for sociodemographic and behavioural risk factors: age (continuous, years), sex, level of education, smoking status (never, former, or current), physical activity (inactive, moderately inactive, moderately active, or active), alcohol consumption (units/day), and overall energy intake (kJ/day). Model 2 additionally adjusted for self-reported clinical risk factors at baseline (blood pressure-lowering medication, lipid-lowering medication, prevalent diabetes, prevalent hypertension, prevalent hypercholesterolemia, past cancer diagnosis, family history of myocardial infarction, family history of stroke, and family history of diabetes).
The decision to include an extensive list of possible confounders in a second model was made after the descriptive analyses showed evidence of increased cardiovascular risk amongst participants who were consuming the least amount of less-healthy food (i.e., possible reverse causation) and because of the failure of the original analytic analyses to demonstrate an association between increasing consumption of less-healthy food and CVD (which might be attributable to reverse causation). We adjusted for indicators that were likely to signal cardiovascular risk to the participant (rather than all measures of cardiovascular risk), as these might influence dietary behaviour (e.g., knowing that one has a diagnosis of hypertension might affect dietary behaviour). In practice, this meant adjusting for self-reported diagnoses (hypertension, hyperlipidaemia, diabetes, and cancer), reported medication usage (for blood pressure and cholesterol), and reported family history (ischaemic heart disease, stroke, and diabetes). These factors are causally related to incident CVD and, given the descriptive data, might contribute to reverse causation. We did not adjust for factors that might be unknown by the participant and might be on the causal pathway between diet and disease (e.g., measured blood pressure and measured cholesterol). While some of the covariates included in Model 2 may act as confounders, they may also be on the causal pathway, i.e., act as mediators (e.g., poor diet leading to hypertension leading to CVD), and thus, adjustment for these factors might be considered overadjustment. In response to comments from peer review, we additionally report Model 2’, which excludes potential mediators, i.e., adjusts for Model 1 covariates, past cancer diagnosis, family history of myocardial infarction, family history of stroke, and family history of diabetes.
In analyses assessing the outcome of mortality, we additionally adjusted for prevalent CVD (self-reported angina, stroke, and heart attack).
To aid interpretation and as a test of an increasing trend across quintiles, we report the significance of the regression coefficient for the quintiled exposure when it was treated as a continuous variable. All analyses were conducted in Stata v13. We used visual plots and Schoenfeld residuals to test the proportional hazards assumption.
In addition, we also tested the association between quintile group of fruit and vegetable consumption (ranked on weight consumed), adjusting for the same set of covariates. Associations between fruit and vegetable consumption and CVD [27–29] are commonly observed, so an association would be expected. This analysis served as a validation of the approach to categorisation of the exposure and the analytic approach. The decision to include this analysis was made retrospectively in light of the initial findings.
We undertook the following sensitivity analyses. First, in light of initial findings, we repeated our primary analysis of combined CVD as an outcome with the separate outcomes of incident myocardial infarction and incident stroke. Second, in response to comments from peer review, we repeated the analysis but did not adjust for total dietary intake. This is sometimes considered appropriate when testing the relationship between dietary patterns and disease if it is thought that dietary patterns mediate their effect on disease through total energy intake.
Third, we used different approaches to the categorisation of less-healthy food consumption: (A) We allocated participants to a quintile group based on the proportion of food weight that was categorised as less-healthy (rather than food and beverage energy, since the relatively high weight of beverages might distort any association; this analysis was preplanned), and (B) we allocated participants to a quintile group based on the mean energy-weighted FSA-Ofcom score of all food items consumed. This latter approach is the same as that used by other authors and was introduced in response to work published after the study was conceived [10–13]. It effectively only tested the first part of the FSA-Ofcom model, the scoring system, treating it as a ‘dietary index’ measure, and did not test the classification system. In addition and in response to comments from peer review, we tested a ‘substitution model’ in which we included the following terms: energy from unhealthy food, energy from unhealthy beverages, energy from healthy beverages, and total dietary energy. The resultant coefficient estimates the hazard ratio when energy from unhealthy food is replaced with energy from healthy food, holding total energy intake constant.
Fourth, we took an alternative approach to confounding variables: (A) After undertaking the initial analysis and noting the inverse association between body mass index (BMI) and consumption of less-healthy food, we additionally adjusted the primary analysis for baseline BMI; and (B) to test for residual confounding by prevalent disease within the mortality analyses, we repeated the mortality analyses excluding participants with prevalent CVD (self-reported angina, stroke, and heart attack). In response to comments from peer review, we have introduced a further set of analyses to address potential reverse causation. First, we excluded all events that occurred within 2 years of follow-up. Second, we excluded—rather than adjusted for—comorbidities at baseline, excluding participants with cardiovascular comorbidities (self-reported hypertension, hyperlipidaemia, blood pressure medication, or lipid-lowering medication) or those with other comorbidities (diabetes and cancer). Third, we excluded participants with a family history of CVD (stroke or heart attack). Finally, we combined all these exclusion criteria and additionally excluded participants with a family history of diabetes, thus restricting the analysis to participants with no reported comorbidities at baseline, with no reported family history of CVD or diabetes, and who did not have an incident event within 2 years of follow-up.
After exclusions (Fig 1), there were 22,992 participants included in the analyses of incident CVD and 24,880 in the analyses of mortality. There were no important differences in the baseline characteristics of participants included and excluded because of missing covariates (Table A in S1 Data). Participants were followed up for a mean of 16.4 years. During follow-up, there were 4,965 incident cases of CVD (1,524 fatal within 30 days). Among a total of 7,139 all-cause deaths, 2,555 deaths were attributed to CVD. The baseline characteristics of the participants are shown in Table 1. Those in quintile group 5 (i.e., highest proportional consumption of less-healthy food) were more likely to be older and male and less likely to have completed higher education (degree or equivalent). Some health indices among quintile group 5 were worse—for example, a greater proportion of participants reported being current smokers. However, some health indices were better—for example, they were less likely to be on medication (antihypertensives or lipid-lowering medication), were less likely to have a family history of heart attack, and had a lower BMI. Reported physical activity did not differ appreciably across the quintile groups.
The quality of diet as assessed by different foods and nutrients showed a gradient across the quintile groups, with those who consumed the highest proportion of less-healthy food also consuming higher absolute quantities of foods or nutrients associated with poor health (e.g., salt, processed meat, saturated fat, and sodium) and lower absolute quantities of foods or nutrients associated with good health (e.g., fish, fruit, and vegetables, as well as a lower ratio of polyunsaturated to saturated fat) (see Table 1). Individuals in quintile group 5 also consumed more energy. At baseline, those in quintile group 5 consumed over twice as much less-healthy food and over 5 times as many less-healthy beverages in comparison to those in quintile group 1.
Table 2 shows the prospective associations between quintile groups of proportional less-healthy food consumption and incident CVD. The unadjusted analyses showed a positive association between consumption of less-healthy food and incident CVD. After adjustment for sociodemographic and behavioural factors (Model 1), there was an inverse (protective association) (test for trend, p = 0.009) between consumption of less-healthy food and incident CVD. After additional adjustment for indicators of cardiovascular risk at baseline (Model 2), there was no association between less-healthy food consumption and incident CVD. The same pattern of findings was observed when we took a different approach to adjustment for confounders, additionally adjusting for BMI (Model 2 + BMI, Table B in S1 Data) or adjusting for a more restricted set of indices of cardiovascular risk, (Model 2′, Table B in S1 Data).
Table 3 shows the prospective association between quintile groups of proportional less-healthy food consumption and mortality. The unadjusted analyses show an association between less-healthy food consumption and cardiovascular mortality. After adjustment for sociodemographic and behavioural factors (Model 1), there was an apparent inverse (protective) association (test for trend, p = 0.03) between consumption of less-healthy food and cardiovascular mortality. After additional adjustment for indicators of cardiovascular risk at baseline (Model 2), there was no association between less-healthy food consumption and cardiovascular mortality.
The unadjusted analyses showed an association between less-healthy food consumption and all-cause mortality. After adjustment for sociodemographic risk factors and behavioural risk factors (Model 1), there was no association. After further adjustment for indicators of cardiovascular risk at baseline (Model 2), a higher risk of all-cause mortality was observed for those in quintile group 5 relative to those in quintile group 1 (hazard ratio = 1.11, 95% CI 1.02–1.20).
An inverse (protective) association between fruit and vegetable consumption (quintile group of consumption by weight) and incident CVD was observed, in unadjusted and all adjusted models (Table 4).
After adjustment (Model 2), no association was observed for the separate outcomes of incident stroke and incident myocardial infarction (Table C in S1 Data). When not adjusting for total dietary energy intake, an inverse (protective) association between quintile of less-healthy food consumption and risk of incident CVD was observed for Model 1, and no association was observed for Model 2 (Table D in S1 Data)
After adjustment (Model 1 and Model 2), no association was observed between proportional less-healthy food consumption (based on proportion of food weight that was categorised as less-healthy) and incident CVD (Table E in S1 Data), nor was an association observed between less-healthy food consumption (Model 2), based on the mean energy-weighted score of all items consumed, and incident CVD (Table F in S1 Data). The ‘substitution model’ indicates that the isocaloric replacement of less-healthy food for healthier food (or vice versa) was not associated with increased risk of CVD (Model 1 and Model 2, Table G in S1 Data).
Further sensitivity analyses attempted to deal with possible reverse causation. Additional adjustment for BMI did not materially alter the findings (Table B in S1 Data), nor did exclusion of participants with comorbid conditions at baseline (n = 21,338 for exclusion of diabetes and cancer, Table H in S1 Data, and n = 17,948 for exclusion of participants with self-reported hypertension or hyperlipidaemia or blood pressure- or lipid-lowering medication, Table I in S1 Data), participants with a family history of CVD (n = 11,481, Table J in S1 Data), or participants who experienced an incident event within 2 years of follow-up (n = 22,737, Table K in S1 Data). The findings were also similar when excluding participants with comorbid conditions or a family history of CVD or who experienced an incident event within 2 years of follow-up (Table L in S1 Data).
Further exclusion of prevalent diseases for the mortality analyses did not appreciably alter the findings (Table M in S1 Data).
In this population-based cohort of older UK adults, we did not detect any significant association between the quantity of less-healthy food consumed, defined using the FSA-Ofcom model, and incident CVD or cardiovascular mortality after adjustment for confounders. Whilst unadjusted models showed positive and significant associations between the quantity of less-healthy food consumed and cardiovascular outcomes (Tables 2 and 3), this was explained by a number of confounding factors, principally age and sex, and as such, we do not consider these crude associations to be meaningful. There was also a suggestion that those who report lower intakes of less-healthy foods were at higher risk of CVD (e.g., high prevalence of diabetes and medication usage among participants in quintile group 1; see Table 1). For this reason, we put more emphasis on the findings of Model 2, which adjusts for indicators of cardiovascular risk at baseline, when considering cardiovascular outcomes.
We did observe an association between less-healthy food consumption and all-cause mortality after adjustment for baseline indicators of cardiovascular risk (Model 2), but not when only adjusting for sociodemographic and behavioural risk factors (Model 1). Given that CVD accounts for a third of all deaths (35.7%) and the absence of associations for CVD, it might be more appropriate to put greater emphasis on the Model 1 findings for the all-cause mortality analyses. Given this and having undertaken multiple tests of significance, we suggest the all-cause mortality Model 2 findings should be treated with caution.
The key strength of this study is defining the exposure in a way that reflects the operational usage of the FSA-Ofcom model in the UK, making use of 7-day diet diaries, which in our sample have been shown to have greater agreement with objective measures of diet than other common methods (24-hour recall or food frequency questionnaires) [19,30]. We have tested associations with both specific outcomes (ischaemic heart disease and stroke), for which there is a greater a priori expectation of an association given the components included in the FSA-Ofcom model, and nonspecific outcomes (all-cause mortality). While these are important health outcomes, we note they are not health outcomes observed in children, who are the intended beneficiaries of the restriction of television advertising of less-healthy food.
Dietary behaviour is self-reported and may be inaccurate or biased. Baseline dietary data were collected in the 1990s. The foods on offer in the 1990s, particularly processed foods, may not reflect the foods that people consume today in the UK or elsewhere. Our study has effectively tested the FSA-Ofcom model across all foods in the diet, whereas the scoring system is only likely to be operationalised (in the UK) on those foods that are heavily advertised (i.e., manufactured or processed foods).
Our findings are notably different to recently published findings from two French cohort studies SU.VI.MAX (SUpplementation en VItamines et MinérauxAntioXydants) (n = 13,017) and NutriNet-Santé (n = 75,801) [11–14]. Whilst some of the outcomes in these publications (e.g., cancer) are different to our primary outcome, others are related (e.g., metabolic syndrome and weight gain) or the same (incident CVD). Besides some differences in the scoring system for cheese, fats, and beverages [13,15], there are a number of differences between the studies in terms of population (the French cohorts are younger, with a mean age of 48.9 and 43.1 years, respectively, and have experienced relatively fewer events, with 511 incident cases of metabolic syndrome and 509 major CVD events, respectively) and dietary ascertainment (the French cohorts both used repeated 24-hour recall) [10,13,14,31]. Habitual differences in diet may also contribute to differences in the finding [32,33]. It should also be noted that the scoring system is operationalised slightly differently in France, e.g., with adjustments made for diet drinks and soft cheeses.
Apart from the studies based on SU.VI.MAX and NutriNet-Santé, the Whitehall II study tested the association between the FSA-Ofcom nutrient profile model and CVD risk. However, this study focused on dietary variety (rather than quantity of consumption of less-healthy food) and found that total food variety and variety of recommended (‘healthy’) foods (but not nonrecommended foods) were associated with reduced coronary heart disease mortality and cancer morality, respectively.[34]
There are several possible explanations for the absence of an association between the FSA-Ofcom model and prospective CVD in the adjusted analyses in our study. First, our findings could be a ‘false negative’, either because of chance or because of limited power. However, fruit and vegetable consumption (as a proportion of total food energy) was significantly associated with CVD, and other EPIC-Norfolk studies have detected significant associations between dietary indices (e.g., Mediterranean Diet Score) or dietary factors (e.g., fish consumption) and incident CVD [35–37]. This suggests that the study should have sufficient power. Nonetheless, we note that the point estimate and confidence intervals observed are still consistent with a small increased hazard ratio for people in quintile group 5 compared to those in quintile group 1—i.e., the FSA-Ofcom model may be weakly associated with disease.
Second, the failure to find an association may reflect insufficient heterogeneity between the quintile groups, although quintile 5 participants consumed approximately 3 times as much less-healthy food (by weight and energy) as quintile group 1. We also note the variation in mean energy-weighted score of food (3.9 in quintile group 1 to 10.1 in quintile group 5) was greater than that observed in the French cohort, so insufficient heterogeneity seems an unlikely reason for our null findings [13]. Third, reverse causation may be a factor. We note that participants in quintile group 1 (lowest proportion of less-healthy food) appeared to be at higher cardiovascular risk (as indicated by medication, family history, and BMI). This might suggest that participants in quintile group 1 were at higher risk of CVD and were choosing to adopt a healthier eating pattern to offset this risk. Although we undertook extensive analyses to account for reverse causation, both adjustment and exclusion, we cannot rule out residual confounding and reverse causation as an explanation for our findings. As some of the covariates that we adjusted for (e.g., diagnosis of high blood pressure) could be on the causal pathway between less-healthy food consumption and CVD, adjustment for these risk factors might have attenuated a hypothetical association between less-healthy food consumption and increased incidence of CVD. However, we did not observe any associations when we excluded these risk factors from our analyses (i.e., restricted the analysis to participants who did not have indices of increased cardiovascular risk at baseline).
Finally, it is possible that our findings indicate a ‘true negative’, i.e., the FSA-Ofcom model is not, or is only weakly, associated with CVD, reflecting potential shortcomings of the model. The model was published in 2004, prior to some key advances in nutritional science [38]. The notion that saturated fat consumption is a risk factor for CVD has been challenged [39,40]. The FSA-Ofcom model may misclassify some foods because it does not account for the cardioprotective effects of mono- and polyunsaturated fats, classifying all oils, including healthier oils (e.g., olive oil) as less-healthy. The model also fails to discriminate between some healthy and less-healthy grains, e.g., between brown and white rice or between wholemeal and white bread [38]. This may explain why estimated fibre intake was not strongly patterned across the quintiles (see Table 1) despite the inclusion of fibre within the scoring algorithm. We note the FSA-Ofcom model is presently being reviewed in light of revised dietary guidelines on sugar intake [41].
While no single study is definitive, our findings do call into question the FSA-Ofcom model’s value for public health, particularly in the UK. We should emphasise that our analysis amounts primarily to an evaluation of the model’s classification of less-healthy foods, not the underlying scoring system. One should be cautious about extrapolating our findings to other variants of the FSA-Ofcom model (e.g., the New Zealand, Australian, or French versions) that apply different scoring and classification systems. We also want to emphasize that our study was not designed to test current dietary guidelines or advice around the consumption of specific ‘unhealthy’ foods. There is a robust evidence base concerning the health risks associated with the consumption of many such food groups (e.g., red meat, sugar-sweetened beverages, and takeaway food) [42–46]. On the basis of our study, it would be inappropriate to conclude that the present dietary advice about the consumption of certain foods, some of which may be labelled unhealthy or less-healthy, is incorrect.
Given the conflicting findings of our study and those based on a French cohort [11–14], further replications in other cohorts and considering other outcomes (e.g., weight change) would be of value. However, the failure to demonstrate a positive association between less-healthy food consumption and CVD in this cohort suggests the FSA-Ofcom model is not consistently discriminating among foods with respect to their association with CVD in the UK context. It may be appropriate for public health officials and scientists to review whether and how the FSA-Ofcom model could be improved for use in the UK and elsewhere, but it would not be appropriate to use the study to undermine present dietary advice.
|
10.1371/journal.pntd.0001254 | Randomized Clinical Trial on Ivermectin versus Thiabendazole for the Treatment of Strongyloidiasis | Strongyloidiasis may cause a life-threatening disease in immunosuppressed patients. This can only be prevented by effective cure of chronic infections. Direct parasitologic exams are not sensitive enough to prove cure if negative. We used an indirect immune fluorescent antibody test (IFAT) along with direct methods for patient inclusion and efficacy assessment.
Prospective, randomized, open label, phase III trial conducted at the Centre for Tropical Diseases (Verona, Italy) to compare efficacy and safety of ivermectin (single dose, 200 µg/kg) and thiabendazole (two daily doses of 25 mg/Kg for two days) to cure strongyloidiasis. The first patient was recruited on 6th December, 2004. Follow-up visit of the last patient was on 11th January, 2007. Consenting patients responding to inclusion criteria were randomly assigned to one of the treatment arms. Primary outcome was: negative direct and indirect (IFAT) tests at follow-up (4 to 6 months after treatment) or subjects with negative direct test and drop of two or more IFAT titers. Considering 198 patients who concluded follow-up, efficacy was 56.6% for ivermectin and 52.2% for thiabendazole (p = 0.53). If the analysis is restricted to 92 patients with IFAT titer 80 or more before treatment (virtually 100% specific), efficacy would be 68.1% for ivermectin and 68.9% for thiabendazole (p = 0.93). Considering direct parasitological diagnosis only, efficacy would be 85.7% for ivermectin and 94.6% for thiabendazole (p = 0.21). In ivermectin arm, mild to moderate side effects were observed in 24/115 patients (20.9%), versus 79/108 (73.1%) in thiabendazole arm (p = 0.00).
No significant difference in efficacy was observed, while side effects were far more frequent in thiabendazole arm. Ivermectin is the drug of choice, but efficacy of single dose is suboptimal. Different dose schedules should be assessed by future, larger studies.
Portal of Clinical Research with Medicines in Italy 2004–004693–87
| Strongyloidiasis is the infection caused by the worm Strongyloides stercoralis. Due to its peculiar life cycle Strongyloides may remain indefinitely in the host, if not effectively cured. Although the disease is usually mild, in case of weakening of the host's immune defenses the worm may invade virtually all organs and tissues (disseminated strongyloidiasis, almost invariably fatal). The treatment must then reach the goal of the complete elimination of the parasite. Small size clinical trials showed similar, high efficacy of the two drugs ivermectin (used as a single dose) and thiabendazole (used twice daily for two consecutive days). All trials used as the criterion for cure the absence of larvae in stool exams. The latter however may easily miss the infection, falsely suggesting that the infection has been cured. This trial, using a test detecting specific Strongyloides antibodies as an additional and more sensitive diagnostic tool, confirms previous reports: the two drugs have similar efficacy but ivermectin is better tolerated and is therefore the first choice. However the cure rate was lower than 70% for the standard, single dose. The authors then conclude that a larger, multi center trial is needed to find the optimal dose schedule of ivermectin.
| Strongyloidiasis is a chronic, soil-transmitted infection caused by Strongyloides stercoralis, a helminth with a worldwide distribution, primarily in tropical and subtropical regions. Foci of low endemicity are also reported in temperate climates, such as the Mediterranean Coast, mostly among elderly patients [1], [2]. Prevalence data indicate that 30–100 million people are infected, but the figure is presumably underestimated [3].
Due to a peculiar life cycle that includes autoinfection (maintenance of parasitism in the absence of any further exposure to an external source), the infection can persist indefinitely, usually with mild and aspecific symptoms [3]–[6]. Nevertheless, disseminated strongyloidiasis, a life-threatening condition, may occur in case of immunosuppression [7], [8]. A suboptimal efficacy of the therapy of chronic strongyloidiasis may result in the persistence of the infection, with the potential risk of disseminated disease at any time. Several reported cases of fatal, disseminated disease had previously been treated and apparently cured [7], [9], [10].
Ivermectin is currently considered the best therapeutic option [11], [12]: trials comparing ivermectin and albendazole demonstrated unsatisfactory efficacy of the latter [13]–[16], while small sized trials comparing thiabendazole and ivermectin showed similar efficacy, but better tolerability of the latter [17], [18].
All trials conducted so far have exclusively relied on direct methods [17]–[22]. Therefore, the efficacy of any regimen could have been overestimated, because negative stool tests after treatment are no proof of eradication of the infection: the sensitivity of direct methods is largely unsatisfactory [3], [23], [24]. On the other hand, serology has been suggested as a reliable tool to monitor response to treatment [25]–[30].
This study was meant to compare the efficacy of ivermectin, administered as a single dose of 200 µg/kg, and thiabendazole, administered in two daily doses of 25 mg/Kg for two days, to cure strongyloidiasis.
The protocol for this trial and supporting CONSORT checklist are available as supporting information: see Checklist S1 and Protocol S1.
This was a prospective, randomized, open label, phase III trial, carried out at the Centre for Tropical Diseases (CTD), Sacro Cuore Hospital, Negrar (Verona, Italy). Eligible patients were male and female subjects older than 5 years and weighing >15 kg, currently living in a non-endemic area; they had to have a diagnosis of strongyloidiasis established by indirect immune fluorescent antibody test (IFAT).
Exclusion criteria were: pregnancy or breastfeeding; CNS diseases; disseminated strongyloidiasis; immunodeficiency (malignancies, chemotherapy or other immunosuppressive treatments); planned travel to endemic countries before follow-up; lack of informed consent.
HIV positive subjects were excluded if CD4+ count was lower than 400/µl.
This research was conducted in full accordance to the Ethical Principles for Medical Research Involving Human Subjects as expressed in the Declaration of Helsinki and following amendments. Eligible patients were asked to meet the study investigator, who gave detailed explanation of the study protocol according to the patient information sheet and requested for written consent from the patient or, in case of minors, from her/his parent(s)/guardians. The study protocol was approved by the local Ethics Committee (Sacro Cuore Hospital Ethics Committee, 5th August, 2004). All interventions (including unscheduled visits) were at no charge to the patients.
Potentially eligible subjects attending the study clinic were identified through laboratory diagnosis of S. stercoralis infection as defined above. Indirect immune fluorescent antibody test (IFAT) was performed in accordance with the procedures described in detail elsewhere [29]. Stool agar plate culture and microscopic examination (after concentration according to Ritchie) were performed if not previously available. Baseline assessment also included routine haematology with WBC differential count and routine chemistry.
Consenting patients were admitted to the clinic for at least three days for a close monitoring of side effects; on admission a Case Report Form (CRF) was filled with the patient's unique ID number. Clinical examination and history were carried out on admission, according to the CRF. Based on the randomization list, patients were given either ivermectin or thiabendazole. Ivermectin (tablets 3 mg) was administered at the single dose of 200 µg/kg on an empty stomach, and patients were instructed to keep fasting for the following 2 hours. Thiabendazole (tablets 600 mg) was administered with food, in two daily doses of 25 mg/Kg for two days. The drug intake was directly observed by a nurse.
The patients were asked to attend the clinic twice after treatment completion: after one month and after four months. At both follow-up visits, clinical history and examination were carried out and a full blood count (FBC) was performed. At the second visit only, IFAT was performed, and so was a stool agar plate culture (if positive on recruitment). As was the routine procedure at CTD laboratory, follow-up serum samples were tested in parallel with those of the initial diagnosis. If the patient did not present for the second follow-up visit, the investigator had to contact her/him and fix another appointment. The second follow-up visit, at which the efficacy outcomes were assessed, was considered still valid up to 6 months from the treatment. Patients who did not present within the 6 months were considered lost to follow-up.
Primary objective was to compare the efficacy of ivermectin, administered as a single dose of 200 µg/kg, and thiabendazole, administered in two daily doses of 25 mg/Kg for two days, to cure strongyloidiasis. Secondary objective was to assess safety and tolerability of both regimens.
Primary outcome was cure at Time 2 (T2: 4 to 6 months after recruitment), defined as follows: negative stool agar culture for S. stercoralis (assessed in case of positivity of any direct stool tests on recruitment), AND: negative IFAT or decrease of two or more antibody titers. Secondary outcome was: patients with adverse reactions (grade 1 to 5 as defined below) to treatment.
All adverse events reported by the patients on days 1 and 2 of treatment were recorded in the patient's CRF, and so were adverse events recorded during scheduled and unscheduled visits. Adverse events for this study purpose were graded as: 0 = None; 1 = Mild: any symptoms possibly related to drug, not necessitating medication; 2 = Moderate: any symptoms possibly related to drug, requiring medication; 3 = Serious: requiring treatment to be discontinued; 4 = Near fatal: requiring intensive care; 5 = Fatal.
The sample size was determined based on the primary outcome. The trial was designed to detect a difference of efficacy of at least 15% with a study power of 80% and p<0.05 for alternative hypotheses, 2-sided and with a minimal efficacy of 70% for the less effective regimen: the required sample size was of 133 subjects in each group. Considering subjects lost to follow-up, a total of 150 patients for each treatment group was initially planned to be enrolled.
Subjects were randomly assigned to one of the following arms with allocation ratio 1∶1. Group A: ivermectin 200 µg/kg in a single dose. Group B: thiabendazole, 25 mg/Kg b.i.d for two days. The randomization list was computer-generated by a biostatistician who was not directly involved in any other operational aspect of the study and handed to the nurse in charge, who was not directly involved in the study, either, and kept the list in a locked drawer. When a patient was considered to meet the inclusion criteria and had given her/his informed consent, the patient was formally recruited by the study investigator (ZB, AA, GM, MA, MB or SM) who was on duty, who then reported the patient's unique ID number and the general data in the CRF. The nurse in charge (or her delegate in her absence) was then asked to indicate the allocation group according to the ID number and treatment was started immediately. As randomization was not in blocks, there was no way for the investigator to guess in advance what the next assigned treatment would be. More rigorous procedures (such as the use of sealed envelopes labelled with the unique ID number and containing the indication of allocation) were not judged necessary.
This was an open label trial that exclusively relied on lab values for the assessment of the primary outcome, therefore blinding of laboratory staff was ensured: the laboratory personnel performing the analyses (stool culture, serology) had no direct contact with the investigators and no information as regards the drug administered to the patient.
Data were double entered with Epi Info software (CDC Atlanta, version 3.3.2) and analysed with the same software and with Stata 9.2 (StataCorp LP, College Station, TX 77845 USA). The two randomised groups were first compared with respect to baseline demographic and clinical data. Proportions were compared through Yates' chi-square test. T test for independent groups was used for continuous variables. Mann-Whitney U test was used for non normal variables. The pattern of compliance to treatment and to follow-up visits was also explored and compliers/non compliers were compared with respect to baseline data. Patients with missing values and patients lost during treatment or at follow-up were to contribute to the analysis only for the time during which data were available.
The analysis of primary as well as secondary outcomes was planned on an intention-to-treat basis (ITT) considering all subjects as originally assigned to the two arms. As all patients were able to conclude their treatment according to plan, and as we subsequently excluded from the analysis of efficacy patients lost to follow-up whose outcome was unknown, this corresponded, de facto, to a per-protocol (PP) analysis [31].
The proportions of patients with the occurrence of the binary, primary and secondary outcomes of interest (as defined above) in each of the two arms were compared through the Yates' chi-square test with continuity correction. Fisher's exact test was used when appropriate. No subgroup analysis was initially planned. Subsequently however, a separate analysis was carried out on subgroups, in order to be able to better compare our results with those of previous trials based on direct diagnostic criteria only.
The study started with the recruitment of the first patient on the 6th December 2004, while the last one was recruited on the 3rd August 2006. At that moment, 223 patients had been included in the study. Recruitment was concluded before the required sample size was attained. The reason was the obvious difference in tolerability observed by the investigators between the two arms. Although this was not an explicitly defined criterion for the early conclusion of the study (as all observed side effects were mild to moderate), the recruitment was interrupted and an interim analysis was carried out in November, 2006, on the 187 patients who had concluded follow-up. The analysis showed a very similar cure rate between the two arms, while the frequency of side effects was much higher in the thiabendazole arm. Then, on 27th November, 2006, the decision to stop recruitment was notified to the Ethical Review Board. After that date, 11 more patients, previously recruited, presented to follow-up until January, 2007, when the data lock occurred, after the second follow-up visit of the last patient (11thJanuary, 2007), therefore the final analysis of the primary endpoint concerned 198 subjects.
The flow of patients assessment and enrollment is reported in the study flow diagram (Figure 1). Briefly, out of 283 patients initially screened for inclusion, 242 were eligible for inclusion, of whom 223 gave their written (or their guardians') informed consent and were recruited. Of the patients recruited, 115 (51.6%) were randomly assigned to ivermectin arm and 108 (48.4%) to thiabendazole. Follow-up was completed by 198 patients (88.8%), 106 (53.5%) assigned to ivermectin and 92 (46.5%) to thiabendazole.
The main baseline characteristics of the randomized population (223 patients) is reported in Table 1. None of the observed differences between the two groups was statistically significant.
In the following table the baseline characteristics of the patients ultimately analysed for efficacy are compared with those of patients lost to follow-up (Table 2). All patients lost to follow-up (25/25) belonged to the group of “residents overseas”, which included immigrants. Compliance to follow-up was higher for ivermectin (106/115 or 92.2%) than thiabendazole (92/108 or 85.2%), but the difference was not statistically significant (p = 0.15).
We first assessed the efficacy on all 198 subjects included (Table 3) who were assessed at follow-up. Based on the primary endpoint (all criteria fulfilled), the subjects cured were 60/106 (56.6%) and 48/92 (52.2%) in ivermectin and thiabendazole arm, respectively (p = 0.53). If we considered as cured, with less stringent criteria, also the subjects with a partial response (negative stool culture and decrease of only one IFAT titer), efficacy would rise to 75/106 (70.8%) and to 67/92 (72.8%), respectively (p = 0.75). If we considered as criteria of cure the direct methods only (negative stool culture at follow-up in subjects who were positive at microscopy and/or culture on recruitment), efficacy would be 30/35 (85.7%) and 35/37 (94.6%), respectively (p = 0.21). We then did the same analyses on the subgroup with IFAT titer ≥80 (virtually giving no false positive results) on recruitment (Table 4). On this sub group, comprising about half the total sample (92 subjects), all criteria were fulfilled by 32/47 subjects (68.1%) in ivermectin arm and 31/45 (68.9%) in thiabendazole arm (p = 0.93). Including subjects with a partial response as defined above, efficacy would be 41/47 (87.2%) and 40/45 (88.9%), respectively (p = 0.81). Finally, considering direct methods only in this subgroup, the cure rate would be 22/24 (91.7%) and 27/27 (100%), respectively (p = 0.22).
As side effects of the two drugs are known to be limited in time, we considered for this outcome all 223 patients included and not only those who completed the follow-up. No serious side effect (grade 3 or more) was observed in any patient. Overall, 103/223 patients complained of any side effect, grade 1 to 2 (Table 5). In ivermectin arm, side effects were observed in 24/115 patients (20.9%), versus 79/108 (73.1%) in thiabendazole arm (p = 0.00). Only 5/115 (4.3%) patients in the ivermectin arm presented effects of grade 2 (requiring medication), while in thiabendazole arm 43/108 patients (39.8) presented effects of grade 2 (p = 0.00). Dizziness was the most frequently reported side effect both in thiabendazole arm (57/79 or 72.2%, followed by nausea and vomiting) and in ivermectin arm (12/24 or 50.0%, followed by day somnolence) (data not reported in tables).
This was the first trial on strongyloidiasis treatment using serology along with direct methods for case inclusion and assessment of efficacy. The latter, based on primary outcome, was lower than 60%, with no significant difference between the two treatment arms. With less strict criteria (including partial response as defined above), efficacy would rise to above 70% for both regimens, still with no significant difference. As the specificity of IFAT, though very high, is not 100% for the lower dilutions [29], the inclusion of some false positives may have occurred and partly explain the low efficacy found. We then analyzed a sub group of patients who had a serologic titer ≥80 (virtually giving no false positive results) [29] on recruitment. Efficacy as defined by primary outcome, and efficacy including partial response as defined above, were significantly higher in this subgroup for both regimens (close to 70% and to 90%, respectively), suggesting a more correct case inclusion. Finally, when we analysed only the patients who had positive stool tests on inclusion, taking culture negativization as a criterion for cure, the efficacy was close to or higher than 90% for both regimens, approaching that found by other studies [15], [17]–[19], [22]. Thus, the lower efficacy found by our study is clearly due to the more strict criteria used to define cure, that include serology. Our data confirm that serology tends to decrease in titer within a few months of effective treatment and can thus be a useful tool for treatment monitoring as was previously suggested [25]–[30]. Considering subjects with serologic titer ≥80 on inclusion, almost 90% had a drop of titer following treatment.
Whatever the criterion used, we were not able to find any significant difference between thiabendazole and ivermectin at standard dose for the cure of S. stercoralis infection. This finding confirms previous, smaller trials [17], [18].
Both drugs appeared to be safe, with no serious side effect in either treatment arm. Nevertheless, thiabendazole caused significantly more side effects and of higher grade.
This trial was not double blind. This cannot have affected the assessment of efficacy, as the primary outcome was entirely based on laboratory investigations and lab staff was kept unaware of the treatments administered. Contrarily, side effect reporting might have been influenced both by the investigator's and the patient's knowledge of the drug received. Results are therefore to be taken with some caution, though the difference between the two arms was clearly too big to be entirely attributable to bias.
Inclusion criteria allowed the recruitment of patients with negative direct tests on stool and positive serology at any IFAT titer. As discussed above, this probably caused some patients without the infection to be erroneously included with a consequent underestimation of the efficacy of both drugs. Subsequent analysis showed that more strict criteria (based on a minimal required cutoff of dilution) should be followed for trial inclusion. We believe the analysis of the subgroup of subjects with IFAT titer ≥80 to provide the more reliable estimate of efficacy. Given this more strict inclusion criteria, the analysis still failed to show any significant difference between the two regimens, but the sample size was originally calculated only to detect a 15% difference between the original groups.
Finally, we still remain with the problem of the lack of a gold standard to define cure. We think that serology should also have a role at least in a scenario like ours, with most probably no more local transmission, where the interpretation of the results is not potentially confounded by reinfection. While awaiting alternative diagnostic methods such as real time PCR [32] to become a reliable alternative, the best option is probably the combination of direct with indirect methods, but the latter need further study to identify the optimal serologic test and cutoff for diagnosis, trial inclusion and treatment follow-up.
Ivermectin is the treatment of choice due to better tolerability, but the single dose efficacy is sub optimal. Some guidelines and the WHO drug formulary have already shifted to a new schedule (200 µg/Kg/day for two consecutive days) [11], [12], while some experts recommend the repetition of treatment after two weeks, on ground of the parasite life cycle [33]. Though it seems reasonable to expect that the use of an increased dose would improve the efficacy of ivermectin, neither of these alternative regimens has ever been validated by a randomized trial to our knowledge, and the last published trial [13] failed to show any significant difference between the single dose and two doses two weeks apart.
Considering that no truly promising new drug is in the pipeline, we are planning a multi center trial on different dose schedules of ivermectin.
|
10.1371/journal.pgen.1004322 | Genetic Dissection of the Drosophila melanogaster Female Head Transcriptome Reveals Widespread Allelic Heterogeneity | Modern genetic mapping is plagued by the “missing heritability” problem, which refers to the discordance between the estimated heritabilities of quantitative traits and the variance accounted for by mapped causative variants. One major potential explanation for the missing heritability is allelic heterogeneity, in which there are multiple causative variants at each causative gene with only a fraction having been identified. The majority of genome-wide association studies (GWAS) implicitly assume that a single SNP can explain all the variance for a causative locus. However, if allelic heterogeneity is prevalent, a substantial amount of genetic variance will remain unexplained. In this paper, we take a haplotype-based mapping approach and quantify the number of alleles segregating at each locus using a large set of 7922 eQTL contributing to regulatory variation in the Drosophila melanogaster female head. Not only does this study provide a comprehensive eQTL map for a major community genetic resource, the Drosophila Synthetic Population Resource, but it also provides a direct test of the allelic heterogeneity hypothesis. We find that 95% of cis-eQTLs and 78% of trans-eQTLs are due to multiple alleles, demonstrating that allelic heterogeneity is widespread in Drosophila eQTL. Allelic heterogeneity likely contributes significantly to the missing heritability problem common in GWAS studies.
| For traits with complex genetic inheritance it has generally proven very difficult to identify the majority of the specific causative variants involved. A range of hypotheses have been put forward to explain this so-called “missing heritability”. One idea—allelic heterogeneity, where genes each harbor multiple different causative variants—has received little attention, because it is difficult to detect with most genetic mapping designs. Here we make use of a panel of Drosophila melanogaster lines derived from multiple founders, allowing us to directly test for the presence of multiple alleles at a large set of genetic loci influencing gene expression. We find that the vast majority of loci harbor more than two functional alleles, demonstrating extensive allelic heterogeneity at the level of gene expression and suggesting that such heterogeneity is an important factor determining the genetic basis of complex trait variation in general.
| Uncovering the genetic basis of quantitative phenotypes is a central, yet unresolved problem in biology. There is a major discrepancy between the heritability estimates of most quantitative traits and the amount of heritable variation accounted for by all variants localized to a causative site. This phenomenon is often referred to as the “missing heritability” problem. Several hypotheses have been offered as possible explanations, including widespread epistasis [1], the infinitesimal model (many, very small effect loci influencing the phenotype of interest that are difficult to detect statistically) [2]–[4], rare alleles of large effect, that are also statistically difficult to detect [5]–[7], and widespread allelic heterogeneity (many independent effects segregating at each causative locus) [7]. This quest to understand the genetic basis of complex traits has given rise to a community-based strategy of creating freely-available genetic resource populations in model organisms such as mice [8]–[10], Arabidopsis thaliana [11], [12], maize [13]–[16], and Drosophila melanogaster [17]–[20]. Those organisms with the greatest genetic resources and with a community of researchers focused on a single system provide a logical starting point toward finding the missing heritability associated with quantitative phenotypes. In addition, the experimental designs of some of these resources are well suited to test different hypotheses for the sources of missing heritability. For example, Bloom et al. [21] used a large segregant pool from a two line yeast cross to demonstrate that epistasis is not a major contributor to the heritability of most traits. In particular, resources that have a well-defined multi-haplotype structure can be used to identify the extent of allelic heterogeneity [22] owing to the ability to estimate trait means for each haplotype at each mapped QTL. By focusing effort on these community resources, the hope is that we will gain a better understanding of the causes of missing heritability problem.
Much of the genetic variation underlying whole organism phenotypes is thought to be due to regulatory variation, i.e., variants influencing gene expression [23]–[26]. Causative loci are linked to whole organism phenotypes through the transcriptome, an interrelated network of transcripts whose abundances influence the resulting phenotype. The transcript abundances of most genes are quantitative traits themselves and have heritabilities comparable to typical whole-organism phenotypes [24], [26], [27]. Increasingly, expression quantitative trait locus (eQTL) mapping is being used to identify the source of genetic variation in transcript abundances with the ultimate goal of linking variation at the nucleotide level to variation in gene expression and to variation in visible phenotypes. Expression QTL studies have shown that most genes have local (cis) eQTL that tend to be located near the transcription start site and to be of fairly large effect. Distant regulatory effects (trans-eQTL) are more difficult to identify, likely because they are more numerous and are of smaller average effect, leaving a great deal of variation in transcript abundance unexplained [23], [24], [26], [27]. There is a growing movement toward identifying the causative quantitative trait nucleotides (QTN) underlying cis-eQTL, often with the assumption there is a single causative site [28]–[30]. However, if most eQTL harbor allelic heterogeneity [31], identifying a single causative variant will cause researchers to miss a significant portion of the genetic variation [7].
Here we describe transcriptome-wide mapping in female head tissue in the Drosophila Synthetic Population Resource (DSPR) [17], [18], one of the major genetic reference panels in the Drosophila model system. Our goals are two-fold. First, we aim to provide a comprehensive map of cis- and trans-eQTL for female head tissue in the DSPR. A key advantage of genetic reference panels is the potential to integrate phenotypes measured at multiple levels on genetically identical individuals. Incorporating eQTL data with visible phenotype data can increase mapping power and help users identify candidate genes [9], [23], [25], [32]. Second, we use the large set of discovered eQTL to quantify the number of alleles segregating at each causative locus, providing an evaluation of the degree of allelic heterogeneity at both cis- and trans-eQTL. The hypothesis that allelic heterogeneity is prevalent in quantitative traits has not been tested directly, in part because it is difficult to do so using a genome-wide association (GWAS) framework. Within loci, linkage disequilibrium makes it very difficult to distinguish between two SNPs tagging two independent causative sites versus a single causative site. In addition, the step-wise regression approaches used, for example [2], [33], to identify multiple SNPs in a gene region associated with a phenotype lack power. The result is that the majority of GWAS that have identified multiple SNPs at a single locus using conditional analysis rarely identify more than two such SNPs despite very large sample sizes e.g. [2] but see [33]. In contrast, mapping in the DSPR and other multi-parental advanced generation intercross mapping panels take a haplotype based approach, providing a natural way to distinguish between multiple alleles at each QTL and a way to ascertain the potential contribution of allelic heterogeneity to the missing heritability problem.
We mapped genome-wide expression variation using trans-heterozygote F1 individuals from 596 crosses between DSPR population A (pA) females and population B (pB) males, thus avoiding mapping variation for inbreeding depression. Gene expression was assayed using Nimblegen 12×135 K arrays, and we analyzed the resulting data using a custom data analysis pipeline (see methods) to identify all significant eQTL.
We identified a total of 7922 eQTLs corresponding to 7850 transcripts out of a total of 11064 transcripts tested (Figure 1). Details for all eQTLs are in Table S1. Of these, 7704 transcripts were associated with a single cis-eQTL, 71 were associated with both cis- and trans-eQTL, and 75 were associated with only trans-eQTL. A small percentage of eQTLs (∼7%; Table 1) were associated with only a single recombinant inbred line (RIL) population (pA or pB; see methods), but for most eQTL fitting both pA and pB was necessary to explain the eQTL signal, indicating that causative variants were present in both populations.
The amount of variation explained by our mapped eQTLs was high (Figure 2), though our stringent, experiment-wise permutation-based correction for multiple tests severely limits our ability to detect QTL of small effect. Not surprisingly, the variance explained by cis-eQTLs was higher than trans-eQTLs [24]. Our cis-eQTLs explained a median of 24% of the phenotypic variance, and 855 eQTL explained more than 50% of the phenotypic variance. Using our heritability estimates for each transcript abundance, we estimated the percentage of the heritability each eQTL explained. The median for the percent heritability explained by each eQTL was 73%. Our trans-eQTLs explained lower levels of variance, the median phenotypic variance explained was 15%, and the median percent heritability explained was 38%. However, if heritability values are underestimated, and/or we overestimate the effects of eQTLs (which is likely due to the Beavis effect [34]), these values will be inflated. This effect is obvious for the set of eQTL estimated to explain greater than 100% of the heritability (Figure 2A).
Our mapping resolution was high (Figure 3). We used two methods for estimating confidence intervals, a 3 LOD drop and the Bayesian credible interval. We excluded confidence intervals that spanned centromeres or occurred near telomeres, because these tend to cover very large physical distances (7% of eQTLs). The Bayesian credible intervals tended to be narrower than 3 LOD drops (median BCI = 110 kb, 0.25 cM; median 3 LOD drop = 240 kb, 0.51 cM), but the range was larger for BCIs (BCI: 0–4.5 Mb, 0–6.5 cM; 3 LOD drop: 20 kb–4.0 Mb, 0.001–3.9 cM). The median number of genes within cis-eQTL CIs was 32 (range 1–551), and within trans-eQTL CIs, the median was 44 (range: 5–479).
We have provided a comprehensive map of eQTLs for female head tissue in the Drosophila model system within the constraints of our statistical power. There is little doubt many smaller effect eQTLs exist that we were not able to identify given our conservative statistical threshold. Our use of trans-heterozygote individuals means that we not only avoid the effects of inbreeding depression, but we have also obtained estimates for all eQTL for both pA and pB DSPR populations. Overall, our results confirm what many other researchers have observed, widespread large effect cis-eQTLs and smaller effect trans-eQTLs [23], [24], [26], [27]. One of the major advantages of a stable genetic panel is the ability to measure multiple traits at multiple levels on genetically identical individuals, which allows for the potential to combine these sources of data to identify causative genes [9], [23], [25], [32]. We expect this dataset to be very useful to DSPR users, particularly those interrogating phenotypes measured in females with relevance to neuroanatomy or behavior. All of the raw and analyzed data are freely available at http://FlyRILs.org/Data. The data have also been deposited in NCBI's Gene Expression Omnibus [35] and are accessible through GEO Series accession number GSE52076.
We identified regions of the genome associated with a high trans-eQTL density to identify eQTL regulating the expression of several other genes (trans hotspots). There were two regions of high trans-eQTL density, TQTLA and TQTLB (Figure 4; Table 2). These clusters regulate several genes distributed throughout the genome, as is apparent in Figure 1. We used a gene ontology term finder [36] to determine whether the sets of genes regulated by these trans-eQTL were related to a common process. The set of 16 genes regulated by TQTLA showed enrichment for circadian rhythm of gene expression (2 of the 16 genes regulated by TQTLA; P = 0.0007). We used principal components analysis on the set of 16 genes to create a composite variable. All 16 genes load fairly evenly on the first principal component (absolute value range: 0.08–0.20). We then correlated this composite variable with expression measures for each gene in the TQTLA region to identify possible candidate genes. The gene timeless (tim) was highly correlated with the TQTLA composite variable (r = 0.90), and it does have a significant cis-eQTL. All other genes in the interval had a correlation with an absolute value of less than 0.5. Additionally, after correlating the expression of each of the 16 transcripts regulated by TQTLA with the expression of all genes in the TQTLA region, timeless showed the maximum pairwise correlation in all 16 cases (absolute value of correlation range:0.35–0.84). The estimated haplotype means follow this pattern and are correlated with the estimated effects for the timeless cis-eQTL in most cases (average absolute value correlation: 0.65; min: 0.03; max: 0.99). The gene timeless (tim) is expressed in the adult central nervous system [37] and is involved in transcriptional regulation of circadian rhythm [38].
Not all genes in the TQTLA interval are included in our expression set. For example, some genes may have been dropped due to the presence of SNPs in probes, or were not included in the Nimblegen probe set to begin with. For TQTLA, 23 genes in the interval are not represented in the expression set. However, none of these genes are associated with any terms involving circadian rhythm, regulation of gene expression, or transcription (http://FlyBase.org) [39], and we therefore do not consider any of these likely candidate genes.
The genes associated with TQTLB are enriched for several GO terms including compound eye pigmentation (2/11 genes; P = 0.005), the umbrella term: single-organism metabolic process (6/11 genes; P = 0.007), and several specific metabolic process terms: tryptophan metabolic process (2/11 genes; P = 0.008), indolalkylamine metabolic process (2/11 genes; P = 0.0008), indole-containing compound metabolic process (2/11 genes; P = 0.002), aromatic amino acid family metabolic process (2/11 genes; P = 0.006). Once again we performed PCA to create a composite variable. sugarbabe (sug) was the gene most highly correlated with the TQTLB composite variable (r = −0.63) and does have a significant cis-eQTL. All other genes in the interval had a correlation with an absolute value of less than 0.4. Loadings were again fairly even (absolute value range for all other genes: 0.08–0.39). Pairwise correlations between the transcripts associated with TQTLB and the expression measures in the interval showed sugarbabe to be most highly correlated in all cases except two: gene CG5321 and gene CG6834 (absolute value of correlation range for all other genes: 0.40–0.52). These two genes were also the two with the lowest loading values on the composite variable. The correlation between the estimated haplotype effects for the cis-eQTL for sugarbabe, and the effects for the trans-eQTLs were moderate (mean absolute value correlation: 0.24; min: 0.005; max: 0.44). The gene sugarbabe (sug) is expressed in the adult head [37], is involved in regulation of transcription [40], is involved in regulation of response to starvation [41], and is part of the insulin-like growth factor signaling pathway [41]. The 21 genes not included in the interval are not associated with any terms involving metabolism, regulation of gene expression, or transcription (http://FlyBase.org) [39].
We have identified two trans hotspots, and, in both cases, we were able to use our expression dataset to narrow the causative gene to a single likely candidate gene. Previous eQTL studies have identified many more trans hotspots that regulate many more genes (hundreds or thousands) than our two identified hotspots (TQTLA: 16 genes; TQTLB: 11 genes; e.g. [27], [42], reviewed in [24], [26]). However, while some of these global regulators of gene expression have been confirmed as true signals, most notably in yeast [43], [44], Kang et al. [43] show how hotspots can result from confounding factors such as batch effects. In our dataset, we employed PCA to correct for possible batch effects [45]. This method has been shown to increase power to detect eQTL [29], [45], [46], however, it makes identifying even true trans global regulators impossible. The signal that results from a global regulator is statistically indistinguishable from an unmeasured batch effect. In addition, even true global regulators can confound the detection of other true eQTLs, and correcting for these true global regulators increases the power to detect these other associations [43], [45]. It is possible to distinguish true trans hotspots from batch effects using biological replicates [43], but for our study we chose to maximize the number of RILs rather than increase replication to maximize our statistical power to map eQTL. As a result, we are unable to detect many trans hotspots in this study. However, our stringent statistical correction does give us increased confidence that the eQTL we do identify are indeed true signals.
The vast majority of our eQTLs appear to be multiallelic (Figure S1). In 95% of cases, the number of alleles estimated at cis-eQTL was 3 or greater. For trans-eQTL this percentage was somewhat lower, at 78%. Figure 5 shows an example of an eQTL where the best model is a two allele model and of an eQTL where the full haplotype model is the best model. In cases where we estimated multiple alleles, we were able to explain additional phenotypic variance compared to the best two allele model (Figure S2), sometimes as much as an additional 27%. We investigated our ability to accurately estimate the number of alleles by performing a simulation designed to provide the highest power to distinguish between different alleles (see methods). Our simulation revealed that our estimator underestimates the number of alleles in 63% of cases, correctly estimates the true number of alleles in 26% of cases, and overestimates the number of alleles in 10% of cases (Figure 6). This bias toward underestimating the number of alleles gets increasingly severe as the true number of alleles increases. Our simulations with a lower effect size (5%) and normally distributed allelic effects both resulted in an even stronger bias toward underestimating the true number of alleles. Our allele number distribution for cis-eQTLs is no doubt composed of a mixture of eQTLs of varied numbers of true alleles. Overall, it is closest to the distribution we obtain for a simulation of ∼5 alleles. So while most of our estimates for cis-eQTL are for 3–4 alleles, many may be determined by many more alleles.
Our results indicate widespread allelic heterogeneity for both cis- and trans-eQTLs. The focus of mapping studies is often to identify the single causative variant underlying a significant signal, the implicit assumption being that the causative loci are biallelic. cis-eQTL in particular, with their large effects, are thought to be more likely than other traits to have a simple genetic architecture and be biallelic [22], [28]–[30]. Baud et al. [22] found some support for this idea when comparing a two allele model to the full haplotype model in hippocampus eQTLs in the heterogeneous stock mouse resource [32]. They found that in 97% of cases, the two allele model was superior for cis-eQTLs while trans-eQTLs were more likely to be multiallelic [22]. However, in contrast to these findings, cis-eQTLs have been found to be multiallelic in Drosophila [47], Arabidopsis [42], and humans [31], [48]. Our results strongly confirm the result of multiallelism in Drosophila with 95% of cis-eQTLs estimated to be due to 3 or more alleles. This result indicates that in Drosophila, widespread allelic heterogeneity exists at one of the most basic levels of genetic variation: cis-regulatory variation.
Widespread allelic heterogeneity is one potential explanation for the missing heritability problem in the study of complex traits. Allelic heterogeneity presents a statistical challenge for GWAS [7]. GWAS utilize natural populations and interrogate each SNP (or other specific variant) for association with the phenotype of interest. At the single gene level, it is difficult to distinguish between simple linkage disequilibrium between a single causative variant and other, nearby neutral SNPs, and multiple independent causative SNPs. If GWAS focus only on the strongest association at a locus, in the presence of allelic heterogeneity that individual variant will account for less of the variation than the entire gene, causing the effect of the locus to be underestimated [7]. In this respect, haplotype-based mapping approaches, such as the one described here, have an advantage because entire haplotypes (and thus an entire set of causative variants associated with a single gene) are tested together. The effect size associated with the causative gene will tend to be larger and easier to detect in this framework. This effect, combined with the more favorable frequencies of alleles in linkage based panels could explain why these studies tend to explain very large proportions of the heritable variation [9], [21], [49], while GWAS grapple with large amounts of missing heritability. However, one drawback of current haplotype-based methods is that they do not have single gene resolution and therefore identifying the causative gene within the QTL interval can be a significant challenge. Furthermore, while identifying the causative loci under allelic heterogeneity is easier with haplotype based methods, the subsequent identification of the causative SNPs within the loci is made much more complicated by heterogeneity [17], [18], [50].
Allelic heterogeneity is typical for Mendelian diseases (http://www.omim.org/) and it has been suggested as the likely model for quantitative traits [51]. There is a growing body of empirical [2], [17], [22], [31], [42], [47], [50] and theoretical [7] support for this idea. For example, one of the largest GWA studies found support for allelic heterogeneity for human height by identifying several cases of multiple SNPs likely associated with the same gene [2]. Even age related macular degeneration, the first successful GWA study [52], has subsequently been shown to harbor multiple functional alleles [53]–[56]. Our results should therefore not be surprising. However, they do suggest the community should focus on developing experimental designs and analytical methods, e.g., [7], that function well under a model of allelic heterogeneity.
We used RILs from the DSPR (http://FlyRILs.org) to map genome-wide expression variation. The DSPR has been described in detail previously. Complete details of the development of the DSPR, founder whole genome re-sequencing, and RIL genotyping are described in [17]. The development of the hidden Markov model to infer the mosaic structure of the RILs and the power and mapping resolution of the DSPR for QTL mapping are described in [18]. Briefly, the DSPR is a multi-founder advanced intercross panel consisting of a set of over 1700 RILs of Drosophila melanogaster. Two 8-way synthetic populations (pA and pB) were created from two independent sets of 7 inbred founder lines (A1–A7 or B1–B7) with one additional line (AB8) shared by both populations. Each synthetic population was maintained as two independent replicate subpopulations (pA1 and pA2 or pB1 and pB2), kept at a large population size, and allowed to freely recombine for 50 generations. At generation 50, each subpopulation gave rise to ∼500 RILs via 25 generations of full-sib mating. The genomes of the original fifteen inbred founder lines have been completely re-sequenced, and the complete underlying founder haplotype structure of all RILs in the panel has been determined via Restriction-Associated DNA (RAD) sequencing along with a hidden Markov model (HMM).
In order to avoid potentially mapping QTL for inbreeding depression, we phenotyped trans-heterozygote F1 individuals from crosses between pA females and pB males. The crosses were done to maintain the subpopulation structure by crossing pA1 to pB2 and pA2 to pB1. In both cases, we arbitrarily crossed pA and pB RILs with the same line number (i.e., pA11*pB21, …, pA1n*pB2n, pA21*pB11, …, pA2n*pB1n). For each of 596 crosses, we generated 4–6 replicate cross vials containing 10 virgin pA females and 10 pB males and cleared the adults after 24–48 hours to maintain roughly equal larval density across experimental vials. Both the inbred RIL parents and the experimental trans-heterozygous cross progeny were raised on standard cornmeal-yeast-molasses media at 25°C, 50% relative humidity, and on a 12∶12 light∶dark regime.
Progeny from each cross vial were allowed to emerge and mate in the source vial for 2–4 days. Then 250–300 females were harvested over CO2 from the multiple replicate vials. Since we did not isolate virgin females on eclosion, females are very likely mated. These experimental females were kept for 24 hours in fresh vials to minimize any effects of the anesthesia before the heads were isolated (3–5 days old). Heads were removed by transferring the females without anesthesia to a 50 ml conical bottom centrifuge tube, freezing in liquid nitrogen, vigorously vortexing, and sieving using dry ice-chilled brass analytical sieves (mesh sizes 0.0165 and 0.0278 inches), separating heads from bodies and from legs and wings. Head samples were stored at −80°C until RNA isolation.
We did not have any technical or biological replicates aside from the effect of pooling 250–300 individuals, collected from multiple source vials, for each sample. This was intentional because we are mainly interested in the variance among RILs. There were two exceptions to this lack of replication. Crosses A1.299×B2.299 and A1.350×B2.350 were prepared independently twice.
RNA was isolated using TRIzol Reagent (Life Technologies), cleaned up using RNeasy Mini spin columns (Qiagen), concentrated—if necessary—using a vacuum centrifuge, and shipped to the Carver Center for Genomics Microarray Center at the University of Iowa for cDNA synthesis and array hybridization. We used Nimblegen 12×135 K arrays to assay genome-wide gene expression. These arrays assay 16,637 transcripts with eight 60 bp probes per transcript. Each array holds 12 different crosses.
All data analysis was performed in R [57]. Initially, we performed standard quantile normalization and corrected for background effects using the normalize and backgroundCorrect functions in the oligo package to correct for any overall array effects [58]–[61]. We then created a custom probe-to-transcript map using the most recent version of the CDS file available at FlyBase (v. 5.48). We blasted all probe sequences against the CDS, requiring an exact match [62], [63]. We eliminated any probe sequences without an exact 60 bp match to a transcript (6842 probes). We did not require a unique match given many transcripts from the same gene share portions of their sequences. Thus a single probe can correspond to multiple transcripts.
Single nucleotide polymorphisms in probe sequences are known to affect array hybridization and thus expression measurement [64]–[68]. We took advantage of the availability of full genome sequences for all 15 founder lines to identify SNPs within probe sequences. We first updated the alignment and SNP calling for the founder re-sequencing data using the Burrows-Wheeler Aligner (BWA) [69] with the following switches: -m 50000000 -R 5000, followed by the SAMtools [70] mpileup command (the initial alignment used Mosaik and a custom SNP caller, see [17]) to obtain an accurate, comprehensive list of SNPs in the founder lines (http://FlyRILs.org/Data, Release 3). We also applied the following filters: 1) at least one founder was fixed for the minor allele and at least three founders were fixed for the major allele (given a coverage of 10×), 2) minimum overall coverage of 90 (5 per sample), and 3) maximum overall coverage of 3600. A large proportion of our probe sequences contained SNPs segregating in the set of DSPR founder lines. Because we have the full genome sequences in silico of all RILs in the panel, we were able to identify all positions in probes that are SNPs in our RIL panel and test for the effect of each SNP on the expression measurement. We discarded any probes containing multiple SNPs (22018 probes). For probes containing a single SNP, we used the haplotype probabilities from the hidden Markov model to infer the probability each RIL harbored the minor allele and assigned a genotype value to each cross by adding the paternal and maternal probabilities. In the case of perfect certainty, genotype values are: 2 = AA, 1 = Aa, and 0 = aa. We then tested for the effect of the SNP on the expression measurement by fitting the following model:where y is the expression measurement, S is subpopulation, M is the cross genotype at the marker, and βs and βm are the corresponding effect estimates. We then eliminated all probes with a p-value less than 0.05 (21141 probes).
Following re-mapping of probes and elimination of probes with SNPs affecting expression, transcripts were associated with a variable number of probes instead of each transcript being associated with exactly 8 probes as in the original NimbleGen array design. We eliminated any transcript associated with fewer than four probes. Next, we performed standard RMA using the basicRMA function in the oligo package [61] to combine probe-specific data and generate a single expression measure per transcript. Many genes are associated with multiple transcripts. Whether the expression of different transcripts can be independently assessed is dependent on how many probes uniquely map to each transcript. We calculated pairwise correlations between each transcript in each set of transcripts associated with a single gene. If all of the pairwise correlations between the set of transcripts were > = 0.95, we used the average expression for the gene. Otherwise, we mapped each transcript separately. We will refer to all expression measures (including those averaged across transcripts for a single gene) simply as transcripts for clarity.
We followed the methods of [29], [46] and used principal components analysis (PCA) to minimize batch effects [45] and increase our power to detect QTL. Following quantile normalization of each transcript to coerce each transcript distribution to be normal, we performed PCA on the entire set of transcripts. We selected the first 10 principal components to correct our expression measurements. The percentage of the variance explained by each remaining principal component was below 1% (Figure S3). We then fit the following modelwhere yi is the ith expression measurement, S is subpopulation, xj is the jth principal component, and βs,i and βj are the corresponding effect estimates. We used the resulting residuals for the remaining analyses. We performed an additional round of quantile normalization on these residuals to ensure normality.
We estimated the narrow-sense heritabilities for all transcripts by fitting a linear mixed model using the polygenic function in the GenABEL package [71]. Briefly, the model includes a random effect polygenic term whose variance is determined by the kinship matrix between RIL crosses. We calculated the kinship matrix using the genome-wide haplotype assignments resulting from the HMM. At each position spaced every 0.025 cM, we calculated the probability of identity by decent and averaged these across the genome to obtain the relationship coefficient. Our kinship matrix is thus estimated over genetic distance. We then used the polygenic function to calculate heritabilities for each transcript [71].
To map eQTLs, we first selected transcripts expressed above background levels. We utilized the two replicated samples, A1.299×B2.299 and A1.350×B2.350, to identify the point where measurements were less repeatable and excluded all transcripts with expression levels below this point (Figure S4). This cutoff excluded approximately 23% of transcripts. For all included transcripts, we performed haplotype-based genome scans by fitting the following model at regularly spaced positions every 10 KB across the genome (11768 positions; http://FlyRILs.org/Data, Release 3).where yr,i is the ith transcript, μ is the grand mean, GA,j are the genotype probabilities for the jth paternal RIL, GB,j are the genotype probabilities for the jth maternal RIL, and βA,j, and βB,j are the corresponding effect estimates. Because we assayed only females, the model for the X chromosome is the same as for the autosomes. At each position, we calculated the F-statistic for the overall effect of genotype and obtained LOD scores.
To identify the statistical significance threshold, we performed 1000 permutations of the expression measures [72]. The entire set of expression measures was permuted together to maintain the correlation structure in the dataset. We used these permutations to determine a conservative genome-wide, experiment-wise 5% significance threshold (threshold = 14.99). We also determined a separate threshold for cis-eQTL. We defined cis-eQTL as QTL occurring within 1.5 cM of the transcription start [18] site for each transcript (1.5 cM is our typical confidence interval width). To define a cis-only threshold, we only included the LOD scores for the positions within 1.5 cM of the transcription start for each gene (threshold = 14.4).
We identified all peaks with LOD scores exceeding the above-defined thresholds. When multiple nearby peaks were identified, we determined whether their 3 LOD drop intervals overlapped, and, if so, only the peak with the highest LOD score was retained. We expect 3 LOD drops to be a conservative estimate of the 95% confidence interval. Standard 2 LOD drops have been shown to be overly narrow for pA×pB cross designs [18]. It should be noted however, that confidence intervals on QTL locations are not true 95% confidence intervals and effect size, sample size, and the number of haplotypes in the model affect the degree of coverage. We also calculated Bayes credible intervals, for which 95% coverage tends to be more consistent [73], [74].
In a pA×pB cross, a mapped QTL may be due to genomic variation at that position in only one population or in both. We identified peaks associated with only a single population using Akaike's Information Criterion (AIC). We calculated the AIC for three models: pA alone, pB alone, and pA & pB. The smallest AIC indicates the model with the best fit. Thus any cases in which the lowest AIC resulted from a reduced model, the QTL peak was concluded to be due to variation in a single population.
We identified trans-eQTLs influencing multiple transcripts by estimating the trans-eQTL density across the genome using a 500 kb sliding window with a step size of 1 kb. Our estimate of density included only unique genes, not transcripts to avoid counting multiple transcripts associated with a single gene as independent events. If trans-eQTL density in a window exceeded the density expected by chance under a Poisson distribution, we concluded it was a significant trans hotspot. This threshold for a Poisson distribution given the total number of trans-eQTLs (147), the window size (500 kb), the size of the genome tested (118 Mb) and the Bonferonni corrected P-value threshold (117,741 tests; P = 4.2×10−7) is a trans-eQTL density greater than 6. We delineated the size of these hotspot regions as the lowermost and uppermost confidence interval bound for any trans-eQTL peak included in a window exceeding a density of 6.
Our initial scan identified 3 trans hotspots but upon further investigation, we determined one to be a false signal resulting from a single gene family. All of the eQTL peaks associated with this hotspot represent 13 members of a single gene family located on the X chromosome: Stellate (Ste). In addition, members of this family also occur at an unlocalized region in the heterochromatin on the X chromosome. The “trans-” eQTL we map regulating this family is located at the very tip of the X chromosome, making it very likely we are tagging this heterochromatic location of Stellate members, and it is in fact an additional cis effect. In fact, all thirteen members show two peaks, one cis peak and a second “trans” peak at the tip of the X, indicating most of our probes for these genes are tagging multiple members of this gene family. In addition, Stellate is expressed in adult males and involved in spermatogenesis (http://FlyBase.org) [39]. It is likely we are seeing high expression due to large numbers of copies of gene family members (∼200 copies) [75]. We therefore excluded this trans hotspot.
We estimated the number of alleles at each eQTL using a model comparison technique similar to the method employed by Yalcin et al. [76] and Baud et al. [22] The major difference in our approach is that we consider models with more than 2 alleles and do not restrict our analysis to specific SNPs in the QTL interval. The merge analysis employed by Baud et al. [22] considered all two allele models associated with a single SNP within the QTL interval. We simply assign different alleles to different haplotypes without those necessarily corresponding to SNPs in the interval. This method also allows us to consider models with several alleles. For each eQTL, at the peak position, we fit all possible models for different numbers of alleles, fitting a maximum of 11337 models at each eQTL. We first estimated the haplotype means at the peak, sorted these means, and then fit all possible models that did not change the order of the haplotype means for 2, 3, 4, 5, 6, 7, 8, and 16 (the full model allowing different estimates for AB8 in pA RILs and AB8 in pB RILs) alleles (Figure S5). We only included haplotypes at the peak that occurred at least 5 times (at a probability of greater than 95%) in our set of crosses. Haplotypes at lower frequencies lead to inaccurate estimates of haplotype means with large standard errors. For each possible allele grouping, individual founder haplotype probabilities in each allele group were summed to obtain a probability each RIL harbored each allele group. For example, if haplotypes A3 and A5 are grouped as a single allele named allele 1, and the probabilities a given RIL cross harbors the A3 or A5 haplotype are 0.90 and 0.03 respectively, then the probability that RIL cross harbors allele 1 is 0.93 (i.e., the probability the RIL cross harbors either A3 OR A5 and thus allele 1). Alleles were only combined within pA and within pB given that the pA and pB sets of probabilities are independent. The model fit was as follows:where yr,i is the ith transcript, μ is the grand mean, na is the number of pA allele groupings, nb is the number of pB allele groupings, GA,c are the genotype probabilities for the cth paternal allele group, GB,d are the genotype probabilities for the dth maternal allele group, and βA,c, and βB,d are the corresponding effect estimates. The model with the lowest P-value was chosen as the best model and the number of alleles associated with this model was recorded. We also explored using Akaike's information criterion (AIC) to choose the best model, however simulations revealed a higher error rate using AIC (see below). Table S3 provides hard coded genotype assignments for all RIL crosses at all significant eQTL.
To test our method of estimating the number of alleles associated with QTL, we simulated QTL stemming from between 2 and 15 different alleles and subsequently estimated the number of alleles using the model comparison methodology described above. We intentionally set up this simulation to make distinguishing different alleles as easy as possible. We performed 1000 iterations for each of 2, 3, 4, 5, 6, 7, 8 and 15 alleles (the full model assuming the same effect for AB8 in the pA and pB panels). For each iteration, we randomly selected 600 pA RILs and 600 pB RILs from the DSPR panel and randomly paired them to create pA-pB crosses. We then simulated a QTL in this set of RIL crosses at a randomly selected position in the genome with the chosen number of alleles. We assigned the different alleles equal effects, because we found equal effects gave higher power to distinguish different alleles compared to pulling effects from a normal distribution (Figure S6). For example, for a four allele model each founder haplotype was randomly assigned an effect of 1, 2, 3, or, 4. We assumed an additive model to calculate a genetic effect for each cross. We generated a set of random normal deviates N(μ = 0, ) to correspond to environmental variance where z = the percent of the phenotypic variance explained by the QTL and is the genetic variance at the QTL. The percent of the total phenotypic variance explained by the QTL was randomly chosen from our observed distribution of phenotypic variance explained by cis-eQTLs. These effects tend to be quite large, however, we found large effects lead to higher power to distinguish different alleles (Figure S7). We then estimated the number of alleles at our simulated QTL as described above. We used two methods to determine the best model: 1) the model with the lowest P-value, and 2) the model with the lowest AIC. Our results showed the method using P-values had a greater accuracy (P-value method: 26% accuracy; AIC method: 19% accuracy). More importantly, the AIC method overestimates the true number of alleles more often, estimating more than two alleles in 83% of cases when the true number of alleles is two (Table S2). We prefer the method that is more conservative, meaning it has a greater tendency to underestimate rather than overestimate the number of alleles, and we therefore use the P-value method in all subsequent analysis (Figure S8). Complete sensitivity information for the different methods and the different simulation models can be seen in Figures S5, S6, S7 and in Table S2.
|
10.1371/journal.ppat.1006341 | The role of microsporidian polar tube protein 4 (PTP4) in host cell infection | Microsporidia have been identified as pathogens that have important effects on our health, food security and economy. A key to the success of these obligate intracellular pathogens is their unique invasion organelle, the polar tube, which delivers the nucleus containing sporoplasm into host cells during invasion. Due to the size of the polar tube, the rapidity of polar tube discharge and sporoplasm passage, and the absence of genetic techniques for the manipulation of microsporidia, study of this organelle has been difficult and there is relatively little known regarding polar tube formation and the function of the proteins making up this structure. Herein, we have characterized polar tube protein 4 (PTP4) from the microsporidium Encephalitozoon hellem and found that a monoclonal antibody to PTP4 labels the tip of the polar tube suggesting that PTP4 might be involved in a direct interaction with host cell proteins during invasion. Further analyses employing indirect immunofluorescence (IFA), enzyme-linked immunosorbent (ELISA) and fluorescence-activated cell sorting (FACS) assays confirmed that PTP4 binds to mammalian cells. The addition of either recombinant PTP4 protein or anti-PTP4 antibody reduced microsporidian infection of its host cells in vitro. Proteomic analysis of PTP4 bound to host cell membranes purified by immunoprecipitation identified transferrin receptor 1 (TfR1) as a potential host cell interacting partner for PTP4. Additional experiments revealed that knocking out TfR1, adding TfR1 recombinant protein into cell culture, or adding anti-TfR1 antibody into cell culture significantly reduced microsporidian infection rates. These results indicate that PTP4 is an important protein competent of the polar tube involved in the mechanism of host cell infection utilized by these pathogens.
| Microsporidia are obligate intracellular parasites that cause disease in immune suppressed individuals such as those with HIV/AIDS and recipients of organ transplants. The microsporidia are defined by a unique invasion organelle, the polar tube. The formation of this organelle and its role in the mechanism of infection remain unknown. Herein, we have identified a role for Encephalitozoon hellem polar tube protein 4 (PTP4) in infection demonstrating that PTP4 can bind to the host cell surface via the host transferrin receptor 1 (TfR1) protein. Interfering with the interaction of PTP4 and TfR1 causes a significant decrease in microsporidian infection of host cells. These data suggest that PTP4 functions as an important microsporidian protein during host cell infection by this pathogen.
| Since the first microsporidium, Nosema bombycis, was discovered in the European silkworm industry in the 19th century [1], more than 1400 species of microsporidia have been identified worldwide [2, 3]. They are ubiquitous obligate intracellular parasites responsible for a variety of diseases both in immune compromised and immune competent individuals. Microsporidia are also responsible for economic losses due to their adverse effects on farming and other industries [4, 5]. Phylogenetic studies suggest that microsporidia are related to fungi, being either a basal branch or sister group [6–13].
Microsporidia have multiple transmission routes including oral transmission of spores through contaminated food and water [14, 15], and vertical transmission [16, 17]. They can infect a wide variety of animals ranging from invertebrate to vertebrate hosts, including humans and insects of economic importance such as the silkworm and honey bee [8, 18]. Encephalitozoon hellem is found in humans and was initially isolated from corneal biopsies and conjunctival scrapings from patients with advanced HIV-1 infection with keratoconjunctivitis [19]. Similar to other members of the family Encephalitozoonidae, E. hellem has been demonstrated to cause disseminated infection presenting with diarrhea, nephritis, keratitis and/or sinusitis [20–22]. Microsporidia possess a unique, highly specialized invasion mechanism that involves the polar tube and spore wall [23]. Despite the description of these pathogens 150 years ago [1], the mechanism of host cell invasion, the structure and formation of both the polar tube infection apparatus and invasion synapse, and the role of microsporidian-specific proteins during the invasion process are not understood.
The polar tube is a highly specialized invasion organelle. Before germination, the polar tube coils around the sporoplasm in the spore [24, 25]. Upon appropriate environmental stimulation, the polar tube will rapidly discharge out of the spore and then interact with and pierce a cell membrane serving as a conduit for the nucleus and sporoplasm passage into the host cell (the entire process taking place in <2 seconds) [26–28]. Since the initial description of the polar tube by Thelohan 100 years ago [24, 25], proteomic and antibody studies have led to the identification of five different polar tube proteins (PTP1 through PTP5) in microsporidia [29–33]. Analysis of protein glycosylation has revealed that PTP1 contains many post translational O-linked mannosylation sites and that these residues can bind concanavalin A (conA) [34, 35]. Pre-treatment of a host cell with mannose has been demonstrated to reduce the infectivity of E. hellem, which is consistent with an interaction between mannosylated PTP1 and an, as yet, unknown host cell mannose-binding molecule [35] that results in the adherence of the polar tube to its host cell [35–40]. PTP2 is found at the same genomic locus as PTP1; and the PTP2 proteins from various microsporidia, despite a high degree of sequence divergence, share common characteristics such as a basic isoelectric point, high lysine content and conservation of cysteine residues [29, 37, 41, 42]. Immunoscreening of an E. cuniculi cDNA library led to the identification of a third polar tube protein, PTP3 [30]. PTP3, along with PTP1 and PTP2, was also found in cross-linked polar tube complexes and these three PTPs have been demonstrated to interact in yeast two hybrid assays [30, 39]. It has been suggested that PTP3 may act as a scaffolding protein for the assembly of other PTPs during the developmental formation of the polar tube [30]. PTP4 and PTP5 were identified in a survey of E. cuniculi proteins, but their function and interactions with other PTPs are unknown [10, 43]. Similar to the PTP1/PTP2 gene cluster, the genes for PTP4 and PTP5 usually are in a cluster in the genome [10, 43]. Despite the identification of these various PTPs, we still do not understand how the polar tube is formed or the mechanisms utilized for host cell infection and attachment.
Herein, we report on our characterization of Encephalitozoon hellem PTP4 and the identification of Transferrin receptor 1 (TfR1) as a candidate host cell receptor for this polar tube protein. The interaction between these molecules is demonstrated to be important for cell infection by this microsporidia. This is the first report of a host cell receptor that is involved in binding of a polar tube protein and infection of host cells. A deeper understanding of the mechanisms of infection and the formation of the microsporidian invasion synapse should provide new therapeutic targets for management of these ubiquitous intracellular pathogens.
Recombinant EhPTP4, (recEhPTP4, S1 Fig) without its signal peptide, was expressed in Escherichia coli as a fusion protein as detailed in the Materials and Methods Section and this recEhPTP4 was used to immunize both mice and rabbits. An EhPTP4 mouse monoclonal antibody (clone F4-6; MAb-EhPTP4) was produced by screening a hybridoma library. Both the rabbit polyclonal antibody (rab-Pc-EhPTP4Ab) and mouse monoclonal antibody (MAb-EhPTP4) to EhPTP4 react to the same antigenic band in E. hellem spore lysates (Fig 1A–1D, S2 Fig). This reactive band has a molecular mass of ~36kDa, which is slightly larger than the predicted EhPTP4 molecular mass of 32kDa. This difference from predicted mass is probably due to post translational glycosylation of PTP4 that is predicted to occur based on in silico analysis of this protein (S1A Fig). Indirect immunofluorescence assay (IFA) and transmission electron microscopy (TEM) have been used to successfully assess the localization of microsporidian proteins [39, 44–46]. IFA using rab-PcAb-EhPTP4 demonstrated that the entire polar tube was labeled by this polyclonal serum (Fig 1E and S3 Fig), proving that EhPTP4 is a polar tube protein; however, MAb-EhPTP4 labeled only the tip of polar tube (Fig 1F and S3 Fig). The different localization patterns seen using the rab-Pc-EhPTP4Ab and MAb-EhPTP4 (S3 Fig) suggests that there is an EhPTP4 epitope which is only exposed at the tip of polar tube and that this epitope is specifically recognized by the MAb-EhPTP4. Immunoelectron microscopy (ImmunoEM) was utilized to examine the location of EhPTP4 in E. hellem spores (Fig 1G and 1H). Consistent with the labeling result using IFA, the gold particles were observed on the polar tube when using rab-PcAb-EhPTP4 (Fig 1G-I and 1G-II) and there was no staining by control rabbit serum (Fig 1G-III). Using MAb-EhPTP4 a more limited region of the polar tube was stained with gold (Fig 1H-I), with the majority of gold particles being found at an area where the polar tube appeared to be connected to the posterior vacuole (Fig 1H-II). This is consistent with previous models of eversion of the polar tube [10, 47] where the tip of the tube should be located near the membrane of the posterior vacuole prior to germination. There were no gold particles seen in the control panel using an unrelated isotype matched MAb (Fig 1H-III). These combined IFA and TEM observations confirm that EhPTP4 is a polar tube protein and the localization of MAb-EhPTP4 to the tip of the polar tube suggests that EhPTP4 may be involved in an interaction with a host cell protein during infection.
Correlated fluorescence and scanning electron microscopy (CLEM) combines wide-field laser/light microscopy with subsequent electron microscopy and can minimize the respective disadvantages of each technique when used individually [48, 49]. CLEM was used to investigate the polar tube during infection. Infected cells were fixed and stained with MAb-EhPTP4 (mouse monoclonal) and rab-PcAb-EhPTP1 (rabbit polyclonal) to locate EhPTP4 at the tip of the tube and EhPTP1 on the entire polar tube respectively. After being incubated with fluorescent secondary mouse and rabbit antibodies, a fluorescence microscopy image was taken, and then a scanning electron microscopy (SEM) image of the same site on the slide was taken. Finally, the fluorescence images that demonstrated localization of EhPTP4 and polar tube were correlated with the SEM images that demonstrated the invasion organelle of the microsporidia at a high resolution. As seen in Fig 2A, the germinated polar tube was labeled with rab-PcAb-EhPTP1 and MAb-EhPTP4 labeled the tip of polar tube, consistent with the results of the IFA and TEM studies. Fig 2B shows a discharged polar tube and the interaction of the polar tube tip with host cell plasma membrane (PM). As is shown in enlarged section (Fig 2C), the tip of polar tube is buried under some fiber-like structures of the host cell (these structures surround the invasion synapse). The mechanism by which the polar tube interacts with the host cell membrane resulting in penetration is currently unknown; however, there is some evidence that host cell actin may be involved in microsporidian penetration of the host cell within the invasion synapse [50]. The location of a specific epitope of EhPTP4 at the tip of the polar tube suggests that EhPTP4 may directly interact with host cell proteins/receptors in the final penetration of the host cell membrane during infection. After the binding of polar tube to host cell membrane, the next step in infection should be the initial penetration of polar tube into the host cell plasma membrane. Fig 2D demonstrates an apical tip of polar tube that has entered the host cell plasma membrane by pushing the host cell plasma membrane into the cell. Our TEM observations support the concept that the polar tube is surrounded by host cell membrane at the invasion site (Fig 2E). These observations are similar to previous reports regarding the mechanism of polar tube penetration in that the tip of polar tube is not piercing or breaking the host plasma membrane, but instead pushing the host cell plasma membrane into the host cell creating a microenvironment into which the microsporidian sporoplasm is extruded from the end of the polar tube [51–54].
These combined IFA, TEM and CLEM studies suggested that EhPTP4 may be involved in an interaction with the host cell membrane during infection. Binding of EhPTP4 to host cells was evaluated using ELISA, IFA and FACS (Fig 3). Increasing concentrations of recEhPTP4 were incubated with fixed RK13 cells on a 96 well plate and binding was detected as described in Materials and Methods. The results demonstrated that incubation of host cells with increasing amounts of EhPTP4 resulted in a proportional increase in binding (Fig 3A) suggesting that EhPTP4 most likely interacts with a protein (or other binding partner) on the host cell membrane; however, as the host cells used in these ELISA experiments were fixed, it is possible that EhPTP4 could have been binding to either an intracellular or extracellular binding site. To address this issue FACS analysis was done using live cells that were not fixed or permeabilized. FACS analysis (Fig 3B) demonstrated that EhPTP4 also bound to the cells and, as live cells were used, this binding was clearly on the host cell surface. RK13 cells were detached by citric saline solution and incubated with EhPTP4-Fc fusion proteins or human Fc proteins. After labeling with PE-conjugated anti-Fc antibody, binding was detected by flow cytometry. As demonstrated in Fig 3B, the binding curve of EhPTP4-Fc was significantly shifted compared to the control human Fc binding curve, confirming that EhPTP4 binds to host cells. When trypsin was used to remove surface proteins from HFF cells, there was no binding of EhPTP4 to host cells (S4A and S4B Fig). When recEhPTP4 was incubated with these RK13 cells followed by IFA the binding signal was clearly visualized on the host cell surface (Fig 3C). In summary, the binding data are consistent with an ability of EhPTP4 to interact with protein(s) on the host cell surface.
Immunoprecipitation (IP) was utilized to identify the potential interacting targets of EhPTP4 on its host cell. RK13 (rabbit) cells were infected with E. hellem spores and harvested after two weeks of infection. The lysate of infected cells was incubated with protein A&G sepharose beads that had been conjugated with mouse polyclonal antibody to EhPTP4 (mo-PcAb-EhPTP4). The proteins that bound to the beads, i.e. the immunoprecipitation sample (IP sample), were then analyzed by SDS-PAGE. Silver staining demonstrated that there were several bands which were unique to the IP sample (lane 1) compared to the control samples (lane 2 and 3) (Fig 4A), and an Oryctolagus cuniculus (rabbit) protein, Transferrin receptor 1 (TfR1) was identified by LC-ESI-MS/MS analysis (Fig 4B and 4C). The same band was also identified using a TfR1 monoclonal antibody (MAb-TfR1; see details in Materials and methods section) (Fig 4D). We repeated the IP using MAb-EhPTP4 (monoclonal anti-PTP4) and the same TfR1 band was identified in this IP using MAb-TfR1 as shown in Fig 4E. Furthermore, pull down assays using antibody to EhPTP4 (MAbEhPTP4) demonstrated that recEhPTP4 can directly interact with recTfR-1 in vitro (Fig 4F and S5 Fig).
IFA co-localization techniques using EhPTP4 Alexa Fluor 488 (green) and TfR1 Alexa Fluor 594 (red) staining on host cells demonstrated signal at the site of infection for both EhPTP4 and TfR1 (Fig 5A). During microsporidian infection we could also find that the tip of the polar tube co-localized with TfR1 on host cell membrane at the site of the invasion synapse (Fig 5B). Therefore, several lines of evidence suggest that TfR1 is a host cell receptor that can interact with EhPTP4. We next examined if this interaction had an effect on the infection process.
To further examine the role of EhPTP4 during microsporidia infection (see S6 Fig for assay details), we performed antibody blocking and protein competition experiments. EhPTP4 rabbit polyclonal antibody (rab-PcAb-PTP4) was incubated with purified spores for 1hr before infecting host cells. We found a significant reduction in infectivity as a result of blocking EhPTP4 interactions with rab-PcAb-PTP4 in the antibody blocking assay (Fig 6A). To perform a protein competition assay, recombinant EhPTP4 protein was directly added into cell culture prior to infection with spores and the number of parasitophorous vacuoles was counted after 9 days post-infection. Similar to what was seen with rab-PcAb-PTP4, adding EhPTP4 recombinant protein to cell culture during infection significantly decreased the infectivity of microsporidia (Fig 6B). Taken together, these data indicate that PTP4 plays a role in infection.
To further evaluate if TfR1 plays a role in infection, we evaluated the effect of alternations of TfR1 expression in host cells on the ability of microsporidia to infect these cells. The CHO cell lines TRVb, which is a TfR1 knockout, and TRVb-1, which expresses human TfR1, were used to examine microsporidia infection. Immunoblot analysis using MAb-TfR1 confirmed that there is no TfR1 protein expressed in the TRVb cell line (lane 1, Fig 7A), and that TfR1 is expressed in the TRVb-1 cell line (lane 2, Fig 7A). Microsporidian infection was dramatically decreased in TRVb cells compared to the infection rate seen in TRVb-1 cells (Fig 7B). In order to examine the role of TfR1 in invasion we evaluated the number of intracellular E. hellem organisms 6 hours after infection (Fig 7C). Microsporidia invasion was significantly decreased in TRVb cells compared to TRVb-1 cells (Fig 7D). In order to further examine the role of TfR1 in infection, an antibody blocking assay and TfR1 competition assay were performed as described in the Materials and Methods section of this paper. Analogous to the infection results in TRVb cells, we also observed a dramatically decreased infection rate in cells whose TfR1 was blocked by polyclonal antibody to TfR1 (Fig 7E). Furthermore the addition of recombinant TfR1 protein (recTfR1) into cell culture during microsporidia infection also decreased host cell infection rates (Fig 7F). This blocking of infection by recTfR1 may be a consequence of the binding of recTfR1 to PTP4 preventing the interaction of PTP4 with native host cell TfR1, thereby, interfering with downstream events due to this interaction. In summary, these data suggest that host cell TfR1 is involved in the microsporidian infection process.
Microsporidia possess a unique invasion apparatus, the polar tube. Under the appropriate environmental stimulation, the polar tube can discharge rapidly out of microsporidia spore, form a hollow tube and serve as a conduit for the passage of sporoplasm and nuclear material into a new host cell [51]. Before germination, within the spore, the polar tube is filled with material that has been hypothesized to consist of unpolymerized polar tube proteins and during the process of germination and extrusion of the polar tube this filled structure transitions to a hollow tube [23, 55]. The length of discharged polar tubes is approximately 2 to 3 times that of the coiled tubes inside the spore and it has been hypothesized that either unpolymerized polar tube proteins are incorporated at the growing tip of polar tube during discharge or that the tube unfolds during eversion [23, 27, 28, 55]. In the process of polar tube eversion, unique immunologic epitopes may be exposed on polar tube proteins [10].
In the current study we have provided a characterization of PTP4 and identified a role for this protein in infection. Both monoclonal and polyclonal antibodies to EhPTP4 recognize the same band in spore lysates by immunoblotting, suggesting that these antisera recognize the same protein and that any difference in staining patterns is not due to recognition of additional proteins by the polyclonal antiserum. A monoclonal antibody to PTP4 (MAb-EhPTP4; clone F4-6) was found to specifically label the anterior end of polar tube while the polyclonal rabbit antibody to EhPTP4 (PcAb-EhPTP4) labeled the entire polar tube. This is consistent with the exposure of a specific epitope on EhPTP4 at the tip of the polar tube that is recognized by this monoclonal antibody, and which defines the location of the tip of the polar tube. Since the tip of polar tube is the place where polar tube sporoplasm exits and the location of an interaction with the host cell membrane in the microenvironment of the invasion synapse, these localization data suggest that PTP4 may be involved in polar tube host-cell interactions in the invasion synapse and in the mechanism of penetration of the host cell by microsporidia. Analyses employing IFA, ELISA and FACS confirmed that PTP4 binds mammalian cells. When EhPTP4 binding was inhibited by antibody treatment of host cells microsporidia infection was reduced (Fig 6A), suggesting that EhPTP4 interaction with the host cell is involved in the infection process.
Despite the identification of several polar tube proteins in the last few years [29–31], it is not known how the PTPs or the polar tube interact with host cells or how penetration occurs. In this manuscript, we provide evidence that TfR1, a host cell membrane protein, interacts with PTP4 and that this interaction is important in the process of infection. Proteomic analysis of PTP4 bound to host cell membranes purified by co-IP identified Transferrin receptor 1 (TfR1) as a potential host cell interacting partner for PTP4. Data obtained by immunoprecipitation, pull down and immunocolocalization provide additional confirmation of an interaction between EhPTP4 and TfR1. Knockout of TfR1 in CHO cells significantly decreased the infectivity of microsporidia. Antibody blocking of TfR1 or the addition of recombinant TfR1 protein to cell culture also reduces microsporidia infectivity.
The Transferrin receptor 1 (TfR1) is involved in iron uptake in many cells. Cells can import iron by internalizing the transferrin-iron complex through clathrin-mediated endocytosis [56]. TfR1 binds iron-loaded transferrin at a neutral pH, and releases the iron in the early acidic endosome and then recycles back to the cell surface [57]. TfR1 is the cell receptor for a variety of viruses including parvoviruses, hepatitis C virus, mammary tumor virus and arenaviruses, and this receptor can be utilized by those virus to bind, invade and infect host cells [58–62].
Interestingly recEhPTP4 was found to bind to both TRVb CHO and TRVb-1 CHO cells (S4C Fig) suggesting that EhPTP4 may bind to host cells by several mechanisms. According to our analysis of the protein sequence of EhPTP4 (S1B Fig), there are several carbohydrate binding domains predicted to occur consistent with previously reported structures of chitin-binding proteins [63]. Both Cys87 and Gly90 of EhPTP4 are conserved amino acid residues that are involved in binding chitin in previously described proteins (S1C Fig), indicating that EhPTP4 might be able to bind chitin. Therefore, a possible explanation for recEhPTP4 binding to both TRVb CHO and TRVb-1 CHO cells is that the chitin domain may allow EhPTP4 to bind to glycoproteins on host cells (or facilitate other receptor interactions). Binding to these alternative sites would explain the ability of recEhPTP4 to bind to TRVb cells. Despite this, it is clear that the presence of functional TfR1 is important in the infection process and that EhPTP4 binding to TfR1 is involved in E. hellem host cell infection. The data also suggest a role for TfR1 in the invasion synapse in addition to any role TfR1 may have in the binding/adherence of the polar tube to the host cell surface.
It has been suggested that the polar tube is extracytoplasmic in the unfired spore [64]. After the discharge of the polar tube is completed, the sporoplasm will flow through the polar tube and appear as a droplet at the distal end [26, 65]. If a spore discharges next to a host cell, it will pierce the cell membrane and transport sporoplasm into the cell [26, 27, 65]. If there are no cells in the area, the droplet will remain attaching to the end of polar tube for a period of time (Fig 8). CLEM clearly demonstrates the PTP4 is located at the place where the polar tube connects with droplet spore contents (Fig 8). As is shown in Fig 1H-II, the end of polar tube which was labeled with EhPTP4 MAb was connected with posterior vacuole area, suggesting that the localization of the end of polar tube in spore is probably next to the posterior vacuole in the microsporidian spore.
On the basis of previous reports [34, 50, 51] and the current data, we have proposed an infection model for microsporidia (Fig 9). According to this model, upon appropriate environment stimulation the polar tube rapidly discharges out of spore and then polar tube protein 1 (PTP1), which was shown to be a mannosylglycoprotein, interacts with mannose binding proteins or lectins on host cell surface [34, 35] allowing the polar tube to adhere to the host cell surface. PTP4 interactions, with TfR1 or with glycoproteins, may also facilitate adherence of the polar tube to the host cell surface. As the polar tube everts it pushes into the host cell creating a protective microenvironment, the invasion synapse, into which the sporoplasm is then extruded. Interactions with PTP1 (and perhaps PTP4) with the host cell membrane during the formation of this invasion synapse allow the creation of an invasion synapse microenvironment that is to be isolated from the external environment. PTP4 epitopes that are exposed at the tip of polar tube can interact with TfR1 in the protected environment of the invasion synapse. This interaction probably triggers clathrin-mediated endocytosis pathways helping to facilitate the progress of infection [51, 66]. When the extruded sporoplasm is injected into the invasion synapse the vacuole membrane will pinch off from host cell membrane and form a parasitophorous vacuole (PV) and any associated TfR1 will be lost from PV membrane immediately after infection being recycled in the cell [67]. Observations by other laboratories suggest that host cell actin may have also has a role in the final invasion event [50]. The precise mechanisms of the EhPTP4 interaction with TfR1, how this complex functions and dissociates in the PV after infection and whether PTP4 interacts with other proteins on the host cells remains an unknown and important area for further investigation.
RK13 (rabbit kidney) cells (ATCC, CCL-37) and CHO cells (TRVb and TRVb-1) (gift of Dr. Colin R. Parrish and Dr. Timothy E. McGraw, Cornell University) were cultured in 10% fetal bovine serum (FBS) (ThermoFisher) Minimum Essential Medium Eagle (MEM) with penicillin-streptomycin at 5% CO2. Human foreskin fibroblasts (HFF) (ATCC, CRL-2522) and 293FT (ATCC, CRL-3216; gift of Dr. Matthew Levy, Albert Einstein College of Medicine) cells were maintained in Dulbecco’s Modified Eagle Medium (DMEM; ThermoFisher Scientific) with penicillin-streptomycin (ThermoFisher Scientific) supplemented with 10% FBS (ThermoFisher Scientific) at 5% CO2. Anti-human TfR1 mouse monoclonal antibody (mo-Mab-HuTfR1; clone H68.4) used for immunoblot and IFA was purchased from Invitrogen. Anti-human TfR1 rabbit polyclonal antibody (rab-PcAb-TfR1) used for blocking and IFA was purchased from Santa Cruz Biotechnology. PE-conjugated anti-human Fc antibody (clone HP6017) used for flow cytometry was purchased from Sony Biotechnology. The lipofectamine 3000 transfection kit used for 293FT cell transfection was purchased from Invitrogen. HRP-conjugated anti-mouse secondary antibodies, Alexa Fluor 488 conjugated goat anti-mouse secondary antibody and Alexa Fluor 594 conjugated goat anti-rabbit secondary antibody were purchased from ThermoFisher Scientific. Rabbit anti-EhPTP1 polyclonal antibody (rab-PcAb-EhPTP1) used for IFA and CLEM was produced previously in our laboratory [33]. Buffered solutions and chemicals were purchased from Sigma Aldrich. All reagents used are commercially available and chemicals were of the highest analytical grade available from the supplier.
RK13 (rabbit kidney) cells were maintained in 10% fetal bovine serum MEM with penicillin-streptomycin at 5% CO2. Confluent monolayers were infected with E. hellem spores. The spores were collected from culture media, purified by passing them through 5 μm size filter (Millipore, Billerica, MA) to remove host cells, concentrated by centrifugation, and stored in sterile distilled water at 4°C. Spores used in these experiments were counted with a hemocytometer (three times/sample and averaged).
The coding region of EhPTP4 was PCR amplified from E. hellem genomic DNA using Q5 High-Fidelity DNA Polymerase (New England Biolabs) with specific primers (S1 Table) and ligated into the vector pMCSG7 containing a TEV cleavable N-terminal hexahistidine (His6) tag. The resulting plasmid was transformed into BL21 (DE3) T1R (Sigma-Aldrich) containing the RIL plasmid (RIL) from Stratagene (CA, USA) containing copies of genes encoding tRNAs for rare codon. Transformed bacteria were grown in Luria Broth (LB; Sigma-Aldrich) containing ampicillin and chloramphenicol (100 μg/μl and 34 μg/μl, respectively), for 5hrs at 37°C. At that time the temperature was lowered to 22°C for overnight growth. The bacterial cells were pelleted and resuspended in buffer A (20 mM Hepes pH 7.6, 500 mM NaCl, 20 mM imidazole, 10% glycerol, 0.01% Tween-20, 0.1% NaN3, containing 1 mM PMSF and 4 u/ml of DNAase I). The cells were lysed using an EmulsiFlex C3 (Avestin) and separation of the lysate from the intact cells was achieved by centrifugation (16500 g, 1hr). The pellet was washed two times in buffer B (20 mM Hepes pH 7.6, 150 mM NaCl, 0.1% Triton X100, 10 mM DTT) and another time in buffer B without Triton X100. The pellet was then solubilized in buffer C (100 mM NaH2PO4, 10 mM Tris, 6 M guanidine at pH 8.0). The supernatant was saved after centrifugation (16500Xg, 1hr) and loaded onto a His-Pure Ni-NTA column (ThermoFisher Scientific) that had been previously equilibrated with buffer C. The column was successively washed with 10 column volumes of buffer D (100 mM NaH2PO4, 10 mM Tris, 8 M urea, 2% Triton X100 at pH 8.0), followed by 5 column volumes of buffer D without Triton X100, and then 5 column volumes of buffer E (100 mM NaH2PO4, 10 mM Tris, 8 M urea, 2% Triton X100 at pH 6.3). The protein was then eluted with 5 column volumes of buffer F (100 mM NaH2PO4, 10 mM Tris, 8 M urea, 2% Triton X100 at pH 4.5). Protein purity was determined by SDS-PAGE and protein concentration determined using a nanodrop spectrophotometer (ThermoFisher Scientific).
The coding region of EhPTP4 was PCR amplified using Q5 High-Fidelity DNA Polymerase (New England Biolabs) employing EhPTP4 specific primers (S1 Table), ligated into vector pFUSE-hIgG1-Fc2 (a gift from Dr. Matthew Levy), and then transfected into 293FT cells using Lipofectamine 3000 Reagent (Thermo Fisher Scientific). A mock transfection was also prepared with the empty vector pFUSE-hIgG1-Fc2. 293FT cells grown to 70% ~90% confluency in six well plate were prepared in 3 ml fresh medium (Dulbecco’s modified Eagle medium with 10% FBS and 1%). The transfected reagent was prepared by diluting 2.5 μg plasmid with 250 μl Opti-MEM medium and 10 μl P3000 reagent, followed by adding additional diluted Lipofectamine 3000 regent. The solution was mixed well and incubated for 5 min at room temperature before adding to the cells. Supernatant containing EhPTP4-Fc fusion protein or Fc protein was harvested 48hrs later after transfection.
For polyclonal antibody production in mice 500 μg of purified recEhPTP4 protein was injected into five BALB/c mice after emulsification of the protein with Freund's complete adjuvant (Sigma-Aldrich) at a dose of 100 μg protein for each mouse. After boosting, using 100 μg protein per mouse in Freund’s incomplete adjuvant (Sigma-Aldrich) monthly for three months, the mice were bled and serum (mo-PcAb-EhPTP4) stored at -20°C until use. Control mouse sera was collected from BALB/c mice prior to immunization (pre-immune serum) as well as BALB/c mice of the same age which had not been immunized with recEhPTP4 (i.e. control polyclonal mouse IgG).
Polyclonal rabbit antibody to recEhPTP4 was produced by Harlan Laboratories (Envigo, USA) following their standard 112 day polyclonal antibody protocol. Briefly, a New Zealand white rabbit was initially immunized with 200 μg of purified recEhPTP4 emulsified with Freund’s complete adjuvant and this was then followed by three monthly injections of 100 μg of purified recEhPTP4 emulsified with Freund’s incomplete adjuvant. The ELISA titer of the rabbit serum was assessed following each boost injection and was over 1:50,000 one month after the final injection. Rabbit serum (PcAb-EhPTP4) was collected one month following the final immunization and stored at -20°C. Prior to immunization serum was collected and screened to confirm that this rabbit did have any endogenous antibody that reacted with E. hellem or EhPTP4. Preimmunization rabbit serum stored -20°C was used as a negative control for experiments using polyclonal rabbit antisera.
For monoclonal antibody production, the spleen cells of mice which were immunized with recEhPTP4 (100 μg protein per mouse) were fused with a myeloma cell line to create hybridoma libraries according to our previously protocol [68]. An ELISA was used to screen the resultant hybridoma libraries. For the ELISA, 0.05 μg per well recEhPTP4 was coated onto 96 well plates, and then the plates were blocked with 3% BSA in TBST for 1hr. The hybridoma cell culture media were then incubated with the EhPTP4 coated plates, and hybridomas that were ELISA positive were then picked and grown in 6 well plates. The selected hybridomas were then screened again by immunoblot using recEhPTP4 and total spore lyses as antigens. The monoclonal lines that corresponded to a single band on immunoblot were then screened by IFA using germinated spores. Hybridoma cells that stained polar tubes in IFA were chosen and then subcloned on agarose and rescreened by ELISA and IFA. The monoclonal hybridoma cell line of interest, clone F4-6 (isotype IgG1), was then cultured in CELLine bioreactor (Integra Biosceinces) for large-scale production of MAb-EhPTP4.
RK13 cells were cultured in six well plates to confluence and infected with 1×106 spores for 3 days. Cells were fixed with 4% paraformaldehyde 0.05% glutaraldehyde in 0.1M sodium cacodylate buffer, dehydrated through a graded series of ethanol, with a progressive lowering of the temperature to -50°C in a Leica EMAFS, embedded in Lowicryl HM-20 monostep resin (Electron Microscopy Sciences), and then polymerized using UV light as we have previously published [26]. Ultrathin sections were cut on a Reichert Ultracut E, immunolabeled with antibodies of interest and then stained with uranyl acetate (Electron Microscopy Sciences) followed by lead citrate. Stained sections were viewed on a JOEL 1200EX transmission electron microscope at 80kv [33].
1×105 RK13 cells were infected with 1×106 spores on glass coverslips imprinted with three location markers. Cells were fixed with 4% paraformaldehyde in PBS for 30 min after three days post-infection. After being washed three times by TBST (TBS plus 0.05%Tween 20), coverslips were blocked by 3% BSA in TBST for 1hr at room temperature. EhPTP4 monoclonal antibody (MAb-EhPTP4; clone F4-6) at a 1:10 dilution and EhPTP1 rabbit polyclonal antibody (rab-PcAb-EhPTP1) at a 1:500 dilution were incubated with cells for 1hr at room temperature. After being washed three times by TBST, Alexa Fluor 488-labeled anti-mouse IgG antibody and Alexa Fluor 594-labeled anti-rabbit IgG antibody were added at 1:500 dilutions. Mouse and rabbit preimmunization sera were used as negative controls. Samples were washed three times by TBST and imaged using a Zeiss AxioObserver microscope equipped with Axiovision software with "shuttle & find" to mark cell locations. After fluorescence imaging, samples were fixed with 2.5% glutaraldehyde, dehydrated in ethanol, critical point dried (Tousimis Samdri 790), and coated with chromium (EMS 150T-ES). The same cells were automatically located in the Zeiss Supra 40 Field Emission Scanning Electron Microscope and imaged with a secondary electron detector. Fluorescence and SEM images were correlated with Zeiss AxioVision.
RK13 cells were grown to confluence in 96-well plate and fixed with 4% paraformaldehyde in PBS for 30 min at room temperature. Non-specific binding sites were blocked by incubating with 3% BSA in TBST buffer. After washing three times with TBST, serial diluted EhPTP4 soluble protein (starting at 100 μg/ml) was added to each well and incubated 1hr at room temperature. BSA at the same concentration as EhPTP4 was used as negative control for all ELISA assays. Unbounded proteins were washed off by washing three times with TBST, and bound proteins were fixed by incubating with methanol for 10 min. After washing with TBST for three times, non-specific binding sites were blocked by incubating plates with 3% BSA in TBST for 1hr at room temperature, EhPTP4 mouse polyclonal sera (mo-PcAb-EhPTP4) was then added at 1:500 dilution, plates were incubated for 1hr at room temperature and then washed three times with TBST. After this wash anti-mouse IgG alkaline phosphatase antibody (Sigma-Aldrich) was added at a 1:7000 dilution, plates were incubated for 1hr at room temperature, washed three times with TBST, and then p-nitrophenyl phosphate substrate (Sigma-Aldrich) was added and absorbance at 405 nm was read using a MRXe microplate reader (DYNEX Technologies). BSA at the same serial dilution was used as negative control.
EhPTP4-Fc fusion proteins used for flow cytometry were produced from 293t cells transfected with the appropriated plasmids. HFF cells were cultured to confluence and detached by treating with 1X Citric saline (135 mM KCl, 15mM sodium citrate, diluted with sterile distilled water) solution for 5 min at 37°C. The detached cells were washed with PBS containing 1% bovine serum albumin (FACS buffer). 1×105 cells were allowed to incubate with the supernatants harvested from 293t cell culture for 1hr. After the cells were washed twice with FACS buffer, PE anti-human IgG Fc (Sony Biotechnology) was added to each sample and allowed to incubate for 30 min. Followed incubation with DAPI for 10 min to identify dead cells, which were removed from the FACS analysis, flow cytometry was performed on a Becton Dickinson LSRII Analyzer Flow Cytometers (BD Biosciences).
To detect the binding of EhPTP4 by IFA, RK13 cells were grown to confluence in four-well chamber slides and fixed by 4% paraformaldehyde in PBS for 30 min at room temperature. After washing three times with TBST, slides were blocked by 3% BSA in TBST for 1hr at room temperature, and 500 μl 20 μg/ml EhPTP4 recombinant protein or BSA was added and incubated for 1hr at room temperature. Unbound proteins were then removed by washing three times with TBST, and bound proteins were fixed with methanol for 10 min. Non-specific binding sites were blocked by incubation with 3% BSA in TBST and EhPTP4 polyclonal sera (mo-PcAb-EhPTP4) was added at a 1:500 dilution in 3% BSA in TBST. After incubating for 1hr at room temperature, slides were washed three times with TBST, Alexa Fluor 488-labeled anti-mouse IgG antibody was added at a 1:500 dilution in 3% BSA in TBST, and then the slide was incubated for 1hr at room temperature. After washing three times with TBST, slides were mounted with ProLong Gold antifade regent and photomicrographs were taken either with a SP5 confocal microscope (Leica) or Microphoto-FXA epifluorescence microscope (Nikon).
To detect the localization of EhPTP4 on microsporidia spores, infected RK13 cells were grown on four-well chamber slides (ThermoFisher Scientific) and fixed with 4% paraformaldehyde in PBS for 30 min. After washing three times with TBST, slides were blocked by 3% BSA in TBST for 1hr at room temperature. Anti-EhPTP4 monoclonal antibody (MAb-EhPTP4) and EhPTP1 rabbit polyclonal antibody (rab-PcAb-EhPTP1 [33]) were added at 1:10 and 1:500 dilutions respectively in 3% BSA in TBST and incubated 1hr at room temperature. After washing three times with TBST, Alexa Fluor 488-labeled anti-mouse IgG antibody and Alexa Fluor 595-labeled anti-rabbit IgG antibody were added at a 1:500 dilution. After washing three times with TBST, slides were mounted with ProLong Gold antifade regent and viewed with a SP5 confocal microscope (Leica). Mouse preimmunization serum was used as a negative control.
RK13 cells were cultured in a T75 flask (Corning) to confluence and then infected with 1×107 spores. The cells were harvested two weeks post-infection and cell lysates were obtained by a modified procedure as reported previously [34]. Infected RK13 cells were spun down and disrupted in 1% SDS lysis buffer containing protease inhibitor (Protease inhibitor cocktail; Thermo Fisher Scientific) with 0.4 g 0.5 μm acid-washed glass beads (Sigma-Aldrich) for 1 min on a Mini-Beadbeater (BioSpec Products). The disrupted host cell suspension was then clarified by centrifugation at 12000 rpm for 10 min.
For immunoprecipitation, 50 μl of Protein A&G agarose beads in PBS was incubated with 20μl of mo-PcAb-EhPTP4 or mouse pre-immune serum at 4°C for 1hr. The protein A&G agarose beads conjugates were washed three times with PBS and resuspended in a final volume of 500 μl cell lyses collected above. The samples were incubated at 4°C for 2hrs with gentle shaking. The samples were then centrifuged, and the pellets were washed three times with PBS. Fifty μl of 2X protein sample buffer (0.5M Tris-HCl pH 6.8, 4.4% SDS, 20% glycerol, 2% 2-mercaptoethanol and 0.01% bromophenol blue) was added to each sample and the samples were then boiled for 5 min. Following centrifugation the supernatant was loaded onto an SDS-PAGE gel and electrophoresed. The gel was then stained with Coomassie Brilliant Blue or silver (using a BioRad Silver staining kit) and analyzed using methods that we have previously reported [68]. Briefly, excised gel bands were reduced, alkylated and digested with trypsin. LC-ESI-MS/MS (liquid chromatography electrospray ionization mass spectrometry) analysis of the peptide digests was then done by C18-Reversed Phase (RP) chromatography using an Ultimate 3000 RSLCnano System (ThermoFisher Scientific) equipped with an Acclaim PepMap RSLC C18 column (2 μm, 100 Å, 75 μm x 15 cm, ThermoScientific, USA). The UPLC was connected to a TriVersa NanoMate nanoelectrospray source (Advion) and a linear ion trap LTQ-XL (ThermoFisher Scientific) mass spectrometer with ESI source operated in the positive ionization mode. MGF files generated from the raw LC-ESI-MS/MS data were searched by Mascot (version 2.5.1, Matrix Science, USA) against NCBInr90_20141124 database (25,782,812 protein sequences) with the following search parameters: trypsin; two missed cleavages; peptide charges of +2 and +3; peptide tolerance of 1.5 Da; MS/MS tolerance of 0.5 Da; carbamidomethylation (Cys) for fixed modification; deamidation (Asn and Gln) and oxidation (Met) for variable modifications. An appropriate decoy database search was utilized to measure the false discovery rate.
The immunoprecipitation was also repeated using with MAb-EhPTP4, briefly 50 μl of Protein G sepharose beads in PBS was incubated with 20 μl of MAb-EhPTP4 or mouse negative IgG at 4°C for 1hr. The protein G sepharose beads conjugates were washed three times with PBS and resuspended in a final volume of 500 μl cell lyses collected above. Samples were incubated at 4°C for 2hrs with gentle shaking, then centrifuged, and the pellets were washed three times with PBS. Fifty μl of 2X protein sample buffer was ere added to each sample and the samples were then boiled for 5 min. The IP samples were then precipitated by using either mo-PcAb-EhPTP4 or MAb-EhPTP4, were run on SDS-PAGE, and then transferred to a PVDF membrane. Immunoblot analysis was then performed using MAb-TfR1 to evaluate the presence of TfR1 in the IP material.
For the pull down assay, 20 μg recTfR-1 and 20 μg recEhPTP4 were incubated at 4°C for 3 hrs in PBS with gentle shaking. Fifty μl of Protein G sepharose beads in PBS were incubated with 20 μl of MAb-EhPTP4 or 20μg of BSA (as a negative control) at 4°C for 1hr with gentle shaking. The protein G sepharose beads conjugates were washed three times with PBS and then incubated with the protein mixture at 4°C for 2hrs with gentle shaking. The samples were then centrifuged, the pellets were washed three times with PBS and 50 μl of 2X protein sample buffer (0.5M Tris-HCl pH 6.8, 4.4% SDS, 20% glycerol, 2% 2-mercaptoethanol and 0.01% bromophenol blue) was added to each sample. Samples were then boiled for 5 min, subjected to SDS-PAGE electrophoresis, transferred to a PVDF membrane and examined by immunoblot using PcAb-TfR1.
For co-localization of recombinant EhPTP4 with Transferrin receptor 1, human foreskin fibroblasts (HFF) cells were seeded in 4 well chamber slides and fixed with 4% paraformaldehyde. After being blocked with blocking buffer (3% BSA in TBST), HFF cells were incubated with 5 μg/ml of recombinant EhPTP4 or BSA for 1hr at room temperature. After being washed three times with TBST, anti-human Transferrin receptor 1 rabbit polyclonal antibody and anti-EhPTP4 mouse monoclonal antibody (MAb-EhPTP4) were incubated with HFF cells for 1hr at the dilution of 1:50 and 1:100 respectively. After washing three times with TBST, Alexa Fluor 488-labeled anti-mouse IgG antibody and Alexa Fluor 594-labeled anti-rabbit IgG antibody were added at a 1:500 dilution, and incubated for 1hr. After washing three times with TBST, slides were mounted with ProLong Gold antifade regent and viewed with a SP5 confocal microscope (Leica).
For co-localization of the tip of polar tube with TfR1, human foreskin fibroblasts (HFF) cells were seeded in 4 well chamber slides and infected with E. hellem. Three days post-infection the cells were fixed and blocked with blocking buffer. After being washed three times with TBST, anti-EhPTP4 mouse monoclonal antibody (MAb-EhPTP4) and anti-EhPTP1 rabbit polyclonal antibody (rab-Pc-EhPTP1) were incubated with these infected HFF cells for 1hr at dilutions of 1:100 and 1:500 respectively. After washing three times with TBST, Alexa Fluor 594-labeled anti-mouse IgG antibody and Alexa Fluor 405-labeled anti-rabbit IgG antibody were added at a 1:500 dilution, and slides were incubated for 1hr. After washing three times with TBST the slides were fixed with methanol for 10 min, rinsed once in TBST to remove methanol and then incubated with FITC conjugated anti-human TfR1 mouse monoclonal antibody (Fisher scientific) for 30 min at 4°C, protected the slides from light. After washing three times with TBST the slides were mounted with ProLong Gold antifade regent and viewed with a SP5 confocal microscope (Leica).
To further analyze the roles of PTP4 and TfR1 in spore infection a protein inhibition assay and an antibody blocking assay were designed. The HFF cell line was used in these assays.
For the protein competition assay, 1×105 HFF cells were cultured in 24 well plates overnight. Purified EhPTP4, TfR1, or BSA (as a negative control) was added to HFF cells at 5 μg/ml and then cells were incubated for 1 hr at room temperature. 1×106 spores were used to infect the HFF monolayers which were then incubated for 9 days until infection was evident. Cells were fixed with 4% paraformaldehyde for 30 min at room temperature. After being washed with PBS for three times, 0.01% Calcofluor White (Sigma-Aldrich) was added into each well and stained for 10 min at room temperature and the parasitophorous vacuole number (infection rate) of each well were counted using Microphoto-FXA epifluorescence microscope (Nikon).
For the EhPTP4 antibody blocking assay, 1×105 HFF cells were cultured in 24 well plates overnight. Anti-PTP4 rabbit polyclonal antibody (rab-PcAb-EhPTP4) or Rabbit seronegative (negative control for polyclonal rabbit antiserum) at a concentration of 5 μg/ml was incubated with 1×106 spores for 1hr at room temperature. Then the antibody and spores mixture were added to the HFF monolayers and cells were incubated for nine days until the number of parasitophorous vacuole were counted as described above.
For the TfR antibody blocking assay, 1×105 HFF cells were cultured in 24 well plates overnight. Anti-HumanTfR rabbit polyclonal antibody (rab-PcAb-HuTfR) or Rabbit seronegative (negative control for polyclonal rabbit antiserum) at a concentration of 5 μg/ml was added to HFF cells and the cells were incubated for 1hr at room temperature. 1×106 spores were added to the HFF monolayers and cells were incubated for nine days until the number of parasitophorous vacuole were counted as described above.
In order to examine the role of TfR1 during the infection of host cells by microsporidia, TRVb cells which do not express endogenous TfR and TRVb-1 cells which stably express the human TfR were utilized (these CHO cell lines were the kind gift of Drs. Parrish and McGraw, Cornell University) [64]. 1×105 TRVb or TRVb-1 cells were cultured in six well plates and infected with 1×106 spores. Fresh media was added every three days until the cells were fixed by 4% paraformaldehyde two weeks post-infection. Samples were washed with PBS followed by staining with 0.01% Calcofluor white (Sigma, USA) for 10 min at room temperature and the parasitophorous vacuole number (infection rate) of each well were counted using Microphoto-FXA epifluorescence microscope (Nikon).
E. hellem spores were purified from the culture supernatant of infected RK13 cells as described above by passage through a 27 gauge needle, followed by a 5 μm Nucleopore filter (Millipore, Billerica, MA) to remove host cells, concentrated by centrifugation, then stained with 0.01% Calcofluor White (Fluorescent Brightener 28; Sigma) for 10 min at room temperature, and then washed 3 times with PBS. These spores were immediately used for infection of CHO cells that had been grown on coverslips in 24 well plates. CHO cells, TRVb-1 and TRVb, were seeded at 1 x105 cells per well into 24 well plates with each well containing a round glass coverslip. The cells were grown overnight to 90% confluence and then infected with 106 labeled spores per well (an MOI of ~4:1). After incubation/infection for 6 hours, the coverslips in the 24 well plates were washed gently with PBS to remove free spores, fixed with 1:1 methanol:acetone for 10 minutes at room temperature, the fixative was then removed and the coverslips were allowed to be air dry at room temperature for 30 minutes. In situ hybridization with an E. hellem Alexa 594-oligo-RNA probe was performed using published methods [69, 70]. Prehybridization of the coverslips was performed with 100μl hybridization buffer (Sigma-catalogue number 11717472001) for 1 hour at 57°C in a humid chamber. Hybridization was then carried out by removing the prehybridization solution and then adding 20uM of the HEL878F probe [70], Alexa Fluor 594-ACTCTCACACTCACTTCAG (ThermoFisher Scientific, HPLC purified custom oligonucleotide), in hybridization buffer (Sigma) at 57°C overnight on a circular rocking platform in a humid chamber. The coverslips were then washed twice with 2X SSC for 15 minutes at 57°C and then mounted with ProLong Gold Antifade Mountant with DAPI (ThermoFisher Scientific). Coverslips were evaluated using Nikon Microphot FXA Microscope employing a triple band D/F/TR filter cube (XF467; Omega Optical, Brattleboro, VT) to determine infectivity rates. Twenty high power (40X) fields (HPFs) were counted for red (intracellular) forms and for the number of CHO cells in each HPF to determine the percent invasion at 6 hours post infection. Only red intracellular forms were counted as having invaded. Any spores that had both red and blue staining were not counted as these were interpreted as representing spores that were either very early in invasion or not able to correctly invade the host cells. Blue spores were not counted as these were interpreted as being extracellular spores adherent to the host cells. Invasion rates were expressed as percent invasion ± SEM and tested for statistical significance as described below.
All animal experiments were conducted according to the U.S.A. Public Health Service Policy on Humane Care and Use of Laboratory Animals. Animals were maintained in an AAALAC-approved facility and all protocols were approved by the Institutional Care Committee of the Albert Einstein College of Medicine, Bronx, New York (Animal Protocols 20150903, 20150904 and 20150905; Animal Welfare Assurance number A3312-01).
No human samples were used in these experiments. Human foreskin fibroblasts were obtained from ATCC.
All of the experiments above were performed in triplicate and repeated at least three times. Each experiment was performed separately with its own negative control. The significance of differences between control and experimental assays was evaluated using two-tailed Student’s T test employing R software for analysis. P values of 0.05 or less were considered statistically significant; P values of 0.01 or less were considered highly significant. Data was also analyzed using the Mann-Whitney U test (nonparametric statistics) and this confirmed the same degree of significance seen with the two-tailed Student’s T test.
The sequence of E. hellem PTP4 has been deposited in the GenBank database under accession number EHEL_071080.
|
10.1371/journal.ppat.0030088 | Regulatory Polymorphisms in the Cyclophilin A Gene, PPIA, Accelerate Progression to AIDS | Human cyclophilin A, or CypA, encoded by the gene peptidyl prolyl isomerase A (PPIA), is incorporated into the HIV type 1 (HIV-1) virion and promotes HIV-1 infectivity by facilitating virus uncoating. We examined the effect of single nucleotide polymorphisms (SNPs) and haplotypes within the PPIA gene on HIV-1 infection and disease progression in five HIV-1 longitudinal history cohorts. Kaplan-Meier survival statistics and Cox proportional hazards model were used to assess time to AIDS outcomes. Among eight SNPs tested, two promoter SNPs (SNP3 and SNP4) in perfect linkage disequilibrium were associated with more rapid CD4+ T-cell loss (relative hazard = 3.7, p = 0.003) in African Americans. Among European Americans, these alleles were also associated with a significant trend to more rapid progression to AIDS in a multi-point categorical analysis (p = 0.005). Both SNPs showed differential nuclear protein-binding efficiencies in a gel shift assay. In addition, one SNP (SNP5) located in the 5′ UTR previously shown to be associated with higher ex vivo HIV-1 replication was found to be more frequent in HIV-1-positive individuals than in those highly exposed uninfected individuals. These results implicate regulatory PPIA polymorphisms as a component of genetic susceptibility to HIV-1 infection or disease progression, affirming the important role of PPIA in HIV-1 pathogenesis.
| Individual risk of acquiring HIV type 1 (HIV-1) infection and developing AIDS is not equal; some people are more prone to HIV/AIDS than others. Susceptibility to HIV-1/AIDS is likely determined by a combination of environmental, viral, and host genetic factors. Genetic variations in host cellular factors involved in HIV-1 cell entry, replication, and host defense have been found to affect susceptibility to HIV-1/AIDS. In this report, we focused on the gene PPIA that encodes cyclophilin A, a human cellular protein that is incorporated into the HIV-1 virion and promotes viral replication. We studied genetic variation in the PPIA gene in persons with different susceptibility levels to HIV-1 infection or different rates of disease progression. We found that individuals who processed two functional variants in the promoter region of PPIA had higher risk of CD4+ T-cell loss or progression to AIDS-defining diseases. We also observed that an additional variant occurred more frequently in HIV-1-infected individuals compared to HIV-1-exposed, but uninfected, individuals. These results suggest that genetic variation in PPIA may influence host susceptibility to HIV-1 infection or disease progression and targeting PPIA might provide therapeutic benefit.
| As an obligate intracellular parasite, HIV type 1 (HIV-1) utilizes host cell factors for its replication. Human cyclophilin A (CypA), also known as peptidyl prolyl isomerase A (PPIA), is a ubiquitous cytoplasmic protein (by convention, we refer to the protein as CypA and the gene as PPIA). CypA has long been known for its incorporation into HIV-1 virions and its important role in facilitating HIV-1 replication in host cells [1,2]. CypA is a member of the cyclophilin family, members of which all possess peptidyl-prolyl cis/trans isomerase activity. Peptidyl prolyl cis/trans isomerases (PPIAses) catalyze the cis/trans isomerization of prolyl peptide bonds and are believed to be involved in protein folding [3]. The incorporation of CypA into the HIV-1 virion capsid is mediated through the direct binding between prolyl peptide bond located in a proline-rich loop of the fourth and fifth helices of the HIV-1 capsid and the active sites of CypA [4,5]. Disruption of CypA incorporation, either by HIV-1 Gag mutations or by cyclosporine A, an immunosuppressive drug that prevents HIV-1 Gag binding to CypA, leads to an attenuation of HIV-1 infectivity [2,6]. Braaten and Luban found that HIV-1 replication was decreased in CypA-null human CD4+ T cells, in which the gene encoding CypA (PPIA) was deleted through homologous recombination [7]. CypA is therefore an important host factor that regulates HIV-1 replication.
Recently, the role of CypA in HIV-1 has gained even greater attention with the discovery of a fusion protein of CypA and TRIM5α, a host restriction factor against HIV-1 [8], which confers HIV-1 resistance in owl monkey [9–11]. Both TRIM5α and CypA recognize and act on the capsid of HIV-1, but apparently confer opposite effects. TRIM5α restricts HIV-1 by promoting premature disassembly of HIV-1 capsid [12], while CypA increases viral infectivity by facilitating proper uncoating. Although the interaction between CypA and TRIM5α is still unclear, it appears that the modulation of HIV-1 infectivity by CypA is independent of TRIM5α [11,13–15]. It has been postulated that binding of CypA to capsid protects HIV-1 from an unknown restriction factor in humans [15].
The study of the influence of human gene variation on susceptibility to HIV-1 infection and progression is an approach that may reveal the in vivo host factor HIV-1 interactions and their epidemiologic importance at the population level. With this approach we have identified several AIDS-modifying variants in genes TRIM5 [16], APOBEC3G [17], and CUL5 [18] that encode human innate HIV-1 restriction factors or related proteins. In a recent study performed in the Swiss HIV Cohort Study, a variant of PPIA was implicated as a potential factor affecting HIV-1 disease progression [19]. As CypA is an essential human protein for completion of the HIV-1 life cycle, we have assessed the influence of genetic variation in the PPIA gene on HIV-1 infection and disease progression in five United States–based HIV-1 natural history cohorts.
The PPIA gene is approximately 7 kb in length, consisting of five exons (Figure 1A). Using PPIA-specific primers we resequenced virtually the entire PPIA gene to screen for single nucleotide polymorphisms (SNPs) in 92 European Americans (EA) and 92 African Americans (AA). Seven polymorphisms were discovered: three in the putative promoter region, one in the 5′ UTR, one in the 3′ UTR, one in intron 4, and one in the second Alu repeat region upstream of the putative promoter. No SNPs were found in the coding regions. These SNPs and four additional SNPs available from dbSNP database located in regions not detected by our sequencing were selected for genotyping in the AIDS cohorts (Figure 1 and Table 1). SNP1 genotypes deviated from the frequencies expected under the Hardy-Weinberg equilibrium (p < 0.001), probably due to its location in the Alu repetitive element; thus, SNP1 was excluded in the analysis. SNP2 was also excluded from analysis due to its rarity. The genotypic frequencies of all other SNPs conform to Hardy-Weinberg expectations.
Allele frequency distributions of PPIA SNPs differed in EA and AA with Fst values ranging from 0.03 to 0.40 (average 0.22). The difference is particularly pronounced for SNPs 7, 8, 9, 10, and 11 (Fst, 0.30–0.40), where the minor alleles are reversed in the two populations (Figure 1).
The extent of linkage disequilibrium (LD) among eight SNPs was assessed by calculating all pairwise D′ values, separately for AA and EA (Figure 2A and 2B). Strong LD was observed among almost all SNPs in AA (D′ range 0.87–1.0) and even stronger LD in EA (D′ range 0.97–1.0). SNP3 and SNP4 were in perfect LD in both populations (D′ = 1, r2 = 1). As each of these two SNPs carries the same information content, SNP4 was selected as a proxy for SNP3 in the analysis. The eight SNPs in both population groups formed a single LD block as defined by the solid spine of LD method [20]. This provided a justification to use all eight SNPs spanning the entire region as one block for subsequent haplotype-based association analyses.
Among AA, SNP6 presented intermediate level of LD (D′ ~0.60) with the SNPs downstream (SNP8–11); SNP6–SNP8 recombinant G-C haplotype had a frequency of 5%. Among EA, SNP5 had weak LD with SNP6–11 (D′ 0.47–0.71); the recombinant haplotype SNP5–SNP6 (G-G) had a frequency of 2.5%. This suggests an occurrence of recombination in the region between SNP5 and SNP8. The presence of multiple Alu elements in PPIA may have introduced the recombination event by promoting genomic instability [21].
Haplotype structure and maximum likelihood haplotype frequencies were estimated with the EM algorithm. There were, in total, eight haplotypes with minor allele frequency >1% in either population (Figure 1B). Five or six haplotypes were present in EA or AA, respectively, accounting for >97% of total sampled chromosomes. Among these, only three common haplotypes (minor allele frequency >5%) were shared between AA and EA (hap1, 2, and 4). Diverse distributions of haplotype frequencies between AA and EA were seen as for SNPs.
We compared the PPIA SNP allele and haplotype frequency distributions among three groups with increasing resistance to HIV-1 infection: seroconverters (SCs), seronegatives (SNs) belonging to an HIV-1 risk group, and those with documented high-risk exposures to HIV-1 who remain uninfected (HREU) (Table 2). The minor allele of SNP5 (G) was carried in 13.7% of SCs, 13.5% of SNs, and 9.5% of HREU, respectively (Mantel-Haenszel trend test, p = 0.21). The reduced carriage of SNP5G in HREU was significant when SCs were compared to HREU (odds ratio = 1.78, p = 0.02), suggesting that the SNP5G allele carriers may have increased susceptibility to HIV-1 infection (Table 2). No distortion of frequency distribution between risk groups were observed for all other SNPs in AA or EA (unpublished data).
We analyzed AA and EA separately since the SNP frequencies differed between the two groups. The Cox proportional hazards model was employed to test the potential differential impact of SNPs on the rates of progression to CD4 <200 or to AIDS clinical diseases (Table 3). To minimize the number of SNPs to be tested, we only tested the unique SNPs represented in each population, i.e., only one SNP was tested for those in perfect LD (r2 = 1). Six and four SNPs were analyzed in AA and EA, respectively. Two SNPs in AA (SNPs 4 and 5) and three SNPs in EA (SNPs 4, 5, and 7) showed significant or near-significant effects (Table 3).
Among AA SCs, the minor allele of SNP4 (the G allele) was associated with accelerated loss of CD4+ T cells (relative hazard [RH] = 3.7, 95% confidence interval [CI] = 1.59–8.63) in a Cox proportional hazards model (Table 3). After stratifying by cohort in the Cox regression analysis, this association became slightly stronger (RH = 4.08, 95% CI = 1.71–9.70). Kaplan-Meier survival analysis presented a clear separation of curves stratified for the SNP4 C/C and C/G genotypes (G/G absent) on progression to CD4 <200 in AA (p = 0.002, log-rank test, Figure 3). Notably, all SNP4 carriers progressed to CD4 <200 within 6 y. The effect on AIDS-1987 was in the same direction but short of significance (Table 3). Identical results were seen for SNP3 (unpublished data). Among EA SCs, a trend of accelerated progression to AIDS-1987 was also observed, though short of significance (RH = 1.34, p = 0.07, Table 3).
To affirm the observed genetic influence of SNP4, we performed a categorical analysis of HIV-1-infected people that included seroprevalents (SPs) in the long-term survivors (Figure 4). SPs who had remained AIDS-free for greater than 7.5 y for AA or 10 y for EA after study enrollment (the median time to AIDS for each) were included in this analysis. The frequency distributions of SNP4 in six discrete time intervals progressing to disease outcome after HIV-1 infection were tested for statistical trend using a Mantel-Haenszel test. As shown in Figure 4, the SNP4 G allele occurred more frequently among fast progressors in both AA and EA. With an increased sample size of 389 patients in AA, a significant trend toward rapid CD4 loss is still preserved (p = 0.04, Figure 4A). Among 970 EA patients, the trend toward rapid progression to AIDS was more pronounced (p = 0.004) (Figure 4B). Thus, the results from SC plus SP groups corroborate the findings from the SC-alone survival analyses, indicating that the accelerating effect of SNP4 is consistent in both AA and EA populations.
In a Swiss Caucasian HIV-1 cohort, SNP5G, referred to as 1650A/G [19], was reported to be associated with a rapid CD4 cell depletion and with a trend toward higher in vitro HIV-1 replication [19]. In our analysis of EA SCs, SNP5G showed a nonsignificant trend toward rapid progression to AIDS-1987 (RH = 1.26, p = 0.08), but had no effect on CD4 cell depletion, while an opposite trend was observed in AA (Table 3). In a categorical analysis of combined samples from SCs and SPs, SNP5G also had no impact on both outcomes in both EA and AA (p > 0.23, unpublished data). Therefore, this study offers no convincing evidence for SNP5 association with disease progression in either EA or AA.
In the haplotype analysis, we first used the Cox regression model to test the global null hypothesis of no association between haplotype frequency and AIDS progression, by comparing a model with all haplotypes to the covariants-only model. This revealed that the global distributions of haplotype frequencies were significantly or near-significantly different for the outcome of CD4 <200 in AA (p = 0.01) and AIDS-87 in EA (p = 0.08). The distortion of frequency distribution was mainly attributable to one haplotype (Hap7) (Table 4). Hap7, the only haplotype carrying SNP4G (Figure 1), was associated with accelerated progression to AIDS in both AA and EA (Table 4), consistent with the results from SNP analysis.
Because TRIM5α interferes with and CypA facilitates post-entry uncoating of the HIV-1 capsid, it is possible that there may be genetic interactions between two genes. Potential interactions were tested only between variants in two genes with functional plausibility or high strength of association. Specifically, PPIA SNP4 was tested for its interaction with TRIM5-rs16934386 and TRIM5-R136Q, which was previously reported using the same patient population to be associated with HIV-1 infection but not progression [22]. In the Cox regression model analysis, we observed no obvious interaction for infection or progression (p > 0.15, unpublished data). This suggests that CypA and TRIM5α likely act independently, consistent with recent in vitro observations [11,13–15,23].
As SNP3, 4, and 5 reside in the regulatory region of PPIA, we performed gel shift assays to assess whether these SNPs differentially bind to transcriptional factors (Figure 5). In this experiment, ~25-bp DNA probes containing either of the SNP alleles were incubated with the nuclear extracts from Th1 cytokine-stimulated human T lymphocytes. Both SNPs showed differential binding affinity to nuclear protein(s) with the same mobility (Figure 5A, top band). Averaged from two independent experiments, a 4-fold decrease or 2.8-fold increase in the band density was observed for the minor allele of SNP3 (G) and SNP4 (G), respectively. The cold unlabeled oligonucleotides with minor alleles were unable to compete with wild-type oligonucleotides, suggesting the bindings are specific. These results indicate that these two SNPs within the promoter alter the binding affinity with certain transcriptional protein(s). On the contrary, no obvious differential binding was observed for SNP5 (Figure 5A).
In silico analysis of the sequence at these three SNP sites by using TESS software also predicted differential binding for SNP3 and 4, but not for SNP5. SNP3 and 4 were each located in one of the consensus-binding motifs of transcription factor SP1, as shown in Figure 5B. The local sequence with SNP3C is identical to the consensus sequence of SP1 site, GGGGCC, whereas SNP3C>G change perturbs this motif. In contrast, the sequence with SNP4C has two mismatch sites compared to an alternative consensus sequence of SP1 site, GAGGCGGGGC, whereas SNP4C>G reduced the mismatch to one, predicting a stronger binding. There is, therefore, a plausible biochemical basis for the differential binding affinity observed in the gel shift experiment. These results also suggest that SNP3 and SNP4 affect transcription, likely through their interactions with SP1. The sequence of the region of human PPIA promoter resided in by SNP3 and SNP4 was compared to the corresponding sequence from other species (Figure 5C). SNP3 and SNP4 are both located in the highly conserved regions. The C allele of SNP3 is present in both primates and rodents. The C allele of SNP4 occurs only in primates and the minor allele G occurs in rodents, suggesting a possible association with speciation. These interspecies data support the hypothesis that SNP3 and SNP4 changes may have functional consequences.
The role of CypA in promoting HIV-1 replication has been well established through extensive in vitro experiments [1,2,6,7]. In this study, we have undertaken a systematic investigation of the association between genetic variations in the PPIA gene and susceptibility to HIV-1 infection and disease progression in HIV-1 natural history cohorts. We found that two promoter variants in perfect LD, SNP3 and SNP4, in PPIA, were associated with rapid HIV-1 disease progression. This effect was revealed for CD4 cell depletion or for AIDS progression, reflecting early or late stage of disease progression, in AA and EA, respectively. The racial difference in the strength or timing of the effects may be due to interactions with other sequence variations that are distributed differently in two populations or to immunologic and viral variables. We further found that both SNPs had differential binding efficiency to nuclear proteins in the gel shift experiment. These findings suggest that the functional promoter SNP3/SNP4 influences HIV-1 disease progression.
PPIA in human encodes a 165-amino acid protein that is highly conserved across species, with 100% and 96% sequence identity to rhesus monkey and mouse, respectively. The nucleotide sequence of PPIA is 98% identical between human and chimpanzee. No variation in the coding sequence of human PPIA was found in this study or in the dbSNP database. This suggests that differential genetic impact of PPIA SNPs and haplotypes on infection and progression are likely due to regulatory region variation affecting protein levels. The regulation and promoter function of human PPIA is still undefined. PPIA is highly expressed in almost all cell types and is considered a housekeeping gene. Sequence analysis revealed that PPIA promoter contains a TATA box and two SP1 binding sites and is highly GC-rich (~70%) with overrepresented CpG dinucleotides [24]. In this study, we identified two SNPs, each located in two additional SP1 binding elements. SP1 is generally believed to function as a transactivating factor for many housekeeping genes [25]. These two SNP changes, SNP3G and SNP4G, always occurring together through LD, had opposite effects on the nuclear protein binding. Assuming the predicted stronger SP1 binding effect of SNP4 is dominant over the weaker SP1 binding of SNP3, a higher PPIA gene expression would be predicted. This would be consistent with more rapid disease course as observed in this study.
Based on our analyses of eight SNPs covering the entire gene region of PPIA and their haplotypes, we found that the PPIA disease-modifying effects were most likely afforded by promoter SNP3 or SNP4 and the associations with the downstream SNPs 7–11 were largely due to tracking of these two SNPs through LD. In a previous report, Bleiber et al. studied the ex vivo HIV-1 infectivity in CD4+ T cells from healthy individuals carrying the SNP4 and SNP5 alleles (named as PPIA-1604 and PPIA-1650, respectively, in the Bleiber's report), as well as their impact on the CD4 gradient (from CD4+ T-cell counts 500/mm3 to 200/mm3) among SPs enrolled in the Swiss HIV cohort, largely comprised of Caucasians [19]. In the ex vivo experiment, both SNPs showed nonsignificant tendencies toward greater HIV-1 replication, with the effects of SNP5 being more pronounced. In the population study, both SNPs were associated with faster CD4+ T-cell depletion in a model also containing two other candidate genes (model 2); with the SNP5 of PPIA effect again being more pronounced. Thus, the detrimental effect of SNP4 observed from these two studies is largely consistent. In contrast, no evidence supporting a role of SNP5 in AIDS progression was found, although the SNP5 G allele was carried significantly less often in HREU than in SC but was similarly distributed in SC and SN. Thus, the role of CypA polymorphism in HIV-1 infection remains inconclusive and warrants further investigation. Taken together, the association and functional results from these two studies point to a role of PPIA genetic variation in HIV-1/AIDS. In future studies, it may be informative to determine the mechanism of association between PPIA variants and host susceptibility to HIV-1 infection or disease progression in view of the interaction between CypA and the capsid domain of the Gag polyprotein.
Population-based genetic association studies provide a powerful approach to uncovering host genes that confer disease susceptibility or resistance. This approach has led to identification of various genetic factors that affect HIV-1/AIDS, and has provided unique insights into the host HIV-1 interaction [26,27]. The successful identification of true genetic associations requires a large sample size (to minimize the type II error), replication in independent studies (to ward off type I error), and plausible functional evidence (to infer causal relationship). In this study, the plausible genetic-modifying role of PPIA is supported by replication in two independent ethnic groups and the demonstration that SNPs 3 and 4 are each associated with altered binding affinities to transcription factors. Moreover, the previous independent study also provides supportive association and functional evidence [19].
Our comprehensive analysis of nearly all variant alleles and their haplotypes in the PPIA gene is, in essence, equivalent to a gene-based approach [28]. A major problem facing genetic association studies is the difficulty in replication, particularly for single SNP associations. One reason for failure to replicate previous studies is due to potential differences in allele frequencies and LD structure across populations. Gene-based replication has recently been advocated as a gold standard for replication studies [28]. The entire gene with prior SNP association is considered the functional unit and is examined for association with effectively all genetic variation in the gene. Nonreplication due to population differences may be minimized as local allele frequencies and LD structure from study populations are considered with this approach [28]. Thus, our comprehensive survey of PPIA genetic variation and LD structure may facilitate future comparisons of replication studies using a gene-based approach.
In summary, through genetic epidemiological and functional approaches, we have identified two promoter SNPs in PPIA as potential genetic modifiers of HIV-1 disease progression. Our findings corroborate the notion that genetic variation of PPIA influences AIDS pathogenesis and provide in vivo evidence that CypA is a critical host protein in interaction with HIV-1. Manipulation of PPIA may be considered as a plausible option for anti-HIV-1 therapeutic development as previously explored [29].
Study participants were enrolled in five United States–based natural history HIV/AIDS cohorts. AIDS Link to the Intravenous Experience (ALIVE) is a community-based cohort of intravenous injection drug users in Baltimore enrolled in 1988–1989 [30], consisting of 92% AA; Multicenter AIDS Cohort Study (MACS) is a longitudinal prospective cohort of men who have sex with men (MSM) from four United States cities: Chicago, Baltimore, Pittsburgh, and Los Angeles, enrolled in 1984–1985 [31], consisting of 83% EA and 10% AA; the San Francisco City Clinic Study (SFCC) is a cohort of MSM originally enrolled in a hepatitis B study in 1978–1980 [32], consisting of 96% EA; Hemophilia Growth and Development Study (HGDS) is a multicenter prospective study that enrolled children with hemophilia who were exposed to HIV-1 through blood products between 1982 and 1983 [33], consisting of 72% EA and 11% AA; the Multicenter Hemophilia Cohort Study (MHCS) is a prospective study that enrolled persons with hemophilia [34], consisting of 90% EA and 6% AA. The participant group is comprised of HIV-1 SCs (infected after study enrollment), SPs (infected at study enrollment), SNs, and HREU. The MACS, MHCS, SFCC, and ALIVE cohorts consist of both SCs and SPs among HIV-1-infected individuals. Due to the potential frailty bias (missing the most rapid progressors to AIDS and death) among SPs, only SCs from these four cohorts were used in the survival analysis. SPs were also included for allele frequency estimation, haplotype inference, and disease categorical analysis. The number of participants studied in each risk or disease category was as follows: SC = 654 EA, 295 AA; SN = 604 EA, 350 AA; HREU = 153 EA, 81 AA; and SP = 1199 EA, 773 AA. Of 290 AA SCs, 237, 42, 5, and 5 were from ALIVE, MACS, MHCS, and HGDS, respectively.
The date of seroconversion after study enrollment was estimated as the midpoint between the last seronegative and first seropositive HIV-1 antibody test; only individuals with less than 2 y elapsed time between the two tests were included in the seroconverter progression analysis. The censoring date was the earliest of the date of the last recorded visit, or December 31, 1995 for the MACS, MHCS, HGDS, and SFCC or July 31, 1997 for the ALIVE cohort to avoid potential confounding by highly effective anti-retroviral therapy (HAART). The censoring date was extended in the ALIVE cohort because of delayed uptake of HAART in this group [30,35].
HIV-1-uninfected individuals were classified into two categories based on individual's documented exposure levels to HIV-1. HREU individuals were those 80 AA and 145 EA with documented high-risk exposure through sharing of injection equipment [36], who had anal receptive sex with multiple partners [37], or transfusions with Factor VIII clotting factor prior to 1984, when heat treatment was initiated [38]. SN individuals (n = 420 and 571, respectively, for AA and EA) are those enrolled in the cohorts who remained HIV-seronegative despite ongoing or prior risk activity.
The study protocols were approved by the Institutional Review Boards of participating institutions and informed consent was obtained from all study participants.
Nucleotide polymorphisms were discovered in a panel of 92 EA and 92 AA, representing the extremes of the distribution for rapid and slow progression to AIDS and HREU. A nonisotopic RNA cleavage assay following PCR was employed to screen for polymorphisms [39]. PPIA has a high degree of homology to multiple processed pseudogenes that varies from 75% to 95% in the exon regions. PPIA also contains six copies of Alu repeats, each with a length of approximately 300 bp, located in the immediate upstream of the putative promoter region and introns [24]. Sequence comparison and BLAST search were performed to select PPIA-specific PCR primers. Overlapping primers covered nearly the entire PPIA gene region including the putative promoter region, 5′ and 3′ UTRs, all five exons, as well as the Alu repeat regions. A part of intron 1 was not covered due to the high GC content. Primer sequences are presented in Table S1 and are numbered according to the GenBank DNA sequence X52851. Additional intronic SNPs were selected from NCBI dbSNP (http://www.ncbi.nlm.nih.gov/SNP) and HapMap databases (http://www.hapmap.org), by considering location, spacing, and allele frequency at least 10%. Haplotype tagging (ht)SNPs were given preference in the SNP selection. The ancestral allele state of the SNPs was based on a reference chimpanzee sequence.
Genotyping was performed using PCR-restriction fragment length polymorphism (PCR-RFLP) assay or TaqMan assays. PCR-RFLP was carried out with 35 cycles of denaturing at 94 °C for 30 s, annealing at 60 °C for 30 s, and extension at 72 °C for 45 s. The PCR product was digested with respective restriction enzymes (New England Biolabs, http://www.neb.com) overnight and then separated on 3% agarose gels. TaqMan assays were obtained from the Assay-by-Demand service of Applied Biosystems (http://www.appliedbiosystems.com). Genotyping primers and conditions were presented in Table S2. Eight water controls were included on each plate to monitor the potential PCR contamination, and 10% of SC and HREU samples were genotyped twice. The genotypes obtained were free of water contamination or of inconsistencies between duplicates.
To assess the difference in allele frequency distribution in two populations, we performed a contingency χ2 test for each marker to test the null hypothesis that the allele frequencies are the same in the two populations. Fst values were estimated by Wier and Cockerham's method [40].
Pairwise LD was quantified using the absolute value of D′. Absolute values of D′ range from 0 for independence to 1 for complete LD between the pairs of loci. LD plots were generated utilizing Haploview (http://www.broad.mit.edu/mpg/haploview) [20]. A triangular matrix of D′ value was used to demonstrate LD patterns within AA and EA. Haplotype blocks were estimated using the solid spine of LD method [20]. Haplotype blocks were defined with a default algorithm based on confidence intervals of D′ [41], or the solid spine of LD method, which creates blocks of SNPs that have contiguous pairwise D′ values of greater than 0.8. With the latter method, the first and last SNPs in a block are in strong LD with all intermediate SNPs, but the intermediate SNPs are not necessarily in LD with each other [20]. Eight SNP haplotype frequencies were inferred separately for each population by means of an expectation-maximization algorithm [42].
Association analyses were conducted using the statistical packages SAS (version 9.0, SAS Institute, http://www.sas.com). EA and AA groups were analyzed separately because allele and haplotype frequencies were quite different between the two groups. Conformity to the genotype frequencies expected under Hardy-Weinberg equilibrium was examined for each SNP. The genetic effects of SNPs on HIV-1 infection susceptibility were assessed by comparing allelic and genotypic frequencies between HIV-1 HREU and HIV-1 SC participants using the chi square or Fisher's exact test. Regardless of the exposure route, persons in the HREU or SN groups were at risk to HIV-1 infection based on their inclusion in a HIV-1 risk group; therefore, we combined participants across cohorts to achieve a reasonable statistical power.
Kaplan-Meier survival statistics and the Cox proportional hazards model (Cox model) were used to assess the effects of SNPs and haplotypes on the rate of progression to AIDS. Two separate endpoints reflecting advancing AIDS pathogenesis were considered for SCs: (1) HIV-1 infection plus a decline of CD4+ T-cell counts <200 cells/mm3 (CD4 <200); (2) the 1987 Center for Disease Control definition of AIDS (AIDS-87): HIV-1 infection plus AIDS-defining illness [43]. The significance of genotypic associations and RH was determined by unadjusted and adjusted Cox model regression analyses. For each SNP, we compared the minor allele genotypes to the most common genotype as a reference group. All p-values were two-tailed. Genetic factors previously shown to affect progression to AIDS were pre-determined to be included as confounding covariates in the Cox model analysis: CCR5 Δ32, CCR2-64I, CCR5-P1, HLA-B*27, HLA-B*57, HLA-B*35Px group (including HLA-B*3502, B*3503, B*3504, and B*5301), and HLA Class I homozygosity for EA (reviewed in [26,27]); HLA-B*57 and HLA Class I homozygosity for AA. CCR2-64I, HLA-B*27, and HLA-B*35Px were not considered as covariates in AA due to no or weak effects in the AA participants, and CCR5 Δ32 was not considered due to its rarity in AA. Analyses were stratified by sex and by age at seroconversion: 0–20, >20–40, and >40 y [27]. Further stratification by cohort was also performed for the exploratory analyses. In the stratified Cox regression model, the overall log-likelihood of hazards obtained is the sum over strata of the stratum-specific hazards, as estimated by the method of partial likelihood. As the same criteria for determining startpoints (seroconversion date) and endpoints and the similar sampling strategy and follow-up settings were used across cohorts, we combined SCs from all cohorts for the survival analysis to increase the power. Although these cohorts differ in routes of HIV-1 transmission, no appreciable effect of mode of infection on AIDS progression has been found through re-analysis of more than ten thousands of SCs from 38 studies around the world (including three used in this study) [44].
To test the association of PPIA haplotypes and HIV-1 disease progression, we first performed a global test in the Cox regression model for each of two disease outcomes separately for AA and EA. The global null hypothesis is that the odds ratios of all haplotypes are equal between cases and controls. Likelihood ratio tests were used to compare a full model with all haplotypes and a base model with only covariates. When the significance of the global test exceeded a relaxed nominal level, p < 0.10, the associations of individual haplotypes were further tested.
To assess the level of correction factor for the number of SNPs, assuming this was a discovery study, we applied spectral decomposition analysis [45]. This multiple testing correction method assesses the equivalent level of independent SNPs taking account of the extensive LD across PPIA. Based on spectral decomposition analysis of SNPs in this study, a corrected p-value of 0.01 would be equivalent to p = 0.05. As this study was a confirmation and extension study of markers with previous positive association, uncorrected p-values were reported.
Cell culture and an electrophoretic mobility-shift assay were performed as described [46]. Freshly explanted human T lymphocytes were obtained from normal donors, purified by isocentrifugation, and activated for 72 h with 1 mg/ml PHA in RPMI 1640 medium containing 10% FCS (Sigma, http://www.sigmaaldrich.com), 2 mM L-glutamine, and penicillin-streptomycin (50 IU/ml and 50 mg/ml, respectively). T lymphocytes were made quiescent by washing and incubating for 24 h in RPMI 1640 medium containing 1% FCS before exposure to cytokines. Cells were then stimulated with 100 nmol/L IL-4 (PeproTech, http://www.peprotechec.com) or 100 nmol/L human rIL-2 (Hoffmann-La Roche, http://www.roche.com) at 37 °C for 10 min. Cell pellets were frozen at −70 °C. The probe sequences were 5′-ttgcgggcggggCcgaacgtggtat-3′ for SNP3C and ttgcgggcggggGcgaacgtggtat-3′ for SNP3G; 5′- ggcgggaggcCaggctcgtgccgtt for SNP4C, and 5′- ggcgggaggcGaggctcgtgccgtt-3′ for SNP4G allele; 5′-aggaaaaccgtgtActattagccatggt-3′ for SNP5A and 5′-aggaaaaccgtgtGctattagccatggt-3′ for SNP5G (SNP site is capped). In cold oligonucleotide competition assay, 100-fold excess of cold unlabeled probe was added as a competitor. The band density was measured by using the software ImageJ (http://rsb.info.nih.gov/ij/index.html). The sequence in and around the SNP sites in the promoter region were searched for the presence of transcription factor binding sites using the program TESS: Transcription Element Search System (http://www.cbil.upenn.edu/tess). Sequence alignment was performed using program ClustalW (http://www.ebi.ac.uk/clustalw).
The Entrez Gene databank (http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?db=gene) accession numbers for the genes mentioned in this paper are PPIA (5478) and TRIM5 (85363). The GenBank (http://www.ncbi.nlm.nih.gov) accession numbers for the human and chimpanzee PPIA genomic sequence are X52851 and NW_001237949, respectively. The Transfac binding site (http://www.cbil.upenn.edu/cgi-bin/tess/tess) accession number is R04195 and R03868 for the SP1 factor. |
10.1371/journal.pcbi.1002390 | Effects of Electrical and Structural Remodeling on Atrial Fibrillation Maintenance: A Simulation Study | Atrial fibrillation, a common cardiac arrhythmia, often progresses unfavourably: in patients with long-term atrial fibrillation, fibrillatory episodes are typically of increased duration and frequency of occurrence relative to healthy controls. This is due to electrical, structural, and contractile remodeling processes. We investigated mechanisms of how electrical and structural remodeling contribute to perpetuation of simulated atrial fibrillation, using a mathematical model of the human atrial action potential incorporated into an anatomically realistic three-dimensional structural model of the human atria. Electrical and structural remodeling both shortened the atrial wavelength - electrical remodeling primarily through a decrease in action potential duration, while structural remodeling primarily slowed conduction. The decrease in wavelength correlates with an increase in the average duration of atrial fibrillation/flutter episodes. The dependence of reentry duration on wavelength was the same for electrical vs. structural remodeling. However, the dynamics during atrial reentry varied between electrical, structural, and combined electrical and structural remodeling in several ways, including: (i) with structural remodeling there were more occurrences of fragmented wavefronts and hence more filaments than during electrical remodeling; (ii) dominant waves anchored around different anatomical obstacles in electrical vs. structural remodeling; (iii) dominant waves were often not anchored in combined electrical and structural remodeling. We conclude that, in simulated atrial fibrillation, the wavelength dependence of reentry duration is similar for electrical and structural remodeling, despite major differences in overall dynamics, including maximal number of filaments, wave fragmentation, restitution properties, and whether dominant waves are anchored to anatomical obstacles or spiralling freely.
| Atrial fibrillation is an abnormal heart rhythm characterized by rapid and irregular activation of the upper chambers of the heart. Atrial fibrillation often shows a natural progression towards longer and more frequently occurring episodes and often occurs in patients with existing heart disease(s). Because atrial fibrillation has several variants, is complex in nature, and evolves over time, it is very difficult and expensive to study comprehensively in large-animal models, in part due to the inherent technical difficulties of imaging whole-atria electrophysiology in vivo. Predictive multiscale computational modeling has the potential to fill this research void. We have incorporated aspects of chronic atrial fibrillation to model some of its various disease states. As such, this study represents the first comprehensive computational study of chronic atrial fibrillation maintenance in a biophysically detailed cell model in a realistic three-dimensional anatomy. Our simulations show that disease-like modifications to cellular processes, as well as to the coupling between cells, perpetuate simulated atrial fibrillation by accelerating the rhythm and/or increasing the number of circulating activation waves. Given the model's ability to reproduce a number of clinically and experimentally important features, we believe that it presents a useful framework for future studies of atrial electrodynamics in response to, e.g., ion channel mutations and various drugs.
| Atrial fibrillation (AF) is a cardiac arrhythmia characterized by rapid and irregular atrial activation. Such desynchronized activation may occur when multiple waves circulate the atria. Unlike ventricular fibrillation, where unsynchronized activation of the ventricles (the main pumping chambers of the heart) causes an immediate and typically fatal loss of blood pressure, atrial fibrillation may be a repetitive, even chronic, disease. In fact, AF is the most common sustained cardiac arrhythmia in the United States and the rest of the developed world [1], [2], with more than 2.3 million sufferers in the U.S. [3]. AF becomes increasingly common with age [2] and is associated with significant mortality and morbidity, such as heart failure and stroke [1].
AF is more prominent in the context of alterations in atrial tissue properties – due to disease, arrhythmias, or age – known as remodeling. In fact, AF itself leads to remodeling, causing electrophysiological (“electrical”), contractile, and structural changes [4]. Although AF can typically be reversed in its early stages, it becomes more difficult to eliminate over time due to such remodeling – hence the expression “AF begets AF” [5].
A central hypothesis for why AF begets AF is that electrical and structural remodeling due to chronic or persistent AF shorten the action potential wavelength, which measures the spatial extent of the action potential. Such wavelength shortening allows more waves to fit in the atria and maintain the arrhythmia [6]. Electrical remodeling primarily shortens the refractory period and the action potential duration (APD) of the atrial action potential, while structural remodeling impedes propagation and hence decreases conduction velocity (CV). Since the wavelength is given as the product of APD and CV (or, alternatively, the product of the effective refractory period and CV), electrical remodeling and structural remodeling both decrease the wavelength, thus potentially perpetuating AF. Additionally, the stability of reentrant waves may be affected by remodeling. Prior modeling work has shown that flattening APD restitution (the dependence of APD on the previous resting interval or Diastolic Interval, DI), which typically occurs as a consequence of electrical remodelling [7], may stabilize reentry [8]. Likewise, diffuse fibrosis, which may occur during structural remodelling [9] may stabilize reentrant waves [10].
Clinically, because electrical and structural remodeling typically present jointly in patients with chronic AF, their effects are difficult to separate. Animal models of primarily electrical remodeling (due to rapid atrial pacing) and predominantly structural remodeling (induced heart failure or mitral regurgitation) exist, however the rapid atrial pacing models also typically develop some degree of structural remodeling while the heart failure animals undergo some concomitant electrophysiological changes [9], [11]. We therefore decided to use computer modeling as a means to investigate the mechanisms of how APD shortening due to electrical remodelling, and CV slowing due to structural remodelling, influence the duration and spatiotemporal dynamics of simulated AF in a computational multiscale model of human electrophysiological dynamics and substrates. Because structural aspects of the complex atrial anatomy are important for, e.g., anchoring waves to anatomical obstacles and thus influencing the duration of reentrant activity, we use an anatomically detailed structural model of the human atria.
We model the atria using the Courtemanche et al. cellular model of human atrial cell electrophysiology [12], with computational cells diffusively coupled to their nearest neighbors in an anatomically derived, three-dimensional structural model of the human atria [13]. As in previous work from our group [14], we increase the conductance of the inward rectifier current, IK1, here by 75%, in order to get the baseline action potential duration and resting membrane potential closer to experimentally observed values. Further, as in our previous work [14], we fix the intracellular concentrations of K+ and Na+ (at 139.0 mM and 11.2 mM, respectively) to avoid long-term drift.
The anatomical model incorporates heterogeneous coupling, resulting in different conduction velocities in different anatomical regions, in agreement with human data [13], [15]. Specifically, the model exhibits fast conduction in Bachmann's bundle, the pectinate muscle network, the crista terminalis, and the limbus of the fossa ovalis (120 cm/s during sinus pacing); slow conduction in the isthmus and the fossa ovalis (36 cm/s), while the remaining (bulk) atrial tissue has intermediate conduction velocity (65 cm/s). The different regions are shown in Fig. S1 in Text S1 (online Supporting Information). The model does not include anisotropy. In human atria, the bulk atrial muscle has a more random fiber orientation than the fast conduction pathways (and also more random than the ventricular myocardium), which have well-organized orientations along the bundles [13]. However, due to the strip-like anatomy of the fast tissues (Fig. S1 in Text S1), anisotropy is predicted to play a minor role there.
This mathematical representation of the atria reproduces basic features such as depolarization time and spatial profile during normal (sinus) pacing [13].
The three-dimensional anatomical model is discretized in a 300×285×210 grid of spatial nodes, with a spacing of Δx = 0.025 cm, and no-flux boundary conditions. The equations were solved using an operator-splitting method [16] with forward Euler integration of both operators. We used a fixed time step of Δt = 0.01 ms for the partial differential equation describing the diffusion of voltage. The cellular model was integrated using an adaptive time step [16].
The code was parallelized using OpenMP, and run on multi-core machines. Simulating 60 s of reentrant activity took 5 days on a 24-core machine (2.66 GHz Intel® Xeon® X7460 processors, 128 GB memory). Because the simulations are this computationally costly, simulations were stopped when reentry terminated or at 60 s (in which case, the arrhythmia was classified as sustained), whichever came first.
Electrical remodeling due to chronic AF was simulated as in previous work, incorporating a 70% decrease in the conductance of the L-type calcium current (ICaL), a 50% decrease in the conductance of the transient outward current (Ito), and a 50% decrease in the conductance of the atrial-specific, ultra-rapid potassium current (IKur) [7]. These values are based on current recordings in cells isolated from human atrial appendages. We refer to this set of values as 100%, or full, electrical remodeling. In order to simulate different degrees of electrical remodeling (10–90%), in some simulations the percentage changes in the three affected conductances were downscaled by the same factor.
We simulate structural remodeling by decreasing the diffusion coefficients (i.e., the coupling strengths between computational cells), which reduces conduction velocity. The three different nominal diffusion constant values (assigned to fast, bulk, and slow conducting tissue) were scaled by the same factor. Maximal structural remodeling was set to a 50% decrease in diffusion, causing the time for full activation of the atria with sinus pacing to increase from 108 ms to 149 ms. However, as experimental and clinical data show a large range of conduction impairment with structural remodelling [9], we investigate two more levels of structural remodeling, using downscaling in diffusion of 70% and 83% (increasing activation time to 119 ms and 130 ms, respectively).
Because the duration of reentrant episodes may depend on where the reentry is initiated, we simulated reentry initiated at three different locations: the left atrial free wall, the left atrium near the left pulmonary veins, and the right atrial free wall. At each of these locations, reentry was initiated using a cross-gradient protocol, using a stimulus current of 80 nA/µF for 1 ms. Because the vulnerable window for reentry initiation is very small in the non-remodeled virtual tissue, we applied a brief hyperpolarizing clamp (−80 mV for 1 ms) to the region of the second wave excitation, 30 ms after its initiation. This allows for earlier reentry into this region and increases the vulnerable window.
The coupling interval (i.e., the time between the first and the second excitation) was varied systematically between simulations in steps of 10 ms within the vulnerable window. The size of the vulnerable window varies with variation in electrical and structural remodeling parameters, but was in the range of 30–60 ms, such that 4–7 reentry simulations were initiated at each location. Applied to three different locations, this means that for a given set of parameters describing the degree of electrical and/or structural remodeling, 12–21 simulations were run with different initiations.
DI and APD were recorded from 16 different locations, spread evenly throughout the atria (see Fig. S2 in Text S1). The APD was measured as the time from the crossing of −70 mV on the upstroke to the crossing of −70 mV during repolarization. Inversely, DI was measured as the time between the crossing of −70 mV during repolarization to the crossing of −70 mV on the next upstroke.
The wavelength (WL) is difficult to measure accurately during reentrant activity, even in computational studies [17], due to wave collisions and irregular wave propagation. We use a method similar to that employed by Graux et al. (Ref. [18]) and determine CV during periodic pacing in the left atrial free wall. However, rather than obtaining CV for very few pacing rates as necessitated in the clinic, we systematically varied the pacing rate to establish the dependence of the CV on diastolic interval. Such restitution curves were obtained for all the combinations of the different levels of electrical and structural remodeling simulated. We later used these restitution curves to estimate local CV for DI values measured during reentry and, finally, to compute the local wavelength as WL = APD×CV. Note that this definition of the wavelength is more practical than WL = ERP×CV, where ERP is the effective refractory period, since ERP measurement requires a series of stimuli and cannot be measured directly during simulated AF. In paced tissue simulations, we found that APD underestimates ERP by 9–16 ms (4–10%) depending on pacing rate and level of remodeling. Hence, our calculations of the wavelength using APD are presumably 4–10% larger than estimates based on ERP.
Atrial fibrillation and flutter can be characterized by the number of wavelets present in the tissue. As in our group's previous work [14], we compute the location of wave tips (filaments) from the crossing of two isopotential curves (the crossing of −30 mV on the upstroke), separated in time by 2 ms. The number of separate filaments is determined by applying a k-means clustering analysis to the filament location data. To characterize individual simulations we use the maximal number of filaments present in that run.
Dominant waves were defined as waves existing for at least five rotations. Their locations were determined directly from the filament location data or (in the case of anchored waves without filaments) based on periodicities in the transmembrane potential from the 16 recording sites, as well as visual inspection of isopotential surface maps.
As shown previously, electrical remodeling leads to shortening of the APD [7]. In particular, in our simulations of tissue strands, full electrical remodeling reduces the APD at 1 Hz pacing from 228 ms to 135 ms, while at 5 Hz the APD is shortened from 134 ms to 103 ms. This reduction in the amount of APD shortening with faster pacing demonstrates the flatter APD restitution occurring with electrical remodeling (see Fig. S3A in Text S1). The CV is unchanged with electrical remodeling (Fig. S3C in Text S1). In contrast, structural remodeling decreases CV (CV restitution slope remains largely unchanged; Fig. S3D in Text S1), while the APD is unchanged (Fig. S3B in Text S1).
As a measure of the effects of electrical and structural remodeling, we focus primarily on the duration of reentrant activity. In our three-dimensional model, in the absence of electrical and structural remodeling, reentrant activity is not sustained: in all simulations of normal tissue, with varying initiation time and location (see Methods), reentry ended 1–3 seconds after initiation. This is consistent with clinical findings in the normal human atria, where AF episodes typically self-terminate soon after initiation.
Fig. 1A shows an example of non-sustained reentrant activity in normal tissue. The wavelength in this case is sufficiently long that the reentrant wave eventually runs into refractory tissue and dies out. In contrast, when simulating full electrical plus structural remodeling, reentry was sustained for 60 s in 18 of 21 simulations. With such electrical plus structural remodeling the wavelength is much shorter than in normal tissue (Fig. 1B), and the reentrant wave in the left atrial free wall does not self-terminate. Videos showing these dynamics with and without remodeling are available as Supporting Information (Videos S1 and S2).
To investigate whether this maintained reentrant activity results from more waves being present versus those present being more stable, we measured the maximal number of filaments in a simulation, as well as the mean duration of reentrant activity, while varying the level of electrical and structural remodeling. Electrical and structural remodeling both lead to increases in the duration of reentrant activity, and the combination of electrical plus structural remodeling gives even longer sustained reentry (Fig. 2A). Structural remodeling also causes an increased number of filaments, while the level of electrical remodeling does not exhibit a clear correlation with the maximal number of filaments (Fig. 2B). Note that the number of filaments present at baseline is consistent with previous experimental and modeling studies [14], [19], [20].
To test the hypothesis that a smaller wavelength perpetuates AF, we determined the dependence of the duration of reentrant activity on the estimated wavelength. For both electrical and structural remodeling, a decrease in WL is associated with longer reentry duration when WL is below a threshold value of around 7 cm (Fig. 3A). Interestingly, the dependence of reentry duration on WL is similar for both electrical and structural remodeling, suggesting that WL is a more important determinant of duration than other factors, such as APD restitution slope, that vary between electrical and structural remodeling.
In contrast, the dependence of the maximal number of filaments on WL is very different for electrical vs. structural remodeling, with more filaments present during structural than electrical remodeling for the same WL (Fig. 3B). This indicates that more conduction block and wavebreaks occur during structural remodeling, but might also be the result of fewer waves being anchored to anatomical obstacles, since an anchored wave does not necessarily have a filament. From analyzing the dynamics of the number of filaments, we found that with structural remodeling, frequent occurrences of conduction block and wavebreaks cause the larger maximal number of filaments relative to electrical remodeling. These wavebreaks often heal, such that the increase in filaments is transient (Figs. S4 and S5 in Text S1).
Taken together, these results demonstrate that the wavelength determines the duration of simulated AF, despite differences in dynamics such as APD restitution, conduction block, and number of filaments.
As mentioned above, in our simulations of periodically paced tissue strands, electrical remodeling shortens APD without changing CV, and structural remodeling decreases CV without affecting APD. If these dependencies hold during reentry in the three-dimensional atrial anatomy, then APD itself should be a marker for reentry duration during electrical remodeling, while CV should correlate with reentry duration during structural remodeling. Such markers might be valuable given the methodological difficulties in determining WL (see Methods).
However, during reentry in the anatomical model with simulated electrical remodeling, there is both a decrease in APD and a concomitant fall in CV (Fig. 4A,C). For structural remodeling, there is a primary decrease in CV (Fig. 4D) and a secondary increase in APD (Fig. 4B). These results show first of all that APD alone (for electrical remodeling) and CV alone (for structural remodeling) are not accurate surrogates for WL during reentry. A similar finding was reported for AF/AFL inductance in a canine model [20].
Importantly, the secondary changes also suggest that the dynamically induced differences in the wave characteristics (APD and CV) may be due to differences in preferred pathways during reentry in the anatomical model, since differences in pathway lengths would affect the size of the excitable gap, and hence cause dynamical changes in DI, APD, and CV.
Our different reentry initiation protocols allow a range of different spatiotemporal dynamics to occur. In all simulations that ran the full 60 s, the reentrant activity settled into a relatively periodic rhythm, with dominant waves remaining in a particular location (often circulating an anatomical obstacle). However, as shown in Table 1, the location of the dominant wave(s) varied significantly. In general, with only electrical remodeling dominant waves tended to be located in the left atrium, while dominant waves were found in the right atrium when we simulated structural remodeling only. With combined electrical and structural remodeling, dominant waves were in either or both atria.
More specifically, for electrical remodeling, 4 of 6 simulations resulted in reentry around the pulmonary veins (Table 1). Fig. 5A shows an example of such dynamics (see also Video S3 in the Supporting Information). A wave is anchored to the left pulmonary veins during the entire rotation (Fig. 5A, left). Another wave front is circulating the right pulmonary veins, but does not remain completely anchored for the entire rotation (Fig. 5A, right). Excitation spreads from the left atrium to the right.
During structural remodeling, all simulations of sustained reentry result in dominant waves rotating around the tricuspid annulus and the inferior vena cava (Table 1). In most cases (5 of 7), the rotation is counter-clockwise around the tricuspid annulus. An example is shown in Fig. 5B (and Video S4), with the wave anchored to the tricuspid annulus on the left, and the wave circulating the inferior vena cava on the right.
With electrical plus structural remodeling, the outcome is more varied. In some cases waves are anchored (to the pulmonary veins, the superior vena cava, or the tricuspid annulus). However, in some simulations, the dominant waves are un-anchored but spiral around the left or the right atrial free walls (Table 1). An example of such a scroll wave is shown in Fig. 5C, left (and Video S5). In this example, there is also a wave anchored around the superior vena cava in the right atrium (Fig. 5C, right, and Video S5), while in other simulations the excitation of the right atrium by the scroll wave in the left atrium is more irregular.
Note that the dominant periods vary with the different interventions. In the baseline (no remodeling) model, the mean period of (non-sustained) reentry is 176 ms. With full electrical remodeling, the shortening of the APD allows a faster mean rhythm of 137 ms. For full structural remodeling, the conduction slowing increases the main period to 205 ms. Interestingly, for full electrical plus structural remodeling, the period is the same as for only electrical remodeling (138 ms), suggesting that the preferred pathways are shorter on average for electrical plus structural remodeling. These values fall within the range seen in patients with paroxysmal and persistent AF [21], [22].
The different rhythms and their occurrence patterns (Table 1) correspond to clinical and experimental observations. The pulmonary veins frequently act as triggers of AF [23], and reentrant waves have been mapped in the pulmonary vein region [24], [25]. In canine models of structural remodeling, and in typical atrial flutter, excitation often occurs around the tricuspid annulus [26], sometimes in concert with reentry around the inferior vena cava [27].
We have incorporated aspects of electrophysiological and structural remodeling as a step in simulating the various disease states of AF. Our simulations show that electrical remodeling alone leads to rapid activation patterns, while structural remodeling causes wavebreaks and wave fragmentation. The decreases in wavelength due to remodeling makes different reentrant pathways possible and cause reentry perpetuation.
Recent years have seen a large increase in modeling atrial-specific aspects of arrhythmogenesis. In particular, there has been increased development of anatomical models, powered by increasing computational speed and data handling (see, e.g., [28] for a review on modeling and [29] for technical details on computational and visualization aspects).
At the cellular level, multiple mathematical models describe the same ionic currents in different representations of human atrial myocytes. The Courtemanche et al. model [12] and the Nygren et al. model [30] are both well-established and have been compared in great detail [31], [32]. Although their simulated behavior can be quite different, we believe that neither is empirically better. Rather they may represent intrinsic variability. Given that neither is obviously better, we have opted to use the Courtemanche et al. model, largely because it is more widely used. The Nygren et al. model was recently updated in terms of some of its potassium currents [33] and its intracellular calcium handling system [34].
The effects of electrical remodeling on action potential morphology, in particular the role of the individual currents involved in remodeling [7], [35], have been studied at the cellular level. In two-dimensional tissue, electrical remodeling accelerates spiral waves generated with both the Courtemanche et al. model [32], [36] and the Nygren et al. model [32]. With electrical remodeling there is also a decrease in spiral wave meandering in the Courtemanche et al. model [32], [36], but not with the Nygren et al. model [32]. Such a decrease in spiral meandering with the Courtemanche et al. model is enhanced with increased IK1 [37] and can indeed occur in simulated atrial tissue with increased IK1 in the absence of electrical remodeling [38]. Increased IK1 also accelerates spiral waves [37], as does another inwardly rectifying current IK,ACh, which is triggered in cholinergic AF [39], [40].
In simulated tissue (Ref. [36], as well as in our simulations (Fig. 4)), electrical remodeling decreases the wavelength in tissue strands and causes arrhythmias in anatomical models to be of longer duration (Fig. 2; [32], [36]). Decreasing the calcium current conductance, which is a main component of electrical remodeling, has similar effects in a human anatomical structure of virtual guinea pig ventricular cells [17], [19]. Combining electrical remodeling and left atrial dilation leads to increased vulnerability to reentry in an anatomical model [41], while combining electrical remodeling and decreased intercellular coupling causes shortened wavelength and sustained spiral wave activity in two-dimensional tissue simulations [42], consistent with our results.
Other lines of study, pursued with complex atrial models, include the effects of myocardial stretch on conduction [43], [44], incorporation of intrinsic APD heterogeneity [45], [46], [47], and simulated ablation [48], [49].
The mechanism underlying AF maintenance is not entirely clear. There are two predominant theories: (i) the multiple wavelet hypothesis and (ii) the “mother rotor” hypothesis. The multiple wavelet theory [50] hypothesizes that AF is composed of multiple interacting electrical wavelets and is maintained by the processes of wavebreak and reentry. The mother rotor theory [51] hypothesizes that, rather than the multiple, equally important wandering wavelets of the multiple wavelet hypothesis, there is one dominant “mother” reentrant wave that sheds and initiates daughter waves as conduction block occurs at multiple sites away from its core.
The dynamics in our simulations are characterized by one or two reentrant waves, rotating fairly periodically around an anatomical obstacle or un-anchored in the (left or right) atrial free wall. Hence, our simulations are more consistent with the mother rotor theory. Further, in the presence of simulated structural remodeling, waves emanating from these dominant waves often exhibit local conduction block and transient fragmentation. Recent high-density mapping studies of patients with persistent AF and structural heart disease also found increased occurrence of conduction block in the right atrium compared to patients without persistent AF and structural heart disease [52]. However, the AF dynamics in the former patient group is characterized by more epicardial breakthroughs and more waves than our simulations, possibly due to increased decoupling of neighboring myocyte bundles [52], [53]. Using non-invasive imaging techniques, Cuculich et al. determined the number of wavelets in AF patients as 1–5, with more wavelets in patients with long-term persistent AF (average 2.6) than in patients with paroxysmal AF (average 1.1) [25], consistent with our findings. However, these patients also had a considerable number of focal sites (possibly due to spontaneous triggering, microreentry, or epicardial breakthroughs) [25], not seen in our simulations.
Atrial flutter (AFl) is usually associated with a single macroreentrant circuit and may be very regular. However, AF can also exhibit a large degree of both temporal periodicity and spatial organization [54], [55]. Hence, our simulated arrhythmias demonstrate characteristics of both AF and AFl. Clinically, AF and AFl may occur in the same patient, and AFl often converts to AF.
By initiating reentry at different locations and at different times, we were able to explore a large range of spatiotemporal dynamics in our simulations. A large variety of ensuing dynamics is also observed clinically and experimentally. While much of this variability may stem from differences among experimental AF models and patient-to-patient variation in underlying heart disease, age, and stage of remodeling, there is considerable variability in activation patterns even among patients with the same underlying condition [22].
Importantly, many features observed in our simulations correlate directly with experimental and clinical characteristics. One example is reentry around the pulmonary veins, seen in our model with electrical remodeling. Another example is reentry around the tricuspid annulus, with 5 of 7 simulations of sustained activity with structural remodeling resulting in the direction of rotation being counter-clockwise. Such right-atrial reentry, involving slow conduction through the isthmus, is typical in patients with AFl [26], [56] often with the majority (19/26) of cases having counter-clockwise rotation around the tricuspid annulus [57].
Further, with structural remodeling in our simulations, waves were also anchored around the inferior vena cava. Such dual-loop reentry involving the tricuspid annulus and the inferior vena cava is often observed in typical atrial flutter [27]. In addition to anatomical reentry, we also observed several examples of meandering scroll waves, which have been mapped experimentally and simulated computationally [58]. Finally we observed incidences of double-wave reentry, also observed clinically [56] and experimentally [59].
Rate gradients, with higher activation frequencies in the left atrium, exist in some animal models of AF [11], [54], [55] and have been observed in patients with both paroxysmal and persistent AF [22], [60], [61], [62]; however, other chronic AF patients do not exhibit such inter-atrial variability [21]. In our simulations of remodeled tissue, there are no significant left-to-right gradients in activation frequency. However, our simulated AF with electrical remodeling alone was primarily a left atrial phenomenon in the sense that most dominant waves were found in the left atrium. This suggests that under some circumstances, the left atrium may be the driver of AF even in the absence of electrophysiological left-to-right heterogeneity (see below).
The extent to which perpetuation of AF depends on the wavelength varies considerably among different studies using different AF models and different methods for wavelength estimation. While several studies, both computational and clinical, have demonstrated a facilitation of AF maintenance with a decrease in wavelength [17], [18], others have not found such a dependence [63].
We found a clear functional dependence of reentry duration on wavelength in our simulations, and found that the wavelength must be below a value of around 7 cm for reentry perpetuation. However, the wavelength has to be considerably shorter for the majority of simulations to be sustained; this critical value for sustenance is about 5 cm. This value is consistent with a previous report of 5 cm using a direct measurement of the wavelength during simulated AF [17]. In contrast, clinical estimates for this threshold tend to be larger, with reported values of 12–13 cm [20], [64]. Importantly, as detailed in Ref. [17], these values are obtained through indirect methods, due to the inherent problem of insufficient spatial mapping, and tend to overestimate the wavelength.
AF and AFl become increasingly common with age, and are also particularly frequent in patients with existing structural heart disease, valvular heart disease, coronary artery disease, ischemic heart disease, hypertension, or a history of heart attacks. Indeed, the majority of AF patients have one or more cardiovascular diseases in addition to AF. Electrical and structural remodeling due to chronic AF occur in concert with substrate changes due to any existing condition(s).
Changes in ionic currents due to chronic AF have been observed consistently in experimental models and in patients [9]. Recent studies have shown differences in several of the outward potassium currents between the left and the right atrial appendages, with some current differences being present in sinus rhythm or paroxysmal AF and others in chronic AF [65], [66]. As it is presently unclear what type of inter-atrial gradients these appendage differences represent, we did not incorporate any left vs. right electrical heterogeneity in our model at this point.
Structural remodeling in chronic AF shows more variability among patients and animal models than does electrical remodeling. Structural remodeling processes also occur on a much slower time scale than electrical remodeling, which may contribute to the substrate variability. The processes include myocyte hypertrophy, fibrosis, and changes in the expression levels of connexin, the protein comprising the gap junctions that couple cells. Fibrosis can manifest in both patchy patterns and diffuse morphologies. Fibrosis and decreases in connexin levels both impede propagation, while atrial dilation and cell hypertrophy increase atrial activation times. Our simulations using decreased coupling between virtual cells, which decrease conduction velocity and increase activation times, represent an unspecified propagation impairment that may be thought of as a decrease in connexin levels or a diffuse fibrosis.
Although contractile remodeling is associated with chronic AF, in addition to electrical and structural remodeling, we do not include such remodeling here. The main reason is that the current data on contractile remodeling in the human atria is insufficient to incorporate into a quantitative model. Further, attempts at including contractile remodeling into the cellular model would almost certainly require a considerable expansion of the intracellular calcium handling system such as that recently developed by Koivumäki et al. [34]. Such an expansion would significantly increase the computational load of the simulations, and as we are already operating on the cusp of computational tractability, we did not incorporate those changes in this study.
The pulmonary vein plays an important role in triggering AF and electrical isolation (ablation) of this area can be an efficient treatment for AF [23]. However, the pro-arrhythmic role of the pulmonary vein region is significantly lessened in chronic AF [61], [67], [68], which is also exemplified by the decreased success of ablation therapy in this patient group. The pulmonary vein region has electrophysiological properties that are different from other regions of the atria, typically showing shorter refractory periods [67], [69]. The details of which ionic currents are responsible for this heterogeneity and how the electrophysiology changes with progression of AF are not known, and hence we did not attempt to include special heterogeneity of the pulmonary vein region in this work. Further, this study focused on the maintenance, rather than the initiation of AF; further studies will be required to help elucidate triggering of AF due to mechanisms such as repolarization alternans and ectopy [14], [70] in remodeled and healthy tissue.
As described in the Methods, the three-dimensional atrial structure includes a gross coupling heterogeneity, with fast, normal, and slow tissues. However, the model does not include anisotropy. More prominent anisotropy during some types of structural remodelling [9] may play a role in AF perpetuation.
We simulated structural remodeling as a decrease in coupling. A first step towards more realistic simulations of fibrosis could include implementing structural remodeling by randomly assigning a variable fraction of cells to be electrically inactive and surrounded by no-flux boundary conditions [10]. Test simulations show that such implementation of structural remodeling does not alter our conclusions. Recent studies have shown how fibrosis may affect repolarization [71] in addition to slowing down conduction - we did not include such effects in this work.
Electrical and structural remodeling both shorten the atrial wavelength; electrical remodeling primarily through a decrease in APD, while structural remodeling primarily slows down conduction. The wavelength determines the mean reentry duration in the same manner for electrical and structural remodeling despite major differences in overall dynamics (maximal number of filaments, conduction block, wave fragmentation, restitution properties, preferred dominant wave pathways, and whether dominant waves are anchored to anatomical obstacles or spiralling freely). As such, these findings have implications for our understanding of the mechanisms by which AF remodeling processes perpetuate AF.
|
10.1371/journal.ppat.1000974 | Adaptive Evolution of Mus Apobec3 Includes Retroviral Insertion and Positive Selection at Two Clusters of Residues Flanking the Substrate Groove | Mouse APOBEC3 (mA3) is a cytidine deaminase with antiviral activity. mA3 is linked to the Rfv3 virus resistance factor, a gene responsible for recovery from infection by Friend murine leukemia virus, and mA3 allelic variants differ in their ability to restrict mouse mammary tumor virus. We sequenced mA3 genes from 38 inbred strains and wild mouse species, and compared the mouse sequence and predicted structure with human APOBEC3G (hA3G). An inserted sequence was identified in the virus restrictive C57BL strain allele that disrupts a splice donor site. This insertion represents the long terminal repeat of the xenotropic mouse gammaretrovirus, and was acquired in Eurasian mice that harbor xenotropic retrovirus. This viral regulatory sequence does not alter splicing but is associated with elevated mA3 expression levels in spleens of laboratory and wild-derived mice. Analysis of Mus mA3 coding sequences produced evidence of positive selection and identified 10 codons with very high posterior probabilities of having evolved under positive selection. Six of these codons lie in two clusters in the N-terminal catalytically active cytidine deaminase domain (CDA), and 5 of those 6 codons are polymorphic in Rfv3 virus restrictive and nonrestrictive mice and align with hA3G CDA codons that are critical for deaminase activity. Homology models of mA3 indicate that the two selected codon clusters specify residues that are opposite each other along the predicted CDA active site groove, and that one cluster corresponds to an hAPOBEC substrate recognition loop. Substitutions at these clustered mA3 codons alter antiviral activity. This analysis suggests that mA3 has been under positive selection throughout Mus evolution, and identified an inserted retroviral regulatory sequence associated with enhanced expression in virus resistant mice and specific residues that modulate antiviral activity.
| APOBEC3 (mA3) is a cytidine deaminase with antiretroviral activity. Genetic variants of mA3 are associated with the restriction factor Rfv3 (recovery from Friend leukemia virus) and with resistance to mouse mammary tumor virus. We sequenced mA3 from laboratory strains and wild mouse species to examine its evolution. We discovered that the mA3 allele in virus resistant mice is disrupted by insertion of the regulatory sequences of a mouse leukemia virus, and this insertion is associated with enhanced mA3 expression. We also subjected the Mus mA3 protein coding sequences to statistical analysis to determine if specific sites are subject to strong positive selection, that is, show an increased number of amino acid replacement mutations. We identified 10 such sites, most of which distinguish the mA3 genes of Rfv3 virus restrictive and nonrestrictive mice. Six of these sites are in two clusters that, in human APOBEC3G, are important for function. We generated a structural model of mA3, positioned these clusters opposite each other along the putative mA3 active site groove, and demonstrated that substitutions at these sites alter antiviral activity. Thus, mA3 has been involved in genetic conflicts throughout mouse evolution, and we identify an inserted regulatory sequence and two codon clusters that contribute to mA3 antiviral function.
| Species susceptible to infectious retroviruses have evolved numerous constitutively expressed antiviral factors that target various stages of the retroviral life cycle. The factors responsible for this intrinsic immunity include 3 that act at post-entry stages of virus replication: Fv1, APOBEC3 and TRIM5α. Fv1 was discovered in mice, [1] and only mice carry Fv1 [2], [3]. TRIM5α was initially identified in primates as an anti-HIV-1 restriction factor [4], [5], and while mice carry TRIM5α related sequences [6], no mouse orthologue with virus restriction activity has been identified. Active APOBEC3 genes, on the other hand, are found in various species including human and mouse, and mouse and human APOBEC3 have antiviral activity against multiple retroviruses [reviewed in 7].
The APOBEC3 editing enzyme is incorporated into budding virions. During reverse transcription in subsequently infected cells, the virion-associated APOBEC3 catalyzes C-to-U deamination, resulting in G-to-A mutations in the viral DNA [8]. The increased mutational load has a major impact on viral fitness, and there is also some evidence that APOBEC3 antiviral activity is enhanced by additional deamination-independent mechanisms that act before proviral integration [9], [10].
APOBEC3 was initially described in primates, and human APOBEC3 paralogues responsible for resistance are present as a cluster of 7 genes on chromosome 22, the most extensively studied of which is APOBEC3G (hA3G). HIV-1 can avoid inhibition by hA3G through the action of one of its viral accessory proteins, Vif (viral infectivity factor), that prevents incorporation of hA3G into the virion [11]. The antiviral activity of hA3G can be observed with Vif-negative HIV-1 and SIV lentiviruses as well as other retroviruses such as equine infectious anemia virus (EIAV) and mouse leukemia viruses (MLVs). In the mouse, there is only a single APOBEC3 copy (mA3) on chromosome 15. Several observations indicate that mA3 functions in antiviral defense: mA3 inhibits infection by several viruses including HIV-1 and mouse retroviruses such as mouse mammary tumor virus (MMTV), intracisternal A-particles (IAPs) and MusD endogenous retroviruses [12]–[14]; mA3 knockout mice are more susceptible to MMTV infection and tumorigenesis [15]; endogenous retroviruses (ERVs) of MLV in the sequenced Mus genome show modifications consistent with APOBEC3 activity [16].
Two recent studies proposed that mA3 is responsible for the Friend virus resistance factor Rfv3 [10], [17]. Rfv3 is one of several host resistance factors that, like Fv1, were discovered in studies with the pathogenic Friend MLV (FrMLV) [18]. Rfv3 was identified as a non-major histocompatibility complex gene that influences the duration of viremia, partly through its effects on the production of virus-neutralizing antibodies [19]. The prototype Rfv3 virus restrictive strain is C57BL, and BALB/c is the prototype non-restrictive strain. The Rfv3 gene map location on chromosome 15 [20] has now been linked to the locus of Apobec3 [10], [17]. That mA3 is responsible for Rfv3 resistance is supported by the observations that mA3 of C57BL restricts FrMLV replication and FrMLV-induced disease more effectively than BALB/c mA3, and that genetic inactivation of mA3 generates an FrMLV susceptible phenotype [10], [17]. It has also been shown that the C57BL mA3 allelic variant is more effective than the BALB/c allele in restricting MMTV [21].
The mA3 genes in prototype Rfv3 restrictive and nonrestrictive strains differ in protein sequence, splicing pattern, and expression level, and all three of these factors may contribute to resistance [10], [17], [21]. Few strains and Mus species have been characterized for these differences [21], so we sequenced mA3 genes from various inbred strains and wild mice representative of the major taxonomic groups of Mus. In this paper, we demonstrate that an MLV long terminal repeat (LTR) disrupts a splice donor site in the mA3 of C57BL and other strains and species and is associated with altered expression levels, we demonstrate strong positive selection of this gene in Mus that involves sites that distinguish the mA3 genes of Rfv3 virus resistant and susceptible mice, we use homology modeling to position the positively selected residues in two clusters on opposite sides of the putative active site groove, and we describe the antiviral activity of mA3 genes carrying mutations at these sites.
Analysis of the antiviral activities of chimeric and wild type C57BL and BALB/c mA3s by Takeda and colleagues [10] indicated that the mA3 anti-FrMLV activity resides in the N-terminal half of the C57BL protein. This 194 amino acid residue segment contains the active Z2-type cytidine deaminase region (CDA) [22], [23], and the translated protein sequences of restrictive C57BL and nonrestrictive BALB/c prototypes differ from one another in this region at nine residues [10]. To determine the distribution of the restrictive variant among mice and to identify novel variants, we sequenced segments of mA3 containing these 9 residues from inbred strains and wild-derived mice representing different taxa and/or mice trapped in different geographic locations (Table S1)(Figure 1A).
In the course of this analysis, we identified a 531 bp sequence inserted into the intron of mA3 of some laboratory strains between exons 2 and 3 (Figure 1A,1B). The insertion was sequenced and identified as an intact retroviral LTR (Figure 1C). This LTR is 96.6% identical to the LTR of the xenotropic gammaretrovirus (X-MLV) NZB-IU-6, an MLV isolated from NZB strain mice [24], [25]. The mA3 LTR insert shows the expected direct repeats characteristic of retroviral insertions, CAT and TATA boxes, and a comparable enhancer region. The LTR is inserted in an antisense orientation, and the site of insertion is the splice donor site at the end of exon 2 (Figure 1D). Part of the splice donor site contributes to the direct repeat flanking the insertion. The insertion alters the last base of the splice donor site, a position that is not highly constrained in the consensus sequence.
We screened 32 laboratory mouse strains for presence of this LTR insertion by PCR (Figure 2). The insertion was identified in 6 strains, including C57BL and the 3 related strains NZB, NZL and NZO. The LTR was absent from other NZB-related strains, from other strains in the C57/C58 series and was also absent from 21 strains from other families of inbred strains. The sequences of exons 2–4 of 13 strains were compared, and the only strains identified as having the C57BL/6 coding sequence, NZB and RF, also carried the LTR insertion. (Figure 2).
The common inbred strains of mice are a mosaic of Eastern European M. m. musculus, Western European M. m. domesticus and Asian M. m. castaneus [26], [27]. Therefore we looked for the sequence polymorphisms associated with the C57BL allele and for the MLV LTR in M. musculus subspecies from breeding stocks established from mice trapped in Old World sites where these commensal (house mouse) subspecies originated, and from M. musculus mice trapped in the Americas where they had been introduced from Europe and Asia (Figure 2). Two wild-derived mice from the Delmarva (Delaware-Maryland-Virginia) Peninsula, CL and LEWES, had this LTR along with the C57BL mA3 coding sequence. PCR fragments diagnostic of the LTR insert were also found in other Maryland mice as well as in two mice trapped in California, one of three M. m. castaneus breeding lines, and three of four lines developed from mice trapped in the former Czechoslovakia. The LTRs sequenced in 4 laboratory strains and 5 wild-derived mice were 99% identical to one another, and the mA3 genes of the LTR+ wild mice had several substitutions compared to the C57BL gene. Thus, the LTR was acquired in Eurasian species, and these LTR modified mA3 genes continued to accumulate mutations after this insertion event.
Previous reports had determined that mA3 mRNAs can lack exon 5 [10], [13], [14], and that BALB/c mA3 can also lack exon 2 [10]. We examined 31 mA3 mRNAs from cultured cells or tissues of 24 different inbred strains and wild-derived M. musculus mice for these splice variants by RT-PCR (Figure 3). mA3 mRNAs isolated from different tissues of the same mouse produced the same pattern of PCR products. Eleven of these 24 mice carry the LTR (Figure 3B), and all 11 mice produced a single PCR product of the size expected for a spliced message lacking exon 5 (Figure 3A). Among the 13 LTR-free mice, two, M. m. molossinus and the LTR− inbred MOLD/RkJ line of this subspecies, produced this same single isoform, while the other 11 LTR− mice additionally produced an exon5+ message that in 10 mice was significantly more abundant than the Δexon5 isoform (Figure 3A). Both sequenced BALB 3T3 mA3s lacked exons 2 and 5, and a third barely detectable smaller PCR product was observed in BALB 3T3 and other LTR− mice of the size consistent with the absence of exons 2 and 5 (Figure 3A). The distribution of the MLV LTR among these mice suggests that the LTR was inserted into the mA3 variant that produces the Δexon5 isoform.
Previous reports had noted that mA3 expression level is significantly higher in C57BL mouse tissues (LTR+) than in BALB/c (LTR−) [10], [21]. We isolated total RNA from the spleens of 11 mice that had been typed for the LTR and for mA3 splicing patterns. Included were mice from 2 breeding lines of M. m. molossinus, the inbred MOLD/RkJ strain and a mouse from a random bred colony, both of which are LTR− and produce the Δexon5 isoform (Figure 3B). Quantitative real-time PCR analysis showed that the 7 LTR+ mice produced 4–20 fold higher levels of mA3 mRNA than did the 4 LTR− mice, including the two M. m. molossinus mice (Figure 3C). These data demonstrate a correlation between the LTR and expression level but not splicing pattern.
We used sequenced segments of mA3 from 4 inbred strains and 21 wild-derived mouse species and subspecies for phylogenetic analysis. The sequences were used to construct phylogenies, and were analyzed with the PAML suite of programs [28] for evidence of adaptive evolution and to identify possible sites of positive selection. Two sets of DNA sequences were analyzed separately: exons 2–4 amplified from genomic DNA or RNA and a set of 8 near full length DNAs generated by RT-PCR (Text S1, S2). The sequences in the smaller dataset of 8 DNAs do not include the extreme 5′ and 3′ends of the gene or exon 5 which was absent from all but 3 of the 8 sequenced mRNAs.
The sequences were used to construct neighbor-joining trees (based on Kimura 2-parameter distances) for the near full-length sequences (Figure S1A) and for the 2–4 exon set (Figure 4A). Modifications to the trees were made based on generally accepted phylogenetic trees [29], [30]. The data-based and taxonomy-based trees were both used for PAML analysis and produced nearly identical statistics (Tables S2,S3). Values of dN/dS along each tree branch were calculated using the free-ratio model of PAML. A dN/dS value >1 suggests that positive selection has acted along that lineage. Several branches of the trees show evidence of positive selection with dN/dS>1, or, when dS = 0, by the identification of 4 or more replacement substitutions.
Likelihood ratio tests indicate that mA3 has a significant probability of having experienced positive selection, and this was the case for all codon frequency models, and for both datasets (Figures 4B and S1B, Tables S2, S3). The Bayes empirical Bayes calculation of posterior probabilities in PAML identified specific mA3 codon positions as having significant probability of positive selection. In the separate analyses of the two datasets, we identified 20 codons as being under positive selection with high posterior probability P>0.95, and 10 of these 20 codons were under very strong positive selection with P>0.99 (Tables 1, S2, S3). Sixteen of these 20 codons are in exons 2–4. Analysis of the smaller set of 8 near full-length genes identified a subset of the positively selected codons identified by analysis of exons 2–4. The full-length sequence analysis also identified 5 additional codons under positive selection with P>0.95 that were not identified in the exon 2–4 analysis: one codon, 142, in exon 3 of the active CDA and four codons, 201, 273, 316 and 371, in the inactive C-terminal CDA (Tables 1, S3).
There are 15 mA3 codons that specify different amino acids in virus restrictive C57BL and sensitive BALB/c mice. Eleven of these codons were found to be under positive selection (P>0.95), and 5 of the codons under very strong positive selection (P>0.99) mapped to two clusters in the active CDA (Figure 5). Because this type of analysis is designed to identify sites involved in diversifying selection (antagonistic interactions with pathogens being a prime example), our results indicate that most of the residues that distinguish C57BL and BALB/c mice identify key sites likely to be involved in genetic conflicts. These results also suggest that mA3 has had a defensive role that predates development of the laboratory strains and involves species in all 4 Mus subgenera.
Homology models for the C57BL N-terminal active CDA sequences were chosen from the LOMETS homology modeling program based on templates that had the highest sequence identity. The search identified several templates with highest confidence, crystal structures determined for the catalytic domain of hA3G (PDB ID 3IR2) [31] and (PDB ID 3IQS, 3E1U) [32]. The hA3G-3IR2 template model, based on the active hA3G C-terminal Z1 deaminase domain, was chosen for detailed analysis because it provides more coverage of the N-terminal Z2 domain of the mouse sequence [23], [31], and because it was the top LOMETS solution overall.
The C57BL mA3 CDA sequence has 36.4% identity to the hA3G CDA (Figure 6A). Superposition of the hA3G-3IR2 crystal structure and the mouse homology model show they share the 5 stranded β-sheet core surrounded by 6 α-helices that is common to known deaminase structures, along with a conservation of active-site loops involved in substrate binding (Figure 6B). The sidechain conformations of the C57BL residues involved in coordinating Zn are identical to their counterparts in the hA3G structure (Figure 6A,6B). The overall fold between the human and mouse structures is nearly the same with the RMSD (root mean square deviation) between backbone atoms of the C57BL mA3 model and the human structure being 0.56Å. The RMSD between all atoms for the mouse model and the human structure is 0.94Å.
Mutagenesis, NMR DNA titration data and structural analysis of hA3G-3E1U and the NMR structure hA3G-2JYW have identified key residues important in deaminase activity and formation of the substrate groove [32]–[34]. Among these key hA3G sites are the catalytic E259, 3 hydrophobic residues and 10 critical residues of which 9 are charged, all of which are within and brimming the groove and all of which are needed for deaminase activity (Figure 6A). N244 and R256 are associated with active center loop 3 (AC loop 3), R213 and R215 are present in active center loop 1 (AC loop 1), residue R313 resides on the floor of the groove and D316, D317, R320 face the substrate groove at or near the end of helix 4. The most obvious difference between hA3G-3IR2 and the mouse model in these functionally important sites is the presence of an 8 residue deletion in the AC loop 3 of the mouse model. hA3G AC loop 3 is an unstructured loop, and the deletion of the majority of the residues in the mouse AC loop 3 suggests they play no critical role; the mouse AC loop 3 structure, however, does conserve the two residues found at the hA3G loop base, N244 and R256, known to be critical for deamination [32], and it is likely that these mouse residues, N66 and I70, serve similar functions in mA3.
In contrast to this difference in AC loop 3, the functionally important AC loop 1 and helix 4 residues in hA3G are retained in mA3, and closely align with the two clusters of residues in mA3 shown here to be under positive selection (Figure 6A, 6B, 6C). On the other side of the substrate groove from selected AC loop 1 residues 34–38 is the region encompassing residues 134–139 in C57BL (and the corresponding region in hA3G); these residues are at the end of helix 4 with some residues participating in the α-helix and the rest as a loop. A solvent accessible surface representation of the mA3 structure indicates the position of the predicted substrate groove, and suggests the location of the two clusters of positively selected residues on opposite sides of this substrate groove (Figure 6D). The residues at the end of helix 4 and the residues in the 34–38 cluster on the other side of the mA3 groove likely serve steric roles in maintaining groove structure and likely also have functional roles based on charge and hydrophobicity that govern substrate interactions.
293T cells were cotransfected with the pLRB302 Friend virus clone and mA3 clones to assess the relative antiviral activities of 4 mA3 clones: the wild type Rfv3 virus resistant C57BL mA3 [13] and three clones with mutations that introduced residues of the Rfv3 virus sensitive BALB/c: M1 (G34R, K37I, G38D), M2 (V134I, Q135R, T139N), M3 (all 6 substitutions) (Figure 7). Cells and virus-containing supernatants were harvested 48 hours after transfection. Cells were analyzed by immunoblotting for mA3 expression, and infectious virus in the supernatants was quantitated by the XC overlay test. For each of the transfected mA3 clones, infectious virus titers decreased in a dose dependent manner relative to increasing expression of mA3 (data not shown). The wild type C57BL mA3 and the BALB-like M3 mutant both showed antiviral activity, but the antiviral activity of M3 was reduced relative to wild type mA3 (Figure 7). The M1 mutant mA3 was found to reduce the infectivity of Friend virus as effectively as wild type C57BL mA3, whereas M2 more closely resembled M3 in antiviral activity suggesting that substitutions in the 134–139 cluster are particularly important for anti-FrMLV activity.
This analysis indicates that mA3 has been involved in genetic conflicts through Mus evolution. This gene shows strong positive selection marked by an increase in replacement versus synonymous substitutions. Six of the 10 codons that evolved under strongest positive selection are in two clusters in the N-terminal catalytically active CDA. Five of these 6 codons specify different amino acids in MLV and MMTV restrictive and nonrestrictive mouse strains, and mutational analysis suggests these residues contribute to antiviral activity. We also demonstrate that the antiviral allelic variant has acquired a retroviral LTR insertion, the presence of which is associated with elevated mA3 expression levels in the spleens of inbred and wild-derived mice.
Retroviral insertions can be important functional components of the host genome, and can clearly affect host gene expression. Examination of spontaneous mutations in the mouse suggested that 10–12% of all mutations are due to ERV insertions [35]. Like the mA3 LTR, most of these mutant-associated ERVs are in reverse orientation in introns, and the responsible mutational mechanisms include two of relevance here: aberrant splicing and enhanced transcription driven by the ERV LTR. While the mA3 LTR is inserted at a splice donor site, it does not alter splicing of the associated intron, and although all mice carrying this LTR produce the same Δexon5 mA3 isoform, the absence of this LTR in at least one mouse species producing that isoform (M. m. molossinus) suggests that the LTR was acquired by mice already preferentially producing this splice variant. As for LTR-driven altered expression levels, two of three previous studies that compared mA3 RNA levels in virus-resistant and susceptible strains reported that mA3 expression levels are significantly higher in mice carrying the LTR+ C57BL allele compared to LTR− BALB/c [21], [17], [10]. Our analysis of mA3 expression levels shows a correlation between the presence of the LTR and elevated expression in a variety of inbred strains and mouse species. Because enhancer activation of cellular genes by viral LTRs can occur with insertions in either orientation and at considerable distance from the cellular promoter, it is thus possible that the enhancer of this inserted LTR sequence drives the elevated expression observed in the LTR+ mice. This elevated expression in conjunction with altered splicing may together have contributed to the evolution of the antiviral C57BL mA3. It has been suggested that the Δexon5 isoform has enhanced antiviral activity due to its resistance to the viral protease [36]; elevated expression of this variant due to subsequent LTR insertion would further boost the survival value of this factor.
It is particularly intriguing that this X-MLV LTR sequence is found in NZB and CZECH mice and one breeding line of M. m. castaneus. These mice are unusual among laboratory strains and wild mice in that they harbor highly active X-MLV ERVs producing infectious virus, and such active ERV expression increases the likelihood of insertional mutagenesis. NZB mice are characterized by lifelong viremia with X-MLVs [37]. M. m. castaneus and CZECH mice are among wild mouse Eurasian populations with highest copy number of X-MLV ERVs [38], and we have isolated infectious X-MLV-related virus from both of these wild mice [39], [40]. If in fact the inserted MLV LTR causes elevated mA3 expression, then this would provide another instance of an ERV sequence that is co-opted by the virus-infected host for an antiviral function, other examples in the mouse being Fv1, Fv4, and Rmcf [41].
In addition to differences in splicing and expression levels, mA3 genes of virus resistant and sensitive mice differ in protein sequence. Our phylogenetic analysis showed that most of these polymorphic sites are under strong positive selection. The alignment of these sites with functionally important residues in the hA3G C-terminal active CDA suggests they serve similar roles in the mouse and that therefore, this function has been important during Mus evolution. That this evolutionarily important function is related to mA3 deaminase activity is supported by the observation that the great majority of these selected residues are in the N-terminal half of mA3 which encodes the active Z2 CDA [22] and that antiviral activity resides in the first 194 amino acids (exons 1–4) [10]. In the predicted mA3 structure, these positively selected residues are positioned in one of two loops assigned functional importance in hA3G, AC loop 1 and a cluster of residues facing AC loop 1 on the other side of the putative substrate groove [32]–[34]. The charged and hydrophobic residues in these regions are positioned to maintain structural integrity of the groove and to interact with one another and the nucleic acid substrate in a way that could contribute to substrate specificity.
Three positively selected residues, G34, K37 and G38, in the mA3 AC loop 1 sequence KNLGYAKGRKD are most likely responsible for providing conformational freedom (in the case of the G34 and G38) and for interacting favorably with the phosphate backbone (in the case of K37). The electrostatic contributions of K37 along with K40 and D41 probably play an important role in determining substrate affinity and specificity while Y35 is in a position to stack with a nucleotide base. The analogous sequence in hA3G is NNEPWVRGRHE (207–217) with R213, H216 and E217 positioned to interact electrostatically with a phosphate backbone and W211 able to stack with a nucleotide base. R39 (mA3) and R215 (hA3G) are positioned similarly in that the residue provides an elaborate H-bonding network defining the shape of AC loop 1 [32].
Five positively selected residues (V134, Q135, D136, E138 and T139) lie in a region that comprises the end of helix 4 and an adjacent loop that define the side of the substrate binding groove opposite of AC loop 1 (mA3 sequence YNVQDPET). Close inspection of this region in the mouse model reveals that the sidechain of D136 is in a position to H-bond with T139 maintaining the helical nature of helix 4 despite the presence of P137. This has the result of allowing Q135 to form the top-side of the groove allowing V134, N133 and Y132 to form the side of the groove with Y132 in position to stack with a nucleotide base. Y132 is invariant in our mouse sequences along with nearby W102 which defines the floor of the groove. The homologous segment of human APOBECs has now been implicated in the distinctive substrate preferences among AID/APOBEC family members which target cytosine within different sequence motifs. A recognition loop responsible for these preferences (hA3G sequence IYDDQGRCQ) lies between the β4 strand and the α4 helix (Figure 6A, residues 314–322) [42]. That this highly variable region controls substrate preferences is also supported by mutational analysis [32], [43]. Alignment of the active CDAs of hA3G and mA3 indicates that this loop overlaps the 134–139 cluster of positively selected residues in mA3. This suggests that genetic conflicts between host and pathogen in this case produced positive selection that may be driven, not by protein-protein interactions, but by the interaction of mA3 and varying ssDNA substrates, a suggestion that is also consistent with the finding that the efficiency of substrate deamination is sensitive to ssDNA secondary structure [44].
Mutational analysis of 6 codons in the two clusters under positive selection showed that introduction of BALB/c residues, particularly in the 134–139 cluster, reduced antiviral activity against Friend MLV. Further studies may determine if the differences associated with overexpressed mA3 in transiently transfected cells have physiological relevance, and whether substitutions at these sites similarly affect restriction of other retroviruses. It has been reported that mA3 shows stronger antiviral activity against HIV-1 than against MLV [45], suggesting that the genetic conflicts responsible for positive selection during Mus evolution may have resulted from interactions with pathogens unrelated to the FrMLV used here.
Previous phylogenetic analysis of hA3G had identified 21 sites under very strong positive selection, 9 of which are in the active CDA [46]. One of these sites, R213, aligns with one of the clusters of residues (positions 34–38) under strong selection in mA3; however, the analysis of hA3G did not identify selection in the region aligning with the second cluster under strong selection in mA3 (positions 134–139), although this segment is a substrate recognition loop that is highly variable among members of the AID/APOBEC family [42]. The additional sites identified to be under positive selection in the hA3G active CDA have no positively selected counterparts in mA3. Among these additional sites in hA3G, two, H248 and K249, lie in AC loop 3 [46]. Mutagenesis and analysis of hA3G structure have implicated this loop in antiviral deamination [32], but much of AC loop 3 is deleted in the mouse, leaving only the key residues at the base of this loop that align with critical residues N244 and R256. The residues at these sites are invariant in our mA3 sequences suggesting their evolution is under purifying selection. The differences in AC loop 3 between hA3G and mA3 and the fact that different residues are under selection in hA3G and mA3 suggests there may be functional differences between these proteins.
Our analysis of the full-length mA3 sequences also identified four sites under positive selection in the C-terminal half of the protein (Tables 1, S1) that carries the Z3 CDA that has been determined to be inactive [22]. It is not clear what role these residues serve. An antiviral role for the C-terminal half of mA3 is suggested by the observation that that the conserved glutamates in the N-terminal Z2 domain and the C-terminal Z3 domain of mA3 are both required for antiviral activity against HIV-1 [45]. Other evidence suggests that the inactive CDA is involved in virus encapsidation [47]. We note that alignment of the mouse Z2 and Z3 CDA regions shows that one of the two selected Z3 codons, P316, aligns with the 134–139 selected cluster of codons in Z2, VQDPET. Another selected codon in the Z3 CDA, T273, aligns with an hA3G segment with two codons under selection in primates [46]. This suggests the possibility that this Z3 CDA may have had deaminase activity in some branches of the Mus lineage.
Further analysis of the C57BL and BALB/c mA3 genes should shed light on the functional roles of the polymorphic residues in the two groove-associated clusters. The information from additional phylogenetic, structural, and functional comparisons will help describe the range of antiviral activity and evolutionary history of this gene. We are currently analyzing additional mA3 mutants for antiviral activity, and using molecular dynamics simulations to describe the structural implications of specific substitutions.
DNA and RNA were isolated from animals and cell lines developed from laboratory mouse strains and from wild mice and wild mouse-derived breeding colonies (Table S1). Many wild-derived mice were obtained from M. Potter (NCI, Bethesda, MD). SAMP8 mice were provided by R. Carp (New York State Institute for Basic Research in Developmental Disabilities, Staten Island, NY), SIM.S mice were obtained from E. Boyce (Memorial Sloan-Kettering Cancer Center, NY), and mice trapped in California were provided by S. Rasheed (University of Southern California, Los Angeles). Mice or DNA samples of M. spretus (SPRET/EiJ), M. m. castaneus (CAST/EiJ), various inbred lines derived from M. m. molossinus, PERA, PERC, PWD, and the inbred strains listed in Figure 2 were obtained from The Jackson Laboratory (Bar Harbor, ME).
A set of African pygmy mouse DNA samples was obtained from Y. Cole and P. D'Eustachio (Depts. Biochemistry and Medicine, NYU, New York); these mice had been classed into 4 species of subgenus Nannomys mice on the basis of skeletal features by J. T. Marshall (Smithsonian Natural History Museum, Washington, DC). A sample of M. m. macedonicus DNA was provided by R. Elliott (Roswell Park, Buffalo). Cell lines used as DNA and RNA sources included NZB-Q and M. fragilicauda cells obtained from J. Hartley (NIAID, Bethesda, MD), cells from some wild mouse species obtained from J. Rodgers (Baylor College of Medicine, Houston, TX), and NIH 3T3, M. dunni [48], SC-1 [49], A9 (C3H/He) [50], and CMT93 (C57BL) (ATCC CCL-223).
All studies in which animals are involved were performed in accordance with the guidelines of the Committee on the Care and Use of Laboratory Animals under an NIAID-approved animal study protocol [51], and all studies and procedures were reviewed and approved by the Institutional Animal Care and Use Committee of the NIH.
APOBEC3 segments were amplified from mouse genomic DNAs or RNAs using primers designed from coding, flanking or intron sequences based on the C57BL genomic sequence (GenBank No. NT_03921) (Figure 1A). Exon 2 was amplified using forward intron primer a: 5′-CTCCTCTCCCTCTGTCTTCCT and reverse primer b: 5′-GGATTCAAGGTATGAGCCACCATGC. Exons 3 and 4 were amplified using primer c: 5′-GCTTCAACAGGGCTCAGAGTGC and primer d: 5′-AGGTTTGGGAGGAGGGAGAAC. Reverse transcription PCR (RT-PCR) was used to amplify near full-length APOBEC3 from total RNA using primer e in exon 1 (5′-GGACCATTCTGTCTGGGATGCAGCCATCG) and primer f in exon 9 (5′-GACATCGGGGGACCAAGCTGTAGGTTTCC) and a shorter RT-PCR fragment was generated using primer a and primer g (5′-GGTTGTAAAACTGCGAGTAAAATTCC). The larger RT-PCR product contained 1083 bp of the full-length 1287 bp mA3 sequence. Most of these products lacked the 99 bp exon 5, and the aligned sequences lack 72 bp at the 5′ end and 33 bp at the 3′ end of the gene. PCR products were sequenced directly in some cases, and in others fragments were first cloned into pCR2.1-TOPO (Invitrogen, Carlsbad, CA) before sequencing (Text S1, S2).
Total RNAs from mouse spleens were isolated using Trizol (Invitrogen). Reverse transcription was carried out at 50°C for 1 hour using 2 µg of total RNA in the presence of Oligo (dT) primer (Ambion, Austin, TX) and SuperScript III (Invitrogen). After reverse transcription, the reaction mixtures were diluted to 1000 µl with DEPC-water. 1 µl of the diluted cDNA were added to a 15 µl PCR reaction mix containing 0.4 µl of 10 µM primers and 2× SYBR Green PCR mix (Applied Biosystems, Foster City, CA ). APOBEC3 transcripts were amplified using primers 5′-GACCATTCTGTCTGGGATGCA and 5′-TTCTAGTCACTTCATAGCACA. β-actin was also measured using primers (5′- GTGGGGCGCCCCAGGCACCA; 5′- CTCCTTAATGTCACGCACGATTTC) as a normalization control. Amplification was done under the condition of 15 s at 95°C and 1 min at 60°C for 50 cycles in a 7300 Real Time PCR System (Applied Biosystems).
HA-tagged mA3 [13] was obtained from the NIH AIDS Research and Reference Reagent Program (Germantown, MD) (catalog no. 10021) and mutagenized using the QuikChange mutagenesis kit (Stratagene, La Jolla, CA) to introduce substitutions at 6 codons. M1 (G34R, K37I, G38D) was generated using primer 5′-CCACTTTAAGAACCTACGCTATGCCATTGATCGGAAAGATACCTTC and its reverse complement. M2 (V134I, Q135R, T139N) was generated using primer 5′-GCTCCCGCCTCTACAACATCCGAGACCCAGAAAATCAGCAGAATCTTTGC and its reverse complement. M3, containing mutations at all 6 codons, was generated by mutating M1 with the primers designed for M2. Mutations were confirmed by sequencing. Attempts to generate stable transfectants of various mouse cells expressing these mA3 variants were not successful. Human 293T cells were co-transfected with 3–4 µg of the pLRB302 clone of Friend MLV [52] obtained from L. Evans (RML, NIAID, Hamilton, MT), and 0.5 or 1.0 µg mA3. At 48 hours after transfection, the culture supernatant was collected and virus infectivity was measured by the XC overlay test [53]. In this test, subconfluent cultures of NIH 3T3 cells were infected with virus dilutions, irradiated 4 days later and overlaid with rat XC cells. Infectivity was determined as plaque-forming units per ml of culture fluid. Infectivity was normalized against reverse transcriptase activity [54] or virus-associated capsid protein in pelleted virus. After electrophoresis on 12.5% SDS-polyacrylamide genes and transfer to polyvinylidene difluoride membranes, capsid protein was detected using polyclonal goat anti-Rauscher MLV p30 antiserum (Viromed Biosafety Laboratories (NCI/BCB Repository), Camden, NJ) and horseradish peroxidase conjugated rabbit anti-goat antibody (Invitrogen catalog # R21459). The transfected 293T cells were lysed and tested for mA3 expression by western immunoblot analysis. Cell lysates were subjected to electrophoresis and western blots were probed with a monoclonal antibody against HA, HA-7 (Sigma catalog #H-3663) and a monoclonal anti-tubulin antibody (Sigma #T-9026).
DNA sequences were aligned using MUSCLE [55] and improved manually. Two phylogenies were produced, one for the full-length sequences and one for the exon 2–4 sequences. In all cases the Kimura 2-parameter distance-based neighbor-joining phylogenies for each set returned by PHYLIP (version 3.68) [56]) were corrected for closer correspondence to the consensus Mus phylogeny [29], [30]. The trees were corrected to make the Nannomys species a monophyletic group and to place M. spretus basal to the M. musculus node.
The codeml program of the PAML4 package [28] was used for maximum likelihood analysis of codon evolution [57]. Both lineage-specific and codon-specific analyses were performed. In the lineage-specific selection analyses, the free ratio model (codon model = 1) was used to calculate branch-specific rates of dN/dS. In this model each branch is assumed to have a specific dN/dS ratio. The likelihood of the phylogeny under this model was tested against the likelihood of the phylogeny under the model of one uniform dN/dS ratio across all branches (codon model 0) using a likelihood ratio test (LRT). The significance of the LRT value was assessed using a chi-squared distribution with 49 degrees of freedom for the exon 2–4 sequence analysis and 12 degrees of freedom for the full-length sequence analysis.
Selection acting on Apobec3 codons was analyzed using two models of equilibrium codon frequencies and four models of codon selection. The two codon frequency models used were the F3x4 model (codon frequencies estimated from the nucleotide frequencies in the data at each codon site) and the F61 (Codon Table) model (frequencies of each of the 61 non-stop codons estimated from the data). The codon selection models were two neutral/negative selection models (M1 and M7) which were compared against corresponding positive selection models which included a category for dN/dS>1 (M2 and M8, respectively). The significance of this additional codon selection category was assessed using LRTs of the phylogeny likelihoods under the neutral and positive selection models. Significance of the test statistics was calculated using a chi-squared distribution with two degrees of freedom. The Bayes empirical Bayes algorithm [58] was used to calculate the posterior probability of individual codons experiencing dN/dS>1.
The C57BL mouse mA3 sequence (GenBank No. NM_030255) was submitted to the LOMETS program [59]. A model constructed using a template with the highest sequence identity was chosen from the top ten solutions ranked by a combination of highest sequence identity, most coverage, Z-score and overall confidence. The model was generated using Modeller v4 [60] and energy optimized in SYBYL7.3 using the AMBER7 ff99 forcefield with AMBER7 ff99 atom types and charges with the Powell method to a termination gradient of 0.05 kcal/mol·Å. The model was examined using Procheck [61] to detect any bad geometries.
mA3 exon 2–4 sequences were given GenBank Accession Nos. GQ901957–GQ901974. Near full length sequences were given GenBank Nos. GQ871500–506.
|
10.1371/journal.pntd.0002009 | Using Geographical Information Systems to Identify Populations in Need of Improved Accessibility to Antivenom Treatment for Snakebite Envenoming in Costa Rica | Snakebite accidents are an important health problem in rural areas of tropical countries worldwide, including Costa Rica, where most bites are caused by the pit-viper Bothrops asper. The treatment of these potentially fatal accidents is based on the timely administration of specific antivenom. In many regions of the world, insufficient health care systems and lack of antivenom in remote and poor areas where snakebites are common, means that efficient treatment is unavailable for many snakebite victims, leading to unnecessary mortality and morbidity. In this study, geographical information systems (GIS) were used to identify populations in Costa Rica with a need of improved access to antivenom treatment: those living in areas with a high risk of snakebites and long time to reach antivenom treatment.
Populations living in areas with high risk of snakebites were identified using two approaches: one based on the district-level reported incidence, and another based on mapping environmental factors favoring B. asper presence. Time to reach treatment using ambulance was estimated using cost surface analysis, thereby enabling adjustment of transportation speed by road availability and quality, topography and land use. By mapping populations in high risk of snakebites and the estimated time to treatment, populations with need of improved treatment access were identified.
This study demonstrates the usefulness of GIS for improving treatment of snakebites. By mapping reported incidence, risk factors, location of existing treatment resources, and the time estimated to reach these for at-risk populations, rational allocation of treatment resources is facilitated.
| Snakebite accidents are a neglected cause of much death and suffering worldwide. The situation is especially severe in rural areas of low income tropical countries, where long distances to well-equipped health care facilities mean that the time needed to reach antivenom treatment is often long. Delays in receiving antivenom treatment of snakebites could lead to severe outcomes, such as death or permanent disability. In this study we demonstrate how Geographical Information Systems (GIS) could be used to allocate antivenom rationally and thereby decrease the impact of snakebite in a cost-effective manner. We map areas with a high risk of snakebite accidents, based on a high reported incidence and environmental conditions favoring snakebites. We then estimate the time needed to reach a facility in which antivenom treatment is available for the population in these high risk areas, thus identifying areas in need of improved treatment accessibility. Based on these maps of the unmet need of antivenom treatment, allocation of antivenom and other resources needed to treat snakebites can be made more efficiently.
| Snakebites are a health problem in several parts of the world [1], especially affecting poor people living in rural areas in tropical countries [2]. The treatment of snakebite envenoming is based on the timely administration of animal-derived antivenoms [3]. However, a number of factors limit the accessibility of antivenoms in various areas of the world [3], [4]. The relatively high cost of some antivenoms make accessing them difficult in low income countries [4]. In addition, long distances to healthcare facilities and incorrect distribution of antivenoms within countries [3], [5], [6] means that for many of the snakebite victims worldwide, specific treatment with safe and effective antivenoms is unavailable, and traditional healers are instead often consulted [1], [7]–[9]. This is an unfortunate situation as antivenom treatment is highly effective at preventing morbidity and mortality caused by snakebite envenoming [3]. In other regions of the world, such as in various countries in Latin America, antivenom is readily available in health centers [5]. This is the case of Costa Rica, where antivenom is widely available throughout the public health system, as a result of domestic production and effective acquisition and distribution schemes [5].
The incidence of snakebites is known to vary widely on a sub-national level due to environmental and demographic factors [10]–[14]. In Central America, an estimated number of 4,000 snakebite cases occur every year [15]. The Central America lancehead pitviper Bothrops asper, locally known as ‘terciopelo’ or ‘barba amarilla’, which is widely distributed in humid, lowland areas (0–600 m.a.s.l., but it might be found up to 1,200 m.a.s.l.) [16], [17], is responsible for 50–80% of the bites and 60–90% of all fatal cases in Central America and northern South America [13]. Epidemiological studies indicate that most victims are agricultural workers and/or rural residents in general [13]. In the 1990s, the snakebite incidence in Costa Rica was approximately 15 per 100,000 inhabitants per year, with a decreasing trend [14]. Snakebite mortality has also been reported to decrease in Costa Rica, an observation attributed to, among other things, improvements in the health care system, including better geographical accessibility to health care facilities [12], [18]. The use of first aid interventions and the attendance to traditional healers for treatment of snakebites is very limited in Costa Rica, at least among those seeking hospital treatment [19].
Identification of areas with high snakebite incidence is an important goal for the design of preventive and therapeutic interventions aimed at reducing the impact of snakebites. Such interventions could include educational programs, antivenom distribution and training of health staff in the management of these envenomations [4]–[6]. Similar issues of rational distribution of treatment and intervention resources are being considered in the struggle against other neglected tropical diseases, with geographical information systems (GIS) increasingly being used as a tool for improved decision-making [20]. Leynaud and Reati [21] described the use of the free GIS-software SIGEpi [22] for identification of areas with a high risk of snakebites and long distances to healthcare in Argentina. It is necessary to further extend the use of this methodology in other countries, in order to get an integrated view of snakebite envenomings on various regions of the world.
The health care system in Costa Rica is considered well developed, in terms of geographical accessibility [23] and insurance coverage [24]. This country has made important achievements in public health, such as long life expectancy and low infant mortality rate [24], [25]. Antivenom is available in all clinics and hospitals and, according to a recent decision by the state social insurance administration that runs all public health services in the country (Caja Costarricense del Seguro Social, henceforth CCSS), it can also be distributed to primary health care teams (Equipos Básicos de Atención Integral en Salud, EBAIS) of which there are approximately 1 per 5,000 inhabitants [24]. This opens the possibility to further improve the accessibility to antivenoms in Costa Rica. However, the distribution of antivenoms to EBAIS has to be carefully analyzed, in order to ensure that mostly EBAIS serving a population with high snakebite risk and limited access to clinics and hospitals will receive antivenoms. This will prevent unnecessary wastage of this precious drug by deploying it to regions where snakebites are infrequent, or where distances to hospitals or clinics where these accidents are treated are short.
The primary aim of the present study is to provide information to assist decision-making concerning for which primary health care facilities (EBAIS) it is suitable to have antivenom, i.e. those that serve a population with a high risk of snakebites and long transport times to hospitals or clinics where antivenom is available. The methodology used in this study might be applied in other countries and regions where snakebite envenoming is a relevant public health hazard, in order to identify vulnerable areas that require interventions aimed at ensuring the access to antivenom and adequate medical treatment. As part of the primary aim, we will demonstrate the benefits of spatial smoothing of the small-area snakebite incidence data for increasing the interpretability of this data. Aside from the primary aim, we will also describe the relationship between some district-level environmental and demographic factors and snakebite incidence in Costa Rica.
Populations in a high risk of snakebite were identified using two approaches: one based on the reported district-level snakebite incidence, and another based on identifying populations with environmental and demographic risk factors favoring snakebites.
For the first approach, an incidence of 30 bites per 100,000 population per year was selected as threshold, on the basis of the overall incidence of snakebite envenomings in Costa Rica (15 cases per 100,000 inhabitants per year [14]). In order to reduce the random noise in the small-area incidence, a Bayesian smoothing method [26] was employed by fitting a Poisson regression to the observed incidence in the period 2003–2007. Smoothing can be explained as modelling the underlying risk around which the observed incidence varies stochastically. As such the smoothed incidence estimates give a more stable picture of the actual risk of snakebites than the raw incidence data, making interpretation easier and giving a better basis for decision-making [26]. Spatial smoothing was also used by Leynaud and Reati [21] in their study of snakebite epidemiology in Argentina. In this study, we present an alternative smoothing method that includes also measured explanatory variables, and allows for estimation of the probability that an incidence threshold is exceeded [27]. We then compare the performance of this smoothing method with the one used by Leynaud and Reati [21]. Measured variables to be included in the smoothing were identified in the literature as important factors influencing snake presence, i.e. environmental factors [16], and snake-human interaction, in this case urban population percentage [13].
For the second approach to identifying areas in possible need of improved antivenom accessibility, rural areas in humid, lowland areas were considered at high risk of snakebites, as this is a habitat suitable for the medically most relevant snake species B. asper [16].
Among the populations identified as in high risk of snakebites by either of these two approaches, those with an estimated minimum transportation time to hospital or clinic exceeding 2 hours were identified as those requiring improved antivenom accessibility, based on the time identified as critical to avoid mortality from snakebites [13].
The centroid of the census enumeration units were used to estimate the location of the population [23]. A 2 km buffer around the census tract centroids were used as a proxy for populated areas. Census enumeration unit population size in year 2000 and information whether the unit was urban or rural was available as attribute information.
Information about snakebite cases 1990–2007 per district of residence were obtained from hospital discharge reports from the CCSS. Duplicate cases, with exactly the same age, sex, month of bite and district of residence (n = 123) were removed from the analysis. Cases reported from health care facilities in which it was highly unlikely that the place of the bite was in the district of residence (n = 141), or when the place of residence was unknown (n = 43) were also removed. These steps left 9,149 snakebites, divided over 413 districts (the smallest administrative area) (Figure 1). The location of B. asper specimens (n = 241) collected to the serpentarium at Instituto Clodomiro Picado, San José, Costa Rica was obtained.
Digital maps of environmental variables (elevation 30×30 m raster, mean annual precipitation, number of dry months, forest coverage, and biotic unit [28]) were obtained from a database compiled by Instituto Tecnológico de Costa Rica [29]. The values for each of the environmental variables were extracted to the census enumeration unit centroids, and the population-weighted district average value for each of the environmental variables was calculated. For forest coverage, the proportion of inhabitants living within 500 meters from forests larger than 5 ha was calculated. A few districts (n = 9) were covered by clouds in the satellite image from which the forest coverage was derived, and these were assigned the mean forest coverage value of the other districts.
The location of lakes, rivers, clinics, hospitals and roads were obtained from the same database as the environmental variables [29]. A list of primary health care facilities, i.e. Equipos Básicos de Atención Integral en Salud (EBAIS) was obtained from the CCSS [30]. EBAIS were geocoded by matching the name and service area of the facility with the name and district of communities in the country, available as a digital map [29]. In most cases, there was a community with the same name as the facility. If there was no such community, the facility was located in the main community of the district. If it was not possible to determine which the main community was, the facility was located in one of the communities located in the center of the service area. Red Cross ambulance stations were geo-coded using a list and map available at the Costa Rican Red Cross Website [31].
In order to analyze which factors were important for snakebite occurrence, and to smooth the snakebite incidence, a Bayesian Poisson regression was used to model the district-level risk of snakebites. Factors chosen for inclusion were those that had previously been identified in literature as either influencing snake presence [16] or important for snake-human interaction patterns [13]. The number of snakebite cases Y per district i were modeled as Poisson variates in the form;where is the intercept, a matrix of five fixed effect district-level explanatory variables (urbanity, forest coverage, elevation, precipitation, and number of dry months), a spatially correlated random effect modelled using a conditional autoregressive (CAR) prior structure [32], which assumed dependence between districts if they shared a border or corner, and correspond to the number of snakebites that would be expected if they were distributed evenly within the population, i.e. the offset. In order to facilitate convergence, continuous variable (precipitation, number of dry months and elevation) were standardized to have mean 0 and standard deviation 1. The coefficients for the fixed effects were assigned non-informative normal distribution priors (mean 0 and precision 0.0001), and the intercept a non-informative flat prior (range −∞ to ∞). The variance of the spatially correlated random effects was assigned a non-informative gamma prior. The model was fitted in WinBUGS 1.4.3 [33].
When the district-level risk factors of snakebites were analyzed, all snakebite data (i.e. 1990–2007) were used. Convergence was reached after 50 000 iterations, after which another 100 000 iterations were performed to estimate the posterior distribution, from which model parameters with 95% credible intervals (Cr.I.) were obtained.
The snakebite incidence dataset was divided into nine temporal periods, six training periods on which the smoothing was performed, and three test periods used to assess the ability of the estimates produced by the smoothing methods to improve identification of future high-incidence districts. Three of the training periods were five-year periods (1990–1994, 1994–1998 and 1998–2002) and three one-year periods (1994, 1998 and 2002).The three test periods were five years (1995–1999, 1999–2003 and 2003–2007).
The first smoothing method (A) was the same Bayesian model with random spatial effects and fixed effect explanatory variables (elevation, number of dry months, precipitation, forest coverage and urban population percentage) as described above. The Bayesian framework makes it possible to estimate the probability that the incidence exceeds a threshold directly from the posterior distribution. For each of the training time periods, the models were run until convergence was reached (after 50,000 iterations) and then for another 50,000 iterations to obtain samples from the posterior distribution. From the posterior distribution, the probability that the incidence exceeded 30 bites per 100,000 inhabitants was calculated.
The second smoothing method (B) was the one used by Leynaud and Reati [21]; the tool for automated spatial Bayesian smoothing of incidence rates (“Suavizador espacial de tasas”) available in SIGEpi [22]. The settings of local smoothing and neighborhood defined as adjacency were used.
The abilities of these smoothed estimates, and the unsmoothed incidence, to identify whether the incidence in a district would exceed 30 bites per 100,000 in the next-coming five-year period (i.e. the test periods 1995–1999, 1999–2003 and 2003–2007) were assessed using the ROC Curve function of IBM SPSS 20 [34]. The AUC (area under the curve) of the ROC (receiver operating characteristics; a plot of sensitivity vs. one minus specificity) is an often used tool to assess the discriminatory ability of tests; an AUC of 0.7–0.9 indicate reasonable discriminatory ability, and >0.9 very good discriminatory ability [35]. For each of the six combinations of training time (1 and 5 years) and smoothing method (A, B and unsmoothed), the mean AUC (with 95% empirical confidence intervals (C.I.)) were computed by simulating plausible AUC values from the uncertainty interval for the three training-test period pairs and calculating the mean of these.
In order to produce smoothed estimates of the underlying snakebite risk in a district, that would be less affected by random noise, and thereby able to more precisely identify the need for antivenom accessibility in that district in the upcoming years, the above model was applied to the most recently available 5-year-period of the snakebite data; i.e. 2003–2007. After reaching convergence after 50,000 iterations, the model was run for another 50,000 iterations, during which the probability that the smoothed incidence exceeded the threshold of 30 bites per 100,000 inhabitants was calculated (Figure 2).
The incidence threshold exceedance probability estimate that corresponded to 90% sensitivity in detecting districts with an incidence above the 30 bites per 100,000 inhabitants threshold in a future 5-year period was on average 10% for three earlier 5-year periods (data not shown). Therefore, this was set as the cut-off probability for identifying high-risk districts, in need of good antivenom accessibility. Rural population residing in an environment suitable for the medically most relevant snake species B. asper (i.e. below 1200 meters of elevation and in Moist, Wet or Pluvial biotic unit [16], [28], [29]) (Figure 3) were also identified as living in high-risk areas (Figure 4).
The mountainous terrain of Costa Rica imposes strong restrictions on human movement. The Euclidean distance (straight line) approximation of the time needed to move from the place of snakebite to the healthcare facility (hospital or clinic) might therefore not accurately capture the real time spent in this transportation. We aimed at constructing a model of the time needed to reach antivenom treatment, i.e. hospitals or clinics, which takes into account the availability and quality of roads, topography, land cover and the mode of transportation used to reach healthcare. We assumed that people would choose ambulance services after a snakebite accident. Travel time was therefore calculated as time with ambulance from closest ambulance station, to place of residence, and from there to closest hospital or clinic. The time estimated should be considered a minimum, ideal, time as it is modeled assuming that the ambulance leaves immediately after the snakebite to meet up the snakebite victim, without any delays.
The time to treatment was estimated using GRASS 6.4.1 [36]. The road data vector layer was converted to a 30×30 m raster and classified according to the following assumed speeds: primary roads, 60 km/h; secondary and urban roads, 40 km/h; and tertiary, local and other roads, 20 km/h. Off-road speed was set to 6 km/h, but the off-road speed will become 3 km/h as this distance is not covered by ambulance transportation and it will be counted twice (see below). Off-road speed in areas covered by forest was assumed to be 50% slower. Streams, rivers and lakes, that were not crossed by roads, were given a speed one fifth of that in open terrain, in order to penalize movement across water. From the elevation raster, a slope raster was constructed. The road/off-road speed raster, and the slope raster were combined to yield a raster of the time needed to travel one raster cell (tr,α) using Equation 1.
(Equation 1)The first part () of Equation 1 calculates the excess distance needed to travel the cell due to change in altitude (i.e. the hypotenuse). The second part (of the equation aims at taking into account the reduced speed associated with moving in undulated terrain, as well as to create barriers in extremely steep slopes. In this part of the equation, we assume that the excess time needed to travel one of the 30 m cells is proportional to the squared slope (in degrees) multiplied by a coefficient βr (0.001 for road travel and 0.02 for off-road). The value of this co-efficient was chosen based on simulation of what values produced reasonable estimations.
The time needed to go from an ambulance station to any cell, and the time to go from a hospital or clinic to any cell was calculated. These two time rasters were then summed to give a raster of the total time needed to reach healthcare facilities using ambulance, i.e. from ambulance station to snakebite victim and from there to healthcare (Figure 5). The mean time to reach healthcare (Figure 5) from populated areas in a high risk of snakebites (Figure 4) was extracted, and categorized into <2 h, 2–3 h and >3 h, in order to visualize the need of improved antivenom accessibility (Figure 6). Finally, the location of EBAIS and roads were added to the map of populations in need of improved antivenom accessibility, to develop small-scale maps of areas in need of improved accessibility to antivenom, which can be used to guide such improvements (Figure 7 and Supporting Information S1).
The non-spatial descriptive statistics of major parts of our dataset correspond very closely to what has been described in detail previously [14], and thus only a brief summary is presented in this article. A total of 9,333 cases were reported for the period 1990–2007, corresponding to an average incidence of 13.8 snakebites per 100,000 inhabitants per year. Seventy-two percent of the victims were male. Regarding age, 27% of victims were below 15 years of age, 32% between 15 and 30 years, 21% between 30 and 45 years, and the remaining 22% corresponded to people older than 45 years.
In multivariate analysis, lowland districts with much precipitation and few dry months generally had a higher snakebite incidence, as did districts with many rural inhabitants, and many inhabitants residing close to larger forests, although the last explanatory factor did not reach 95% significance (Table 1).
The unsmoothed incidence observed in a one-year period had an acceptable ability to identify which districts would have a high snakebite incidence in the next five-year period (Table 2), but discriminatory ability was clearly improved by using smoothing. There were only minor differences between the estimates produced by the different smoothing methods A and B in discriminatory performance. When the training time was only one the incidence threshold exceedance probability of method A had a tendency to be better than the smoothed incidence produced by method B (p = 0.07). When the training time was five years, both methods had a borderline significantly better discriminatory ability than the unsmoothed incidence observed in five years (p = 0.02 for method A and p = 0.10 for method B).
There were marked geographical differences in the reported incidence, with high rates especially along the southern Pacific coast, and in parts of the Caribbean and northern lowlands (Figure 1). The districts with a high incidence corresponded to a large extent to the humid lowland areas identified as suitable habitat for B. asper (Figure 4); 91% of the rural population in high-incidence districts lived in this type of environment. However, only 51% of the rural population in humid lowland areas lived in districts with a high snakebite incidence. According to the population distribution of the 2000 census, 18% of the total population lived in districts with a high snakebite incidence and an additional 13% lived in rural, humid lowlands in districts not reporting an incidence above the threshold.
Hospitals, clinics, ambulance stations and roads were strongly aggregated in the great metropolitan area around the cities of San José, Heredia, Cartago and Alajuela, where a major part of the population lives. There is however also a network of roads and healthcare facilities in the more peripheral regions of the country (Figure 5). Of the population living in areas with a high risk of snakebites, 92.5% were estimated to have a minimum transportation time of less than 2 hour to hospitals or clinics, 5% were estimated to delay 2–3 hours to hospitals or clinics, and 2.5% more have transportation times higher than 3 hours. On the south Pacific coast around Golfo Dulce, around the Talamanca highlands in the southeast, and along the northern border, there are populations in high risk of snakebites and with long transportation times to antivenom treatment (Figure 6). Figure 7 provides a close-up view of one target region (Golfo Dulce), including all the information presented in the previous maps, as well as the location of primary health care facilities (EBAIS). This map demonstrates the type of map that can be prepared and used to identify vulnerable places where access to hospitals or clinics is delayed, thus setting the stage for the design of more effective ways to guarantee a more rapid access to antivenom treatment at the local level. Similar maps for other target regions are available as a supplement (Supporting Information S1).
The spatial distribution of district-level snakebite incidence in Costa Rica largely followed the expected pattern, based on previous studies and on the distribution of the most important venomous snake in the country, B. asper. Incidence was higher in rural, humid lowlands, notably in the southern part of the country [14]. Geographical accessibility to antivenom treatment was generally good, however, in some areas there is a need of improved treatment accessibility. These are areas where hospitals and clinics are located relatively far from some of the areas with a likely high snakebite incidence.
The potential bias introduced by using data reported from the healthcare system to analyze snakebite incidence, due to the use of traditional medicine and dysfunctional reporting routines, are well known from other parts of the world [9], [37]. However, in contrast to the situation in many other developing countries in Latin America and elsewhere, traditional medicine is not widely used in Costa Rica [38] and the formal health care system is well developed and largely accessible [23]. The percentage of births attended by skilled personnel was 98.7% in 2007 [39], highlighting a highly developed and utilized formal healthcare system. A survey among snakebite victims receiving treatment at hospitals in 1996 found that only 2.9% of them had received any type of empirical treatment before reaching formal healthcare attention [19]. The literacy rate in Costa Rica is 96% [39], and educational campaigns of various sorts over several decades have raised awareness about the importance of seeking formal healthcare after snakebites among the Costa Rican public [19]. Based on these facts, we assume that the degree of utilization of traditional healthcare after snakebites is very low in Costa Rica and that, on this basis, there will not be much underestimation of the true snakebite incidence in the statistics available from the Ministry of Health. Nevertheless, our data may still suffer from under- or misreporting because of reporting errors, such as missing discharge reports.
Assuming equal distribution of snakebite risk within districts is a strong assumption. Even if districts are the smallest administrative unit, several spatial processes could still lead to large incidence variations within districts. One specific example of where problems are likely to arise due to within-district variation is in districts with high proportion of urban residents. Even though snakebite incidence might be high among the rural population of such districts, this might go unnoticed due to the large urban population among which there are few snakebites. By mapping census enumeration units with risk factors favoring snakebite occurrence, the impact of the above two limitations can be reduced as this risk-factor based approach is not dependent on the quality of the gathered incidence data or arbitrary district divisions.
The information about the location of the health care system must be regarded with caution; previous centrally available information about primary health care facilities has been found incorrect in an earlier study [23]. Furthermore, the method of locating primary health care facilities (EBAIS) by matching facility and community names is not infallible as there was not always a community with the same name as the EBAIS. However, this was mostly a problem in the urban areas, and thus of smaller importance for this study. The data about roads was generally old, and as there has likely been some improvement in road availability and quality since the data was gathered, our estimates of the time needed to reach treatment might be biased towards overestimation.
The estimated transportation times to reach hospital or clinic were much lower than those observed in previous studies of the time to hospital treatment of snakebites in Costa Rica. In a hospital-based study of all snakebites in 1996 [19], the time to reach hospital was recorded for approximately 70% of the patients. Of these, 61% reached hospital within 3 hours and 20% after more than 5 hours. However, the estimates cannot be fully compared with the transportation times observed in this study as they were recorded at hospitals, meaning that a major proportion of the patients could have received antivenom treatment at clinics and subsequently been transferred to a hospital, something that would delay the time to reach hospital substantially. Saborio et al. [40] found that among children admitted to the hospital in Limon on the Caribbean coast in 1985–1995, 50% received medical treatment within 3 hours, whereas the mean time was 6.8 hours, indicating very long transportation times for some snakebite victims in this area, parts of which are also estimated to have long times to treatment by our model. Another reason why it could be incorrect to compare the estimated times with the time observed in these studies is that they are at least 15 years old, and there have been improvements in ambulance and health care facility accessibility, telecommunications, and possibly road network since then. It should however be further emphasized that our model aim at estimating the ideal time to reach treatment, and that in reality there could be several unaccounted-for causes of longer times, such as problems in communicating with ambulance stations, temporarily impassable roads or unavailable ambulance services, etc. Even though there is a discrepancy between the estimated and observed times to treatment, we consider that the time-to-treatment estimation model provides important hints about the location of areas where the accessibility to antivenom treatment is more difficult. If the minimum time to treatment is estimated to be 2 hours in our analysis, there is an imperative for improved accessibility as the actual time to treatment will probably be longer.
It is however important to remember that healthcare accessibility cannot be reduced to a purely spatial concern. Logistical issues, such as effective communication with ambulance facilities, availability of ambulances, and problems with other forms of transportation in the communities need to be taken into account as well. Furthermore, geographical accessibility is just one dimension of health care access [41]; economic, social and cultural dimensions need to be also considered, something that is easily missed when doing analyses based on maps only. Research gathering empirical evidence on the actual, current time needed to reach treatment, and determinants of this time, would provide important information for the identification of vulnerable regions and for improving access to snakebite treatment in Costa Rica.
The estimated minimum times to reach antivenom treatment were generally short, compared to the actual times to reach health facilities after snakebites reported in previous studies in Costa Rica [19], [40]. However, our analysis allowed the identification of some areas where accessibility to antivenom treatment needs to be improved. The specific strategies to be implemented to accomplish this demand a case by case analysis on a local basis, but a feasible alternative might be the distribution of antivenoms to some EBAIS, the strengthening of the training of health staff in antivenom use, and the organization of the work in such a way that antivenom is available at all times. There is a risk of over-interpreting the messages transmitted through these maps and forget how sensitive it is to data errors and assumptions of, for example, road speeds. Based on these limitations, we advise that the maps should be interpreted with care, and that the expert knowledge of actual conditions provided by health care officials at a local level is also taken into account when making decisions about allocation of treatment resources.
The snakebite incidence data available in this study, country-wide, based on the smallest administrative unit and probably reliable, are not available in many of the countries where this type of study needs to be conducted, owing to the large underestimation of snakebite incidence and mortality by hospital statistics [37], [42], [43]. For data available as small area counts, Bayesian smoothing techniques have a well-known ability to improve interpretability [26], as further demonstrated in this study. Thus, using Bayesian smoothing, the interpretability of the gathered data can be increased so that more accurate estimations of area-level incidence can be made from sparse data. Leynaud and Reati [21] used a spatial Bayesian smoothing technique available in SIGEpi [22]. We compared this technique with a Bayesian smoothing technique that also allowed for variation in district risk factor composition, and enabled estimation of the probability that an incidence threshold value was exceeded. We found that there were benefits of employing such smoothing techniques to improve the interpretability of the raw incidence when the data material was sparse, whereas there was no significant difference when the data material was increased (in this case five years of observation time instead of one year).
Large-scale approaches to identifying areas in need of antivenom could also benefit from using GIS. Available household-based incidence surveys, hospital and mortality records etc., could be mapped and used to construct geostatistical models which, based on the spatial variation of snakebite burden and its relationship with other spatially varying factors, predict snakebite burden in areas for which there is no data available. By taking this spatial approach, the sparse data available could be better utilised than predicting snakebite burden in non-surveyed areas by extrapolating information to country or Global Burden of Disease Region, as was the method used in the most recent review of global snakebite burden [1]. There has been an attempt to use such methods to map snakebite in West Africa [10], and useful methods have been further developed in studies predicting burden of other tropical diseases, such as soil-transmitted helminth infection [44] and malaria [45]. However, in order for these approaches to be feasible, there is still a large need for more data on snakebite burden, especially in sub-Saharan Africa, where a recent systematical review found only a small number of studies [46]. Large-scale studies such as those recently conducted to estimate snakebite mortality and incidence in India [43] and Bangladesh [37], respectively, provide an important source of data for producing snakebite burden maps, especially if the geographical coordinates of the survey clusters are available.
Mapping the availability of treatment is another challenge to implementing our method on a large scale; antivenom availability in many areas of low income countries is known to be poor [3], but information on antivenom availability on facility or even country level is to our knowledge not easily available, but requires further data collection.
GIS not only offers the possibility to improve the interpretation of incidence data through spatial smoothing, but also to identify areas which, on the basis of environmental risk factors, could be expected to have a high snakebite incidence. This approach could be especially useful when incidence data are considered unreliable. We compared the map of areas with an environment considered as suitable habitat for Bothrops asper, with a map of specimen collection locations, and found that these had good congruence, except for the Nicoya Peninsula, where environmental degradation could have led to species disappearance [17]. The method of determining high-risk areas by environmental determinants requires prior knowledge about the habitat of the snake species and should ideally be complemented by such field studies of actual snake distribution. Another important application of GIS in the struggle for reducing the impact of snakebite envenoming is the ability to analyze geographical accessibility to treatment, an important factor for the outcome of the bite. Leynaud and Reati [21] used Euclidean distances to hospitals and roads to analyze access to treatment, a common and readily implemented choice. We used a more demanding method aiming at estimating the time required for a snakebite victim to reach healthcare that takes into account topography, land use, road type and location of ambulance services. This theoretically allows for a more detailed analysis of the time needed to reach treatment, especially in mountainous study areas.
Our study demonstrates that using GIS it is possible to facilitate rational decision-making on localization of treatment resources against snakebite by overlaying the risk of snakebite accidents, estimated using reported data and/or presence of risk factors, transport times to existing hospitals or clinics, and the location of possible additional facilities to which treatment resources could be allocated. GIS is a promising tool for devising cost-effective interventions aimed at reducing the public health impact of snakebite envenoming.
|
10.1371/journal.ppat.1005708 | Bacterial Manipulation of NK Cell Regulatory Activity Increases Susceptibility to Listeria monocytogenes Infection | Natural killer (NK) cells produce interferon (IFN)-γ and thus have been suggested to promote type I immunity during bacterial infections. Yet, Listeria monocytogenes (Lm) and some other pathogens encode proteins that cause increased NK cell activation. Here, we show that stimulation of NK cell activation increases susceptibility during Lm infection despite and independent from robust NK cell production of IFNγ. The increased susceptibility correlated with IL-10 production by responding NK cells. NK cells produced IL-10 as their IFNγ production waned and the Lm virulence protein p60 promoted induction of IL-10 production by mouse and human NK cells. NK cells consequently exerted regulatory effects to suppress accumulation and activation of inflammatory myeloid cells. Our results reveal new dimensions of the role played by NK cells during Lm infection and demonstrate the ability of this bacterial pathogen to exploit the induction of regulatory NK cell activity to increase host susceptibility.
| Natural killer (NK) cells are an innate immune cell population known to promote antiviral immunity through cytolysis and production of cytokines. Yet, some pathogens encode proteins that cause increased NK cell activation. Here, using a model of systemic infection by the bacterial pathogen Listeria monocytogenes (Lm), we show that NK cell activation increases host susceptibility. Activated NK cells increased bacterial burdens in infected tissues despite their early production of the pro-inflammatory cytokine IFNγ. We found that the ability of NK cells to exacerbate infection was independent from their production of IFNγ and instead due to subsequent production of the anti-inflammatory cytokine IL-10. A single bacterial protein, p60, was sufficient to elicit NK cell production of both early IFNγ and delayed IL-10. IL-10-production by NK cells has been shown to occur in other systems, but our studies are first to show how this “regulatory” response impacts the course of a bacterial infection. We found that IL-10 producing NK cells suppress accumulation and activation of inflammatory myeloid cells. Our studies suggest that the exploitation of NK cell regulatory activity provides selective pressure for the evolution of pathogen proteins that promote NK cell activation.
| Immune defense against diverse pathogens requires timely recruitment of monocytes to sites of infection and activation of their antimicrobial functions [1]. IFNγ promotes antimicrobial activation of myeloid cells and is required for innate resistance to numerous pathogenic bacteria, including Lm [2,3]. Lm is an intracellular pathogen that causes severe and life-threatening infections primarily in elderly, pregnant, and immune compromised individuals [2–4]. In murine models of infection, Lm elicits a robust innate immune response that is characterized by IFNγ production by activated NK cells [5,6]. Memory-phenotype T cells can also serve as an early source of IFNγ following Lm and other bacterial infections [7,8]. NK cells protect against several viral infections and mediate anti-tumor immune responses in animal models and human patients [9,10]. Thus, NK cells have also often been assumed to confer protection during bacterial infections. However, there is a paucity of experimental evidence supporting a protective role for NK cells during in vivo antibacterial immune responses. Moreover, it has been puzzling that an Lm expressed virulence protein, p60, promotes NK cell activation and IFNγ production during infection [5,11].
NK cells were the first described innate lymphoid cell (ILC) population [12]. Activation of NK cell effector functions is regulated by germ line-encoded activating and inhibitory receptors [9]. Inhibitory NK cell receptors recognize host MHC or MHC-like molecules. Activating receptors recognize diverse stress-induced host proteins and, in some cases, microbe-encoded proteins [10]. Cytokines produced in response to infections also regulate NK cell activity. During Lm infection, the p60 protein appears to stimulate NK cell activation indirectly by promoting cytokine secretion from dendritic cells [11]. An abrupt increase in cytokines or activating receptor ligands or an encounter with target cells that have lost expression of ligands for inhibitory receptors licenses NK cells for cytolytic activity and secretion of IFNγ [9]. Some older studies provided evidence that depletion of NK1.1+ cells (which include both NK cells and NKT cells) increases host resistance to Lm infection [13,14]. The contributions of NK versus NKT cells to this phenotype and the mechanism for how these cells limit host resistance to Lm have not been described.
With appropriate stimulation, human and mouse NK cells have been observed to produce the immune regulatory cytokine IL-10 [15,16]. During Lm infection IL-10 has been shown to suppress both innate and adaptive immune responses and increases host susceptibility [17]. It is not known whether NK cells might be a crucial source of the IL-10 mediating these suppressive effects. However, one study provided evidence to suggest NK cells might produce IL-10 during Lm and Toxoplasma infections [18]. Additionally, IL-10 production by NK cells was shown to impair immunity during infection with the parasite Leishmania [19]. NK cells also have been shown to limit T cell responses during infections by MCMV and LCMV [20,21]. It is not known if or how NK cells affect adaptive immunity during Lm infection in wildtype mice. However, mice with a point mutation in NKp46 demonstrated hyper-activation of NK cells that correlated with reduced T cell responses to Lm-expressed ovalbumin [22]. These prior studies raised the hypothesis that NK cells responding to Lm infection might suppress host resistance through the production of IL-10, thus providing a rationale for Lm to express a protein that promotes NK cell activation.
Here, we used the murine model of systemic Lm infection to investigate how activation of NK cells and NK cell production of IFNγ impacts host susceptibility to this bacterial pathogen. Our results confirmed that NK cell activation exerts pro-bacterial effects. These effects were independent from IFNγ production and, in fact, NK cell IFNγ had no discernable effect on host resistance. Rather, we found that NK cells responding to Lm infection rapidly switched from IFNγ production to the secretion of IL-10. The secreted p60 virulence protein was sufficient to drive IL-10 production by mouse and human NK cells. IFNγ signaling in the NK cells dampened their IL-10 production. IL-10 producing NK cells were sufficient to dampen resistance to Lm infection and this regulatory activity was selectively associated with the suppression of inflammatory myeloid cell recruitment and activation. These data demonstrate the ability of a bacterial pathogen to exploit NK cell activation for selective suppression of innate immune responses during establishment of infection.
To investigate the impact of NK cells during bacterial infection, mice were depleted of NK1.1+ cells by a single injection of purified monoclonal Ab (αNK1.1) at 24 h prior to i.v. infection with 104 live Lm (~0.5 LD50). This protocol eliminated splenic CD3-NK1.1+NKp46+ NK cells from the time of infection (0 h post infection; hpi) through 96 hpi (S1A Fig). At 96 hpi, Lm burdens in the depleted mice were observed to be 10–100 fold lower than in mice treated with an isotype control Ab (IgG2a) (Fig 1A). NK1.1 cell depletion was also effective at reducing bacterial burdens following low dose Lm infection (S1B Fig), and prolonged survival of mice infected with ~2.5 LD50 (Fig 1B). To address whether depletion of NK1.1+ NKT cells contributed to these effects, NKT cell-deficient B6.cd1d-/- mice were infected with Lm. Unlike αNK1.1 treatment, the absence of NKT cells in B6.cd1d-/- mice had no significant effect on Lm burdens (Fig 1C). Antisera specific for the ganglioside asialo-GM1 (αGM1) also depletes NK cells but does not deplete conventional NKT cells [23]. Lm burdens in B6 mice treated with αGM1 before infection were identical to those in mice treated with αNK1.1 (Fig 1D). Depleting just the subset of NK cells expressing Ly49C/I also significantly reduced Lm burdens, though not to the extent as seen with αNK1.1 (S1C Fig). However, treatment of mice with a non-depleting monoclonal Ab that binds the NK cell surface markers NKG2A/C/E (αNKG2) did not impact Lm burdens (Fig 1D). These data indicated that removing NKT cells had no effect on Lm burdens whereas depletion of NK cells or a subset of NK cells dramatically suppressed bacterial survival and growth in host tissues. We conclude that the presence of NK cells acts to increase host susceptibility to Lm and that the protective effects of αNK1.1 treatment are due to depletion of NK cells.
NK cells were the largest population staining positive for intracellular IFNγ at 24 hpi (S1D and S1E Fig), which corresponded to the peak of their IFNγ production as determined by intracellular staining (Fig 1E). Serum IFNγ concentrations were reduced significantly in mice depleted of NK1.1+ cells, particularly at 24 hpi (Fig 1F). T cells also stained positive for intracellular IFNγ and were likely the source of residual serum IFNγ in the depleted mice (S1D and S1F Fig). We have previously shown that production of type I interferon down-regulates IFNGR, reducing host resistance to Lm [24]. To evaluate whether the effects of αNK1.1 treatment were dependent on type I interferon or IFNγ signaling, we evaluated Lm burdens in mice lacking expression of the type I interferon receptor (B6.ifnar1-/-) or the IFNγ receptor (B6.ifngr1-/-). Despite the extreme differences in susceptibility of these mouse strains to Lm, depletion of NK1.1+ cells was protective in both (Fig 1G and 1H). We also found that NK cell depletion did not impact Lm burdens until 72 hpi (Fig 1I), well after the peak of IFNγ production by NK cells (Fig 1E). Finally, when early NK cell IFNγ production was allowed to occur prior to αNK1.1 treatment, NK1.1+ cell depletion remained highly effective at reducing Lm burdens (Fig 1J). These results indicated that in mice with an intact T cell compartment NK cell production of IFNγ has no discernable impact on host resistance or susceptibility to Lm, arguing the pro-bacterial effects of NK cells are not due to hyper-production of IFNγ.
NK cells were previously shown to have the capacity to produce IL-10 at late stages of infections by the parasites Leishmania donovani and Toxoplasma gondii and viruses such as murine cytomegalovirus (MCMV) [18,19,25]. In one of these studies, IL-10-gfp+ NK cells were also observed at 4 dpi with Lm in Vert-X IL-10 GFP-reporter mouse [18]. We thus speculated that NK cells might be a source of IL-10 during Lm infection. Consistent with this hypothesis, we observed elevated serum IL-10 concentrations by 72–96 hpi in control but not αNK1.1-treated mice (Fig 2A). To more directly assess the potential for IL-10 production by NK cells, we evaluated gfp staining in NK cells from tiger IL-10 GFP-reporter mice infected with Lm [26]. In tiger IL-10 GFP-reporter mice the 3’UTR of IL-10 remains unaltered, unlike Vert-X and other IL-10 GFP-reporter mouse strains that contain a mRNA-stabilizing sequence [27]. This preserves post-transcriptional regulation of the labile il10 transcripts and thus permits more reliable detection of IL-10-producing cells over time without ex-vivo re-stimulation. At 72 hpi, gfp staining was selectively increased in NK cells from spleens, livers, and blood of Lm-infected reporter mice versus uninfected reporter mice or Lm-infected control B6 mice (Fig 2B and 2C). IL-10-gfp reporter expression was not observed at 24 hpi (Fig 2C). Intracellular staining for IFNγ confirmed the differing kinetics of IFNγ and IL-10-gfp production and demonstrated that only a small population of NK cells produced both cytokines (Fig 2D and S2 Fig). Thus, the timing of NK cell IL-10-gfp reporter activity correlated well with NK cell-dependent increases in serum IL-10 (Fig 2A) and Lm burdens (see Fig 1I). These data further suggested that most NK cells responding to Lm infection are committed to either IFNγ or IL-10 production at these time points during the infection.
We previously reported that the N-terminal LysM domain (LysM1) of the secreted Lm virulence protein p60 indirectly promotes NK cell IFNγ secretion during systemic infection and in cell culture studies. This protein domain stimulates DC production of cytokines including IL-12 and IL-18 that together with cell-contact stimulate NK cell IFNγ production [5,11]. Given the protective effects of IFNγ, it has not been clear how the pathogen might benefit from stimulating these responses. We thus asked whether Lm expression of p60 might also promote NK cell IL-10 secretion. Consistent with this hypothesis, serum IL-10 was significantly reduced in mice infected with an Lm strain deficient in p60 (Δp60; Fig 3A). However, the Δp60 strain is also attenuated in vivo, though this does not impair Lm infection of cultured macrophages or DCs [11,28]. To more directly investigate whether p60 expression permits Lm to stimulate NK cell IL-10 production we used a co-culture system consisting of bone marrow-derived DC (BMDC) and purified splenic NK cells [29]. Use of IL-10-deficient BMDC ensured any IL-10 in these cultures was derived from NK cells. BMDC were infected with wt Lm or the Δp60 strain. At 1hpi the BMDC were washed and media containing gentamicin was added. Purified splenic NK cells were added at 2 hpi (Fig 3B). Under these conditions, any IFNγ produced in the cultures is dependent on NK cells ([11,29]). Consistent with our prior findings, BMDCs infected with Δp60 Lm elicited significantly less NK cell-dependent IFNγ than those infected with wt Lm (Fig 3C). As shown above, NK cell IFNγ production during systemic Lm infection peaks at 24 hpi (Fig 1E and 1F), while IL-10 production peaks later (Fig 2A). Consistent with these results, IL-10 was not detected in culture supernatants at 24 hpi, but was reproducibly detected by 72 hpi (Fig 3D). As seen in serum, IL-10 concentrations in culture supernatants were also significantly reduced following infection with Δp60 Lm. These data suggested Lm expression of p60 stimulates DC to promote serial NK cell secretion of IFNγ and IL-10.
To specifically investigate whether the region of p60 protein that stimulates NK cell IFNγ secretion also promotes secretion of IL-10, BMDC were primed with TLR agonists and stimulated with a recombinant p60 fragment (L1S) that contains the LysM1 domain (Fig 3E). As previously reported [11], IFNγ secretion was observed selectively in 24 h co-cultures where BMDC and NK cells were stimulated with a priming agent (LPS) and L1S protein (Fig 3F). L1S also induced NK cell IL-10 secretion, but again this was selectively observed in the 72 h cultures (Fig 3G). As for NK cell production of IFNγ, IL-10 production required stimulation with both TLR agonist and L1S protein. Thus, our data suggested L1S stimulates primed BMDCs to promote NK cell activation for sequential IFNγ and IL-10 production. To determine whether human NK populations were also responsive to L1S, DCs were cultured 7 days from healthy donor PBMCs, then stimulated with a priming agent (pI:C) ± L1S and purified autologous NK cells as in Fig 3E. The results using human cells paralleled those above. Co-cultures with cells from 4/4 donors produced IFNγ at 24 h (Fig 3H) and IL-10 at 72 h (Fig 3I) in response to the L1S treatment. Thus, the L1S fragment of the Lm p60 protein is necessary and sufficient to stimulate the ability of primed DCs to promote early IFNγ and delayed IL-10 secretion from both murine and human NK cells.
Because IL-10 production by both mouse and human NK cell cultures was delayed relative to IFNγ production we considered whether production of IFNγ might inhibit NK cell IL-10 secretion. Subsequent to L1S or control stimulation, recombinant IFNγ was added to the co-cultures. NK cell IL-10 secretion was significantly reduced in the cultures treated with IFNγ (Fig 3J). To confirm these effects were due to IFNγ signaling in the NK cells, we established co-cultures using B6.il10-/- BMDC and NK cells purified from spleens of B6.ifngr1-/- mice. The IFNGR-deficient NK cells produced 4–5 fold more IL-10 than wt B6 NK cells (Fig 3K). Secreted IL-10 was also detected earlier in the cultures with IFNGR-deficient NK cells (Fig 3L). These results suggest that early IFNγ production may contribute to the observed delay in IL-10 production by the responding NK cells.
To further investigate the relationship between pro-bacterial effects of NK cells and their production of IL-10, NK cell depletion was performed in IL-10-deficient (B6.il10-/-) mice. Analysis of the infected B6.il10-/- animals revealed that liver bacterial burdens were comparable to those seen in B6 mice depleted of NK cells prior to infection (Fig 4A). NK cell depletion failed to further reduce Lm burdens in the B6.il10-/- mice. Thus, NK cell depletion and IL-10 deficiency had similar and non-additive effects on susceptibility to Lm, suggesting production of IL-10 by NK cells might be responsible for the increased host susceptibility. To further test this, we performed adoptive transfer experiments. CD45.2+ or CD45.1+ B6.il10-/- recipients were infected with Lm then respectively transferred with NK cells from the spleens of naïve wt CD45.1 (B6.ptprca) or IL-10 deficient CD45.2 (B6.il10-/-) mice (Fig 4B). At 96 hpi (72 h after transfer) small populations of donor CD45.1+ wt and CD45.2+ IL-10 deficient NK cells could be detected in spleens of the B6.il10-/- CD45.2+ and CD45.1+ recipients, respectively (Fig 4C), demonstrating persistence of the transferred cells. The detection of IL-10 protein in lysates of splenocytes from the B6.il10-/- recipients of wt, but not IL-10-deficient, NK cells indicated the transferred cells were activated to produce IL-10 (Fig 4D). This IL-10 production in the presence of wt NK cells was also associated with 10–100 fold increases in Lm burdens in livers and spleens of the B6.il10-/- mice (Fig 4E).
To establish whether IFNγ production by the NK cells might mediate pro- or anti-bacterial effects in the recipient mice, groups of B6.il10-/- mice received NK cells from the spleens of IFNγ deficient (GKO, B6.ifng-/-) mice. The GKO NK cells produced IL-10, as measured in spleen lysates (Fig 4D), and GKO NK cells sufficed to increase Lm burdens (Fig 4E). There was no significant difference in burdens of mice receiving wt or GKO NK cells and both NK cell types increased burdens in the il10-/- mice to a level near that seen in wt mice infected with Lm (compare Fig 4A and 4E). Experiments using donor NK cells labeled with CFSE further confirmed that the take of GKO, WT, and il10-/- NK cells was similar in the il10-/- recipients (S3A Fig). Thus, the ability of NK cells to produce IL-10 is a crucial factor governing Lm survival and replication during systemic infection and these pro-bacterial effects are independent from NK cell production of IFNγ.
IL-10 is known to suppress M1-type myeloid cell activation as well as the production of IFNγ by activated T cells [15]. Activated myeloid and T cells are both important for mediating resistance to Lm infection [2,3]. Thus, we asked whether NK cell IL-10 production might suppress myeloid or T cell responses during Lm infection. A Ly6C+CD11b+ inflammatory myeloid cell population was observed to accumulate by 3–4 dpi with Lm infection in both control and NK cell-depleted mice (Fig 5A, top). In both cases, the accumulating cells were a mixture of Ly6G+ neutrophils and Ly6G-CD11cl° inflammatory monocytes (Fig 5A, bottom). However, we observed significantly more of these inflammatory cells in spleens (Fig 5B) and livers (Fig 5C) of the mice lacking NK cells. Increased numbers of Ly6C+CD11b+ cells were also seen in spleens of Lm-infected il10-/- mice and these numbers were reduced when mice received NK cells capable of producing IL-10 during the experiments shown in Fig 4 (S3B Fig). These results suggest NK cell IL-10 suppresses the accumulation of these inflammatory myeloid cell populations at sites of bacterial infection. Both Ly6G- inflammatory monocytes and Ly6G+ granulocytes can ingest and kill bacteria to mediate protection against Lm [30,31]. Thus, blunting the accumulation of these inflammatory cell populations might itself promote bacterial infection. NK cell IL-10 also appeared to suppress the activation of these recruited myeloid cells, as we observed increased serum concentrations of IL-12p70 (a product of activated myeloid cells) in mice depleted of NK cells (Fig 5D). Further, in vivo depletion of NK cells before Lm infection led to an increased production of reactive oxygen species (ROS) by adherent splenocytes cultured from spleens of infected mice as measured by increased fluorescence of the ROS-sensitive dye, DCF (Fig 5E). These data together suggest regulatory NK cell activity suppresses accumulation and activation of inflammatory myeloid cells and thus their ability to ingest and kill Lm or infected cells.
In contrast to the effects on myeloid cell responses, depleting NK cells did not notably affect the activation of T cells specific for Lm-expressed antigens when measured at 7 dpi or following secondary Lm challenge (S4 Fig). Here, mice were treated with control and αNK1.1 antibodies before immunization with an ovalbumin expressing Lm strain (Lm-OVA). NK cell depletion did not change the number of Lm- or OVA- peptide responsive CD4+ or CD8+ T cells that stained positive for intracellular IFNγ+ 7 d later (S4A–S4C Fig). We also failed to observe altered expansion of T cells in response to secondary Lm challenge and the NK cell-depleted mice developed protective immunity (S4D and S4E Fig). These results suggest that the increased bacterial burdens associated with NK cell regulatory activity are primarily due to suppression of inflammatory myeloid cell, but not T cell, mediated immune responses.
Lm infection elicits a potent NK cell response, but whether and how NK cells might impact host resistance to Lm and other bacteria has remained poorly understood. One notion often stated in the literature is that NK cells protect against bacterial infections through their production of IFNγ. Certainly, IFNγ is crucial for resistance to Lm [32]. However, as shown here, NK cell IFNγ production peaks and wanes rapidly. T cells were also a significant early source of IFNγ production in our studies, and memory-phenotype T cells are known to be capable of antigen-independent IFNγ production that can mediate protection against Lm [7]. Memory-phenotype T cells transferred into IFNγ-/- mice conferred significantly more protection than transferred NK cells, though the latter conferred a low degree of protection in the absence of any other source of IFNγ [33]. Thus, IFNγ production by T cells appears to obviate IFNγ production by NK cells. Consistent with this interpretation, depleting NK cells in wildtype mice did not increase Lm burdens. Instead, we confirmed several older reports showing that depletion of NK1.1+ cells reduces severity of systemic infections by this and other bacteria [14,34,35]. The older studies did not discriminate the effects of NK versus NKT cells, but we showed here that the absence of NKT cells had no effect on Lm burdens during systemic infection. These results support the conclusion that NK cells are not a crucial source of early IFNγ production and that NK cells instead mediate pro-bacterial effects. Our data further demonstrate the mechanism responsible for these pro-bacterial effects. We found that NK cells responding to Lm produce IL-10 and this production is necessary and sufficient to increase Lm burdens. Further, the p60 virulence protein of Lm drives NK cell activation and IL-10 production. NK cells consequently suppressed innate myeloid cell responses. These findings together suggest Lm promotes NK cell activation to exploit their regulatory effects on antibacterial myeloid cell responses.
Despite the evidence that NK cell activation has deleterious effects during systemic Lm infection we are aware of a few seemingly contradictory reports. In studies where Lm was introduced into the footpad of mice, NK cell depletion was shown to modestly increase bacterial burdens in the draining lymph nodes [36,37]. Lm does not normally infect the host through the skin and little is known about the sequence of immune responses in this model. Perhaps footpad-inoculated Lm is unable to exploit NK cell IL-10 production, for example due to differences in the pattern or kinetics of inflammatory cell recruitment. Another example where NK cell deficiency was reported to increase susceptibility was in mice doubly deficient for the common gamma chain (γc) and Rag2 [38]. The γc is important for cellular responses to several cytokines, including IL-2, 4, 7, 9, 15 and 21. IL-15 signaling is particularly important for development and survival of NK and memory CD8+ T cells [39,40]. The increased susceptibility in the γc-/- x rag2-/- mice might thus be interpreted to indicate a protective role for NK cells. However, mice singly deficient for γc or Rag2 did not demonstrate increased susceptibility to Lm infection [38]. Thus, the reported susceptibility of mice doubly deficient for Rag2 and γc is not simply a result of NK cell deficiency. It was suggested that NK cells may be a key source of protective IFNγ in the absence of T cells [38]. It is also notable that recent studies indicate that Rag protein expression in NK cells modulates their survival and functional activity [41]. Regardless, it has since been shown that IL-15-/- mice exhibit increased resistance to systemic Lm infection [42].
It has been previously postulated that NK cells might exacerbate Lm infection through overproducing IFNγ [14]. Three lines of evidence from our studies argue against this model: (1) We showed that mice lacking IFNGR1 expression were still protected by depletion of NK1.1+ cells despite their increased susceptibility overall. (2) Depletion of NK cells was protective when initiated subsequent to peak NK cell IFNγ production. (3) Transfer of GKO NK cells were as effective as wt NK cells at increasing Lm burdens in IL-10 deficient recipients. However, Jablonska and colleagues recently argued that NK cell secretion of IFNγ contributes to pro-bacterial effects based on the finding that mice treated with a low dose of anti-IFNγ antibody had heightened resistance [43]. These apparently discrepant results could be explained by antibody stabilization of IFNγ to enhance its signaling, as has been shown to occur when cytokines such as IL-2 and IL-15 are complexed with antibodies or soluble receptor subunits [44,45]. We thus interpret the available data as evidence that NK cells exert pro-bacterial effects independent from their IFNγ production.
We instead found that the key mechanism by which NK cells increase susceptibility to Lm is through production of IL-10. How does NK cell-derived IL-10 suppress resistance to bacterial infection? It is well established that mice entirely deficient for IL-10 are highly resistant to Lm infection. This resistance is associated with increased innate and adaptive immune responses [17]. Consequently, we evaluated the effects of NK cell depletion on both T and myeloid cell responses to Lm infection. We failed to observe any effect of NK cells on the T cell response to Lm infection, suggesting that NK cell IL-10 is not responsible for the previously observed suppression of T cell responses. Consistent with this result, CD4+ T cells were recently shown to be a crucial source of IL-10 that regulates memory CD8+ T cell responses during LCMV infection [46]. In contrast, our findings here indicated that NK cell IL-10 suppresses both accumulation and activation of myeloid cells. Coincident with the timing of NK cell IL-10 production in control Lm infected mice (3–4 dpi), we found that depleting NK cells increased inflammatory monocyte and neutrophil accumulation in spleens and livers, increased serum IL-12p70, and increased ROS production in cultures of adherent splenocytes. Activated myeloid cells are the primary source of IL-12, and its production is known to be suppressed by IL-10 [47,48]. ROS production is also a correlate of M1-type activation, is suppressed by IL-10, and correlates with macrophage and neutrophil bactericidal activity [49,50]. Further, IL-10 blockade is known to increase macrophage bactericidal activity against Lm [49,51,52]. Production of IL-12 or ROS may not themselves mediate reduced Lm burdens in NK cell-depleted mice, but certainly indicate enhanced myeloid cell activation. Accumulation and activation of myeloid cells is crucial for immune resistance to infections by Lm and many other intracellular pathogens [1]. Hence, the impairment of these processes likely accounts for the ability of NK cell IL-10 to increase susceptibility during Lm, and possibly other, bacterial infections. Presumably, this dampening of inflammatory responses benefits the host in other settings, such as in the context of inflammatory diseases.
Lm infection is not unique in its ability to stimulate IL-10 secretion by NK cells. NK cell IL-10 production was previously observed at late stages of chronic infection by Leishmania donovanni [19], and during infections by Toxoplasma gondii [18], and murine cytomegalovirus (MCMV) [53]. During T. gondii infection, NK cell activation is regulated by cytokine (IL-12) production and this IL-12 is driven by ligation of TLR11 and 12 by the parasite profilin protein [54]. IL-12 production during T. gondii [18] induces NK cell expression of the aryl hydrocarbon receptor (Ahr) transcription factor to drive il10 transcription [55]. IL-12 also drives il10 transcription in NK cells during chronic infection by the parasite L. donovanii and IL-10 producing NK cells were shown to increase parasite numbers in this model [19]. However, it is not yet clear whether a specific L. donovanii protein drives the NK cell response to this infection. NK cell IL-10 production during MCMV infection is largely seen in the Ly49H+ NK cell subset [25]. Ly49H is an activating receptor that responds to the virus-encoded M157 protein [56,57]. Work here showed that the Lm p60 protein was important for promoting regulatory NK cell activity. Our prior work found that Lm expression of the p60 protein increases both bacterial replication in host tissues and NK cell production of IFNγ early after systemic infection [5]. Recombinant p60 protein was subsequently shown to stimulate production of IL-18 by primed BMDC to stimulate cell contact-dependent NK cell production of IFNγ [11,58]. In the present paper Lm expression of p60 increased NK cell secretion of IL-10 in vivo and in BMDC/NK cell co-cultures. Stimulation of primed murine BMDCs or human PBMC-derived DCs with a recombinant fragment of p60 protein (L1S) likewise sufficed to trigger NK cell secretion of IL-10. These data suggest Lm uses p60 to actively promote NK cell IL-10 secretion. Together with the prior work in other pathogen models, these data with p60 further suggest that the presence of NK cell stimulating proteins might be a marker for pathogens that have evolved to exploit NK cell regulatory activity.
The kinetics of NK cell IL-10 secretion was found here to occur only after reductions in NK cell IFNγ secretion, both during systemic infection and in cell co-cultures infected with Lm or stimulated with recombinant L1S protein. Similar delays were observed in other models where NK cell IL-10 production occurs [18,19,53]. For example, during Leishmania donovani infection NK cell IFNγ production is stimulated in an IL-12-dependent manner from 24 hr infected mice, followed by NK cell IL-10 production from 21 day infected mice [19]. Similarly, during MCMV infection IL-2 and IL-12 drove NK cell IFNγ as well as subsequent IL-10 production in culture ex vivo [53]. These studies suggest that the switch from IFNγ to IL-10 production is a consequence of NK cell activation in multiple infections. However, we do not believe this switch is a “hard-wired” response to activation given previous reports showing certain cytokine stimulation protocols which induce IFNγ production by human NK cells fail to also trigger IL-10 secretion [59,60]. Recent work from Biron and colleagues suggested that during MCMV infection this delay reflects a requirement for NK cell proliferation to open the il10 locus to transcriptional machinery [53]. Proliferation might similarly contribute to IL-10 production by NK cells during Lm or other infections, though this remains to be determined. Defining the mechanistic basis for sequential production of IFNγ and IL-10 was not the purpose of our studies, but we did observe that IFNγ acts to suppress NK cell IL-10 secretion. Further, we found that NK cells deficient in IFNγ signaling secreted increased quantities of IL-10. These results suggest early NK cell IFNγ production may be important for suppressing or delaying NK cell IL-10 secretion. In T cells an initial pro-inflammatory response is necessary for switching from IFNγ to IL-10 production [61–63]. However, we found that IFNγ-deficient NK cells remained capable of producing IL-10 and were as effective as wt NK cells at increasing susceptibility in IL-10-/- mice. Thus, cell-intrinsic IFNγ production does not appear to be an essential stimulus for NK cell IL-10 secretion. What additional factor(s) might contribute to driving NK cell transitioning from IFNγ to IL-10 production during Listeria and other infections remains to be determined.
The impact of NK cell regulatory activity on human health and disease is not yet known. However, human NK cells were previously shown to produce IL-10 [59], and we showed that Lm p60 could drive IL-10 secretion by human NK cells. Thus, NK cell IL-10 production may well impact human susceptibility to infections and other diseases, including Lm infection. Severe Lm infections primarily occur in elderly and pregnant individuals. Ageing is associated with an increase of circulating NK cells in humans [64], and NK cells are a major cell population in the placenta of pregnant humans and animals [65]. Perhaps then, the increased prevalence of NK cells and their IL-10 production is an important factor governing the susceptibility of these populations. Pregnant individuals also have increased susceptibility to infections by T. gondii, Cytomegalovirus, and Leishmania [66]. Depletion of NK cells or more selective approaches to suppress their acquisition of regulatory activity could thus prove useful in some clinical settings. Improved understanding of how Lm p60 and other pathogen factors induce pro-inflammatory and regulatory NK cell responses will be an important step in defining the potential benefits, risks, and feasibility of manipulating NK cells in the context of infectious, autoimmune, and cancerous diseases.
Female mice were used at 8–12 weeks of age. C57BL/6J, B6.il10-/-, B6.ptprca, B6.ifngr1-/-, B6.ifnar1-/-, B6.ifng-/- (GKO) and B6.IL-10-gfp (tiger) mice were purchased from Jackson Labs. B6.cd1d-/- mice were from Dr. Laurent Gapin (Univ. Colorado). Mice were maintained in the National Jewish Health Biological Resource Center and University of Colorado Office of Laboratory Animal Resources.
Mice were treated i.p. with PBS or PBS containing 100 μg of purified Ab or 100 μl of rabbit antisera to the ganglioside asialoGM1 (α-GM1, Wako USA). Endotoxin free IgG2a control (C1.18) and αNK1.1 (PK136) Abs were purchased (Bio X Cell). Anti-Ly49C/I (clone 5E6) and anti-NKG2A/C/E (20D5) Abs were purified from hybridoma supernatants using protein A affinity chromatography. Unless otherwise stated, Abs were given in a single dose at 24 h before infection. Depletions were confirmed by flow cytometry.
L. monocytogenes (Lm; strain 10403s), congenic p60-deficient [28], and OVA-expressing [67] Lm (OVA-Lm) were thawed from frozen stocks and diluted for growth to log phase in brain-heart infusion or tryptic soy broth (MP Biomedicals), then diluted in PBS and given to mice i.v. in the lateral tail vein. Unless stated otherwise, mice received a single sublethal dose of 104 CFU. Lm was given at 4000 CFU for infection in B6.ifngr1-/- mice, and a lethal dose of 5 x 104 CFU for analysis of survival ± NK cell depletion. OVA-Lm was given at 5000 CFU for immunizations. Challenges used a lethal dose of Lm-OVA (105 CFU). For CFU counts, organs were harvested into 0.02% Nonidet P-40, homogenized for 1 min with a tissue homogenizer (IKA Works, Inc.) and serial dilutions were plated on BHI or TSB agar plates.
Spleens were harvested into media containing penicillin/streptomycin at 100 U/ml then transferred to a solution of 1 mg/ml type 4 collagenase (Worthington) in HBSS plus cations (Invitrogen). After a 30 min incubation at 37°C, spleens were disrupted by passage through a 70 μm cell strainer and the resulting single cell suspensions treated with RBC lysis Buffer (0.15 M NH4Cl, 10mM KHCO3, 0.1 mM Na2EDTA, pH 7.4) for 2 min. Prior to intracellular staining, splenocytes were incubated 3–4 h in RP10 media (RPMI 1640, 10% FBS, 1% L-glutamine, 1% Sodium Pyruvate, 1% Penicillin, 1% Streptomycin and 0.1% β-ME) containing Brefeldin A (GolgiPlug; BD Biosciences). No additional ex vivo stimulation was included for NK cell analyses. For T cells, 1 μM concentrations of synthetic OVA257–264 (SIINFEKL) or LLO190–201 (NEKYAQAYPNVS) peptides were included during the incubation. Blood cells were harvested into HBSS plus cations and heparin (Sigma). Cells were treated twice with RBC lysis Buffer for 1 min. Liver cells were harvested and treated with collagenase in the same manner as spleen cells. Following passage through a 70 μm cell strainer, cells were re-suspended in 40% Percoll in HBSS minus cations. The 40% Percoll was underlayed with 60% Percoll, and cells were collected from gradient following centrifugation. RPMI with 5% FBS was added to cells to pellet, and cells were treated with RBC lysis Buffer for 1 min.
Anti-CD16/32 (2.4G2 hybridoma supernatant) was added to block Fc receptors prior to staining, which used FACS buffer (1% BSA, 0.01% NaN3, PBS) containing fluorophore-labeled antibodies purchased from eBioScience or BioLegend. Staining antibodies included anti- CD3 (clone 145 2C11), CD4 (clone RM-4-5), CD8 (clone 53–6), CD11b (M1/70), CD11c (N418), CD27 (clone LG.7F9), CD45.1 (clone A20), IFNγ (clone XMG1.2), Ly6C (HK1.4), Ly6G (1A8), NK1.1 (clone PK136), and NKp46 (clone 29A1.4). After surface staining, cells were fixed in 2–4% paraformaldehyde for direct analysis with or without saponin treatment for intracellular staining. To amplify IL-10-gfp signal [53], fixed and permeabilized cells were stained with a rabbit monoclonal anti-GFP followed by goat anti-rabbit IgG Alexa Fluor 488 (both from Life Technologies). At least 100,000 data events per sample were collected using an LSRII (BD Biosciences). FlowJo software (Treestar) was used for analysis of flow data.
Bone marrow-derived DC (BMDC) were cultured from B6.il10-/- mice and infected with Lm or stimulated with recombinant L1S protein purified as previously described [11,29]. Briefly, bone marrow was cultured 6d in GM-CSF and 3 x 105 BMDC (>90% CD11c+) per well were cultured overnight in 24 well plates. For infections, log phase wt or Δp60 Lm were added at a multiplicity of one bacterium per BMDC. One h later, cells were washed and gentamycin was added at 10 μg/mL. For L1S stimulation, BMDC were activated 1 h by treatment with 20 μg/ml poly I:C (Invivogen, San Diego, CA) or 10 ng/ml LPS (L8274 Sigma-Aldrich, St. Louis, MO) and purified L1S protein was added to the BMDC at 30 μg/ml. For IFNγ treatment, 50 ng/mL IFNγ (Life Technologies) was added at 1 hr post L1S stimulation. Purified splenic NK cells were negatively sorted using the EasySep Mouse NK cell Enrichment Kit (19755 Stemcell Technologies) and added to cultures 2 h after Lm or L1S treatments at a ratio of .1:1 (NK cells:BMDC). Purified NK cell populations were >80% NK1.1+CD3-. To culture human DCs, adherent PBMCs were obtained from unrelated donors and grown in RPMI 1640 (Invitrogen) supplemented with 10% human AB serum (Innovative Research, Novi, MI), 0.01M HEPES, 0.02mg/ml gentamicin, 200 IU/ml IL-4 (eBioscience), and 100 IU/ml GM-CSF (R&D Systems, Minneapolis, MN). After 6 days, DC were plated at 105 cells/well in 96 well plates, primed, and stimulated with L1S and polyI:C as above. Purified human NK cells were negatively sorted from PBMCs using the EasySep Human NK cell Enrichment Kit (19015 Stemcell Technologies). Supernatants were harvested for analysis of IFNγ and IL-10 at 24 or 72 h cytokines using commercial ELISAs (BD Biosciences).
Blood was collected by cardiac puncture. Serum from clotted blood was collected and frozen prior to analysis using commercial ELISAs for IFNγ, IL-12p70, or IL-10 (BD Biosciences). Equal numbers of splenocytes were homogenized in 1 mL of 0.02% Nonidet P-40 with protease inhibitor cocktail (ThermoFisher) using a dounce for measurement of IL-10 in homogenates.
NK cells were purified from spleens of naïve B6.ptrpca, B6.il10-/-, and B6.ifng-/- (GKO) mice (>80% purity) using the EasySep Negative Selection Mouse NK Cell Enrichment Kit (Stemcell Technologies). Each recipient received 1.5–2 x 106 live NK cells i.v. at 24 hpi.
Single cell splenocyte suspensions were prepared as above and 106 cells were added per well to a 96 well round bottom suspension plate (Greiner) followed by centrifugation at 500 x g for 5 min. Pelleted splenocytes were resuspended in DPBS (Gibco) containing 10μM 2',7'-dichlorodihydrofluorescein diacetate (DCF; ThermoFisher) and 1% DMSO. After a 30 min incubation at 37°C, spenocytes were washed twice in DPBS and resuspended at 100 μl/well in phenol red-free DMEM (Gibco) then transferred to 96 well white bottom Nunc F96 plates (ThermoFisher). Plates were incubated in a 37°C Biotek Synergy HT plate reader and fluorescence read at 15 min intervals using a 528/20 emission filter and Gen5 software.
Graphing and statistical analyses used Prism (GraphPad Software). Statistical tests included t-tests, analysis of variance (ANOVA), and linear regression with Pearson correlation. P<0.05 was considered significant.
The Animal Care and Use Committees for National Jewish Health (Protocol# AS2682-08-16) and the University of Colorado School of Medicine (Protocol# 105614(05)1E) approved all studies. These protocols adhere to standards of the United States Public Health Service and Department of Agriculture.
|
10.1371/journal.pgen.1004877 | Antagonistic Cross-Regulation between Sox9 and Sox10 Controls an Anti-tumorigenic Program in Melanoma | Melanoma is the most fatal skin cancer, but the etiology of this devastating disease is still poorly understood. Recently, the transcription factor Sox10 has been shown to promote both melanoma initiation and progression. Reducing SOX10 expression levels in human melanoma cells and in a genetic melanoma mouse model, efficiently abolishes tumorigenesis by inducing cell cycle exit and apoptosis. Here, we show that this anti-tumorigenic effect functionally involves SOX9, a factor related to SOX10 and upregulated in melanoma cells upon loss of SOX10. Unlike SOX10, SOX9 is not required for normal melanocyte stem cell function, the formation of hyperplastic lesions, and melanoma initiation. To the contrary, SOX9 overexpression results in cell cycle arrest, apoptosis, and a gene expression profile shared by melanoma cells with reduced SOX10 expression. Moreover, SOX9 binds to the SOX10 promoter and induces downregulation of SOX10 expression, revealing a feedback loop reinforcing the SOX10 low/SOX9 high ant,m/ii-tumorigenic program. Finally, SOX9 is required in vitro and in vivo for the anti-tumorigenic effect achieved by reducing SOX10 expression. Thus, SOX10 and SOX9 are functionally antagonistic regulators of melanoma development.
| For the development of future cancer therapies it is imperative to understand the molecular processes underlying tumor initiation and expansion. Many key factors involved in these processes have been identified based on cell culture and transplantation experiments, but their relevance for tumor formation and disease progression in the living organism is often unclear. Therefore, genetically modified mice spontaneously developing tumors present indispensable models for cancer research. Here, we address this issue by studying the formation of melanoma, the most fatal skin tumor in industrialized countries. To this end, we use a transgenic mouse model to elucidate cellular and molecular mechanisms regulating congenital nevus and melanoma initiation. We show that a transcription factor called SOX10 promotes melanoma formation by repressing an anti-tumorigenic program involving the activity of a related factor, SOX9. When SOX10 is inactivated, SOX9 becomes upregulated and induces cell cycle arrest and death in melanoma cells. Furthermore, upon experimental elevation of SOX9 levels, SOX10 activity is suppressed, revealing an antagonistic relationship between SOX9 and SOX10 in melanoma initiation. Knowledge of how an anti-tumorigenic program can be stimulated by modulating the activities of these key factors might help to design novel therapeutic strategies.
| Sox (Sry (sex determining region Y)-related HMG box) genes encode a family of transcription factors that are characterized by a conserved high-mobility group (HMG) domain mediating their binding to DNA in a sequence-specific manner [1–3]. While the majority of Sox proteins functions as transcriptional activators, some members of the Sox family including Sox9 and Sox10 may also act as transcriptional repressors [4–6]. Sox genes play key roles in embryonic development and are major determinants of stem cell behavior, regulating cell fate decisions and maintaining cellular identity [3]. Their crucial role in normal tissue formation and homeostasis is evident from the fact that several mutations in Sox genes are causative for developmental diseases, and accumulating evidence demonstrates the important functional role of Sox family proteins in a variety of cancers [7–10].
A common feature of SoxE group proteins, which includes Sox9 and Sox10, is their expression in neural crest (NC) cells during embryonic development [2, 11]. NC cells are a transient embryonic cell population that gives rise to most of the peripheral nervous system, chondrocytes and osteoblasts of craniofacial structures, smooth muscle cells of the cardiovascular system, and melanocytes, the pigmented cells of the skin [12]. While Sox9 is expressed in premigratory NC cells and in the pharyngeal apparatus, Sox10 is found in NC cells at the time of their emigration and is essential for their self-renewal and survival [12–16]. Loss of Sox10 results in absence of most NC derivatives, whereas Sox10 haploinsufficiency causes Waardenburg Hirschsprung syndrome, characterized by aganglionic megacolon, pigmentary abnormalities and often deafness due to loss of sensory innervation [13, 17–20]. In the melanocytic lineage, Sox10 is expressed during all stages of development as well as in the adult and is required in different species for the generation and homeostasis of embryonic and adult melanocytes in vitro and in vivo [13, 21–25].
In contrast, loss of Sox9 in the NC does not lead to general defects in NC-derived structures, but specifically affects the development of mesectodermal derivatives, such as smooth muscle cells and craniofacial bones and cartilage [11, 26–28]. Furthermore, heterozygous mutations in Sox9 in both mice and humans, result in campomelic dysplasia, a syndrome associated with dwarfism, skeletal malformations, cleft palate, XY sex reversal and often hermaphroditism [28–30]. However, data on Sox9 expression in melanocytes are inconsistent, and a functional implication of Sox9 in melanocyte formation has not been provided so far [23].
Based on the assumption that mechanisms of tumor formation might be related to those underlying the generation of the cell type, from which the tumor develops, we and others have recently addressed the function of Sox10 in melanoma. These studies demonstrated a crucial role of Sox10 in the pathogenesis of giant congenital naevi and melanoma in both mice and humans by regulating proliferation and survival of melanocytic cells and maintenance of their cellular identity [9; 31]. However, the precise molecular mechanisms mediating Sox10 function in melanoma remain to be investigated.
Here we show that in contrast to Sox10, Sox9 appears to be expressed at very low levels only and is functionally not required in melanocyte stem cells, committed melanocytes, and melanoma cells. However, Sox9 expression is elevated upon Sox10 deletion in mouse and human melanoma cells [9], and critically contributes to the anti-tumorigenic effects observed upon Sox10 inactivation in giant congenital naevus and melanoma.
While SOX10 expression and function has been well established in adult melanocytes, naevi, and melanoma tissue in human and mice [9; 23], studies on SOX9 expression in melanocytic cells are controversial. SOX9 was reported to be expressed in cultured human melanocytes in vitro [32], human melanocytes in vivo [33], and in human melanoma [34–36]. Other reports, however, failed to reveal Sox9 mRNA and protein expression in melanoblasts and differentiated melanocytes during development and postnatally [21,37]. Given the close relationship between SoxE factors, one conceivable explanation for these discrepancies might be that antibodies raised against a given SoxE protein fail to discriminate between SOX9 and SOX10 epitopes. Indeed, when we performed immunohistochemistry on murine skin to test the specificity of various anti-SOX9 antibodies, several of them recognized both melanocytes and epithelial cells in the outer root sheet (a region in hair follicles known to express and functionally require Sox9; [37–38] (S1 Fig.). In contrast, the antibody sc-20095 exclusively detected protein expression in epithelial cells but not in melanocytes. Of note, in human melanoma cell lines in vitro, all antibodies tested but sc-20095 not only recognized SOX9, but also a protein of the molecular weight of SOX10 and detected by a SOX10-specific antibody (S2A-S2D Fig.).
To further investigate the specificity of anti-SOX9 antibodies, we performed SOX10 knockdown in human melanoma cell lines in vitro and analyzed SOX10 and SOX9 expression using Western blot analysis (S2E-S2K Fig.). As shown in S2E-S2K Fig., different anti-SOX9 antibodies used in earlier studies detected SOX10 protein expression, which was lost upon SOX10 knock-down. The only anti-SOX9 antibody, which did not display cross-reactivity with SOX10 protein, was sc-20095.
Therefore, we chose to reassess SOX9 expression in human melanocytes and melanocytic skin lesions using the specific SOX9 antibody sc-20095. Double immunostaining for SOX9 and MITF (Microphthalmia-associated transcription factor), an established marker of melanocytes [39] revealed no detectable SOX9 expression in human skin melanocytes in vivo (Fig. 1A-G). In contrast, SOX10 was readily detectable in human melanocytes (S3A-S3B Fig.). Moreover, while SOX10 was expressed in 100% of human giant congenital naevi, SOX9 expression was not detected in the same set of patient samples (n = 17; (Fig. 1H; S3D Fig.); [9]). Likewise, all samples of a melanoma tissue microarray composed of 56 primary melanoma biopsies revealed strong SOX10 expression (Fig. 1I; [9]). SOX9 expression, however, was found in only 41% (23/56) of the primary melanoma samples, in which it was expressed in a few scattered cells accounting for less than 10% of all melanoma cells (Fig. 1I). In contrast, expression of SOX9 was readily detectable in the epithelial lineage of normal skin as well as in basal cell carcinoma, an epithelial skin cancer (Fig. 1E, F; S3C Fig.; [37,40]). To investigate the mRNA expression of SOX10 and SOX9 in a large set of human melanoma samples, we used of the TCGA (The Cancer Genome Atlas) database. Interestingly, the vast majority of human melanoma samples displayed much higher SOX10 than SOX9 expression (Fig. 1J) and only very few samples were characterized by a SOX9 high / SOX10 low expression pattern (Fig. 1K). Thus, SOX10 but not SOX9 is prominently expressed in normal human melanocytes, human giant congenital naevi, and primary melanoma.
To corroborate our findings in an experimentally amenable system, we extended the analysis of Sox9 expression to mouse melanocytes, taking advantage of a previously described iDct-GFP mouse line (Fig. 2A,B; [41]). Doxycycline-induced GFP-labelled melanocytes were isolated via fluorescence-activated cell sorting (FACS) and subjected to RNA-Seq analysis (Fig. 2A). While Sox10, Mitf and Tyr were expressed at high levels (Sox10 reads were 1292, 1372, 1776 and 2488 at E15.5, E17.5, P1 and P7, respectively), the expression of Sox9 was extremely low (Sox9 reads were 68, 65, 105 and 128 at E15.5, E17.5, P1 and P7, respectively). These data are in accordance with earlier studies on Sox9 mRNA and protein expression in murine melanoblasts and melanocytes [21,37] and suggest that in contrast to Sox10, Sox9 expression is virtually absent in the melanocytic lineage during mouse embryogenesis and postnatally.
Melanocyte stem cells are found in a specialized region of the hair follicle called bulge and give rise to melanocyte progenitors and differentiated melanocytes [42]. The latter are located in the lower hair follicle portion termed bulb, where they transfer pigment to the growing hair. When melanocyte stem cells are functionally impaired, they fail to generate melanocytes, which results in hair graying [43]. To further investigate the expression of Sox10 and Sox9 in the melanocytic lineage of the mouse skin, we made use of Dct::LacZ transgenic mouse line expressing LacZ driven by the dopachrome tautomerase (Dct) promoter that allows genetic tracking of melanocyte stem cells and their derivatives in the hair follicle (Fig. 2C; [44]). Sox10 expression was detected in X-Gal-positive melanocyte stem cells located in the bulge region (Fig. 2E, upper panels) as well as in differentiated melanocytes in the hair follicular bulb (Fig. 2E, lower panels). Similarly to the situation in human melanocytes, immunostaining with the specific anti-Sox9 antibody sc-20095 demonstrated absence of Sox9 expression in X-Gal-positive melanocyte stem cells and their progeny (Fig. 2D, upper and lower panels). Sox9 expression was restricted to cells of the epithelial lineage, namely the outer root sheath and the epithelial stem cell compartment in the bulge area (Fig. 2D; S4 Fig.), in agreement with previous reports [37,38,45].
To address the function of Sox10 and Sox9 in the melanocytic lineage in vivo, we conditionally ablated either Sox10 (Fig. 2H) or Sox9 (Fig. 2F) using Tyr-CreERT2 transgenic mice [46] carrying floxed alleles of Sox10 [47] and Sox9 [27], respectively. Tamoxifen-induced homozygous deletion of Sox10 in Sox10fl/fl Tyr-CreERT2 mice resulted in progressive hair graying, revealing an essential function of Sox10 for melanocyte stem cell homeostasis (Fig. 2I, [23]). In contrast, homozygous deletion of Sox9 in the melanocytic lineage did not cause hair graying even after a prolonged period after tamoxifen-induced gene deletion (Fig. 2G). Thus, these two closely related genes are not only differentially expressed but also play distinct roles in the biology of melanocytes in vivo.
To functionally assess the role of SoxE factors in melanomagenesis, we first performed immunohistochemical staining for Sox9 and Sox10 of giant congenital naevi formed in Tyr::NrasQ61K mice and in melanoma derived from giant congenital naevi in Tyr::NrasQ61KInk4a−/− mice [9,48]. In contrast to the widespread expression of Sox10 displayed by mouse naevi and melanoma tissue (Fig. 3B,D, [9]), immunostaining of both naevi and primary melanoma sections did not reveal any detectable expression of Sox9 protein (Fig. 3A,C; S4B Fig.), consistent with the data that we have obtained for human giant congenital naevi and melanoma (Fig. 1H-I). Despite the lack of detectable Sox9 expression, low levels of Sox9 might be functionally implicated in the formation of melanocytic lesions arising in Tyr::NrasQ61K mice. To address this issue, we generated Tyr::NrasQ61K Sox9fl/fl Tyr-CreERT2 mice and conditionally deleted both Sox9 alleles by tamoxifen treatment of the mice (Fig. 3E). However, skin hyperpigmentation induced by oncogenic NRas was not affected in Tyr::NrasQ61K Sox9fl/fl Tyr-CreERT2 mice and was comparable to that presented by their control Tyr::NrasQ61K littermates (Fig. 3F). These data reveal that in contrast to Sox10 (Fig. 3G,H}, Sox9 is not required for the formation of melanocytic lesions.
To gain insight into the possible interplay between SOX10 and SOX9 during tumor progression, we quantified the expression levels of SOX10 and SOX9 in normal human melanocytes, in cells from giant congenital naevi, and in a melanoma cell line (M010817; [49]) using quantitative RT-PCR analysis (Fig. 4A-C). Notably, NRASQ61K-associated tumor progression was associated with an increase in SOX10 expression (Fig. 4B). Cells from giant congenital naevi showed a 5-fold increase in SOX10 expression when compared to normal melanocytes, while M010817 melanoma cells displayed a 10-fold increase in SOX10 expression when compared to normal melanocytes (Fig. 4B). In striking contrast to the increase in SOX10 expression, SOX9 expression levels were low in human melanocytes and further decreased with melanoma progression (Fig. 4C).
To assess whether the disparate expression of SOX9 and SOX10 is a general feature of human melanoma samples, we analyzed the endogenous expression of these SOXE proteins in a large set of human melanoma cell lines previously categorized into cells with proliferative and invasive signatures, respectively [49]. Of note, all cell lines with a proliferative signature were characterized by high SOX10/low SOX9 mRNA expression (Fig. 4D). However, many cell lines with an invasive signature displayed the opposite expression pattern and showed low SOX10/high SOX9 mRNA expression. These date were confirmed on the protein level by Western blot analysis of several proliferative and invasive cell lines (Fig. 4E).
Interestingly, an inverse correlation between SOX10 and SOX9 expression has previously been observed in several systems, including cultured human melanocytes, where upon the induction of differentiation, SOX10 levels were reduced concomitantly with an increase in SOX9 levels [32]. Thus, high expression of a given SoxE transcription factor might be causative for reduced expression of the related SoxE factor. We therefore asked whether deregulation of SOX10 leads to changes in SOX9 expression and found a significant upregulation of SOX9 mRNA expression upon SOX10 knockdown in human giant congenital naevus cells (Fig. 4F-H). This is in analogy to the upregulation of SOX9 mRNA previously observed in melanoma cells upon SOX10 knockdown [9]. Moreover, SOX10 knockdown also resulted in upregulation of SOX9 protein levels in human melanoma cells (S5A Fig.). The combined data indicate that SOX10 normally suppresses SOX9 expression in cells from melanocytic lesions.
To address whether the findings in human cells in vitro also apply to the in vivo situation in mice, we isolated melanocytic progenitors using fluorescence-activated cell sorting (FACS) from the skin of Tyr::NrasQ61K and Tyr::NrasQ61K Sox10LacZ/+ mice that lack one allele of Sox10 (Fig. 4I) and subsequently measured Sox9 expression levels using quantitative RT-PCR (Fig. 4K, L). Elevated Sox10 levels mediated by oncogenic NRas in melanocytic cells from Tyr::NrasQ61K mice [9] were associated with decreased Sox9 mRNA expression as compared to normal melanoblasts wild-type for NRas. However, reduced Sox10 levels brought about by Sox10 heterozygosity resulted in upregulation of Sox9 expression, as revealed by comparing cells from Tyr::NrasQ61K with cells from Tyr::NrasQ61K Sox10LacZ/+ mice. Thus, in cells derived from both human and mouse melanocytic lesions, reduced Sox10 levels were accompanied by elevated Sox9 expression.
Based on our findings it is conceivable that increased levels of SOX9 might mediate at least some of the anti-tumorigenic effects observed upon SOX10 loss-of-function in melanoma. To address this hypothesis, we overexpressed SOX9 in human M010817 melanoma cells in vitro and compared the gene expression profile of these cells with that of SOX10 knock-down melanoma cells, using the parental M010817 cell line as control [9]. Unsupervised hierarchical clustering revealed that overexpression of SOX9 led to a gene expression profile closely resembling the SOX10 knockdown-signature, which included in both conditions regulators of cell cycle progression, apoptosis, and melanocytic and mesectodermal differentiation (Fig. 5A). Among others, these data suggest that while SOX10 acts as an inhibitor of apoptosis in melanoma cells, SOX9 elicits a proapoptotic response. Similarly, SOX10 and SOX9 appear to play antagonistic functions in the regulation of the cell cycle, melanocytic differentiation, and mesectodermal differentiation (Fig. 5A).
Notably, SOX9 overexpression resulted in decreased expression of a number of genes associated with melanocytic differentiation, such as MLANA, MITF, DCT, TYR, and importantly SOX10. To confirm the downregulation of SOX10 upon SOX9 overexpression observed in microarray analysis (Fig. 5A) also on the protein level, we performed Western blot analysis and observed a pronounced downregulation of SOX10 protein upon SOX9 overexpression in human melanoma cell lines (Fig. 5B). Moreover, chromatin immunoprecipitation assays in human melanoma M010817 cells indicated that SOX9 binds to the promoter of SOX10, suggesting a direct regulation of the SOX10 gene by SOX9 (Fig. 5C). Thus, whereas SOX10 loss-of-function leads to increased SOX9 expression (Fig. 4), high levels of SOX9 suppress SOX10 expression, revealing cross-regulatory interactions between these two transcription factors.
Next, we addressed whether the cross-regulation between SOX10 and SOX9 is functionally relevant in human melanoma cells. To this end, we performed RNAi experiments to test whether interfering with SOX9 overexpression upon SOX10 knockdown could rescue M010817 melanoma cells (Fig. 5D-F). Using two independent sets of siRNAs, the elevated SOX9 levels observed in SOX10 knockdown cells were reverted below control levels by means of concomitant SOX9 knockdown (Fig. 5E). Importantly, both SOX9 gain-of-function and SOX10 knockdown promoted apoptosis (Fig. 5F; Fig. S5B). However, SOX10 knockdown cells were rescued when SOX9 expression was simultaneously downregulated, resulting in numbers of apoptotic cells comparable to those found in control cells (number of Annexin V-positive apoptotic M010817 cells: control, 9.95±0.9%; SOX10 siRNA #1, 19.96±0.13%; SOX10 siRNA #2 18.8±0.49%; combination of SOX10 siRNA #1/SOX9 siRNA, 9.45±0.79%; combination of SOX10 siRNA #2/SOX9 siRNA. 10.75±0.4%) (Fig. 5F; S5B Fig.). These data indicate that at least in vitro, SOX9 plays a key role in mediating the cellular phenotype obtained in human melanoma cells upon suppression of SOX10.
As in human melanoma cells in vitro, reducing Sox10 levels in vivo elicits an anti-tumorigenic effect, preventing melanocytic hyperplasia in Tyr::NrasQ61KSox10fl/+Tyr-CreERT2 mice [9]. To functionally test whether upregulation of Sox9 is the key factor accountable for the lack of hyperplasia in these mice in vivo, we performed simultaneous conditional ablation of both Sox10 and Sox9 genes in the Tyr::NrasQ61K mice (Fig. 6A, B). In agreement with our previous observations [9], the skin of the snout, paws and the back was noticeably lighter in color in Tyr::NrasQ61KSox10fl/+Tyr-CreERT2 mice as compared to their control Tyr::NrasQ61K littermates (Fig. 6A, left and middle panels). In contrast, macroscopic examination of the skin of Tyr::NrasQ61KSox10fl/+Sox9fl/flTyr-CreERT2 animals showed pronounced hyperpigmentation, a hallmark of the skin phenotype found in Tyr::NrasQ61K mice (Fig. 6A, right panels). Thus, conditional deletion of Sox9 rescued the effect of Sox10 haploinsufficiency in Tyr::NrasQ61K mice.
As shown in Fig. 2, conditional ablation of both alleles of Sox10 in the normal melanocytic lineage using Tyr-CreERT2 resulted in loss of functional melanocytes and hair graying (Fig. 2I). Likewise, homozygous deletion of Sox10 in Tyr::NrasQ61K mice not only prevented NrasQ61K-stimulated skin hyperpigmentation, but also led to hair greying in these mice (Fig. 6B, left panels). To determine whether these phenotypes involve Sox9 activity and whether, therefore, loss of Sox9 could rescue the effects of homozygous Sox10 deletion, we generated Tyr::NrasQ61KSox10fl/flSox9fl/flTyr-CreERT2 animals. Surprisingly, loss of Sox9 in Sox10 homozygous conditional knock-out animals efficiently restored skin hyperpigmentation as seen on the snout and paws of Tyr::NrasQ61KSox10fl/flSox9fl/flTyr-CreERT2 animals when compared to mice lacking Sox10 but not Sox9 in the melanocytic lineage (Fig. 6B, right panels).
We have previously demonstrated that skin hyperpigmentation in Tyr::NrasQ61K mice is due to hyperplasia of ectopically located pigment cells emerging in the dermis, around the upper part of the hair follicle [9]. To measure the degree of hyperpigmentation in Tyr::NrasQ61K mice with conditional Sox10 deletion (heterozygous and homozygous) versus mice simultaneously lacking both Sox9 alleles and one or both Sox10 alleles in the melanocytic lineage, we quantified the percentage of hair follicles associated with ectopically located melanocytic cells in the back skin of these mouse lines. In Tyr::NrasQ61K mice, more than 90% of all hair follicles displayed ectopic pigment cells (Fig. 6D). In contrast, in the absence of one or both alleles of Sox10, there were almost no hair follicles with ectopic pigment cells, despite NrasQ61K expression (Fig. 6C, D). Strikingly, however, the percentage of hair follicles associated with ectopic melanoblasts in Tyr::NrasQ61KSox10fl/+Sox9fl/flTyr-CreERT2 animals was reverted to numbers similar to those found in Tyr::NrasQ61K mice (93±1.8% and 96±1%, respectively) (Fig. 6C, D). Moreover, even in the absence of both Sox10 alleles, loss of Sox9 rescued the NrasQ61K–dependent appearance of melanocytic cells found outside hair follicles (93.5±3%) (Fig. 6D). These data reveal a key role of Sox9 in preventing melanoma initiation and provide novel insights into the functional interplay between Sox10 and Sox9 during melanoma formation.
Our study identifies the structurally related transcription factors SOX10 and SOX9 as functionally antagonistic regulators of postnatal melanocyte and melanoma development. Although we did not find SOX9 to be expressed in the melanocytic lineage when SOX10 is present, SOX9 expression becomes evident upon SOX10 inactivation in naevus and melanoma cells. In this context, SOX9 appears to promote the major cellular processes induced by SOX10 loss-of-function, namely stop of proliferation and apoptosis. Intriguingly, SOX9 and SOX10 are engaged in a cross-regulatory feedback loop whereby SOX9, which is induced upon SOX10 inactivation, itself suppresses SOX10, thus strengthening an anti-tumorigenic program.
In many cell lineages and tissues, SOX10 and SOX9 are co-expressed and functionally redundant [50]. We propose that this is not the case in melanocytic cells and that SOX9, unlike SOX10, is neither required for normal melanocyte stem cell homeostasis nor for formation of congenital nevi and primary melanoma. Our findings disagree with some previously published studies reporting SOX9 expression in the normal and tumor-associated melanocytic lineage [32–34,36,51]. However, as we demonstrate here, most previously used anti-SOX9 antibodies display cross-reactivity with SOX10, owing to the close relationship between these two SoxE factors. Having identified anti-SOX10 and anti-SOX9 specific antibodies, we reveal virtually exclusive expression patterns of these transcription factors in the normal human skin and in a large set of melanoma biopsies and cell lines. While SOX10 expression is restricted to neural crest derivatives, including melanoblasts, differentiated melanocytes, and virtually all human naevus and melanoma biopsies tested (Fig. 1; S3 Fig.; [9], SOX9 expression in melanocytic cells was restricted to few scattered cells in a subset of melanoma biopsies. In contrast, SOX9 was strongly expressed in epithelial cells of the hair follicle, which are devoid of SOX10 expression.
In support of these data, Sox10 protein expression in the mouse skin is detected in vivo throughout all stages of melanocyte development from stem cells to differentiated melanocytes in the hair follicular bulb (Fig. 2). Mice lacking Sox10 in the melanocyte lineage display hair graying, indicating that Sox10 is necessary for maintenance of melanocyte stem cells and committed melanoblasts [23] (Fig. 2). Likewise, Sox10 is required for the establishment of giant congenital naevi as well as melanoma [9]. In contrast, murine Sox9 appeared not to be expressed in melanocytic cells of the normal skin, nevi, and primary melanoma, while it was readily detectable in epithelial cells in accordance with previous reports [33,37,38]. Importantly, loss of function analyses failed to reveal a crucial role of Sox9 in normal melanocytes, as conditional deletion of Sox9 did not affect generation and long-term maintenance of melanocytes in vivo and did not result in hair graying, a phenotype characteristic for the loss of Sox10. Likewise, lack of Sox9 did not prevent emergence of melanocytic lesions induced by oncogenic NRasQ61K (Fig. 3). These data demonstrate that Sox10 and Sox9 are not only expressed in different cellular compartments in the skin, but also play distinct roles in normal and transformed melanocytes.
In cell types other than melanocytes Sox9 and Sox10 can act redundantly. For instance, in oligodendroglial progenitors, concomitant expression of Sox9 can compensate for the loss of Sox10 [52,53]. Similarly, in avian and Xenopus embryos, Sox9 and Sox10 are co-expressed in premigratory neural crest cells and are both able to induce ectopic neural crest cell formation upon forced expression in chicken neural tube [11,24,54]. In addition, the two factors are able to cross-regulate each other at this early stage of neural crest formation. Interfering with Sox10 function leads to inhibition of Sox9 expression [55], suggesting that Sox10 is required for the expression of Sox9 in pre-migratory neural crest. On the other hand, Sox9 overexpression in Xenopus embryos leads to upregulation of Sox10 expression [24], suggesting that Sox9 can also act upstream of Sox10. As development proceeds, however, Sox10 expression persists in the trunk neural crest and is downregulated in cranial neural crest cells giving rise to mesectodermal structures, while Sox9 expression is absent in trunk neural crest cells but present in the cranial neural crest [24,26]. These divergent expression patterns are established by signaling pathways differentially regulating transcription of Sox9 and Sox10, respectively. In particular, TGFβ (transforming growth factor β) simultaneously triggers induction of Sox9 and reduction of Sox10 expression [56]. Accordingly, mice lacking Sox10 display phenotypes that are distinct from those obtained upon loss of Sox9 [13,54,56–58]. In particular, Sox10 but not Sox9 is expressed in and required for the generation of melanoblasts during mouse embryogenesis [13, 21]. Finally, in agreement with our expression studies on human skin, humans carrying mutations in SOX9 display campomelic dysplasia affecting the skeleton and reproductive system but not melanocytes, whereas patients with mutations in SOX10 often exhibit pigmentary anomalies [20,50,59].
However, the divergent functions of Sox10 and Sox9 in the skin appear not to be simply due to their differential expression patterns. Depending on the cellular context, these two transcription factors can also elicit different responses in one and the same cell lineage rather than playing redundant roles. Studies in Xenopus embryos demonstrated that while expression of Sox10 at the two-cell stage was sufficient to activate the expression of Trp-2 (Dct) and the induction of melanocytic precursors, the expression of Sox9 failed to do so [24]. In mice, loss of Sox9 promotes apoptosis and other phenotypes in neural crest cells, but Sox10 is maintained in these cells and cannot rescue the Sox9-mutant phenotype [54]. Moreover, while Sox9 activates the expression of genes involved in the induction of osteochondrogenesis in neural crest cells in pharyngeal arches [26,28,60], Sox10 is involved in the specification of a glial and melanocytic gene expression program [13]. In this context, Sox10 and Sox9 play antagonistic roles, in that Sox9 promotes cells cycle exit and mesenchymal fates, while Sox10 activates proliferation and suppresses mesenchymal fate acquisition [56]. Accordingly, Sox10 inactivation results in induction of Sox9-dependent fates in postmigratory neural crest cells. This interplay between Sox10 and Sox9 functions during normal neural crest development is highly reminiscent of our findings in melanocytic lesions, where Sox10 also promotes proliferation and survival, while Sox9 counteracts these cellular processes. Indeed, in human melanoma cells, loss of SOX10 not only resulted in upregulation of SOX9 expression, but also in global transcriptional changes highly similar to the changes observed upon SOX9 overexpression (Fig. 5), indicating that these factors appear to play opposing functions in melanoma. Interestingly, a study by Passeron et al. revealed that overexpression of SOX9 prevents melanoma formation [35] by increasing the expression of the CDK inhibitor p21 and subsequent cell cycle arrest. Likewise, a recent report by Pavan and colleagues established that the expression of p21 and p27 were increased upon SOX10 knockdown [31]. Thus, our data might provide one explanation for the anti-tumorigenic effect of SOX9, namely by downregulation of SOX10. This is in accordance with our previously published results on the essential role of SOX10 for melanoma initiation and progression [9]. Of note, the anti-tumorigenic effect elicited by suppressing SOX10 was abolished by concomitant SOX9 inactivation both in human melanoma cells as well as in mice. Thus, antagonistic SOX10/SOX9 constitutes a key node in the genetic network underlying melanomagenesis. Nonetheless, it is conceivable that further cues mediate SOX10-pro- and SOX9-anti-tumorigenic effects, respectively. Moreover, our data do not exclude a role of SOX9 at later stages of melanoma disease progression, in particular during metastasis formation by invasive cells. Indeed, while we could attribute a SOX10 high/SOX9 low signature to proliferative human melanoma cell lines and to all human and murine melanoma tissues analyzed, several human melanoma cell lines reported to display invasive features [49] exhibited SOX10 low/SOX9 high expression. Although this remains to be shown, these invasive cell lines with SOX10 low/SOX9 high expression might have been established by capturing or inducing invasive tumor cells that appear to be rather rare in biopsies of bulk tumor tissue. Likewise, apart from experimentally reducing SOX10 levels, other stimuli such as UV exposure might also lead to upregulation of SOX9 [51]. In any case, our discovery of the antagonistic interaction between SOX10 and SOX9, together with the further characterization of their mode of action in melanoma cells, might not only provide new mechanistic insights into how SoxE group proteins are regulated and act in the context of melanoma initiation and maintenance, but might also point to novel strategies for melanoma therapies.
All analyses involving human skin, giant congenital naevi and melanoma tissue were performed in accordance with the ethical committee in canton Zurich, Switzerland. TMA containing melanoma tissue was constructed as previously described [61].
Tyr::NrasQ61K [48] were provided by F. Beermann (EPFL Lausanne, Switzerland). Dct-LacZ mice were described previously [44]. Sox10fl/fl mice were described previously [47]. Sox9fl/fl mice [27] were a kind gift from G. Scherer (Institute of Human Genetics, Freiburg, Germany). Tyr-CreERT2 line [46] was provided by L. Chin (The University of Texas MD Anderson Cancer Center, Houston, Texas, USA). Rosa26-lacZ mice were obtained from Jackson laboratory. All animal experiments were performed in accordance with Swiss law and have been approved by the veterinary authorities of Zurich.
Mice were subjected to intraperitoneal injections of tamoxifen (T5648, Sigma), diluted with the mixture of ethanol and sunflower oil (1:9 ratio). Tamoxifen was injected for 5 consecutive days.
Immunohistochemistry on paraffin sections was performed as previously described [9]. Briefly, skin samples were fixed in 4% buffered paraformaldehyde and embedded in paraffin. For immunohistochemistry, antigen retrieval was performed in citrate buffer (pH 6.0) for 10 minutes at 110°C in HistoPro (Rapid Microwave Histoprocessor, Milestone, USA). The following primary antibodies were used: anti-Sox10 (goat, 1:200, Santa Cruz Biotechnology, Santa Cruz, CA), anti-Sox10 (mouse, 1:200, R&D), anti-Sox9 (rabbit, 1:100, sc-20095, Santa Cruz Biotechnology, Santa Cruz, CA), anti-Sox9 (rabbit, 1:100, ab36748, Abcam), anti-Sox9 (M00006662, Abnova), anti-Sox9 (AB5535, Millipore), anti-Sox9 (GTX 109661, GenTex), anti-MITF (mouse, clone 6D3, 1:500) was a kind gift from Heinz Arnheiter (NIH, USA). Images were captured with a Leica DMI 6000B Microscope and using LAS AF (Leica Application Suite Advanced Fluorescence) software. For whole mount X-Gal staining, skin samples were fixed with 4% buffered paraformaldehyde, washed with PBS and subjected to X-Gal staining solution overnight at 37°C. After several washing steps, tissue was paraffin embedded and sectioned. 5 μm thick sections were further counterstained with eosin solution and mounted.
Total RNA was isolated using Trizol according to manufacturer’s instructions (Invitrogen). 1 μg aliquots of RNA were reverse transcribed with Reverse Transcription System (Promega) according to the manufacturer’s instructions. Data collection and analysis were performed by ABI Viia7 Fast Real-Time PCR Systems (Applied Biosystems). Gene expression values of averaged triplicate reactions were normalized to RPL28 expression levels. RPL28 primers are as follows: 5’-GCAATTGGTTCCGCTACAAC-3’ and 5’-TGTTCTTGCGGATCATGTGT-3’. The expression of SOX10 and SOX9 was measured using primers purchased from QIAGEN: SOX10 (Hs_SOX10_1_SG); SOX9 (Hs_SOX9_1_SG).
Cells derived from patients with giant congenital naevi were sequenced for NRAS. Primers for sequencing for Exon 1 (mutation G12) and Exon 2 (mutation Q61K) of NRAS gene were as follows: NRAS_1F 5’-ATAGAAAGCTTTAAAGTACTG-3’ and NRAS_1R 5’-TTCCTTTAATACAGAATATGG-3’, NRAS_2F 5’-CCCCTTACCCTCCACAC-3’ and NRAS_2R 5’-AACCTAAAACCAACTCTTCCCA-3’.
Silencing RNA (siRNA) transfection was carried out using INTERFERin transfection solution according to the manufacturer’s protocol (Polyplus-transfection, Illkirch, France). Cells were transfected with 10 nM of siRNA (Qiagen) for 96 hours before RNA was extracted or used for FACS analysis. As control siRNA, the All-Star negative siRNA sequence (Qiagen) was used, and gene-specific siRNAs targeting siSOX10 (SI00729414, SI00729421) and siSOX9 (SI00007595, SI00007609) were obtained from Qiagen. Transfection of DNA was carried using JetPEI transfection solution according to the manufacturer’s protocol (Polyplus-transfection, Illkirch, France). Cells were transfected with 1 ug of pCMV6-SOX9 (Origene SC321884) or empty vector for 96 hours before RNA was extracted or used for FACS analysis.
Melanocytes were purified by FACS from doxycycline-treated iDct-GFP mice as previously described [41]. Total RNA was prepared from FACS-sorted cell fractions containing GFP-positive melanoblasts/melanocytes according to standard Illumina RNA-Seq paired-end protocol and sequenced on the Illumina GAIIx to 80 bp per read.
Total RNA was isolated from melanoma cell cultures using TRIzol according to the manufacturer’s instructions (Invitrogen). Total RNA was amplified and biotin-labelled using the Message Amp II-Biotin Enhanced aRNA Amplification Kit (Ambion, Austin, TX, USA). Biotin-labelled RNA was hybridized to Affymetrix HG-U133 plus 2.0 oligonucleotide microarrays following the manufacturer’s protocol (Affymetrix, Santa Clara, CA, USA). After hybridization, microarrays were washed and stained using a GeneChip Fluidics Station 450 (Affymetrix) and then scanned using a GeneChip Scanner 7G (Affymetrix). Raw data was processed by R using the affycoretools package (RMA). Gene expression datasets for SOX10 knockdown were obtained from NCBI GEO GSE37059. Gene expression analysis was performed by R using the limma package. P-values were adjusted by FDR p-value adjustment. For melanoma cell lines analysis (proliferative vs invasive): Normalized expression values were downloaded from GSE4840 containing microarray data for twenty three melanoma cell cultures. Pearson’s product moment correlation (r) was calculated for the SOX10 and SOX9 expression values across all twenty three samples. P-value was determined from the t statistic calculated from r.
Skin tissue (from back skin) was digested with a mixture of Dispase (Roche) and Collagenase I (Worthington) for 1 hour at 37°C and enzymatic reaction was stopped by addition of DMEM media supplemented with 10%FCS as previously described [9]. Subsequently, single cell suspension was filtered through 40 μm strainers (BD). For cell cycle analysis, Click-iT EdU Alexa Fluor 647 Flow Cytometry Assay Kit (Invitrogen) was used. Cells were labeled with PI according to manufacturer’s protocol and DNA content was measured using a BD FACSCanto II flow cytometer (BD Biosciences) and a BD FACSDiva software (BD Biosciences). For measurement of apoptosis, Annexin V-PE Apoptosis Detection Kit I (BD Pharmingen, 559763) was used. FACSAria sorter and FACS DiVa software (BD Biosciences) were used for cell sorting.
ChIP analysis was performed as previously described [62]. Sox9 antibody was from Santa Cruz Biotechnology (sc-20095, Santa Cruz Biotechnology). SOX10 promoter sequences were amplified with forward primer (5’-CCTCTGCCTCGTGTGACTAC-3’) and reverse primer (5’-TCCTGTCTGGAGTGGGCTG-3’).
|
10.1371/journal.pbio.3000210 | To integrate or not to integrate: Temporal dynamics of hierarchical Bayesian causal inference | To form a percept of the environment, the brain needs to solve the binding problem—inferring whether signals come from a common cause and are integrated or come from independent causes and are segregated. Behaviourally, humans solve this problem near-optimally as predicted by Bayesian causal inference; but the neural mechanisms remain unclear. Combining Bayesian modelling, electroencephalography (EEG), and multivariate decoding in an audiovisual spatial localisation task, we show that the brain accomplishes Bayesian causal inference by dynamically encoding multiple spatial estimates. Initially, auditory and visual signal locations are estimated independently; next, an estimate is formed that combines information from vision and audition. Yet, it is only from 200 ms onwards that the brain integrates audiovisual signals weighted by their bottom-up sensory reliabilities and top-down task relevance into spatial priority maps that guide behavioural responses. As predicted by Bayesian causal inference, these spatial priority maps take into account the brain’s uncertainty about the world’s causal structure and flexibly arbitrate between sensory integration and segregation. The dynamic evolution of perceptual estimates thus reflects the hierarchical nature of Bayesian causal inference, a statistical computation, which is crucial for effective interactions with the environment.
| The ability to tell whether various sensory signals come from the same or different sources is essential for forming a coherent percept of the environment. For example, when crossing a busy road at dusk, seeing and hearing an approaching car helps us estimate its location better, but only if its visual image is associated—correctly—with its sound and not with the sound of a different car far away. This is the so-called binding problem, and numerous studies have demonstrated that humans solve this near-optimally as predicted by Bayesian causal inference; however, the underlying neural mechanisms remain unclear. We combined Bayesian modelling, electroencephalography (EEG), and multivariate decoding in an audiovisual spatial localisation task to show that the brain dynamically encodes multiple spatial estimates while accomplishing Bayesian causal inference. First, auditory and visual signal locations are estimated independently; next, information from vision and audition is combined. Finally, from 200 ms onwards, the brain weights audiovisual signals by their sensory reliabilities and task relevance to guide behavioural responses as predicted by Bayesian causal inference.
| In our natural environment, our senses are exposed to a barrage of sensory signals: the sight of a rapidly approaching truck, its looming motor noise, the smell of traffic fumes. How the brain effortlessly merges these signals into a seamless percept of the environment remains unclear. The brain faces two fundamental computational challenges: First, we need to solve the ‘binding’ or ‘causal inference’ problem—deciding whether signals come from a common cause and thus should be integrated or instead be treated independently [1,2]. Second, when there is a common cause, the brain should integrate signals taking into account their uncertainties [3,4].
Hierarchical Bayesian causal inference provides a rational strategy to arbitrate between sensory integration and segregation in perception [2]. Bayesian causal inference explicitly models the potential causal structures that could have generated the sensory signals—i.e., whether signals come from common or independent sources. In line with Helmholtz’s notion of ‘unconscious inference’, the brain is then thought to invert this generative model during perception [5]. In case of a common signal source, signals are integrated weighted in proportion to their relative sensory reliabilities (i.e., forced fusion [3,4,6–10]). In case of independent sources, they are processed independently (i.e., full segregation [11,12]). Iin a particular instance, the brain does not know the world’s causal structure that gave rise to the sensory signals. To account for this causal uncertainty, a final estimate (e.g., object’s location) is obtained by averaging the estimates under the two causal structures (i.e., common versus independent source models) weighted by each causal structure’s posterior probability—a strategy referred to as model averaging (for other decisional strategies, see [13]).
A large body of psychophysics research has demonstrated that human observers combine sensory signals near-optimally as predicted by Bayesian causal inference [2,13–16]. Most prominently, when locating events in the environment, observers gracefully transition between sensory integration and segregation as a function of audiovisual spatial disparity [12]. For small spatial disparities, they integrate signals weighted by their reliabilities, leading to cross-modal spatial biases [17]; for larger spatial disparities, audiovisual interactions are attenuated. A recent functional MRI (fMRI) study showed how Bayesian causal inference is accomplished within the cortical hierarchy [14,16]: While early auditory and visual areas represented the signals on the basis that they were generated by independent sources (i.e., full segregation), the posterior parietal cortex integrated sensory signals into one unified percept (i.e., forced fusion). Only at the top of the cortical hierarchy, in anterior parietal cortex, the uncertainty about the world’s causal structure was taken into account and signals were integrated into a spatial estimate consistent with Bayesian causal inference.
The organisation of Bayesian causal inference across the cortical hierarchy raises the critical question of how these neural computations unfold dynamically over time within a trial. How does the brain merge spatial information that is initially coded in different reference frames and representational formats? Whereas the brain is likely to recurrently update all spatial estimates by passing messages forwards and backwards across the cortical hierarchy [18–20], the unisensory estimates may to some extent precede the computation of the Bayesian causal inference estimate.
To characterise the neural dynamics of Bayesian causal inference, we presented human observers with auditory, visual, and audiovisual signals that varied in their spatial disparity in an auditory and visual spatial localisation task while recording their neural activity with electroencephalography (EEG). First, we employed cross-sensory decoding and temporal generalisation matrices [21] of the unisensory auditory and visual signal trials to characterise the emergence and the temporal stability of spatial representations across the senses. Second, combining psychophysics, EEG, and Bayesian modelling, we temporally resolved the evolution of unisensory segregation, forced fusion, and Bayesian causal inference in multisensory perception.
To determine the computational principles that govern multisensory perception we presented 13 participants with synchronous audiovisual spatial signals (i.e., white noise burst and Gaussian cloud of dots) that varied in their audiovisual spatial disparity and visual reliability (Fig 1A and 1B). On each trial, participants reported their perceived location of either the auditory or the visual signal. In addition, we included unisensory auditory and visual signal trials under auditory or visual report, respectively.
Combining psychophysics, EEG, and computational modelling, we addressed two questions: First, we investigated when and how human observers form spatial representations from unisensory visual or auditory inputs, which generalise across the two sensory modalities. Second, we studied the computational principles and neural dynamics that mediate the integration of audiovisual signals into spatial representations that take into account the observer’s uncertainty about the world’s causal structure consistent with Bayesian causal inference.
Combining psychophysics, multivariate EEG pattern decoding, and computational modelling, we next investigated the computational principles and neural dynamics underlying audiovisual integration of spatial representations using a general linear model (GLM)-based wAV and a Bayesian modelling analysis. As shown in Fig 3, both analyses were applied to the spatial estimates that were either reported by participants (i.e., behaviour, Fig 3B left) or decoded from EEG activity patterns independently for each poststimulus time point (i.e., neural, Fig 3B right, for further details, see the Methods section and the Fig 3 legend).
The GLM-based wAV analysis quantifies the influence of the true auditory and true visual location on (1) the reported or (2) EEG decoded auditory and visual spatial estimates in terms of an audiovisual weight index wAV.
The Bayesian modelling analysis formally assessed the extent to which (2) the full-segregation model(s) (Fig 3C, encircled in light blue, red or green), (2) the forced-fusion model (Fig 3C, yellow), and (3) the Bayesian causal inference model (i.e., using model averaging as decision function, encircled in dark blue; see supporting material S1 Table for other decision functions) can account for the spatial estimates reported by observers (i.e., behaviour) or decoded from EEG activity pattern (i.e., neural).
Integrating information from vision and audition into a coherent representation of the space around us is critical for effective interactions with the environment. This EEG study temporally resolved the neural dynamics that enable the brain to flexibly integrate auditory and visual signals into spatial representations in line with the predictions of Bayesian causal inference.
Auditory and visual senses code spatial location in different reference frames and representational formats [26]. Vision provides spatial information in eye-centred and audition in head-centred reference frames [27,28]. Furthermore, spatial location is directly coded in the retinotopic organisation in primary visual cortex [29], whereas spatial location in audition is computed from sound latency and amplitude differences between the ears, starting in the brainstem [27]. In auditory cortices of primates, spatial location is thought to be represented by neuronal populations with broad tuning functions [30,31]. In order to merge spatial information from vision and audition, the brain thus needs to establish coordinate mappings and/or transform spatial information into partially shared ‘hybrid’ reference frames, as previously suggested by neurophysiological recordings in nonhuman primates [30,32]. In the first step, we therefore investigated the neural dynamics of spatial representations encoded in EEG activity patterns separately for unisensory auditory and visual signals using the method of temporal generalisation matrices [21]. In vision, spatial location was encoded initially at 60 ms in transient neural activity associated with the early P1 and N1 components and then turned into temporally more stable representations from 200 ms and particularly from 350 ms (Fig 2, upper right quadrant, S2 Fig). In audition, spatial location was encoded by relatively stable EEG activity from 95 ms and particularly from 250 ms, which is associated with the auditory long latency P2 component [22–24] (S3 Fig).
Activity patterns encoding spatial location generalised not only across time but also across sensory modalities between 160 and 360 ms. As indicated in Fig 2, SVR models trained on visual-evoked responses generalised to auditory-evoked responses and vice versa (upper left and lower right quadrant, significant cross-sensory generalisation encircled by thick grey line). These results suggest that unisensory auditory and visual spatial locations are initially represented by transient and modality-specific activity patterns. Later, at about 200 ms, they are transformed into temporally more stable representations that may rely on neural sources in frontoparietal cortices that are at least to some extent shared between auditory and visual modalities [22,33,34].
Next, we asked when and how the human brain combines spatial information from vision and audition into a coherent representation of space. The brain should integrate sensory signals only when they come from a common event but should segregate signals from independent events [1,2,12]. To investigate how the brain arbitrates between sensory integration and segregation, we presented observers with synchronous audiovisual signals that varied in their spatial disparity across trials. On each trial, observers reported either the auditory or the visual location. Our results show that a concurrent yet spatially disparate visual signal biased observers’ perceived sound location towards the visual location—a phenomenon coined spatial ventriloquist illusion [17,35]. Consistent with reliability-weighted integration, this audiovisual spatial bias was significantly stronger when the visual signal was more reliable (Fig 1C left, grey solid versus dashed lines). Furthermore, observers reported different locations for auditory and visual signals, and this difference was even greater for large- relative to small-spatial-disparity trials. This significant interaction between spatial disparity and task relevance indicates that human observers arbitrate between sensory integration and segregation depending on the probabilities of different causal structures of the world that can be inferred from audiovisual spatial disparity.
Using EEG, we then investigated how the brain forms neural spatial representations dynamically post stimulus. Our analysis of the neural audiovisual weight index wAV shows that the spatial estimates decoded from EEG activity patterns are initially dominated by visual inputs (i.e., wAV close to 90°). This visual dominance is most likely explained by the retinotopic representation of visual space that facilitates EEG decoding of space leading to visual predominance (for further discussion, see the Methods section). From about 65 ms onwards, visual reliability significantly influenced wAV (Fig 4A): as expected, the location of the visual signal exerted a stronger influence on the spatial estimate decoded from EEG activity patterns when the visual signal was more reliable than unreliable. By contrast, the signal’s task relevance influenced the audiovisual weight index only later, from about 190 ms (Fig 4B). Thus, visual reliability as a bottom-up stimulus-bound factor impacted the sensory weighting in audiovisual integration prior to top-down effects of task relevance. We observed a significant interaction between task relevance and spatial disparity as the characteristic profile for Bayesian causal inference from about 310 ms: the difference in wAV between auditory and visual report was significantly greater for large- than for small-disparity trials (Fig 4D, Table 2). Thus, spatial disparity determined the influence of task-irrelevant signals on the spatial representations encoded in EEG activity from about 310 ms onwards. A task-irrelevant signal influenced the spatial representations mainly when auditory and visual signals were close in space and hence likely to come from a common event, but it had minimal influence when they were far apart in space. Collectively, our statistical analysis of the audiovisual weight index revealed a sequential emergence of visual dominance, reliability weighting (from about 100 ms), effects of task relevance (from about 200 ms), and finally the interaction between task relevance and spatial disparity (from about 310 ms, Fig 4A–4D).
This multistage process was also mirrored in the time course of exceedance probabilities furnished by our formal Bayesian model comparison: The unisensory visual segregation (SegV) model was the winning model for the first 100 ms, thereby modelling the early visual dominance. The audiovisual forced-fusion model embodying reliability-weighted integration dominated the time interval of 100–250 ms. Finally, the Bayesian causal inference model that enables the arbitration between sensory integration and segregation depending on spatial disparity outperformed all other models from 350 ms onwards. Hence, both our Bayesian modelling analysis and our wAV analysis showed that the hierarchical structure of Bayesian causal inference is reflected in the neural dynamics of spatial representations decoded from EEG. The Bayesian causal inference model also outperformed the audiovisual full-segregation (SegV,A) model that enables the representation of the location of the task-relevant stimulus unaffected by the location of the task-irrelevant stimulus. Instead, our Bayesian modelling analysis confirmed that from 350 ms onwards, the brain integrates audiovisual signals weighted by their bottom-up reliability and top-down task relevance into spatial priority maps [36,37] that take into account the probabilities of the different causal structures consistent with Bayesian causal inference. The spatial priority maps were behaviourally relevant for guiding spatial orienting and actions, as indicated by the correlation between the neural and behavioural audiovisual weight indices, which progressively increased from 100 ms and culminated at about 300–400 ms. Two recent studies have also demonstrated such a temporal evolution of Bayesian causal inference in an audiovisual temporal numerosity judgement task [38] and an audiovisual rate categorisation task [39].
The timing and the parietal-dominant topographies of the AV potentials (see S2 and S3 Figs) that form the basis for our spatial decoding (and hence for wAV and Bayesian modelling analyses) closely match the P3b component (i.e., a subcomponent of the classical P300). Although it is thought that the P3b relies on neural generators located mainly in parietal cortices [40,41], its specific functional role remains controversial [42]. Given its sensitivity to stimulus probability [43–45] and discriminability [46] as well as task context [42,47,48], it was proposed to reflect neural processes involved in transforming sensory evidence into decisions and actions [49]. Most recent research has suggested that the P3b may sustain processes of evidence accumulation [50] that are influenced by observers’ prior [51], incoming evidence (i.e., likelihood [52]), and observers’ belief updating [53]. Likewise, our supplementary time-frequency analyses revealed that alpha/beta power, which has previously been associated with the generation of the P3b component [54], depended on bottom-up visual reliability between 200 and 400 ms and top-down task relevance between 350 and 550 ms post stimulus (see S5 Fig, S2 Table and S1 Text), thereby mimicking the temporal evolution of bottom-up and top-down influences observed in our main wAV and Bayesian modelling analysis.
Yet, our main analysis took a different approach. Rather than focusing on the effects of visual reliability, task relevance/attention, and spatial disparity directly on event-related potentials (ERPs) or time-frequency power, the wAV analysis investigated how these manipulations affect the spatial representations encoded in EEG activity patterns, and the Bayesian modelling analysis accommodated those effects directly in the computations of Bayesian causal inference. Along similar lines, two recent fMRI studies characterised the computations involved in integrating audiovisual spatial inputs across the cortical hierarchy [14,16]: whereas low level auditory and visual areas predominantly encoded the unisensory auditory or visual locations (i.e., full-segregation model) [55–64], higher-order visual areas and posterior parietal cortices combined audiovisual signals weighted by their sensory reliabilities (i.e., forced-fusion model) [65–68]. Only at the top of the hierarchy, in anterior parietal cortices, did the brain integrate sensory signals consistent with Bayesian causal inference. Thus, the temporal evolution of Bayesian causal inference observed in our current EEG study mirrored its organisation across the cortical hierarchy observed in fMRI.
Fusing the results from EEG and fMRI studies (see caveats in the Methods section) thus suggests that Bayesian causal inference in multisensory perception relies on dynamic encoding of multiple spatial estimates across the cortical hierarchy. During early processing, multisensory perception is dominated by full-segregation models associated with activity in low-level sensory areas. Later audiovisual interactions that are governed by forced-fusion principles rely on posterior parietal areas. Finally, Bayesian causal inference estimates are formed in anterior parietal areas. Yet, although our results suggest that full segregation, forced fusion, and Bayesian causal inference dominate EEG activity patterns at different latencies, they do not imply a strictly feed-forward architecture. Instead, we propose that the brain concurrently accumulates evidence about the different spatial estimates and the underlying causal structure (i.e., common versus independent sources) most likely via multiple feedback loops across the cortical hierarchy [18,19]. Only after 350 ms is a final perceptual estimate formed in anterior parietal cortices that takes into account the uncertainty about the world’s causal structure and combines audiovisual signals into spatial priority maps as predicted by Bayesian causal inference.
Sixteen right-handed participants participated in the experiment; three of those participants did not complete the entire experiment: two participants were excluded based on eye tracking results from the first day (the inclusion criterion was less than 10% of trials rejected because of eye blinks or saccades; see the Eye movement recording and analysis section for details), and one participant withdrew from the experiment. The remaining 13 participants (7 females, mean age = 22.1 years; SD = 3.0) completed the 3-day experiment and are thus included in the analysis. All participants had no history of neurological or psychiatric illnesses, had normal or corrected-to-normal vision, and had normal hearing.
All participants gave informed written consent to participate in the experiment. The study was approved by the research ethics committee of the University of Birmingham (approval number: ERN_11_0470AP4) and was conducted in accordance with the principles outlined in the Declaration of Helsinki.
The visual (‘V’) stimulus was a cloud of 20 white dots (diameter = 0.43° visual angle, stimulus duration: 50 ms) sampled from a bivariate Gaussian distribution with vertical standard deviation of 2° and horizontal standard deviation of 2° or 12° visual angle presented on a dark grey background (67% contrast). Participants were told that the 20 dots were generated by one underlying source in the centre of the cloud. The visual cloud of dots was presented at one of four possible locations along the azimuth (i.e., −10°, −3.3°, 3.3°, or 10°).
The auditory (‘A’) stimulus was a 50-ms-long burst of white noise with a 5-ms on/off ramp. Each auditory stimulus was delivered at a 75-dB sound pressure level through one of four pairs of two vertically aligned loudspeakers placed above and below the monitor at four positions along the azimuth (i.e., −10°, −3.3°, 3.3°, or 10°). The volumes of the 2 × 4 speakers were carefully calibrated across and within each pair to ensure that participants perceived the sounds as emanating from the horizontal midline of the monitor.
In a spatial ventriloquist paradigm, participants were presented with synchronous, spatially congruent or disparate visual and auditory signals (Fig 1A and 1B). On each trial, visual and auditory locations were independently sampled from four possible locations along the azimuth (i.e., −10°, −3.3°, 3.3°, or 10°), leading to four levels of spatial disparity (i.e., 0°, 6.6°, 13.3°, or 20°; i.e., as indicated by the greyscale in Fig 1A). In addition, we manipulated the reliability of the visual signal by setting the horizontal standard deviation of the Gaussian cloud to a 2° (high reliability) or 14° (low reliability) visual angle. In an intersensory selective-attention paradigm, participants reported either their auditory or visual perceived signal location and ignored signals in the other modality. For the visual modality, they were asked to determine the location of the centre of the visual cloud of dots. Hence, the 4 × 4 × 2 × 2 factorial design manipulated (1) the location of the visual stimulus (−10°, −3.3°, 3.3°, 10°; i.e., the mean of the Gaussian), (2) the location of the auditory stimulus (−10°, −3.3°, 3.3°, 10°), (3) the reliability of the visual signal (2°, 14°; SD of the Gaussian), and (4) task relevance (auditory-/visual-selective report), resulting in 64 conditions (Fig 1A). To characterise the computational principles of multisensory integration, we reorganised these conditions into a 2 (visual reliability: high versus low) × 2 (task relevance: auditory versus visual report) × 2 (spatial disparity: ≤6.6° versus >6.6°) factorial design for the statistical analysis of the behavioural and EEG data. In addition, we included 4 (locations: −10°, −3.3°, 3.3°, or 10°) × 2 (visual reliability: high, low) unisensory visual conditions and 4 (locations: −10°, −3.3°, 3.3°, or 10°) unisensory auditory conditions. We did not manipulate auditory reliability, because the reliability of auditory spatial information is anyhow limited. Furthermore, the manipulation of visual reliability is sufficient to determine reliability-weighted integration as a computational principle and arbitrate between the different multisensory integration models (see Bayesian modelling analysis section).
On each trial, synchronous audiovisual, unisensory visual, or unisensory auditory signals were presented for 50 ms, followed by a response cue 1,000 ms after stimulus onset (Fig 1B). The response was cued by a central pure tone (1,000 Hz) and a blue colour change of the fixation cross presented in synchrony for 100 ms. Participants were instructed to withhold their response and avoid blinking until the presentation of the cue. They fixated on a central cross throughout the entire experiment. The next stimulus was presented after a variable response interval of 2.6–3.1 s.
Stimuli and conditions were presented in a pseudo-randomised fashion. The stimulus type (bisensory versus unisensory) and task relevance (auditory versus visual) was held constant within a run of 128 trials. This yielded four run types: audiovisual with auditory report, audiovisual with visual report, auditory with auditory report, and visual with visual report. The task relevance of the sensory modality in a given run was displayed to the participant at the beginning of the run. Furthermore, across runs we counterbalanced the response hand (i.e., left versus right hand) to partly dissociate spatial processing from motor responses. The order of the runs was counterbalanced across participants. All conditions within a run were presented an equal number of times. Each participant completed 60 runs, leading to 7,680 trials in total (3,840 auditory and 3,840 visual localisation tasks—i.e., 96 trials for each of the 76 conditions were included in total; apart from the four unisensory auditory conditions that included 192 trials). The runs were performed across 3 days with 20 runs per day. Each day was started with a brief practice run.
Stimuli were presented using Psychtoolbox version 3.0.11 [69] (http://psychtoolbox.org/) under MATLAB R2014a (MathWorks) on a desktop PC running Windows 7. Visual stimuli were presented via a gamma-corrected 30” LCD monitor with a resolution of 2,560 × 1,600 pixels at a frame rate of 60 Hz. Auditory stimuli were presented at a sampling rate of 44.1 kHz via eight external speakers (Multimedia) and an ASUS Xonar DSX sound card. Exact audiovisual onset timing was confirmed by recording visual and auditory signals concurrently with a photodiode and a microphone. Participants rested their head on a chin rest at a distance of 475 mm from the monitor and at a height that matched participants’ ears to the horizontal midline of the monitor. Participants responded by pressing one of four response buttons on a USB keypad with their index, middle, ring, and little finger, respectively.
To address potential concerns that results were confounded by eye movements, we recorded participants’ eye movements. Eye recordings were calibrated in the recommended field of view (32° horizontally and 24° vertically) for the EyeLink 1000 Plus system with the desktop mount at a sampling rate of 2,000 Hz. Eye position data were on-line parsed into events (saccade, fixation, eye blink) using the EyeLink 1000 Plus software. The ‘cognitive configuration’ was used for saccade detection (velocity threshold = 30°/sec, acceleration threshold = 8,000°/sec2, motion threshold = 0.15°) with an additional criterion of radial amplitude larger than 1°. Individual trials were rejected if saccades or eye blinks were detected from −100 to 700 ms post stimulus.
Participants’ stimulus localisation accuracy was assessed as the Pearson correlation between their location responses and the true signal source location separately for unisensory auditory, visual high reliability, and visual low reliability conditions. To confirm whether localisation accuracy in vision exceeded performance in audition in both visual reliabilities, we performed Monte-Carlo permutation tests. Specifically, we entered the subject-specific Fisher z-transformed Pearson correlation differences between vision and audition (i.e., visual–auditory) separately for the two visual reliability levels into a Monte-Carlo permutation test at the group level based on the one-sample t statistic with 5,000 permutations [70].
Continuous EEG signals were recorded from 64 channels using Ag/AgCl active electrodes arranged in a 10–20 layout (ActiCap, Brain Products GmbH, Gilching, Germany) at a sampling rate of 1,000 Hz, referenced at FCz. Channel impedances were kept below 10 kΩ.
Preprocessing was performed with the FieldTrip toolbox [71] (http://www.fieldtriptoolbox.org/). For the decoding analysis, raw data were high-pass filtered at 0.1 Hz, re-referenced to average reference, and low-pass filtered at 120 Hz. Trials were extracted with a 100-ms prestimulus and 700-ms poststimulus period and baseline corrected by subtracting the average value of the interval between −100 and 0 ms from the time course. Trials were then temporally smoothed with a 20-ms moving window and downsampled to 200 Hz (note that a 20-ms moving average is comparable to a finite impulse response [FIR] filter with a cutoff frequency of 50 Hz). Trials containing artefacts were rejected based on visual inspection. Furthermore, trials were rejected if (1) they included eye blinks, (2) they included saccades, (3) the distance between eye fixation and the central fixation cross exceeded 2°, (4) participants responded prior to the response cue, or (5) there was no response. For ERPs (S2 and S3 Figs), the preprocessing was identical to the decoding analysis, except that a 45-Hz low-pass filter was applied without additional temporal smoothing with a temporal moving window. Grand average ERPs were computed by averaging all trials for each condition first within each participant and then across participants.
For the multivariate pattern analyses, we computed ERPs by averaging over sets of eight randomly assigned individual trials from the same condition. To characterise the temporal dynamics of the spatial representations, we trained linear SVR models (LIBSVM [72], https://www.csie.ntu.edu.tw/~cjlin/libsvm/) to learn the mapping from ERP activity patterns of the (1) unisensory auditory (for auditory decoding), (2) unisensory visual (for visual decoding), or (3) audiovisual congruent conditions (for audiovisual decoding) to external spatial locations separately for each time point (every 5 ms) over the course of the trial (S2, S3 and S4 Figs). All SVR models were trained and evaluated in a 12-fold-stratified cross-validation (12 ERPs/fold) procedure with default hyperparameters (C = 1, ε = 0.001). The specific training and generalisation procedures were adjusted to the scientific questions (see the Shared and distinct neural representations of space across vision and audition section and the GLM analysis of audiovisual weight index wAV section for details).
Combining psychophysics, computational modelling, and EEG, we addressed two questions: First, focusing selectively on the unisensory auditory and unisensory visual conditions, we investigated when spatial representations are formed that generalise across auditory and visual modalities. Second, focusing on the audiovisual conditions, we investigated when and how human observers integrate audiovisual signals into spatial representations that take into account the observer’s uncertainty about the world’s causal structure consistent with Bayesian causal inference. In the following sections, we will describe the analysis approaches to address these two questions in turn.
First, we investigated how the brain forms spatial representations in either audition or vision using the so-called temporal generalisation method [21]. Here, the SVR model is trained at time point t to learn the mapping from, e.g., unisensory visual (or auditory) ERP pattern to external stimulus location. This learnt mapping is then used to predict spatial locations from unisensory visual (or auditory) ERP activity patterns across all other time points. Training and generalisation were applied separately to unisensory auditory and visual ERPs. To match the number of trials for auditory and visual conditions, we applied this analysis to the visual ERPs pooled over the two levels of visual reliability. The decoding accuracy as quantified by the Pearson correlation coefficient between the true and decoded stimulus locations is entered into a training time × generalisation time matrix. The generalisation ability across time illustrates the similarity of EEG activity patterns relevant for encoding features (i.e., here: spatial location) and has been proposed to assess the stability of neural representations [21]. In other words, if stimulus location is encoded in EEG activity patterns that are stable (or shared) across time, then an SVR model trained at time point t will be able to correctly decode stimulus location from EEG activity patterns at other time points. By contrast, if stimulus location is represented by transient or distinct EEG activity patterns across time, then an SVR model trained at time point t will not be able to decode stimulus location from EEG activity patterns at other time points. Hence, entering Pearson correlation coefficients as a measure for decoding accuracy for all combinations of training and test time into a temporal generalisation matrix has been argued to provide insights into the stability of neural representations whereby the spread of significant decoding accuracy to off-diagonal elements of the matrix indicates temporal generalisability or stability [21].
Second, to examine whether and when neural representations are formed that are shared across vision and audition, we generalised to ERP activity patterns across time not only from the same sensory modality but also from the other sensory modality (i.e., from vision to audition and vice versa). This cross-sensory generalisation reveals neural representations that are shared across sensory modalities.
To assess whether decoding accuracies were better than chance, we entered the subject-specific matrices of the Fisher z-transformed Pearson correlation coefficients into a between-subjects Monte-Carlo permutation test using the one-sample t statistic with 5,000 permutations ([70], as implemented in the FieldTrip toolbox). To correct for multiple comparisons within the two-dimensional (i.e., time × time) data, cluster-level inference was used based on the maximum of the summed t values within each cluster (‘maxsum’) with a cluster-defining threshold of p < 0.05, and a two-tailed p-value was computed.
To characterise how human observers integrate auditory and visual signals into spatial representations at the behavioural and neural levels, we developed a GLM-based analysis of an audiovisual weight index wAV and a Bayesian modelling analysis that we applied to both (1) the reported auditory and visual spatial estimates (i.e., participants’ behavioural localisation responses) and (2) the neural spatial estimates decoded from EEG activity pattern evoked by audiovisual stimuli (see Fig 3 and [14,16]).
|
10.1371/journal.ppat.1002942 | ABO Blood Groups Influence Macrophage-mediated Phagocytosis of Plasmodium falciparum-infected Erythrocytes | Erythrocyte polymorphisms associated with a survival advantage to Plasmodium falciparum infection have undergone positive selection. There is a predominance of blood group O in malaria-endemic regions, and several lines of evidence suggest that ABO blood groups may influence the outcome of P. falciparum infection. Based on the hypothesis that enhanced innate clearance of infected polymorphic erythrocytes is associated with protection from severe malaria, we investigated whether P. falciparum-infected O erythrocytes are more efficiently cleared by macrophages than infected A and B erythrocytes. We show that human macrophages in vitro and mouse monocytes in vivo phagocytose P. falciparum-infected O erythrocytes more avidly than infected A and B erythrocytes and that uptake is associated with increased hemichrome deposition and high molecular weight band 3 aggregates in infected O erythrocytes. Using infected A1, A2, and O erythrocytes, we demonstrate an inverse association of phagocytic capacity with the amount of A antigen on the surface of infected erythrocytes. Finally, we report that enzymatic conversion of B erythrocytes to type as O before infection significantly enhances their uptake by macrophages to observed level comparable to that with infected O wild-type erythrocytes. These data provide the first evidence that ABO blood group antigens influence macrophage clearance of P. falciparum-infected erythrocytes and suggest an additional mechanism by which blood group O may confer resistance to severe malaria.
| Plasmodium falciparum malaria is considered to be one of the strongest forces for evolutionary selection pressure on the human genome. Different red blood cell variants associated with a survival advantage to P. falciparum infection have undergone positive selection. Blood group O is found more frequently in malaria-endemic regions and has been associated with protection against severe malaria and death. However the biological basis of protection remains unclear. In this study, we investigated innate immune clearance of P. falciparum-infected erythrocytes by macrophages as a possible mode of protection. We show that macrophages clear P. falciparum-infected O erythrocytes more avidly than infected A and B erythrocytes. We also report that enzymatic conversion of infected blood group B red cells to type as “O” like erythrocytes significantly enhances their uptake by macrophages to a level comparable to that observed with infected O wild type erythrocytes. These data provide the first evidence that clearance of P. falciparum-infected erythrocytes is influenced by human ABO blood groups and suggest a new mechanism by which blood group O may contribute to protection against severe malaria.
| Plasmodium falciparum malaria is responsible for an estimated 1.24 million deaths annually, with the majority of deaths occurring in individuals before reproductive age [1]. P. falciparum malaria predated the development of modern Homo sapiens and has co-evolved with human populations [2], [3]. It is considered to be one of the strongest forces for evolutionary selection of the human genome [2], [3]. In populations where P. falciparum infection is highly prevalent, common erythrocyte polymorphisms, such as deficiencies in globin synthesis, membrane proteins and erythrocyte enzymes, are associated with protection against severe and fatal disease [4]–[6].
Recent evidence suggests that the ABO blood group system has also been subject to malaria-related selection pressure [2], [3]. The ABO phenotype is determined by a polymorphic gene that encodes an enzyme, ABO glycosyltransferase that conjugates A- or B-specific sugar residues onto the precursor molecule known as the H antigen. If functionally active ABO glycosyltransferase is inherited via the co-dominant A or B alleles, transfer of either α-1,3-linked N-acetylgalactosamine or α-1,3-linked galactose produces A and B antigens, respectively, resulting in the non-O blood groups (A, B and AB). Molecular evidence indicates that the predominant O allele arose as the result of a loss-of-function mutation at nucleotide position 261 [7], [8]. Consequently, in O erythrocytes, the H antigen is left unaltered and ends with an α-1,2-linked fucose moiety that lacks the terminal α-1,3-linked monosaccharides [9].
Host pathogen interactions have been proposed as an important evolutionary force shaping the global distribution of ABO blood groups [10]. There is strong epidemiological evidence that the ABO phenotype may modulate disease severity and outcome of P. falciparum malaria, with blood groups A and B associated with increased disease severity compared to blood group O [11]–[15]. This association is consistent with the higher prevalence of group O observed in malaria-endemic sub-Saharan Africa compared to many parts of the world where malaria is not endemic, suggesting that blood group O may be a selected, protective adaptation against severe and fatal infection [2], [16], [17].
While several studies have reported that individuals with blood groups A and B are more likely to develop severe malaria, the mechanisms underlying the putative protection afforded by blood group O remain unclear [11], [12], [15]. Proposed mechanisms of protection parallel those implicated in other erythrocyte polymorphisms and include decreased erythrocyte invasion and reduced erythrocyte rosetting [12], [18]. Several studies have examined the association of ABO blood groups with rosetting and have reported that infected O erythrocytes exhibit fewer or smaller rosettes in vitro [12], [19]–[23] and in vivo [24] than infected A and B erythrocytes. Decreased rosetting may reduce microvascular obstruction that is believed to contribute to the pathogenesis of severe disease [12], [13], [16]. However, alternative mechanisms may also exist that contribute to the protective effect to P. falciparum malaria observed in individuals with blood group O.
P. falciparum infection and intracellular growth induce profound changes to the erythrocyte membrane resembling red cell senescence [25]. These changes, including hemichrome formation and band 3 aggregation resulting in erythrophagocytosis, may be accelerated in the presence of underlying erythrocyte disorders [26], [27]. Based on these observations, we hypothesized that enhanced senescence and phagocytosis of infected O erythrocytes, resulting in improved innate clearance and lower parasite densities, may provide an alternative explanation for protection observed in blood group O individuals. Studies of other erythrocyte polymorphisms associated with malaria-endemic areas, including sickle cell trait, beta-thalassemia, G6PD trait, and pyruvate kinase deficiency [26]–[28] that have reported increased phagocytosis of P. falciparum-infected variant erythrocytes, are also consistent with this hypothesis. Here, we show that P. falciparum parasites invade and mature similarly in group A, B and O erythrocytes. However, compared to P. falciparum-infected A and B erythrocytes, infected O erythrocytes display enhanced hemichrome deposition, band 3 aggregation and increased macrophage phagocytosis in vitro and in vivo. These data suggest that enhanced clearance of infected O erythrocytes may represent a novel mechanism by which blood group O contributes to protection against severe malaria.
To determine if ABO polymorphism influences malaria parasite invasion and maturation, we examined P. falciparum (ITG clone) parasite invasion and development in A, B, and O erythrocytes in vitro. No statistically significant differences in parasite invasion of group A, B and O erythrocytes were observed during two growth cycles (Figure 1). In addition, there were no significant differences noted in intracellular maturation (from ring stage to trophozoite stage parasites) within A, B and O erythrocytes (Figure 1). Similar results were obtained using two other clones (3D7, E8B) of P. falciparum malaria (data not shown).
To test the hypothesis that infected O erythrocytes may be preferentially phagocytosed compared to infected A or B erythrocytes, infected A, B and O erythrocytes at ring- or mature-stage were co-incubated with human monocyte-derived macrophages for 2 hours and the phagocytic index was determined blinded to the erythrocyte blood group. These data were then normalized to the average phagocytic index of infected A erythrocytes. The phagocytic uptake of ring-stage infected O erythrocytes (Figure 2A; 1.43±0.16, mean±SEM) was significantly higher than ring-stage infected A (1.09±0.08, p = 0.022) and B erythrocytes (0.75±0.076, p = 0.007). Similarly, the mean uptake of mature-stage infected O erythrocytes (Figure 2B; 2.3±0.29, mean±SEM) was significantly greater than the uptake of infected A (1.0±0.07, p = 0.001) and infected B (1.02±0.16, p = 0.026) erythrocytes. By contrast there were no significant differences observed in the uptake of control uninfected A, B, or O erythrocytes (Figure 2A, B). Similar results were observed with parasite clone 3D7 (not shown) suggesting that these were not a P. falciparum clone-specific phenomenon. We next examined the influence of different blood group macrophage donors on infected erythrocyte uptake at mature-stage. No significant differences were observed in the preferential uptake of infected O erythrocytes by macrophage donors of A versus O blood group (Figure 2C, p = 0.568, two-way ANOVA, main effect: macrophage donor). Macrophages obtained from either group O or A donors displayed enhanced phagocytosis of P. falciparum-infected O erythrocytes compared to infected A erythrocytes (Figure 2C, p = 0.002, two-way ANOVA, main effect: blood group). These data indicate that infected O erythrocytes are preferentially cleared by macrophages independent of the macrophage donor blood group.
Previous studies have established that murine macrophages express CD36 and bind and mediate the update of P. falciparum infected erythrocytes. This model has proven to be useful to investigate the molecular basis for the interaction of CD36 with malaria-infected erythrocytes as well as the interactions of other pattern recognition receptors such as TLRs with their respective cognate ligands in vitro [29], [30] and in vivo [31]. We used a previously established murine model system [31] to investigate phagocytosis of P. falciparum-infected A, B and O erythrocytes in vivo by macrophages in the peritoneal cavity of C57BL/6 mice. Three hours after intraperitoneal injection with infected or uninfected (as controls) A, B and O erythrocytes, peritoneal lavage was performed, and phagocytosis of infected and uninfected erythrocytes by peritoneal macrophages was quantified. In agreement with our in vitro observations, the phagocytic uptake of infected O erythrocytes (4.97±1.04, mean±SEM) in vivo was 3- to 4-fold greater than the phagocytic uptake of infected A (1.00±0.08, p = 0.01) or infected B (1.49±0.29, p = 0.04) erythrocytes (Figure 3A, B). There were no significant differences observed in the uptake of uninfected A, B and O erythrocytes (Figure 3A).
There are various subgroups of each ABO blood group, with the two most common subgroups of A being A1 and A2. The differences between them are both qualitative and quantitative: A antigens are more complex on erythrocytes of the A1 phenotype and have an approximate 5× greater antigen site density than on A2 cells, thus leaving the latter with a greater number of unmodified H antigens [32], [33]. The quantitative spectrum of H antigen expression is therefore O>A2>A1>Bombay (where the latter is a genetically H-deficient phenotype) (Figure 4A) [8], [9], [34]. Based on our above observations, we postulated that an increase in H antigen expression, and a corresponding decrease in A antigen levels, would be associated with increased macrophage phagocytosis of infected erythrocytes. Within the spectrum of H antigen expression, we also predicted that phagocytosis of infected A2 erythrocytes would be greater than that of infected A1 erythrocytes, but less than infected O erythrocytes. In agreement with this hypothesis, we observed a dose-dependent effect with presumed H antigen expression on the phagocytosis of P. falciparum infected cells, with phenotypically less A antigen expression (or increasing H antigen expression) associated with increased phagocytosis (Figure 4B; Spearman's correlation, r = 0.80, p<0.0001).
To confirm that increased relative H antigen expression and decreased A/B antigen expression was associated with enhanced phagocytosis of infected O erythrocytes, we enzymatically converted B erythrocytes to type as O erythrocytes using a specific glycosidase, B-zyme (Bacteroides fragilis, α-galactosidase). This enzyme has been shown to efficiently convert virtually all B antigen on B erythrocytes to H antigen at neutral pH, without changing the sialic acid content on the cell membrane [35]. By treating B erythrocytes with this α-galactosidase, we cleaved the terminal α1–3-linked galactose residues responsible for blood group B specificity, converting these cells serologically to type as blood group O [35]. As controls, untreated group B erythrocytes were treated with the conversion buffer alone and control O erythrocytes were treated with the conversion buffer plus B-zyme. Efficient enzyme conversion of blood group B to O was achieved as demonstrated by flow cytometric analysis (Figure 5A). Positive reactions with anti-B were also demonstrated serologically with B erythrocytes while negative reactions were obtained with O erythrocytes and B erythrocytes converted with B-zyme.
Consistent with our previous findings, there was increased uptake of infected untreated O erythrocytes when compared to infected untreated B erythrocytes (Figure 5B, p = 0.042, Mann-Whitney with Bonferroni correction for multiple comparisons). Moreover, the phagocytosis of infected B erythrocytes was significantly enhanced following cleavage of the B antigen (p = 0.008, Figure 5B) and comparable to mock-treated and infected O erythrocytes. The phagocytic index of infected O erythrocytes was not significantly altered by B-zyme treatment (p>0.05). Additionally, there was no significant difference observed in the uptake of control uninfected, treated or untreated, B and O erythrocytes (Figure 5B).
Phagocytic recognition and clearance of senescent erythrocytes has been reported to depend, at least partly, on increased expression of erythrocyte senescence antigens such as phosphatidylserine (PS) and aggregated band 3 [36]. Increased outer leaflet exposure of PS has been associated with enhanced macrophage erythrophagocytosis [36]. To examine potential mechanism(s) underlying enhanced uptake of infected O erythrocytes, we initially investigated P. falciparum-induced PS exposure by comparing Annexin-V staining on infected O, A, and B erythrocytes (Figure 6). Although infected erythrocytes had increased PS expression compared to uninfected erythrocytes, there were no significant differences observed in PS levels on infected O compared to infected A and B erythrocytes (Figure 6).
In contrast, ABO blood groups influenced P. falciparum-induced hemichrome deposition and band 3 aggregation. Figures 7A and B shows the presence of membrane-bound hemichromes in uninfected and infected A, B, and O erythrocytes. No differences in hemichrome levels were observed in uninfected erythrocytes maintained in the same culture conditions as infected erythrocytes. However, increased hemichrome deposition was detected in infected ring-stage (p = 0.005 and p = 0.038, Figure 7A) and mature-stage (p = 0.013 and p = 0.024, Figure 7B) O erythrocytes, compared to infected A and B erythrocytes, respectively.
Since hemichromes are thought to bind to and oxidize the band 3 cytoplasmic domain inducing band 3 clustering [25], [36], we compared band 3 aggregation in infected O, A and B erythrocytes using gel filtration chromatography (Figure 7C) and eosin-5-maleimide fluorescence (Figure 7C insert), a specific label for band 3. We observed that extracts from infected O erythrocytes fractionated by gel filtration, display an earlier and higher protein peak in membrane extracts compared to infected A or B erythrocytes. The absorption spectrum of the heme-containing fractions corresponded to that of hemichromes [37]. The same fractions contained aggregated band 3, which was localized by labeling infected ABO erythrocyte membranes with the specific fluorescent band 3 label eosin-5-maleimide as described [38]. The observed chromatographic co-elution of hemichromes and aggregated band 3 is indicative of hemichrome-induced clustering of band 3 [25].
This study provides the first evidence that the phagocytic uptake of P. falciparum-infected erythrocytes is influenced by ABO blood group antigens. These data provide a new putative mechanism by which blood group O may contribute to protection against severe malaria. In order to define potential mechanisms of protection associated with blood group O, we investigated P. falciparum invasion and growth in ABO erythrocytes, as well as examined a role for differential clearance of infected ABO erythrocytes in vitro and in vivo. We found no difference in the invasion or maturation of P. falciparum parasites in A, B or O erythrocytes. However, we did observe enhanced phagocytosis of infected O erythrocytes by human macrophages (Figure 2) that was attributable to increased hemichrome deposition and band 3 aggregation (Figure 7). This observation was dependent on P. falciparum infection as no differences were observed in the baseline uptake of uninfected A, B or O erythrocytes. Preferential phagocytosis of infected O erythrocytes was independent of the donor ABO blood group (Figure 2C). We extended these observations to an in vivo model, and demonstrated increased macrophage uptake of infected O erythrocytes in vivo compared to infected A or B erythrocytes (Figure 3A, B) [29]. Taken together, our data suggest that there are differences in phagocytic clearance of infected O versus infected A and B erythrocytes which may contribute to reduced parasite burdens and improved malaria outcomes in blood group O individuals.
In order to investigate whether blood group antigens might directly affect phagocytic uptake, we performed phagocytosis assays on infected erythrocytes that varied in their relative expression of H and A antigens. We observed a relationship between phagocytic index and lower levels of immunodominant A or B expression (or higher reciprocal levels of erythrocyte H antigen expression) on infected erythrocytes (Figure 4B). Differences in blood group terminal monosaccharides may influence phagocytosis either directly on the basis of H antigen density, or indirectly by the absence of the A or B antigens. Given that no differences in the baseline uptake of uninfected A, B or O erythrocytes were observed, the preferential phagocytosis of infected O erythrocytes is therefore dependent on P. falciparum infection and may be attributable to group-specific differences in parasite-encoded erythrocyte membrane proteins or other P. falciparum-induced structural modifications to the erythrocyte membrane. Given the possibility that other inter-individual ABO-associated differences might have accounted for the observed ABO effect on phagocytosis, we examined the uptake of infected erythrocytes that had and had not been enzymatically modified to resemble O erythrocytes (Figure 5). B erythrocytes were chosen since enzymatic conversion of A erythrocytes to O results in not only the common H antigens of types 1 and 2 regularly found on wild type O erythrocytes but also a qualitatively different H antigen of type 3 found on A erythrocytes [39]. In these experiments B- zyme α-glycosidase treatment of B erythrocytes removed the terminal α-1,3-galactose from blood group B antigens, resulting in loss of anti-B recognition and conversion to erythrocytes which type as group O (Figure 5A). Subsequent infection of B-zyme-converted erythrocytes resulted in enhanced macrophage uptake to levels observed with infected wild type O erythrocytes, and different from the levels seen with unmodified B erythrocytes (Figure 5B). B-zyme treatment per se was not responsible for this effect, as treatment of uninfected O or B erythrocytes, or infected O erythrocytes, had no affect on their uptake by macrophages.
There are a number of potential explanations for how A/B/H antigens could modify macrophage recognition and uptake. ABO may influence the differential expression of parasite ligands such as PfEMP-1, or the steric accessibility of other parasite-dependent pattern recognition motifs. The H antigen found at high levels on O erythrocytes may alternatively act as a co-receptor to a parasite ligand, or influence other parasite-induced erythrocyte modifications (for example, increased senescence antigen expression by infected O erythrocytes). Recent evidence suggests that ABH antigens can stabilize sialylated glycan clusters on the erythrocyte membrane in a manner that is unique for each blood group [40]. In this way ABH antigens can differentially modulate cellular interactions without being a direct ligand themselves by modifying other cell surface glycans and making them more or less accessible for binding. Cohen et al. have shown that by stabilizing such structures, ABH antigens can also modulate interaction with pathogens such as P. falciparum [40]. Therefore, it is possible that ABH antigens may non-covalently alter the expression or presentation of other cell surface glycans including parasite encoded proteins such as P. falciparum erythrocyte membrane protein-1 (PfEMP-1). PfEMP-1, an important parasite virulence factor [41], has also been shown to demonstrate differential expression on the erythrocyte membrane in erythrocyte disorders, including hemoglobin C, associated with protection to severe malaria [42]. It is therefore plausible that modified expression of PfEMP-1 or other parasite ligands on O erythrocytes, results in increased interaction with macrophage pattern recognition and phagocytic receptors and enhances uptake. Our data are consistent with a model whereby infected O erythrocytes bind more avidly to phagocytic cell receptors resulting in enhanced uptake (Figure S1).
In addition to a putative role for ABH antigens in modifying parasite-erythrocyte interactions, phagocytic recognition and clearance of erythrocytes have also been associated with increased expression of erythrocyte senescence antigens such as aggregated band 3 and phosphatidylserine (PS) [43], [44]. Although PS exposure has been reported to be elevated in variant erythrocytes, (e.g., sickle cell trait) where it may serve as a senescence signal for accelerated clearance [45], we found no significant difference in P. falciparum-induced PS expression on infected red cells to account for the observed preferential uptake of infected O erythrocytes.
With respect to band 3, there are approximately 1 million ABH glycan antigen sites on each erythrocyte, and many are presented on this protein [43]. Increased band 3 aggregation has been reported in sickle cell and β-thalassemic erythrocytes, contributing to erythrocyte membrane modification and enhanced phagocytic uptake by macrophages. Whether the absence of immunodominant sugars is more permissive to malaria-induced band 3 aggregation is unknown, as ABO effects have not previously been specified in such studies. In the present study we observed increased hemichrome formation and band 3 aggregation in infected O erythrocytes compared to infected A and B erythrocytes. The mechanism by which O erythrocytes might be more susceptible to malaria-induced oxidant stress is not known. Erythrocytes under increased oxidative stress, such as that induced by malaria parasite invasion and growth, may show preferential oligomerization/phosphorylation of less-glycosylated band 3 fractions [46]. This possibility is consistent with reports that band 3 displays an increased tendency to cluster in congenital dyserythropoietic anemia type 2 which is characterized by band 3 under-glycosylation, and with the irreversible cross linking observed in poorly glycosylated band 3 fractions in G6PD-deficient erythrocytes [46]–[48]. Glycosylation of band 3 appears to be a restraint to its oxidative cross-linking, clustering and subsequent phagocytic uptake. Collectively these observations provide a putative molecular mechanism for the observed enhanced uptake of infected O erythrocytes.
In summary, we have demonstrated a novel mechanism by which blood group O may contribute to protection against severe disease. The present model is complementary to, and not incompatible or inconsistent with, the decreased rosetting of infected O erythrocytes reported by others [12]. Both increased phagocytosis and decreased rosetting of blood group O may contribute functionally to reduced parasite burden, decreased infected erythrocyte adhesion to the endothelium and decreased microvascular obstruction, all of which are believed to play important mechanistic roles in the pathophysiology of severe falciparum malaria.
Whole blood was donated from healthy malaria-naïve individuals after informed consent using a protocol approved by the University Health Network Research Ethics Board. Animal use protocols were reviewed and approved by the Faculty of Medicine Advisory Committee on Animal Services at the University of Toronto according to the Guide to the Care and Use of Experimental Animals (Canadian Council on Animal Care, 1993).
Endotoxin-free RPMI 1640 and gentamicin were purchased from Invitrogen Life Technologies (Burlington, ON, Canada). Human AB serum was purchased from Wisent Inc (St-Bruno, Quebec, Canada). Diff-Quik staining kit and fetal bovine serum (FBS) were purchased from Fisher Scientific (Ottawa, ON, Canada). FBS and human AB serum were heat-inactivated for 30 minutes at 55°C before use. Alanine was purchased from Sigma Aldrich (Oakville, Ontario, Canada). Mycoplasma removal agent was purchased from MP Biochemical (Solon, Ohio, USA). Ficoll-Paque and Percoll were purchased from GE Healthcare (Baie D'Urfé, Québec, Canada). NOVACLONE blood grouping reagent was purchased from Dominion Biologicals Ltd (Dartmouth, Nova Scotia, Canada). All other reagents were purchased from common commercial sources.
C57BL/6 mice used in this study were 6–10 weeks old and were purchased from Charles River Laboratories (Hollister, CA) and maintained under pathogen-free conditions with a 12-h light cycle.
P. falciparum (clones ITG, 3D7 and E8B) was cultured as previously described [47]. Cultures were treated with Mycoplasma-Removal Agent, confirmed to be Mycoplasma-free (MycoAlert Mycoplasma Detection Kit, Lonza) and synchronized by alanine treatment.
Whole blood was obtained from hematologically healthy laboratory staff members (11 group A, 4 group B and 6 group O). Individuals with underlying red cell traits or disorders, or previous malaria exposure were excluded. Erythrocytes were separated from whole blood as previously described [27]. Briefly, whole blood was layered on an 80% Percoll gradient [80% (vol/vol) Percoll, 6% (w/v) mannitol, 10 mM glucose and 10% (vol/vol) PBS 10×] and spun for 15 minutes at 3000 RPM at 24°C. The erythrocyte pellet was collected and washed in R-0G media (RPMI 1640 medium supplemented with 10 mM glucose and 10 g/L gentamicin) and resuspended in equal volumes of parasite growth medium R-10G (RPMI-1640 containing 20 mM glucose, 2 mM glutamine, 6 g/L HEPES, 2 g NaHCO3, 10 g/L gentamicin, 10% human AB serum and 1.35 mg/L hypoxanthine, pH 7.3). Serum was isolated by centrifugation at 1500 RPM for 10 minutes at 24°C, and 200 µl aliquots were stored at −20°C for future use. Each aliquot was thawed only once and discarded after use.
Blood samples were tested by standard hemagglutination techniques with commercially available anti-A and anti-B reagents approved for diagnostic use [39]. Donors expressing A antigens were further typed using Dolichos biflorus lectin to differentiate between the A1 and A2 subgroups.
Human monocytes were purified from the peripheral blood of healthy donors and cultured on glass cover slips in 24-well polystyrene plates as previously described [49]. Briefly, whole blood was diluted 1∶1 with warm PBS, layered onto Ficoll (25 mL/15 mL) and centrifuged at 1800 RPM for 30 minutes at 20°C. The peripheral blood mononuclear cell (PBMC) layer was washed 3 times with cold RPMI-1640 and resuspended in R-10G FBS media (RPMI-1640 medium containing L-glutamine and HEPES supplemented with 10% heat-inactivated FBS and 25 mg/L gentamicin). 1.25×106 PBMCs were pipetted onto 24-well plate containing coverslip and incubated at 37°C for 1 hour. Wells were washed twice to remove non-adherent cells and the monocytes were cultured in R-10G FBS for 5 days at 37°C to allow differentiation into macrophages.
To assess parasite invasion and maturation, schizont stage P. falciparum-infected erythrocytes from synchronized cultures were purified on a Percoll-mannitol gradient [27], [50] and mixed with erythrocytes of different blood groups (A, B and O) in R-10G as described [27]. Invasion of erythrocytes was assessed at 24 hours, and 72 hours, and maturation was assessed at 48 hours, and 96 hours. Slides were stained with Diff-Quik, and 1000 erythrocytes were examined microscopically. Percent parasitemia was determined as follows: [number of parasites÷number of total erythrocytes counted]×100.
Uninfected and infected ring-stage or mature-stage parasitized erythrocytes were incubated with 50% fresh autologous serum for 30 minutes at 37°C. Erythrocytes were then washed twice, resuspended at 10% hematocrit, and incubated with macrophages adherent to glass coverslips at a target-to-effector ratio of 40∶1(ring-stage) or 1∶20 (mature-stage). Phagocytosis assays were performed as described previously [27] and were counted and analyzed blinded to the erythrocyte blood group.
To assess phagocytosis of infected A, B and O erythrocytes in vivo, 50×106 infected erythrocytes, or uninfected erythrocytes as a control, were injected into the peritoneal cavity of C57BL/6 mice as previously described [29]. Three hours after injection, peritoneal cells were collected, and washed with R-0G media. The cells were then suspended in cold water to lyse and remove non-internalized erythrocytes. Cells were then resuspended in 500 µl of R-10G media. 150 µl aliquots of peritoneal cells from each mouse were placed on a glass coverslip in a 24 well plate, allowed to adhere for 30 minutes at 37°C and stained with Diff-quick. In addition, 200 µl of the suspension were cytospun at 800 RPM for 10 minutes and stained with Diff-quick. Images from these slides were acquired with an Olympus BX41 microscope and an Infinity2 camera at 1000× magnification.
Flow cytometric detection of H antigens was performed as previously described [34], using a FITC-conjugated monoclonal anti-H (BRIC231) antibody. Blood groups A1, A2, O and H negative control (Bombay), were tested simultaneously and ten thousand events were collected. In the histogram FITC-derived fluorescence is displayed on the x axis on a logarithmic scale and the number of cells is on the y axis.
Removal of B antigens was achieved as previously described [34]. Briefly, erythrocytes were prewashed 1∶1 and 1∶4 vol/vol in glycine buffer (200 mM glycine and 3 mM NaCl, pH 6.8). The conversion reaction consisting of a 30% suspension of erythrocytes in glycine buffer with addition of bacterially-derived GH110 family α-galactosidase (B-zyme from Bacteroides fragilis) [34], [51]. The bacterial glycosidase was incubated for 60 min during gentle mixing at 26°C, followed by 4 repeated washing cycles with 1∶4 vol/vol of saline by centrifugation at 1000 RPM. To verify the removal of B antigen after enzyme conversion flow cytometric detection of ABO antigens was performed as recently outlined [51], using the IgM anti-B clone 9621A8 (Diagast, Loos, France) as primary antibody and phycoerythrin (PE)-conjugated rat-anti-mouse Ig kappa light chain (Becton Dickinson, San Jose, CA, USA) as secondary antibody. Treated and untreated erythrocytes were tested simultaneously and control cells of known phenotype (B, Bweak subgroup and O cells were included in each run to confirm sensitivity and specificity of the assay, as previously shown) [51]. Ten thousand events were collected and log fluorescence data was gated on a linear forward scatter versus linear side scatter dot plot.
A, B and O uninfected and P. falciparum ring-stage and mature-stage infected erythrocytes were washed in PBS supplemented with 0.5% BSA and 0.1% azide and lysed with Tris-HCl/EDTA (pH 8.0) in the presence of protease inhibitors cocktail (Roche Diagnostics GmbH). Where indicated, erythrocytes were incubated for 30 minutes at room temperature in the dark in PBS-glucose containing 10 µmol/L eosin-5-maleimide in order to label band 3 in situ [52]. Membrane pellets were extracted as described [38]. Hemichromes were quantified using the Winterbourne equation [37]. Tween-20 detergent-extracted membrane proteins (500 µL) were then loaded onto Sepharose CL-6B column equilibrated with 10 mmol/L Tris buffer and separated at a flow rate of 0.760 mL/min. The effluent was collected in 1.2 mL fractions. Total proteins in the fractions were assayed using Bedford reagent at 595 nm and labeled band 3 was assayed using fluorometry (Ex-522 nm and Em-550 nm). Aggregated band 3 was assayed in the Tween-20 fractions of uninfected and mature stage-separated infected erythrocytes previously labeled by the band 3-specific fluorescent label eosin-5-maleimide as described [53]. In order to quantify the percentage of aggregated band 3, eosin-5-maleimide-labeled band 3 in Tween-20 fractions, the fluorescence value measured in the high-molecular-weight fractions was normalized to the total fluorescence measured in all fractions.
Statistical analysis was performed using GraphPad Prism 4 software (San Diego, CA, USA). To confirm the normal distribution of data, all continuous data sets were assessed using the Kolmogorov-Smirnov test. Data sets that displayed normal distribution were analyzed by Student's t-test (two-tailed) or a one-way ANOVA as appropriate. Data sets that did not display normal distribution were analyzed by the Mann-Whitney rank sum test. Multiple comparisons were corrected using the Bonferroni method. A general linear model was used to analyze experiments with multiple independent variables (e.g., macrophage and erythrocyte group). To test the dose-dependent effect of the A and H antigen, we used a Spearman's rank correlation on the individual data points (phagocytic index) correlate with decreasing level of A antigen. Data are presented as box plots representing the median, inter-quartile range and range or as bar graphs representing the mean±SEM.
|
10.1371/journal.ppat.1000055 | Control of Viremia and Prevention of AIDS following Immunotherapy of SIV-Infected Macaques with Peptide-Pulsed Blood | Effective immunotherapies for HIV are needed. Drug therapies are life-long with significant toxicities. Dendritic-cell based immunotherapy approaches are promising but impractical for widespread use. A simple immunotherapy, reinfusing fresh autologous blood cells exposed to overlapping SIV peptides for 1 hour ex vivo, was assessed for the control of SIVmac251 replication in 36 pigtail macaques. An initial set of four immunizations was administered under antiretroviral cover and a booster set of three immunizations administered 6 months later. Vaccinated animals were randomized to receive Gag peptides alone or peptides spanning all nine SIV proteins. High-level, SIV-specific CD4 and CD8 T-cell immunity was induced following immunization, both during antiretroviral cover and without. Virus levels were durably ∼10-fold lower for 1 year in immunized animals compared to controls, and a significant delay in AIDS-related mortality resulted. Broader immunity resulted following immunizations with peptides spanning all nine SIV proteins, but the responses to Gag were weaker in comparison to animals only immunized with Gag. No difference in viral outcome occurred in animals immunized with all SIV proteins compared to animals immunized against Gag alone. Peptide-pulsed blood cells are an immunogenic and effective immunotherapy in SIV-infected macaques. Our results suggest Gag alone is an effective antigen for T-cell immunotherapy. Fresh blood cells pulsed with overlapping Gag peptides is proceeding into trials in HIV-infected humans.
| Effective immunotherapies for HIV are needed. We assessed a simple technique, reinfusion of fresh blood cells incubating with overlapping SIV peptides (Overlapping Peptide-pulsed Autologous ceLls, OPAL), in 36 randomly allocated SIV-infected monkeys. We analyzed this therapy for the stimulation of immunity, control of virus levels, and prevention of AIDS. The OPAL immunotherapy was safe and stimulated remarkable levels of T-cell immunity. Levels of virus in vaccinated monkeys were 10-fold lower than in controls, and this was durable for over 1 year after the initial vaccinations. The immunotherapy resulted in fewer deaths from AIDS. We conclude this is a promising immunotherapy technique. Trials in HIV-infected humans of OPAL therapy are planned.
| Several attempts at immunotherapy of HIV using more conventional vaccines have thus far been poorly immunogenic and weakly efficacious in human trials [1],[2],[3],[4]. The use of professional antigen-presenting cells such as dendritic cells to deliver HIV immunotherapies has shown strong immunogenicity efficacy in macaques and pilot humans studies but is limited to highly specialized facilities [5],[6],[7]. A simple intermittent immunotherapy that reduces the need for long-term antiretroviral therapy (ART) would be a quantum advance in treating HIV.
We recently reported the robust T-cell immunogenicity of treating unfractionated whole blood or peripheral blood mononuclear cells (PBMC) with overlapping peptides of SIV, HIV-1 or hepatitis C virus in outbred pigtail monkeys [8],[9]. We termed this simple immunotherapy OPAL (Overlapping Peptide-pulsed Autologous ceLls). This technique is attractive since there is no prolonged ex vivo culture of antigen-presenting cells, robust CD4 and CD8 T-cell responses to both structural and regulatory proteins can be induced, and peptide antigens are simple to manufacture to high purity. This study assessed whether OPAL vaccination improves the outcome of SIV-infected monkeys.
Considerable debate exists regarding the most effective antigens to target for T-cell based therapeutic HIV vaccination. It has been widely believed that broader immunity to multiple proteins would be more efficacious [10],[11]. In contrast, recent studies highlight the effectiveness of Gag-specific T cell immunity in comparison to T cell immunity to other antigens. We therefore also assessed whether narrower responses induced only to SIV Gag are as effective as more broadly targeting all 9 SIV proteins.
Juvenile pigtail macaques (Macaca nemestrina) free from Simian retrovirus type D were studied in protocols approved by institutional animal ethics committees and cared for in accordance with Australian National Health and Medical Research Council guidelines. All pigtail macaques were typed for MHC class I alleles by reference strand mediated conformational analysis and the presence of Mane-A*10 confirmed by sequence specific primer PCR as described [12],[13]. 36 macaques were injected intravenously with 40 tissue culture infectious doses of SIVmac251 (kindly provided by R. Pal, Advanced Biosciences, Kensington, MD) as described previously [14],[15] and randomized into 3 groups of 12 animals (OPAL-Gag, OPAL-All, Controls) 3 weeks later. Randomization was stratified for peak SIV viral load at week 2, weight, gender and the MHC I gene Mane-A*10 (which is known to enhance immune control of SIV) [15]. Animals received subcutaneous injections of dual anti-retroviral therapy with tenofovir and emtricitibine (kindly donated by Gilead, Foster City, CA; both 30 mg/kg/animal) for 7 weeks from week 3: daily from weeks 3–5 post-infection and three times per week from weeks 6–10. This dual ART controls viremia in the majority of SIV-infected macaques [16],[17],[18],[19],[20].
Two animal groups (OPAL-Gag and OPAL-All) were immunized with OPAL immunotherapy using PBMC as previously described [8]. Briefly, peripheral blood mononuclear cells (PBMC) were isolated over Ficoll-paque from 18 ml of blood (anticoagulated with Na+-Heparain). All isolated PBMC (on average 24 million cells) were suspended in 0.5 ml of normal saline to which either a pool of 125 SIVmac239 Gag peptides or 823 peptides spanning all SIVmac239 proteins (Gag, Pol, Env, Nef, Vif, Tat, Rev, Vpr, Vpx) were added at 10 µg/ml of each peptide within the pool. Peptides were 15mers overlapping by 11 amino acids at >80% purity kindly provided by the NIH AIDS reagent repository program (catalog numbers 6204, 6443, 6883, 6448-50, 6407, 8762, 6205). To pool the peptides, each 1 mg vial of lyophilised 15mer peptide was solubilized in 10–50 µl of pure DMSO and added together. The concentration of the SIV Gag and All peptide pools was 629 and 72 µg/ml/peptide respectively, although each peptide was pulsed onto cells at 10 µg/ml regardless of vaccine type. The peptide-pulsed PBMC were held for 1 hr in a 37°C waterbath, gently vortexed every 15 minutes and then, without washing, reinfused IV into the autologous animal. Peptide concentrations and timing of incubation were adapted from effective stimulation of T cell responses in vitro. Control macaques did not receive vaccine treatment. This was done since (a) we had not previously observed any significant VL changes with non HIV/SIV peptide sets ([8],[9] and unpublished data), (b) reinfusion of blood cells pulsed with irrelevant sets of peptides would result in some level of immune activation and drive higher viral loads in controls, artificially magnifying any reductions in the active treatment groups, (c) reinfusion of control peptide pulsed cells might have obscured any unexpected safety problems of the procedure.
SIV-specific CD4 and CD8 T-cell immune responses were analysed by expression of intracellular IFNγ as previously described [21]. Briefly, 200 µl whole blood was incubated at 37°C with 1 µg/ml/peptide overlapping 15mer SIV peptide pools (described above) or DMSO alone and the co-stimulatory antibodies anti-CD28 and anti-CD49d (BD Biosciences/Pharmingen San Diego CA) and Brefeldin A (10 µg/ml, Sigma) for 6 h. Anti-CD3-PE, anti-CD4-FITC and anti-CD8-PerCP (BD, clones SP34, M-T477 and SK1 respectively) antibodies were added for 30 min. Red blood cells were lysed (FACS lysing solution, BD) and the remaining leukocytes permeabilized (FACS Permeabilizing Solution 2, BD) and incubated with anti-human IFNγ-APC antibody (BD, clone B27) prior to fixation and acquisition (LSRII, BD). Acquisition data were analyzed using Flowjo version 6.3.2 (Tree Star, Ashland, OR). The percentage of antigen-specific gated lymphocytes expressing IFNγ was assessed in both CD3+CD4+ and CD3+CD8+ lymphocyte subsets. Responses to the immunodominant SIV Gag CD8 T-cell epitope KP9 in Mane-A*10+ animals were assessed by a Mane-A*10/KP9 tetramer as described [13]. Total peripheral CD4 T-cells were measured as a proportion of lymphocytes by flow cytometry on fresh blood.
Plasma SIV RNA was quantitated by real time PCR on 140 µl of plasma at the University of Melbourne (lower limit of quantitation 3.1 log10 copies/ml) at all time-points using a TaqMan probe as previously described [21],[22] and, to validate these results with a more sensitive assay, on pelleted virions from 1.0 mL of plasma at the National Cancer Institute (lower limit of quantitation 1.5 log10 copies/ml) as previously described [23].
The primary endpoint was the reduction in plasma SIV RNA in OPAL-immunized animals compared to controls by time-weighted area-under-the-curve (TWAUC) for 10 weeks following withdrawal of ART (i.e. samples from weeks 12 to 20). This summary statistical approach is recommended for studies such as these involving serial measurements [24]. We compared both active treatment groups (OPAL-Gag and OPAL-All) to controls separately and together. The primary analysis was restricted to animals that controlled viremia on the ART at week 10 (VL<3.1 log10 copies/ml), since control of VL is an important predictor of the ability of animals to respond to immunotherapies [8],[25]. A pre-planned secondary virologic endpoint was studying all live animals adjusting for both VL at the end of ART (week 10) and Mane-A*10 status. Group comparisons used two-sample t-tests for continuous data, and Fisher's exact test for binary data. Survival analyses utilised Cox-regression analyses.
Prior to initiating the study, we estimated the standard deviation of the return of VL after treatment interruption would be approximately 0.8 log10 copies of SIV RNA/mL plasma [5],[16],[17],[18],[19],[20]. In this intensive study we estimated that 2 of the 12 monkeys within a group may have confounding problems such as incomplete response to ART or death from acute SIV infection. A 10 control vs 10 active treatment comparison yields 80% power (p = 0.05) to detect a 1.0 log10 difference in TWAUC VL over the first 10 weeks. An estimated comparison of 10 control vs all 20 actively treated animals (OPAL-Gag plus OPAL-All) gave 80% power to detect differences of 0.87 log10 copies/ml VL reduction.
This study was conducted according to a pre-written protocol using Good Laboratory Practice Standards from the Australian Therapeutic Goods Administration as a guide. Protocol deviations were minor and did not affect the results of the study. Partial data audits during the study did not raise any concerns about the study conduct.
OPAL immunotherapy was studied in SIV-infected pigtail macaques receiving ART. Pigtail macaques have at least an equivalently pathogenic course of SIV infection as alternate rhesus macaque models [14],[26]. Thirty-six macaques were infected with SIVmac251 and 3 weeks later treatment with the antiretrovirals tenofovir and emtricitabine for 7 weeks was initiated. The animals were randomly allocated to 3 groups stratified by peak plasma SIV viral load (VL), Mane-A*10 status (an MHC class I gene that improves VL in SIV-infected pigtail macaques [15]), weight and gender. Macaques were immunized 4 times under the cover of antiretroviral therapy (weeks 4, 6, 8, 10) with autologous fresh PBMC mixed for 1 hour ex vivo with 10 µg/ml/peptide of either 125 overlapping SIV Gag 15mer peptides only (OPAL-Gag), 823 SIV 15mer peptides spanning all 9 SIV proteins (OPAL-All) or un-immunized. The macaques were initially followed for 26 weeks after ceasing ART on week 10.
All 36 macaques became infected following SIVmac251 exposure and had a mean peak VL of 7.1 log10 copies/ml (Table S1). Prior to vaccination, 4 animals died during acute SIV infection with diarrhoea, dehydration, lethargy, anorexia and weight loss. The vaccinations were well tolerated, with no differences in mean weights, haematology parameters, or clinical observations in OPAL immunized animals compared to controls (data not shown).
There was striking SIV-specific CD4+ and CD8+ T-cell immunogenicity after the course of vaccination in the OPAL immunized animals. Mean Gag-specific CD4 and CD8 T-cell responses 2 weeks after the final immunization were 3.0% and 1.9% of all CD4 and CD8 T cells respectively in the OPAL-Gag group. Mean Gag-specific CD4 and CD8 T-cell responses 2 weeks after the final immunization were 0.84% and 0.37% in the OPAL-All group and 0.15% and 0.29% in controls (Fig. 1A, B). The Gag-specific T cells in the OPAL-All immunized animals, but not control or OPAL-Gag only immunized animals, also had elevated T-cell responses to all other SIV proteins. Mean Env, Pol and combined regulatory protein-specific CD4/CD8 responses were 2.5%/11.8%, 0.8%/0.3% and 1.5%/2.4% respectively in the OPAL-All group compared to ≤0.4% for all CD4/8 responses to non-Gag antigens in control and OPAL-Gag groups (Fig. 1C, D and Fig. 2). The kinetics of induction of non-Gag CD4 and CD8 T cell responses in the OPAL-All group was similar for induction of Gag-specific T cell immunity. Stronger CD8 T-cell responses to non-Gag proteins correlated with reduced CD8 T-cell responses to Gag (Fig. 1E). Thus, although a larger number of SIV proteins were recognized in the OPAL-All immunized animals, Gag responses were reduced in comparison to only immunizing with Gag peptides.
Although the short linear peptides were primarily used to induce T cell immunity, we also studied serial plasma samples for SIV-specific antibodies. All animals seroconverted following SIV infection, as shown by Western Blot (Fig. 3A). No significant enhancement of Gag or Env antibody responses occurred with the OPAL vaccinations (Fig. 3B, C). There was a dip in mean Gag antibody responses during the period of ART in all groups consistent with reduced viral antigen during this period. In addition to the lack of difference in mean Gag (p26) or Env (gp36) responses shown in Figure 3B and 3C, there were also no significant different antibody responses to p16, p68, gp125 and gp140 across the vaccine groups (not shown).
The 7-week period of ART controlled VL to below 3.1 log10 copies/ml in 26 of the remaining 32 animals by week 10 (Table S1). The pre-defined (per-protocol) primary VL endpoint analyses was performed on animals controlling viremia on ART (26 animals). The 6 animals that failed to control viremia on ART had higher peak VLs at week 2 (mean±SD of 7.74±0.33 compared to 6.94±0.52 for animals controlling viremia on ART, p<0.001) and higher VL following ART withdrawal (5.98±0.53 vs 4.28±0.90, p<0.001). Control of VL is likely to be important in achieving optimal results from immunotherapy of infected macaques [8],[20].
The primary endpoint comparison of VL between combined OPAL-All and OPAL-Gag treatment groups in the 10 weeks after ART withdrawal was 0.5 log10 copies/ml lower than controls (p = 0.084, Fig 4, Table 1). Each vaccination group (OPAL-All and OPAL-Gag) had very similar reductions in VL. By 6 months after ART withdrawal, the mean difference in VL between control and OPAL-immunized groups was 0.93 log10 copies/ml (p = 0.028, Table 1).
As a secondary endpoint, we also analysed all 32 remaining animals by adjusting for VL control on ART and Mane-A*10 status. There was a significant difference in VL between controls and vaccinated macaques with these analyses at both 10 and 26 weeks off ART (p = 0.050, 0.016 respectively, Table 1).
To confirm the virologic findings using a sensitive independent VL assay, frozen plasma (1 ml) from study week 32 was shipped to the National Cancer Institute (NCI) in Maryland, USA. Drs M Piatak and J Lifson kindly analysed the samples for SIV RNA blindly using an assay with a limit of quantitation of 1.5 log10 copies/ml (Table S1) [23]. The University of Melbourne and NCI assays were tightly correlated (r = 0.97, p<0.001) and showed an almost identical mean reduction in viremia in vaccinees compared to controls at this time (0.82 vs 0.88 log10 copies/ml respectively).
To further assess the durability of SIV control and prevention of disease with OPAL immunotherapy, we re-boosted all 32 animals in the same randomized groups 3 times with the identical procedure (at week 36, 39, 42) without ART cover and followed the animals for an additional 6 months. Despite the lack of ART cover, SIV-specific T cell immunity was dramatically enhanced in immunized animals 2 weeks after the last vaccination, similarly to the primary vaccination (Figs 1, 2). The T cell responses to Gag were again highest in the OPAL-Gag group with broader responses in the OPAL All group. The pattern of enhancement of T cell immunity was similar for the first and second vaccination sets (Figs 1, 2).
We again sampled plasma for viral load every 3–6 weeks. To account for the death of animals from AIDS, we used a “last observation carried forward” analysis for missing VL data. Significant viral control was maintained throughout the follow up period of just over 1 year off ART (Fig 4A, Table 1). In animals which controlled VL on ART, there was a mean 0.98 log10 copies/ml difference between controls and vaccinees 54 weeks after coming off ART (p = 0.019 for time-weighted analysis).
Twelve of the remaining 32 animals developed incipient AIDS and were euthanised during the extended follow up. All 6 animals that did not control viremia on ART required euthanasia. Of the 6 euthanised animals which did control viremia on ART, 5 were in the control group and one in the OPAL-Gag group. OPAL immunotherapy resulted in a survival benefit, analysing either the 26 animals that controlled viremia on ART (p = 0.053, Fig 4B, Table 1) or all 32 animals, adjusted for Mane-A*10 status and control of viremia on ART (p = 0.02, Table 1).
In summary, OPAL immunotherapy, either using overlapping Gag SIV peptides or peptides spanning the whole SIV proteome was highly immunogenic and resulted in significantly lower viral loads and a survival benefit compared to unvaccinated controls. The virologic efficacy in OPAL-immunized macaques was durable for 12 months after ART cessation. Our findings on OPAL immunotherapy were observed despite the virulent SIVmac251-pigtail model studied [14] and provide strong proof-of-principle for the promise of this immunotherapy technique.
The OPAL immunotherapy approach is simpler than many other cellular immunotherapies, particularly the use of dendritic cells. The use of DNA, CTLA-4 blockade and viral vector based approaches are also now showing some promise in macaque studies [17],[27], although such approaches have not yet been translated into human studies. This study added peptides to PBMC, however we have shown an even simpler technique, adding peptides to whole blood is also highly immunogenic, a technique that will be more widely applicable ([8] and unpublished studies).
This is one of the largest therapeutic SIV vaccine studies yet reported. Although it may have been ideal to have studied irrelevant peptide-pulsed autologous cells as an additional control group, we were concerned that this may have magnified the therapeutic effect or obscured any safety concerns. In the end, the vaccination process was both safe and effective.
How well the findings on OPAL immunotherapy translate to humans with acute HIV-1 infection will be determined by clinical trials. Virus-specific CD4 T cells are typically very weak in HIV-infected humans or SIV-infected macaques; dramatic enhancement of these cells were induced by OPAL immunotherapy and this may underlie its efficacy [28]. We measured IFNγ-producing T cells in this study since we had not developed polyfunctional ICS assays prior to initiating the study. However, recent cross-sectional polyfunctional ICS assays suggests OPAL immunotherapy can also induce T cells capable of also expressing the cytokines TNFα and IL-2, the chemokine MIP1β and the degranulation marker CD107a (unpublished data).
A ∼1.0 log10 reduction in VL would result in a substantial delay in progressive HIV disease in humans and allow a reasonable time period without the requirement to reintroduce ART [29] if these findings are confirmed in human trials. Both the control and vaccinated macaques were treated with ART early in this study (3 weeks after infection), which alone can be associated with a transiently improved outcome in humans [30]. None-the-less, a massive loss of CD4+ T cells in the gut occurs within 2 weeks of infection [31]. Although it may be challenging to identify humans within 3 weeks of infection, this is when HIV-1 subjects typically present with acute infection. The durable control of viremia exhibited by the vaccinated animals is interesting and consistent with other recent macaque studies [27], suggesting the need for re-immunization may not be substantial. We cannot attribute the durable control of viremia to the second set of immunizations; there was only a marginal, non-significant, increase in the difference in VL between OPAL vaccinees and controls before and after the second immunization series. Further studies are required to address the timing and benefit of ART cover during boosting immunizations with OPAL immunotherapy.
Control of viremia was similar for the OPAL-Gag and OPAL-All groups. Gag-specific CD4 and CD8 T-cell responses in OPAL-Gag animals 5.1- and 3.5-fold greater than those in the OPAL-All animals, despite an identical dose of Gag overlapping peptides. This suggests antigenic competition between peptides from Gag and the other SIV proteins. Inducing immunodominant non-Gag T-cell responses by multi-protein HIV vaccines may limit the development of Gag-specific T-cell responses [21]. A large human cohort study demonstrated Gag-specific T-cell responses were the most effective in controlling HIV viremia [32]. Useful subdominant T cell responses may be particularly susceptible to dominant non-Gag T cell responses [33],[34]. The utility, if any, of inducing T-cell responses to non-Gag proteins (i.e. excluding Gag peptides from the vaccine antigens) can be addressed in future studies of this flexible vaccine technology. Therapeutic HIV vaccines may not need to aim for maximally broad multi-protein HIV-specific immunity.
OPAL immunotherapy with Gag peptides is proceeding into initial trials in HIV-infected humans. Additional peptides can readily be added into standard consensus strains mixes to cover common strain or subtype variations between strains with this technology [35]. Additional technologies such as toggling variable amino acids peptides may provide further T cell immunogenicity with this general technology [36]. Immunotherapy with peptides delivered onto fresh blood may have potential applicability for other chronic viral diseases such as hepatitis C virus infection and some cancers such as melanoma [37]. |
10.1371/journal.ppat.1004063 | A Human Lung Xenograft Mouse Model of Nipah Virus Infection | Nipah virus (NiV) is a member of the genus Henipavirus (family Paramyxoviridae) that causes severe and often lethal respiratory illness and encephalitis in humans with high mortality rates (up to 92%). NiV can cause Acute Lung Injury (ALI) in humans, and human-to-human transmission has been observed in recent outbreaks of NiV. While the exact route of transmission to humans is not known, we have previously shown that NiV can efficiently infect human respiratory epithelial cells. The molecular mechanisms of NiV-associated ALI in the human respiratory tract are unknown. Thus, there is an urgent need for models of henipavirus infection of the human respiratory tract to study the pathogenesis and understand the host responses. Here, we describe a novel human lung xenograft model in mice to study the pathogenesis of NiV. Following transplantation, human fetal lung xenografts rapidly graft and develop mature structures of adult lungs including cartilage, vascular vessels, ciliated pseudostratified columnar epithelium, and primitive “air” spaces filled with mucus and lined by cuboidal to flat epithelium. Following infection, NiV grows to high titers (107 TCID50/gram lung tissue) as early as 3 days post infection (pi). NiV targets both the endothelium as well as respiratory epithelium in the human lung tissues, and results in syncytia formation. NiV infection in the human lung results in the production of several cytokines and chemokines including IL-6, IP-10, eotaxin, G-CSF and GM-CSF on days 5 and 7 pi. In conclusion, this study demonstrates that NiV can replicate to high titers in a novel in vivo model of the human respiratory tract, resulting in a robust inflammatory response, which is known to be associated with ALI. This model will facilitate progress in the fundamental understanding of henipavirus pathogenesis and virus-host interactions; it will also provide biologically relevant models for other respiratory viruses.
| Nipah virus (NiV) is a highly pathogenic zoonotic virus that causes fatal disease in humans and a variety of other mammalian hosts including pigs. Given the lack of effective therapeutics and vaccines, this virus is considered a public health and agricultural concern, and listed as category C priority pathogen for biodefense research by the National Institute of Allergy and Infectious Diseases. Both animal-to-human and human-to-human transmission has been observed. Studies on the molecular mechanisms of NiV-mediated pathogenesis have been hampered by the lack of biologically relevant in vivo models for studying the initial host responses to NiV infection in the human lung. We show here a new small animal model in which we transplant human lung tissue for studying the pathogenesis of NiV. We showed that NiV can replicate to high levels in the human lung. NiV causes extensive damage to the lung tissue and induces important regulators of the inflammatory response. This study is the first to use a human lung transplant for studying infectious diseases, a powerful model for studying the pathogenesis of NiV infection, and will open up new possibilities for studying virus-host interactions.
| Nipah virus (NiV) is a member of the genus Henipavirus (family Paramyxoviridae) that causes severe and often lethal respiratory illness and encephalitis in humans resulting in case fatality rates of up to 92% [1]. The first human cases of NiV infection were identified during an outbreak of severe febrile encephalitis in Malaysia and Singapore in 1998–1999 [2], [3]. More recently, outbreaks have occurred in Bangladesh and India almost yearly since 2001 [1], [4]. NiV can cause Acute Lung Injury (ALI) in humans, and human-to-human transmission has been observed in recent outbreaks of NiV [5], [6], [7]. Data on the histopathology of the lungs of NiV cases is limited to necropsy findings in the respiratory tract of NiV infected cases and include hemorrhage, necrosis and inflammation in the epithelium of the small airways but not in the bronchi [8].
Endothelial cells have been identified as a major target for NiV and most studies have focused on the role of the endothelium in NiV pathogenesis [9], [10], [11]. However very limited data is available on the host responses following NiV infection in the human lung. While the endothelium plays an important role in the terminal stages of NiV infection, the role of the respiratory epithelium in the early stages of infection is critical; however, it remains largely unexplored. The specific sites of henipavirus infection in the human respiratory tract are still unknown as well as the molecular mechanism by which these viruses cause disease in humans. We have previously shown that NiV can efficiently infect human respiratory epithelial cells from the trachea, bronchi and small airways resulting in the induction of key inflammatory mediators that have been implicated in leukocyte recruitment and ALI [12]. Similarly, in animal models (hamster, ferret and African green monkey), NiV can replicate to high titers in the lungs of these animals and cause acute and severe respiratory distress [13], [14], [15], [16], [17]. Human xenograft mouse models have previously been used to study tissue development and cancer as well as the pathogenesis of infectious agents [18], [19], [20]. The majority of viral pathogenesis studies involving human xenograft mice focus on Human Immunodeficiency Virus (HIV) or Human Cytomegalovirus (HCMV) in humanized mice that have been grafted with human hematopoietic stem cells and thymus [21], [22].Here we report the first characterization of NiV infection of the human respiratory tract using a human lung xenograft model to gain further insight into the mechanisms of NiV pathogenesis in humans. Our results showed that NiV replicates to high titers in the human lung and that infection results in the induction of a robust host response.
Severely immunodeficient NSG mice served as hosts to support the successful engraftment of human lung xenografts. We implanted 6 small fragments of human fetal lung in the dorsal subcutaneous space, 3 on each side of the spine. Following transplantation, human fetal lung xenografts typically increased in size by 2–10 fold over 3 months. The human fetal lung xenografts rapidly grafted and developed mature structures similar to those seen in adult lung (Figure 1A). These mature structures included bronchi that were partially surrounded by cartilage (Figure 1B), lined with ciliated pseudostratified columnar epithelium (Figure 1C) and surrounded by longitudinal elastic fibers. The bronchi divide into bronchioles and terminal bronchioles lined by cuboidal to flat epithelium (Figure 1D). The distal respiratory tract comprises of primitive alveolar spaces that are lined with both cells that have flat (type 1) and larger rounded (type 2) pneumocyte morphology (Figure 1E). The alveolar walls of xenografts were thicker than those of normal adult human lungs. The human graft was well vascularized with the presence of arteries, veins, and capillaries (Figure 1F). Finally, the expression patterns of ephrin B2, the receptor of NiV, was similar to that seen in normal human lung tissue [23] (Figure 1G) with expression on bronchial epithelium, alveolar cells and vasculatures (Figure 1H).
Natural NiV infection involves exposure to the virus through the respiratory epithelium. To mimic this route in human lung xenografts that lack air exchange, tissues were directly injected with NiV. Following direct intragraft injection of the 3 lung tissues on the left side (primary infection) within each mouse, NSG mice did not show any signs of morbidity or mortality during our observation period of 10 days. In addition, two non-grafted NSG mice that were challenged intradermally as controls, with the same dose as xenografts, did not develop any clinical signs. Primary infection of the human lung xenografts resulted in detection of infectious NiV as early as 1 day post infection (Figure 2A) and NiV replicated to high titers (107 TCID50/gram tissue) by day 3 post infection. NiV titers remained high until the end of the experiment at 10 days post infection. Importantly, high titers of NiV were also detected in the other 3 lung tissues (on the right side of each mouse) that were not initially infected through direct intragraft injection as early as 3 days post infection. This finding clearly demonstrates that the virus can spread from infected human lung grafts to uninfected grafts in the same mouse, most likely through viremia (secondary infection; Figure 2A). The presence of viremia was further supported by the observation that infectious NiV was detected, albeit at lower levels, in several mouse tissues including lung, brain, heart, spleen and kidney at various time points post infection (Figure 2B). In fact, viremia was detected in a blood sample of 1 animal on day 10 post infection in which a low level (300 TCID50/mL) of infectious NiV was determined (Table 1). Interestingly, virus was not detected in organs from non-grafted NSG mice that were challenged intradermally with the same dose (Table 1), suggesting the NSG mouse tissues are probably not intrinsically susceptible to NiV. In order to confirm that the human lung xenografts could be infected via the hematogenous route, we next challenged 2 lung-engrafted NSG mice with NiV via the IP route. The IP challenge with NiV in this model confirmed that infection resulted in detectable viremia in 1 animal with virus spreading to the human lung xenografts in both, replicating to high titers and resulting in histopathological changes similar to those observed with intragraft challenge (Table 1). Together, these data suggest that following intragraft infection, the human lung is highly susceptible to NiV infection and results in viremia and subsequent spread to other organs in the absence of disease.
In order to study the histopathological changes associated with NiV infection in the human lung, tissue sections were stained with hematoxylin and eosin (H&E). No gross pathologic lesions were observed in the human lung grafts. Since NSG mice exhibit multiple defects in innate and adaptive immunity [24], NiV infection in human lung grafts did not result in significant inflammation. Histopathological changes in the human lung tissues following NiV infection were independent of the route of infection (intragraft, indirect or intraperitoneal) and included small focal areas with syncytia and necrosis as early as day 3 pi (Figure 3A). These areas rapidly expanded to large areas with hemorrhages and significant loss of architecture of the small airways by day 10 pi (Figure 3B). The main histopathological features of NiV infection in these tissues were the characteristic syncytia formation (Figure 3C) and areas of necrosis (Figure 3D). Syncytia formation could be observed in bronchial epithelium (Figure 3C), alveolar epithelium (Figure 3E) and vascular endothelium (Figure 3F). In addition, fibrinoid necrosis was observed in some of the vasculature as well as recruitment of granulocytes (Figure 3F). In agreement with the absence of clinical signs, NiV infection did not result in histopathological changes in any visceral mouse tissue (Table 1).
In order to identify the cells targeted by NiV in the human lung, viral nucleocapsid protein (N) expression in human lung grafts was examined with immunohistochemistry. Expression of NiV N coincided with the focal areas of histopathological changes on day 3 pi and showed intense staining (Figure 4A). By day 10, widespread expression of NiV N was observed throughout the human lung tissues (Figure 4B). NiV primarily targeted the respiratory epithelium of the bronchi and bronchioles, interstitial mesenchymal cells (Figure 4C), and the small airways (Figure 4D). Cells targeted in the small airways were primarily cuboidal, which is consistent with type-2 pneumocyte morphology, although cells with type-1 pneumocyte morphology also showed reactivity. In addition to the respiratory epithelium, NiV replication also involved the vasculature (Figure 4E). In agreement with the observation that low levels of infectious virus were detectable in several mouse tissues, small focal areas of viral antigen primarily focused in small airway epithelium could be detected in mouse lungs (Figure 4F) but not in other organs tested (Table 1). Tropism of viral antigen in mouse lung was similar between tissues infected by direct injection or following IP challenge (data not shown).
Although focal areas were generally not centered around vessels (Figure 5A) during the early stages of infection, when NiV infection involved the vasculature, CD31-positive endothelial cells were a specific target of infection (Figure 5B). Similar findings were observed in animals challenged via the IP route (Table 1).
In order to elucidate the host responses following NiV infection in the human lung, the expression of several cytokines and chemokines was determined in homogenates of human lung xenografts following direct infection with NiV (Figure 6). Since human immune cells were absent in this model, any expression of human cytokines or chemokines was primarily the result from NiV infection of human epithelial and endothelial cells. NiV infection in human lung resulted in the expression of several cytokines/chemokines, including eotaxin-1, G-CSF, GM-CSF, TNFα, VEGF, IP-10, IL-1β and IL-6 starting by day 5 pi (Figure 6). Expression of GM-CSF, TNFα, IP10 and IL-1β peaked on day 5 post infection and gradually declined over time. IL-6 and eotaxin-1 expression peaked at day 7 pi, whereas G-CSF initially peaked on day 5 but remained high throughout infection. Interestingly, VEGF expression continued to increase over time concomitant with the increased hemorrhaging and remodeling of the lung. The cytokine and chemokines profiles were similar between lung xenografts following primary (direct) or secondary (indirect) infection (data not shown).
Nipah virus is an emerging zoonotic virus that can cause severe respiratory distress and encephalitis in humans [1]. Despite intensive studies in vitro and in animal models, little is known about the mechanisms governing the development of NiV-related respiratory disease in humans; this is due to difficulties in obtaining human samples where the disease is endemic. To address this important limitation, the goal of the present study was to characterize a novel human lung xenograft model to study the pathogenesis of NiV infection in human lung in vivo.
Studies on the molecular mechanisms of NiV-mediated pathogenesis have been hampered by the lack of biologically relevant in vitro models for studying the initial host responses to NiV infection in the human lung [7], [13], [15]. To fill this gap, we recently showed that NiV can efficiently replicate in primary epithelial cells from the human respiratory tract [12]. While this is an attractive model to study the early steps of NiV entry in the host, it lacks the complexity of the microenvironment in the lung. In the current model, we show that human fetal lung tissues grafted on an immunocompromised mouse develop into more mature human lung tissues within 3 months after implantation. Transplanted lung tissues rapidly vascularized and developed bronchioles, lined with columnar epithelium, and alveolar-like spaces closely resembling those seen in normal human lung tissue.
The prototype strain of NiV (Malaysia) was used in this study. While the outbreaks in Malaysia and Singapore have primarily been associated with the development of severe febrile encephalitis with a case fatality rate of 38%, respiratory symptoms were observed (40% of lethal cases) [8]. Interestingly, the more recent outbreaks in Bangladesh and India are associated with a higher prevalence of respiratory disease as well as a significantly higher case fatality rate of 67% to 92% [1], [6]. It is currently unknown whether differences in respiratory involvement are due to genetic difference between the Malaysia and Bangladesh strains of NiV or whether confounding factors are involved, however both NiV strains can replicate efficiently and cause respiratory distress in animals [25], [26]. In addition, no histopathological data is available for human cases of the Bangladesh strain of NiV; therefore, the Malaysia strain was used in this study to allow for comparisons of histopathology and viral tropism. In humans, vasculitis and fibrinoid necrosis in the lungs was observed in the majority of fatal cases of NiV infection [8]. Multinucleated giant cells were occasionally observed in alveolar spaces and showed prominent immunostaining for viral antigen, along with alveolar hemorrhage, edema and pneumonia. Bronchial epithelium rarely showed histopathological changes. Nipah viral antigen could also be observed in the vasculature and rarely in bronchiolar epithelium [8]. We believe that the lack of viral antigen in the bronchial epithelium in fatal human cases is most likely due to timing of sampling. We previously showed, using a hamster model, that the bronchial epithelium is initially targeted by NiV early on during infection followed by rapid spread to the interstitium and involvement of pulmonary vessels [13].
In the present model, NiV replicated to high titers following intragraft injection, and virus was found to primarily replicate in respiratory epithelium of the bronchi and small airways. This is consistent with our previous finding that human respiratory epithelium is highly susceptible to NiV infection [12]. In animal models, the lung is the primary target organ of NiV infection following intranasal challenge [13], [14]. In addition to the respiratory epithelium, NiV replication was also found in the endothelium, a type of cell that has been identified as an important target for NiV [11]. The infection of the vascular system is thought to occur in the late stages of disease and lead to systemic spread of these viruses to other organs, including brain and kidney [13], [27]. In our model, systemic spread of the virus was indicated by similar titers and replication kinetics of NiV in directly inoculated lung grafts and grafts not directly injected with virus. This suggests that following infection of the lung, NiV quickly becomes viremic and spreads to other organs. In addition, systemic infection through intraperitoneal injection with NiV also resulted in infection of the human lung grafts, thus confirming hematogenous spread of the virus.
Interestingly, NiV infection in NSG mice engrafted with human lung tissues did not result in clinical signs despite evidence of replication in mouse organs, including lung and brain. Previous studies have shown that NiV infection in type I IFN receptor knock-out mice and aged mice is lethal [28], [29]. Aged mice have been shown to mount an aberrant IFN response [30]. While the NSG mice are immunocompromised, they can mount an IFN response, which seems sufficient to protect against lethal disease in our studies. Alternatively, it is possible that NSG mice are not susceptible to NiV and that the virus measured in mouse organs is only found in the blood. Also, the small foci of viral antigen detected in mouse lungs could correspond to emboli of infected human cells that slough off from the infected grafts.
Several cytokines and especially IL-6, IP-10 and VEGF were upregulated during NiV infection in the human lung. Interestingly, the levels of cytokines observed in our xenograft lung model are similar to those observed in the lungs of fatal cases of influenza virus A (H1N1) [31]. Upregulation of inflammatory mediators such as TNF-α, IP-10, IL-1β and IL-6 in the lungs was previously shown to play a role in the pathogenesis of lethal NiV in hamsters [13], as well as in the development of ALI with other respiratory virus infections, including SARS-CoV and influenza virus (H5N1) [30], [32], [33]. VEGF has an important role in ALI pathogenesis by acting as a growth factor and increasing vascular permeability [34]. We previously showed that VEGF is expressed by human respiratory epithelial cells during NiV infection [12]. This suggests that VEGF may be partly responsible for the increased pulmonary hemorrhage, endothelial destruction, and alveolar remodeling in an emphysema-like phenotype as observed in our model. Since these inflammatory mediators also play an important role in the recruitment of immune cell, our data suggests that inflammation could be observed in this model when human immune cells are present.
In addition to implanting human lung xenografts, the NSG mice have been used to engraft the human hematopoietic system to study hematopoiesis, immunity, inflammatory disease and human-specific pathogens. This humanized NSG mouse model routinely contains >25% human CD45+ cells in the peripheral blood 12 weeks post engraftment of hematopoietic stem cells [35]. Many of the inflammatory mediators expressed in the current study play an important role in immune cell recruitment [36], [37]. The ability to engraft human immune cells will allow us to study the effect of these mediators on specific immune cells populations. Future studies will make use of the fully humanized lung xenograft model to study the role of the inflammatory response in the pathogenesis of the different henipavirus strains in the human lung.
In conclusion, these data confirm that the human lung is highly susceptible to NiV infection. NiV is capable of replicating to high titers in the human lung and targets both respiratory epithelium and endothelium. Infection results in the characteristic syncytial formation and extensive lung damage. Key inflammatory mediators such as IL-6, IP-10, G-CSF and GM-CSF are expressed during infection. This model will allow for more detailed studies of the pathogenesis of respiratory disease caused by henipavirus infection. Furthermore, these data point to several inflammatory mediators that potentially play critical roles in henipavirus pathogenesis, which may be valuable as candidates for future studies of the mechanism of henipavirus pathogenesis and as potential targets for treatment.
Approval for animal experiments was obtained from the Institutional Animal Care and Use Committee, University of Texas Medical Branch (protocol number 0905041). Animal work was performed by certified staff in an Association for Assessment and Accreditation of Laboratory Animal Care (AAALAC) approved facility. Animal housing, care and experimental protocols were in accordance with NIH guidelines of the Office of Laboratory Animal Welfare. Discarded tissue from deceased human fetuses was obtained via a non-profit partner (Advanced Bioscience Resources, Alameda, CA) as approved under exemption 4 in the HHS regulations (45 CFR Part 46). Need for informed consent was waived by the UTMB Institutional Review Board.
NiV (Malaysia strain) was kindly provided by the Special Pathogens Branch of the Centers for Disease Control and Prevention, Atlanta, Georgia, United States. The virus were propagated on Vero cells in Dulbecco's minimal essential medium supplemented with 10% fetal calf serum (Hyclone, Logan, UT), L-glutamine, penicillin and streptomycin at 37°C in a humidified CO2 incubator (5%). All infectious work was performed in a class II biological safety cabinet in a biosafety level 4 laboratory (BSL4) at the Galveston National Laboratory.
NOD.Cg-Prkdcscid Il2rgtm1Wjl/SzJ mice, also known as NOD/SCID/γcnull or NSG mice (Jackson Laboratories), 3–5 weeks of age, were housed in a sterile microisolator environment. Mice were engrafted with human lung tissue (Advanced Bioscience Resources, Alameda, CA). Six fragments of human fetal lung from the same donor (15–19 weeks of age) were sutured to muscle fascia in the dorsal subcutaneous space in each mouse (∼0.5 cm from the spine, three on each side). Animals received appropriate post-surgery treatment including antibiotics and analgesics.
Twelve weeks post-engraftment, animals were transferred to an ABSL-4 facility. Prior to infection, animals were anesthetized by chamber induction (5 Liters/min 100% O2 and 3–5% isoflurane). Three of the six lung tissues were inoculated via intragraft injection of 105 TCID50 NiV in a 50 µl volume. Animals were monitored daily for weight loss and clinical signs. Groups of 3 animals were euthanized on days 1, 3, 5, 7 and 10 post infection, and samples for virus isolation and histological examination were procured from whole blood (EDTA vacutainer), human lung tissues, and mouse liver, spleen, kidney, lung, heart and brain. In a separate experiment, 2 animals were injected via the intraperitoneal route with 105 TCID50 NiV and euthanized 10 days post infection. Control groups were NSG mice without a lung xenograft and challenged via the intradermal route with 105 TCID50 NiV on the back of the mouse at the same location the human lung xenografts would be.
Whole blood was tested for presence of infectious virus by 10-fold diltutions as described below. Tissue samples were weighed and homogenized in 10 equivalent volumes of DMEM to generate a 10% solution. The solution was centrifuged at 10,000 rpm under aerosol containment in a table top centrifuge for 5 min to pellet insoluble parts. Virus titration was performed using a TCID50 assay on 96-well plates (1×104 Vero cells per well) with 100 µL inocula (cleared homogenate or whole blood) from 10-fold serial dilutions. Plates were incubated for 3 days at 37°C, and wells were scored for cytopathic effect (CPE). Virus concentrations were calculated as TCID50 per gram of tissue.
All tissue samples were immersion-fixed in 10% neutral buffered formalin for at least 7 days under BSL4 conditions. Prior to removal from the BSL4 laboratory, formalin was changed and specimens were processed under BSL2 conditions by conventional methods, either embedded in paraffin, sectioned at 5 µm thickness and stained with hematoxylin and eosin (H&E) or embedded in Tissue Tek and frozen sections cut at 3–8 µm thickness and used for immunofluorescent (IF) staining. Tissues for immunohistochemistry (IHC) were stained as previously described using a rabbit anti-NiV-nucleoprotein (N) antibody (kindly provided by Dr. C. Broder, Uniformed Services University, Bethesda, MD) [13]. Tissues for IF were stained with a rabbit anti-NiV-N antibody, a biotinylated anti-CD31 (eBioscience), anti-collagen IV labeled with Alexa 647 (eBioscience), anti ephrin B2 (Santa Cruz Biotechnology) or ephrin B3 (R&D Systems) and Hoechst for nuclear staining. NiV N in mouse tissue could only be detected following immunofluorescent staining, likely due to the limit of detection by conventional IHC. An Alexa 546 labeled secondary antibody (Life Technologies) was used for detection of the anti NiV N antibody as well as anti ephrin B2 and B3 antibodies and an Alexa 488 conjugated streptavidin (R&D Systems) was used for detection of the anti-CD31 antibody.
Cytokine/chemokine concentrations in the homogenates of NiV infected human lung tissues were determined using a Milliplex Human Cytokine PREMIXED 28 Plex Immunoassay Kit (Millipore, Billerica, USA). Prior analysis, samples were inactivated on dry ice by gamma-radiation (5 MRad). The assay was performed according to the manufacturer's instructions. The concentration of the following 28 cytokines were determined using the Bio-Plex 200 system (BioRad): Epidermal Growth Factor (EGF), Granulocyte-Colony Stimulating Factor (G-CSF), Granulocyte Macrophage-Colony Stimulating Factor (GM-CSF), interferon (IFN)-α2, IFNγ, Interleukin (IL)-1α, IL-1ß, IL-2, IL-3, IL-4, IL-5, IL-6, IL-7, IL-8, IL-10, IL-12(p40), IL-12(p70), IL-13, IL-15, IL-17A, chemokine ligand 3-like 1 (CCL3L1 or MIP-1α), chemokine ligand 4 (CCL4 or MIP-1ß), chemokine ligand 10 (IP-10 or CXCL10), chemokine ligand 11 (CCL11 or Eotaxin-1), chemokine ligand 13 (CCL13 or MCP-1), Tumor Necrosis Factor (TNF-α), Lymphotoxin alpha (TNFß) and Vascular Endothelial Growth Factor A (VEGF).
|
10.1371/journal.pgen.0030039 | Function-Altering SNPs in the Human Multidrug Transporter Gene ABCB1 Identified Using a Saccharomyces-Based Assay | The human ABCB1 (MDR1)-encoded multidrug transporter P-glycoprotein (P-gp) plays a major role in disposition and efficacy of a broad range of drugs including anticancer agents. ABCB1 polymorphisms could therefore determine interindividual variability in resistance to these drugs. To test this hypothesis we developed a Saccharomyces-based assay for evaluating the functional significance of ABCB1 polymorphisms. The P-gp reference and nine variants carrying amino-acid–altering single nucleotide polymorphisms (SNPs) were tested on medium containing daunorubicin, doxorubicin, valinomycin, or actinomycin D, revealing SNPs that increased (M89T, L662R, R669C, and S1141T) or decreased (W1108R) drug resistance. The R669C allele's highly elevated resistance was compromised when in combination with W1108R. Protein level or subcellular location of each variant did not account for the observed phenotypes. The relative resistance profile of the variants differed with drug substrates. This study established a robust new methodology for identification of function-altering polymorphisms in human multidrug transporter genes, identified polymorphisms affecting P-gp function, and provided a step toward genotype-determined dosing of chemotherapeutics.
| Patients often show varied drug responses ranging from lack of therapeutic efficacy to life-threatening adverse drug reactions. Drug therapy would be greatly improved if it were possible to predict individual drug sensitivity and tailor drugs to patients' genetic makeup. Like all other organisms, humans have a set of transporters and enzymes to detoxify and eliminate foreign molecules including drugs. Understanding the function of genetic variants in these proteins is a key goal toward personalized medicine. To that end, we examined the functional consequences of naturally occurring genetic variants in P-glycoprotein, the most versatile human multidrug transporter. A novel method was developed and employed that can identify function-altering variants in human transporters. This methodology was robust and powerful in that the functional effect of genetic variants can be directly assessed in yeast where all confounding variables in humans are excluded. Surprisingly, the majority of single amino acid substitutions were found to cause alterations in resistance to three tested anticancer agents. This study extends the impact of yeast-based medical research to a new niche, pharmacogenomics.
| Patients vary widely in their drug responses including unpredicted adverse drug reactions that cause a significant loss of lives and a huge toll on health-care costs [1]. Rational selection and dosage optimization of anticancer agents are particularly important due to their narrow therapeutic index and inherent cytotoxicity. Membrane transporters affect drug disposition and response by determining whether or not the level of drug is maintained within the therapeutic index. Of the known human transporters, P-glycoprotein (P-gp) is of particular clinical relevance in that this multidrug efflux pump has a broad range of substrates, including structurally and functionally divergent drugs in common clinical use [2–4]. P-Gp belongs to the ATP-binding cassette (ABC) superfamily [5] and is encoded by the human ABCB1 gene (also known as multidrug resistance 1 gene [MDR1]). Multidrug resistance caused by ABCB1 amplification is a major obstacle in cancer chemotherapy. In fact, the ABCB1 gene was originally identified because of its amplification in tumor cells that had acquired cross-resistance to multiple cytotoxic anticancer agents [2,6–9]. P-Gp is expressed in many tissues, suggestive of a broad physiological role [10,11] and functions by pumping cytotoxic drugs and xenotoxins out of cells into the intestinal lumen, bile, and urine, and thus limiting distribution of such compounds to other organs.
Genetic heterogeneity of the ABCB1 gene may be a potent determinant of interindividual variability in resistance to multiple drugs including anticancer agents. Furthermore, P-gp can act alone or in combination with other genetic variants, particularly polymorphisms in CYP3A4, a cytochrome P450 monooxygenase that metabolizes a wide range of drugs [12,13]. Naturally occurring null mutations in P-gp have been reported in mice and dogs but not in humans [14,15]. Animals carrying a null ABCB1 variant are viable unless challenged by drugs that are substrates for P-gp. Likewise, there may be unidentified human ABCB1 variants that cause a total loss of function. Numerous ABCB1 single nucleotide polymorphisms (SNPs) have been identified. However, the correlation of SNPs with ABCB1 expression and P-gp function in clinical pharmacokinetics has been inconclusive. A synonymous 3435C>T SNP has been heavily studied, but its function remains under debate [16]. Moreover, to date there have been no naturally occurring nonsynonymous substitutions with a validated functional consequence [17]. Robust functional assays of P-gp variants at the cellular and molecular levels are needed to address their impact on clinical pharmacokinetics.
Since human populations are outbred, and each individual is heterozygous for several million polymorphisms, the impact of ABCB1 variants is difficult to separate from the potential contributions of other variations in an individual. Yeast cells offer an excellent context for functional analysis of foreign eukaryotic transport proteins [18]. Expressing human proteins and their variants in yeast allows the function of individual variants to be assessed directly. The human P-gp can be functionally expressed in the yeast Saccharomyces cerevisiae, where it exports at least some of the same compounds that it exports in human cells [19]. A typical assay for human P-gp function in yeast involves testing its ability to restore growth to cells in the presence of compounds that would otherwise block their growth. This functional complementation in yeast allows the impact of ABCB1 variants found in human populations to be assessed.
This study tested the functional consequences of ABCB1 genetic variants found in ethnically diverse populations (Figure S1) [20]. From this dataset (http://pharmacogenetics.ucsf.edu or http://www.pharmgkb.org), we prioritized nonsynonymous SNPs by their predicted impact on P-gp function, selected ten haplotypes carrying high-priority SNP(s), and determined the level of resistance caused by these ABCB1 variants to clinically important drugs. For those variants that altered function, subsequent experiments tested the mechanism of these effects.
As the first step toward functional analysis of the nonsynonymous variants of human P-gp, we tested the sensitivity of yeast strains harboring mutations in major endogenous multidrug transporter genes, PDR5, SNQ2, and YOR1. Combinatorial deletions of these three genes confer sensitivity to a variety of toxic compounds including two anticancer agents, daunorubicin and doxorubicin, which are substrates for human P-gp [21]. The double mutant pdr5 yor1 (JRY8008) displayed increased sensitivity relative to wild-type cells toward doxorubicin, whereas another double mutant pdr5 snq2 (JRY8004) displayed increased sensitivity toward daunorubicin and doxorubicin. The strain that exhibited the greatest drug sensitivity was the pdr5 snq2 yor1 triple deletion mutant (JRY8012) (Figure 1A) (see Table S1 for the strain list). This result was reminiscent of bacterial multidrug efflux pumps that produce greater drug resistance in combination than alone [22].
To address the function of human P-gp in yeast, we used a plasmid (pJR2702) that contains a cDNA for the human ABCB1 gene expressed from the promoter for the S. cerevisiae STE6 gene on a multicopy vector [19]. The yeast STE6 gene encodes an ABC transporter that mediates the export of the a-factor pheromone in MATa cells. The cloned cDNA carried the G185V SNP of ABCB1, and therefore site-directed mutagenesis was used to restore it to the most common allele, referred to as the ABCB1 reference allele in the Pharmacogenetics of Membrane Transporters dataset (pJR2703) (http://pharmacogenetics.ucsf.edu or http://www.pharmgkb.org). Cells expressing the ABCB1 reference cDNA from the multicopy plasmid in the pdr5 snq2 yor1 strain showed highly increased resistance towards daunorubicin and doxorubicin relative to that of the pdr5 snq2 yor1 strain (Figure 1B). Thus the P-gp reference was functionally expressed in these yeast cells.
The Pharmacogenetics of Membrane Transporters study identified fourteen nonsynonymous SNPs in 247 healthy individuals from an ethnically diverse population (Figure S1) [20]. These SNPs comprised 25 haplotypes including 15 haplotypes in which the phase relationship of the SNPs was inferred but not directly resolved. SNPs were prioritized for functional analysis by two criteria: the degree of evolutionary conservation [23] and the biochemical severity of the alteration. The extent of evolutionary sequence conservation and thus inferred constraint at a particular residue was observed across ten mammalian species. The severity of missense changes was estimated by the Grantham scale [24], which formulates the difference in codon substitutions based on chemical dissimilarity of the encoded amino acids. Grantham values range between 5 and 215, with higher values indicating more radical chemical changes.
Out of the 14 nonsynonymous SNPs in the dataset [20], we chose seven SNPs for functional characterization (Table 1). We first focused on the five SNPs with highest Grantham values (>80): M89T, L662R, R669C, A893S, and W1108R. The M89T polymorphic site was not evolutionarily conserved, but the other four sites were highly conserved. In addition, the P1051A SNP was chosen because of its conservation despite a low Grantham value, and the S1141T SNP was included due to its relatively high allele frequency (11% in African Americans) and evolutionary conservation. Although A893S, S1141T, and R669C SNPs are common variants (minor allele frequency ≥1% in at least one major ethnic group), the remaining four chosen variants are observed only once among 494 alleles from different populations. These rare variants (minor allele frequency <1%) were included because rare adverse drug reactions may be due to highly penetrant but rare variants. The alignment and allele count of ABCB1 haplotypes based on the 14 nonsynonymous SNPs identified in the previous resequencing project are presented in Table S2. From the standpoint of functional impact, the R669C SNP was particularly interesting. First, this Arg-to-Cys substitution had the highest Grantham value (180) among the fourteen SNPs. Second, this SNP was observed twice in the African American population exhibiting a 1% allele frequency, whereas the four chosen rare variants occurred only once. Third, the R669C SNP may be in phase with the W1108R variant. One of the two R669C SNPs was detected in an individual whose ABCB1 gene also contained the W1108R variant, potentially resulting in haplotype R669C-W1108R. This observation prompted us to test whether a R669C-W1108R allele had a unique phenotype relative to alleles carrying each individual SNP.
We constructed plasmids expressing P-gp variants by site-directed mutagenesis on the reference plasmid to evaluate the effect of selected SNPs and their combinations on P-gp function. These plasmids (pJR2703–pJR2712), along with two control vectors (YEp352 and pJR2713), were transformed into the pdr5 snq2 yor1 strain (JRY8012). These yeast strains carrying plasmids with ABCB1 variants (JRY8025–JRY8036) were examined for their level of resistance to daunorubicin and doxorubicin on solid medium. Different P-gp variants displayed higher levels of resistance (A893S-M89T, L662R, and R669C) or lower levels of resistance (A893S, S1141T, A893S-R669C, A893S-P1051A, W1108R, and W1108R-R669C) relative to the P-gp reference (Figure 2A and 2B). The alleles varied widely in their ability to survive on high concentrations of daunorubicin and doxorubicin. The replacement of Arg669 by Cys led to one of the most drastic gain-of-function effects on the ability of P-gp to confer drug resistance. This allele's elevated resistance was compromised when in combination with W1108R. Cells expressing truncated P-gp (see Materials and Methods) were indistinguishable from cells transformed with an empty vector with respect to drug resistance.
To quantify the extent of drug cytotoxicity in liquid medium, median effective concentration (EC50) values were measured for daunorubicin and doxorubicin for each P-gp variant in liquid culture (Figure 2C). For the majority of the variants, these results were consistent with those observed in the plate assay. However, the plate assay was more sensitive, allowing variants that were indistinguishable from each other in the liquid assay to be ranked. There was a discrepancy between the two drug resistance phenotypes with the A893S and A893S-R669 variants: the variants showed a slightly higher level of drug resistance relative to that of the reference in the liquid assay, but a lower survival in the plate assay. This difference presumably reflects the nature of the two assays: the plate assay measures the level of cell survival on a relatively high fixed concentration of the drug, whereas the liquid assay determines growth rate over multiple drug concentrations. In the plate assay, all variants for daunorubicin and six variants for doxorubicin exhibited statistically significant differences (p < 0.05) (Figure 2B; Table S3). In the liquid assay, three variants for daunorubicin (A893S-R669C, A893S-M89T, and R669C) and five variants for doxorubicin (A893S, S1141T, A893S-M89T, L662R, and R669C) exhibited statistically significant increases in EC50 values (p < 0.05) (Figure 2C; Table S4).
To determine whether the observed differences in drug resistance were due to differences in protein level, we measured the protein level of each P-gp variant by immunoblotting. The mouse anti-P-gp antibody detected P-gps with an apparent molecular mass of 125 kDa, the expected size of unglycosylated P-gp, in membranes from yeast cells transformed with plasmids carrying reference and variant ABCB1 genes, but not in membranes from control cells transformed with an empty vector. The amount of P-gp reference and variants differed by no more than 1.5-fold (Figure 3A). The correlation coefficient of the extent of daunorubicin cytotoxicity of each variant relative to the protein level of each variant was 0.227 (Figure 3B). Thus the P-gp variants were present at comparable levels and altered drug cytotoxicity in the variants was not due to the differences in protein levels for P-gp.
In principle, the differing drug resistance of the variants might reflect differences in their subcellular localization if the SNPs affected the P-gp trafficking. To test this possibility, strains carrying green fluorescent protein (GFP) fused in frame to the C terminus of each P-gp variant were evaluated for their subcellular localization patterns. Fluorescence microscopy indicated that the fusion proteins were localized to both the plasma membrane and the vacuolar membrane in living cells (Figure 3C). The localization patterns were growth-phase–dependent: GFP fluorescence was observed mostly in the plasma membrane in mid-log phase cells and became more concentrated in vacuoles when the cells were grown into the stationary phase. The cells carrying each of the GFP-fused P-gp variants fluoresced to similar extents from the same subcellular location under each growth phase. Thus differences in subcellular localization were unlikely to underlie the differences in drug resistance associated with the variants.
The relative resistance of each P-gp variant to the structurally similar drugs, daunorubicin and doxorubicin, were quite similar (Figure 2). Because P-gp can confer cellular resistance to a variety of cytotoxic drugs, we tested whether P-gp variants might exhibit different resistance profiles when tested with additional P-gp substrates, valinomycin and actinomycin D, which are structurally dissimilar from daunorubicin and doxorubicin. Due to the limited solubility of valinomycin in synthetic (CSM)–Ura culture medium, determining the EC50 values was not possible. However, determining the EC30 proved sufficient to distinguish among the P-gp variants for valinomycin resistance (Figure 4A). Although some alleles showed similar trends of resistance for valinomycin and daunorubicin/doxorubicin, others (e.g., S1141T, W1108R, and W1108R-R669C) were qualitatively different in their resistances.
Yeast MATa ste6 strains, which lack the a factor pheromone transporter, are reported to be more sensitive to actinomycin D than wild-type strains [25]. This prompted us to investigate the interesting possibility that MATa cells are intrinsically more resistant to actinomycin D than MATα cells. Indeed MATα cells were dramatically more sensitive to actinomycin D (EC50 15 μg/ml) than MATa cells (EC50 48 μg/ml). To see if the cytotoxicity profile pattern of P-gp variants is changed with actinomycin D, all variants were tested in a MATa ste6 strain (JRY8572) for their levels of resistance to actinomycin D (JRY8573–JRY8584) (Figure 4A).
We tested the statistical significance of all comparisons between the reference and each variant for each drug (Table S4). Five variants (S1141T, A893S-R669C, A893S-M89T, L662R, and R669C) exhibited a statistically significant increase in EC50 or EC30 values for two or more drugs. The A893S and A893S-P1051A variants caused an increase in resistance only for doxorubicin and valinomycin, respectively. The compromising effect of W1108R on R669C was obvious in resistance for all four drugs (Figures 2 and 4A). To see if the relative resistance profile of the P-gp variants to one substrate was predictive of the relative resistance profile to other substrates, we determined the correlation coefficient for all combinatorial pairs of the four relative resistance profiles (Figure 4B). The resistance profiles of three anticancer agents (daunorubicin, doxorubicin, and actinomycin D) were highly correlated to each other, whereas the resistance profile of valinomycin exhibited a relatively low degree of correlation with those of the other three drugs.
To understand the correlation between ABCB1 polymorphisms and altered cellular pharmacokinetics, we have developed functional assays of P-gp variants in yeast cells. The function of nonsynonymous SNPs was quantitatively measured in isolation from all other variations in the human genome in a yeast-based in vivo assay. The most sensitive measure of drug transport was a colony-counting assay, which provided both qualitative and quantitative measures of drug resistance in yeast expressing reference and variant P-gp. We observed multiple differences caused by the P-gp variants in the level of resistance to the anticancer agents, daunorubicin, doxorubicin, and actinomycin D, and the potassium ionophore valinomycin. The functional consequences of five ABCB1 polymorphisms were previously unknown: the M89T, L662R, R669C, and S1141T variants were associated with increased resistance to two or more drugs; and the W1108R variant strongly mitigated the impact of R669C on gain of P-gp function (Figures 2 and 4A). Due to its high allele frequency (11% in African Americans), the S1141T SNP in particular deserves further attention to define its clinical significance. As measured by plating efficiency in an acute exposure test, the difference between the reference and most sensitive (W1108R) alleles was approximately 30-fold. In a chronic exposure involving growth in the presence of the drug, like most quantitative comparisons of the activity of single amino acid substituted P-gp mutants in the published data, the differences among the P-gp variant alleles in EC50 or EC30 values were modest in most cases. The functional variations can be magnified in clinical practice, especially for anticancer agents due to ABCB1 amplification in cancer patients. In previous studies, the A893 variant, which is the most common SNP, caused either no significant functional impact [20,26,27] or increased P-gp function for digoxin efflux [28]. The data shown here were able to detect an effect of this allele and uncovered unexpected complexity in the response. In the acute assay, A893S cells were significantly more sensitive to both daunorubicin and doxorubicin than cells with the reference allele (Figure 2A and 2B). In contrast, in the chronic assay the A893S allele was indistinguishable from the reference allele with respect to daunorubicin and slightly more resistant to doxorubicin (Figure 2C).
Like variants of facilitated drug influx pumps in the solute-carrier superfamily, P-gp variants that increased function were common. Most random changes in protein sequence are expected to be deleterious or neutral. The significant enhancement of function common to the alleles tested here may reflect a recent adaptation of human populations to local conditions like toxin exposure, leading to selective pressures on medically relevant phenotypes. Interestingly, in Europeans CYP genes encoding drug-metabolizing enzymes show strong signals of very recent positive selection [29].
Despite its distinct chemical structure, the resistance profile of actinomycin D showed a high level of correlation with those of the other anticancer agents, daunorubicin and doxorubicin (Figure 4B). Valinomycin, which lowers the mitochondrial membrane potential, inducing apoptosis in some cell lines [30], exhibited a low correlation in resistance relative to other drugs, presumably reflecting differences among P-gp variants in recognition or transport of the drugs. The resistance profiles of the S1141T, W1108R, and W1108R-R669C variants showed the largest variation across substrates. Based on this finding, we speculate that the region containing W1108 and S1141 contributes to the substrate discrimination activity of P-gp. To date, all mutations that alter substrate specificity of P-gp have been located in the transmembrane domains [16]. In contrast, all seven SNPs for which functional consequences were determined in this study are located either in the extracellular region (M89T) or in the cytoplasmic region (the remaining six variants).
We used two widely accepted criteria for predicting the functional effect of uncharacterized SNPs to prioritize for functional characterization (Table 1). Our data on functional consequences revealed that these predictions were sound: four functional SNPs (L662R, R669C, W1108R, and S1141T) scored highly on both criteria, while the two SNPs (A893S and P1051A) that showed no significant functional impact had lower scores on evolutionary conservation and chemical dissimilarity, respectively. One exception was the M89T variant that altered function despite being poorly conserved among mammals.
Most previous functional studies focused on the impact of individual SNPs rather than that of haplotypes. However, in at least some cases, drug response correlates with the patients' haplotypes rather than individual SNPs [31,32]. We tested SNP interactions to see if a compound allele consisting of two SNPs has a unique phenotype different from those of single-SNP alleles. Indeed, it is striking that the strong impact of R669C on P-gp function diminished almost completely when combined with W1108R (Figures 2 and 4A). In contrast, the W1108R variant either alone or with A893S contributed no significant alterations in EC50 or EC30 values. This result highlighted the importance of testing the impact of all substitutions in a gene together and suggests that compensatory SNPs may exist in nature.
SNPs in the ABCB1 gene have been implicated in altering drug response or susceptibility to diseases such as Parkinson's disease [33], inflammatory bowel disease [34], and renal epithelial tumors [35]. However, in many such cases, the reported effects of ABCB1 polymorphisms are conflicting or inconsistent [26,36–38]. This inconsistency may have several causes. First, P-gp expression levels may be modified by nongenetic factors, such as diet and comedications, especially when surgical specimens are studied. Second, previous studies with mammalian cell lines rely on transient expression assays, which swamp the subtle effects of SNPs by variable levels of expression. Third, only a few coding SNPs have been functionally tested, such as A893S and N21D, which our analysis predicted would have a weak functional impact [26]. The use of yeast to evaluate the function of nonsynonymous coding SNPs bypasses these issues and allows the function of single coding SNPs and haplotypes to be assessed directly, independent of all other variations in their original human genome. This “in yeast pharmacogenetics” can function as a robust screening and phenotyping tool to characterize additional SNPs in ABCB1 and presumably other human multidrug transporter genes.
During the course of these studies, we observed that MATα cells were highly sensitive to actinomycin D, whereas MATa cells were resistant. This was apparently due to expulsion of the drug by the a cell-specific Ste6 transporter. Perhaps chemical exposures in ecological niches or the consequences of treatment with therapeutics might lead to the extreme mating-type biases observed with some fungal pathogens. For example, the mating-type–specific niches occupied by Cryptococcus neoformans may reflect the ability to transport toxins out of the cell in certain environments [39].
S. cerevisiae strains used in this study are listed in Table S1. Standard rich medium (YPD), CSM, and synthetic medium lacking nutritional supplement(s) (CSM–Ura, CSM–His, and CSM–Ura–Trp) were prepared as described [40]. Yeast cells were grown routinely at 30 °C.
A P-gp-expressing plasmid, pJR2702 (alias pYKM77; a multicopy-number vector), was kindly provided by Jeremy Thorner (University of California, Berkeley, California, United States) and used for constructing expression plasmids for ABCB1 bearing different SNPs. A cDNA for the human ABCB1 coding sequence (GenBank accession number M14758.1) was cloned into a multicopy URA3-marked plasmid with the 2 μm origin of replication (YEp352) and expressed from the yeast STE6 promoter (pJR2702). Substitutions at the SNP position were carried out in pJR2702 by site-directed mutagenesis with primers designed to generate individual haplotypes (Table S5), using the QuikChange site-directed mutagenesis kit from Stratagene (http://www.stratagene.com). We introduced five single SNP alleles and four compound alleles consisting of a two-SNP haplotype into the reference plasmid (pJR2703), creating plasmids pJR2704 to pJR2712 (Table S1). As a negative control, a −1 frameshift mutation at codon 1,200 of the ABCB1 sequence (1,280 amino acids) was constructed; this cDNA encodes a truncated product of 1,228 amino acids expected to be nonfunctional when expressed (pJR2713). Presence of the desired substitution in the plasmids was verified by DNA sequencing. These eleven constructs, along with another control lacking the entire ABCB1 sequence (pJR1016), were transformed into a MATa yeast strain lacking three different ABC transporter genes (Δpdr5 Δsnq2 Δyor1, JRY8012), resulting in strains JRY8025 to JRY8036 (Table S1).
Daunorubicin and doxorubicin were kindly provided by Robert Schultz in the Developmental Therapeutics Program of the National Cancer Institute, National Institutes of Health (NIH) (Rockville, Maryland, USA). Valinomycin and actinomycin D were from Sigma (http://www.sigmaaldrich.com). For drug cytotoxicity assays, stock solutions of the drug were prepared at 10 mM in 5% DMSO for daunorubicin and doxorubicin, in 98% ethanol for valinomycin, and in 100% DMSO for actinomycin D.
In the spotting assay, cultures from each strain were grown to midexponential phase, titrated to the same concentration (~107 cells per 1 ml), and serially diluted 5-fold. Aliquots (4 μl) from the dilution series were spotted onto a CSM–Ura plate containing the indicated concentration of the drug. Control plates lacking the drug contained the solvent control at the same concentrations as CSM–Ura plates containing the drug. In the plate assay, cultures from each strain were grown to midexponential phase and titrated to the same concentration (~105 cells per 1 ml). Aliquots (100 μl) were spread onto a CSM–Ura plate containing the indicated concentration of the drug. The same aliquots were further diluted 20-fold (~5,000 cells per 1 ml) and spread onto control plates lacking the drug. After incubation for three days, colony numbers per plate were counted.
Drug resistance was further assayed quantitatively in 96-well microtiter plates (Corning, http://www.corning.com), containing equal volumes (200 μl) of CSM–Ura liquid medium with different concentrations of the drug. Yeast transformants grown to stationary phase in CSM–Ura were diluted to an OD600 of 0.1. Equal volumes (200 μl) of these diluted cultures containing increasing concentrations of the drug were added to wells and incubated at 30 °C for 24 h in a Tecan microtiter plate reader. Cell growth was monitored in the absence of the drug in the presence of the same solvent as a negative control. For the experiments with liquid medium, the EC50 (median effective concentration) value was defined as the drug concentration that reduced growth of the treated cells to 50% of growth of the control cultures as judged by OD600 when the increase in OD600 of the control cultures was 0.7 (midexponential phase). To rule out the possibility that variations in copy number affect the observed differences in drug resistance for the vectors bearing each P-gp variant, all measurements were examined in a series of independent transformants for each of the P-gp variants.
Membrane fractions of yeast cells with the plasmids bearing ABCB1 variants (JRY8025–JRY8036) were prepared as described [41].
The mouse monoclonal anti-P-gp antibody C219, kindly provided by Michael Gottesman (National Cancer Institute, NIH, Bethesda, Maryland, United States), was used in immunoblots to quantify the level of P-gp variants in yeast. A rabbit antibody against the Gas1 protein, kindly provided by Randy Schekman (University of California, Berkeley, California, United States), served as a marker of membrane proteins. Human P-gp and yeast Gas1 protein were detected simultaneously on the same blot using infrared-labeled secondary antibodies visualized at two different fluorescence channels, 700 and 800 nm. The blot was developed and quantified by Odyssey Infrared Imaging System (LI-COR Biosciences, http://www.licor.com) following the manufacturer's protocol.
A codon-optimized GFP gene for yeast, yEGFP1 [42], was amplified by PCR with oligonucleotide primers designed to allow in-frame fusion to the 3′ end of ABCB1 reference and its variants in a yeast expression vector by recombination following transformation into yeast [43]. The presence of yEGFP in the construct was verified by colony PCR and DNA sequencing.
For fluorescence microscopy, cells were grown in synthetic medium without tryptophan to minimize autofluorescence. Imaging was done at room temperature using an Olympus IX-71 microscope equipped with 100× NA1.4 objectives and Orca-II camera (http://www.olympusamerica.com). ImageJ (http://rsb.info.nih.gov/ij) was used for manipulation of images.
The probability of a statistically significant difference between the mean values of two datasets was determined by one-way ANOVA with Dunnett's post-test using GraphPad Prism version 4.03 for Windows, GraphPad Software (http://www.graphpad.com).
The Entrez (http://www.ncbi.nlm.nih.gov/Entrez) accession numbers for the genes described in this paper are 5243 for human ABCB1, 1576 for human CYP3A4, 854324 for yeast PDR5, 851574 for yeast SNQ2, 853198 for yeast YOR1, 853671 for yeast STE6, and 855355 for yeast GAS1.
The RefSeq (http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?db=OMIM) accession number for human ABCB1 cDNA carried in plasmid pJR2702 is M14758.1.
The Online Mendelian Inheritance in Man (http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?db=OMIN) accession numbers are 168600 for Parkinson's disease, 266600 for inflammatory bowel disease, and 144700 for renal epithelial tumors. |
10.1371/journal.ppat.1007711 | Molluscum contagiosum virus MC80 sabotages MHC-I antigen presentation by targeting tapasin for ER-associated degradation | The human specific poxvirus molluscum contagiosum virus (MCV) produces skin lesions that can persist with minimal inflammation, suggesting that the virus has developed robust immune evasion strategies. However, investigations into the underlying mechanisms of MCV pathogenesis have been hindered by the lack of a model system to propagate the virus. Herein we demonstrate that MCV-encoded MC80 can disrupt MHC-I antigen presentation in human and mouse cells. MC80 shares moderate sequence-similarity with MHC-I and we find that it associates with components of the peptide-loading complex. Expression of MC80 results in ER-retention of host MHC-I and thereby reduced cell surface presentation. MC80 accomplishes this by engaging tapasin via its luminal domain, targeting it for ubiquitination and ER-associated degradation in a process dependent on the MC80 transmembrane region and cytoplasmic tail. Tapasin degradation is accompanied by a loss of TAP, which limits MHC-I access to cytosolic peptides. Our findings reveal a unique mechanism by which MCV undermines adaptive immune surveillance.
| The presentation of antigenic peptides by classical MHC-I to cytotoxic T-cells is a cornerstone of antiviral immunity. As such, viruses have devised a plethora of strategies to target MHC-I or cellular components involved in MHC-I antigen presentation in order to block effective T-cell surveillance. Molluscum contagiosum virus (MCV) is a human-specific poxvirus that produces skin lesions that can persist for months on a healthy individual. Herein, we demonstrate that MCV encodes a protein, MC80, which disrupts MHC-I antigen presentation by associating with the peptide loading complex, a set of proteins responsible for transporting peptides into the endoplasmic reticulum and loading them into a groove on MHC-I proteins. These MHC-I/peptide complexes are then trafficked to the plasma membrane for presentation to T-cell receptors. MC80 shares sequence similarity with classical MHC-I, and like MHC-I, requires β2m to function. We have found that MC80 engages a peptide loading complex specific chaperone, tapasin, which leads to its ubiquitination and subsequent degradation along with the TAP peptide transporter. This prevents peptides from being trafficked into the endoplasmic reticulum and severely limits peptide loading onto MHC-I. Thus, MCV-infected cells expressing MC80 can be protected from T cell killing due to the sabotage of cell surface MHC-I/peptide presentation.
| Molluscum contagiosum virus (MCV) is a phylogenetically distinct poxvirus with a significant global disease burden [1,2]. MCV infections are thought to be restricted to humans, producing cutaneous lesions which often lack signs of an inflammatory response and persist for months to years in otherwise healthy individuals. This is in stark contrast to well characterized orthopoxviruses and parapoxviruses, which generally have a broader tropism and present as acute inflamed infections [1,3]. The persistence of MCV appears to be coupled to its ability to remain undetected by the immune system, as inflammatory responses have been implicated in spontaneous regression of MCV lesions [4,5]. Consistently, immunodeficient individuals are prone to exacerbated MCV infections [6,7]. The clinical features of MCV infections highlight a unique interplay between the virus and the human immune system.
MCV has clearly become well adapted to its niche, encoding a repertoire of proteins which are capable of evading the immune responses of the human epidermis. Yet, of the 59 open reading frames (ORFs) which distinguish MCV from orthopoxviruses [8,9], only nine have been well characterized [10,11,12,13,14,15,16,17,18]. Studies regarding the pathogenesis and immune evasion mechanisms employed by MCV have been severely limited by the lack of an animal model or cell line to propagate the virus [19]. Nevertheless, the ORFs unique to MCV likely play a major role in ensuring human-specific epidermal persistence, and should thus be more thoroughly characterized.
Among the MCV ORFs without a known function, MC80R shares moderate sequence similarity with the α1–3 domains of classical MHC-I. Given the central role of MHC-I and MHC-I-like proteins in inhibiting natural killer cells (NKs), it was long suspected that MC80 may be involved in NK subversion [9,20]. However, unlike classical MHC-I which is presented on the cell surface to NK and T cells, MC80 appeared to be retained in the endoplasmic reticulum (ER) [20]. As MC80 did not come to the cell surface, its host target and function have remained elusive.
Several large DNA viruses, particularly herpesviruses, have been found to repurpose the MHC-I fold in order to evade cell-mediated immune defenses. The specific function of each of these viral MHC-I-like proteins is highly related to its cellular localization. Cell surface viral MHC-I-like proteins (e.g. murine cytomegalovirus (MCMV) m157, human cytomegalovirus (HCMV) UL18) generally function as ligands for NK-inhibitory receptors without concurrently presenting viral peptides to cytotoxic T lymphocytes (CTL) [21]. However, intracellular MHC-I-like proteins (e.g. MCMV m145/m152/m155) are not directly exposed to NKs or CTLs. Instead, these proteins tend to localize to the ER/Golgi/lysosomal compartments where they retain or lead to the degradation of NK-activating ligands and, in some cases, classical MHC-I. Additionally, secreted viral MHC-I-like proteins have been identified which act as competitive antagonists of TNFα signaling and NKG2D-mediated NK activation (tanapox 2L and cowpox OMCP, respectively) [22,23,24,25]. Thus, through prolonged co-evolution, large DNA viruses have used the MHC-I fold for a diverse array of immune evasion functions.
As opposed to viral MHC-I-like proteins, vertebrates utilize classical MHC-I to display a repertoire of peptides on the cell surface to surveilling CTLs [26,27]. This pivotal role demands rapid yet stringent quality control in the assembly of MHC-I/peptide complexes. To accomplish this, host cells utilize a multi-subunit peptide loading complex (PLC); comprised of tapasin (Tpn), transporter associated with antigen processing (TAP), ERp57, and calreticulin (CRT) [28,29]. In the ER, nascent MHC-I heavy chains (HC) are initially stabilized by a chaperone, calnexin (CNX), and subsequently by heterodimerization with β2m [30]. However, HC/β2m must assemble with a high-affinity peptide, usually via the PLC, in order to efficiently traffic from the ER to the cell surface [31]. HC/β2m accomplishes this by directly binding Tpn/ERp57, which plays a central role in both transiently stabilizing the unloaded conformation of MHC-I and bridging the interaction between MHC-I and TAP [32,33]. TAP functions by transporting short cytosolic peptides into the ER, allowing PLC-associated MHC-I to sample the repertoire of proteasome-degraded proteins within the cell [34,35]. Once an MHC-I molecule has been loaded with a high-affinity peptide, it dissociates from the PLC and traffics to the cell surface. Given the critical role played by the PLC in peptide loading, both MHC-I itself and components of the PLC provide attractive targets for viruses seeking to subvert CTL responses [28]. However, the magnitude of MHC-I downregulation depends on the specific target and the MHC-I alleles expressed by the host cell. Understanding these viral mechanisms provides insight into the pathogenesis of viral infections, as well as the underlying cellular pathways that these viruses exploit.
Here we demonstrate that expression of MC80 results in ER-retention and consequent surface downregulation of classical MHC-I in human and mouse cells. Mechanistically, we found that MC80 interacts with Tpn via its luminal domain and targets Tpn for ER-associated degradation in a transmembrane (TM)- and cytoplasmic tail-dependent manner. The loss of Tpn coincides with a loss of TAP, further impeding the assembly of MHC-I with high-affinity peptides. Our findings reveal a strategy employed by MCV to disrupt antigen presentation and thereby CTL responses by exploiting MHC-I fold recognition by the PLC.
While MCV does not encode an ORF with sequence-similarity to any viral protein known to downregulate MHC-I, previous studies have suggested that MCV may be downregulating MHC-I and β2m in human lesions [5,36]. Given that some poxviruses do not appear to subvert MHC-I antigen presentation [1,37], we hypothesized that the MCV ORF(s) responsible for MHC-I downregulation may be unique to MCV. Additionally, as MHC-I traffics through the ER/Golgi to the plasma membrane, we limited our initial screen to the four MCV-specific ORFs which are predicted to encode type-1 transmembrane proteins (MC3, MC33, MC80, and MC157). We cloned the respective MCV-1 variants into an IRES-GFP retroviral vector (pMXsIG), replacing each predicted signal peptide with the mouse β2m signal peptide and an N-terminal Flag tag. Following transient transfection of human embryonic kidney (HEK-293T) cells with these constructs or vector control, we found that MC80 dramatically decreased the level of cell surface MHC-I by 2–4 days post-transfection (Fig 1A).
MC80 is well conserved among known MCV strains, sharing 24–36% amino acid identity to the ectodomains of human classical and non-classical MHC-I. The MC80 ORF has at least two potential start codons N-terminal to the MHC-I like α1 domain, termed MC80L and MC80S, which both provide unusually long signal peptides (S1A Fig). While the ectodomains of MC80 share moderate sequence-similarity with MHC-I, functionally distinct regions exhibit varying levels of conservation (S1A Fig). Briefly, residues involved in peptide binding and PLC interactions are not well conserved between MC80 and classical MHC-I (S1A Fig) [38,39,40]. In contrast, residues known to be involved in β2m binding are well-conserved, consistent with a previous study which found that MC80 associates with β2m [20,41]. This same study found that MC80 did not traffic to the cell surface, which we were able to recapitulate by flow cytometry and EndoH sensitivity assays, indicating that MC80 is retained in the ER (S1B and S1C Fig).
The cell surface half-life of MHC-I can be greater than 24 hours depending on the specific cell line and peptide(s) displayed [42,43]. We investigated the kinetics of MHC-I surface downregulation using a transient transfection system, finding that maximal expression of MC80 at 24 hours post transfection (hpt) did not coincide with maximal downregulation of MHC-I surface expression (Fig 1B). Instead, we found a significantly lower level of surface MHC-I at 72hpt than 24hpt, even though there was less MC80 at the later time point. Additionally, our data demonstrated that continued expression of MC80 was necessary to maintain MHC-I downregulation (Fig 1B). As extended HLA class I half-lives may have played a role in the apparent time-dependence of MC80-mediated MHC-I downregulation (Fig 1B), we employed a stable retroviral transduction system to achieve a steady state of MHC-I downregulation in further analyses. This time-dependence may also provide insight into why the previous MC80 study reported no change in surface expression of HLA-A2 12 hours post-infection of an MC80-expressing vaccinia virus [20].
Since viral MHC-I evasion mechanisms can downregulate MHC-I with varying levels of promiscuity [44,45], we next sought to determine the specificity of MC80-mediated MHC-I downregulation. Flow cytometry analysis demonstrated that expression of MC80 markedly decreased the surface levels of classical MHC-I in multiple human cell lines, including HEK 293T, Hela (human cervical cancer cell line), and HFF-1 (a human foreskin fibroblast cell line) (Fig 2, S2 Fig). A comparable effect was observed with untagged MC80 constructs with the canonical signal peptide, indicating that the Flag tag did not significantly affect MC80 function (Fig 2A). Interestingly, MC80 also downregulated all tested alleles of classical MHC-I in murine cell lines (Fig 2B, 2C, 2D and 2E). Additionally, of the non-classical MHC-I proteins examined, MC80 significantly decreased the surface expression of Qa-1 but did not significantly affect surface expression of CD1d or the NKG2D ligands MICA and Rae1a (Fig 2C, 2D and 2E). Like classical MHC-I and Qa-1, CD1d requires β2m for stable expression, indicating that MC80 is unlikely to function by competing for β2m. Instead, MC80 appears to be specifically downregulating peptide-binding MHC-I through a cellular component/pathway that is conserved between humans and mice.
As viruses are well known to have strategies to downregulate classical MHC-I by altering the cellular trafficking of MHC-I [28], we next examined the maturation state of Ld by EndoH sensitivity in the presence or absence of MC80 (Fig 3A). We found that MC80 did not appear to affect the steady state expression level of Ld. However, compared to the 35% of EndoH-resistant Ld in the control, we found that Ld was completely EndoH-sensitive in MC80-expressing cells (Fig 3A, S3A Fig). This indicates that MC80 interferes with MHC-I trafficking to the Golgi, which consequently decreases the extent of surface MHC-I.
Virally-encoded MHC-I saboteurs are further dichotomized into PLC-dependent and PLC-independent mechanisms; as exemplified by the cowpox virus CPXV012 and CPXV203 proteins, respectively [46,47,48,49]. To determine which strategy is utilized by MC80, we next tested whether MC80 downregulates MHC-I in murine embryonic fibroblasts (MEF) expressing SIINFEKL, a Kb-specific peptide of egg ovalbumin, either in the cytosol or in the ER. Cytosolic SIINFEKL requires TAP to be transported into the ER for MHC-I loading, whereas ER-SIINFEKL is able to load onto MHC-I independent of TAP function. Because CPXV203 does not require the PLC in order to downregulate MHC-I surface expression, it dramatically affects Kb/SIINFEKL expression in both cell lines (S3B Fig). However, CPXV012 only induces significant downregulation of Kb/SIINFEKL in cells expressing cytosolic SIINFEKL, as its mechanism of action is dependent on TAP [49]. Similar to CPXV012, MC80-mediated downregulation of Kb/SIINFEKL could be rescued by expression of SIINFEKL in the ER (Fig 3B, S3B Fig). SIINFEKL localization had only a marginal effect on the surface expression of Db in the presence of MC80, indicating that this effect was specific to Kb/SIINFEKL (Fig 3B).
While TAP- and Tpn-deficient cells display low levels of MHC-I on the cell surface, these levels can be further decreased by PLC-independent viral mechanisms, such as CPXV203. However, we found that MC80 functionally relies on the presence of TAP and Tpn, as murine MHC-I alleles were not further downregulated by MC80 in TAP- or Tpn-deficient cells, relative to vector control (Fig 3C). Together, (1) the lack of MHC-I-maturation, (2) the specific rescue of MHC-I by an ER targeted peptide, and (3) the TAP-/Tpn-dependence collectively suggest that MC80 sabotages the PLC-assisted peptide transport/loading of MHC-I in the ER.
Given the central role of Tpn in peptide loading and the partial conservation of Tpn-binding residues within MC80 (S1A Fig), we hypothesized that MC80 may subvert peptide loading by competitively binding Tpn to block the interaction between MHC-I and the PLC. A Flag-IP of MC80 followed by western blotting (WB) for PLC components supported this hypothesis; demonstrating that Tpn, TAP, CRT and CNX co-immunoprecipitate (co-IP) with the full length MC80 constructs (Fig 4B). While the N-terminal Flag-tag did not appear to impact MHC-I downregulation in Fig 2A, we observed that an anti-Flag WB of F-MC80S (N-terminally Flag-tagged MC80S) produced a laddering effect in non-reduced samples. Therefore, we used C-terminal Flag-tagged constructs, encoding the canonical signal peptide of MC80, for all further co-IP experiments.
Notably, a truncated MC80S protein which lacks the putative TM and cytoplasmic tail (sol MC80S-F) could also co-IP Tpn, CRT and CNX but not TAP1 in murine cells (Fig 4B). This suggests that MC80 primarily associates with the PLC via the luminal domain and that the interaction between MC80 and TAP1 may be further stabilized by the TM and cytoplasmic tail. Using TAP1-deficient MEFs treated with interferon gamma (IFNɣ), we were able to recapitulate the association of the MC80 luminal domain with Tpn, CRT and CNX. However, in the absence of Tpn, we could not detectably co-IP TAP1 or CRT with MC80, even when the Tpn-/- MEFs were treated with IFNɣ. These data suggest that MC80 interacts with CNX and Tpn via the luminal domain; and the latter association may bridge the interaction of MC80 with CRT and TAP1, reminiscent of classical MHC-I assembly in the ER. A co-IP of PLC components with MC80 in 293T cells demonstrated that soluble MC80S-F interacts with both Tpn and TAP1, further suggesting that the TM and tail of MC80 may not be necessary for the association of MC80 and TAP (S4A Fig). Despite the observed PLC-associations, flow cytometry demonstrated that soluble MC80 constructs do not markedly downregulate surface MHC-I in MEFs or HFF-1s (Fig 4C, S2B Fig). As binding to the PLC appears to be insufficient for MC80 function, Tpn-competition/blockade is unlikely to be the mechanism of MHC-I downregulation, as we had hypothesized.
To determine the role of the transmembrane and tail in MC80-mediated MHC-I downregulation, we assessed the steady state levels of PLC components in the presence of various MC80 constructs. Remarkably, we found that MC80L and MC80S, but not soluble MC80S, dramatically reduced the steady state levels of Tpn and TAP compared with vector control in MEFs and HEK 293Ts (Fig 5A and 5C). However, CNX, CRT, and ERp57 were not downregulated by the expression of MC80 in MEFs. While the soluble MC80S construct slightly increased β2m levels, they appeared unchanged by active forms of MC80. These data suggest that MC80 selectively destabilizes Tpn and TAP in a TM/tail-dependent manner. While full-length forms of MC80 downregulated Tpn, the completion of tapasin degradation appears to be cell-line specific. For instance, the MEF-Ld cells in Fig 4B (left panel) were not treated with drug or cytokine to rescue Tpn levels, yet all MC80 constructs associated with Tpn. However, only soluble MC80S-F appeared to associate with Tpn in untreated TAP-/- MEFs and 293T cells, presumably due to the degradation of Tpn by the functional MC80 constructs (S4 Fig). Therefore, to demonstrate the ability of functional MC80 to associate with Tpn in the absence of TAP, we treated the Tpn-/- and TAP-/- MEFs with mIFNɣ (Fig 4B; right panel). F-MC80L also downregulated Tpn in Hela cells in the presence of IFNɣ, but did not markedly affect the steady-state levels of TAPBPR, a structural relative of Tpn known to interact with classical MHC-I (Fig 5D) [40].
Given the interdependence of Tpn and TAP, we next sought to determine whether MC80 primarily targeted one component or both equivalently. Using TAP1-/- MEFs, we were able to recapitulate the MC80 TM/tail-dependent degradation of Tpn observed in wildtype MEFs (Fig 5B). However, the level of TAP in Tpn-/- cells expressing MC80 was comparable to the vector control. This was more readily observable when the Tpn-/- cells were treated with IFNɣ for 24hr prior to harvesting, to upregulate TAP expression (Fig 5B, right panel). Thus, while the destabilization of TAP by MC80 depends on the presence of tapasin, the MC80-mediated loss of tapasin is independent of TAP, indicating that tapasin is the primary target of MC80 in murine cells. TAP destabilization is potentially a consequence of the loss of Tpn, given that both this and previous studies demonstrate that TAP is generally unstable in the absence of Tpn (Fig 5B, left panel) [50,51]. We hypothesize that tapasin is also the primary target in human cells due to the homology of murine and human PLC components. However, as murine TAP is apparently more Tpn-dependent than human TAP, our data does not rule out the possibility that MC80 directly targets TAP for degradation in human cells.
While previous work demonstrated that MC80 associates with β2m, it was not clear whether this association was functionally relevant. To determine whether β2m was necessary for MC80-mediated downregulation of Tpn, we used a classical MHC-I (H2-Kb/H2-Db) and β2m triple knock-out MEF cell line (3KO) with or without stably transduced β2m. As observed in wild-type MEFs, the soluble form of MC80S-F did not cause Tpn degradation in either cell line. However, β2m expression was necessary for MC80S-F to induce Tpn degradation (Fig 6). In addition, the level of MC80 in 3KO cells without β2m is lower than that in 3KO cells with β2m transcomplementation. Given that the cistronically-translated GFP was expressed at similar levels, these findings indicate that, (1) β2m can stabilize MC80, (2) β2m is required for MC80 function, and (3) MC80-mediated degradation of Tpn is classical MHC-I-independent.
The majority of eukaryotic protein degradation is mediated through proteasomal and lysosomal pathways [52]. Autophagy has also been implicated in trafficking proteins from the ER to the lysosome/autophagosome for degradation [53]. To determine which host degradation pathway was being exploited by MC80, we assessed the effects of two inhibitors of proteasomal degradation (MG132 and Epoxomicin) and one inhibitor of lysosomal degradation (chloroquine). We also assessed an Atg5-/- murine microglial cell line, which is deficient in classical autophagy. Due to the toxicity of the tested drugs and the slow intrinsic turnover of Tpn, we treated the cells with IFNɣ prior to drug exposure to increase the synthesis of Tpn. Following a nine hour incubation, MG132 treatment partially but significantly rescued the expression of Tpn in the presence of MC80 in murine cells (Fig 7A, S5A Fig). The proteasome-dependence of MC80 was also demonstrated in hIFNɣ-stimulated Hela (human) cells using the more specific inhibitor, Epoxomicin (Fig 7D, S5C Fig). In contrast, neither the Atg5-/- cell line nor chloroquine treatment had a discernable effect on MC80 function (S5A and S5B Fig). Furthermore, upon treatment with IFNɣ and MG132, anti-Tpn antibody co-immunoprecipitated ubiquitinated bands corresponding to the size of multi/poly-ubiqutinated Tpn, specifically in the presence of MC80 (Fig 7B). This suggests that the expression of MC80 leads to the ubiquitination of Tpn. Aside from ubiquitination, ER-associated degradation (ERAD) requires retrotranslocation of targeted proteins to the cytoplasm for proteasomal degradation to occur. While multiple retrotranslocation complexes exist in the ER, we have observed that MC80 associates with Derlin-1 in the presence of MG132 (Fig 7C). Given that MC80 is retained in the ER, these data suggest that MC80 selectively destabilizes Tpn by recruiting ER-associated degradation (ERAD) components for the ubiquitination and retrotranslocation of Tpn (Fig 7E).
Through this study, we demonstrate that (1) the MHC-I-like MCV protein, MC80, associates with the PLC via its luminal domain; (2) MC80 remains localized to the ER, where it induces the degradation of Tpn and TAP to impede peptide loading and consequent MHC-I surface expression; (3) MC80 requires β2m to degrade Tpn; (4) MC80 primarily targets Tpn, with our data suggesting that MC80 induces ERAD through a mechanism that requires its transmembrane and cytoplasmic tail. Taken together, these findings support a model wherein MC80 directly interacts with Tpn via its luminal domain and presumably recruits cellular ERAD machinery via its transmembrane and/or tail to facilitate the degradation of Tpn; which secondarily destabilizes TAP. The loss of Tpn/TAP in turn dramatically affects the ability of nascent MHC-I to load high affinity peptides and subsequently traffic to the cell surface. Our experiments indicate that MC80L and MC80S both downregulate classical MHC-I, associate with Tpn, and degrade Tpn/TAP in human and mouse cells. Thus, the extended MC80 signal peptide does not appear necessary for MHC-I sabotage but may have an as-yet-unknown independent function. To our knowledge, this is the first example of a viral protein that primarily targets Tpn for degradation.
Given the central role of Tpn in PLC organization and function, it is not surprising that multiple virally-encoded proteins have been found to undermine its function to evade CTL killing. HCMV US3 directly competes for Tpn binding to prevent peptide loading, while adenovirus E3-19K obstructs the TAP interface to prevent Tpn from bridging classical MHC-I to the PLC [54,55]. Unlike these mechanisms, our data indicates that soluble MC80 can associate with the PLC, but does not prevent MHC-I presentation. Therefore, at the expression levels tested in our retroviral system, MC80 does not appear to be functioning as a competitive inhibitor of PLC-mediated peptide-loading. Instead, we find that Tpn is degraded in the presence of MC80; but Tpn levels can be partially rescued in MC80-expressing cells by inhibiting proteasomal degradation. We also found that Tpn was multi/poly-ubiquitinated in the presence of MC80, suggesting that MC80 induces the ER-associated degradation of Tpn. The fact that we can detect the association between functional MC80 and Tpn in our wild-type MEFs and interferon-induced TAP-/- cells indicates that the downstream steps of ERAD (ubiquitination, retrotranslocation, and/or degradation) may be rate-limiting. However, this hindrance may be cell-line/species specific, as the association between MC80 and tapasin is only detectable for soluble (non-functional) MC80 in HEK 293T cells and untreated TAP-/- MEFs. In comparison to other viral MHC-I-evasion mechanisms which utilize ERAD, HCMV US2/US11 are only known to target MHC while MHV68 mK3 primarily targets MHC-I with only a slight effect on TAP levels [56,57]. Recently, a virally encoded ER-resident ubiquitin E3 ligase, RHVP pK3, was found to degrade MHC-I, Tpn, and TAP [58]. However, the degradation of Tpn and TAP were found to be secondary effects of the pK3-mediated degradation of MHC-I. Conversely, MC80 appears to degrade Tpn independent of TAP or MHC-I. Thus, the MC80-mediated destabilization of Tpn is distinct from other known viral MHC-I-evasion mechanisms.
While US2 and US11 specifically target MHC, their mechanisms are reminiscent of MC80. Indeed, while none of these viral proteins appear to have a ubiquitin E3 ligase domain, all three appear to induce ubiquitination-mediated ERAD. The luminal domains of US2 and US11 are able to associate with MHC-I, but without their transmembrane or tail domains these viral proteins cannot induce MHC degradation [59,60]. The transmembrane and tail sequences that distinguish US2 and US11 are thought to be responsible for recruiting distinct ERAD pathways. Of note, Derlin-1 associates with US11 and is essential for US11-mediated, but not US2-mediated, ER-associated degradation of MHC-I [61,62,63]. Intriguingly, MC80 contains two glutamic acid residues in its predicted TM domain [64], which are exceedingly rare in human type I TM proteins [65]. While MC80 is also found to associate with Derlin-1, the observation that neither US2 nor US11 TMs contain a negatively-charged residue raises the question of whether MC80 co-opts a distinct ERAD pathway. The current data cannot rule out the possibility that MC80 is able to actively degrade TAP through the proximal interaction with Tpn or direct interaction with TAP. However, considering that the loss of Tpn has been previously shown to destabilize TAP in human and murine cells, it is attractive to speculate that the loss of TAP is a secondary effect of specific ubiquitination and degradation of Tpn [50,51].
Interestingly, MC80 is retained in the ER despite lacking a putative ER-retention motif [66]; even when expressed as a truncated protein without a TM or tail. MC80's ability to associate with β2m and members of the PLC suggest that it maintains an MHC-I-like fold, and thus may exploit host machinery which canonically retains unloaded MHC-I in the ER. The association of soluble MC80 with CNX and CRT supports this hypothesis, as both chaperones have been previously shown to bind and retain immature MHC-I in the ER [29,30,31]. Furthermore, four of the eight residues critical to sequence-independent association of MHC-I with peptides are not conserved in MC80, suggesting that MC80 may not bind peptides in an analogous manner. Potentially, the divergence of MC80 from MHC-I functions to mimic the peptide-receptive conformation of MHC-I to continually associate with CNX/PLC. As such, a structural analysis of MC80 may provide insight into the mechanism underpinning PLC-assisted peptide loading of classical MHC-I; particularly regarding the poorly understood interaction between MHC-I and Tpn.
Recent studies have made significant progress toward a structural understanding of MHC-I peptide loading by employing a protein with sequence similarity to Tpn, TAPBPR [40,67]. Like Tpn, TAPBPR is capable of peptide editing through association with MHC-I; however, its functional role in antigen presentation has not yet been fully resolved. It is interesting to note that, while expression of MC80 in Hela cells treated with IFNɣ dramatically decreased Tpn levels, TAPBPR levels remained unchanged compared to vector control. We hypothesize that this specificity is a result of the fact that E3 ligases usually conjugate ubiquitin with lysine residues [68]; whereas the cytoplasmic tail of Tpn has four lysines, the cytoplasmic tail of TAPBPR does not have any. However, it is possible that MC80 preferentially interacts with Tpn over TAPBPR; or that TAPBPR may be present in other cellular compartments where MC80 is absent. Regardless, MC80 expression appears to cause the specific degradation of Tpn and not TAPBPR.
While MC80 was originally predicted to be involved in NK-subversion, the mechanism described herein suggests that MC80 is involved in subverting CTL responses by downregulating MHC-I, which may in turn increase NK killing [69]. However, our data does not preclude MCV from encoding additional ORFs which subvert NK and CTL responses through distinct mechanisms. One such MCV protein, MC148, is known to function as an inhibitor of CCR8-mediated chemotaxis, limiting T cell migration into sites of infection [70]. MC80 likely works in concert with MC148 to prevent the activation of surveilling T cells, specifically those which have been able to localize to the MCV lesion. Comparatively, cowpox virus encodes at least seven distinct proteins suspected of antagonizing host chemokines [71], while also downregulating MHC-I expression by two independent mechanisms [72], and encoding at least one separate protein to prevent NK activation [24]. We therefore believe it unlikely that MC80 and MC148 make up the complete repertoire of immune evasion proteins that allow for apparent MCV subversion of both T and NK cell surveillance.
Murine embryonic fibroblasts (MEF) cell lines including B6/WT3, TAP1-deficient (TAP1-/-; also referred to as FT1-), tapasin-deficient (Tpn-/-), triple knock-out (Kb-/-, Db-/-, β2m-/-; 3KO), and L cells were gifts from Dr. Ted Hansen, and have been described previously [73]. Baf3 cells [74], a murine proB lymphocyte was obtained from Dr. Anthony French. The BV2 microglial cell line was a gift from Dr. Anthony Orvedahl, and have been described previously [75]. The murine T lymphocyte cell line RMA (ATCC: TIB-39), human embryonic kidney 293T cell line (HEK 293T; ATCC: CRL-3216), human cervical cancer cell line (Hela; ATCC: CCL-2), and human foreskin cell line (HFF-1; ATCC: SCRC-1041), were obtained from the American Type Culture Collection (ATCC, Manassas, VA). Hela cells used in this work were stably transfected with HLA-A2, indicated as Hela-A2. Cyt-SIINFEKL and ER-SIINFEKL MEFs were produced by stably transfecting a construct encoding SIINFEKL peptide conjugated to ubiquitin or a signal peptide, respectively. Atg5-KO BV2 cell line isolation was performed as described [76]. All cell lines were cultured in 5% CO2 at 37°C with RPMI-1640 (Gibco) supplemented with 10% fetal bovine serum (Gibco), 2mM L-glutamine, 10mM HEPES pH 7.2, 1mM sodium pyruvate, and 100U/mL penicillin/streptomycin. Where indicated, prior to harvesting for immunoprecipitation and immunoblotting, cells were cultured for 24–48 h with 100–150 units/mL of mouse or human interferon gamma (mIFNɣ, Invitrogen; hIFNɣ, R&D Systems), followed by an 8-9h incubation with 100nM Epoxomicin, 30μM MG132 (Calbiochem, MA) or 100μM chloroquine (Sigma). Cells were harvested and washed in PBS containing 20mM iodoacetamide twice before freezing cell pellets at -80°C for storage prior to processing. Where indicated dithiobis-succinimidyl propionate (DSP; Pierce) was added to wash buffer at a concentration of 2mM.
The MC80 (MCV genotype 1) sequence was PCR amplified starting at M1 and M33 for constructs without the N-terminal Flag-tag or starting at Q18 and H72 for constructs with the N-terminal Flag-tag. Constructs with an N-terminal Flag-tag were inserted in frame with the canonical mouse β2m signal peptide and a Flag-tag into the pMXsIG vector (CellBioLabs) by overlap PCR and Gibson Assembly (NEB). In constructs that lacked an N-terminal Flag-tag, including the untagged construct and C-terminally Flag-tagged constructs, the native signal peptide was used for both the long and short forms of MC80. The soluble construct only included up to residue A342 of MC80, to truncate the predicted transmembrane and cytoplasmic tail. All constructs were confirmed by Sanger sequencing (GeneWiz). Retrovirus-containing supernatants were produced as per manufacturer instructions using either (i) the pVPack-GP and pVSVG vector (Agilent) in 293T cells to generate virus which infects human cell lines or (ii) the retroviral-packaging plat-E cells [77] to generate virus which infects murine cell lines. When necessary, retrovirally-transduced cells were enriched by cell sorting for GFP-positive cells (MoFlo).
Rabbit anti-mouse TAP1 and ERp57, rabbit anti-human TAP1 and tapasin, and hamster anti-mouse tapasin (5D3) were gifts from Dr. Ted Hansen and have been described previously [33,35,78]. The rabbit anti-Derlin-1 antibody was a gift from Dr. Yihong Ye and has been described previously [62]. Anti-human tapasin (TO-3), anti-human TAPBPR (42-L), and anti-ubiquitin (P4D1) antibodies were purchase from Santa Cruz Biotechnology. Anti-β-actin (AC-74) and anti-Flag (M2); anti-GFP; rabbit anti-CRT and rabbit anti-calnexin; anti-CD1d and anti-Qa-1 were purchased from Sigma, Covance, Stressgen, and BD Pharmingen, respectively. Rae1a and MICA were detected using an NKG2D-tetramer (a gift from Dr. Sytse Piersma). All MHC-I mAbs including 11-4-1 (α-H-2Kk), B8-24-3 (α-H-2Kb), 30-5-7 (α-H-2Ld), B22/249 (α-H-2Db), 25-D1-16 (α-H-2Kb-SIINFEKL), BBM.1 (α-β2m), and W6/32 (HLA-ABC) were previously described and available from the ATCC collection.
Staining was performed as described previously [79]. Phycoerythrin-conjugated goat anti-mouse IgG (BD Pharmingen) was used to visualize primary antibody staining. Intracellular staining was conducted using the BD cytofix/cytoperm kit (BD Pharmingen) following the manufacturer’s instructions. All flow cytometric analyses were performed using a FACSCalibur (Becton Dickinson). Data was analyzed using FlowJo 10 (Tree Star) and Prism 7 (GraphPad). Relative surface MHC-I expression % was calculated using the equation: [Mean fluorescence intensity (MFI) of MC80 positive population (GFP+)/MFI of MC80 negative population (GFP-)]*100. Error bars represent the standard deviation of 2 to 3 independent replicates.
Cells were lysed in PBS buffer containing 20mM iodoacetamide, 1% IGEPAL CA-630 (Sigma), and cOmplete protease inhibitors (Roche). For coimmunoprecipitations, IGEPAL CA-630 was replaced with digitonin (Wako). After lysis for at least 30min on ice, homogenized lysates were incubated for 1hr with a saturating concentration of antibody that was either directly conjugated to resin or associated via resin-conjugated protein A. Beads were then washed four times with 0.1% IGEPAL-CA-630 or digitonin, and eluted with Flag peptide or LDS sample buffer (Invitrogen). If endoglycosidase H treatment followed the immunoprecipitation, bound proteins were instead eluted by boiling in 10mM TrisCl, pH 6.8, 0.5% SDS. Supernatants were then incubated with an equal volume of 100mM sodium acetate, pH 5.4, and 1–5 μU endoglycosidase H (NEB) for 1hr at 37°C. Immunoblotting was performed following SDS-PAGE separation of precipitated proteins or cell lysates as described previously [79]. Following primary blotting with mouse or hamster primary antibodies, membranes were blotted with biotin-conjugated goat anti-mouse IgG (Invitrogen) or goat anti-hamster IgG (Jackson ImmunoResearch), respectively, followed by blotting with streptavidin-horseradish peroxidase (Invitrogen). In cases where the biotin-conjugated anti-mouse system produced a high background, m-IgGk BP-HRP was substituted (Santa Cruz Biotechnology). For rabbit primary antibodies, HRP-conjugated mouse anti-rabbit IgG light chain (Jackson ImmunoResearch) was used instead. Specific proteins were visualized by chemiluminescence using ECL (Thermo).
Statistical significance compared with the control group was calculated using ANOVA with Dunnett’s multiple comparisons test or unpaired t test and annotated as * = P<0.05; ** = P<0.01; *** = P<0.001; **** = P<0.0001. Protein alignments were conducted using the ESPRESSO webserver [80]. Signal peptide cleavage sites were predicted using the SignalP 4.1 webserver [81].
|
10.1371/journal.ppat.1007924 | Schistosoma mansoni soluble egg antigen (SEA) and recombinant Omega-1 modulate induced CD4+ T-lymphocyte responses and HIV-1 infection in vitro | Parasitic helminths evade, skew and dampen human immune responses through numerous mechanisms. Such effects will likely have consequences for HIV-1 transmission and disease progression. Here we analyzed the effects that soluble egg antigen (SEA) from Schistosoma mansoni had on modulating HIV-1 infection and cytokine/chemokine production in vitro. We determined that SEA, specifically through kappa-5, can potently bind to DC-SIGN and thereby blocks DC-SIGN mediated HIV-1 trans-infection (p<0.05) whilst not interfering with cis-infection. DCs exposed to SEA whilst maturing under Th2 promoting conditions, will upon co-culture with naïve T-cells induce a T-cell population that was less susceptible to HIV-1 R5 infection (p<0.05) compared to DCs unexposed to SEA, whereas HIV-1 X4 virus infection was unaffected. This was not observed for DCs exposed to SEA while maturing under Th1 or Th1/Th2 (Tmix) promoting conditions. All T-cell populations induced by SEA exposed DCs demonstrate a reduced capacity to produce IFN-γ and MIP-1β. The infection profile of T-cells infected with HIV-1 R5 was not associated with down-modulation of CCR5 cell surface expression. We further show that DCs maturing under Tmix conditions exposed to plant recombinant omega-1 protein (rω-1), which demonstrates similar functions to natural ω-1, induced T-cell populations that were less sensitive for HIV-1 R5 infection (p<0.05), but not for X4 virus infection. This inhibition associated again with a reduction in IFN-γ and MIP-1β expression, but additionally correlated with reduced CCR5 expression. We have shown that SEA parasite antigens and more specifically rω-1 can modulate HIV-1 infectivity with the potential to influence disease course in co-infected individuals.
| Parasitic helminths have developed a number of strategies to evade, skew and dampen human immune responses. Such effects will likely have consequences for HIV-1 transmission and disease progression. Here we analyzed the effect that soluble egg antigen (SEA) from Schistosoma mansoni had on HIV-1 infection in vitro. We determined that SEA, through kappa-5, can potently block DC-SIGN mediated HIV-1 trans-infection of CD4+ T-lymphocytes, but not block cis-infection. Dendritic cells (DC) exposed to SEA during maturation under Th2 skewing conditions, induce T-cell populations that are less susceptible to HIV-1 R5 infection compared to cells induced by unexposed DCs. HIV-1 X4 infection was unaffected. This restricted infection profile was not associated with down-modulation of CCR5 surface expression or observed differences in cytokine/chemokine production. Using recombinant omega-1, an abundant component of SEA, HIV-1 R5 infection was similarly inhibited with no effect on HIV-1 X4 infection levels. Hence SEA possesses antigens, namely omega-1, that can modulate HIV-1 infection and potentially influence disease course in co-infected individuals.
| Humans encounter numerous pathogens throughout their life-time, encompassing bacteria, fungi, parasites and viruses with many infections occurring concomitantly. Since CD4+ T-lymphocytes are the main cell-type infected with human immunodeficiency virus type 1 (HIV-1), the immune responses mounted against the array of co-infecting pathogens will likely influence HIV-1 transmission and disease progression. Helminthic parasites such as Schistosoma mansoni (S. mansoni) are pertinent in this context, due to their ability to evade, dampen and skew the human immune system including the modulation of CD4+ T-lymphocyte responses. Moreover, many areas endemic for S. mansoni infection have high HIV-1 prevalence rates indicating that co-infection is likely.
Cells are infected with HIV-1 through the initial binding of its trimeric gp120 envelope protein to CD4, after which it interacts with numerous chemokine receptors, typically CCR5 or CXCR4, and undergoes entry [1]. CCR5 using viruses (R5) are those predominantly transmitted and later in disease in approximately 50% of individuals the virus switches to utilizing CXCR4 (X4) as a co-receptor [2]. Following transmission the virus rapidly disseminates to lymph nodes and especially to the gut associated lymphoid tissue (GALT). The GALT is a major reservoir for CD4+CCR5+ memory T-cells and approximately 80% of these cells are lost in the first weeks following HIV-1 infection [3,4]. Direct infection of cells via the CD4 molecule and co-receptors is termed cis-infection. An array of C-type lectins (CLR) expressed on myeloid cell lineages have been shown to successfully capture HIV-1 and pass the virus to activated CD4+ T-cells, referred to as trans-infection [5]. One such lectin known to strongly support trans-infection is dendritic cell specific ICAM3-grabbing non-integrin (DC-SIGN) which is expressed to high levels on dendritic cells (DCs). This molecule has been implicated in supporting HIV-1 transmission as well as virus dissemination [5,6]. DC-SIGN is known to bind many glycosylated structures including a large array of pathogen antigens as well as numerous host proteins found in bodily secretions [7–10]. Indeed, these molecules have the capacity to interfere with HIV-1 trans-infection.
CD4+ effector memory Thelper (Th) cells consist of three major subsets; Th1 cells induced by viral infections, Th2 cells induced by parasitic infections and Th17 cells induced by bacterial and fungal infections [11]. Remarkably, expression of HIV-1 co-receptors is not directly linked to the HIV-1 susceptibility of these cells. For instance, Th1 cells express high levels of CCR5 but also produce MIP-1α (CCL3), MIP-1β (CCL4) and RANTES (regulated upon activation normal T-cell expressed and secreted) (CCL5), the natural ligands for CCR5, thereby limiting R5 infection in these cultures [12,13]. Contrary to this, Th2 cells express lower levels of CCR5 but due to the limited production of MIP-1α, MIP-1β and RANTES these cultures have been shown to be infected with HIV-1 more easily [12,13]. This generalized view that Th2 cells are more susceptible to HIV-1 infection than Th1 cells is no longer supported. In a review by Mariana et al. it was stated that HIV-1 susceptibility of CD4+ T-cells varied dependent on the in vitro stimulation of these cells [14]. More recent studies have correlated pathogen specific CD4+ T-cell phenotypes to HIV-1 susceptibility. Cytomegalovirus (CMV) and Mycobacterium tuberculosis (Mtb) infections both result in the induction of Th1 cells [15,16]. However, the Mtb specific T-cells are lost early during HIV-1 infection while the CMV specific T-cells are lost later in disease [17]. This discrepancy was explained by differences in cytokine expression profiles, where Mtb specific cells possess a high IL-2 and low MIP-1β expression pattern, the reverse phenotype was observed in CMV specific CD4+ T-cells [17]. Human papilloma virus specific CD4+ T-lymphocytes have also been shown to be lost early after HIV-1 infection [18,19].
Helminths, including S. mansoni, are known to skew immune responses towards a Th2 phenotype, which according to the above hypothesis would be detrimental to those individuals co-infected with HIV-1 [20,21]. This has led to the assumption that treating S. mansoni in co-infected individuals would be beneficial for their HIV-1 disease. Clear epidemiological evidence to-date is lacking, as treatment studies have been reporting contradictory findings [22]. A treatment program in Ethiopia found that deworming S. mansoni infected HIV-1 patients led to a decrease in viral loads [23], whilst another study in Uganda reported the opposite [24]. Similar inconsistencies have been found for other markers associated with HIV-1 disease progression as reviewed in [21], with only one exception. Women infected with S. haematobium and who have egg induced lesions in their genital tract were found to be at higher risk of HIV-1 infection [25,26].
In S. mansoni infections the eggs play a crucial role in disease as they induce lesions and skew CD4+ T-lymphocyte responses. An adult S. mansoni pair typically lay up to 300 eggs a day which migrate to the gut lumen in order to be expelled [27]. One of the best studied antigen mixtures of S. manoni is soluble egg antigen (SEA) which is an extract derived from homogenized eggs and consists of hundreds of proteins of which many are glycosylated [28]. SEA has accordingly been shown to bind many glycan receptors including DC-SIGN, mannose receptor (MR) and macrophage galactose type-lectin (MGL) [28,29]. Through binding to these receptors SEA alters the DCs response to TLR4 ligand, LPS (lipopolysaccharide) and TLR3 ligand PolyI:C [30]. Although, SEA itself cannot fully mature immature DC (iDCs), while antigen processing is similar to LPS matured DCs [31]. Furthermore, SEA exposed DCs are known to induce Th-cell responses that are skewed towards a Th2 phenotype, even when a Th1 cell response is required [32]. Recently it has been demonstrated that omega-1 (ω-1), one of SEA’s main components, is able to drive Th2 cell responses [33–35]. Omega-1 is a member of the T2 RNase family which enters the cell through binding MR and subsequently degrades cellular mRNA and rRNA products. Both the RNase activity and the glycan group are essential for Th2 skewing [33].
Infection with either S. mansoni or HIV-1 has major implications for the host due to the longevity of infection and extent of damage these pathogens cause to the immune system. The complex pathogen interactions encountered in co-infected individuals makes it difficult to determine the effect of S. mansoni on HIV-1 infection. Consequently, our study focused on the effects of SEA on HIV-1 infection, where we specifically address whether SEA can interfere with cis- or trans-infection of CD4+ T-lymphocytes as well as whether the effects exerted by SEA on DC maturation can modulate the T-cell population’s susceptibility to HIV-1 infection.
It has been reported that SEA binds several C-type lectin receptors and competes with monomeric HIV-1 gp120 for binding [28,36], but has not been tested for inhibiting HIV-1 capture and transfer. We confirmed that our derived SEA binds DC-SIGN using a DC-SIGN binding ELISA, where increasing concentrations of SEA results in a dose-dependent increase of DC-SIGN-Fc binding (Fig 1A). To determine whether this interaction interferes with HIV-1 binding to DC-SIGN we performed a gp140 competition ELISA. Here DC-SIGN-Fc is incubated with SEA before addition to an ELISA plate coated with trimeric gp140. We found that concentrations as low as 0.2μg/ml SEA resulted in a 50% reduction of DC-SIGN-Fc binding to gp140 (Fig 1B). Since the trimeric gp140 protein closely resembles the functional HIV-1 envelope protein, this data suggests that SEA can prevent HIV-1 from interacting with DC-SIGN.
We next tested whether SEA could interfere with HIV-1 cis-infection of CD4+ lymphocytes (direct infection). Therefore, CD4+ enriched (CD8+ depleted) T-cell blasts were incubated with SEA (1, 5 or 25μg/ml) before HIV-1 SF162 (CCR5 using virus) or LAI (CXCR4 using virus) was added. Neither SF162 (R5) nor LAI (X4) viral outgrowth, measured as the concentration of HIV-1 capsid protein (CA-p24) in the culture supernatant, was affected by the presence of SEA (Fig 1C). This indicates that SEA does not interfere with HIV-1 binding to CD4, CCR5 or CXCR4. Additionally, the presence of SEA had no visible effect on cell counts and/or cell viability of the CD4+ enriched T-cells.
Since incubating DC-SIGN-Fc with 1μg/ml SEA provided a 70% reduction in its capacity to bind gp140 (Fig 1B), we pre-incubated Raji DC-SIGN cells with the same concentration to identify whether SEA can block HIV-1 trans-infection of CD4+ T-cells. Viral outgrowth of both SF162 (R5) and LAI (X4) was approximately 80% reduced in co-cultures where the Raji DC-SIGN cells were pre-incubated with SEA (p<0.01 and p<0.05, respectively) (Fig 1D). Similarly, we tested the effect of SEA on immature monocyte derived DCs (iDCs) and observed a significant reduction (p<0.01) in SF162 (R5) outgrowth but only when iDCs were pre-incubated with 100μg/ml SEA (Fig 1E). This higher SEA concentration was likely required because iDCs express higher levels of DC-SIGN than Raji DC-SIGN cells, have multiple other CLRs capable of binding HIV-1 and have a rapid receptor turnover. Mannan, known to bind DC-SIGN and block HIV-1 trans-infection, was tested as a positive control and provided similar inhibitions (p<0.001).
SEA consists of many molecules of which IPSE/α-1, kappa-5 (κ5) and omega-1 (ω-1) are the major components. To determine whether any of these could be responsible for blocking HIV-1 trans-infection, SEA depleted of each product was evaluated using the gp140 competition ELISA. Depletion of κ5 from SEA partially restored the binding capacity of DC-SIGN to gp140 (p<0.001) whereas depletion of IPSE/α-1 or ω-1 did not abrogate the effect (Fig 1F). We next tested whether plant derived recombinant omega-1 (rω-1) or recombinant κ5 (rκ5), which carry similar glycosylation profiling as natural derived products [37], could bind to DC-SIGN in the standard binding ELISA assay. SEA showed a dose dependent binding to coated DC-SIGN as did rκ5, whilst rω-1 demonstrated no binding, supporting the previous finding where depletion of natural derived κ5 from SEA removed DC-SIGN binding activity (Fig 1G). Furthermore, when testing rω-1 and rκ5 in a Raji-DC-SIGN mediated capture transfer experiment, only rκ5 inhibited HIV-1 infection of CD4+ T-lymphocytes (p<0.001) (Fig 1H).
In conclusion, SEA potently binds DC-SIGN, via κ5, and prevents DC-SIGN mediated capture and transfer of HIV-1 to CD4+ T-lymphocytes, whilst having no effect on direct infection of CD4+ lymphocytes.
Besides the direct effect of blocking HIV-1 trans-infection, SEA may also affect HIV-1 infection indirectly. SEA exposed DCs have been shown to promote the development of T-cells that are skewed towards a Th2 phenotype [32]. Since it has been shown that differentially skewed Th cell populations have variant HIV-1 infection profiles [12,13], it is likely that SEA can modulate HIV-1 infection and/or replication. Here we aimed to i) determine whether the presence of SEA during DC maturation alters the polarizing capacity resulting in a modified Th cell profile and ii) identify whether such a modulated Th cell phenotype has an altered susceptibility for HIV-1. We therefore established an in vitro model system where we could assess the Th cell populations, induced by DCs matured in the absence or presence of parasite products, for cytokine/chemokine production and HIV-1 infection (Schematic shown Fig 2A). Here iDCs (donor A) were matured in Th1/Th2 mixed (Tmix)-, Th1- or Th2-promoting conditions (LPS, LPS+IFN-γ or LPS+PgE2, respectively) either in the absence or presence of SEA. Subsequently, the matured DCs were washed, to remove the SEA, and co-cultured with naïve CD4+CD45RA+ T-cells (donor B) and Staphylococcus Enterotoxin B (SEB) for 8 days resulting in memory Th cell populations with specific phenotypes. These cells were then infected with HIV-1 SF162 (R5) or LAI (X4) and monitored for infection over time (Fig 2A and S1 Fig). In our model, Th cell populations induced by DC matured under Tmix and Th2 promoting conditions harbour on average 4.4% ±0.4 and 4.6% ±0.6 SF162+ cells, respectively, whilst Th cells induced by DCs matured under Th1 promoting conditions have a lower level of infection, 3.1% ±0.7 (p = 0.09, compared to Tmix cell culture) (Fig 2B, black symbols). When SEA was present during DC maturation the percentage of SF162+ T-cells was found to be similar in Tmix, Th1 and Th2 cell cultures, 3.5% ±0.6, 4.6% ±0.9 and 2.8% ±0.5, respectively (Fig 2B, blue symbols). A 2-way ANOVA revealed that there is a significant interaction (p = 0.041) between SEA and the Th cell subset infected with SF162, meaning that SEA affects SF162 infection depending on the Th cell subset. More in depth analysis of the different groups demonstrated that a Th2 cell population induced by SEA exposed DCs have a significantly lower percentage of SF162+ T-cells than a Th2 cell population induced by SEA unexposed DCs (p = 0.044). In contrast, SEA exposure of DCs inducing Tmix or Th1 populations did not affect the percentage of SF162+ T-cells.
No such effect was observed in LAI (X4) infected Th cell populations. The percentage of LAI+ cells in Th cell cultures induced by DCs, matured under Tmix, Th1 and Th2 promoting conditions, were all found to be similar, 14.4% ±2.2, 16.7% ±4.1 and 12.8% ±1.4, respectively, in the absence of SEA or 13.5% ±2.1, 17.8% ±2.5 and 10.6% ±2.4, respectively, in the presence of SEA (Fig 2C). Analysing the effect of SEA on infection of different Th cell populations using a 2-way ANOVA demonstrated no significant differences. Our data demonstrate that the addition of SEA to DCs maturing under Th2 promoting conditions result in the induction of a Th cell population with a reduced susceptibility to HIV-1 (R5) infection.
In order to explain the reduced SF162 (R5) infection in Th2 cell cultures induced by SEA exposed DCs we characterized the cytokine/chemokine (IFN-γ, IL-4, IL-2, TNF-α, MIP-1β) expression profiles of the different Th cell populations. First we demonstrated that neither SF162 (R5) nor LAI (X4) infection alters the potential of the T-cells to produce cytokines in Tmix cells (Fig 3A). Next, we determined the effect of exposing DCs to SEA on Th cell responses and found a significant reduction of IFN-γ producing T-cells in all three conditions irrespective of HIV-1 infection (Fig 3B). Although IL-4 production of the T-cells remained similar, the ratio, IL-4/ IFN-γ did increase when T-cells were induced by DCs matured in the presence of SEA (Fig 3B and S1 Table). Besides IFN-γ, the percentage of MIP-1β+ cells was also significantly reduced in all three T-cell populations where SEA had been present during DC maturation (Fig 3C).
Interestingly, when comparing the cytokine profile of the SF162 infected (p24+) T-cells to the profile of the total population we found a higher percentage of IFN-γ and MIP-1β producing T-cells in the infected fraction (for all three T-cell conditions) (Fig 3D and 3E). Analysing the FACS plots revealed it is mainly the percentage of MIP-1β+ and IFN-γ+MIP-1β+ cells that is enhanced (Fig 3D). Combining data from several donors shows a significant increase in MIP-1β+ and IFN-γ+ cells in the p24+ fraction (p<0.05) (Fig 3F). This increase was not observed for IL-4 nor for any cytokine monitored in the LAI (X4) infected fractions (Fig 3E). However, since this increase in MIP-1β+ and IFN-γ+ cells in the SF162 (R5) infected fraction is found in all three T-cell populations induced by SEA exposed DCs, it is unlikely that this phenomena contributes to the altered SF162 (R5) susceptibility of Th2 cell cultures induced by SEA exposed DCs. Hence, the cytokine/chemokine production of the T-cells does not explain the significant reduction in SF162 (R5) infection found in Th2 cell cultures induced by DCs exposed to SEA. Nevertheless, the high percentage of MIP-1β producing T-cells in Th1 cell cultures induced by unexposed DCs may explain the trend of reduced SF162 infection levels in this population (p = 0.09) (Fig 2B).
The reduced HIV-1 infection was only observed for SF162 (R5) and not for LAI (X4), suggesting that CCR5 expression levels may be altered in the Th cell populations induced by SEA exposed DCs. We determined the percentage of cells expressing CCR5 on the surface as well as the quantity of CCR5 per cell. When comparing the Th2 cell cultures induced by DCs, either exposed or unexposed to SEA, we found no difference in the percentage of cells with a high expression of CCR5, 72% versus 76.6% respectively, as determined by the gate based on the marker or the amount of CCR5 expressed on each cell, geometric mean of 1365 and 1705, respectively (Fig 4A). The interesting finding that the SF162- but not LAI-infected fraction had a significantly higher percentage of IFN-γ+ and MIP-1β+ T-cells than the total population (Fig 3E) led us to speculate that SF162 (R5) infection resulted in upregulation of these cytokines through CCR5 engagement rather than the specific cell phenotype being targeted by SF162 (R5) virus. To examine this possibility we added either monomeric SF162 gp120 or RANTES (a CCR5 ligand) to LAI (X4) infected cultures and determined the percentage of IFN-γ and MIP-1β producing T-cells. Addition of monomeric SF162 gp120 (1 and 3μg/ml) or RANTES (0.7 and 5μg/ml) did not increase the percentage of IFN-γ+ or MIP-1β+ T-cells compared to 9.3% and 12.8% found in the LAI infected fraction in the absence of these compounds (Fig 4B and 4C). Our data indicates that CCR5 signalling is not the cause of heightened IFN-γ+ and MIP-1β+ T-cells in the SF162+ fraction. To conclude, we found no difference in CCR5 expression between Th2 cell cultures induced by DCs, either exposed or unexposed to SEA. Furthermore, it is unlikely that signalling through CCR5 causes the increased percentage of IFN-γ and MIP-1β producing T-cells in the SF162 infected fraction compared to the total population.
We have demonstrated that Th2 cell cultures induced with SEA exposed DCs harbour a significantly lower percentage of SF162 infected cells compared to Th2 cell cultures induced by unexposed DCs. Since the Tmix cell population is more relevant to the in vivo scenario and where Th1 and Th2 cross-talk likely resides to influence responses we were interested to follow this further. Indeed, Tmix cell culture SF162 infection levels were lower (although not statistically significant) when DCs were exposed to SEA (Fig 2B). Since omega-1 has been described as the main Th2 skewing component of SEA we aimed to study the effect of its recombinant form (rω-1) on the different cell subsets [33,34,37]. In an initial experiment we identified that the addition of rω-1 (3µg/ml) in the Tmix and Th2, but not the Th1, setting reduced infection with SF162 virus (S2A Fig). When focussing further on the Tmix population we identified that the presence of 2µg/ml or 4µg/ml, but not 1µg/ml), rω-1 during DC maturation resulted in Tmix cell cultures with a reduced susceptibility for SF162 (R5) infection, 40% and 60% less infected, respectively (Fig 5A). Moreover, the presence of rω-1 during DC maturation did not affect the induced T-cell population’s susceptibility to LAI (X4) virus. Plotting data from independent experiments (n = 5) confirms that there is significantly less infection with SF162 (R5) virus in Tmix cell cultures induced by DCs exposed to rω-1 during maturation, whilst not being observed with LAI (X4) infections (Fig 5B).
Analysing the cytokine/chemokine profile of Tmix cell populations induced by rω-1 exposed DCs resulted in a similar pattern as observed for T-cell populations induced by SEA exposed DCs. There was a significant reduction in the percentage of IFN-γ and MIP-1β producing T-cells whilst the percentage of IL-4 producing cells remained unaltered (Fig 5C). Based on the cytokine/chemokine profile of these cells it does not appear that rω-1 leads to enhanced Th2 skewing. What we did find was a modest, 10% reduction in the percentage T-cells with a high expression of CCR5 (indicated by the gate based on the marker) in Tmix cell cultures induced by rω-1 exposed DCs. Similarly, we observed a reduced level of CCR5 surface expression per cell, geometric mean of 1,209 versus 2,063, respectively which potentially plays a role in the reduced infection with R5 virus (Fig 5D). Collectively, these results indicate that ω-1 is the component responsible for the reduced SF162 (R5) infection observed in Th2 cell populations induced by SEA exposed DCs and that the molecule can similarly reduce infection of Tmix cells at higher concentrations.
Helminthic parasites are known to possess immunomodulatory properties and specifically skew immune responses towards a Th2 phenotype [32]. Here we studied the effect’s SEA has on modulating HIV-1 infection in vitro. Although SEA did not affect cis-infection of CD4+ enriched T-cell blasts, it efficiently blocked DC-SIGN mediated HIV-1 trans-infection through the binding of Kappa-5 to DC-SIGN. We demonstrated that exposing DCs maturing under Th2 promoting conditions to SEA induce Th cells that were less susceptible to R5 but not X4 infection and that rω-1 possessed the same activity with DCs maturing under Tmix promoting conditions. Collectively, our data demonstrates that SEA has the capacity to influence numerous mechanisms associated with HIV-1 transmission and pathogenesis, suggesting that the S. mansoni infection has the potential to modulate HIV-1 infection as well as disease course.
Receptive anal intercourse carries the highest high risk for HIV-1 transmission, likely due to the nature of the mucosal barrier and the immune cells residing at this site [2,38]. DCs will be amongst the first cells encountering HIV-1 with the potential of transferring virus to CD4+ T-lymphocytes either locally or in adjacent lymph nodes [5]. The presence of SEA in the gut wall may therefore influence trans-infection and interfere with the transmission process. Similarly, Th cell responses induced in localised lymph nodes by DCs exposed to SEA, more so ω-1, could have a phenotype that is less susceptible to infection with R5 HIV-1, again limiting the likelihood of transmission [39]. Moreover, regardless of the route of HIV-1 infection, a massive loss of CD4+ memory T-lymphocytes in the GALT is observed within the first few weeks of infection [4]. This reservoir is not replenished after peak viremia and although the initial cell loss is not correlated with HIV-1 disease progression, microbial translocation caused by loss of the GALTs function is [4,40]. The potential reduction in HIV-1 infection of GALT tissue exposed to SEA/ω-1 may therefore reduce not just cellular infection profiles but also the effects of bacterial translocation.
Several cytokines and CC-chemokines have been associated with alterations in HIV-1 infection and disease progression. MIP-1α, MIP-1β and RANTES are the natural ligands for CCR5, hence in vitro T-cell cultures containing high levels of these chemokines have lower levels of HIV-1 R5 infection [41,42] as has been observed ex vivo [17]. In contrast, we observed a significantly lower percentage of Th cells capable of producing IFN-γ and MIP-1β when they were induced by DCs matured under Tmix, Th1 or Th2 promoting conditions in the presence of SEA or rω-1 (Figs 3C and 5C). The HIV-1 R5 infected fraction however, had a higher percentage of Th cells producing IFN-γ and MIP-1β then the total population. This suggests either that these cells were targeted or that IFN-γ and MIP-1β were induced upon HIV-1 R5 infection. Addition of monomeric HIV-1 R5 gp120 or recombinant RANTES to HIV-1 X4 infected cell cultures did not induce a similar effect, suggesting the latter is not the case. Targeting of these cells seems equally unlikely since MIP-1β competes with HIV-1 R5 for CCR5 binding [12,42–44]. HIV-1 R5 may induce IFN-γ and MIP-1β via another mechanism, for example some forms of the viral protein Nef induce MIP-1β production in macrophages [45,46] while the viral protein Tat can induce MIP-1β secretion by neural progenitor cells [47]. Although it may seem counter intuitive that HIV-1 infection stimulates expression of MIP-1β, the function of this chemokine is to recruit CD4+ T-lymphocytes to the site of infection, thereby enhancing the number of target cells for HIV-1 replication [48]. It has also been shown that IL-2+ expressing cells are rapidly lost as a consequence of HIV-1 infection due to high levels of infection [17,49]. Since IL-2 is added to our cell culture system to ensure T-cell survival, we are unable to analyze the effects of IL-2 on HIV-1 infection with our system.
Besides cytokines and chemokines, we explored whether differences in surface expression levels of CCR5 could explain reduced infection with HIV-1 R5 but not X4. In Th2 cell cultures induced by SEA exposed DCs we found similar percentages of cells expressing CCR5 and similar levels of CCR5 on the surface of each cell (geometric mean) compared to Th2 cell cultures induced by unexposed DCs. As observed with SEA, rω-1 also reduced the DCs capacity to produce cytokines and co-stimulatory molecules in response to stimulus [33–35]. The mechanism via which a Th2 response is induced is still unknown, with one hypothesis being that the lack of stimuli provided by the DCs to the naïve T cells pushes them into default mode, namely the Th2 phenotype [35,50,51]. Yet, rω-1 exposed DCs induced Tmix cell cultures with even lower levels of HIV-1 R5 infection showed a reduced percentage of cells expressing CCR5 as well as a lower level of CCR5 expression per cell. Consequently, although CCR5 is not the main mechanism responsible for reduction in HIV-1 R5 infection it will likely contribute to the overall reduction of infection observed in Tmix cell cultures induced by DCs exposed to rω-1.
There are several hypotheses to explain the reduced infection of CD4+ T-lymphocytes, one of which is the capacity of SEA and rω-1 exposed DCs to induce regulatory T (Treg) cells [52,53]. Treg cells are susceptible to HIV-1 although several studies demonstrate that infection is restricted. It has been demonstrated that Treg cells were less susceptible to HIV-1 (R5) infection than effector memory T-lymphocytes while X4 viruses gave higher or similar levels of infection [54]. Further comparisons revealed these cells expressed similar levels of CD4, CCR5 and CXCR4 on their surface and produced similar levels of CC-chemokine production. Although no explanation was provided, the cellular activation status as well as FoxP3 interference patterns were thought to play a role. Some studies support the notion that FoxP3 inhibits HIV-1 infection by interfering with HIV-1’s LTR transcription activation [55,56] while others demonstrate enhanced HIV-1 production in FoxP3 positive cells [57,58]. The role of FoxP3 seems to be dependent on the viral strain as well as the culture protocol used for the cells. Another hypothesis is that virus production may be limited but not susceptibility [59], or that viral restriction factors, such as APOBEC3G may limit infectivity [60]. Further phenotypic characterisation of T-cells induced by SEA exposed DCs and more specifically rω-1 will undoubtedly identify additional cellular differentiation markers that associate with HIV-1 infection and/or replication with potential implications for induction of viral latency. A more in-depth analysis of SEA and rω-1 alterations to expression patterns of specific transcription factors associated with Th cell differentiation will further highlight the mechanisms responsible and associate such differences with the altered cytokine/chemokine expression patterns.
As well as SEA and rω-1 driving the induction of Th2 phenotypes [32–35], SEA has been shown to down-modulate DC TLR4 as well as LPS induced signalling [30]. More in-depth analysis with human DCs has demonstrated that SEA exposed DCs have enhanced expression of the negative regulators SOCS1 and SHP1 which result in impaired DC maturation and induction of CD4+ lymphocyte proliferation [61]. Similarly, in a murine model system it has been shown that SEA treated DC possess impaired LPS mediated maturation as identified through reduced expression of co-stimulatory molecules [62]. Our results support these findings and indicate that SEA/rω-1, through interacting with DCs, can induce CD4+ lymphocytes responses with reduced proliferative/activation phenotypes with potential consequences for HIV-1 infection and/or replication.
Reduced infection of CD4+ T-lymphocytes with R5 but not X4 HIV-1 may restrict initial viral replication or slow the rate at which HIV-1 switches its co-receptor phenotype. Large scale monitoring of study cohorts would need to be performed to identify whether such effects indeed do exist and which would be limited by factors such as parasite load and corresponding egg count. This complexity is confirmed by the discrepancies found in epidemiological studies of co-infected individuals [21,22]. The findings that HIV-1 R5 viruses replicated better in Th2 than Th1 cells resulted in the interpretation that treating S. mansoni infection may benefit co-infected patients [20,21]. Whether this can correlate to the in vivo setting needs to be determined and as always with such findings, antigenic concentration and the localisation of the source will be instrumental in determining to what capacity co-infection is beneficial. The effects of helminthic infections have been shown to dampen immune inflammatory responses with considerable consequences for diseases associated with such effects [63]. This would suggest that SEA antigens, including κ5 and ω-1, possess the physiological capacity to influence HIV-1 infection and disease. These skewing effects must ultimately be considered in the context of parasite antigens having the capacity to activate immune responses which provides a complex balance where different antigens have variant effects within co-infected individuals. We provide evidence here that Th cells induced by SEA/rω-1 exposed DCs are less susceptible to R5 HIV-1 infection, suggesting that helminthic infections may be beneficial when considering HIV-1. Deciphering the mechanisms may provide a means towards modulating immune responses beneficial for limiting viral transmission or reducing viral loads. Furthermore, these results should be considered in the context of HIV-1 vaccine trials being conducted in regions of the world where S. mansoni infections are endemic.
SEA was prepared and isolated as described previously [32]. Kappa-5 (κ5) was isolated by soybean agglutinin (SBA; Sigma, Zwijndrecht, the Netherlands) affinity chromatography as described previously [64]. The remaining SEA fraction was labeled as κ5 depleted SEA. Omega-1 and IPSE/alpha-1 were purified from SEA via cation exchange chromatography as previously described [35,49,65]. Omega-1 was then separated from IPSE/alpha-1 by affinity chromatography using specific anti-IPSE/alpha-1 monoclonal antibodies coupled to an NHS-HiTrap Sepharose column according to the manufacturer's instructions (GE Healthcare). Purified components were concentrated and dialyzed. Omega-1–depleted SEA or IPSE-depleted SEA were prepared by adding back purified IPSE/alpha-1 or omega-1, respectively, to the remaining SEA fraction left from the cation exchange chromatography. The purity of the preparations was controlled by SDS-PAGE and silver staining. Protein concentrations were tested using the Bradford or BCA procedure [66]. Plant produced rω-1 and rκ5 were purified from apoplast fluid using HS POROS 50 strong cation exchange (CEX) resin (Applied Biosystems) as previously described [34]. Apoplast fluids were transferred over G25 Sephadex columns to exchange for CEX binding buffer (20 mM phosphate buffered saline (pH 6.0) containing 100 mM NaCl). The plant derived rω-1 and rκ-5 molecule is similar to helminthic derived proteins with similar post-translational modifications, specifically glycosylation profiling, as observed in parasites.
SF162 (R5), BAL (R5) and LAI (X4) are replication competent HIV-1 subtype B viruses. SF162and BAL are molecular cloned isolates obtained from HIV-1 infected patients and which utilise CCR5 as co-receptor. LAI represents a molecular cloned virus isolated from an HIV-1 infected patient and utilises the CXCR4 co-receptor for infection. Viruses were passaged on CD4+ enriched T lymphocytes and tissue culture infectious dose (TCID50) values were determined by limiting dilutions on CD4+ enriched T-lymphocytes according to the Reed and Muench method, as previously described [67].
The DC-SIGN binding ELISA was performed as described [10]. Briefly, the components of interest (SEA, rω-1 or rκ5) were coated at 5 fold dilutions (50μg/ml– 0.01μg/ml on an ELISA plate after which 333ng/ml DC-SIGN-Fc (R&D systems) was added. Subsequently, DC-SIGN-Fc was detected by a secondary goat-anti-human-Fc HRP labelled antibody (Jackson Immunology), diluted 1:1000 and using standard ELISA procedures. In the gp140 competition ELISA, 10μg/ml anti-HIV-1 gp120 antibody D7324 (Aalto BioReagents Ltd) was coated on an ELISA plate after which trimeric HIV-1 gp140 (JR-FL SOSIP.R6-IZ-D7324) was added to the plate as previously described [68,69]. Meanwhile, 333ng/ml DC-SIGN-Fc (R&D systems) was incubated with SEA at limiting dilutions. Next, the mixture was added to the gp140 coated plate and using a secondary HRP labelled goat-anti-human-Fc antibody (Jackson Immunology) diluted at 1:1000, DC-SIGN-Fc binding to the gp140 plate was determined. A more detailed description can be found [68]. The capsid p24 ELISA was performed as standard [8]. In short, an ELISA plate was coated O/N with 10#x00B5;g/ml sheep anti-p24-specific antibody (Aalto Bio Reagents Ltd.). Culture supernatant was added, followed by 4ng/ml mouse anti-HIV-1-p24 alkaline phosphatase conjugate antibody (Aalto Bio Reagents Ltd.). Development solution Lumi-phos plus (Lumigen Inc.) was used according to the manufacturer's protocol. The standard curve consists of a serial dilution of Escherichia coli-expressed recombinant HIV-1-p24 (Aalto Bio Reagents Ltd.).
Raji DC-SIGN cells (a kind gift from Prof TBH Geijtenbeek, University of Amsterdam), an immortalized B cell line stably transfected with DC-SIGN [5] was cultured in RPMI 1640 (Invitrogen) supplemented with 10% FCS, 100U/ml penicillin and 100U/ml streptomycin. Peripheral blood mononuclear cells (PBMCs) were isolated using ficoll-hypaque density centrifugation from buffy coats purchased from the Netherlands Blood bank (Sanquin). PBMCs of three CCR5 wild-type homozygous donors were pooled and cultured in RPMI 1640 containing 10% FCS, 100U/ml penicillin, 100U/ml streptomycin, 100U/ml recombinant IL-2 and cells were activated with 2µg/ml phytohemagglutinin (PHA) (Sigma). To enrich for CD4+ T-lymphocytes, the CD8+ T-lymphocyte population was depleted using CD8 dynabeads (Life Technologies) according to manufacturer’s protocol at day 5. These cells were used for experiments depict (Fig 1), as for the remaining experiments cells were isolated from fresh blood.
Monocytes were isolated from fresh blood using lymphoprep (Nycomed) and subsequent percoll (GE Healthcare) density gradient (34, 47.5 and 60% of standard isotone percoll) centrifugation. For 6 days monocytes were cultured in IMDM (Gibco) containing 5% FCS, 86μg/ml gentamycin (Duchefa), 500U/ml GM-CSF (Schering-Plough) and 10U/ml IL-4 (Miltenyi) differentiating them into immature DCs (iDCs). From the PBMC fraction the CD4+ T-cells were isolated using the CD4 T-cell isolation MACS kit (Miltenyi Biotec., 130-091-155) according to manufacturer’s protocol. Subsequently, the CD4+CD45RA+CD45RO- naïve T-cells were isolated from the CD4+ T-lymphocytes using anti-CD45RO-PE (DAKO, R084301) and anti-PE beads (Miltenyi-Biotec, 130-048-801), described in detail [70].
CD4+ enriched T-cells (2x105 cells/well) were incubated with 25, 5 or 1μg/ml SEA or medium (control) for 2h in a 96-well flat-bottomed culture plate (Greiner Bio-One), subsequently medium, SF-162 (TCID50/ml of 200) or LAI (TCID50/ml of 200) was added to the well. At day’s 4, 7 and 12 supernatants were harvested and cells were fed with fresh media. HIV-1 capsid p24 was determined in the supernatants using a standard ELISA protocol.
Raji DC-SIGN cells (5x104cells/well) were pre-incubated with 1μg/ml SEA for 2h after which SF162 (R5) or LAI (X4) viruses were added at an end concentration of 200 TCID50/ml. After a further 2h incubation the cells were washed 3 times with PBS and co-cultured with enriched CD4+ T-lymphocytes (2x105 cells/well). Viral outgrowth was measured by monitoring capsid p24 at day 4, 7 and 12 in harvested culture supernatant. HIV-1 capture/transfer by iDCs was conducted in a similar manner with minor modifications; 20 or 100μg/ml SEA was utilized and the incubation steps were shortened to 30min.
A co-culture system was developed where monocytes were pre-incubated with immune skewing reagents in the presence or absence of parasite product (SEA or rω-1), before washing and adding to isolated memory T-cells and subsequently monitoring HIV-1 infection profiles or markers of CD4 cell phenotype. Routinely, iDCs from donor A were matured for 2 days with 100ng/ml LPS (Sigma-Aldrich) (which typically generates DCs that induce a Thmix-cell phenotypes), 100ng/ml LPS and 100U/ml IFN-γ (UCytech) (which typically generates DCs that induce a Th1-cell phenotype) or 100ng/ml LPS and 10μM PgE2 (Sigma-Aldrich) (which typically generates DCs that induce a Th2-cell phenotype) either in the absence or presence of 25μg/ml SEA or 1–4μg/ml rω-1. After thoroughly washing the matured DCs 3 times, 5x103 of cells were co-cultured with 2x104 CD4+CD45RA+CD45RO- T-cells (naive) from donor B in 96-well flat bottom culture plates and 10pg/ml Staphylococcus enterotoxin B (SEB) (Sigma-Aldrich) in IMDM, 10% FCS, 86μg/ml gentamycin (Duchefa). The addition of SEB in combination with allogeneic stimulation of cells provides for maximal T-cell stimulation and proliferation in order to best achieve the numbers required for analysis. Cells were split at day 5 and day 7 with 20U/ml IL-2 being added to the medium. At day 8, 5x104 cells/well were plated on 96-well flat bottomed plates and infected with SF162 (TCID50/ml 1000) or LAI (TCID50/ml 200). Since iDCs don’t survive this length of time in culture under these conditions the measured levels of infection represent direct cis-infection of Th-cells, indeed monitoring cell phenotypes indicated that only CD4+ lymphocytes were present at time of infection. Day 5 and 7 post-infection cells were re-stimulated for 6h with 10ng/ml PMA (Sigma-Aldrich), 1μg/ml ionomycin (Sigma-Aldrich), 10μg/ml brefeldin A (Sigma-Aldrich) supplemented with 0.1μg/ml T1294 (Pepscan Therapeutics BV) to prevent new infections during this period. Subsequently, cells were fixed in 3.7% formaldehyde and stored for no more than 1 week at 4⁰C in FACS buffer (PBS+ 2%FCS) for flow cytometry analysis measuring intracellular viral p24 antigen as well as an array of cell surface markers or intracellular cytokines/chemokines. Where monomeric gp120 (SF162) (Immune Technology Corp.) or RANTES (Sigma-Aldrich) were added the experiments were performed similarly. These compounds (concentrations in text) were added to the T-cell culture 24h after LAI infection and four days later the cells were re-stimulated. Experiments with SEA show the results of at least five different donor combinations and for experiments using rω-1 four different donor combinations are depicted in figures (see respective Fig legends).
To measure intracellular activation markers, cells were permeabilized for 30min with PermWash at RT (BD Pharmingen) and subsequently stained with p24-PE (Beckman Coulter) and IFN-γ-FitC, IL2-PerCp-Cy5.5, IL-4-APC, TNF-α-PE-CF594, Mip-1β-AlexaFluor700 (all from BD Bioscience) for 30min at 4°C at a pre-determined dilution. Next, these cells were washed once with PermWash and taken up in FACS buffer (PBS+ 2%FCS) after which they were analyzed using a FACS Canto II machine. For determining CCR5 surface expression, T-cells obtained from our T-cell outgrowth model were fixed in 3.7% formaldehyde before HIV-1 infection and stored for no more than 1 week at 4°C in FACS buffer. These cells were stained in FACS buffer with CCR5 PE-Cy7 (Biolegend) for 30min at RT, washed with FACS buffer and measured. Shown in histograms are cytokine producing T-cells with samples taken 5 and 7 days after HIV-1 infection with the optimal time point being determined by the percentage of live cells (>50%) and level of HIV-1 infection, which varied per donor and per virus. Gating strategies and flow cytometric controls are represented (S3A Fig and S3B Fig, respectively).
All cells were isolated from anonymized buffy coats purchased through the Nederlands Blood Bank (Sanquin) where IRB approval was granted and donors signed a specifically consented in the additional use of samples for research purposes as outlined in the research contract between the AMC and Sanquin (number NVT0202.02).
Data was analyzed using an unpaired t-test when comparing two groups and a 1-way ANOVA when comparing several groups (Fig 1). For the remaining figures the data is corrected for the systematic differences between donors using factor correction [71]. Subsequently, within each specific T-cell population the effect of induction with SEA/rω-1 exposed or unexposed DCs was compared using a paired T-test whereas data between two T-cell population’s data was compared using an unpaired T-test. Additionally for Fig 2, a 2-way ANOVA was performed to determine whether there was an interaction between the T-cell population infected and the usage of SEA.
|
10.1371/journal.pbio.2005479 | Nonmonotonic recruitment of ventromedial prefrontal cortex during remote memory recall | Systems-level consolidation refers to the time-dependent reorganisation of memory traces in the neocortex, a process in which the ventromedial prefrontal cortex (vmPFC) has been implicated. Capturing the precise temporal evolution of this crucial process in humans has long proved elusive. Here, we used multivariate methods and a longitudinal functional magnetic resonance imaging (fMRI) design to detect, with high granularity, the extent to which autobiographical memories of different ages were represented in vmPFC and how this changed over time. We observed an unexpected time course of vmPFC recruitment during retrieval, rising and falling around an initial peak of 8–12 months, before reengaging for older 2- and 5-year-old memories. This pattern was replicated in 2 independent sets of memories. Moreover, it was further replicated in a follow-up study 8 months later with the same participants and memories, for which the individual memory representations had undergone their hypothesised strengthening or weakening over time. We conclude that the temporal engagement of vmPFC in memory retrieval seems to be nonmonotonic, revealing a complex relationship between systems-level consolidation and prefrontal cortex recruitment that is unaccounted for by current theories.
| Our past experiences are captured in autobiographical memories that allow us to recollect events from our lives long after they originally occurred. A part of the brain’s frontal lobe, called the ventromedial prefrontal cortex (vmPFC), is known to be important for supporting autobiographical memories, especially as memories become more remote. The precise temporal profile of the vmPFC’s involvement is unclear, yet this information is vital if we are to understand how memories change over time and the mechanisms involved. In this study, we sought to establish the time course of vmPFC engagement in the recollection of autobiographical memories while participants recalled memories of different ages during functional magnetic resonance imaging (fMRI). Using a method that detects the brain activity patterns associated with individual memories, we found that memory-specific neural patterns in vmPFC became more distinct over the first few months after a memory was formed, but then this initial involvement of vmPFC subsided after 1 year. However, more remote memories (2 years and older) appeared to reengage vmPFC once again. This temporal profile is difficult to accommodate within any single existing theory. Consequently, our results provoke a rethink about how memories evolve over time and the role played by the vmPFC.
| We possess a remarkable ability to retrieve, with ease, one single experience from a lifetime of memories. How these individual autobiographical memories are represented in the brain over time is a central question of memory neuroscience that remains unanswered.
Consolidation takes place on two levels, which differ on both a spatial and temporal scale. On a cellular level, the stabilisation of new memory traces through modification of synaptic connectivity takes only a few hours [1] and is heavily dependent upon the hippocampus [2–5]. On a much longer timescale, the neocortex integrates new memories, a form of consolidation termed ‘systems-level’ [6]. The precise timeframe of this process is unknown. A related long-standing debate that has contributed to this uncertainty is whether or not the hippocampus ever relinquishes its role in autobiographical memory retrieval. One theory asserts that the hippocampus is not involved in the retrieval of memories after they have become fully consolidated to the neocortex [7]. Alternate views maintain that vivid, detailed autobiographical memories retain a permanent reliance on the hippocampus for their expression [8–12].
An undisputed feature of systems-level consolidation, however, is the strengthening of neural representations in the neocortex over time. Clarity on the time course of systems-level consolidation is therefore more likely to be achieved through scrutiny of its neocortical targets. While theoretical accounts often fail to specify these cortical locations, animal experiments have consistently implicated the medial prefrontal cortex. While this region has been associated with the formation [13, 14] and recall of recently acquired memories [15–17], in rodents, it appears to be disproportionately involved in the retrieval of memories learned weeks previously [18–26]. The dependency on this region, which emerges over time, is facilitated by postlearning activation [27] and structural changes [28–30].
The evolutionary expansion of prefrontal cortex in humans makes it challenging to make direct anatomical comparisons with rodents, but the ventromedial prefrontal cortex (vmPFC) has been proposed as a homologous site of long-term memory consolidation [31]. It may appear surprising that an association between impaired autobiographical memory retrieval and vmPFC lesions has only recently started to be more precisely characterised [32]. However, there are a number of confounding factors in this field [33]—nonselectivity of vmPFC lesions, methodological differences in memory elicitation, and the tendency of patients with vmPFC damage to recollect events that have never occurred, a phenomenon known as confabulation [34].
Numerous functional magnetic resonance imaging (fMRI) studies of vmPFC activity during autobiographical memory recall have been conducted but with inconclusive results. Delay-dependent increases in retrieval-related activity have been observed in some studies [35, 36] but not others [37–39]. Autobiographical memory, in particular, induces robust vmPFC engagement [40], but it is unclear whether this activity increases [41], decreases [42], or remains constant in accordance with memory remoteness [43–52].
A powerful method of fMRI analysis that can help to bridge the empirical gap between the human and animal literatures is multivoxel pattern analysis (MVPA), because of its increased sensitivity to specific neural representations [53]. Using this approach, Bonnici and colleagues [54] demonstrated that remote 10-year-old autobiographical memories were more detectable in the vmPFC than recent 2-week-old autobiographical memories, consistent with its proposed role as a long-term consolidation site. This difference was not apparent in other cortical areas, nor did it emerge from a standard univariate analysis. A follow-up study 2 years later with the same participants and memories demonstrated that the original 2-week-old memories were now as detectable in the vmPFC as the remote memories [55]. This suggested the recent memories had been fully consolidated in the vmPFC after just 2 years and perhaps even sooner.
The identification of this 2-year time window represented an opportunity to resolve the time course of systems-level consolidation with high precision. To do so, we sampled memories from 4-month intervals spanning a 2-year period and compared their neural representations using fMRI. As opposed to the pattern-classification approach employed by Bonnici and colleagues [54] to decode the neural signatures of individual memories, we used representational similarity analysis (RSA) [56]. This method compares the consistency of neural patterns across repetitions of a single memory against all other unrelated memories to detect its unique informational content in a region of interest (ROI). Differences in the strength of memory representations across time periods were interpreted as delay-dependent engagement of the vmPFC. To verify observed time-sensitive differences, we followed the neural evolution of individual memories in a follow-up study with the same participants and memories 8 months later. The selection of numerous time points characterised the consolidation process with unprecedented temporal resolution, while the longitudinal design was an opportunity not only to replicate these findings but to observe systems-level consolidation in action.
Systems-level consolidation is generally assumed to be an incremental process; therefore, we considered a gradual linear trajectory of vmPFC recruitment as the most likely outcome. The alternative hypothesis was a rapid strengthening of vmPFC neural representations in the first few months after an event. The results conformed to neither scenario and revealed an unexpected temporal relationship—a transient recruitment of the vmPFC beginning in the months following the initial experience, followed by an enduring signature of more remote memories. The second, longitudinal experiment confirmed this finding. This is the first demonstration, to our knowledge, of such a temporal dissociation in vmPFC-mediated memory retrieval.
One week prior to the fMRI scan, with the assistance of personal photographs, participants (n = 30) verbally recalled and rated the characteristics of autobiographical memories from 8 time periods: memories that were 0.5 months old (0.5 M, i.e., 2-week-old memories), 4 M, 8 M, 12 M, 16 M, 20 M, 24 M, and also 60 M old—these latter memories serving as a definitive benchmark for remote (5-year-old) memories (see Materials and methods, Fig 1A). Two memories from each time period that were sufficiently vivid, detailed, specific, and unique in time and place were chosen for subsequent recall in the scanner. This meant that there were 2 full sets of memories. Participants created a short phrase pertaining to each autobiographical memory, which was paired with the photograph to facilitate recall during the subsequent fMRI scan.
The nonmonotonic pattern we observed in the fMRI data did not manifest itself in the subjective or objective behavioural data. In fact, the only difference in those data was higher ratings for the most recent 0.5 M old memories. However, these were paradoxically the most weakly represented memories in the vmPFC, meaning the neural patterns were not driven by memory quality. The objective scoring of the memories confirmed comparable levels of detail provided for all memories, without any significant drop in episodic detail or increase in the amount of semantic information provided as a function of time. Therefore, the amount or nature of the memory details were not contributing factors.
Nevertheless, to verify that the results genuinely represented the neural correlates of memory purely as a function of age, one would need to study the effects of the passage of time on the individual neural representations. Therefore, we invited the participants to revisit 8 months later to recall the same memories again both overtly and during scanning; 16 of the participants agreed to return. In order to generate specific predictions for the neural representations during Experiment 2, we took the actual data for the 16 subjects from Experiment 1 who returned 8 months later (Fig 5 green line, in which the nonmonotonic pattern is still clearly evident) and shifted them forwards by 2 time points to simulate the expected pattern 8 months later (Fig 5 pink dotted line). Note that for the 28 M and 32 M time periods in Experiment 2, we assumed they would have the same level of detectability as 24 M old memories, given the absence of data relating to these time periods from Experiment 1. We further assumed the neural representations between 60 M and 68 M would be unchanged.
A comparison of the original and simulated neural representation scores yielded a number of clear hypotheses about how memory representations would change over time in the vmPFC. Two-week-old memories should become detectable 8 months later, while the original 4 M and 8 M old memories should not differ in their representational strength. Twelve-month-old memories from Experiment 1 should be significantly less detectable, whereas 16 M old memories should remain unchanged. The original 20 M old memories should be better represented at 28 M, whereas the 24-and 60-month old-memories from Experiment 1 were not predicted to change over time.
One week prior to the fMRI scan, with the assistance of the personal photographs and previously chosen phrases that were used as cues in Experiment 1, the participants verbally recalled and rated the characteristics of their autobiographical memories just as they had done 8 months previously (see Materials and methods and Fig 6A).
This study exploited the sensitivity of RSA to detect not only the extent to which memories of different ages were represented in the vmPFC but how these representations changed over time. During Experiment 1, we observed detectability in vmPFC for memories at 4 M to 12 M of age, which was also evident at 24 M and 60 M. As expected, recent 0.5 M old memories were poorly represented in vmPFC in comparison. Curiously, however, the same lack of detectability in vmPFC was observed for memories that were 16 M to 20 M old. This pattern persisted across separate sets of memories and was replicated in a follow-up study 8 months later with the same participants and memories. Behavioural data failed to account for these time-dependent representational changes in either experiment, and other regions failed to show a significant change in memory representations over time. These findings are difficult to accommodate within any single theoretical account of long-term memory consolidation [9, 12, 61–63], as neocortical recruitment is generally assumed to involve an ascending linear trajectory. Consolidation has been characterised as fluid and continuous [64], but the nonmonotonic vmPFC engagement observed here suggests additional complexity in its temporal recruitment.
Over the course of consolidation in this study, the vmPFC twice alternated between disengagement and engagement, indicative of 4 separate stages. Below we consider, based on the latest theoretical developments and empirical research on systems-level consolidation and vmPFC functioning, the time-dependent processes which could underlie such a nonmonotonic pattern.
The current findings have potential implications for the two dominant theoretical perspectives on systems-level consolidation. Standard consolidation theory [7] predicts that the passage of time promotes the strengthening of neural representations in the neocortex, but the duration of this process in humans is poorly specified. The current results suggest this process is accomplished over a relatively fast timescale on the order of months. The alternative perspective on consolidation, multiple trace theory and transformation hypothesis [10], posits that over time, consolidation promotes the emergence of schematic, gist-like representations in the neocortex, which complement the original detailed memory. The reengagement of the vmPFC at 2 years in this study may reflect the emergence of these generalised representations to facilitate specific recall at more remote time points. Therefore, the consolidation of new memories in the neocortex may be reasonably rapid, whereas the transformation of these engrams may take place over a much longer timescale.
Using an autobiographical memory paradigm to study consolidation is preferable to laboratory-based episodic memory tests by virtue of its ecological validity, availability of temporally distant stimuli, clinical significance, and context-dependent equivalence to animal tasks. However, studying autobiographical memory carries with it potential confounds that can affect interpretation of results. In the sections that follow, we consider why these factors cannot account for our observations.
Older memories may yield a higher RSA score if they are more consistently recalled. Here, however, participants actually rated 0.5 M memories as more consistently recalled than 60-month-old memories. Older memories were not impoverished in detail when compared to the detail available for recent memories. Moreover, an inspection of interview transcripts across experiments revealed participants rarely offered new details for previous memories when retested, countering the suggestion that increased detectability of old memories may arise from the insertion of new episodic or semantic details [77]. The consistency in recalled detail across experiments could be attributable to participants recalling in Experiment 2 what they had said during Experiment 1. However, whether or not participants remembered by proxy is irrelevant, as they still recalled the specific details of the original event, removing forgetting as a potential explanation of changes in neural patterns over time.
Retrieving a memory initiates reconsolidation, a transient state in which memories are vulnerable to interference [78, 79]. Therefore, repeated retrieval may cause this process to have an influence on neural representations. However, all memories were recalled 1 week before the fMRI scan, so if such an effect was present, it would be matched across time points. Retrieval at this stage may also accelerate consolidation [80], yet if this were a major influence, we would likely have found 0.5 M memories to be more detectable than they were. Further repetition of memories within the scanner in Experiment 1 took place over a timescale that could not affect consolidation processes or interpretation of the initial neural data. Nevertheless, this could arguably affect vmPFC engagement over a longer period of time [81] and thus perturb the natural course of consolidation, influencing the results of Experiment 2. However, given that 7 out of the 8 specifically hypothesised temporally sensitive changes in neural representations were supported, an altered or accelerated consolidation time course appears highly unlikely. Again, recall recency was matched in Experiment 2 by the memory interview, and recall frequency between experiments was low.
Taking a more general and parsimonious perspective, the ratings demonstrate that, naturally, all memories are recalled on an occasional basis (Table A in S1 Table); therefore, it seems highly unlikely that a mere six repetitions within a scanning session would significantly alter the time course of systems-level consolidation. It should also be noted that successful detection of neural patterns relied on the specific content of each memory rather than being due to generic time-related retrieval processes (S4 Fig). One alternative to the current two-experiment longitudinal design to limit repetition across experiments would be to have a different group of participants with different memories for the second experiment. However, the strength of the current approach was the ability to track the transformation in neural patterns of the same memories over time.
An alternative interpretation of the time-sensitive vmPFC engagement is a systematic bias in the content of selected memories—for example, annual events coinciding across all participants, such as a seasonal holiday. However, recruitment took place over a period of 5 months in an evenly spaced manner, ensuring that such events did not fall into the same temporal windows across participants. The occurrence of personal events such as birthdays was also random across participants. The use of personal photographs as memory cues also limited the reliance on time of year as a method for strategically retrieving memories. Furthermore, the nature of memory sampling was that unique, rather than generic, events were eligible, reducing the likelihood of events that were repeated annually being included. Memory detectability was high at 12-month intervals such as 1, 2, and 5 years in this study, suggesting perhaps it is easier to recall events that have taken place at a similar time of year to the present. However, this should have been reflected in behavioural ratings and equivalently strong neural representations for recent memories, but neither was observed. Most importantly, if content rather than time-related consolidation was the main influence on memory detectability, then we would not have observed any change in neural representation scores from Experiment 1 to Experiment 2, rather than the hypothesised shifts that emerged.
A related concern is that memories across time differ in nature because they differ in availability. Successful memory search is biased towards recency, meaning there are more events to choose from in the last few weeks than in remote time periods. Here, this confound is circumvented by design, given that search was equivalently constrained and facilitated at each time point by the frequency at which participants took photographs, which was not assumed to change in a major way over time. These enduring ‘snap-shots’ of memory, located within tight temporal windows (see Materials and methods), meant that memory selection was not confounded by retrieval difficulty or availability. It could also be argued that selection of time points for this study should have been biased towards recency, given that most forgetting occurs in the weeks and months after learning. However, it is important to dissociate systems-level consolidation from forgetting, as they are separate processes that are assumed to follow different time courses. Memory forgetting follows an exponential decay [82], whereas systems-level consolidation has generally been assumed, until now, to be gradual and linear [83]. Our study was concerned only with vivid, unique memories that were likely to persist through the systems-level consolidation process.
A further potential concern regarding memory selection is that recent and remote memories that are comprised of equivalent levels of detail must be qualitatively different in some way. For example, selected remote memories must have been highly salient at the time of encoding to retain such high levels of detail. However, the underlying assumption that individual memories invariably become detail impoverished over time does not necessarily hold. While the volume of memories one can recall decreases over time [84], the amount of details one can recall from individual consolidated memories can actually increase over a 1-year delay [85]. While generalised representations are thought to emerge over the course of consolidation, they do not necessarily replace the original detailed memories [10], and the equivalent level of detail provided by participants across the two experiments here would suggest that memory specificity can be preserved over time. Furthermore, the possibility that remote memory selection may still be biased towards more salient memories is rendered unlikely by the method of memory sampling employed here. Because memories were chosen only from available photographic cues, the salience of recent and remote events was determined at the time of taking the photograph and not during experimentation. These photographs served as potent triggers of remote memories that were not necessarily more salient than recent memories and that may not have otherwise come to mind using a free-recall paradigm. In addition, one would expect more salient remote memories to score higher than recent memories on subjective ratings of vividness, personal significance, or valence, but this was not the case. Therefore, stronger neural representations at more remote time points were likely due to consolidation-related processes rather than qualitative difference between recent and remote experiences at the time of encoding.
Given that the medial prefrontal cortex is often associated with value and emotional processing [86], could these factors have influenced the current findings? Humans display a bias towards consolidating positive memories [87], and remembered information is more likely to be valued than that which is forgotten [88]. Activity in vmPFC during autobiographical memory recall has been found to be modulated by both the personal significance and emotional content of memories [89]. However, in the current two experiments, memories were matched across time periods on these variables, and the selection of memories through photographs taken on a day-to-day basis also mitigated against this effect. In the 8 months between experiments, memories either remained unchanged or decreased slightly in their subjective ratings of significance and positivity, suggesting that these factors are an unlikely driving force behind the observed remote memory representations in vmPFC. For example, if recent memories in Experiment 1 were not well represented in vmPFC because they were relatively insignificant, there is no reason to expect them to be more so 8 months later, yet their neural representation strengthened over time nonetheless.
A methodological discrepancy between this experiment and that conducted by Bonnici and colleagues [54] is the additional use of a photograph to assist in cueing memories. One possible interpretation of the neural representation scores is they represent a role for the vmPFC in the maintenance of visual working memory following cue offset. However, the prefrontal cortex is unlikely to contribute to maintenance of visual information [90]. Furthermore, if this was the driving effect behind neural representations here, the effect would be equivalent across time periods, yet it was not.
There is, however, an obvious inconsistency between the findings of the current study and that of Bonnici and colleagues [54]. Unlike that study, we did not detect representations of 0.5 M old memories in vmPFC. It could be that the support vector machine classification–based MVPA used by Bonnici and colleagues [54] is more sensitive to detection of memory representations than RSA; however, the current study was not optimised for such an analysis, because it necessitated an increased ratio of conditions to trials. Nonetheless, the increase in memory representation scores from recent to remote memories was replicated and additionally refined in the current study with superior temporal precision. One observation that was consistent with the Bonnici findings was the detection of remote memories in the hippocampus, which also supports theories positing a perpetual role for this region in the vivid retrieval of autobiographical memories [10, 12]. However, the weak detectability observed at more recent time points may reflect a limitation of the RSA approach employed here to detect sparsely encoded hippocampal patterns, which may be overcome by a more targeted subfield analysis [91].
There are, however, distinct advantages to the use of RSA over pattern-classification MVPA. RSA is optimal for a condition-rich design, as it allows for the relationships between many conditions to be observed. For example, in the current experiment, a visual inspection of the group RSA matrix (S1 Fig) does not reveal an obvious clustering of recent or remote memories that would indicate content-independent neural patterns related to general retrieval processes. The approach employed by Bonnici and colleagues [54] assessed the distinctiveness of memories within each time point from each other in order to detect memory representations. Should the neural patterns of a single memory become more consistent over time, yet also more similar to memories of the same age because of generic time-dependent mechanisms of retrieval, pattern classification would fail to detect a representation when one is present. In the current study, however, the two can be assessed separately, revealing memories at each time point become distinct from both memories of all other ages (Fig 4A) and identically aged memories (Fig 4C). The machine-learning approach employed by Bonnici and colleagues [54] to decode memory representations also requires the division of data into ‘training’ and ‘testing’ sets to classify unseen neural patterns [53]. This reduces the number of trials available for analysis, which would have been suboptimal for the current design because it would have necessitated an increased number of conditions and fewer trials per memory, whereas this restriction is not a necessity for RSA. Finally, because the pattern classification approach used by Bonnici and colleagues [54] compared memories from each time point directly to each other, they could not be analysed independently. In the current RSA design, the two sets of memories could be analysed separately from each other to ascertain if the temporal patterns could be replicated in an independent set of data. As is evident in Fig 4B, the nonmonotonic pattern of vmPFC recruitment was present in both sets of memories. The suitability of each MVPA method, therefore, depends on the study design and the research questions being posed.
In the light of our hypotheses, Experiment 2 generated one anomalous finding. Twenty-four-month-old memories from Experiment 1 were no longer well represented 8 months later. Why memories around 32 M of age are not as reliant on vmPFC is unclear, but unlike other time periods, we cannot verify this finding in the current experiment, as we did not sample 32 M memories during Experiment 1.
The current results revealed that the recruitment of the vmPFC during the expression of autobiographical memories depends on the exact stage of systems-level consolidation and that retrieval involves multiple sequential time-sensitive processes. These temporal patterns were remarkably preserved across completely different sets of memories in one experiment and closely replicated in a subsequent longitudinal experiment with the same participants and memories. These findings support the notion that the vmPFC becomes increasingly important over time for the retrieval of remote memories. Two particularly novel findings emerged. First, this process occurs relatively quickly, by 4 months following an experience. Second, vmPFC involvement after this time fluctuates in a highly consistent manner, depending on the precise age of the memory in question. Further work is clearly needed to explore the implications of these novel results. Overall, we conclude that our vmPFC findings may be explained by a dynamic interaction between the changing strength of a memory trace, the availability of temporally adjacent memories, and the concomitant differential strategies and schemas that are deployed to support the successful recollection of past experiences.
This study was approved by the local research ethics committee (University College London Research Ethics Committee, approval reference 6743/002). All investigations were conducted according to the principles expressed in the Declaration of Helsinki. Written informed consent was obtained for each participant.
|
10.1371/journal.ppat.1000835 | Selective Condensation Drives Partitioning and Sequential Secretion of Cyst Wall Proteins in Differentiating Giardia lamblia | Controlled secretion of a protective extracellular matrix is required for transmission of the infective stage of a large number of protozoan and metazoan parasites. Differentiating trophozoites of the highly minimized protozoan parasite Giardia lamblia secrete the proteinaceous portion of the cyst wall material (CWM) consisting of three paralogous cyst wall proteins (CWP1–3) via organelles termed encystation-specific vesicles (ESVs). Phylogenetic and molecular data indicate that Diplomonads have lost a classical Golgi during reductive evolution. However, neogenesis of ESVs in encysting Giardia trophozoites transiently provides basic Golgi functions by accumulating presorted CWM exported from the ER for maturation. Based on this “minimal Golgi” hypothesis we predicted maturation of ESVs to a trans Golgi-like stage, which would manifest as a sorting event before regulated secretion of the CWM. Here we show that proteolytic processing of pro-CWP2 in maturing ESVs coincides with partitioning of CWM into two fractions, which are sorted and secreted sequentially with different kinetics. This novel sorting function leads to rapid assembly of a structurally defined outer cyst wall, followed by slow secretion of the remaining components. Using live cell microscopy we find direct evidence for condensed core formation in maturing ESVs. Core formation suggests that a mechanism controlled by phase transitions of the CWM from fluid to condensed and back likely drives CWM partitioning and makes sorting and sequential secretion possible. Blocking of CWP2 processing by a protease inhibitor leads to mis-sorting of a CWP2 reporter. Nevertheless, partitioning and sequential secretion of two portions of the CWM are unaffected in these cells. Although these cysts have a normal appearance they are not water resistant and therefore not infective. Our findings suggest that sequential assembly is a basic architectural principle of protective wall formation and requires minimal Golgi sorting functions.
| The protozoan Giardia lamblia is the leading cause for parasite-induced diarrhea with significant morbidity in humans and animals world-wide, and is transmitted by water-resistant cysts. Giardia has undergone substantial reductive evolution to a simpler organization than the last common eukaryotic ancestor, which makes it an interesting model to investigate basic cellular mechanisms. Its secretory system lacks a Golgi, but trophozoites induced to differentiate to cysts generate organelles termed encystation-specific vesicles (ESVs). Previous work shows that ESVs are most likely minimal pulsed Golgi-like compartments for exporting pre-sorted cyst wall material. We tested whether the sorting function associated with classical trans Golgi networks was also conserved in these organelles. By tracking immature and processed forms of the three cyst wall proteins during differentiation we discovered a novel sorting function which results in partitioning of ESV cargo and sequential secretion of the cyst wall material. Using live cell imaging we identified reversible formation of condensed cores as a mechanism for cargo partitioning. These observations suggest that the requirement for sequential secretion of extracellular matrix components protecting Giardia during transmission has prevented the complete secondary loss of the machinery to generate Golgi cisterna-like maturation compartments; indeed, the preserved functions have been placed under stage-specific control.
| Infectious parasite stages transmitted to a new host via the oral route (cysts, oocysts, eggs) require a highly resistant extracellular matrix to protect them in the environment and during passage through the stomach. The diplomonad Giardia lamblia (syn. G. intestinalis, G. duodenalis) is an intestinal protozoan and a leading cause for parasite-induced diarrheal disease [1]. Trophozoites in the small intestine or in culture undergo stage-differentiation to a cyst form in response to environmental cues, e.g. changes in pH, bile and/or cholesterol concentration [2],[3]. The members of this phylum have undergone strong reductive evolution resulting in minimization or loss of cellular systems and organelles such as mitochondria, peroxisomes and the Golgi apparatus [4]–[6], but despite significant advances in phylogenetic analysis their point of divergence during evolution remains elusive [7]. Comparative genomic data suggest that the complexity of cellular organization in the last common eukaryotic ancestor with respect to compartments and membrane transport was considerable [4]. Specifically, the central organelle for maturation and sorting of excretory/secretory proteins, a classical Golgi apparatus likely with a typical stacked configuration of functionally distinct cisternae, appears to have been present in this hypothetical cell. Thus, reductive evolution is the most parsimonious, albeit still unproven, explanation for the absence of a Golgi organelle and Golgi functions in Giardia trophozoites [8].
In Giardia trophozoites, secreted proteins appear to traffic directly from the endoplasmic reticulum (ER) to the target organelle or the plasma membrane [9]. In contrast, in cells differentiating to cysts secretory cargo is delayed for many hours in specialized organelles termed encystation-specific vesicles (ESVs) which arise de novo [5],[10]. ESVs contain only presorted cyst wall material (CWM) and exclude constitutively secreted proteins even during neogenesis. The CWM rapidly polymerizes upon secretion and forms the protective cyst wall (CW) on the parasite surface at 20–24 h post induction (p.i.) of differentiation in vitro. The CWM biopolymer has a surprisingly low complexity considering its effectiveness as a biological barrier. It consists of three paralogous cyst wall proteins (CWP1–3) and simple β1–3 GalNAc homopolymer chains [11]. The glycan portion constitutes ∼60% of the CW [12], but where it is synthesized and how it is exported and incorporated into the cyst wall structure is unknown. Galactosamine synthesis from glucose and its incorporation as a polymer is mediated by pathways whose components are upregulated transcriptionally and allosterically during encystation [10], [13]–[16]. Synthesis of CWP mRNA peaks at ∼7 h p.i. and protein export from the ER to ESVs is completed after 8–10 h p.i. [17] in parasites encysting in vitro. CWPs are sorted away from constitutively secreted proteins presumably during ER export, thus ESVs contain only presorted cargo [9]. The pulsed synthesis and sorting of the CWPs to ESVs the cargo is delayed by many hours in the newly formed ESV organelle system which is best described as transient Golgi cisterna analogs [18], even though the compartments have no morphological similarity to a classical Golgi with biochemically distinct cisternae. Previously, we and others have shown transient association of COPI components with ESVs [5], ESV sensitivity to brefeldin A [9],[10], and dependence of ESV genesis and maturation on giardial Sar1 and Arf1 GTPases, respectively [18]. Taken together, there is increasing support for a model depicting ESVs as developmentally regulated, minimized Golgi-like organelles which undergo simultaneous maturation before being consumed during secretion of their cargo [8]. If confirmed, ESVs could be considered the most simply organized Golgi system identified as yet. ESVs arise stage-specifically and lack Golgi glycosyl transferases and typical structural or morphological landmarks which define this organelle in most other eukaryotes. This makes it impossible to test directly whether ESVs are indeed Golgi analogs or whether they arose independently during evolution. Thus, this issue can only be addressed by accumulation of circumstantial evidence and rigorous experimental testing of predictions based on this model.
Although few details are known it is reasonable to assume that export of the CWM is delayed in ESVs for several hours to allow for post translational maturation before it is secreted in fluid form to cover the entire cell surface where it eventually polymerizes. Proteolytic processing of CWP2 which has a 121 residue C-terminal extension rich in basic amino acids [19] is the only modification of CWPs described in any detail. Although the evidence clearly implicates a cysteine protease, there is a controversy as to which enzyme is responsible [20],[21]. In addition to processing, the enzymatic formation of disulfide [17] and isopeptide [22] bonds between CWPs appears to play a major role in the export process.
In the present study we address the question whether assembly of the giardial cyst wall requires an additional sorting step. This idea follows from a central prediction of our working model [23],[24], namely that ESVs as the only Golgi-like organelles in Giardia ultimately mature to a stage corresponding to the trans Golgi compartment of the classical Golgi whose principal function is sorting of mature cargo into distinct transport intermediates. However, a fundamental difference between ESVs and conventional Golgi cisternae is that the giardial organelles contain only a single type of pre-sorted cargo, the CWM, all of which is believed to be simultaneously secreted to the cell surface. In principle this should make a sorting step at this stage unnecessary except if the CWM were divided into distinct subfractions, for example as a result of post translational processing. This hypothesis is testable by analyzing the fate of all CWPs (pro-proteins and mature forms) by (quantitative) confocal fluorescence microscopy and Western blot. With this approach we discovered a completely novel cargo sorting function in mature ESVs resulting in partitioning of the proteinaceous CWM into two clearly defined fractions. Using specific antibodies and conditionally expressed epitope-tagged variants of CWPs we show that this processing/sorting mechanism which results in the sequential secretion of the CWM to the cell surface is necessary for the functional integrity of the cyst wall as a protective extracellular matrix.
Previous investigations of transport and secretion of the CWM provided evidence for proteolytic processing of the C-terminal extension of CWP2 [25]. The data suggested cleavage of the entire C-terminal extension of ∼13 kDa. However, the small C-terminal portion of the native or the transgenic CWP2 has never been visualized directly [26]. Processing of CWP2 was found to occur before secretion of the CWM but has not been correlated with expression kinetics or maturation and morphology of ESVs. To determine the temporal and spatial distribution of pro-CWP2 and its mature products we engineered a dually tagged CWP2 variant (Flag-CWP2-HA) for conditional expression under the CWP1 promoter (Figure 1A, B) [18]. Western analysis showed stage-specific expression of pro- Flag-CWP2-HA and appearance of a large processed form with a MR reduced by ∼5 kDa between 8 h and 10 h post induction (p.i.) (Figure 1A). The data are consistent with removal of a short C-terminal portion with the attached HA tag (ΔC-HA) from Flag-CWP2-HA which appears to be nearly complete at 12 h p.i. We were unable to resolve ΔC-HA on SDS-PAGE although it is readily detected in immunofluorescence microscopy analysis (IFA) (see below). To make a rough determination of the proteolytic cleavage site we expressed two modified Flag-CWP2-HA variants containing deletions from N244 to A272 (ΔPS) or A300 to V359 (ΔPS3) in the C-terminal domain (Figures 1B and S1). Western blot analysis of protein from transgenic cells at 10 h p.i. confirmed processing of ΔPS but not of ΔPS3. Combined with the apparent mass difference after removal of ΔC, this is consistent with cleavage of Flag-CWP2-HA ∼50–60 amino acids from the C-terminus.
During the 20–24 h process of encystation, ESVs arise de novo through export of CWM from the ER and attain their final dimension of ∼500 nm between 8 h and 10 h p.i [9]. As a standard marker to follow this organelle development we and others use a commercially available mAb against CWP1. The signal observed in optical sections of ESVs generated by confocal IFA in encysting trophozoites at 10–12 h p.i. showed a characteristic ring-like distribution of the anti-CWP1 antibody (Figure 1C) compared to an even staining of organelle contents typical for earlier stages (see also Figure 2A). The ring-like distribution of the CWP1 marker was a transient phenomenon and disappeared after a few hours. The simplest explanation for this observation was that the CWM in more mature ESVs became condensed during maturation which limited penetration of the anti-CWP1 antibody into the organelle. This was also consistent with the conspicuous electron density of ESV contents (Figure 1C). However, unlike in secretory granules of other unicellular organisms, e.g., rhoptries of Toxoplasma gondii [27] or dense core granules of Tetrahymena termophila [28], no lattice-like structure is detected in transmission EM micrographs of ESVs. Together with the demonstrated exchange of a soluble CWP1::GFP reporter within an ESV organelle network [18] at this stage of the encystation process this would argue against condensation in ESVs. As an alternative explanation we therefore considered that CWP2 and/or CWP3 could be involved in the formation of a putative core which excludes CWP1. To test this we performed high-resolution confocal IFA in cells expressing the Flag-CWP2-HA reporter under stage-specific control. In differentiating transgenic cells (Figure 2) we labeled developing ESVs using the anti-CWP1 antibody (red) and detected the N- or the C-terminus of the CWP2 reporter with the anti-HA or the anti-Flag antibody, respectively (green). At 6 h p.i. the HA and the CWP1 signals overlapped completely in ESVs (Figure 2A, merged image) as documented by co-localization analysis based on the three-dimensional reconstruction of all optical sections (scatter plot). Until at least 8 h p.i. the tagged CWP2 reporter is not processed (Figure 1A). Consistent with this, the Flag and the CWP1 signals also overlapped completely in ESVs (data not shown). This was still true in cells at 12 h p.i. although both markers now showed the typical ring-like distribution of the proteins at the periphery of ESVs (Figure 2B). ΔC-HA, on the other hand, had a completely different distribution at this stage and localized inside the ring-like staining pattern of the anti-CWP1 antibody (Figure 2C). This visual assessment was confirmed by quantitative analysis of the confocal image stack which demonstrated significant loss of signal overlap (scatter plot). The same characteristic distribution was found when a CWP2 monoclonal antibody (mAb) was used in combination with anti-HA instead of the mAb against CWP1 (Figure S2). As also shown below the anti-CWP2 mAb reacts with an epitope in the N-terminal portion of CWP2. The combined data was direct evidence for the physical separation of the ΔC-HA and Flag-N products consistent with proteolytic cleavage of the pro-protein as documented in Figure 1A. Thus, the small ΔC-HA fragment was a first marker localizing to a putative core of ESVs. The observed cargo partitioning was unaffected by swapping of epitope tags on the CWP2 reporter (data not shown). To complete this analysis of CWPs in fixed cells we conditionally expressed a HA-CWP3 reporter cloned in the same vector. Analysis of tagged CWP3 by Western blot indicated that, as for the closely related CWP1, this protein was not processed by proteolytic cleavage (data not shown). Interestingly, by confocal IFA, HA-CWP3 appeared also clearly segregated from CWP1 and localized to the same central portion of ESVs as did ΔC-HA (Figure 2D). Taken together, this is direct evidence for a partitioning of the CWM inside ESVs into two separate and physically distinct fractions, each containing a CWP2-derived product.
Partitioning of CWP1 from ΔC-HA within ESVs, together with the fluid nature of a CWP1::GFP reporter documented previously [18], strongly suggested that these components of the CWM assumed different physical states. Since HA-CWP3 showed the same distribution as ΔC-HA, this allowed us to directly test the hypothesis that the mechanism for cargo partitioning was indeed formation of a condensed core in ESVs. Fluorescence recovery after photobleaching (FRAP) was used to quantify the degree of mobility of a CWP3::GFP reporter in the ESV organelle system in living transgenic cells. In analogy to the experiment with CWP1::GFP [18], exchange of CWP3::GFP between ESVs was used as a measure of condensation and core formation. We tested this in cells prior to appearance of mature ESVs at 6 h p.i. and found clear evidence for recovery of fluorescence in bleached organelles (Figure 3A, quantitative analysis). Recovery showed similar kinetics as observed previously for CWP1::GFP [18], which proved that in principle the soluble CWP3::GFP reporter could be transported between ESVs. In contrast, in cells at 12 h p.i. which contained maturing ESVs, recovery of fluorescence was consistently absent (Figures 3B and S3A, B). Note also the higher rate of fluorescence loss in control organelles (6 h time point) due to dilution of the GFP pool within the ESV system during the recovery period (compare quantitative analyses in Figures 3A and 3B). To show that CWP3::GFP is immobilized in ESV cores but CWP1::GFP is not we performed fluorescence loss in photobleaching (FLIP) experiments in transgenic cells at 12 h p.i. We used six rapid cycles to bleach fluorescence in all but one ESV and quantified fluorescence loss in this organelle as a measure of diffusion in the ESV system (Figure S3C, D). Consistent with previous observations of CPW1::GFP mobility and the FRAP analysis of CWP3::GFP presented herein, we find rapid diffusion and signal loss in ESVs containing CWP1::GFP compared with the sustained fluorescence of CWP3::GFP in ESVs at this stage of encystation. Taken together, this is direct evidence for virtually complete immobilization of the CWP3 reporter in mature ESVs and consistent with core formation and loss of solubility. Combined with previously reported FRAP data using CWP1::GFP [18] this strongly supports the idea that a hallmark of maturing ESVs is partitioning of the CWM into two fractions with distinct physical properties.
Sorting mechanisms based on selective condensation of secretory cargo and formation of condensed cores in the trans Golgi network (TGN) and in post Golgi vesicles of mammalian cells have been described by the “sorting by retention model” [29],[30]. In analogy, condensation of CWM components in maturing ESVs suggested that this cargo is selected for differential secretion. Indeed, in transgenic cells at 14–16 h p.i. dual labeling revealed that cargo partitioning in ESVs gave way to actual sorting of cargo into separate compartments (Figure 4A, B). The fraction consisting of CWP1 and the large N-terminal portion of the processed CWP2 (collectively termed CWMfl), presumably remained in a fluid state throughout, and appeared to be concentrated in small compartments with peripheral localization in the cell. Tagged ΔC and CWP3 proteins, collectively termed CWMco), were detected in organelles with a more central localization. Thus, whilst the mechanism for partitioning of the CWM within ESVs (i.e. physical separation of the two fractions) is clearly condensation, the cellular machinery for the subsequent sorting of CWMco and CWMfl into distinct organelles remains to be identified, but possibly involves coat protein complexes. This idea is also based on localization studies which show that clathrin (CLH) is specifically recruited to membranes of maturing ESV (Figure S4A) [8],[31]. CLH is not upregulated during encystation but the protein appears to re-localize from the membranes of the endosome-lysosome-like peripheral vesicle organelles to ESVs. CLH is most abundant on maturing ESVs with evidence for a condensed core, and appears to lose this association as sorting progresses (Figure S4B, C). Whether clathrin is directly involved in sorting of CWMfl or has another role remains to be determined. Classical clathrin coated pits on ESV membranes, at least, have never been demonstrated by electron microscopy.
The significance of this sorting event only became evident when newly formed cysts were analyzed by IFA at 16 h p.i. (Figure 4C). The CWMfl fraction (represented here by CWP1) was secreted quantitatively whereas CWMco (represented by ΔC-HA) remained in internal compartments. Correspondingly, CWMfl and CWMco lost colocalization completely as CWMfl was deposited on the outer face of the plasma membrane during morphological differentiation of the trophozoites into cysts (Figure 4C). Yet, if the cysts were allowed to mature longer and were harvested at 24 h p.i., all cysts showed partial and some even full recovery of marker colocalization at the cyst wall (Figure 4D, E) as documented in the quantitative analysis (scatter plots). This suggested that ΔC-HA, as well as CWP3-HA or CWP3::GFP (Figure S5A–C) were secreted with clearly different kinetics, suggesting a requirement for sequential deposition of the CWM fractions.
Partitioning of CWM, core formation and processing of tagged and endogenous CWP2 all appeared to take place around 10–12 h p.i. Together with an idea presented recently by the Lujan laboratory that CWP2 coordinated export of CWM [25], the simplest explanation was that this change of physical property was triggered by the release of ΔC. To test this we inhibited processing of the Flag-CWP2-HA reporter as well as endogenous CWP2 by treating encysting cells with the protease inhibitor E64 shown to block giardial cysteine protease 2 (CP2) [20]. The Western analysis of parasites harvested at 12 h p.i. confirmed complete inhibition of processing in these conditions (Figure 5A, B). Surprisingly, using labeled anti-CWP1 antibody as a marker we found that cyst formation was not significantly impaired. More detailed analysis of fixed transgenic cells by IFA showed that in cysts derived from treated cells, unprocessed Flag-CWP2-HA remained in internal vesicles containing the CWMco fraction (Figure 5C). General cargo partitioning and sorting of the CWMfl and CWMco fractions and sequential secretion appeared to be unaffected despite the changed composition. This suggested that in treated cells CWP1 was the only family member which was exported during the formation of the first layer of the CW, followed by the components of CWMco which now included the unprocessed CWP2 (Figure 5C). Altogether, the results indicate that proteolytic cleavage of CWP2 is not necessary to induce core formation. This leaves two possibilities for the role of CWMco components: CWP3 can induce condensation alone, or alternatively, through interaction with the ΔC portion of CWP2 independent of processing. Sequestration of unprocessed Flag-CWP2-HA in the condensed core and in CWMco compartments of E64 treated cells suggests the presence of a dominant sorting signal in the short ΔC domain. An additional conclusion from these experiments was that building of the first layer of the cyst wall whose likely function is to provide structural stability to the morphologically transformed cell appeared to be independent of CWP2 processing and trafficking. Interestingly, the truncated ΔPS3 variant of the Flag-CWP2-HA reporter, which was not processed because it lacks the cleavage site, showed an identical distribution in maturing cysts derived from untreated cells (Figure S5D) as the wild type variant in cells treated with E64. This might also indicate that cleavage has to be very precise for the N-terminal part of CWP2 to be exported with CWMfl, or that cleavage and partitioning are coupled processes.
Although two CWM fractions appeared to be secreted sequentially in differentiating cells treated with E64, we suspected that the viability of these cysts was compromised. We tested water resistance as a quantifiable hallmark of correctly formed cysts by exposing mature cysts derived from E64-treated cells and from untreated controls to cold water for >24 h. Quantification of survival rates (Figure 5D) shows that the number of viable cysts from treated cells was reduced by ∼90% after exposure to water although their cyst walls remained apparently intact. This is direct evidence that correct composition of the sequentially secreted CWM, which is achieved by processing of CWP2 and by sorting of the two products in maturing ESVs, is essential for the biological activity of cysts.
Efficient formation of water-resistant cysts of Giardia is a major contributing factor for the world-wide distribution of this extremely successful parasite. The simple organization and genetic tractability of Giardia allow for the study of basic principles of Golgi compartment neogenesis, sorting and regulated secretion in an uncluttered system [8],[32]. More importantly, by looking for universally conserved paradigms of protein trafficking and organelle organization in the minimal secretory system of Giardia we uncovered a completely unknown mechanism for cyst wall formation. The regulated secretory pathway in Giardia is established from ER-derived transport intermediates [9]. As the only Golgi-like compartments in Giardia, ESVs are exceptional since they contain only CWM and no constitutively secreted proteins [9], [23], [33]–[35]. Thus, ESVs constitute a laterally connected network of maturation compartments which is clearly distinguishable from the ER and whose synchronous maturation can be tracked during the entire 20–24 h of the differentiation process in vitro. The exported CWM has a very low complexity: Three paralogous CWPs are very likely the major proteins of the extracellular portion of the giardial cyst wall [26],[36], in addition to a simple β1–3 GalNAc homopolymer glycan [11],[12], whose manner of integration with CWPs is unknown. A cysteine-rich membrane protein (HCNCp) which may localize also to the plasma membrane of cyst forms [37] could function as a possible link between the cyst wall and the cell surface. Thus, the structural and organizational minimization in Giardia provides unique opportunities to investigate basic principles of extracellular matrix formation. Compared to Giardia cysts, environmentally resistant infectious stages of other pathogenic protozoa have more elaborate cyst walls. While the first layer of the giardial CW is secreted rapidly [8], the Entamoeba invadens CW is built more gradually from soluble secreted material and is anchored by a plasma membrane bound Gal/GalNAc lectin. This in turn binds to seven Jacob glycoproteins [38],[39], which cross-link chitin fibrils to establish a structural scaffold. In an elegant study, Chatterjee et al. [40] showed that construction of this structural part of the matrix, which also includes a chitinase [41],[42] involved in its remodeling, was followed by incorporation of Jessie3 proteins which provide the “mortar” that seals it. Sequential secretion of distinct CWM fractions from different secretory organelles was observed during establishment of the three distinct layers of the Eimeria oocyst wall [43],[44]. Based on these and other models we postulate that sequential assembly of multi-layered cyst walls from protein and carbohydrate is a universally conserved albeit polyphyletic trait required for full protection of infectious stages.
CWP2 with its prominent C-terminal extension (Figure S1) was postulated to act as an escorter for the other CWPs during export [25]. We have used a dually tagged CWP2 reporter (Flag-CWP2-HA) to investigate processing and trafficking of CWP2. In contrast to a previous report which postulated the removal of the entire C-terminal domain which is unique to CWP2 [21], our Western analysis and the examination of deletion variants indicated removal of only ∼5 kDa. Localization of C-terminally tagged CWP2 in the cyst wall by Sun and coworkers was interpreted as the presence of pro-CWP2 [26], which could mean that processing may not be required for incorporation into the matrix. Our results showed that both CWM fractions receive a portion of this domain rich in basic amino acids, indicating that it fulfills several functions. We also find evidence for the presence of a dominant sorting signal in the ΔC domain (see also below). Our data suggest that proteolytic cleavage of CWP2 is a discrete process that marks the transition from the ESV genesis to the ESVs maturation phase. Encystation is not completely synchronous in an induced population because parasites need to complete the S-G2 transition of the cell cycle in order to exit the proliferation cycle and differentiate [45]. However, our observations indicate that the transition into the maturation phase starts at ∼10 h post induction in the large majority of cells (Figure 6) and coincides with a marked downregulation of CWP synthesis [17].
Cargo partitioning provides a more sophisticated explanation for the incomplete staining by anti-CWP1 antibodies which was observed in maturing ESVs. Considering that ΔC contains a dominant targeting signal, processing can be interpreted as liberating the large N-terminal domain which remains soluble and can be secreted to the outer cyst wall layer. Since only CWP2 was processed and both its products could be detected in IFA all major players were followed either by epitope tagging or using a specific mAb in the case of CWP1 and CWP2. Investigation of cargo partitioning by high resolution confocal microscopy yields correspondingly clear cut results showing distinct localizations for these factors in either the center or the periphery of ESVs which can be quantitatively analyzed for co-localization. In addition to providing a sorting mechanism for partitioning of CWM, selective condensation theoretically allows for differential post-translational modification of components in fluid and condensed fractions. This could partially offset the lack of a stacked cisternal organization of this organelle system.
Biogenesis of secretory granules is still poorly understood [30]. Formation of immature granules occurs at the TGN in endocrine, exocrine and neuronal cells by sorting granule proteins from constitutively secreted cargo. Condensation of soluble proteins is organized by aggregation factors such as chromogranin A of neuroendocrine cells which drive granule formation independently of coat protein complexes [46],[47]. Core formation in secretory granule biogenesis is dependent on inherent biophysical properties of cargo components and aided by acidic pH and high Ca2+ in these organelles. Our attempts to disrupt or delay this process in ESVs by inhibiting acidification of organelles using ammonium chloride or the H+-ATPase inhibitor bafilomycin, or by depleting intracellular calcium were not successful (Konrad and Hehl, unpublished). This suggests that an inherent tendency to aggregate is the dominant driving force of CWMco condensation. Alternatively, interaction with an as yet unidentified component could prevent circulating CWMfl components from becoming condensed. Further maturation of secretory granules in higher eukaryotes entails sorting and removal of non-granule proteins by vesicular traffic involving AP1/clathrin [48]. We have observed important recruitment of clathrin to membranes of mature ESVs (Figure S4A) [9] but this alone does not prove any involvement in the sorting of CWMfl. Furthermore, because over-expression of a clathrin hub fragment during encystation had no effect on cyst formation (Stefanic and Hehl, unpublished data) the role of this coat protein remains to be determined. More interestingly, expression of a dominant-negative Arf1 homolog which also recruits AP1/clathrin prevented secretion of CWP1 from ESVs but not morphological transformation which suggests an essential function in late steps of regulated secretion [18].
One of the principal questions in connection with ESV maturation was whether a condensed core was formed. We were able to address this directly using a CWP3::GFP reporter to compare the physical state of CWP3 as a marker for the core with that of the closely related CWP1 protein whose dynamics was investigated previously [18]. FRAP and FLIP experiments provided the key piece of evidence for the interpretation of the results obtained by fluorescence microscopy which revealed progression from cargo partitioning to sorting and sequential secretion.
E64-inhibitable proteolytic processing of pro-CWP2 has been described previously by Touz and coworkers [21] using an antibody which binds to its N-terminal portion. In agreement with these findings we detected an initial retention of unprocessed CWP2 in internal compartments. In contrast, by using an extended experimental approach, i.e., co-labeling of markers for both CWM fractions, we documented the sequential nature of CWM secretion. Interestingly, our data showed that cyst formation was completed even when processing of endogenous and recombinant CWP2 was blocked (see Figure 5). Treatment of encysting cells with E64 did not affect stage-differentiation, secretion of CWM, or formation of an extracellular matrix, even though pro-CWP2 was retained in the CWMco fraction. Retention of ΔPS3, which differs from the mature N-terminal CWP2 fragment by only few amino acids, points to a surprisingly stringent dependence on precise cleavage of pro-CWP2 or on cleavage itself, which contrasts with the overall robustness of the sorting process. Detailed investigation of this step, including identification of the proteolytic cleavage site, will be necessary to resolve the link between processing and sorting of the two pro-CWP2 products.
Together with results from previous investigations the sorting data presented herein provide a novel scenario for regulated export of CWM in Giardia (Figure 6). The complete pathway requires two discrete sorting steps which are both consistent with our Golgi model for ESVs: I) Sorting of CWM from constitutively secreted proteins at ER exit sites [9], and concomitant export of CWPs to ESVs. II) Partitioning and sorting of the mature CWM into two fractions shortly before secretion. Processing of CWP2 coincides with, but is not required for, condensed core formation in ESVs. The subsequent separation and sequential secretion of the physically distinct CWMfl and CWMco fractions is consistent with maturation of ESVs to TGN analogs and a “sorting by retention” mechanism for separating differentially secreted cargo [29]. In addition to being a prerequisite for subsequent sorting, condensation of CWMco may serve to sequester this soluble content cargo from modifying factors. CWMfl components, however, continue to circulate and could theoretically intersect with other compartments such as the ER or PVs. Secretion of CWMfl is completed in only a few minutes simultaneously with loss and/or resorption of flagella, disassembly of the ventral disk and nuclear division [8],[49], and most likely provides primarily structural stability to the differentiated cell. Unlike reported previously [50], we find that encysting trophozoites adhere quite well in vitro, until just prior to secretion of the CWM (Trepp, Spycher and Hehl, unpublished) when they lose attachment as the cytoskeleton is reorganized. This allows the newly formed cyst walls to reach full function before cysts are finally shed into the environment.
Integration of chitin with early and late secreted proteins in encysting E. invadens is essential for establishing a fully functional cyst wall [40]. How the unique β1–3 GalNAc homopolymer chains which provide the bulk of the CW carbohydrate are integrated into this structure during encystation in Giardia remains unknown. Three factors, i.e., the fibrillar nature of the polymerized CWM in the outer cyst wall as shown in scanning EM [11], studies showing that this material is composed of carbohydrate and protein [15],[51], and the absence of specialized vesicles containing large amounts of this carbohydrate, raise the question how this material is exported. This still awaits resolution, mainly because no known lectin reacts with the carbohydrate portion of the giardial CW with sufficient specificity. The Phaseolus lunatus lectin (LBA) has been reported to bind to the G. muris CW [11], but reactivity with the G. lamblia CW is poor and inconsistent. More importantly, fluorochrome-conjugated LBA weakly labeled the nuclear envelope but not ESVs or other large organelles in encysting cells which might be involved in export of CW carbohydrate (Hehl, unpublished). Whatever the route of carbohydrate export, evidence for extensive covalent cross-linking [12],[24],[52] of CWPs and carbohydrate chains underscore the importance of a structurally resistant CW.
The simplest explanation for the sequential secretion of CWM components is that the fibrillar shell of polymerized CWMfl requires sealing to become fully protective and infectious. Based on the high proportion of intermediate stages (Figure 4D) found in cyst preparations at 16–24 h p.i., export of the CWMco fraction appears to be considerably slower than secretion of CWMfl, most likely because the former must be decondensed for secretion. In light of the low complexity of the CWM, investigation of the biochemistry of its reversible condensation and subsequent polymerization is expected to reveal fundamental aspects of biopolymer export and assembly. CWPs and their inherent tendency to aggregate may be the primary driving force for cargo partitioning and ultimately for polymerization on the surface. It is likely that carbohydrates play a much more important role in coordinating sequential secretion than merely providing a means for cross-linking the protein components of the CW.
Trophozoites of the Giardia lamblia strain WBC6 (ATCC catalog number 50803) were grown under anaerobic conditions in 11 ml culture tubes (Nunc, Roskilde, Denmark) containing TYI-S-33 medium supplemented with 10% adult bovine serum and bovine bile according to standard protocols [17]. For chemical fixation or protein extraction parasites were harvested by chilling the culture tubes on ice for 30 minutes to detach adherent cells, and collected by centrifugation at 1000×g for 10 minutes. Cells were then resuspended in phosphate-buffered saline (PBS) and counted.
Encystation was induced using the two-step method as described previously [17], by cultivating the cells for 44 hours in medium without bile and subsequently in medium with porcine bile and a pH of 7.85.
Circular plasmid DNA of expression vectors was linearized at the SwaI restriction site [18] and 15 µg of cut DNA were electroporated into 5·106 freshly harvested trophozoites on ice using the following settings: 350 V, 960 µF, 800Ω. Linearized plasmids were targeted to the Giardia lamblia triose phosphate isomerase (Gl-TPI) locus (see below) and integration occurred by homologous recombination under selective pressure of the antibiotic puromycin (Sigma, St. Louis, MO) for 5 days. Transgenic cell lines were maintained and analyzed without antibiotic.
For the inducible expression of tagged proteins in Giardia, a previously described vector pPacV-Integ was used which allows for the expression of fusion proteins with a N-terminal HA-tag under the control of the CWP1 promoter [18]. For the expression of double-tagged CWP2 (Flag-CWP2-HA), a Flag-tag was fused downstream of the stretch coding for the CWP1 signal peptide using oligonucleotide primers 44 and 768 (Table S1) to PCR amplify the CWP1 promoter including the CWP1 signal peptide from genomic DNA. The PCR product was ligated into the XbaI and NsiI sites upstream of the CWP2 coding sequence. The CWP2 open reading frame (ORF) without the stretch coding for the signal sequence (E21 - R362) was PCR amplified using oligonucleotide primers 760 and 756. The latter included the sequence coding for the HA epitope tag. This fragment was ligated in frame using the NsiI and PacI sites of the pPacV-Integ expression cassette to generate the basic Flag-CWP2-HA vector. All constructs were sequenced prior to transfection.
CWP2 deletion constructs: To express CWP2 lacking N244–A272 (ΔPS), two DNA fragments (coding for E21–R243 and H273–R362) were amplified with oligonucleotides 760 and 842, and 844 and 756, respectively, and ligated via the introduced EcoRI site, and used to replace the original NsiI - PacI fragment in the Flag-CWP2-HA vector. The same strategy was used to generate ΔPS3 lacking A300–V359 of the CWP2 ORF: the sequence coding for the E21–T299 fragment of the CWP2 ORF was PCR amplified using primers 760 and 877 and used to replace the NsiI - PacI fragment in the Flag-CWP2-HA vector.
CWP3 constructs: the CWP3::GFP expression construct was made by replacing the CWP1 sequence in a previously used construct CWP1::GFP [18] with the CWP3 ORF and promoter region PCR amplified with primers 936 and 937. An HA-tagged CWP3 (HA-CWP3) fragment was made by PCR amplifying the region coding for M17-R247 of the CWP3 ORF using primers 856 and 420 and replacing the NsiI-Pac fragment in the pPacV-Integ expression vector.
Encystation of trophozoites was induced in the presence of 30 µM E64 or an equal volume of the solvent (control). To determine the number of viable cysts in preparations cells were harvested after 48 h, washed with PBS and incubated in ddH2O for >48 hours at 4°C. For the quantification of cyst viability cells were stained with a mixture of acridin orange (4 ug/ml) and ethidium bromide (0.1 mg/ml) in PBS for 10 min at room temperature and washed once in PBS. Cells were mounted on a slide and imaged on a Leica DM-IRBE microscope using a 40× lens (Leica Microsystems GmbH, Wetzlar, Germany). For each condition two sets of 13 randomly selected fields were digitally recorded (Diagnostic Instruments Inc., USA) and processed with the Metaview software package (Visitron Systems GmbH, Puchheim, Germany). Percentage values of replicates were averaged.
For the preparation of total cell extracts Giardia parasites were harvested as described above. The cell pellet was dissolved in SDS sample buffer to obtain of 2·105 cells in 50 µl and boiled for 3 minutes. Dithiothreitol (DTT) was added to a final concentration of 7.75 µg/ml before boiling. SDS-PAGE on 12% polyacrylamide gels and transfer to nitrocellulose membranes was done according to standard techniques. Nitrocellulose membranes were blocked in 5% dry milk/0.05% TWEEN-20/PBS and incubated with the primary antibodies (anti-HA, anti-Flag, anti CWP2 mAb) at the appropriate dilution in blocking solution. Bound antibodies were detected with horseradish peroxidase-conjugated goat anti-mouse IgG (Bio-Rad, Hercules, CA) and developed using Western Lightning Chemiluminescence Reagent (PerkinElmer Life Sciences, Boston, MA, USA). Data collection was done in a MultiImage Light Cabinet with AlphaEase FC software (Alpha Innotech, San Leonardo, CA, USA) using the appropriate settings.
Immunofluorescence analysis: Chemical fixation and preparation for fluorescence microscopy was performed as described [9]. Briefly, cells were washed with cold PBS after harvesting and fixed with 3% formaldehyde in PBS for 40 min at 20°C, followed by a 5 min incubation with 0.1 M glycine in PBS. Cells were permeabilized with 0.2% triton X-100 in PBS for 20 min at room temperature and blocked overnight in 2% BSA in PBS. Incubations of all antibodies were done in 2% BSA/0.2% Triton X-100 in PBS. Cells were incubated with directly coupled mouse monoclonal antibodies, i.e., Alexa488-conjugated anti-HA (Roche Diagnostics GmbH, Manheim, Germany; dilution 1∶30), Cy3 conjugated anti-Flag (Sigma, St. Louis, MO 1∶30), or Texas Red-conjugated anti-CWP1 (Waterborne™, Inc., New Orleans, LA, USA; dilution 1∶80) for 1 h at 4°C. CLH was detected with a Giardia-specific polyclonal antibody [31]. Post incubation washes were done with 0.5% BSA/0.05% triton X-100 in PBS. Labeled cells were embedded for microscopy with Vectashield (Vector Laboratories, Inc., Burlingame, CA, USA) containing the DNA intercalating agent 4′-6-Diamidino-2-phenylindole (DAPI) for detection of nuclear DNA. Immunofluorescence analysis was performed on a Leica SP2 AOBS confocal laser-scanning microscope (Leica Microsystems, Wetzlar, Germany) equipped with a glycerol objective (Leica, HCX PL APO CS 63× 1.3 Corr). Confocal image stacks were recorded with a pinhole setting of Airy 1 and twofold oversampling. Further processing was done using the Huygens deconvolution software package version 2.7 (Scientific Volume Imaging, Hilversum, NL). Three-dimensional reconstructions and quantitative analysis of co-localization were done with the Imaris software suite (Bitplane, Zurich, Switzerland). Alternatively, a standard fluorescence microscope (Leica DM IRBE) and MetaVue software (version: 5.0r1) was used for data collection.
Live cell microscopy, fluorescence recovery after photobleaching (FRAP) and fluorescence loss in photobleaching (FLIP) analysis: For live cell microscopy, induced cells expressing the CWP3::GFP chimera were harvested at 6 or 12 h p.i. and transferred to 24-well plates at a density of 6·106/ml. After incubation on ice for 5–8 h, oxygenated cells were sealed between microscopy glass slides and warmed to 21°C or 37°C. Under these conditions, the encysting cells were stable and even continued to complete encystation. For FRAP, FLIP and time-lapse series, images were collected on a Leica SP2 AOBS confocal laser-scanning microscope (Leica Microsystems, Wetzlar, Germany) using a 63× water immersion objective (Leica, HCX PL APO CS 63× 1.2 W Corr). Fluorescence in selected regions of interest was quantified using the corresponding Leica software suite. The pinhole was set to Airy 2 in order to increase the thickness of the optical sections to accommodate an entire ESV in the z-plane. Quantifiable criteria for cell viability were active attachment to substrate and continuous beating of the ventral and anterolateral flagella pairs. FRAP experiments were performed with the same settings as used for the CWP1::GFP reporter [18] with Leica FRAP software module to set bleaching parameters and to quantify fluorescence recovery.
Electron microscopy: Encysting parasites were prepared for TEM as described previously [18]. To achieve uniform orientation, ultrathin sections were cut parallel to the sapphire surface, stained with uranyl acetate and lead citrate and examined in a CM12 electron microscope (Philips) equipped with a slow-scan CCD camera (Gatan, Pleasanton, CA, USA) at an acceleration voltage of 100 kV. Recorded pictures were processed further with the Digital Micrograph 3.34 software (Gatan, Pleasanton, CA, USA).
GiardiaDB accession numbers are given for the following genes: CWP1 GL50803_5638, CWP2 GL50803_5435, CWP3 GL50803_2421, clathrin heavy chain (CLH) GL50803_102108.
|
10.1371/journal.pgen.1008306 | Marcksb plays a key role in the secretory pathway of zebrafish Bmp2b | During vertebrate early embryogenesis, the ventral development is directed by the ventral-to-dorsal activity gradient of the bone morphogenetic protein (BMP) signaling. As secreted ligands, the extracellular traffic of BMP has been extensively studied. However, it remains poorly understood that how BMP ligands are secreted from BMP-producing cells. In this work, we show the dominant role of Marcksb controlling the secretory process of Bmp2b via interaction with Hsp70 in vivo. We firstly carefully characterized the role of Marcksb in promoting BMP signaling during dorsoventral axis formation through knockdown approach. We then showed that Marcksb cell autonomously regulates the trafficking of Bmp2b from producing cell to the extracellular space and both the total and the extracellular Bmp2b was decreased in Marcksb-deficient embryos. However, neither the zygotic mutant of marcksb (Zmarcksb) nor the maternal zygotic mutant of marcksb (MZmarcksb) showed any defects of dorsalization. In contrast, the MZmarcksb embryos even showed increased BMP signaling activity as measured by expression of BMP targets, phosphorylated Smad1/5/9 levels and imaging of Bmp2b, suggesting that a phenomenon of “genetic over-compensation” arose. Finally, we revealed that the over-compensation effects of BMP signaling in MZmarcksb was achieved through a sequential up-regulation of MARCKS-family members Marcksa, Marcksl1a and Marcksl1b, and MARCKS-interacting protein Hsp70.3. We concluded that the Marcksb modulates BMP signaling through regulating the secretory pathway of Bmp2b.
| Bone morphogenetic proteins (BMPs) are extracellular proteins which belong to the transforming growth factor-β (TGF-β) superfamily. BMP signaling is essential for embryonic development, organogenesis, and tissue regeneration and homeostasis, and tightly linked to various diseases and tumorigenesis. However, as secreted proteins, how BMPs are transported and secreted from BMP-producing cells remains poorly understood. In this study, we showed that Marcksb interacts with a molecular chaperon–Hsp70.3 to mediate the secretory pathway of BMP ligands during early development of zebrafish. Moreover, we discovered a novel phenomenon of “genetic over-compensation” in the genetic knock-out mutants of marcksb. To our knowledge, this is the first report that reveals the molecules and their related trafficking system mediating the secretion of BMPs. Considering the wide distribution of BMP and MARCKS within the human body, our work may shed light on the studies of BMPs secretion in organogenesis and adult tissue homeostasis. The finding of MARCKS in controlling BMP secretion may provide potential therapeutic targets for modulating the activity of BMP signaling and thus will be of interest to clinical research.
| Early vertebrate development involves the formation and patterning of body plan, such as dorsoventral axis formation and anteroposterior axis formation. Bone morphogenetic protein (BMP) signaling gradient is critical for the specification of ventral and posterior cell fate [1]. Like other morphogens, the formation of BMP signaling gradient depends on several factors, including the graded transcription and secretion of BMP ligands, the extracellular transport of BMP ligands and the interaction between BMP ligands and their antagonists [2]. In zebrafish, the secreted ligands Bmp2b and Bmp7a act as heterodimers and bind to their receptors type I and type II to transduce signal and to phosphorylate the regulatory Smads (Smads 1, 5, and 9), which in turn regulate BMP target genes with Smad4 in the nuclei [3, 4].
As secreted ligands, the extracellular traffic of BMP homolog Dpp has been extensively studied in Drosophila. The long-range distribution of Dpp is mainly dependent on restricted extracellular diffusion [5], which process is regulated by glypican members of heparin sulfate proteoglycans [6]. In zebrafish, it was reported that BMP gradient is mainly determined by the graded expression of BMP ligands [7]. The secretion of several morphogens, such as WNTs, FGF-2 and Hedgehog has been studied in different animal models [8–11]. Recent study implies that the release of Dpp is regulated by inwardly rectifying potassium channel and calcium transients [12]. However, it remains poorly understood how the secretory pathway, including the intracellular trafficking and the secretion to extracellular space, of BMP ligands is regulated.
The myristoylated alanine-rich C-kinase substrate (MARCKS) is a ubiquitous substrate for protein kinase C (PKC). Two conserved domains within the MARCKS proteins are known to be critical for their functions: the N-terminal myristoylated domain helps anchoring MARCKS to the plasma membrane; and the phosphorylation site domain (PSD) domain serves as the site for MARCKS binding to actin filaments and calcium/calmodulin [13–17]. A notable function of MARCKS is to regulate the secretion of different substances including airway mucin [18, 19]. The well-studied regulated mucin secretion process via MARCKS involves its PKC and calcium/calmodulin dependent phosphorylation, high binding affinity with F-actin and membrane phosphoinositides, and interaction with intracellular molecular chaperons [20–23]. The MARCKS family proteins have also been reported to play various roles in gastrulation movements in Xenopus [24] and zebrafish [25], and the morphogenesis of neural tube in mouse [26] and chick [27]. However, the potential roles of MARCKS in morphogen secretion and embryonic patterning has never been studied and reported.
In this study, we unveiled a role of a MARCKS family member–Marcksb in dorsoventral patterning by regulating the BMP signaling activity through interacting with Heat-shock protein 70 (Hsp70) to control the secretion of BMP ligands. Interestingly, unlike the marcksb knockdown embryos showing dorsalization, the maternal-zygotic mutants of marcksb (MZmarcksb) showed mild ventralization, suggesting that genetic over-compensation arises in the MZmarcksb embryos. We further proved that the transcription of other MARCKS family members were strongly activated during oogenesis of MZmarcksb females, and Hsp70.3 –the MARCKS interaction protein was up-regulated at shield stage in MZmarcksb embryos, suggesting a sequential compensation of different genetic factors.
We previously identified zebrafish marcksb which is important for gastrulation movements [25]. To further understand the role of MARCKS family genes in early embryonic development, we examined the expression patterns of all the four members of MARCKS family–marcksa, marcksb, marcksl1a and marcksl1b during early embryogenesis. Among these four genes, marcksb is the only one showing maternal expression and is the most highly expressed one at the time of zygotic genome activation (S1 Fig).
We then injected the morpholino (MO) blocking the translation of marcksb into zebrafish embryos and evaluated their phenotypes. The MO-injected embryos (morphants) showed spindle-like shape at bud stage (Fig 1A) and 77.9% showed dorsalization at 1 day post-fertilization (dpf) (Fig 1A and 1B). The defect of dorsalization in marcksb morphants was rescued by the injection of morpholino-insensitive marcksb mRNA (Fig 1A and 1B). Whole-mount in situ hybridization (WISH) analysis further confirmed the dorsalization defects in marcksb morphants, as revealed by the ventral expansion of otx2 expression (labeling neural ectoderm) (Fig 1C) and chordin expression (labeling dorsal organizer) (Fig 1D). Accordingly, the expression level and region of ventral markers foxi1 (labeling non-neural ectoderm) (Fig 1E) and eve1 (labeling ventral margin) (Fig 1F) were strongly reduced.
To understand whether inhibition of marcksb could affect the development of ventral tissues, we performed a tail organizer graft assay as described previously (Fig 1G) [28]. We transplanted the wildtype ventral margin cells to the animal pole of wildtype host, and as expected, 5 out of 27 host embryos had extra tails (Fig 1H). In contrast, when the ventral margin cells of marcksb morphants were grafted to wildtype embryos, they failed to induce any extra tail structures (Fig 1I). Taken together, our data indicate that marcksb is required for the specification of ventral cell fate in zebrafish.
As zygotic BMP signaling plays a pivotal role in specifying the ventral cell fate, we next examined the BMP signaling activity in marcksb-depleted embryos. WISH showed that the expression of two direct transcriptional targets of BMP signaling—szl and ved were decreased in marcksb morphants compared to wildtype embryos (Fig 2A and 2B). We then performed immunofluorescence to measure the nuclei-enriched phosphorylation level of Smad1/5/9 (p-Smad1/5/9). The data showed that the relative intensity of p-Smad1/5/9 was lower in marcksb morphants than that in wildtype embryos (Fig 2C and 2D). Moreover, knockdown of marcksb could restore the ventralization phenotype in bmp2b-overexpressed embryos (5 pg of bmp2b mRNA per embryo) (Fig 2E). Accordingly, the injection of bmp2b caused robust expression expansion of szl and ved at dorsal region, and this dorsal expansion could be inhibited by knockdown of marcksb (Fig 2F and 2G). Altogether, our data indicate that marcksb knockdown leads to attenuation of BMP signaling and marcksb is required for the normal activation of BMP signaling.
We then conducted ectopic overexpression experiments of marcksb. Since marcksb was strongly maternally expressed (S1 Fig), injection of moderate dosage of marcksb mRNA (200pg per embryo) did not result in any visible effects, whereas injection of extremely high dosage of marcksb mRNA (1000 pg per embryo) led to ventralization (Fig 3A and 3B).
To test whether phosphorylation of Marcksb is required for the activation of BMP signaling, two mutated forms of Marcksb, the HA-tagged S4D-Marcksb (phosphomimetic type) and S4N-Marcksb-HA (non-phosphorylatable type) were generated according to previous study [29], and their mRNA were injected into one blastomere at 16-cell stage (Fig 3C-a, b). The wildtype Marcksb-HA mainly localized at the cell membrane (Fig 3C-c). In accordance with the notion that phosphorylation of Marcksb leads its translocation from the cell membrane to the cytoplasm [19], S4D-Marcksb mainly located inside cytoplasm (Fig 3C-d) and the S4N-Marcksb mainly co-localized with membrane-labeled EGFP (Fig 3C-e). We then examined szl expression in the embryos overexpressed with mutated marcksb. When compared with wildtype embryos, marcksb overexpressed embryos showed mildly increased expression of szl (Fig 3D-b), while both S4D-marcksb and S4N-marcksb overexpressed embryos showed decreased expression of szl (Fig 3D-c, d and 3E). These data suggest that both types of mutated Marcksb caused a dominant negative effect on regulating the BMP signaling activity.
To establish a sensitive way to examine the effects of marcksb-overexpression, we overexpressed marcksb in chd_MO injected embryos (chd morphants) in which BMP signaling was slightly enhanced. As expected, all the chd morphants showed moderate ventralization (Fig 3F and 3G). Strikingly, injection of moderate dosage of marcksb mRNA resulted in severe ventralization in chd morphants, although injection of the same dosage of marcksb mRNA did not result in any visible phenotype in wildtype embryos (Fig 3F and 3G). This phenomenon was further proved by WISH analysis of szl in those embryos at shield stage (Fig 3H-a, b and 3I). Strikingly, the elevated BMP signaling activity in chd morphants was dramatically inhibited by overexpression of both types of mutated Marcksb (Fig 3H-a, c, d and 3I). Thus, our data suggest that the phosphorylation and de-phosphorylation switch of Marcksb is tightly related to the activation of BMP signaling.
Next, we asked whether Marcksb regulates the BMP signaling through the BMP secretory pathway. We first constructed tagged Bmp2b by insertion of mCherry or Myc tag right after the pro-domain of ligand protein according to previous study [30]. The overexpression of both myc-bmp2b and mcherry-bmp2b caused similar ventralization defect (Fig 4A a-c). To further confirm that fusion of mCherry to the N-terminal of Bmp2b does not interfere its in vivo function, we used the mCherry-bmp2b to rescue the mutant of bmp2b (bmp2bta72a/ta72a). We did individual genotyping for embryos of bmp2bta72a/ta72a and mcherry-bmp2b injected bmp2bta72a/ta72a. We found that all the bmp2bta72a/ta72a were dorsalized (Fig 4-d), while injection of mcherry-bmp2b mRNA could either rescue the dorsalization of bmp2bta72a/ta72a or cause ventralization (Fig 4A e-f). These data demonstrate that the insertion of myc- or mCherry-tag dose not interfere the biological function of Bmp2b.
To test whether Myc-Bmp2b or mCherry-Bmp2b can be properly cleaved and secreted, we transfected the plasmids containing either myc-bmp2b or mcherry-bmp2b and collected the cells and growth medium for immuno-analysis. For Myc-bmp2b, we found that the precursor (49 KD) was enriched in the cell lysis while the matured Myc-bmp2b (15 KD) in the medium (Fig 4B). For mCherry-Bmp2b, we transfected the cultured cells with plasmid containing mCherry alone as a control. Similarly, the precursor of mCherry-Bmp2b (74 KD) was mainly observed in the cell lysis while the matured mCherry-Bmp2b (41 KD) in the medium (Fig 4C). These data demonstrate that the insertion of Myc or mCherry does not interfere the proper cleavage and secretion of Bmp2b.
To investigate whether Marcksb regulates the secretion of Bmp2b, we then performed mosaic injection assay (Fig 4D-a). In the mcherry-bmp2b overexpressed embryos, the mCherry-Bmp2b could be detected outside the overexpressed-cells (Fig 4D-b and 4E). Strikingly, in the marcksb morphants, the level of mCherry-Bmp2b outside their producing cells was significantly decreased (Fig 4D-c and 4E). To further confirm that the above extracellular signal was from mature Bmp2b-mCherry, we detected the embryonic and extracellular mCherry-Bmp2b using immunoblotting. We revealed that there was mainly the precursor of mCherry-Bmp2b in the embryonic cells of wildtype or marcksb morphants and there were only properly cleaved matured mCherry-Bmp2b fusion proteins in the extracellular space, and the extracellular mCherry-Bmp2b was less in the marcksb morphants than that in wildtype embryos (Fig 4F). Moreover, it appeared that the total cleaved mature mCherry-Bmp2b of marcksb morphants was less and the precursor of mCherry-Bmp2b was more than that from wildtype embryos (Fig 4F). Thus, based on the above data, we conclude that marcksb is likely required for the intracellular trafficking and/or secretion of Bmp2b in which the cleavage of the Bmp2b precursor may be involved.
To investigate whether Marcksb regulated the secretion of Bmp2b in a cell-autonomous manner, we performed a transplantation assay. When the transplanted embryos developed to shield stage, we sorted the transplanted embryos of wildtype-to-wildtype into three groups according to the location of labeled descendants—ventral, lateral and dorsal (Fig 4G-a-d). Although the locations of labeled cells were different in those three groups, we did not observe any difference on the extracellular level of Bmp2b suggesting a similar capability of Bmp2b secretion from ventral to dorsal regions (Fig 4G-e-g). Subsequently, we transplanted the myc-bmp2b-overexpressed cells into the marcksb morphant host and found that they were capable of secreting Bmp2b in marcksb morphants as the myc-Bmp2b could be detected abundantly outside of the producing cells (62%, n = 21, Fig 4G-h). In contrast, the extracellular level of Bmp2b was significantly less in embryos with the marcksb-depleted cells transplanted to the wildtype host (100%, n = 21, Fig 4G-i). These data indicate that marcksb cell-autonomously regulates secretory pathway of Bmp2b.
To further unveil the role of marcksb on BMP signaling and dorsoventral patterning, we generated marcksb mutant by CRISPR/Cas9 mediated knockout (Fig 5A). After screening and verification by sequencing, we obtained two types of mutations–marcksbihb199/ihb199 (https://zfin.org/ZDB-ALT-180302-14) and marcksbihb200/ihb200 (https://zfin.org/ZDB-ALT-180302-15), both of which were predicted to shift their opening reading frames. There were no differences between these two alleles in phenotype analysis in subsequent studies. Therefore, we only presented the results of marcksbihb199/ihb199 in the following part. To our surprise, the homozygous zygotic mutants (Zmarcksb) did not show any early patterning defects and they could be raised up to adulthood, and we further generated maternal-zygotic mutant (MZmarcksb). WISH analysis showed that the expression of marcksb was dramatically decreased from 2-cell stage to shield stage in MZmarcksb, indicating that both maternal deposition and zygotic expression of marcksb were severely reduced in MZmarcksb (Fig 5B). This might be due to the failure of ribosome binding to mutated marcksb mRNA in the MZmarcksb embryos [31]. Surprisingly, MZmarcksb did not show any visible dorsoventral defects. However, we observed that the distance between the leading edges of enveloping layer (EVL) and deep cell layer (DCL) was enlarged in MZmarcksb during epiboly (Fig 5C). At bud stage, some MZmarcksb embryos showed a yolk bulge phenotype. A yolk droplet could be squeezed out of the body in some of the MZmarcksb embryos (Fig 5D arrow). These results indicated that MZmarcksb does not have dorsoventral defects but has moderate epiboly defects probably due to mild disorder of F-actin assembly [32].
As MZmarcksb did not show any dorsoventral defects as marcksb morphants, we speculated that there was genetic compensation occurring in the MZmarcksb [33, 34]. To challenge the hypothesis, we first injected the marcksb_MO into the MZmarcksb. The injected MZmarcksb showed to be marcksb_MO resistant and has no obvious dorsalization defect (Fig 6A). WISH analysis also showed that the expression of BMP targets—szl and ved (Fig 6B) and dorsal and ventral ectoderm markers—otx2 (S2A and S2B Fig) and foxi1 (S2E and S2F Fig) did not show any obvious difference between the MZmarcksb embryos with or without marcksb_MO injection. All these indicate that MZmarcksb is a null mutant of marcksb which does not respond to marcksb_MO and the marcksb morphant phenotype in wildtype embryos is a specific effect.
However, when we carefully compared the expression of szl and ved in wildtype and MZmarcksb embryos, we found a slight increase of expression levels of szl and ved in the MZmarcksb embryos, suggesting an elevation of BMP signaling activity in MZmarcksb. To further confirm this finding, we detected and compared the nuclear localization of P-Smad1/5/9 in MZmarcksb and wildtype embryos at shield stage. Consistent with the WISH results, the intensity of P-Smad1/5/9 was significantly increased in MZmarcksb (Fig 6C and 6D). Additionally, we performed the live imaging of Bmp2b by mosaic injection of mcherry-bmp2b mRNA. We found higher amount of mCherry-Bmp2b outside their producing cells in MZmarcksb in comparison with that in wildtype embryos (Fig 6E and 6F). Finally, we found that the chd_MO injection only led to moderate ventralization phenotype (V1-V2) in wildtype embryos, but it resulted in very severe ventralization (V3-V4) in MZmarcksb embryos (Fig 6G). Consistently, WISH analysis showed that knockdown of chd caused more robust increase of szl and ved expression in MZmarcksb than those in the wildtype embryos (Fig 6H and 6I). Taken together, these data strongly suggest that genetic compensation occurred in the MZmarcksb embryos, and moreover, the BMP signaling activity was even “over-compensated”.
To better understand the compensation network in MZmarcksb, we carried out RNA-Seq analysis of the MZmarcksb mutant at shield stage (S1 Dataset). Consistent with WISH analysis of marcksb (Fig 5B), RNA-Seq data showed that the expression level of marcksb was significantly reduced in MZmarcksb (Table 1). We also found that bmp7a was up-regulated in MZmarcksb, which is consistent to our observation that BMP signaling activity was slightly enhanced in MZmarcksb embryos (Table 1), as bmp7a is a transcriptional target of the BMP signaling [35, 36]. To dig out the main compensation factors, we searched for the list of differentially expressed genes (S1 Dataset) and found that hsp70.3 was the second most up-regulated gene after hsp90aa1.2 on the list of up-regulated genes in MZmarcksb.
We then performed RT-qPCR analysis of all the MARCKS genes and the hsp70.3 in MZmarcksb, maternal mutant of marcksb (Mmarcksb), marcksb morphants and wildtype embryos. Interestingly, in MZmarcksb, all the other MARCKS members, marcksa, marcksl1a and marcksl1b were all significantly up-regulated at 1-cell stage (Fig 7A), but not at shield stage (Fig 7B). Moreover, hsp70.3 was up-regulated in MZmarcksb at shield stage but not 1-cell stage (Fig 7A and 7B). These data suggest a phenomenon of sequential genetic response by MARCKS family members and hsp70.3 to maternal-zygotic loss of marcksb from oogenesis to early embryogenesis. Interestingly, this phenomenon could also be seen in the Mmarcksb embryos, suggesting that the genetic responses is independent of zygotic activation of marcksb in Mmarcksb (Fig 7A and 7B). Unlike the upregulation of hsp70.3 in MZmarcksb or Mmarcksb embryos, the expression of hsp70.3 was significantly decreased in marcksb morphants. To further confirm whether this genetic compensation persists even after wildtype zygotic gene activation of marcksb, we knocked down marcksb in Mmarcksb embryos and found that those embryos did not show any dorsalization defect (S2C, S2D, S2G, S2H and S2I–S2L Fig). Together, these results suggest that the MARCKS family members and hsp70.3 were up-regulated sequentially from oogenesis to early embryogenesis to response to the genetic loss of marcksb, and these genetic responses appear to be independent of zygotic activation of marcksb.
We then asked whether the other three MARCKS family members or hsp70.3 could compensate the function of marcksb in the absence of functional Marcksb, we injected marcksa, marcksl1a, marcksl1b or hsp70.3 mRNAs individually into marcksb morphants and we found that all of them could partially rescue the dorsalization defect of marcksb morphants (Fig 7C). These results suggest that the other MARCKS family members have the potential to replace the role of the Marcksb in the MZmarcksb mutant and hsp70.3 may have some genetic interaction with MARCKS family genes in regulating BMP signaling activity.
As it was reported previously that MARCKS interacts with HSP70 to regulate mucin secretion in human airway epithelial cells [37], We performed in vitro co-IP analysis to test whether the zebrafish MARCKS also bind to Hsp70.3. The Hsp70.3 had the highest binding affinity to Marcksb and moderate binding affinity to Marcksl11a and Marcksl1b. However, the binding affinity between Hsp70.3 and Marcksa is rather weak (Fig 7D), which is in consistent to the relatively low rescue efficiency of marcksa-overexpression in marcksb morphants (Fig 7C).
To further address whether hsp70.3, marcksa, marcksl1a and marcksl1b over-compensate the BMP signaling activity in MZmarcksb embryos, we performed loss-of-function analysis of those genes in MZmarcksb. We found that the expression of szl was severely decreased in MZmarcksb injected with moderate dosage of hsp70_MO (previously published morpholinos against all three variant splicing isoforms [38]) or a combination of morpholinos against marcksa, marcksl1a and marcksl1b (previously published morpholinos, for abbreviation, a_l1a_l1b_MOs [39, 40]) (Fig 7E and 7F), while the same dosage of hsp70_MO or a_l1a_l1b_MOs only led to slightly decreased szl expression in wildtype embryos (S3 Fig). To further verify the compensatory role of Hsp70.3 and other MARCKS members in MZmarcksb, we performed the experiments with CRISPR/Cas9 knockout method using the gRNAs against hsp70, marcksa, marcksl1a and marcksl1b. All the gRNAs were validated by sequencing of the target sites (S4 Fig). We found that the expressions of szl and ved were severely decreased in MZmarcksb embryos injected with either hsp70_gRNA or a mixer of MARCKS gRNAs (S5 Fig), which was similar to the observations from their MOs mediated knockdown in MZmarcksb. All these data revealed that hsp70.3, marcksa, marcksl1a and marcksl1b over-compensated the BMP signaling activity in MZmarcksb embryos.
We then performed BMP imaging in MZmarcksb using mCherry-fused Bmp2b as reporter. Although we previously observed a higher level of extracellular Bmp2b in the MZmarcksb than that in the wildtype embryos (Fig 6E and 6F), knockdown of either hsp70 or a combination of marcksa, marcksl1a and marcksl1b remarkably reduced the secreted Bmp2b level in MZmarcksb (Fig 7G and 7H). These lines of evidence demonstrated that the genetic over-compensation was due to the cooperation between the other members of MARCKS family and the molecular chaperone–Hsp70.3, which might even lead to mildly enhanced Bmp2b secretion level and BMP signaling activity in MZmarcksb embryos.
To investigate whether Marcksb interacted with Hsp70.3 to regulate the secretory pathway of Bmp2b in wildtype embryos, we performed a series of genetic interaction experiments. Firstly, we knocked down hsp70 by injection of full dosage of hsp70_MO. We found that knockdown of hsp70 led to inhibition of BMP signaling activity shown by decreased expression of szl and ved, which could be partially rescued by morpholino-resistant mRNA injection (S6 Fig). In the embryos co-injected with sub-dosage of marcksb_MO and hsp70_MO, a series of criteria were performed for careful evaluation: spindle shape of morphological defect was visible at early-somite stage (Fig 8A); the expression of BMP targets szl and ved was dramatically decreased (Fig 8B and 8C); the expression of neuronal dorsal marker otx2 was expanded to the ventral region (Fig 8D); the expression of epidermal marker foxi1 was decreased (Fig 8E). By contrast, in the embryos injected with either marcksb_MO or hsp70_MO alone did not show such defects (Fig 8A–8E). The Bmp2b live imaging was performed by transplantation of wildtype cells or cells injected with sub-dosage of hsp70_MO or marcksb_MO either alone or together into wildtype host. To our expectation, there were very few signals of the mCherry-Bmp2b outside the source cells in the hsp70 and marcksb double morphants, unlike that the mCherry-Bmp2b could be efficiently secreted from the source cells in the wildtype, or the embryos injected with sub-dosages of hsp70_MO or marcksb_MO (Fig 8F). All these data demonstrate that Marcksb interacts with Hsp70.3 to regulate the secretory process of Bmp2b in wildtype embryos, and BMP signaling activity is over-compensated in MZmarcksb embryos likely by mildly enhanced secretory pathway involving MARCKS family members and Hsp70.3 (Fig 9).
Bmp2b acts as a major morphogen to specify ventral cell fate during early embryogenesis. In this study, we found that the secretory pathway of Bmp2b requires a MARCKS family member-Marcksb and its interaction protein Hsp70.3. Interestingly, we revealed that a phenomenon of genetic over-compensation, which has seldomly described in previous studies, happened in the MZmarcksb, which was achieved by sequential up-regulation of the other MARCKS family members and Hsp70.3.
MARCKS is known to be involved in regulating secretion of many proteins in various cell types. The role of MARCKS in mucin secretion in the airway has been intensively studied [19, 22, 37, 41–46]. The translocation of MARCKS from the cell membrane to the cytoplasm upon phosphorylation by PKCδ is the initial step allowing MARCKS binding to the mucin granules [44], and this binding requires the interaction among translocated MARCKS, Hsp70 and Cysteine string protein (CSP) [22, 37]. After dephosphorylated by protein phosphatase I and 2A, MARCKS mediates the mucin granules binding to the myosin V and move along the cytoskeleton to the cell membrane [20]. In our study, the interaction between MARCKS and Hsp70 and the Phosphorylation of Marcksb both affect the extracellular level of Bmp2b, which indicates that MARCKS acts similarly to its role in mucin secretion in the intracellular trafficking and the secretion of Bmp2b.
The maturation of TGF-β superfamily ligands, such as BMPs, requires endoproteolytic cleavage of the prodomain from BMPs precursors (ProBMPs) which coincides with the intracellular trafficking process [47, 48]. Our data show that Bmp2b is properly cleaved before it being secreted to the extracellular space, as only properly cleaved Bmp2b is detected in the extracellular medium. The extracellular level of Bmp2b is much lower in marcksb-deficient embryo, indicating that marcksb is required for the secretory pathway of BMP ligands. Besides, we also noticed that the proBmp2b level was slightly increased and the cleaved Bmp2b level was slightly decreased in the embryonic lysis of marcksb-deficient embryo when compared with wildtype embryo. In consideration of the key role of MARCKS in intracellular trafficking system, we propose that proBmp2b would not traffic properly to the place where it is cleaved without the help of MARCKS. Moreover, the defective cleavage of proBMPs may further interfere dimerization, folding, and secretion of the active ligands [49, 50]. Therefore, it is possible that marcksb and hsp70 are required for one or several steps in the whole secretory pathway of BMPs, which mainly includes the intracellular trafficking along with endoproteolytic cleavage and the secretion to extracellular space.
Embryonic gastrulation includes dynamic events of cell migration and cell fate determination, both of which some molecules are involved in. One example is that the ventral to dorsal BMP signaling gradient transducing through Alk8 and Smad5 can create loose cell-cell adhesiveness at ventral region and allow ventral cells migrating dorsally [51]. This effect of BMP signal is different from its classical role in ventral cell fate determination and possibly is achieved by transcriptional activation of gene regulating cadherin function [51]. Our study provides another example on how one molecule could act on both morphogenesis and cell fate determination. It is widely accepted that MARCKS is required for gastrulation movements, which might be related to its binding with phosphoinositides [52] and F-actin [24]. Although the previous MARCKS knockdown studies in Xenopus and zebrafish mainly focused on its function on gastrulation movements [24, 25], they could not exclude the possibility that MARCKS family members also participate in embryonic patterning before or during gastrulation. In the present study, we also observed epiboly defects in both marcksb morphants and marcksb mutants, which is consistent to its classical role in regulation of cell migration. For the first time, however, we revealed that zebrafish marcksb is also required for dorsoventral patterning, and the function is achieved by interacting with Hsp70.3 to regulate the secretory process of BMPs, a type of morphogen crucial for ventral cell fate specification. Therefore, our study provides new insights into how a classical factor involved in cell migration also acts on cell fate determination.
In this study, we faced the genetic compensation responding to gene knockout which was reported recently [33]. Interestingly, the transcription of other MARCKS family members were activated during oogenesis in MZmarcksb females, probably driven by non-sense mRNA decay mechanism [53, 54], and Hsp70.3 –the MARCKS interaction protein was up-regulated at shield stage which was presumably driven by zygotic activation in MZmarcksb embryos, suggesting a sequential compensation of different genetic factors via different mechanisms. Knockdown of either hsp70.3 or a combination of marcksa, marcksl1a and marcksl1b can efficiently block the activity of BMP signaling and reduce the extracellular level of Bmp proteins in the MZmarcksb embryos, which indicates that both Hsp70 and other MARCKS proteins collaborate closely to respond to the genetic loss of marcksb. In our case, the genetic compensation raised both from genes with sequence homology, and from genes within the same functional network, which support the recently proposed working model [33].
Interestingly, MZmarcksb showed a higher level of secreted Bmp2b (Fig 6E and 6F) and was sensitive to the knockdown of Bmp2b antagonist Chordin (Fig 6G–6I), suggesting that the genetic compensation could even lead to elevated output of the overall products and mild enhancement of certain biological process. This phenomenon has never been demonstrated in previous studies. In addition, the detection of maternal expression of other MARCKS family members in MZmarcksb suggests that they may have switched from zygotic genes to maternal genes in the genetic adaption process during oogenesis.
The experiments involving zebrafish followed the Zebrafish Usage Guidelines of the China Zebrafish Resource Center (CZRC) and were performed under the approval of the Institutional Animal Care and Use Committee of the Institute of Hydrobiology, Chinese Academy of Sciences under protocol number IHB2014-006.
Embryos were obtained from the natural mating of zebrafish of the AB genetic background (from the China Zebrafish Resource Center, Wuhan, China; Web: http://zfish.cn) and maintained, raised, and staged as previously described [55].
For overexpression of proteins, short peptides tags, mCherry or EGFP was inserted in frame after amino acid 295 of Bmp2b according to a previous study [4, 30]. The tagged Bmp2b were inserted into the pCS2+ vector for mRNA synthesis. The constructs of S4N-marcksb and S4D-marcksb were generated by PCR of the construct of marcksb-HA with mutation on the primer pairs. The primer pair for S4N-marcksb were F: AACGGTTTCAACTTTAAGAAGAACGCCAAAAAAG and R: CAGCTTGAACGGCTTCTTAAAGTTGAATCG. The primer pair for S4D-marcksb were F: GACGGTTTCGACTTTAAGAAGGACGCCAAAAAAGAAG and R: CAGCTTGAACGGCTTCTTAAAGTCGAATCGCTTTTTG (mutated bases in the primer pairs were underlined). Capped mRNA was synthesized using the mMessage mMachine Kit (Ambion). The previously validated morpholino antisense oligonucleotides (MOs) targeting the following genes were used: marcksa [39], marcksb [25, 39], chordin [56], hsp70.3 [38], marcksl1a [40], marcksl1b [40]. mRNA and MOs were injected into the yolk at the one-cell stage or into one-cell at 32- to 64-cell stage for mosaic injection. Doses for RNAs and MOs were indicated in the text or figures.
The mutants of marcksb were generated using CRISPR/Cas9 mediated mutagenesis. The gRNA target for marcksb was designed by CRISPRscan [57]. Capped mRNA of zebrafish codon optimized Cas9 [58] and gRNAs of marcksb were synthesized by in vitro transcription using the mMESSAGE mMACHINE kit (Ambion). 500pg Cas9 mRNA and 50pg gRNAs were co-injected at one-cell stage for each embryo. The gRNA target sequence is as follows: 5’-GGAGCACAAATCTCCAAAAACGG-3’ (the PAM sequence is underlined). The target region was amplified using specific primers of marcksb (fwd: 5’-GCGTTGTATCTCGCATCTCAT-3’ and rev 5’-CACACCCCCTCATAACATCA-3’). The PCR products were subject to Sanger sequencing for direct evaluation of the targeting efficiency and identification of mutation [59].
The gRNA targets for hsp70 (hsp70.1, hsp70.2 and hsp70.3), marcksa, marcksl1a and marcksl1b were designed by CRISPRscan [57]. The gRNA target sequences for the above genes were as follows: hsp70: 5’-CCTTTAATCCTGAAGAGATTTCC-3’ marcksa: 5’-GGCACCGCACCAGCAGAGGATGG-3’; marcksl1a: 5’-GGAGAAGCAGTGGCAGCGGACGG-3’; marcksl1b: 5’-GGATCCCAGGCATCAAAGGGAGG-3’ (the PAM sequence is underlined). 500pg Cas9 mRNA and 50pg gRNAs were co-injected at one-cell stage for each embryo.
Digoxigenin-labeled antisense RNA probes were synthesized by in vitro transcription. Whole-mount in situ hybridization (WISH) was performed as described [3, 60].
For tail organizer transplantation assay, donor embryos were either injected with egfp mRNA or a combination of egfp mRNA and marcksb_MO at 1-cell stage. Donor embryos were them raised till the shield stage. Approximately 30 donor cells from ventral margin were transplanted to the animal pole of wildtype host embryos of sphere or dome stage as described [28]. Embryos were raised till 1 dpf for evaluation.
The Bmp2b secretion assay was performed either by mosaic injection or transplantation. For mosaic injection, 50 pg memGFP mRNA and 50 pg mCherry-bmp2b mRNA with or without 1 ng marcksb_MO were injected into one blastoderm cell of a 16-cell to 32-cell stage embryo. The injected embryos were raised till shield stage for confocal imaging.
For transplantation method, 50pg myc-bmp2b mRNA and 150 pg memGFP mRNA with or without 6 ng marcksb_MO were injected into the wildtype fertilized egg. Approximately 30 donor cells at the dome to sphere stage were randomly transplanted into wildtype or marcksb-morphant host embryos at the equivalent stage. The correspondent donors and hosts were indicated (Fig 4G-a). Transplanted embryos were screened at shield stage for position identification of donor cells. Embryos were fixed at 60%-epiboly for immunofluorescence staining.
Immunofluorescence was performed as described [61]. Generally, embryos were fixed in 4% Paraformaldehyde for overnight at 4 oC. Embryos were permeabilized by serial treatments with distilled water for 5 minutes at room temperature, cold acetone for 5 minutes at -20 oC, distilled water for 5 minutes at room temperature. For immunofluorescence of P-Smad1/5/9, all the steps before adding secondary antibody should be performed under 4 oC. Anti-Phospho-Smad1/5/9 (D5B10) Rabbit mAb (CST) was used at dilution 1:500. Anti-Myc (Santa Cruz) was used at dilution 1:500. Anti-rabbit Alexa Fluor 568 were used as secondary antibody (Molecular probes) at dilution 1:500. Embryos were counterstained with DAPI (5mg/ml in stock, 1:5000 diluted with PBS for working solution) for 1 hour. After immunofluorescence, the embryos were kept in 50% glycerol-50%PBS with 1mg/ml anti-fade reagent phenylenediamine (Sigma) avoid from light at 4 oC.
Confocal images were acquired using a laser-scanning confocal inverted microscope (SP8, Leica) with a LD C-Apo 40×/NA 1.1 water objective. Z-stacks were generated from images taken at 0.5 μm intervals, using the following settings (2048x2048 pixel, 400MHz). For detection of p-Smad1/5/9 signal, confocal images were acquired by the same scope using Lan-Apo 20x/NA 0.75 objective at zoom 0.75. Z-stacks were generated from images taken at 3 μm intervals, using the following settings (1024x1024 pixel, 400MHz). Embryos of shield to 60% epiboly stage were mounted in 0.5% low-melting agarose and positioned with animal pole to the bottom.
The Fiji software was used to quantify the average fluorescent intensity of P-Smad and the secreted Bmp2b protein [62].
For quantifying the P-Smad1/5/9 intensity, all the embryos were firstly orientated as dorsal region to the right. 8-bit image of each channel was transformed into 32-bit image. The threshold was made using default method and the background was set to NaN. A rectangle selection tool was used to select an area covering the whole embryo. The selected area was added to the ROI manager. The intensity from ventral to dorsal region of both P-Smad1/5/9 and DAPI were measure by plot profile function of Fiji. The data were then exported into Microsoft Excel and calculated. The ratio of P-Smad to DAPI from ventral region to dorsal region was plotted using GraphPad Prism 7.
For quantifying the secreted mCherry-Bmp2b or myc-Bmp2b, the 8-bit image was first transformed into 32-bit. The threshold was made using default method and the background was set to NaN. A polygon selection tool was used to select an area covering the outside of Bmp2b-source cell. A total selected area (Areatotal) and the area (Areathreshold) limited to the threshold were measured. The secreted Bmp2b (Bmp2bsecreted) was calculated by the formula: Bmp2bsecreted = Areathreshold / Areatotal. The data was plotted using GraphPad Prism 7 as scatterplots with median for small sample size studies [63].
Co-immunoprecipitation experiments were performed as described previously [64]. For immunoprecipitation assays, the cDNAs of marcksa, marcksb, marcksl1a and marcksl1b were cloned into pCS2+MTC (C-terminal multiple myc tag) and hsp70.3 was cloned into pCGN-HAM (N-terminal multiple HA tag) vectors. HEK293T cells were transiently transfected with the indicated constructs of interest using VigoFect (Vigorous Biotechnology, China) at dosage of 10 μg plasmid for cells covering about 70% surface of culture bottle (Nest, 100mm cell culture Dish).
For immunoblotting of intracellular and extracellular Bmp2b in cultured 293T cells, 10 μg of endotoxin-free pCS2-myc-bmp2b or pCS2-mcherry-bmp2b were transfected for cells covering 70% surface of culture bottle. After 8 hours, we replaced the growth medium (DMEM (high glucose, Biological Industries, 01-052-1ACS) with 10% FBS (Biological Industries, 04-001-1A)) with serum-free high-glucose DMEM and cultured cells for another 12 hours. One bottle of cells and growth medium were collected separately. Cells was lysed with 500μl RIPA buffer (50 mM Tris at pH 7.4, 150 mM NaCl, 1% NP-40, 0.5% deoxycholate, 1 mM NaF, 1 mM EDTA and protease inhibitors) at 4 oC. The protein concentration was measured by Enhanced BCA Protein Assay Kit (Beyotime Biotechnology, P0010). About 50 μg protein was loaded to a lane for immunoblotting. The growth medium was centrifuged at 300g for several minutes to precipitate the cells and the supernatant was collected and concentrated by Centrifugal ultrafiltration tube (Amicon Ultra UFC9001096 and UFC5010BK).
For immunoblotting of embryonic and extracellular Bmp2b in vivo, zebrafish embryos were either injected with 10 pg mCherry-bmp2b mRNA per embryo or co-injected with 10 pg mCherry-bmp2b mRNA and 6 ng marcksb_MO per embryo. Each of 300 embryos at shield stage were harvested and dissociated by pipetting in 350 μL calcium-free Ringer’s solution. The cells were collected by centrifugation at 300 g for several minutes. The cells were then lysed with RIPA, vortexed vigorously, added with 5xSDS loading buffer (Beyotime Biotechnology, P0015), incubated for 10 minutes at 95 oC and used for immunoblotting. 5 embryos were loaded for each lane. 300 μL supernatant was incubated with mouse anti-mCherry antibody (Abclonal, AE002) embedded Protein G beads (Life, Dynabeads protein G, 10003D) (10 μL antibody for 50 μL beads) overnight at 4 oC. After washing 3 times with PBS with 0.02% Tween-20, the beads were added with RIPA and 5xSDS loading buffer, incubated for 10 minutes at 95 oC and used for immunoblotting.
For immunoblotting, anti-Myc (Santa Cruz Biotechnology, 1:2000), anti-HA (Sigma-Aldrich, 1:5000), anti-mCherry (Abclonal, AE002) antibodies were used.
Two hundred Embryos of either wildtype or MZmarcksb at shield stage were divided into two groups as replicates. The RNA was extracted using Trizol according to the manufacturer’s manual. Then the RNA was purified using RNA purification kit (Tiangen, China). The RNA samples were quantified and integrity was assessed by the Agilent 2100 Bioanalyser. The RNA integrity Numbers (RIN) of all RNA samples were >8.0. The RNA libraries were prepared using the Illumina TruSeq RNA sample preparation kit v2. The amount of input RNA is 1 μg. The average final library size is 309 bp. Sequencing was performed on Illumina Miseq with read length of 150 bp paired-end (PE) at the Analysis and Testing Center of Institute of Hydrobiology, Chinese Academy of Sciences. Clean data were mapped to zebrafish reference genome GRCz10 Ensembl release 87 using HISAT2 with default parameters [65]. Cuffquant and Cuffnorm from Cufflinks software package were used to calculate the normalized gene expression level [66, 67]. The differential expression analysis was performed using DEseq2 [68]. The original RNA-seq data has been deposited to the BioProject with accession number PRJNA432757 (https://www.ncbi.nlm.nih.gov/bioproject/PRJNA432757).
Wildtype, MZmarcksb and Mmarcksb embryos at 1-cell stage and shield stage, and marcksb morphants at shield stage were collected for RNA extraction and reverse-transcription with about 60~70 embryos per sample and at least biological triplicate. The BioRad CFX Connect Real-Time System was used for transcript quantification. Samples were tested in technical triplicate for each gene, and resultant Cq values were averaged. Primer efficiencies and gene expression levels were calculated according to the previous study [69]. eef1a was selected as reference gene. Data were processed using 2-ΔΔCq method. All RT-qPCR gene-specific primers are listed in S1 Table.
|
10.1371/journal.pgen.1005478 | Reduced Crossover Interference and Increased ZMM-Independent Recombination in the Absence of Tel1/ATM | Meiotic recombination involves the repair of double-strand break (DSB) precursors as crossovers (COs) or noncrossovers (NCOs). The proper number and distribution of COs is critical for successful chromosome segregation and formation of viable gametes. In budding yeast the majority of COs occurs through a pathway dependent on the ZMM proteins (Zip2-Zip3-Zip4-Spo16, Msh4-Msh5, Mer3), which form foci at CO-committed sites. Here we show that the DNA-damage-response kinase Tel1/ATM limits ZMM-independent recombination. By whole-genome mapping of recombination products, we find that lack of Tel1 results in higher recombination and reduced CO interference. Yet the number of Zip3 foci in tel1Δ cells is similar to wild type, and these foci show normal interference. Analysis of recombination in a tel1Δ zip3Δ double mutant indicates that COs are less dependent on Zip3 in the absence of Tel1. Together these results reveal that in the absence of Tel1, a significant proportion of COs occurs through a non-ZMM-dependent pathway, contributing to a CO landscape with poor interference. We also see a significant change in the distribution of all detectable recombination products in the absence of Tel1, Sgs1, Zip3, or Msh4, providing evidence for altered DSB distribution. These results support the previous finding that DSB interference depends on Tel1, and further suggest an additional level of DSB interference created through local repression of DSBs around CO-designated sites.
| Meiosis is the type of cell division used by sexually reproducing organisms to create gametes (eggs and sperm, in animals). During meiosis, the two copies of each chromosome swap segments of DNA, forming reciprocal exchanges called crossovers. Crossovers are needed to help ensure that each gamete inherits a copy of every chromosome. Exchange occurs at deliberately induced double-strand DNA breaks, a subset of which become crossovers. In this study we investigate the role of the Tel1/ATM checkpoint kinase in modulating meiotic recombination in budding yeast. We find that in the absence of Tel1, recombination is increased, crossover distribution is altered, and crossovers are less dependent on the Zip3 protein, which mediates the major crossover pathway in yeast. We also find evidence which we infer indicates that Tel1, the helicase Sgs1, and the crossover-promoting proteins Zip3 and Msh4 influence how breaks are positioned throughout the genome. These results are consistent with a role for Tel1 in regulating the spacing of breaks along chromosomes. Our results also suggest that crossover-committed sites may suppress break formation in surrounding areas. Such a feedback mechanism would allow cells to achieve a sufficient number of crossovers without sustaining excess DNA breaks, which are inherently risky.
| Sexual reproduction depends on meiosis, a specialized type of cell division that produces haploid gametes from diploid cells. Recombination between homologous chromosomes is a key feature of the first meiotic division. A subset of recombination events creates reciprocal exchanges known as crossovers (COs), which help ensure that homologs segregate properly in meiosis I. Recombination also includes non-reciprocal events called noncrossovers (NCOs). The number and distribution of COs are highly regulated to ensure proper chromosome segregation. A striking feature of the CO landscape is the non-random spacing of COs, a phenomenon known as interference (reviewed in [1]). As a result of interference, COs tend to be relatively evenly spaced along chromosomes. Although interference was first reported over a century ago as the decreased probability that a CO would occur if another CO occurred nearby [2], its mechanistic underpinnings are still not well understood.
Both COs and NCOs arise from double-strand DNA breaks (DSBs) induced by the Spo11 enzyme [3]. How each DSB’s fate is determined is poorly understood, but several findings indicate that a decision is made prior to formation of stable strand invasion intermediates [4,5,6]. Formation of both COs and NCOs begins with resection of DSBs to expose 3’ single-stranded tails that can invade homologous duplex DNA (Fig 1A). At sites of future COs, initial strand invasion is followed by formation of stable intermediates known as single-end invasions and double Holliday junctions (dHJs) [4,6]. Normal timing and levels of these CO-specific intermediates require the ZMM proteins (Zip2-Zip3-Zip4-Spo16, Msh4-Msh5, Mer3) [5]. Upon pachytene exit, dHJ-containing intermediates are resolved to form COs. In contrast, NCOs appear prior to pachytene exit, without formation of stable intermediates, and without the need for ZMMs [4,5,6]. Thus COs and NCOs show distinct timing, intermediates, and genetic dependencies, but how the repair pathway is initially chosen at each DSB is unknown.
In budding yeast, a subset of COs is associated with cytologically observed foci known as synapsis-initiation complexes (SICs) [7,8]. SICs contain the ZMM proteins and appear to promote polymerization of the synaptonemal complex (SC). Multiple lines of evidence indicate that SICs form at CO-committed sites. [9,10,11,12]. SICs, like COs, show interference [9,13,14,15,16]. Strikingly, however, in certain deletion mutants the distribution of SICs (cytological interference) is normal even though CO interference as assessed genetically is defective (e.g. zip1Δ, msh4Δ, and sgs1Δ) [9]. Based on these findings a two-phase model for establishment of CO interference has been proposed (Fig 1B) [5,9]. First, DSBs are formed and designated as future sites of COs or NCOs, with SICs marking CO-committed sites. Second, these sites are processed into their respective products. According to this model zip1Δ, msh4Δ, and sgs1Δ cause defects in the implementation phase without disrupting the initial CO/NCO decision. SICs thus provide a readout of repair pathway choice.
Formation of SICs requires the presence of Spo11-induced DSBs [8,10]. SICs are seen in the processing-defective rad50S strain, in the recombination-defective dmc1Δ strain, and in haploid cells, indicating that normal DSB processing and interhomolog recombination are not required for SIC formation [7,8,17,18], thus prompting us to ask whether recombination pathway choice hinges on events immediately after break induction.
In mitotic cells, where the response to DSBs has been extensively characterized, the earliest known events after DSB formation are the binding and activation of proteins involved in the DNA damage response, including Mre11-Rad50-Xrs2 (MRX), Tel1, Mec1, and the 9-1-1 complex (Ddc1-Mec3-Rad17 in budding yeast) [19]. MRX and Tel1 are recruited to unresected DSBs, while Mec1 and 9-1-1 respond to single-stranded DNA (ssDNA). Since SICs are seen in the processing-defective rad50S mutant, we reasoned that Tel1, which responds to unprocessed DSBs, might play a role in SIC formation.
Tel1/ATM is known to control meiotic DSB levels. In mice, loss of ATM causes a dramatic increase in DSB frequency [20]. In flies, mutation of the ATM ortholog tefu causes an increase in foci of phosphorylated H2AV, suggesting an increase in meiotic DSBs [21]. Measurements of DSB frequency in tel1Δ yeast have given conflicting results, with three studies showing an increase [22,23,24] and two showing a decrease [25,26]. All but one of these studies relied on mutations that prevent DSB repair (rad50S or sae2Δ) to enhance detection of DSBs. These mutations may themselves influence the number and distribution of DSBs, confounding interpretation of the results. The one study that examined DSB levels in tel1Δ single mutants found a convincing increase in DSBs [23].
Tel1/ATM also influences the outcome of recombination. In mice, loss of ATM causes meiotic arrest due to unrepaired DSBs [27,28,29]. Infertility due to a failure to produce mature gametes is a feature of the human disease ataxia telangiectasia, suggesting that ATM is also required for meiotic DSB repair in humans. Meiotic progression in Atm−/− mice can be partially rescued by heterozygosity for Spo11 [30,31]. Compared to Spo11 +/− alone, Spo11 +/− Atm−/− spermatocytes show synapsis defects and higher levels of MLH1 foci, a cytological marker for COs [30]. In these spermatocytes the spacing of MLH1 foci is less regular and the sex chromosomes often fail to form a CO in spite of greater overall CO frequency. These results point to a role for ATM in regulating the distribution of COs. In yeast, examination of recombination intermediates at the HIS4LEU2 hotspot found that Tel1 is required for efficient resection of DSBs when the overall number of DSBs genome wide is low [32]. Under these conditions, the preference for using the homolog as a repair template was decreased in the absence of Tel1.
Tel1 also regulates DSB distribution (reviewed in [33]). In budding yeast DSBs are distributed non-uniformly throughout the genome, falling into large “hot” and “cold” domains spanning tens of kb, as well as smaller hotspots of a few hundred bp or less [3]. DSBs, like COs, are thought to show interference. Direct measurement of DSBs at closely spaced hotspots found that the frequency of double cuts on the same chromatid was lower than expected under a random distribution [23]. These calculations could only be done in repair-defective mutants due to detection issues, but nevertheless provide the most compelling evidence to date of DSB interference. This study found that DSB interference in yeast depends on TEL1. The existence of DSB interference was originally proposed based on the observation that introduction of a new hotspot greatly reduces DSB frequency in nearby areas [34,35,36,37]. It remains unknown whether this hotspot-hotspot competition and DSB interference represent the same phenomenon. A careful examination of recombination products at the HIS4LEU2 hotspot found evidence that DSBs also inhibit each other in trans, i.e. between chromatids, and that trans inhibition depends on Tel1 [24]. The authors proposed that spreading of trans inhibition along chromosomes could contribute to even spacing of DSBs.
Several proteins with key meiotic roles are subject to Tel1/Mec1-dependent phosphorylation, although in many cases the individual contribution of Tel1 (separate from Mec1) has not been tested. These include the axial protein Hop1, the Spo11 accessory factor Rec114, histone H2A, Sae2, and Zip1 [22,38,39,40]. Tel1-dependent phosphorylation of Rec114 may at least partially account for Tel1 regulation of DSB levels, although this has yet to be definitively tested [22]. Loss of Tel1 causes only a mild defect in spore viability and little or no delay in meiotic progression [39,41].
Multiple lines of evidence indicate that interactions between homologs influence DSB formation (reviewed in [42]). Experiments in worms first led to the proposal that nascent COs inhibit additional DSBs on the same chromosome [43,44]. This mechanism would allow DSB formation to continue until each chromosome has achieved a CO. Studies of worms, mice, and yeast indicate that some aspect of homolog engagement, possibly SC formation, leads to inhibition of DSBs [45,46,47,48]. High-resolution mapping of DSBs in synapsis-defective yeast found a change in the genome-wide distribution of DSBs in populations of cells [47]. To our knowledge, no previous studies have assessed whether regular spacing of DSBs along individual chromosomes is dependent on synapsis or other interhomolog interactions.
Our lab and others have developed techniques for mapping recombination products genome-wide in budding yeast [49,50,51,52]. We mate two yeast strains, S96 and YJM789, with sequence differences at about 65,000 sites. After recovery of the four progeny of a single meiosis, we use next-generation sequencing or microarrays to genotype progeny. The resulting map allows us to deduce the locations of all COs and nearly all NCOs with a median resolution of 81 bp.
Using this technique, we show here that loss of Tel1 causes an increase in recombination along with decreases in CO interference and the CO/NCO ratio. Yet the number of SICs in tel1Δ cells is similar to wild type, and these SICs show normal interference. These results suggest that in the absence of tel1Δ, a substantial number of COs arises from a ZMM-independent pathway. Our analysis of recombination in tel1Δ zip3Δ confirms this conclusion. Furthermore, we also see a change in the distribution of all recombination products in tel1Δ, zip3Δ, msh4Δ and sgs1Δ, which we infer indicates a change in DSB distribution. Since SIC distribution is normal in these strains (except zip3Δ, which lacks SICs) this result implies that DSB interference is not required for proper patterning of CO precursors. We argue that the opposite is true: the CO patterning process contributes to DSB interference, as CO-designated sites repress formation of additional DSBs in surrounding areas.
To investigate the role of Tel1 in meiotic recombination, we identified recombination products genome-wide in the progeny of 14 tel1Δ hybrid diploids. Eight tetrads were genotyped at high resolution by next-generation sequencing and used for analysis of both NCOs and COs, while six were genotyped at lower resolution and used for analysis of COs only. As wild-type controls, we used data from 46 tetrads genotyped by high-density tiling array [51] and six wild-type tetrads sequenced in our lab [53]. As expected based on analysis of recombination at a single hotspot [24], deletion of TEL1 significantly increases the overall rate of recombination (Fig 2A). This finding is also consistent with reports that DSB levels are increased in tel1Δ [22,23,24], although our data should be taken only as a rough estimate of DSB levels, since not all DSBs produce detectable products. In addition, there is potential for selection bias in our results since we are only able to assay cells that complete meiosis and produce viable spores. In the case of tel1Δ this bias is expected to be mild since the defects in sporulation and spore viability are quite modest (Fig 3E). We find that NCOs are increased disproportionately: the mean number of NCOs per tetrad increases by 60%, while COs increase by only 23%, resulting in a significantly lower CO/NCO ratio in tel1Δ compared to wild type (Fig 2B; p = 0.009; Student’s t-test).
The lower CO/NCO ratio in tel1Δ suggests that loss of Tel1 alters repair pathway choice. Another readout of pathway choice is CO interference, which refers to the relatively even spacing of COs in wild type. One way to assess interference is to analyze distances between adjacent COs (Fig 2C). In wild-type cells, inter-CO distances are well fit by a gamma distribution [50]. The value of the shape parameter γ of the best-fit distribution indicates the strength of interference, with γ > 1 indicating positive interference and γ = 1 indicating random distribution. γ is reduced from 2.0 in wild type to 1.6 in tel1Δ (Fig 2C), revealing a decrease in interference. Since γ is sensitive to changes in CO density, we also analyzed interference using the coefficient of coincidence (CoC) method in which the frequency of COs in two intervals is compared with the expected frequency of double COs under an assumption of no interference. Interference expressed as 1 –CoC also shows a significant decrease in tel1Δ (Fig 2D; p < 0.0001, chi-square test).
tel1Δ cells show a striking increase in complex products containing discontinuous gene conversion (GC) tracts or genotype changes on multiple chromatids (Fig 3A). To classify recombination products, we merge changes within 5 kb of each other into a single “event” that is assigned to one of eight event types (E1-E8). Except for E8, all of the types were previously defined [53]. 5 kb was chosen based on prior analysis of wild-type tetrads showing that events within 5 kb have distinct properties suggesting they arise from a single DSB [49]. All figures use a 5 kb threshold for merging unless otherwise specified. Results without merging are qualitatively similar and are shown in S1, S2B, S3, S6 and S8A Figs. In our classification system, “simple NCOs” (E1) and “simple COs” (E2) are products without any other genotype switches within 5 kb. A “CO with discontinuous GC” (E3) is a CO with a nearby GC tract (within 5 kb). A “discontinuous NCO” (E4) contains two GC tracts within 5 kb of each other. We also identify three categories of “minority” events (E5-E7), which are ambiguous products that could arise more than one way. For example, a minority event on three chromatids (E6) could be two closely spaced COs or a CO with a nearby NCO. In the current study we add a new category, E8, containing 4:0 tracts. These may represent cases of two overlapping NCOs, or may arise from pre-meiotic recombination. In wild type, complex events (categories E3-E7) account for about 14% of all meiotic recombination products; in tel1Δ they represent 22% of products, a statistically significant difference (p < 0.0001, Student’s t test). We see a similar increase in complex events in sgs1Δ ([53] and Fig 3A). The phenotypes of tel1Δ and sgs1Δ show several other similarities. Both mutants have higher recombination frequency, a decrease in the CO/NCO ratio, and a moderate decrease in CO interference [53,54,55,56]. These similarities suggested that Tel1 and Sgs1 might act together in regulating recombination.
Sgs1 is thought to control recombination pathway choice by unwinding nascent strand invasion intermediates unless they are protected by ZMMs [54]. Deletion of SGS1 rescues CO levels in zmm mutants [54,55]. We find that tel1Δ and sgs1Δ rescue crossing over in zip3Δ to similar extents (Fig 3B). However, in other ways tel1Δ and sgs1Δ show dramatically different phenotypes. First, loss of Zip3 causes a striking increase in NCOs. This increase is largely suppressed by sgs1Δ but not by tel1Δ (Fig 3B). Second, zip3Δ displays abnormally long GC tracts associated with COs (Fig 3C and [53]). This tract lengthening is suppressed by sgs1Δ [53] but only partially suppressed by tel1Δ. Third, a notable feature of recombination in sgs1Δ is the presence of a population of very short NCOs that we propose arise from aberrant SDSA [53]. This cohort of short NCOs is not seen in tel1Δ (Fig 3D). Together these results indicate that Sgs1 and Tel1 have distinct roles in regulating recombination.
To determine whether Tel1 acts upstream or downstream of SIC formation we measured the number and positions of Zip3 foci on chromosome IV or on all chromosomes in pachytene spreads of wild type and tel1Δ (Fig 4A). We find that tel1Δ cells show no increase in Zip3 foci compared to wild type in spite of greater numbers of COs and DSBs (Fig 4B and 4C). Since the number of foci in tel1Δ could be underestimated if foci are less intense and thus more difficult to detect, we determined whether the intensity of foci is similar in wild type and tel1Δ. By mixing both strains on a single slide, we control for slide-to-slide variation in staining. The two strains were labeled with arrays of tet operators on chromosomes of dramatically different size, allowing the genotype of individual cells to be identified after imaging. We find that the intensity of Zip3 foci in tel1Δ is slightly higher than in wild type (Fig 4D), indicating that the lack of increase in focus abundance is not caused by detection problems. Detection of Zip3 foci could also be impaired if foci are closer together in tel1Δ, causing adjacent foci to appear as a single merged focus. However, we find that the median distance between pairs of adjacent foci is similar in the two strains (0.42 μm in wild type vs. 0.44 μm in tel1Δ, a difference that is not statistically significant (S4A Fig)). We would also expect an increase in focus size if many more adjacent foci were unresolvable in tel1. This is not the case since the size of individual foci is the same in the two strains (S4B Fig). Together these results indicate that tel1Δ does not cause an increase in Zip3 foci. Zip3 foci in tel1Δ also show normal interference as determined by CoC analysis (Fig 4E).
SC length has been shown to correlate with the number of cytologically distinguishable CO-committed sites in worms and mammals [57,58] and not necessarily with the total CO number [59,60,61]. We find that the mean length of chromosome IV SC is 6% shorter in tel1Δ than in wild type (Fig 4F; p = 0.0004, Student’s t test). Thus in yeast, SC length parallels the number of SICs and not the overall number of COs.
The lack of increase in SIC abundance in tel1Δ is unexpected because three previously tested mutants with higher levels of COs (sgs1Δ, pch2Δ, and ndj1) had more SICs, while mutants with fewer COs (msh4Δ and zip1Δ) had fewer SICs [9,17]. By comparing the number of COs on chromosome IV in our recombination mapping experiments with the number of Zip3 foci on chromosome IV, we calculated a ratio of SICs to COs (Fig 5A). This ratio should be viewed as a rough estimate, since the measurements of SICs and COs were performed in different strains (isogenic and hybrid diploids, respectively). In wild type, the SIC/CO ratio is 0.63, implying that the majority of COs occur at SICs. In tel1Δ this ratio is reduced to 0.40, suggesting that non-SIC-associated COs are the major class. For comparison we determined the SIC/CO ratio in two other mutants with increased CO levels, sgs1Δ and ndj1Δ. For this analysis we compared the number of Zip2 foci on chromosome XV with the number of COs on that chromosome, both from published studies [9,50]. We find no significant change in the SIC/CO ratio in these mutants compared to wild type (Fig 5B). These results reveal a specific role for Tel1 in regulating the fraction of SIC-associated COs.
We considered the possibility that the failure of tel1Δ cells to make more Zip3 foci than wild type might be caused by DSB processing defects. A role for Tel1 in resection of meiotic DSBs has been suggested [32,39,62] Yet high levels of Zip3 foci are seen in the resection-defective rad50S strain (Fig 5C and [7]). These results indicate that resected ends are not required for formation of SICs.
Non-ZMM associated COs, often called Class II COs, are assumed to lack interference [63,64,65]. A possible reason for decreased CO interference in tel1Δ is that non-ZMM-associated COs, which represent a minority of events in wild-type cells, make up a larger share of events in tel1Δ. To further test this we compared the effect of deleting ZIP3 on CO abundance in wild type and tel1Δ (Fig 5D). To adjust for different DSB frequencies, we normalized CO numbers by expressing them as a percent of all interhomolog events. The percent of events resolved as COs drops from 72% in wild type to 39% in zip3Δ. As predicted, the decrease in COs between tel1Δ (67%) and tel1Δ zip3Δ (49%) is more modest. Thus COs in tel1Δ show less ZMM dependence than in wild type. An even more dramatic decrease in ZMM dependence is seen in sgs1Δ: CO frequency is similar in sgs1Δ (67%) and sgs1Δ zip3Δ (61%). We conclude that in tel1Δ, SICs are still at least partially functional in terms of promoting the CO fate, since loss of Zip3 in tel1Δ causes a decrease in COs. The opposite is true in sgs1Δ: SICs are either not fully functional or not functionally relevant in terms of promoting COs, since very little effect was seen upon deleting ZIP3.
In cells lacking the SC central element Zip1, synapsis is lost and axes are held together at a few sites per chromosome, termed axial associations. The exact nature of these links is unknown, but they are thought to correspond to SIC-marked sites [8]. In the zip1Δ sgs1Δ double mutant, axes are held closely together by a dramatic increase in the number of axial associations, a phenomenon referred to as pseudosynapsis [56]. Given the similar numbers of recombination products in tel1Δ and sgs1Δ (Fig 3A), we tested whether pseudosynapsis also occurs in zip1Δ tel1Δ. We find strikingly distinct phenotypes in zip1Δ sgs1Δ and zip1Δ tel1Δ (Fig 5E). In zip1Δ sgs1Δ, virtually no regions of axial separation are seen, whereas many sites of axis separation are visible in zip1Δ tel1Δ, similar to zip1Δ alone. This is consistent with the finding that SICs are increased in sgs1Δ but not in tel1Δ, and supports the idea that axial associations occur at SICs. Alternatively, the close association of axes in zip1Δ sgs1Δ may arise from aberrant structures, such as trapped recombination intermediates, found only in zip1Δ sgs1Δ and not in zip1Δ tel1Δ.
To test whether Tel1 mediates DSB interference we examined the distribution of all recombination products in our tel1Δ tetrads, using all interhomolog events as a proxy for DSBs. A potential concern relating to this analysis is that we are unable to detect some recombination events. These include intersister events, estimated to arise from 15–30% of all DSBs [66], and NCOs falling between markers or in which mismatch repair restored the original genotype, together estimated to include 30% of interhomolog NCOs [51]. However, failure to detect a percentage of the DSB population per se should not affect the calculated strength of interference since CoC does not vary significantly with event density [15], a fact that we verified by randomly removing events from a wild-type data set to simulate loss of detection (S7 Fig). The inability to detect some events would only be problematic if the undetected events were distributed non-uniformly throughout the genome. Previous analysis of the genome-wide distribution of COs and NCOs found good agreement between recombination frequencies in wild type and DSB frequencies in dmc1Δ [51], indicating that the distribution of detectable interhomolog events reflects the underlying DSB distribution.
We find that the distribution of all interhomolog events in wild type displays interference, and this interference is decreased (from 0.37 to 0.21) in tel1Δ (Fig 6A; p = 0.0007; chi-square test). We infer that Tel1 mediates DSB interference, in agreement with physical assays [23].
Unexpectedly, we find that the combination of all interhomolog products in zip3Δ, msh4Δ, and sgs1Δ also shows reduced interference (from 0.37 in wild type to 0.14, 0.11, and 0.21, respectively; p = 0.0003, 0.004, and 0.002 respectively). These results suggest that DSB interference is defective in these mutants. These three mutants are known to disrupt CO interference, but to our knowledge they have not been proposed to affect DSB-DSB spacing. Based on these results, we hypothesize that CO designation and/or formation of a SIC suppresses formation of DSBs nearby. Several previous studies point towards the existence of feedback between interhomolog interactions and DSB formation [43,44,45,46,47,48] and indicate that there is considerable temporal overlap between DSB and SIC formation [47,67,68]. We suggest that, beyond controlling the levels of DSBs, some aspect of CO designation also shapes the pattern of DSBs along individual chromosomes.
One potential question in interpreting these results is whether reduced interference among COs would automatically be expected to cause reduced interference among all detectable products, even without an underlying change in DSB interference. To test this we performed a simulation in which DSB interference was established entirely independently of CO interference. All DSB positions were first selected (with interference), and then CO positions were selected (with additional interference) from the DSBs, with the remaining DSBs becoming NCOs. We then randomly removed 20% of all events to simulate intersister repair, and 30% of the remaining NCOs to simulate loss of detection due to restoration and lack of markers. Results are shown for a wild-type level of CO interference with various levels of DSB interference (Fig 6B, left), and for the same conditions without CO interference (Fig 6B, right). These simulations illustrate several points. First, in the presence of CO interference, the strength of interference between all detectable recombination products is slightly higher than the true DSB interference among all four chromatids. This is due to preferential detection of COs (i.e., we detect essentially all COs, which strongly interfere, but we fail to detect some NCOs, which do not). Second, the level of interference between NCOs varies with the strength of DSB and CO interference. At low levels of DSB interference, selection of strongly interfering COs from an almost randomly spaced pool of DSBs results in NCOs that show negative interference, i.e. a tendency to cluster. At high levels of DSB interference, imposition of CO interference enhances the regular spacing of both COs and NCOs. In this model, to achieve a level of interference between all products equivalent to what is observed in wild type, it is necessary to impose strong DSB interference (1-CoC = 0.32). At this level of DSB interference, NCOs show strong interference. In contrast, NCOs in wild type do not show significant interference (Fig 6A). In wild type, interference for NCOs alone is 0.1, which does not differ significantly from no interference (p = 0.18). In addition, there are no statistically significant differences between wild type and any of the mutants in the strength of interference between NCOs. This lack of interference among NCOs lends support to the notion that DSB interference is at least partially driven by DSB suppression near COs. If DSB interference arose entirely independently of COs, we would expect NCOs to show interference.
Third, these simulations show that complete loss of CO interference only slightly reduces the interference among all detectable events (Fig 6B, compare left and right panels). This reduction is too small to account for the observed reductions in tel1Δ, zip3Δ, msh4Δ, and sgs1Δ.
It should be noted that in these simulations, DSB interference was applied to all four chromatids equally; i.e., a DSB on one chromatid suppressed DSBs equally along the same chromatid and along the three other chromatids, a situation that might not occur in vivo. We have separately simulated situations where DSB interference exclusively affects DSBs on the same chromatid or on the same pair of sister chromatids (S8B Fig). We found that it was not possible to recapitulate the observed strength of DSB interference among all four chromatids when the simulated DSB interference only affected DSBs on the same chromatid. Simulations in which DSB interference acted on a chromatid and its sister were capable of recapitulating the wild-type level of interference among all events on all chromatids, but this simulation again predicted much stronger interference among NCOs than is actually observed. In reality, DSB interference may arise from a combination of same-chromatid, intersister, and interhomolog effects, but our simulations suggest that none of these scenarios can account for the observation of very weak interference among NCOs if we assume DSB interference is entirely independent of CO designation. These results do not rule out that DSB interference may be partially created upstream of CO designation, but they suggest that such a mechanism does not solely account for the observed distribution of events.
A previous study of the HIS4LEU2 hotspot found many tetrads with multiple COs and/or GC tracts in both wild type and tel1Δ (20% and 36% of detectable recombination products, respectively) interpreted as arising from multiple DSBs [24]. To test whether the complex recombination events we observed in tel1Δ could be caused by closely spaced DSBs, we modeled a total loss of DSB interference by randomizing the positions of COs and GC tracts in our unmerged tel1Δ or wild-type data. GC tracts falling within the boundaries of a CO were not randomized since they are assumed to arise from the same DSB as the CO.
In the simulation, we incorporated the DSB landscape, such that the probability of an event falling in a particular area was determined by the frequency of DSBs in that region [69]. We then merged genotype changes within 5 kb into a single event and classified them as event types E1-E8. Zhang et al. [24] classified recombination products as T0, T1, or T2 based on the inferred number of initiating DSBs. We consider our event types E3-E8 as equivalent to T2 events (inferred to arise from two DSBs). Some of these event types could not be detected by Zhang et al. due to the limited number of markers available at HIS4LEU2. Surprisingly, we find that events inferred to arise from two DSBs occur more frequently in wild type than expected based on random chance (Fig 6C). If a specific mechanism existed to prevent these events, we would expect the opposite: these events should be more frequent in randomized data than in real tetrads. The high number of these events may reflect the fact that such events could arise from a single DSB; for example, three-chromatid events could result from two ends of a DSB invading different chromatids. Such multi-chromatid events were proposed to underlie the high number of complex products potentially arising from two DSBs in the sgs1-ΔC795 mutant [24]. Alternatively, DSBs in both wild type and tel1Δ might show negative interference, i.e. a tendency to cluster. If so, this effect would presumably operate only over short distances (less than 5 kb), since we see positive interference when genotype changes within 5 kb are treated as a single event (Fig 6A). In accordance with this, concerted formation of DSBs on the same chromatid within an approximately 8 kb range was observed in tel1Δ cells by a physical assay [23].
Due to the ambiguous origins of two- and three-chromatid events, we separately analyzed four-chromatid events (E7). We consider these more likely to be cases of more than one DSB occurring in trans (i.e. on different chromatids), since only a very aberrant recombination event could produce genotype switches on all four chromatids from a single DSB. We find that the frequency of four-chromatid events in wild type is significantly lower than the frequency expected due to random chance (Fig 6D; p = 0.0007; Student’s t test). In contrast, the frequency of these events in tel1Δ is statistically indistinguishable from the frequency expected due to random chance (Fig 6D; p = 0.78) These results support the conclusions of Zhang et al. that a Tel1-dependent mechanism suppresses the occurrence of more than one DSB per quartet of chromatids. As noted by Zhang et al. and Garcia et al. [23,24], trans inhibition could operate either between sister chromatids or between homologs. Our analysis of E7 products cannot distinguish between these two models, since we are unable to determine whether the initiating DSBs occurred on homologs or sisters. In theory, E8 products (4:0 tracts), which are increased in tel1Δ, may represent cases where DSBs occurred on both sisters. However, such products can also arise from premeiotic gene conversions. We find that the majority of E8 events have perfectly overlapping endpoints (i.e., gene conversion tracts beginning and ending at the same markers on both chromatids). Of the 4:0 tracts that are not part of a complex event, 72% (in wild type) or 74% (in tel1Δ) have perfect overlap. Such a high degree of overlap would not be expected if the majority of these events represented independent NCOs. Therefore we suspect that the tel1Δ-dependent increase in these events may arise from an increase in premeiotic recombination. Some, but not all, previous studies of recombination in vegetatively growing tel1Δ cells have found an increase [70,71,72].
Our simulations show that complex products arising from multiple DSBs are expected to occur more often in hot genome regions compared to cold regions (S8C and S8D Fig). This trend may explain the unusually high number of complex events seen by Zhang et al. at HIS4LEU2, an artificial hotspot with higher DSB frequency than natural hotspots.
Our data indicate that Tel1 is required for an early step in recombination pathway choice (Fig 7). In the absence of Tel1, the ratio of COs to NCOs, CO interference, and the dependence of COs on ZIP3 are all decreased, indicating that a greater proportion of recombination events occurs via non-ZMM-dependent mechanisms. The abundance of SICs is also similar to wild type, which is surprising given the higher levels of DSBs and COs in tel1Δ. Zhang et al [16] found modestly increased numbers of SICs in tel1Δ in the SK1 strain background, (11% increase on chromosome XV). Given the differences in strain backgrounds and chromosomes analyzed, these may represent essentially the same result. In SK1, the increase in SICs was smaller than the increase in DSBs (50% increase at HIS4LEU2 in a rad50S background) and COs (23% increase at HIS4LEU2) previously reported in SK1 [24]. Thus both studies point to the conclusion that the number of SICs per CO is reduced in tel1Δ.
Taken together, our results suggest two non-mutually-exclusive mechanisms for the modulation of recombination by Tel1. One possibility is that in tel1Δ there are two distinct populations of DSBs: a normal cohort of DSBs repaired as in wild type, and a population of “excess” DSBs repaired via non-ZMM-dependent pathway(s). Another model consistent with our results is that tel1Δ causes a general defect in commitment of DSBs to the ZMM-dependent CO pathway. The wild-type-like number of foci in tel1Δ may be the net result of a decrease in SIC-forming ability partially offset by an increase in the abundance of DSBs. If Tel1 does promote SIC formation, other factors must have functional overlap with Tel1 in this role, since SICs show normal abundance in tel1Δ. We speculate that Tel1 phosphorylation of ZMMs may promote their recruitment to specific DSBs. All of the ZMM proteins contain multiple SQ/TQ sites, the consensus sequence for Tel1/Mec1 phosphorylation. Mutation of the four SQ/TQ sites in Zip3 reduces its association with DSB hotspots and reduces CO frequency in some intervals, suggesting its ability to form a SIC is impaired [11]. However, zip3-4AQ causes only a mild decrease in COs and no loss of spore viability, indicating that other relevant Tel1 targets in addition to Zip3 must exist.
Our results confirm that interference among CO-committed sites is not defective in tel1Δ, as previously reported [16]; instead, poor CO interference arises from the fact that many COs in tel1Δ occur via a non-ZMM pathway. Our analysis of recombination outcomes in tel1Δ zip3Δ provides experimental evidence for the prediction that in mutants with higher levels of DSBs without an increase in SICs, “extra” DSBs would be channeled into ZMM-independent repair pathways [15].
In previous observations of Atm−/− Spo11 +/− mouse spermatocytes [30], MLH1 served as a cytological marker for CO positions. Loss of ATM caused a decrease in interference between MLH1 foci, whereas Zip3 foci in yeast show normal interference (this study and [16]). MLH1 foci are often assumed to mark all COs rather than only ZMM-associated COs [73], although this view is not universally accepted (for example, [74,75,76].) If the view that MLH1 foci mark all COs is correct, the decreased interference between MLH1 foci would be consistent with our genome-wide mapping of tel1Δ recombination products, which showed decreased overall CO interference. Alternatively, ATM may play distinct roles in CO patterning in mammals and yeast.
COs are often categorized as Class I (ZMM-dependent) or Class II (Mus81-Mms4 dependent), with only Class I COs participating fully in CO interference. In tel1Δ the additional non-ZMM COs may be typical Class II COs dependent on Mus81-Mms4, or may form by another mechanism. Regardless of the mechanism, due to not participating in ZMM-dependent CO patterning, they would be expected to show decreased CO interference. Class II COs are often described as “non-interfering”, but as noted by Zhang et al. this terminology is probably inaccurate [16]. Since all sites of recombination are influenced by DSB interference, even Class II COs are expected to show weak interference.
The distribution of all events in tel1Δ is consistent with a decrease in interference between DSBs. Effects of tel1Δ on DSB spacing have been previously reported [23,24], but it was not necessarily obvious that this would be detectable at the level of all recombination products genome wide. Garcia et al. observed a defect in DSB interference along single chromatids, but could not assay interference among all four chromatids in a homolog pair [23]. The genetic analysis by Zhang et al. observed trans inhibition among all four chromatids at a particular hotspot, but could not determine whether such inhibition extends laterally along chromosomes [24]. It is thus striking that a defect in interference among all recombination products is detectable in our data among all four chromatids and at distances of tens of kb. This supports the proposal of crosstalk between homologs in determining DSB positions [24].
The distribution of all events in zip3Δ and msh4Δ also implies a decrease in interference between DSBs. The inferred decrease in DSB interference in zip3Δ and msh4Δ suggests that CO designation and/or formation of a SIC suppresses formation of DSBs nearby (Fig 7a). Consistent with this model, recent analysis of the genome-wide DSB distribution in a population of zip3Δ cells found that regions with the greatest change in DSB frequency in zip3Δ were enriched for Zip3 binding in wild type [47]. This strongly suggests that the influence of Zip3 on DSBs is at least partially a local effect, and is not solely attributable to chromosome-wide or nucleus-wide effects such as altering the timing of synapsis. Importantly, this model explains why CO-NCO pairs show interference while NCO-NCO pairs do not [51]. One implication of this model is that earlier-forming DSBs would have a greater tendency to become CO-designated sites compared to later-forming DSBs. In support of this, Zip3 localization is reduced at hotspots believed to represent late-forming DSBs [11]. A prediction of the model is that any mutation causing changes in SIC distribution or defects in SIC formation will also cause changes in DSB distribution. This may explain a recent observation in hed1Δ dmc1Δ cells, which have a reduced number of SICs. In this mutant CO distribution is altered such that the difference in recombination rates between adjacent hot and cold regions is diminished [18]. This was interpreted as indicating a change in the distribution of DSBs, with cold regions sustaining more DSBs as a result of delayed pairing or synapsis. We suggest that decreased SIC formation may also contribute to this change in DSB distribution.
The defective DSB interference inferred to occur in sgs1Δ may also be mechanistically related to SICs. In the absence of Sgs1, SICs form but appear to be uncoupled from sites of COs. This conclusion is based on the fact that SICs in sgs1Δ show normal interference while COs do not (Fig 6A and [9,55]) and that loss of ZMMs in sgs1 mutants does not significantly diminish CO frequency (Fig 3B and [53,54,55,56]). We speculate that the CO-promoting function of SICs and their putative DSB interference function are both impaired by lack of Sgs1.
How might a CO-designated site suppress nearby DSBs? Several studies have proposed that SC formation, which proceeds from SICs, inhibits DSBs [45,46,47,48]. Axial proteins including the Spo11 accessory complex Rec114-Mei4-Mer2 and HORMAD proteins are excluded from synapsed regions, suggesting mechanisms by which synapsed chromosomes could become refractory to DSB formation [22,48]. Alternatively, an inhibitory signal other than synapsis, such as modification of axial proteins, might spread from CO-designated sites. We note that in yeast, the presence of a homolog is not strictly required for SIC formation [8]. This leaves open the possibility that ZMMs may influence the DSB landscape through mechanisms not involving interactions between homologs. Regardless of the exact molecular nature of the signaling events, such a mechanism would allow cells to create a sufficient number of COs to promote proper chromosome segregation without sustaining excess DSBs, which are inherently risky.
A key question raised by these results is whether Tel1 and ZMMs influence DSB distribution via distinct mechanisms. In our data, the inferred level of DSB interference in tel1Δ zip3Δ double mutants is lower than in either single mutant, implying action through different pathways, but the difference is not statistically significant, possibly due to the small size of the data sets. Another observation that suggests Tel1 and ZMMs control DSBs through different mechanisms is their behavior in sae2Δ or dmc1Δ backgrounds: the tel1Δ-dependent increase in DSBs persists in sae2Δ or dmc1Δ, while zmm-dependent increases do not [5,23,47]. However, the aforementioned ZMM experiments measured only DSB levels and not DSB interference [47], which may represent distinct phenomena. One piece of evidence that is difficult to reconcile with Tel1 controlling DSB interference independently of SICs is the fact that NCOs alone do not show a significant level of interference (Fig 6A). This suggests that if SIC-independent DSB interference exists, it is weak, at least when DSBs on all four chromatids are considered. However, some aspect of DSB interference may act only along a particular chromatid or pair of sisters, and such an effect might operate independently of SICs; this effect would be very difficult to detect in our data.
In spite of low inferred DSB interference, normal SIC interference is seen in tel1Δ, msh4Δ, and sgs1Δ [9]. This result implies that proper patterning of SICs does not require an orderly array of DSBs, and further suggests that DSB interference might not contribute significantly to CO interference in wild type. In tel1Δ, poor DSB interference apparently contributes to poor CO interference because many COs occur at non-SIC-marked sites. However, in wild type it is still unclear whether DSB interference plays a role in CO interference.
Previous studies indicated that wild-type cells limit the occurrence of DSBs on multiple chromatids at a particular hotspot and argued that Tel1 mediates this trans inhibition [23,24]. Whether such trans inhibition operates between homologs, sisters, or both has been controversial. Zhang et al. argued that trans inhibition most likely represented inhibition between homologs, whereas Garcia et al. suggested the opposite, based partly on re-analysis of Zhang et al.’s data. Our analysis of recombination products containing genotype switches on all four chromatids supports the existence of a mechanism limiting multiple DSBs per four chromatids. Since we are unable to determine which chromatids sustained the initiating DSBs, we cannot distinguish whether this one-per-quartet constraint arises from trans inhibition between homologs, between sisters, or both.
Our simulations of DSB distributions along chromosomes indicate that multi-DSB events are expected to be more frequent in hot regions compared to cold ones. As a corollary, changes in the frequency of multiple DSBs observed at HIS4LEU2 or any other locus in mutant strains may reflect a change in the relative hotness of the hotspot or a change in the overall DSB landscape, rather than loss of a specific regulatory mechanism limiting re-cutting. In light of this, experiments involving one or a few hotspots should be interpreted with caution, especially if performed in rad50S or sae2Δ strains in which DSBs are restricted to a more limited number of hotspots than in wild type [77].
Strain genotypes are listed in S1 Table. For recombination mapping, diploids were made by mating S96 and YJM789 haploids. All chromosome spreads were in the BR1919-19B background. Strain construction is described in Supporting Materials and Methods.
DNA was prepared for Illumina sequencing using a NextFlex kit (BIOO) with Illumina-compatible indices or as described [49] with 4-base or 8-base inline barcodes. Samples were sequenced in 50-base single-end runs on an Illumina Genome Analyzer or Illumina HiSeq 2000 or 2500 at the Vincent J. Coates Genomic Sequencing Laboratory (UC Berkeley) or the Center for Advanced Technology (UCSF). Genotype determination was performed essentially as described using the ReCombine package [49], but no insertions/deletions were genotyped. Briefly, after genotyping, CrossOver v6.3 was used to detect recombination products without merging close genotype switches. Products within 5 kb were then merged into a single event and sorted into one of seven categories as described [53], but with the addition of the new E8 category containing 4:0 tracts. Only the six wild-type tetrads sequenced in our lab were used to calculate the number of E8 products, since the number of E8s per tetrad was significantly different in the 46 wild-type tetrads genotyped by Mancera et al. Other event types did not show such differences. E8s were not used in any subsequent calculations, including calculations of “total events”, since we consider them likely to arise prior to meiosis.
Raw sequence data have been deposited in the NIH Sequence Read Archive under accession number SRP044001. Data for wild type, sgs1Δ, zip3Δ, msh4Δ, and four out of six sgs1Δ zip3Δ tetrads were previously deposited under accession numbers SRP028549 (wild type) and SRP041214 (all other strains). Additional processed data is deposited in Dryad Digital Repository (doi:10.5061/dryad.bj042).
Chromosome spreads were made as described [78]. Wild-type, tel1Δ, and rad50S cells were collected after 15–21 hours in 2% potassium acetate at 30°C. zip1Δ, zip1Δ sgs1Δ, and zip1Δ tel1Δ cells were collected after 19–21 hours. Antibody staining is described in Supporting Materials and Methods. Images were collected on a DeltaVision microscope (Applied Precision). SC lengths and Zip3 focus positions were measured using the 3D model module in Softworx (Applied Precision). To measure focus intensities, foci were found via the Threshold and Watershed functions in ImageJ. The total signal in each focus was measured by the AnalyzeParticles function in ImageJ.
Gamma distributions were fitted to inter-event distances [50]. For calculations of CoC, the genome was divided into 25 kb bins. The frequency of events in each bin was calculated, as well as the frequency with which any two pairs of bins on the same chromosome both contained events in the same tetrad. The expected frequency of such double events under a model of no interference is the product of the individual event frequencies in the two bins. The CoC for each pair of bins is the ratio of the observed frequency of double events to the expected frequency. This ratio was calculated for all bin pairs with a non-zero expected frequency, and results were averaged for all bin pairs separated by a given distance. In Fig 6 and S8 Fig, only the results for adjacent bin pairs are plotted. A chi-square test was used to compare expected and observed double COs. Measurements of cytological interference were performed essentially as above, but chromosome IV was divided into 0.1 μm bins.
For simulations in Fig 6B, a Python script was used to generate 1000 simulated tetrads for each set of conditions. The genome was divided into bins of 100 bp, and the number of DSBs in each tetrad was chosen from a normal distribution based on observed event frequencies in wild type. DSB positions were sequentially chosen, and a gamma hazard function was used to reduce the probability of DSBs in nearby bins after each DSB position was selected. After selection of DSB positions, CO positions were chosen by an analogous process, using a gamma hazard function to reduce the probability of COs at DSBs located in nearby bins. CO selection continued until 64% of DSBs had been selected as COs. After CO selection all remaining DSBs were considered NCOs. To simulate failure to detect some events, 20% of all events were randomly deleted, and then 30% of the remaining NCOs were randomly deleted. Interference between all simulated events (before deletion of “undetectable” products) is reported as “DSBs” in Fig 6B. For Fig 6B, all four chromatids were treated as a single entity; i.e., DSB interference was applied equally to all four chromatids. Simulations of same-chromatid-only or intersister-only DSB interference are in S8B Fig. Scripts used to simulate tetrads and calculate interference have been deposited in Dryad Digital Repository (doi:10.5061/dryad.bj042).
The DSB landscape was incorporated into randomized tetrads by using DSB frequencies measured by sequencing of Spo11-oligos [69]. The genome was divided into non-overlapping bins of 2 kb, and the DSB signals for all nucleotide positions in each bin were added together and used to set the probability of events occurring in that bin. For analysis of complex event frequency in S8C and S8D Fig, bins within 10 kb of a telomere were not used because they contain lower-than-expected numbers of complex events; this is because the number of possible events for merging (within 5 kb) is limited on one side by the chromosome end.
|
10.1371/journal.ppat.1002767 | Broad Spectrum Pro-Quorum-Sensing Molecules as Inhibitors of Virulence in Vibrios | Quorum sensing (QS) is a bacterial cell-cell communication process that relies on the production and detection of extracellular signal molecules called autoinducers. QS allows bacteria to perform collective activities. Vibrio cholerae, a pathogen that causes an acute disease, uses QS to repress virulence factor production and biofilm formation. Thus, molecules that activate QS in V. cholerae have the potential to control pathogenicity in this globally important bacterium. Using a whole-cell high-throughput screen, we identified eleven molecules that activate V. cholerae QS: eight molecules are receptor agonists and three molecules are antagonists of LuxO, the central NtrC-type response regulator that controls the global V. cholerae QS cascade. The LuxO inhibitors act by an uncompetitive mechanism by binding to the pre-formed LuxO-ATP complex to inhibit ATP hydrolysis. Genetic analyses suggest that the inhibitors bind in close proximity to the Walker B motif. The inhibitors display broad-spectrum capability in activation of QS in Vibrio species that employ LuxO. To the best of our knowledge, these are the first molecules identified that inhibit the ATPase activity of a NtrC-type response regulator. Our discovery supports the idea that exploiting pro-QS molecules is a promising strategy for the development of novel anti-infectives.
| The disease cholera, caused by the pathogenic bacterium Vibrio cholerae, is a major health concern in developing regions. In order to be virulent, V. cholerae must precisely control the timing of production of virulence factors. To do this, V. cholerae uses a cell-cell communication process called quorum sensing to regulate pathogenicity. In the current work, we identify and characterize new classes of small molecules that interfere with quorum-sensing-control of virulence in multiple Vibrio species. The molecules target the key quorum-sensing regulator LuxO. These molecules have the potential to be developed into new anti-infectives to combat infectious diseases of global importance.
| Quorum sensing (QS) is a process of bacterial cell-cell communication that relies on the production, release, detection, and response to extracellular signaling molecules called autoinducers. QS allows groups of bacteria to synchronously alter behavior in response to changes in the population density and species composition of the vicinal community. QS controls collective behaviors including bioluminescence, sporulation, virulence factor production, and biofilm formation (Reviewed in [1], [2]). Impairing virulence factor production or function has gained increasing attention as a method to control bacterial pathogenicity. The advantage of anti-virulence strategies over traditional antibiotics is presumed to be reduced pressure on bacteria to develop resistance [3]–[5]. Because QS controls virulence in many clinically relevant pathogens, disrupting QS is viewed as a promising possibility for this type of novel therapeutic development [6]–[8].
Many pathogenic Gram-negative bacteria use acylhomoserine lactones (HSLs) as QS autoinducers, which are detected by either cytoplasmic LuxR-type or membrane-bound LuxN-type receptors [9]. To date, efforts to interfere with HSL QS in Gram-negative bacteria have yielded several potent antagonists [10]–[15]. While these strategies are exciting, some globally important Gram-negative pathogens do not use HSLs as autoinducers. Thus, additional strategies that target non-HSL based QS systems are required. Here, we describe the identification and characterization of a set of small-molecule inhibitors that act on the non-HSL QS system of Vibrio cholerae by targeting two independent steps in the signal transduction pathway.
V. cholerae is the etiological agent of the disease cholera and its annual global burden is estimated to be several million cases [16]. V. cholerae produces and detects two QS autoinducer molecules called CAI-1 and AI-2. CAI-1 ((S)-3-hydroxytridecan-4-one) is produced by the CqsA synthase [17], [18] and AI-2 ((2S, 4S)-2-methyl-2,3,3,4-tetrahydroxytetrahydrofuran borate) is produced by the LuxS synthase [19], [20]. Detection of CAI-1 and AI-2 occurs through transmembrane receptors CqsS and LuxPQ, respectively [21], [22]. CqsS and LuxPQ are two-component proteins that possess both kinase and phosphatase activities (Figure 1 shows the CqsA/CqsS system). At low cell density (LCD), when the receptors are devoid of their respective ligands, their kinase activities predominate, resulting in the phosphorylation of the response regulator LuxO. LuxO∼P is the transcriptional activator of four genes encoding small regulatory RNAs (sRNAs), Qrr1-4 [23]. The Qrr sRNAs target the mRNAs encoding the quorum-sensing master transcriptional regulators AphA and HapR. At LCD, facilitated by the RNA chaperone Hfq, Qrr1-4 stabilize and destabilize the aphA and hapR mRNA transcripts, respectively [23]. Therefore, AphA protein is made while HapR protein is not (Figure 1). When autoinducer concentration increases above the threshold required for detection (which occurs at high cell density (HCD)), binding of the autoinducers to their cognate receptors switches the receptors from kinases to phosphatases (Figure 1). Phosphate flow through the signal transduction pathway is reversed, resulting in dephosphorylation and inactivation of LuxO. Therefore, at HCD, qrr1-4 are not transcribed, resulting in cessation of translation of aphA and derepression of translation of hapR. This QS circuitry ensures maximal AphA production at LCD and maximal HapR production at HCD. AphA and HapR each control the transcription of hundreds of downstream target genes [24], [25]. Hence, reciprocal gradients of AphA and HapR establish the QS LCD and HCD gene expression programs, respectively (Figure 1).
In pathogens that cause persistent infections, QS commonly activates virulence factor production at HCD. However, in V. cholerae, which causes an acute disease, HapR production at HCD represses genes important for biofilm formation and virulence factor production [22], [26]–[30]. This peculiar pattern of virulence gene regulation can be understood in terms of the disease caused by V. cholerae [31]. Following successful V. cholerae infection, the ensuing diarrhea washes huge numbers of bacteria from the human intestine into the environment. Thus, expression of genes for virulence and biofilm formation at LCD promotes infection, while repression of these genes by autoinducers at HCD promotes dissemination. Thus, molecules that activate QS have the potential to repress virulence in V. cholerae. Moreover, QS plays an essential role in virulence in other pathogenic vibrios including Vibrio parahaemolyticus, Vibrio alginolyticus, and Vibrio vulnificus [32]–[35]. The components of the QS circuits in these species are similar to those of V. cholerae. Therefore, QS-activating molecules identified for V. cholerae could be broadly useful for controlling diseases caused by other vibrios.
Here, we report the identification of a set of small molecules that activate the QS system of V. cholerae. We classify the QS-activating molecules as either QS receptor agonists or LuxO inhibitors. Because we have already reported analyses of QS receptor agonists, we focus here on the LuxO inhibitors. At LCD, LuxO∼P activates production of the Qrr sRNAs, which repress HapR; inhibitors of LuxO thus activate QS due to derepression of HapR. LuxO belongs to the NtrC protein family, σ54-binding transcriptional activators that rely on ATP hydrolysis to promote open complex formation [36]. The LuxO inhibitors identified here function uncompetitively to perturb LuxO ATPase activity. Genetic analysis of LuxO mutants that are insensitive to the inhibitors suggests that the inhibitors interact with a region adjacent to the ATP binding pocket. Finally, using a set of phenotypic assays, we show that the inhibitors broadly activate different vibrio QS circuits and, in turn, repress virulence factor production and reduce cytotoxicity. Because LuxO is conserved among vibrio QS circuits, the molecules we characterize here are capable of inhibiting HSL-based and non-HSL-based vibrio QS systems. Numerous NtrC-type proteins homologous to LuxO act in two-component signaling systems and their roles in controlling nitrogen metabolism, virulence, motility, and other important processes have been extensively studied (Reviewed in [37]). To the best of our knowledge, there exists no previous report of a chemical probe that modulates the activity of a NtrC-family response regulator.
We are interested in identifying small molecules that activate QS in V. cholerae, in order to induce the HCD state and thus repress virulence factor production. To do this, we developed a whole-cell high-throughput screen that relies on QS-dependent induction of bioluminescence (lux) in V. cholerae [22]. We exploited V. cholerae mutants genetically locked into the LCD state and carrying the lux operon from V. harveyi to screen for molecules that induce light production, indicating that they activate QS responses. We performed the screen on two different LCD mutants. The first mutant lacks the two autoinducer synthases, CqsA and LuxS. Therefore, both CqsS and LuxPQ QS receptors function as kinases and constitutively phosphorylate LuxO, resulting in transcription of the Qrr regulatory RNAs, and repression of translation of HapR (see INTRODUCTION). In the absence of HapR, there is no transcription of the heterologous lux operon, and thus, this strain is dark. The second strain carries the luxOD47E allele. This luxO mutation mimics LuxO∼P, rendering LuxO constitutively active [23], [38]. Therefore, HapR is repressed and the strain is dark. We anticipated identifying two classes of molecules that could induce light production: Class 1) Molecules that induce bioluminescence in the double synthase mutant but not in the luxOD47E mutant. These compounds are predicted to be QS receptor agonists; and Class 2) Molecules that induce bioluminescence in both the double synthase mutant and the luxOD47E mutant. Class 2 compounds likely target QS components that lie downstream of the receptors. We screened 90,000 molecules and identified eight Class 1 compounds and three Class 2 compounds (Figures 2A and 2B). The EC50 of Class 1 compounds are comparable to that of CAI-1 and generally lower than those of Class 2 compounds (Figure 2C). These differences support the idea that the two classes of molecules potentiate QS responses by distinct mechanisms. None of the compounds affected cell growth (Figure S1).
To determine which QS component each compound acts on, we first tested the eight Class 1 compounds against V. cholerae mutants that lack only the CqsS receptor or only the LuxPQ receptor. All eight Class 1 compounds induced light production in the ΔluxPQ strain but not the ΔcqsS strain; hence, these eight molecules function as CqsS agonists (Figure S2). Interestingly, none has structural homology to the native CAI-1 autoinducer [17], [18], [39], [40] (Figure 2A). The Class 1 molecules are currently being characterized and are not discussed further here.
The three Class 2 compounds that activate QS in both of the LCD screening strains likely act downstream of the QS receptors. These three compounds are structurally homologous (Figure 2A); therefore, they may function by an identical mechanism. Here, we focused on the compound displaying the highest potency (i.e., compound 11, Figures 2A and 2C). Class 2 compounds could potentially target one or more of the V. cholerae QS cytoplasmic components that function downstream of the receptors: LuxO, σ54, Hfq, and/or Qrr1-4. We reasoned that if these compounds interfere with LuxO or σ54, transcription of qrr1-4 would decrease in the presence of the inhibitors. By contrast, if the compounds target Hfq or act directly on Qrr1-4, they should not affect qrr1-4 transcription. GFP production from a qrr4-gfp transcriptional fusion decreased ∼3-fold when the luxOD47E strain was treated with compound 11 (Figure 2D). This result suggests that compound 11 targets either LuxO or σ54. If the target of compound 11 is σ54, transcription of other σ54-dependent genes should be affected when V. cholerae is treated with the compound. We examined transcription of the σ54-dependent gene vpsR [41] and found that it did not change significantly in the presence of compound 11 (data not shown). These results suggest that compound 11 targets LuxO.
The three identified Class 2 compounds share a 5-thio-6-azauracil core and only their side chains vary (Figure 2A). In addition, several 5-thio-6-azauracil analogs with other modifications on their side chains displayed weak or no activity in the screen. Therefore, differences in the hydrocarbon side chains must be responsible for the corresponding differences in potency with compounds harboring branched side chains displaying greater potency (i.e., compound 11, Figure 2C). To explore the relationship between structure and activity, we synthesized a focused library of compounds bearing the conserved 5-thio-6-azauracil core, and we altered the branching in the side chains. We measured activities using bioluminescence in the V. cholerae luxOD47E mutant. Several of the side chain modifications decreased potency (as shown by an increase in EC50, Figure 3). However, increasing steric bulk by incorporation of a tert-butyl carbinol side chain led to a 3-fold enhancement in potency (i.e., compound 12, Figure 3). Thus, the activity of the 5-thio-6-azauracil compounds within this series is highly sensitive to the structural features of the alkyl side chain. In the focused group of molecules we investigated, a bulky, hydrophobic terminal t-butyl moiety is optimal.
NtrC-type response regulators including LuxO possess three biochemical activities: phosphoryl-group accepting activity, DNA-binding activity, and ATP hydrolyzing activity [36]. We investigated which of these activities is inhibited by compounds 11 and 12. First, using whole-cell bioluminescence assays, we found that both compounds activate QS in V. cholerae strains expressing either wild type LuxO or LuxO D47E (Figures 2B and 3). Wild type LuxO is activated by phosphorylation via the QS cascade, and the LuxO D47E variant, which mimics LuxO∼P, while not phosphorylated is constitutively active [22], [23], [26], [38]. Because both wild type LuxO and LuxO D47E are vulnerable to inhibition, it cannot be the ability of LuxO to participate in phosphorylation or dephosphorylation that is impaired by compounds 11 and 12.
LuxO, as a NtrC-type response regulator, binds to σ54-dependent promoters to activate transcription. Compounds 11 and 12 could prevent LuxO from binding to DNA, and in so doing, prevent qrr transcription. To investigate this possibility, we used electrophoretic-mobility-shift and fluorescence anisotropy assays to probe the LuxO interaction with qrr promoter DNA. Even in the presence of a high concentration (200 µM) of the inhibitors, no significant change in LuxO D47E binding to qrr4 promoter DNA occurred as judged by mobility shift (Figure 4A). Quantitative fluorescence anisotropy assays revealed that, in the presence and absence of the LuxO inhibitors, LuxO D47E interacts with the qrr4 promoter DNA with an identical binding constant (∼300 nM) (Figure 4B). Thus, binding to DNA is not altered by the inhibitors.
Finally, we examined whether compounds 11 and 12 affect LuxO ATPase activity. To do this, we used a coupled-enzyme assay [42] to assess the rate of ATP hydrolysis by LuxO in the presence and absence of the compounds. Both compounds inhibit ATP hydrolysis in a dose-dependent manner (Figures 5A–C). Using traditional Michaelis-Menton enzyme kinetic analyses, we found that both compounds decrease the Km and the Vmax of the LuxO ATPase reaction (Figures 5B and 5C). The Lineweaver-Burk plots of curves derived from control reactions and from inhibitor-containing reactions display parallel slopes (Km/Vmax), indicating that compounds 11 and 12 function as uncompetitive inhibitors (Figures 5B and 5C), suggesting they bind to the pre-formed LuxO-ATP complex to inhibit ATP hydrolysis. Indeed, inhibition of LuxO ATPase by the analogs we identified or synthesized (as represented by % inhibition) is correlated with their potency (EC50) in inducing QS in the luxOD47E mutant (Figure 5D). We conclude that the LuxO inhibitors discovered here activate QS in V. cholerae by specifically inhibiting the ATPase activity of LuxO. Presumably, in the presence of the inhibitors, LuxO is incapable of participating in open complex formation at the qrr promoters, which prevents transcription of the Qrr sRNAs. In turn, translation of HapR is derepressed and the QS response occurs prematurely.
Compounds 11 and 12 likely bind to LuxO at an allosteric site that negatively regulates ATP hydrolysis activity. To determine where compounds 11 and 12 bind, we screened for LuxO mutants refractory to inhibition. To do this, we engineered random mutations into the cloned luxOD47E gene and introduced the mutant library into a V. cholerae ΔluxO strain carrying the lux operon. We screened for clones that conferred a dark phenotype in the presence of compound 12, hypothesizing that such mutants harbor alterations in the inhibitor binding-site. Four such mutants were identified (Figure 6A). These LuxO D47E variants all possess an active ATPase and are functional, as judged by their ability to repress light production in the absence of inhibitor (Figure 6A). Sequencing revealed that the four LuxO D47E mutants carry I211F, L215F, L242F, or V294L alterations, implicating these residues as important for binding of the inhibitors. We mapped these four alterations onto the existing crystal structure of ATP-bound Aquifex aeolicus NtrC1 (PDB:3M0E) [43], which has high sequence homology to LuxO (Figure 6B). The four residues we identified in the screen map to three regions that abut the Walker B motif (D245, E246, L247, and C248 in LuxO) (Figure 6B). In other NtrC-type proteins, mutations in this region have been shown to prevent ATP hydrolysis (See DISCUSSION). These four luxO mutations were introduced into wild type LuxO and the resulting mutants are similarly resistant to inhibition (Figure S3). Thus, binding of compounds 11 and 12 to this region may induce a conformational change in the nearby ATP-binding pocket that inhibits ATP hydrolysis.
As mentioned, LuxO is a conserved member of vibrio QS circuits. We therefore wondered if, similar to what we found in V. cholerae, compounds 11 and 12 could activate QS in other Vibrio species. To test this idea, we exploited two well-characterized phenotypes controlled by QS: light production in V. harveyi and colony opacity in Vibrio parahaemolyticus [44]–[46]. In V. harveyi, light production is induced by QS and a V. harveyi luxOD47E mutant is dark. Treatment of V. harveyi luxOD47E with compounds 11 and 12 induced light production 10,000-fold, indicating that these compounds are indeed active in V. harveyi (Figure 7A). In V. parahaemolyticus, the HapR ortholog, OpaR, controls colony opacity. OpaR production is repressed at LCD by LuxO∼P via the V. parahaemolyticus Qrr sRNAs. V. parahaemolyticus mutants that produce low and high levels of OpaR form translucent and opaque colonies, respectively [32], [46]. Thus, V. parahaemolyticus is naturally translucent at LCD and opaque at HCD. McCarter et al [32] recently identified a constitutively active LuxO mutant (LM4476, luxO*) in V. parahaemolyticus that confers a constitutive translucent colony morphology (Figure 7B, left). By contrast, an isogenic V. parahaemolyticus ΔluxO strain (LM9688) forms opaque colonies (Figure 7B, left). When the luxO* mutant is plated on medium containing compound 11 or compound 12, the colonies switch from translucent to opaque, a phenotype indistinguishable from the ΔluxO mutant (Figure 7B, right). These results suggest that compounds 11 and 12 inhibit V. parahaemolyticus LuxO from repressing the OpaR-dependent QS program. We conclude that the LuxO inhibitors identified in this study are broadly capable of activating QS in Vibrio species that employ LuxO as the central QS regulator.
In pathogenic vibrios, HapR and its homologs (e.g., V. parahaemolyticus OpaR and V. vulnificus SmcR) function as repressors of virulence factor production at HCD [32]–[34]. For example, in V. cholerae, the genes encoding the key V. cholerae virulence factors, the CTX toxin and the Toxin Co-regulated Pilus (TCP), are targets of HapR repression at HCD [17], [27], [30]. V. parahaemolyticus uses Type Three Secretion Systems (TTSS) for pathogenesis, and at HCD, OpaR represses the expression of one of the TTSS operons (TTSS-1) [32], [47]. Thus, luxO mutants that constitutively produce HapR (V. cholerae) or OpaR (V. parahaemolyticus) are attenuated in virulence [22], [30], [32]. The previous section shows that our LuxO inhibitors are active in multiple vibrios. To test whether the inhibitors can disrupt the QS-controlled virulence outputs of pathogenic vibrios, we assayed their effects on TcpA production in V. cholerae and production and secretion of VopD, a TTSS-1 effector protein, in V. parahaemolyticus. Western blot analysis showed that, in a V. cholerae luxOD47E strain, HapR and TcpA levels increased and decreased, respectively, in the presence of compound 12 (Figure 8A). Likewise, exposing the V. parahaemolyticus luxO* mutant to compound 12 resulted in decreased production and secretion of VopD (Figure 8B).
To begin to explore whether repression of these in vitro virulence phenotype translates to repression of the in vivo phenotype, we exploited an established V. parahaemolyticus cytotoxicity assay [48] to investigate whether pathogenicity could be inhibited by treatment with the LuxO inhibitors. We infected cultured HeLa cells with the untreated or compound 12-treated V. parahaemolyticus luxO* mutant and assayed HeLa cell lysis by measuring lactate dehydrogenase released from the host cytoplasm. Consistent with a previous report [32], the V. parahaemolyticus luxO* mutant is more cytotoxic to HeLa cells than the isogenic ΔluxO mutant (Figure 8C). At 2 to 3 hours post-infection, HeLa cell lysis was significantly lower in samples infected with the luxO* mutant treated with compound 12 than in samples infected with the luxO* mutant that had not been treated (average cytotoxicity is ∼30% and ∼100% for treated and untreated, respectively, p<0.01). At that time point, the cytotoxic capability of the Compound 12-treated luxO* mutant is slightly higher than that of the isogenic ΔluxO mutant (Figure 8C). At 4-hour post-infection, the compound 12-treated luxO* mutant was equally toxic (∼100%) as the untreated the luxO* mutant, while the ΔluxO mutant caused only ∼60% HeLa cells lysis. This residual cytotoxicity is consistent with earlier results showing that the ΔluxO mutant is not completely impaired for cytotoxicity [32]. Thus, the level of in vitro inhibition of TTSS-1 (Figure 8B) is a good indicator of the ex vivo inhibition of cytotoxicity (Figure 8C). The increase in cytotoxicity in Compound 12-treated V. parahaemolyticus that occurred at late time points could be due to incomplete inhibition of LuxO, uptake, or degradation of the compound by the HeLa cells. Nonetheless, the progression of V. parahaemolyticus killing of mammalian cells is impaired by compound 12, consistent with the notion that virulence factor production can be controlled by small molecule inhibitors of LuxO.
As part of a continuing effort to identify molecules that modulate QS in bacteria, we have identified two classes of molecules that activate QS in V. cholerae. These newly identified molecules serve two important purposes. First, they can be used as novel chemical probes to study QS signal transduction mechanisms. Second, from a practical standpoint, because QS represses virulence factor production in many pathogenic Vibrio species, molecules that activate QS, which decreases virulence, have the potential to be developed into anti-virulence agents to combat infectious diseases caused by pathogenic vibrios.
The first class of molecules identified here acts on the V. cholerae CqsS receptor. These molecules, surprisingly, do not resemble the native CAI-1 family of ligands (Figure 2A). Previous studies revealed that CqsS receptors from different vibrios possess distinct ligand detection specificities. The V. cholerae receptor is promiscuous in detecting a range of CAI-1-type molecules, while the V. harveyi receptor is relatively stringent [39]. Interestingly, none of the Class 1 molecules identified here activates QS in V. harveyi, lending support to the idea that CqsS receptors, although sharing extensive homology, possess different overall stringencies for ligands. We altered a single specificity-determining residue in the V. cholerae CqsS receptor (Cys 170) to the corresponding amino acid (Phe) in the V. harveyi receptor. This alteration is sufficient to increase stringency in detection of CAI-1 type molecules [39], [49], however, it did not abolish detection of the Class 1 molecules (Figure S4). Identification of CqsS receptor mutants with altered selectivity to the Class 1 molecules will provide additional insight into the molecular basis of ligand-CqsS interactions.
The second class of molecules identified, and the focus of this work, act on LuxO, the central QS regulator that controls transcription of the four Qrr sRNA genes. LuxO, which is a member of the NtrC family of two-component response regulators, possesses an N-terminal regulatory receiver domain, a central ATPase domain (AAA+ type), and a C-terminal DNA-binding domain. Three inhibitors have previously been identified that target non-NtrC type response regulators, AlgR1 of Pseudomonas aeruginosa [50], WalR in low-GC Gram-positive bacteria [51], and DevR in Mycobacterium tuberculosis [52]. The molecules function by perturbing phosphorylation (AlgR1 and WalR) and DNA binding (DevR). Our LuxO inhibitors, by contrast, function by an uncompetitive mechanism, presumably by binding to the pre-formed LuxO-ATP complex to prevent ATP hydrolysis. Thus, multiple families of response regulator can be selectively inhibited using small molecules. Furthermore, all three known response regulator activities; phosphorylation, DNA binding, and ATPase, are potential targets for inhibition. Analyses of LuxO inhibitor-resistant mutants suggest that our inhibitors bind to a region close to the predicted Walker B motif. Additional support for this idea comes from studies of other NtrC-type proteins, which show that mutations that affect ATP hydrolysis but do not interfere with ATP binding also map to the Walker B motif and to amino acid residues preceding the conserved GAFTGA domain [43], [53], [54]. Indeed, one of the LuxO inhibitor-resistant mutations identified here (L242F) lies immediately upstream of the predicted Walker B motif, while both the I211F and L215F mutations map to the helix containing the GAFTGA domain. In addition, the residue identified in the final inhibitor-resistant mutant, V294L, is predicted to sit facing the putative catalytic arginine (R306). The GAFTGA domain is important for interaction with the σ54-RNAP holoenzyme [55]. Thus, it was possible that the mutations we isolated in this region (I211F and L215F) suppress inhibition by compounds 11 and 12 by stabilizing the LuxO-σ54-RNAP interaction without affecting inhibitor binding. If this were the case, the ATPase activity of the purified LuxO D47E/I211F and D47E/L215F variants would be inhibited by these compounds. However, we purified LuxO D47E/I211F protein and found that the ATPase activity is not inhibited (Figure S5). This result is consistent with the idea that these mutations abolish inhibitor binding and, in so doing, prevent ATP hydrolysis.
High sequence conservation in the ATPase domain exists between different NtrC-type response regulator family members. Thus, we were interested to test whether the LuxO inhibitors could inhibit other NtrC-type response regulators. Compounds 11 and 12 only modestly inhibit (∼10%) the ATPase activity of purified E. coli NtrC at 250 µM (a concentration at which >80% of the LuxO ATPase activity is inhibited, Figure S6). This finding is surprising because the key residues (I211, L215, L242, and V294) that, when mutated, confer resistance to the inhibitors in LuxO are all present in E. coli NtrC. Thus, NtrC must possess additional structural features that render it resistant to inhibition. Structural comparisons between these two related RRs, coupled with identification of inhibitor-sensitive NtrC mutants, should allow us to understand the basis of the differences in inhibitor sensitivity.
Two-component signaling (TCS) proteins are widely distributed in bacteria. In addition to their global importance in microbial physiology, the absence of TCSs in mammalian cells makes them attractive drug targets in pathogenic bacteria [56], [57]. Even though significant effort has been devoted to identifying novel TCS inhibitors, to date, none has been developed into a new class of anti-infective. Problems such as undesirable properties associated with lead molecules have been encountered [56], [57]. In particular, inhibitors that generally target the conserved hydrophobic kinase domains of TCS histidine kinases suffer from drawbacks such as low cell permeability, poor selectivity, and unfavorable non-specific off-target effects (e.g. membrane damaging) [58]–[60]. By contrast, approaches to target the sensory domains of histidine kinases have yielded a handful of promising TCS inhibitors. For instance, LED209, an antagonist of the QseC histidine kinase, which regulates motility and pathogenicity in enterohaemorrhagic E. coli, reduces virulence in several pathogens both in vitro and in vivo [61]. In addition, in Staphylococcus aureus, inhibitory Agr peptide analogs antagonize the AgrC histidine kinase receptors and block abscess formation in an experimental murine model [62].
Targeting response regulators as a broad-spectrum anti-infective strategy has been considered challenging because response regulator functions, such as phosphorylation and DNA binding, are thought to be specific. In spite of this, a handful of molecules that inhibit particular response regulator functions have been reported [50]–[52]. For example, as mentioned, Walrycins, molecules that inhibit the phosphorylation of the essential WalR response regulator, are active in suppressing growth in multiple Gram-positive bacteria [51]. In the context of our work, the ATPase domain is highly conserved between all members of the NtrC response regulator family. Therefore, molecules that specifically target the ATPase domain of a single response regulator in this family (e.g., LuxO) could potentially be developed into general inhibitors of NtrC-family of proteins. Because NtrC-type proteins control virulence, nitrogen metabolism, motility, and other vital processes in bacteria [37], targeting the ATPase domain offers an additional route for anti-TCS drug development.
The LuxO inhibitors identified here possess certain favorable drug-like characteristics: potent inhibition, water-solubility, good stability, and cell-permeability. The molecules also display low host-cell cytotoxicity (undetectable cytotoxicity at 500 µM). These broadly-active LuxO inhibitors are not broad-spectrum NtrC-type inhibitors. Microarray analyses reveal that fewer than 40% of genes affected by the inhibitors are non-LuxO targets (data not shown). Nonetheless, our LuxO inhibitors could be used as preliminary scaffolds for building a general NtrC-type RRs inhibitor. Future improvements to these molecules will be focused on the structure-activity relationships of the thio-azauracil core, combined with simultaneously screening for molecules that inhibit LuxO and other NtrC type response regulators.
Although NtrC is not affected by the inhibitors discovered here, multiple LuxO response regulators from different Vibrio species are targeted by our inhibitors. Vibrio species detect a wide array of autoinducers (HSLs, CAI-1, and AI-2), thus, molecules that interrupt QS in Vibrio species by targeting the cognate receptors/synthases are likely to be autoinducer-specific and will have a limited spectrum. By contrast, because LuxO is nearly identical in all Vibrio species, our inhibitors can broadly activate vibrio QS irrespective of what type of autoinducer is detected. More importantly, we showed here that treatment of V. cholerae and V. parahaemolyticus with the LuxO inhibitors reduces virulence factor production and impedes cytotoxicity. Thus, our LuxO inhibitors, upon refinement, can at a minimum be used broadly to control virulence factor production in a variety of Vibrio species that use QS to repress pathogenesis.
The central ATPase module of the NtrC-type RR is classified as AAA+ type [63]. This module is present in multiple domains of life. For example, AAA+ ATPases are important in functions including protein unfolding and degradation (ClpXP, FtsH, and p97), organelle function and maintenance (PEX1 and VPS4), replication and recombination (RuvBL1 and helicases), and intracellular transport (Dyneins). Some eukaryotic AAA+ ATPases have been proposed to be drug targets [64]. Therefore, it will be particularly fascinating to investigate whether the thio-azauracil core discovered here can be developed into an inhibitor of AAA+ ATPases across different domains.
Antagonizing QS in bacteria represents a promising new approach that is an alternative to traditional antibiotics [8], [12], [14], [15], [61], [65]. Likewise, using pro-QS agents to treat acute infections, in which bacteria use QS to repress virulence, should be further explored. Using the native CAI-1 ligand, we previously showed that V. cholerae virulence factor production is repressed in vitro [17]. In the same vein, we show here that our synthetic pro-QS molecules reduce virulence by inhibiting LuxO. March et al reported that pretreatment with commensal E. coli over-producing the V. cholerae autoinducer CAI-1 increases the survival rate of mice following V. cholerae infection [66], which further supports the idea of QS potentiators as drugs. Use of CAI-1, LuxO inhibitors, or other QS-activating molecules as prophylactics could conceivably prevent V. cholerae or other pathogenic vibrios from initiating the LCD virulence gene expression program that is required for colonization. In this scenario, inhibiting the launch of virulence factors would provide sufficient time for the host immune system to eliminate the pathogen. In contrast to traditional antibiotics that target essential bacterial processes, growth is not affected by interfering with QS, so development of resistance could potentially be minimized [8], [14].
All V. cholerae strains are derivatives of wild type C6706str [67]. All V. harveyi strains are derivatives of wild type V. harveyi BB120 [68]. V. parahaemolyticus strains were generously provided by Dr. Linda McCarter. Escherichia coli S17-1 pir, DH5α, and Top10 were used for cloning. The relevant genotypes of all plasmids and strains are provided in Supporting Table S1. Unless specified, E. coli and V. cholerae were grown in LB medium at 37°C and 30°C with shaking, respectively. V. harveyi and V. parahaemolyticus were grown in LM medium at 30°C with shaking. Colony opacity of V. parahaemolyticus was monitored on LM with 2% agar. Unless specified, antibiotic concentrations are as follows: ampicillin, gentamicin, and kanamycin, 100 mg/L; chloramphenicol and tetracycline, 10 mg/L; streptomycin, 5 g/L; polymyxin B, 50 U/L.
The 90,000 molecule library was supplied by the High-Throughput Screening Resource Center of the Rockefeller University. The V. cholerae strains BH1578 (ΔcqsA ΔluxS pBB1) and BH1651 (luxOD47E pBB1) were grown overnight in LB medium with tetracycline and diluted 25-fold. The diluted cultures were dispensed into 384-well microtiter plates containing screening molecules that were previously added to each well. The final concentration of each compound was ∼20 µM. Light production was measured on an Envison Multilabel Reader after 6-hour incubation at 30°C without shaking. Compounds that induced light production >100-fold were reordered from suppliers and tested.
Overnight cultures of reporter strains were grown in LM medium (for V. harveyi) or LB with tetracycline (for V. cholerae carrying pBB1) and diluted 20-fold with sterile medium. Bioluminescence and OD600 were measured in an Envison Multilabel Reader following 4-hour incubation at 30°C with shaking. Synthetic molecules were dissolved in DMSO and supplied at varying concentrations to the reporter strains. DMSO was used as the negative control.
The open reading frame encoding V. cholerae LuxO D47E was amplified by PCR and cloned into plasmid pET28B that had been previously digested with NdeI and BamHI. The resulting plasmid was transformed into E. coli BL21 Gold (DE3) resulting in strain WN133. Strain WN133 was grown in LB with kanamycin at 30°C with shaking until the OD600 of the culture reached ∼1.0. IPTG was added at a final concentration of 200 µM, and the culture was incubated for an additional 4 hours at 30°C with shaking. Cells were harvested by centrifugation, suspended in lysis buffer (20 mM Sodium phosphate buffer pH 7.4, 0.5 M NaCl, 10% glycerol, and 5 mM imidazole), and lysed using a Cell Cracker. Soluble materials were loaded onto a HiTrap chelating column charged with nickel, the column was washed extensively with lysis buffer, and His6-tagged V. cholerae LuxO D47E enzyme was eluted using a linear gradient of increasing concentration of imidazole dissolved in lysis buffer. Fractions containing LuxO D47E were pooled and concentrated with an Amicon Untra-15 filter. Protein was snap-frozen in liquid nitrogen and stored at −80°C. Protein concentrations were determined by UV absorbance at 280 nm. E. coli NtrC and other LuxO D47E variants were purified using the same method.
A modified coupled-enzyme assay was used to measure the rate of ATP hydrolysis by LuxO D47E [42]. Briefly, ADP released from ATP by LuxO D47E is reacted with phosphoenolpyruvate (PEP) to form pyruvate using pyruvate kinase (PK). Pyruvate is reacted with NADH to form NAD and lactate using lactate dehydrogenase (LDH). The rate of NAD production is followed at 340 nm using a spectrophotometer. ATP hydrolysis rates were inferred from the absorbance change observed (εNADH,340−εNAD,340 = 6220 M−1 cm−1 for NADH) [42]. The rates of ATP hydrolysis by LuxO D47E were measured in reactions containing 100 mM Sodium phosphate buffer pH 7.4, 5 mM MgCl2, 0.2 mM NADH, 1 mM PEP, 5–20 units of PK/LDH mix (Sigma), and 10 µM LuxO D47E. ATP and inhibitors were added to the reactions at indicated concentrations. The rate of ATP hydrolysis was monitored for 5 minutes. Data were fitted using Graphpad Prism to obtain the kinetic parameters. Percent ATPase inhibition was calculated using the following formula:
Electrophoretic mobility shift assays to study LuxO and Qrr promoter DNA interactions were performed as described in [69]. Fluorescence anisotropy assays using LuxO D47E were modified from [70].
The luxOD47E allele was removed from plasmids harbored in WN133 with the enzymes XbaI and BamHI and ligated into pEVS143 [71] that had been previously digested with AvrII and BamHI. The luxOD47E reading frame of the resulting plasmid (WN2029) was randomly mutated using the GeneMorph II Random Mutagenesis Kit. The resulting mutagenized luxOD47E plasmid library was introduced into a V. cholerae ΔluxO strain by conjugation. Individual colonies from this V. cholerae luxOD47E mutant pool were arrayed into 96-well plates containing LB medium with 100 µM compound 12. The V. cholerae ΔluxO strain harboring non-mutated luxOD47E was grown in the absence of compound 12 to provide the reference for background light production. Following overnight static incubation at 30°C, clones that produced light comparable to the background were selected and re-tested in the presence and absence of compounds 11 and 12. DNA sequencing was used to determine the alterations in luxOD47E for inhibitor-resistant mutants. Site-directed mutageneses were performed with the QuikChange II XL Site-Directed Mutagenesis Kit to uncouple multiple mutations.
Overnight cultures of the V. cholerae luxOD47E strain were diluted 1000-fold in AKI medium containing the indicated concentrations of compound 12. The cultures were statically incubated at 37°C for 4 hours and subsequently shaken for 4 more hours at 37°C. Cells were collected by centrifugation, TcpA from different samples was analyzed by Western blot as previously described [17]. Overnight cultures of the V. parahaemolyticus luxO* strain (LM4476) were washed and diluted 50-fold in LM medium with 10 mM MgCl2 and 10 mM sodium oxalate in the presence of the indicated concentrations of compound 12. The cultures were grown for 4 hours with shaking at 37°C. Viable cell count showed that all cultures contained ∼1×109 CFU/mL after incubation. Cells were collected by centrifugation, and the secreted and cytoplasmic VopD from different samples were analyzed by Western blot as previously described [47].
Cytotoxicity assays were modified as previously described [48]. HeLa cells (2×104 cells/well) were cultured for 48 hours at 37°C and 5% CO2 in a 96-well plate containing DMEM with 10% fetal bovine serum prior to infection. V. parahaemolyticus strains were grown as described above for VopD analysis and used in the infection assays. Immediately prior to V. parahaemolyticus infection, DMSO or compound 12 (500 µM) was added to the HeLa. Serially diluted bacteria were added to HeLa cells at multiplicity of infection of 10. Lactate dehydrogenase release from HeLa cells was assayed between 1–4 hours after infection using the CytoTox 96 nonradioactive cytotoxicity kit (Promega).
All chemical syntheses and analytical methods are provided in the Supporting Text S1.
|
10.1371/journal.pntd.0001469 | Impact of Continuous Axenic Cultivation in Leishmania infantum Virulence | Experimental infections with visceral Leishmania spp. are frequently performed referring to stationary parasite cultures that are comprised of a mixture of metacyclic and non-metacyclic parasites often with little regard to time of culture and metacyclic purification. This may lead to misleading or irreproducible experimental data. It is known that the maintenance of Leishmania spp. in vitro results in a progressive loss of virulence that can be reverted by passage in a mammalian host. In the present study, we aimed to characterize the loss of virulence in culture comparing the in vitro and in vivo infection and immunological profile of L. infantum stationary promastigotes submitted to successive periods of in vitro cultivation. To evaluate the effect of axenic in vitro culture in parasite virulence, we submitted L. infantum promastigotes to 4, 21 or 31 successive in vitro passages. Our results demonstrated a rapid and significant loss of parasite virulence when parasites are sustained in axenic culture. Strikingly, the parasite capacity to modulate macrophage activation decreased significantly with the augmentation of the number of in vitro passages. We validated these in vitro observations using an experimental murine model of infection. A significant correlation was found between higher parasite burdens and lower number of in vitro passages in infected Balb/c mice. Furthermore, we have demonstrated that the virulence deficit caused by successive in vitro passages results from an inadequate capacity to differentiate into amastigote forms. In conclusion, our data demonstrated that the use of parasites with distinct periods of axenic in vitro culture induce distinct infection rates and immunological responses and correlated this phenotype with a rapid loss of promastigote differentiation capacity. These results highlight the need for a standard operating protocol (SOP) when studying Leishmania species.
| Protozoan of the genus Leishmania undergo several developmental transitions during its life cycle. Leishmania alternates between two morphologically distinct forms, promastigotes (insect stage) and amastigotes (vertebrate stage). Most of the available information about Leishmania spp. has been obtained from studying in vitro cultured promastigotes, an excellent experimental model for the different developmental stages present in the insect vector. Although promastigotes are grown in a controlled environment, the maintenance of long term culture results in loss of virulence, which can lead to a misinterpretation and often contradictory experimental results. It is then of great interest to unravel the defects arising from sustained axenic parasite culture in laboratory settings. The authors demonstrate a correlation between the maintenance of parasite culture with a growing defect of the promastigote form to differentiate in the mammalian amastigote form. This research provides a biological explanation for the loss of virulence due to sustained parasite culture and discusses the impact for all experimental work done with visceral Leishmania species.
| Protozoan parasites of the genus Leishmania undergo several developmental transitions during their life cycle. Ingestion of infected macrophages during a blood meal by the sandfly vector leads to the release of intracellular amastigotes into the vector's midgut. This abrupt change in environment induces the transformation into extracellular procyclic promastigotes. The procyclic form within the vector midgut replicates and ultimately transforms into virulent metacyclic promastigotes, in a complex process that encompass parasite migration towards the upper gut of sandfly vector [1]. In laboratory conditions, it is possible to achieve indefinite promastigote growth outside the sandfly using several established media; the procyclic forms correspond to promastigotes in exponential phase of growth that will eventually pass into a stationary phase, a fraction of these stationary parasites differentiates into the metacyclic form, with properties resembling those of sand fly promastigotes [2], [3]. Although stationary-phase promastigotes with undefined in vitro passages are commonly used without limitations, it has been demonstrated that continuous culture over time induces loss of virulence. In fact, long-term in vitro culture of promastigotes was one of the first empirical approaches to efficiently identify parasite virulence genes leading to the experimental development of attenuated strains [4]. Similarly, long-term in vitro growth of drug-resistant parasites was suggested to mediate a loss of the resistance phenotype [5]. This can be due to either loss of virulence factors induced by the lack of a survival pressure or due to disadvantageous adaptations to the media resulting in phenomena similar to clonal selection [6]. Either way, alterations in the physiology of the parasite that are induced by long-term growth in these media may lead to misinterpretation and contradictory results. Thus, one must carefully consider the influence of the different laboratorial factors in order to minimize these variables.
The current study is based upon the hypothesis that maintenance of Leishmania spp. in axenic in vitro culture results in a progressive loss of virulence quickly generating a significant bias towards the experimental data. We have compared the in vitro and in vivo infections and focused on the influence that axenic parasite growth and long-term maintenance can have on in vitro infections outcome. Our results demonstrate, for the first time, that the loss of virulence caused by the maintenance of axenic promastigotes in culture can be the result of a growing inability to differentiate into amastigote forms. Moreover, the induction of differentiation from promastigote to amastigote and then back to promastigote forms both in vitro and in vivo was capable to restore parasite virulence. Overall, our study demonstrated the need of a standard operating protocol (SOP) to study visceral Leishmania spp. highlighting the crucial importance for proper control of parasite cultures in studies focusing on the mammalian stage, such as drug development or vaccine trials.
Ten to twelve-week-old female Balb/c mice were obtained from Instituto de Biologia Molecular e Celular (IBMC; Porto, Portugal) animal facilities. Under laboratory conditions, the animals were maintained in sterile cabinets and allowed food and water ad libitum. Animal care and procedures were in accordance with institutional guidelines. All conducted experiments were done in accordance with the IBMC.INEB Animal Ethics Committee and the Portuguese Veterinary Director General guidelines. RS has an accreditation for animal research given from Portuguese Veterinary Direction (Ministerial Directive 1005/92). A cloned line of virulent L. infantum (MHOM/MA/67/ITMAP-263) was grown at 26°C in RPMI 1640 medium (Lonza, Swtzerland) supplemented with 10% heat-inactivated Fetal Bovine Serum - FBS (Lonza, Switzerland), 2 mM L-glutamine, 100 U/ml penicillin, 100 mg/ml streptomycin and 20 mM HEPES buffer. The MHOM/MA/67/ITMAP-263 clone (zymodeme MON-1) was originally isolated from the bone marrow of a human patient in Morocco and cloned by micromanipulation. In some experiments, a previously uncharacterized field attenuated L. infantum strain was used (species confirmed by pteridine reductase 1 sequencing and currently under ongoing characterization in our laboratory). To minimize the possibility of clonal bias, we have performed three independent recoveries of parasite from Balb/c mice for these experiments. All cultures were initiated at 106 parasites/ml and passed each 5 days. Promastigote to amastigote differentiation was achieved by culturing 107 stationary phase promastigotes/ml at 37°C in a cell free culture medium (MAA20) [7]. Amastigote to promastigote differentiation was performed by culturing 107 axenic amastigotes/ml in complete RPMI medium for 4 days at 27°C. In alternative, spleens of infected Balb/c mice were placed in similar culture conditions for 7 days.
Metacyclic promastigotes were purified from cultures with 3, 5 or 9 days or from 5-day cultures with 4, 21 and 31 (P4, P21 and P31) in vitro passages by Ficoll density gradient, as previously described [8]. Briefly, 6 ml of 40% Ficoll was overlaid by 6 ml of 10% Ficoll in RPMI base. Then, 6 ml of PBS containing 1.2×109 parasites was placed at the top of the Ficoll gradient. The step gradient was centrifuged for 10 minutes at 370 g at room temperature without brake. The metacyclics promastigotes were recovered from the layer between 0% and 10% Ficoll solution. Metacyclic promastigotes, identified by morphological criteria, i.e., short and slender with a long flagellum twice the body length using phase contrast on a Nikon Eclipse 80i.
Total RNA was isolated from cells with the Trizol® reagent (Invitrogen, Barcelona, Spain), according to the manufacturer's instructions. Briefly, parasites were washed with ice-cold phosphate-buffered saline (PBS), harvested and homogenized in 800 µl of Trizol by pipetting vigorously. After addition of 160 µl of chloroform, the samples were vortexed, incubated for 2 min at room temperature and centrifuged at 12.000 g, for 15 min, at 4°C. The aqueous phase containing RNA was transferred to a new tube and RNA precipitated with 400 µl of isopropanol for at least 10 min at room temperature. Following a 10 min centrifugation at 12.000 g, the pellet was washed with 1 ml of 75% ethanol and resuspended in 10 µl of 60°C heated RNase free water. The RNA concentration was determined by using a Nanodrop spectrophotometer (Wilmington, DE, USA) and quality was inspected for absence of RNA degradation or genomic DNA contamination, using the Experion RNA StdSens Chips in the ExperionTM automated microfluidic electrophoresis system (BioRad Hercules, CA, USA). RNA was stored at −80°C until use. RT was performed with equal amounts of total extracted RNA (1 µg) obtained from parasites recovered from different experimental conditions by using Superscript II RT (Gibco BRL) and random primers (Stratagene). Real-Time quantitative PCR (qPCR) reactions were run in duplicate for each sample on a Bio-Rad My Cycler iQ5 (BioRad, Hercules, CA, USA). Primers sequences were obtained from Stabvida (Portugal) and thoroughly tested. qPCR was performed in a 20 µl volume containing 5 µl of complementary cDNA (50 ng), 10 µl of 2× Syber Green Supermix (BioRad, Hercules, CA, USA), 2 µl of each primer (250 nM) and 1 µl H2O PCR grade. Specific primers for histone H4 (forward: 5′ ACACCGAGTATGCG -3′; reverse: 5′- TAGCCGTAGAGGATG-3′; LinJ35.1400 histone H4: Gene ID 5073031), Small Hydrophilic Endoplasmic Reticulum-associated Protein (SHERP) (forward: 5′ CAATGCGCACAACAAGAT -3′; reverse: 5′- TACGAGCCGCCGCTTA-3′; LinJ23.1190 SHERP: Gene ID 5069222) and rRNA45 (forward: 5′CCTACCATGCCGTGTCCTTCTA -3′; reverse: 5′- AACGACCCCTGCAGCAATAC -3′) [9] were used for amplification. After amplification, a threshold was set for each gene and cycle threshold-values (Ct-values) were calculated for all samples. Gene expression changes were analyzed using the built-in iQ5 Optical system software v2.1 (Bio-Rad laboratories, Inc). The results were normalized using as reference gene the rRNA45 rRNA sequence [9].
Purified and non-purified promastigotes at a density of 105/ml were washed and suspended in Annexin V binding buffer. Parasites were incubated at room temperature for 15 minutes with AnnexinV-Cy5 (BD Pharmingen, San Diego, CA) and 7-AAD (Sigma). Parasites subjected to Ultra Violet light during 30 minutes and kept in culture for 4 hours were used as positive control. In amastigote differentiation, 2×106 cells with 1 µM of propidium iodide (PI) were used. Data were collected in a BD FACScalibur cytometer (20.000 gated events) and analyzed by FlowJo software (Ashland, OR).
Purified and non-purified promastigotes (6×107/ml) were washed two times, suspended in PBS containing 5 µM of carboxyfluorescein succinimidyl ester (CFSE) (Invitrogen Molecular probes, Eugene, Oregon) and incubated at 37°C for 10 minutes. Labeled parasites were washed, incubated at 4°C for 5 minutes. Parasites were washed again to remove the exceed CFSE dye and suspended in culture medium before proceeding to macrophage infections. For promastigote to amastigote differentiation and proliferation analysis, 107 CFSE-labeled promastigotes were placed on 1 ml of MAA20. Each day, 100 µl of culture added with 1 µM of PI was analyzed by flow cytometry. Axenic amastigotes, identified by the absence of visible flagella and oval shape body, were observed in phase contrast on a Nikon Eclipse 80i.
Cell suspension of bone marrow was obtained by flushing the femurs of susceptible Balb/c mice. The cell suspension was cultured in Dulbecco's modified Eagle's medium (DMEM) (Lonza, Switzerland), supplemented with 10% heat-inactivated FBS (Lonza, Switzerland), 2 mM L-glutamine, 100 U/ml penistreptomycin and 1 mM sodium pyruvate. After overnight incubation at 37°C, non-adherent cells were recovered (300 g for 10 min, at room temperature) and cultured in 24-well culture dishes at 2×105cells/ml in supplemented DMEM. For bone-marrow derived macrophages (BMMø) differentiation 10% L-929 cell conditioned medium (LCCM) was added at days 0 and 4. At day 7 of culture, CFSE labeled promastigotes were incubated with the BMMø at a 10∶1 ratio. After 4 hours, infection was stopped the infection rates were determined at 4, 24 and 48 hours post-infection by a BD FACScalibur cytometer and analyzed by FlowJo software. In some experiments, BMMø were infected and submitted to lipopolysaccharide (LPS) stimulus. Briefly, four hours after infection, infection was stopped and 1 µg/ml of LPS (Sigma) added. Twenty-four hours post-infection, BMMø culture supernatants were collected for cytokine quantification by Enzyme-Linked Immunosorbent Assay - ELISA (TNF-α, IL-12p40, IL-6 and IL-10), using commercial sandwich immunoassay kits (Biolegend and BD, San Diego, CA). Also, BMMø were recovered at 24 hours post-infection for surface co-stimulatory markers analysis. Thus, BMMø were stained with CD40-PE and MHCII-APC at 4°C, during 30 minutes in the dark. The cells were then washed in PBS and suspended in 200 µl of PBS-2% FBS. Data were collected by a BD FACScalibur cytometer and analyzed by FlowJo software.
Promastigotes recovered from stationary culture with 4, 21 and 31 in vitro passages stationary-culture were collected, washed and suspended in sterile PBS. A volume of 200 µl of PBS containing 108 parasites was injected intraperitoneally. Mice of each group were sacrificed at 56 days post-infection. The parasite burden in the spleen and liver was determined by limiting dilution as previously described [10].
The data was analyzed using the non-parametric Kruskal-Wallis test followed by Dunn posttest for multiple comparisons when necessary.
We started by clarifying our in vitro model of L. infantum infection in relation to the parasite development stage. L. infantum parasites recovered from the spleen of infected Balb/c mice were used to start axenic cultures at a 106 parasites/ml. The first task was to clearly define the culture time frame in which we can recover stationary parasites. Performing basic cell cycle analysis we excluded the use of the parasites until 2 days of culture because there was still significant active division (Fig. S1A). In order to evaluate the infectivity of stationary L. infantum, we used CFSE-labeled stationary promastigotes recovered at 3, 5 and 9 days of in vitro growth and BMMø as infection cellular target. Our data demonstrate that 3rd culture-day L. infantum promastigotes were significantly less infectious when compared with the 5th and 9th days of culture (Fig. S1B). These differences were already observed at 4 hours post-infection indicating a deficient parasite uptake with 3rd culture-day L. infantum promastigotes. Intraphagolysosomal adaptation mechanisms do not appear to be involved in the infection differences since similar infection percentages reductions were found between 4 and 24 hours (15.6±0.4 for 3rd culture-day; 15.0±1.2 for 5th culture-day; 18.1±1.1 for 9th culture-day, when comparing 4 with 24 hours post-infection). Many reports now relate virulence with parasites culture viability [11], [12]. In order to lay down the hypothesis that differences in infectivity could be attributed to non-viable parasites, we evaluated the percentages of apoptotic or necrotic parasites by AnnV/7AAD labeling [13]. Nevertheless, no significant differences were found between all culture days (data not shown).
Several groups have already reported that long-term in vitro cultivation (more than 12 months) of Leishmania spp. leads to a totally avirulent promastigote population [6], [14]. According to these findings, we decided to evaluate if the sustained maintenance of L. infantum promastigotes in axenic culture at shorter time periods lead to distinct BMMø in vitro infection rates with distinctive immunologic phenotypes. In order to accomplish this, we maintained L. infantum promastigotes recovered from the spleen of infected mice for 4, 21 and 31 passages, which are equivalent to 20, 105 and 155 division events, considering simple exponential growth. The long-term maintenance of L. infantum in culture did not modify the promastigote growth behavior (Fig. 1A) neither their viability that was always superior to 90% (data not shown). Taking into account the distinct infection profiles depicted in Fig. S1B, we chose 5-day culture promastigotes to compare infectivity. When non-purified parasite cultures were used to in vitro infect BMMø, a marginal but significant loss of infectivity at 48 hours for P21 and P31 when compared to P4 was observed (Fig. 1B). Metacyclic forms have been understood to be the most infective parasite form [8], [15]. Therefore, we enriched the promastigote culture recovered from P4, P21 and P31 in metacyclics recovered by Ficoll density gradient, herein referred as Ficoll-purified promastigotes [16], and analyzed their infectivity in primary BMMø cells (Fig. 1C). When Ficoll-purified metacyclic promastigotes from these cultures were used, differences were abrogated irrespective of the passage used (Fig. 1C). To confirm the enrichment of metacyclic promastigotes in this fraction and as an internal control of our experimental conditions, we analyzed the expression of two genes, SHERP and histone H4 that can be used to evaluate metacyclogenesis. SHERP gene is found to be up-regulated in infective metacyclic promastigotes [17]. On the other hand, higher expression of histone H4 is associated with exponential phase promastigotes [18]. Indeed, the qPCR analysis demonstrated a significant increase in the SHERP mRNA transcripts in Ficoll-purified promastigotes when compared with non-purified parasite cultures (Fig. S2). These results suggested that the maintenance of L. infantum in axenic cultures resulted in a virulence leakage affecting their infectivity probably by the loss of metacyclic parasites. Nevertheless, no significant differences were found between the percentage of Ficoll-purified promastigote recovery from each culture (P4: 6.6±0.1%; P21: 3.4±1.6%; P31: 4.0±0.9%).
We have above demonstrated that the BMMø infection by L. infantum promastigotes depends not only upon the days of culture but is significantly modulated by their axenic culture period. However, we were unable to detect any major changes on macrophage activation status when submitted to L. infantum infection (data not shown), which can be explained by the Leishmania silent entry mechanism [19]. Therefore, we hypothesized that, when facing an inflammatory stimulus, axenic cultures with high passage number should be less successful in subverting macrophage effector functions being less capable of promoting infection. In order to investigate this hypothesis, we incubated BMMø cells with Ficoll-purified or non-purified parasites from distinct culture periods, which were 4 hours later submitted to LPS stimulation. As before, we observed a decrease of the infection rate with the augmentation of parasite in vitro passages (Fig. 2A). This difference was minimized if Ficoll-purified promastigotes were used instead, although it was still statistically significant at 48 hours post-infection (Fig. 2B). LPS stimulation rapidly induces a surface up-regulation of MHCII molecules and co-stimulatory marker CD40. The analysis of these markers demonstrated that high passage number parasites had lower capacities to counteract the LPS activation stimulus (Fig. 2C). Once again, these differences were abrogated when metacyclic-enriched populations were used (Fig. 2C). We have also evaluated the levels of secreted IL-6, IL-12p40 and TNF-α and of the anti-inflammatory IL-10 cytokine. We found that the capacity to control LPS-induced cytokines was variable depending on the number of parasite passages, likely reflecting its distinct virulence. While significant differences were found with P31 parasites in the BMMø secretion levels of IL-6 and TNF-α (Fig. S3A and S3B, respectively), the major modifications were observed at the IL-12p40 and IL-10 levels (Fig. S3C and S3D, respectively). Indeed, P4 parasites were more capable to down-regulate IL-12p40 secretion induced by LPS stimulus, while increasing IL-10 cytokine secretion. This demonstrates that high passage parasites failed to counteract the secretion of pro-inflammatory cytokines induced by LPS in a similar manner. Moreover, if a pro-inflammatory/IL-10 ratio is constructed, a strong correlation was observed between shorter axenic culture maintenance periods and lower pro-inflammatory/IL-10 ratios (Table 1). Since the metacyclic enrichment diminished the differences observed in parasite infection rate and co-stimulatory markers found with stationary-phase promastigotes in different passages, we further investigated whether the cytokine bias was similarly altered. Indeed, the use of Ficoll-purified metacyclic enriched promastigotes, whatever their source, shifted the cytokine environment towards an anti-inflammatory ratio. Although some statistical differences were found between Ficoll-purified promastigotes from different passages, all displayed lower pro-inflammatory/IL-10 ratios when compared with LPS stimulation (Table 1). Overall, these data suggests that when Leishmania-infected BMMø are faced with an inflammatory stimulus, there is a specific overall loss of modulatory capacity that seems to be related to the highly immunoregulatory population of metacyclic parasites.
We have above demonstrated that sustained axenic parasite culture results in a rapid loss of in vitro virulence. Previous reports demonstrating an in vivo loss of virulence were based on long-term, usually more than 1 year, parasite maintenance [6], [14]. However, we have observed a clear in vitro defect after only 21 passages. Therefore, we decided to validate the observed phenotype by performing in vivo infections using the susceptible Balb/c mice model. Six weeks after the infectious challenge with non-purified stationary-phase promastigotes recovered from P4, P21 or P31 cultures, a significant difference was found between P4 and high passage number parasite infections in the liver (Fig. 3A). Similarly, we have observed a significant lower parasite burden in the spleen of P31 infected mice that P4 infections (Fig. 3B). These results confirm the observed in vitro loss of virulence with parasite culture maintenance.
To elucidate the biological mechanisms that account for the loss of virulence due to long term parasite culture, we started by hypothesizing two major potential reasons: decrease number of metacyclic promastigotes or inadequate differentiation into amastigote forms. The quantification of Ficoll-purified promastigotes described above did not show any significant differences among the passages suggesting similar metacyclic quantities. However, the Ficoll density gradient assay is not a specific and sensible test for the quantification of metacyclic promastigotes in a culture but rather a method for its enrichment. Thus, to evaluate the hypothetical deficit on the generation of metacyclic promastigotes, we have performed in vitro infections using BMMø as targets, where we substitute 5% or 10% of non-purified stationary-phase promastigotes from each passage with similar percentages of Ficoll-purified fractions of P4 cultures to increase the total percentage of metacyclic promastigote. As a positive control, we used a naturally attenuated L. infantum from which we were unable to recover metacyclic promastigotes. Indeed, although the promastigotes of this L. infantum strain presents a similar axenic growth curve (Fig. S4), we were always unable to recover by Ficoll density gradient relevant number of promastigotes (lower that 0.1% of initial culture) from stationary-phase cultures. The quantification of CFSE-positive BMMø demonstrated that increasing the percentage of Ficoll-purified promastigotes did not significantly enhance, at any time point, the percentage of infected BMMø for P4 (Fig. 4A), P21 (Fig. 4B) or P31 (Fig. 4C) promastigotes. However, the opposite was observed with the naturally attenuated strain, where a significant increase of infected macrophages was observed at 48 hours post-infection (Fig. 4D). These results demonstrated that the lack of virulence originated from sustained parasite culture cannot be reverted by the addition of enriched-metacyclic fractions. This excludes a defect in the capacity to generate metacyclic promastigotes as the inherent biological cause for the virulence loss. Therefore, we investigated the potential role of inadequate capacity to differentiate in the amastigote form. CFSE-labeled stationary-phase promastigotes recovered from each passage were placed on MAA20 [7] and followed for three days. We evaluated the promastigote differentiation by light microscopy, axenic amastigotes proliferation by CFSE labeling and overall viability by PI staining. All cultures presented axenic amastigotes-like cells after 3 days of differentiation (data not shown). However, high passage number promastigotes displayed a striking decrease of differentiated cells. To quantify these differences, we have assessed the progressive diminution in the intensity of CFSE staining after the differentiation process. In fact, while P4 promastigotes progressively diminished CFSE fluorescence (Fig. 5A and B), high passage number promastigotes exhibit a severe defect to proliferate as amastigotes forms, as observed in both histogram curves (Fig. 5A) and quantification of mean fluorescence CFSE intensity (Fig. 5B). Moreover, this defect was not correlated with a difference on cell death, since similar percentages of viable parasites were found for all cultures during the differentiation process (Fig. 5C). Interestingly, the naturally attenuated strain did not display any significant change to differentiate when compared with P4 promastigotes, suggesting a distinct mechanism of loss of virulence that is not related with the capacity to generate axenic amastigotes.
It is a current empirical methodology to pass Leishmania spp. promastigotes in experimental models to maintain virulence. We have hypothesized that the differentiation process from promastigotes to amastigotes forms would select the most virulent parasites in a heterogeneous culture assuring the continuity of competent and adapted parasites. Therefore, to explore this assumption we have differentiated promastigotes from each passage number in amastigotes both in axenic and in vivo conditions. Axenic amastigotes were obtained by differentiating promastigotes in MAA20 medium [7] for a period of 3 days. The viable axenic amastigotes were maintained axenically in culture for 10 days, after which were re-differentiated to promastigotes forms. In alternative, we recovered L. infantum parasites from the spleen of infected Balb/c mice by allowing amastigote to promastigote differentiation for a period of 7 days. All these promastigotes were sub-cultured for 4 passages and used to infect BMMø. Remarkably, we did not observe any difference between infections whatever the initial parasite source used (Fig. 6A and B). Again, we used as a control the naturally attenuated L. infantum strain. Although an increase of virulence was observed after the differentiation protocol (Fig. 6A), when compared to non-differentiated parasites (Fig. 4D), we observed a general lower infection percentage with the exception of 4 hours post-infection. Overall, these results demonstrate that the defect on virulence due to sustained parasite maintenance can be recovered either by in vitro or in vivo full differentiation to amastigote and back to promastigotes forms.
Visceral Leishmania infections studies have been the center of some controversy which can occasionally be traced back to the use of distinct in vitro promastigotes culture conditions. The plasticity of the Leishmania genome [20] is an important variable to consider when axenic promastigotes are used for in vitro or in vivo studies. Thus, parasite phenotypic plasticity allows it to adapt to the environment generating discrepancies between studies in different laboratories even when using the same Leishmania strain.
In the current work, we have started by investigating in our in vitro model of L. infantum infection the relation between the parasite development stage and its infectivity. The first step was to discard logarithmic parasites because they are not ultimately responsible for the infection [3], so we used a basic cell cycle analysis to discard multiplying parasites (less than 10% of total population are in S/G2 phase after the third day of culture). This data correlated clearly with basic morphological visualization. Stationary L. infantum cultures in day 5 and 9 induced higher BMMø infection rates than day 3 parasites. This difference in infectivity might translate the time frame required for becoming truly metacyclic parasites [21], which was also corroborated by the less amount of metacyclic recovered (data not shown). Since the presence of apoptotic parasites is essential for a virulent inoculum of Leishmania promastigotes [22], we decided to quantify the percentage of apoptotic and dead parasites in each case to remove this possible bias from our analysis. In fact, for all time frames tested, the differences in infectivity were not related to apoptotic or dead parasites in the non-purified or Ficoll-purified populations, with culture viability always higher than 90%.
Some authors described that in vitro maintenance for long periods constitute an important factor for the loss of virulence in L. infantum [14] and L. major [6] promastigotes. Nonetheless, this loss of virulence is a reversible phenomenon, since serial passages on susceptible mice allow the parasite to recover a virulence phenotype [23]. In the present study, we complemented the previous observations by comparing the impact of continuous in vitro culture on Leishmania promastigote virulence and also into the capacity of host macrophage manipulation. Our data clearly demonstrated a loss of L. infantum virulence related to the augmentation of in vitro culture periods although no modification was observed in the axenic promastigote growth behavior. This significant loss of infectivity was observed as soon as 105 days of successive (21 passages) culture and worsened with parasite maintenance in culture. In fact, 20 passages was the soonest time point where we could have a significant reproducible loss of infection. There is a grey area between passage 9 and passage 20 where we can have variation of infection in a manner that probably reflects the initial parasite inoculum recovered from the mammal. This was also observed after in vivo infection where we had a significant decrease in parasite burden after 21 and 31 passages, confirming the in vitro observations.
The percentage of metacyclic in a heterogeneous stationary-phase is an important factor in the parasite infectivity since they are significantly more infective than the non-purified population. We used a Ficoll density gradient methodology to enrich the percentage of metacyclic promastigotes [24]. Beyond the morphological changes, during the Leishmania spp. differentiation process, modifications also occur in gene expression and in the composition of parasite surface that help to characterize metacyclic promastigotes. Thus, we have evaluated the enrichment of metacyclic promastigotes in the Ficoll-purified fraction by microscopy (data not shown) and through qPCR analysis of the SHERP and histone H4 gene expression. The augmentation of SHERP gene expression in Ficoll-purification supported an enriched metacyclic population. The use of Ficoll-purified promastigotes abolished the differences found among the different passages. Yet, significant differences were found for P21 and P31, at 48 hours post-infection when facing an inflammatory stimulus, showing that the phenomenon of loss of virulence, although less prominent in the metacyclic-enriched parasites, was not restricted to the unpurified culture. Since the differences at the infection level were significant, we examined if there was a potential effect on the macrophage activation status. In the presence of a strong inflammatory stimulus, L. infantum is able to suppress certain LPS-derived pro-inflammatory cytokine responses in an active parasite-specific process while it augments the production of some anti-inflammatory cytokines (Silvestre et al, unpublished data). Indeed, the addition of LPS to Leishmania spp. infected cells was demonstrated to synergistically induce the secretion of the anti-inflammatory IL-10 cytokine in monocytes [25] and in macrophages [26]. The functional polarization of macrophages into IL-10 producers characterized as M2 cells [27] has been long understood to play a crucial role in the success of parasite infection process [28]. Our results demonstrated a growing defect of high passage parasites to modulate the LPS stimulatory effect. Furthermore, it is clear from the inflammatory profile depicted in Table 1 that metacyclic enriched fractions are always significantly more effective in abrogating a macrophage response to the inflammatory stimuli than their non-purified counterparts revealing the metacyclic parasites as a highly immunomodulatory population with a distinct profile from the non-purified population. There is a distinct and significant loss of immunomodulatory properties from P4 to P21 that becomes stable after P21. This loss of immunomodulatory properties seems reminiscent of phenomenon of transient gene expression similar to what happens under drug pressure [29], being lost upon the terminus of immunological pressure. Indeed, this might be happening in just a few passages of axenic culture. In an attempt to explain the loss of virulence mechanism, some authors referred to a reduction of metallo and cysteine peptidases activity, important for virulence, in L. braziliensis [30], [31] and in L. amazonensis [32] or mitochondrial defects [33] during long-term culture. However, others have been unable to detect any differences in the parasite enzymatic profile with long in vitro periods of cultivation [34], [35].
One can speculate that the overall loss of immunomodulatory properties over time, and in consequence loss of virulence might reflect a diminution of the number of metacyclic parasites in the population. Although the percentage of Ficoll-recovered promastigotes was quite similar among the three tested passages, these fractions do not constitute a pure metacyclic population, so we decided to complement older passage parasites with Ficoll recovered metacyclic promastigotes to access if the loss of virulence could be reverted by exogenous addition of metacyclic parasites from an early passage. No improvement in the overall infection was observed, although, when these same metacyclic parasites were added to an avirulent field strain, from which we were unable to recover metacyclics, there was an improvement on the infection.
Another possibility to explain the virulence loss was the possibility of a defective promastigote to amastigote differentiation. Our data clearly demonstrated that high passage number promastigotes displayed decrease capacity in differentiating, which was not correlated with decreased cellular viability. This incapacity translates into fewer parasites able to differentiate leading to a less capable population to face host cell response. Ultimately, this results in lower parasite burdens in vitro and in vivo. Moreover, these axenic amastigotes recovered after differentiation were morphologically indistinguishable and retained similar growth capacity (data not shown). To ultimately state and confirm the importance of promastigote to amastigote differentiation as a driving selective force for virulence, we showed that parasites passed through the amastigote stage, either in vitro or in vivo, revert the loss of virulence. This fact in conjunction with the remarkable loss of immunomodulatory properties leads to the early loss of virulence detected in our model.
We do not rule out metacyclogenesis related defects as a driving force for a virulence. In relation to metacyclogenesis it has been argued that a successful and complete differentiation is dependent on the presence of large amount of metacyclic promastigotes [36]. However, it is still not clear whether this process is an essential step in the differentiation in vitro, since procyclic promastigotes appear to differentiate with equal efficiency as metacyclics [37]–[40]. Indeed, our results with the naturally attenuated strain support this notion. Our data do not rule out that P21 or P31 metacyclic promastigotes could not display any sort of biochemical or protein expression defect that may impact the differentiation process. Similarly, we cannot reject the idea of longer periods of sustained cultured originating defective metacyclic cultures. However, during our study, the loss of virulence was related to a specific defect on promastigote to amastigote differentiation.
Overall, our data demonstrated that the loss of virulence is linked with decreased capacity to differentiate in amastigote forms, which may probably be originated from the absence of a complete life cycle. Therefore, special care must be taken when performing experiments with axenic Leishmania promastigotes. The systematic and rigorous control of Leishmania culture conditions should be considered as a keystone for each experimental protocol. The differences found in infectivity accompanied by disparate effects at the macrophage activation levels point to significant differences at biochemical and structural level, enlarging the effects of careless parasite maintenance to other experimental fields. This information is extremely relevant especially for those developing new drug and vaccine approaches. In such cases the immune response to the parasite is the essence of the experimental procedure.
|
10.1371/journal.pntd.0004599 | Muscle Tissue Damage Induced by the Venom of Bothrops asper: Identification of Early and Late Pathological Events through Proteomic Analysis | The time-course of the pathological effects induced by the venom of the snake Bothrops asper in muscle tissue was investigated by a combination of histology, proteomic analysis of exudates collected in the vicinity of damaged muscle, and immunodetection of extracellular matrix proteins in exudates. Proteomic assay of exudates has become an excellent new methodological tool to detect key biomarkers of tissue alterations for a more integrative perspective of snake venom-induced pathology. The time-course analysis of the intracellular proteins showed an early presence of cytosolic and mitochondrial proteins in exudates, while cytoskeletal proteins increased later on. This underscores the rapid cytotoxic effect of venom, especially in muscle fibers, due to the action of myotoxic phospholipases A2, followed by the action of proteinases in the cytoskeleton of damaged muscle fibers. Similarly, the early presence of basement membrane (BM) and other extracellular matrix (ECM) proteins in exudates reflects the rapid microvascular damage and hemorrhage induced by snake venom metalloproteinases. The presence of fragments of type IV collagen and perlecan one hour after envenoming suggests that hydrolysis of these mechanically/structurally-relevant BM components plays a key role in the genesis of hemorrhage. On the other hand, the increment of some ECM proteins in the exudate at later time intervals is likely a consequence of the action of endogenous matrix metalloproteinases (MMPs) or of de novo synthesis of ECM proteins during tissue remodeling as part of the inflammatory reaction. Our results offer relevant insights for a more integrative and systematic understanding of the time-course dynamics of muscle tissue damage induced by B. asper venom and possibly other viperid venoms.
| The local pathology induced by viperid snakes is characterized by a complex of alterations as consequence of direct and indirect effects of the toxins present in the venom, as well as the host response to tissue damage, and constitutes a dynamic process of degenerative and reparative events. The pathogenesis of local effects induced by Bothrops asper venom has been studied by traditional methodologies. Recently, proteomic analysis of wound exudates collected in the vicinity of affected tissue has become a powerful tool to study the pathogenesis of local envenoming from a more integrative perspective. Thus, in the present study we analyzed the dynamics of the local effects induced by B. asper venom in the gastrocnemius muscle of mice through a proteomic and immunochemistry approach in order to identify biomarkers of tissue damage and repair during the course of envenoming. Our results showed an early presence of cytosolic and mitochondrial proteins in exudates as compared to cytoskeletal proteins, which reflect the rapid cytotoxic effect of venom, followed by the action of endogenous proteinases in the cytoskeleton of damaged muscle fibers later on in the course of envenoming. On the other hand, the early presence of extracellular matrix components and the increment of some of them in exudates, reflect the rapid microvascular damage and hemorrhage induced by the venom, followed by the action of endogenous matrix metalloproteinases (MMPs) during tissue remodeling as part of the inflammatory response. Overall our study allowed the identification of key biomarkers of tissue damage and repair as part of the pathological effects induced by B. asper venom in skeletal muscle, which offer relevant insights for a better understanding of the complex dynamics of local pathology induced by viperid snakebite envenoming.
| The viperid snake Bothrops asper is responsible for most snakebite cases in Central America and some regions of Mexico and South America [1,2]. The local pathology induced by viperid snakes is characterized by edema, blistering, hemorrhage, lymphatic vessel damage, and necrosis of skin and muscle, some of which can be attributed to the degradation of extracellular matrix (ECM) [1,3]. Such alterations develop very rapidly after the bite, and in some cases can lead to permanent tissue damage, regardless of the application of antivenom treatment. Significant efforts have been undertaken over the last several decades to identify the toxins responsible for these effects, as well as to characterize the pathogenesis of these alterations [3–5]. Nevertheless, the complexity of this pathology demands further analyses into hitherto unknown aspects of tissue damage and the complex interplay between degenerative and early reparative events. As envenoming is a dynamic event, it is critical to investigate the process over time, which is the main focus of this study.
The pathogenesis of local effects induced by B. asper venom has been studied by traditional methodologies, such as histological and ultrastructural analyses, immunohistochemical methods, and quantification of particular components and tissue markers in tissue homogenates or fluids, as a consequence of the action of crude venom and purified toxins [3,6–12]. Despite significant advances in the study of local tissue damage with these approaches, subtle changes in key biomarkers of tissue damage and repair during the course of envenoming remain to be identified and characterized. Moreover, since the venom is a highly complex mixture of components functioning over time, relevant information related to synergistic action of toxins could be missed when working only with isolated toxins; therefore, studies with crude venom may better advance our understanding from a predominantly reductionist to a holistic view of these multifactorial time-dependent phenomena.
Recently, proteomic analysis of exudates collected around the affected tissue has become a new methodological tool to study the pathogenesis of local tissue damage induced by snake venom from a more integrative perspective [13–17]. This approach has been used to study the alterations caused by B. asper snake venom [15], and some of its toxins, such as a myotoxic phospholipase A2 (PLA2) and a hemorrhagic snake venom metalloproteinase (SVMP) [13,14,16]. Moreover, proteomic analysis has allowed the comparison between the action of different types of hemorrhagic and non-hemorrhagic SVMPs [13,16,17]. These studies have identified differences in the species and abundance of intracellular proteins, ECM components, and other proteins present in exudates, which offer new insights in the mechanism of action of these toxins, and in the tissue damage induced by the venom [13–16]. However, these studies have been carried only at early time periods in the course of envenoming and therefore provide only a narrow window within the whole scenario of local pathology.
In the present study we analyzed the time-course variation in the protein composition and abundance of wound exudates collected from mouse gastrocnemius muscle injected with B. asper snake venom utilizing a proteomic and immunochemistry approach, in conjunction with histological analysis of tissue alterations, with the aim of identifying biomarkers of tissue damage and tissue remodeling characteristic of early and late stages of envenoming. This approach allowed the identification of key differences in some intracellular proteins and ECM components over time, which underscores the rapid cytotoxic and hemorrhagic effect of venom, followed by the action of endogenous proteinases associated with tissue remodeling later on in the course of envenoming. These results offer relevant insights for a better understanding of the complex pathological phenomena of viperid snakebite envenoming.
B. asper venom was obtained from more than 40 adult specimens collected in the Pacific region of Costa Rica and maintained at the serpentarium of Instituto Clodomiro Picado. After collection, venoms were pooled, lyophilized, and stored at -20°C until used.
CD-1 mice with a body weight between 18 and 20 g were used for the in vivo studies. All the experimental protocols involving the use of animals were approved by the Institutional Committee for the Care and Use of Laboratory Animals (CICUA) of the University of Costa Rica (protocol approval number CICUA 025–15), and meet the International Guiding Principles for Biomedical Research Involving Animals (CIOMS).
Groups of four CD-1 mice (18–20 g) were injected in the right gastrocnemius with 50 μg of B. asper venom, dissolved in 50 μL of 0.12 M NaCl, 0.04 M phosphate, pH 7.2 solution (PBS). Control mice were injected with PBS alone. After 1, 6 and 24 h of injection, mice were sacrificed by CO2 inhalation and samples of the injected muscles were resected and added to 10% formalin solution in PBS. After 48 h fixation, routine processing of tissues was performed, followed by embedding in paraffin. Sections of 5 μm thickness were obtained for each sample and stained with hematoxylin–eosin for light microscopic observation.
Groups of five CD-1 mice (18–20 g) were injected in the right gastrocnemius with 50 μg of B. asper venom, dissolved in 50 μL of PBS. After 1, 6 and 24 h of injection, mice were sacrificed by CO2 inhalation, a 5 mm incision was made with a scalpel in the skin overlying the injected muscle, and a heparinized capillary tube was introduced under the skin to collect the wound exudate fluid. An approximate volume of 20–50 μL of exudate was collected from each mouse. Exudate samples were then pooled and lyophilized for further analysis.
Wound exudates were collected as previously described and centrifuged at 5000 g for 3 min. The CK activity of supernatants was determined using a commercial kit (CK-Nac, Biocon Diagnostik, Germany). CK activity was expressed in International Units /L (IU/L).
Lyophilized wound exudate samples were resuspended in water and protein quantification was performed using micro BCA protein assay kit (Thermo Scientific). Twenty micrograms of protein was precipitated with acetone, resuspended in Laemmli buffer and separated in a 5–20% precast electrophoresis gel (Bio-Rad). The gel was stained with Coomassie Brilliant Blue and lanes were cut into 8 equal size slices. Gel slices were destained for 3 h and the proteins were reduced (10 mM dithiothreitol, DTT) and alkylated (50 mM iodoacetamide) at room temperature. Gel slices were then washed with 100 mM ammonium bicarbonate, dehydrated with acetonitrile and dried in a speed vac, followed by in-gel digestion with a solution of Promega modified trypsin (20 ng/μL) in 50 mM ammonium bicarbonate for 30 min on ice. Excess trypsin solution was removed and the digestion continued for 18 h at 37°C. The resulting tryptic peptides were extracted from gel slices with two 30 μL aliquots of a 50% acetonitrile/5% formic acid solution. These extracts were combined and dried to 15 μL for mass spectrometric (MS) analysis.
LC/MS/MS was performed using a Thermo Electron Orbitrap Velos ETD mass spectrometer system. Analytical columns were fabricated in-house by packing 0.5 cm of irregular C18 Beads (YMC Gel ODS-A, 12 nm, I-10-25 um) followed by 7.5 cm Jupiter 10 μm C18 packing material (Phenomenex, Torrance, CA) into 360 x 75 μm fused silica (Polymicro Technologies, Phoenix, AZ) behind a bottleneck. Samples were loaded directly onto these columns for the C18 analytical runs. 7 μL of the extract was injected, and the peptides were eluted from the column at 0.5 μL/min using an acetonitrile/0.1M acetic acid gradient (2–90% acetonitrile over 1 h). The instrument was set to Full MS (m/z 300–1600) resolution of 60,000 and programmed to acquire a cycle of one mass spectrum followed by collision-induced dissociation (CID) MS/MS performed in the ion trap on the twenty most abundant ions in a data-dependent mode. Dynamic exclusion was enabled with an exclusion list of 400 masses, duration of 60 seconds, and repeat count of 1. The electrospray voltage was set to 2.4 kV, and the capillary temperature was 265°C.
The data were analyzed by database searching using the Sequest search algorithm in Proteome Discoverer 1.4.1 against the Uniprot Mouse database from July 2014. Spectra generated were searched using carbamidomethylation on cysteine as a fixed modification, oxidation of methionine as a variable modification, 10 ppm parent tolerance and 1 Da fragment tolerance. All hits were required to be fully tryptic. The results were exported to Scaffold (version 4.3.2, Proteome Software Inc., Portland, OR) to validate MS/MS based peptide and protein identifications, and to visualize multiple datasets in a comprehensive manner. Confidence of protein identification in Scaffold is shown as ≥ 95% confidence (green coloration) and 80% to 94% confidence (yellow coloration). Relative quantization of proteins was performed by summing all data from the 8 gel slices for a particular sample in Scaffold and then displaying the Quantitative Value from the program. This number gives an average total of non-grouped spectral counts for a protein divided by the total non-grouping spectral counts for the 8 mass spectral runs from the gels slices from each lane (http://www.proteomesoftware.com/). The Quantitative Value allows a relative quantitative comparison between a specific protein from different samples and relative abundance between proteins for a particular exudate sample.
For immunoblotting, 100 μg protein of each exudate sample were separated under reducing conditions on 4–15% Tris–HCl polyacrylamide gradient gels, and transferred to nitrocellulose membranes. Immunodetection was performed by incubating the membranes overnight at 4°C stirring with rabbit anti-type IV collagen polyclonal antibody at a dilution of 1:200 (Abcam ab19808), rabbit anti-nidogen 1 polyclonal antibody at a dilution of 1:500 (Abcam ab14511), rabbit anti-laminin polyclonal antibody at a dilution of 1:1,000 (Thermo PA1-32130), rabbit anti-type VI collagen polyclonal antibody at a dilution of 1:2,000 (Millipore AB7821), rabbit anti-type I collagen polyclonal antibody at a dilution of 1:1,000 (Abcam ab21286), or rabbit anti-fibronectin polyclonal antibody at a dilution of 1:3,000 (Abcam ab2413). The reaction was developed using anti-rabbit peroxidase antibody at a dilution of 1:5,000 (Jackson ImmunoResearch) and the chemiluminescent substrate Lumi-Light (Roche). Images were captured with the ChemiDoc XRS+ System (BioRad) and the analysis was performed with the ImageLab software (BioRad).
The pathological alterations induced by B. asper venom were studied on mouse gastrocnemius muscle tissue over a time period of 24 h. Tissue sections from control mice injected with PBS showed normal histological features of skeletal muscle tissue with transverse bundles of muscle fibers, surrounded by connective tissue and normal vascular and nerve structures (Fig 1A). Tissue sections from mice injected with B. asper venom showed intense hemorrhage at 1 h (Fig 1B) and 6 h (Fig 1C), evidenced by the presence of abundant erythrocytes in the interstitial space surrounding muscle fibers. After 24 h of injection of B. asper venom, the hemorrhage decreased since the amounts of extravascular erythrocytes was reduced as compared to previous time intervals (Fig 1D), consistent with previous observations [7].
Moreover, tissue sections from mice injected with venom revealed prominent necrosis of skeletal muscle fibers at the first hour interval (Fig 1B and 1C). After 24 h following venom injection, the bundles of muscle fibers appeared partially lost and disorganized with a hyaline appearance (Fig 1D). These pathological observations also agree with previous studies [7,11,12]. Additionally, an infiltration of inflammatory cells was observed in tissue sections, especially after 6 h and 24 h of venom injection with a marked increment at 24 h (Fig 1D). The predominant cell type was polymorphonuclear leukocytes, although a proportion of mononuclear cells, i.e. macrophages, were also observed at 24 h. These observations also agree with previous studies [9].
On the other hand, CK activity of exudate samples collected after injection of venom was 228,776 ± 47,137 IU/L at 1 h, 162,344 ± 23,371 IU/L at 6 h, and 23,371 ± 11,660 IU/L at 24 h (Fig 1E). CK is a marker of plasma membrane damage and cell death of skeletal muscle fibers; hence it appears that myotoxic activity of the venom is highest at one hour, decreasing afterwards. These results are in agreement with the muscle tissue damage observed in the histological analysis, which occurs early on in the course of envenoming.
Wound exudate samples collected from mice injected with B. asper venom were decomplexed by SDS-PAGE for subsequent proteomic analysis. From the mass spectrometric analysis of the gel bands, a total of 537, 578, and 486 proteins were identified in exudates at 1 h, 6 h, and 24 h, respectively, with protein identification probability above 95% and minimum of two peptides (S1 Table). The most abundant proteins identified based on their Quantitative Value (see http://www.proteomesoftware.com/ for full description of term) were classified within the following groups: intracellular proteins (Table 1 and S2 Table), ECM proteins (Table 2), membrane proteins (S3 Table), coagulation factors (S4 Table), and proteinase inhibitors (S5 Table). Within each group, the proteins were organized by those that changed at least three fold as compared to another time and proteins which did not show significant change between the three time intervals, i.e. those whose amounts did not differ more than threefold between times.
A total of 222 intracellular proteins (Table 1 and S2 Table) and 13 membrane proteins (S3 Table) were detected in exudates, thus demonstrating direct or indirect cellular damage induced by the venom. The most abundant intracellular proteins detected in exudates were hemoglobin subunit beta-2 and creatine kinase M-type, in agreement with the hemorrhagic and myotoxic activity of B. asper venom, respectively. Moreover, the creatine kinase M-type identified in the exudates was detected at the highest level at 1 h, and decreased over time until reaching a six fold reduction at 24 h. These results are in agreement with the CK activity of exudates and the muscle tissue damage observed in the histological analysis. In contrast, there was a trend for cytoskeletal proteins, such as actin, myosin, and tropomyosin, to increase in the exudates over time, while most of cytosolic and mitochondrial proteins appeared at the first hour of venom injection, and decreased afterwards.
Of serum proteins, a total of 10 coagulation factors (S4 Table) and 14 proteinase inhibitors (S5 Table) were detected in exudates at various times. Fibrinogen beta and gamma chains appeared in the exudates at the first hour following venom injection and their abundance increased over time. Other coagulation factors detected whose amounts changed at least three fold as compared to values at other time were coagulation factor X, XII, and XIII. The inter alpha-trypsin inhibitor was the only proteinase inhibitor that increased at least threefold at 24 h as compared to 1 h and 6 h.
A total of 24 ECM proteins were identified in exudates, of which 21, 24, and 13 proteins were detected at 1 h, 6 h, and 24 h, respectively (Table 2). Most of these proteins showed a differential abundance greater than three-fold between samples collected at different times. The most abundant basement membrane (BM) protein detected in the wound exudates was BM-specific heparan sulfate proteoglycan core protein (perlecan), followed by alpha 1 and 2 chains of type IV collagen. Most of the BM components, such as heparan sulfate proteoglycan, type IV collagen and nidogen-2, appeared in the exudates at 1 h, and the amount decreased over time, largely becoming undetectable at 24 h. Conversely, the amount of laminin γ-1 detected in the exudates increased at 6 h and 24 h, and the amount of nidogen-1 increased at 6 h as compared to 1 h and 24 h (Fig 2A). Other collagens, such as types VI, XV, and XVIII collagens, were present in the exudates at 1 h and 6 h, but were not detected at 24 h. Interestingly, type I collagen was also detected in the exudates at the first hour and its abundance increased at 24 h (Fig 2B). Other ECM proteins detected in the exudates whose abundance were greater at 6 h as compared to 1 h and 24 h were type III collagen, fibrillin 1 and 2, chondroitin sulfate proteoglycan 4, and type XII collagen. On the other hand, thrombospondin 1 appeared in the exudates at 1 h and decreased over time. Other ECM proteins detected in the exudates whose amounts did not vary more than threefold between times were fibronectin, thrombospondin-4, vitronectin, dermatopontin, proteoglycan 4, type XIV collagen, and lumican (Table 2).
In order to determinate whether SVMP or endogenous proteases are active in the wound exudates, proteolytic activity assays of exudate samples were performed. Exudate samples collected from mice injected with B. asper venom at 1 h showed the highest proteolytic activity on gelatin fluorescein conjugate compared with samples collected at 6 h and 24 h. (Fig 4A). When exudates collected at 1 h were incubated with polyclonal antibody against the SVMP BaP1, the proteolytic activity of the exudates was almost completely inhibited since only 8% of the activity remained.
Using zymography several gelatinolytic bands were detected corresponding to proteins of 50–150 kDa in the exudate samples collected at different times (Fig 4B). An increase of two main bands of about 100 kDa and 60 kDa was observed in exudates collected at 6 h and 24 h. These molecular masses are consistent with the latent forms of matrix metalloproteinases (MMP) 9 and 2, respectively [22]. Therefore, the zymography showed an increase of proteolytic activity of endogenous MMPs in exudate over time. Furthermore, bands corresponding to the molecular mass of the SVMP BaP1 were not detected in the zymographic analysis of exudate samples.
Envenoming by venomous snakes gives rise to a complex pathophysiology by virtue of the complexity of the venoms and the fact that the toxins in the venom produce manifold effects in the tissues. Proteomic analysis of wound exudates collected in the vicinity of affected tissue constitutes a powerful approach to study the pathogenesis of tissue damage induced by snake venoms from a more comprehensive perspective [13–17], thus complementing histological, ultrastructural and biochemical analyses. This methodological tool has been used to investigate the early alterations provoked by B. asper venom [15] and some of its toxins, especially myotoxic PLA2s and hemorrhagic SVMPs [13,14,16], as well as the inhibitory effects of antivenoms and low molecular mass inhibitors [15]. However, analyses in these previous studies were performed at a single time interval after injection, thus precluding the understanding of these events from a time-course perspective. In this study, we investigated the dynamics of local effects induced by B. asper venom in the gastrocnemius muscle of mice at various time intervals.
Our histological and biochemical observations agree with previous studies showing a rapid development of myonecrosis and hemorrhage, followed by an inflammatory process characterized by the infiltration of neutrophils and macrophages at later time intervals [4,5,7,9,10]. Previous works on venom-induced myonecrosis have quantified CK activity in plasma [6], where the highest levels were observed at 3 h post envenoming. In contrast, in exudates, highest CK levels are higher than in plasma, and peak activity occurs at 1 h instead of 3 h. This difference may be attributed to the kinetics of absorption of this enzyme into the circulation after its release from damaged muscle fibers, since exudate was collected close to the venom-injected muscle. In agreement with previous pathological and proteomic studies, intracellular proteins were abundant in exudates, as a consequence of the cytotoxic effect of venom, especially on skeletal muscle fibers [7,11,14,15]. The time-course analysis of the intracellular proteins in exudate underscores that most of the cytosolic and mitochondrial proteins appear early on due to the rapid action of myotoxic PLA2s and PLA2 homologues in muscle tissue [7,11,14,15,23], followed by a decrease of these proteins. Most of these proteins are derived from the cytosol of skeletal muscle fibers since myotoxic PLA2s induce a rapid disruption of the integrity of muscle cell plasma membrane [4,7,24]. The high CK activity of exudates at 1 h and our histological observations corroborate the early onset of myonecrosis in the course of envenoming and agree with proteomic analysis.
In contrast to cytosolic proteins, most of the cytoskeletal proteins, such as actin, myosin, and tropomyosin, are more abundant in exudates collected at later time periods. This late increment suggests that the presence of cytoskeletal protein fragments in the exudate depends on the action of proteinases that release these structural components from damaged cells. A prominent calcium influx in muscle cells occurs after venom-induced plasma membrane damage [25,26]. An increased calcium concentration in the cytosol results in the activation of calpains, which might hydrolyze cytoskeletal components [27]. Subsequently, proteinases derived from inflammatory cells arriving at the necrotic tissue may also contribute to proteolysis of muscle cytoskeletal proteins [9,10]. Thus, the proteomic analyses reveals two ‘waves’ of release of intracellular proteins to exudates: an early release of cytosolic and mitochondrial proteins, which depends on the rapid myotoxin-induced membrane damage, and a more delayed release of cytoskeletal protein fragments, which is due to proteolytic degradation.
The presence of cell membrane-associated proteins may be evidence of direct or indirect cellular damage induced by the venom. Moreover, proteolysis of these components, either by venom or endogenous proteinases, may cause their ‘shedding’ and diffusion to the exudate compartment. It is tempting to speculate that, in addition to being a passive reflection of venom-induced plasma membrane damage, the release of these protein fragments may also play a functional role in cellular signaling associated with inflammatory and reparative events. However, the pathological relevance of the hydrolysis of these proteins in the overall mechanism of local tissue damage induced by snake venoms has not been established and needs further study.
The presence of ECM proteins in wound exudates reflects the cleavage by either venom-derived proteinases or endogenous proteinases, such as MMPs, generated in the course of the inflammatory response. The degradation of ECM is a relevant component of viperid venom-induced tissue damage, and proteomic analysis has been particularly useful in revealing a complex pattern of hydrolysis [14–16]. Previous studies detected B. asper venom components in muscle homogenates of mice during the first week after experimental envenoming [28]; however, the activity of these toxins has not been previously addressed. When assessing the proteinase activity of exudates, highest activity was detected in samples collected after 1 h of envenoming; here we demonstrate that this is mainly due to the action of SVMPs, since antibodies against BaP1, the most abundant proteinase in B. asper venom [29], almost fully inhibited exudate-induced proteolysis. However, this enzymatic activity decreased over time, probably as a consequence of diffusion of venom components from the injected muscle or of inhibition by plasma or tissue-derived proteinase inhibitors. It is likely that activity at later time intervals, i.e. 24 h, is mostly due to endogenous MMPs generated in the course of the inflammatory response, such as MMP-9 and MMP-2, which was confirmed by the detection with zymography of the wound exudates, although it remains possible that some venom proteinases persisting in the tissue may also contribute to this observation. These results agree with previous studies which demonstrated an increase in the expression of MMP-9 and MMP-2 in muscle tissue injected with B. asper venom [28] or with purified SVMP and PLA2 toxins [30]. Taken together these findings suggest that the hydrolysis of ECM is mainly due to SVMPs in the early stages of envenoming, while endogenous MMPs participate later in the course of envenoming.
A large body of experimental evidence indicates that BM and related ECM components that provide stability to microvessel structure are the key targets of hemorrhagic SVMP [5,13,14,16,31–34]. Moreover, SVMP-induced hemorrhage occurs very fast after injection [12,35–37]. Therefore, the presence of ECM components in wound exudates during first hour as compared to later time periods may offer important insights for understanding the mechanism of action of hemorrhagic SVMPs. Regarding BM components, the presence of degradation products of perlecan, type IV collagen, nidogen, and laminin in wound exudates underscores a rapid and drastic damage of BM structure. In particular, perlecan and type IV collagen are abundant in exudates after the first hour, when hemorrhagic events have occurred, and then their amounts decrease over time, as they were not detected by proteomic analysis at 24 h. Western blot analysis of exudate confirmed the presence of fragments of type IV collagen 1 h after venom injection.
Perlecan is the most abundant BM protein detected in the wound exudates during the first hour. In previous proteomic studies, we have found that the relative amount of perlecan in wound exudates induced by a hemorrhagic SVMP was greater as compared to a non-hemorrhagic one [13], but similar when compared to other hemorrhagic SVMPs [16], and its presence was abolished when B. asper venom was previously incubated with batimastat [15], a metalloproteinase inhibitor. Such findings, together with our data, suggest that degradation of perlecan in early stages of envenoming may play an important role in the hemorrhagic mechanism of SVMPs. This proposal agrees with the known structural role of perlecan in BM [38–42]. In addition, mutations in the perlecan gene in mice have been associated with loss of BM integrity in different tissues [43–45], including microvasculature of brain and skin, which cause severe bleedings due to dilatation and rupture of microvessels [45].
On the other hand, previous proteomic studies using similar models have not detected type IV collagen in wound exudates induced by B. asper venom or its toxins [13–16]. However, several studies using in vitro and in vivo models have identified type IV collagen as one of the most likely key components degraded by hemorrhagic SVMPs and associated with the initial microvessel destabilization and hemorrhage [5,13,16,34,46,47]. Our present proteomic results did identify fragments of type IV collagen in exudates at 1 h, thus agreeing with previous immunohistochemical and immunochemical evidence [13,16]. In addition, according to Western blot analysis, degradation products of type IV collagen appear in exudates in samples collected at 1 h. Such early appearance of type IV collagen and perlecan strongly suggest that their degradation is due to the direct proteolytic activity of SVMPs. The hypothesis that type IV collagen is a key target in the hemorrhagic mechanism of SVMPs is compatible with the structural role of this collagen in the mechanical stability of BM, as it is stabilized by covalent cross-links [41,42,48–51]. In addition, mutations on type IV collagen genes have been associated with pathological alterations in microvessels and with hemorrhage in brain, kidney and lungs in mice and humans [52–58]. Thus, the rapid hydrolysis of perlecan and type IV collagen after injection of B. asper venom supports the view that BM destabilization leading to hemorrhage is likely to depend on the degradation of these mechanically-relevant components.
Nidogen 2 appeared in early time periods in wound exudates, in agreement with previous proteomic studies [13,15], and then it decreased over time in our proteomics analysis. Since nidogen 2 is more abundant in the BM of blood vessels [59], and its time-course dynamics of appearance in exudates is similar to that observed for type IV collagen and perlecan, the release of nidogen 2 might be associated with vascular BM damage. In contrast, taken together the proteomic and Western blot analyses showed that nidogen 1 increased over time in wound exudates. In addition, nidogen 1 and 2 have been detected in plasma of healthy mice [60], which could explain the presence of nidogen 1 in the wound exudates according to Western blot results. On the other hand, laminin γ1, which is widely distributed [61,62], also increased over time in wound exudates. The time-course variation of the molecular masses of immunoreactive bands in the cases of nidogen 1 and laminin underscores the dynamics of degradation of these components over time. Furthermore, Escalante et al. [13] demonstrated similar patterns of degradation for nidogen and laminin in muscle tissue induced by hemorrhagic and non-hemorrhagic SVMPs.
Our observations allowed the analysis of the time-course dynamics of the hydrolysis of non-fibrillar collagens associated with the BM, such as types VI, XV, and XVIII collagens. As in the case of type IV collagen and perlecan, hydrolysis of these components was highest at 1 h, hence indicating a rapid degradation, probably by venom proteinases. These collagens connect the BM with fibrillar collagens of the matrix [39,63], and are known to play a relevant role in the mechanical stability and integration of BM with connective tissue [39,63]. Hence, the hydrolysis of these components by SVMP might be also critical for capillary wall destabilization, as have been previously proposed [13,64]. Alternatively, the increase of these collagens in exudates might be consequence of BM damage after the hydrolysis of other components, such as type IV collagen and perlecan. Type VI collagen is more abundant in the BM of muscle cells [65–67]; thus the increment of non-degraded type VI collagen chains in exudates could reflect synthesis de novo during reparative and regenerative events in muscle tissue. The role of the degradation of these collagens in the initial destabilization of BM induced by hemorrhagic SVMP is an issue that should be further investigated.
Other ECM components of interest detected in the proteomic analysis are collagen I and fibronectin. Collagen I is a fibril-forming collagen distributed in non-cartilaginous connective tissues such as skin and connective tissue of muscle [63]. According to proteomic results, the relative abundance of collagen I in exudates is higher at 24 h as compared to 1 and 6 h. This late hydrolysis of collagen I could be result of the action of endogenous MMPs synthesized during the course of inflammation in the damage tissue. Fibronectin was detected in the exudates both in proteomic and immunochemical analyses. This protein can be found in two forms: plasma fibronectin, which is a soluble molecule synthesized by hepatocytes, and cellular fibronectin, which is produced in the tissues and is incorporated in the ECM [62]. Thus, the presence of fibronectin in exudates could be either a consequence of plasma exudation or hydrolysis from the ECM. According to proteomic analysis, the amount of fibronectin in exudates does not change over time; however, on the basis of Western blot analysis, it appears to be more degraded at early time periods most likely due to the action of SVMPs.
Taken together, our observations highlight a dual pattern of ECM protein degradation and appearance in exudates. Types IV and VI collagens, perlecan, nidogen and fibronectin show a higher degradation early on in the course of envenoming, correlating with the rapid action of SVMPs upon venom injection, as demonstrated by the proteinase activity of exudates. The rapid action of SVMPs on various key components of the BM is likely to be causally related to microvessel damage and hemorrhage. In the case of the fibrillar collagen I, it seems to be degraded predominantly by endogenous MMPs at later time periods, during the inflammatory reaction that ensues in the tissue as a consequence of venom-induced damage, as evidenced by zymography.
The presence of abundant plasma proteins in the exudate, as revealed by proteomic analysis, is a consequence of plasma exudation as a result of edema and increment in vascular permeability induced by the venom [68,69]. Some of the plasma proteins detected are acute-phase proteins, proteinase inhibitors and coagulation factors. Of interest is the increase of fibrinogen and the inter α-trypsin inhibitor heavy chains over time.
The presence of fibrinogen in exudates might be secondary to the inflammatory exudation induced by the venom since this protein is typically found in plasma at high concentrations [70]. Previous proteomic studies have found fibrinogen in the wound exudates induced by B. asper venom and its toxins, especially SVMP BaP1, early in the course of envenoming [14,15]. Our data show an increase of fibrinogen in wound exudates over time. This increment might be consequence of fibrin clot formation in capillary walls, due to vascular damage induced by SVMPs [12,14,35], and also to fibrin formation in the extravascular interstitial space, followed by fibrinolysis by endogenous proteinases [70], thus explaining their higher amounts in exudates collected at later time intervals.
The inter-α-trypsin inhibitor heavy chains are mainly secreted into the blood by the liver as serum protease inhibitor whose concentration increases in inflammatory conditions [71]. The effect of these proteins in tissues has been associated with both inflammatory and anti-inflammatory activities [71–73]. Moreover, these proteins can be covalently linked to hyaluronan, exerting functions on cell migration and ECM remodeling under physiological and pathological conditions [74,75]. Thus, the increase of inter-α-trypsin inhibitor heavy chains detected in exudates might be due to an acute-phase inflammatory response and to the tissue inflammation as a consequence of venom-induced damage.
In conclusion, the proteomic analysis of wound exudates performed in this study provides a more complete understanding of the time-course dynamics of muscle tissue damage induced by B. asper venom. These observations, together with Western blot and histology data, provide a more integrated view of venom-induced local tissue damage (Fig 5). The early presence of cytosolic and mitochondrial proteins in exudates, as compared to the later increase of cytoskeletal proteins, confirms the rapid cytotoxic effect of venom, followed by the action of endogenous proteinases in the cytoskeleton of damaged muscle fibers. On the other hand, the early presence of BM and other BM-associated ECM components in exudates, together with venom-derived proteolytic activity of exudates, strongly suggest the hydrolysis of these components by SVMPs in the early stages of envenoming. In contrast, the increment of some ECM proteins in the exudates at later time intervals is likely to be due to the action of endogenous MMPs or to their synthesis de novo during tissue remodeling associated with inflammation and reparative processes. Finally, the time-course of appearance in wound exudates of type IV collagen and perlecan supports the role of the hydrolysis of these BM components in the mechanism of microvascular damage induced by hemorrhagic SVMP.
|
10.1371/journal.pgen.1004111 | Mechanistically Distinct Mouse Models for CRX-Associated Retinopathy | Cone-rod homeobox (CRX) protein is a “paired-like” homeodomain transcription factor that is essential for regulating rod and cone photoreceptor transcription. Mutations in human CRX are associated with the dominant retinopathies Retinitis Pigmentosa (RP), Cone-Rod Dystrophy (CoRD) and Leber Congenital Amaurosis (LCA), with variable severity. Heterozygous Crx Knock-Out (KO) mice (“+/−”) have normal vision as adults and fail to model the dominant human disease. To investigate how different mutant CRX proteins produce distinct disease pathologies, we generated two Crx Knock-IN (K-IN) mouse models: CrxE168d2 (“E168d2”) and CrxR90W (“R90W”). E168d2 mice carry a frameshift mutation in the CRX activation domain, Glu168del2, which is associated with severe dominant CoRD or LCA in humans. R90W mice carry a substitution mutation in the CRX homeodomain, Arg90Trp, which is associated with dominant mild late-onset CoRD and recessive LCA. As seen in human patients, heterozygous E168d2 (“E168d2/+”) but not R90W (“R90W/+”) mice show severely impaired retinal function, while mice homozygous for either mutation are blind and undergo rapid photoreceptor degeneration. E168d2/+ mice also display abnormal rod/cone morphology, greater impairment of CRX target gene expression than R90W/+ or +/− mice, and undergo progressive photoreceptor degeneration. Surprisingly, E168d2/+ mice express more mutant CRX protein than wild-type CRX. E168d2neo/+, a subline of E168d2 with reduced mutant allele expression, displays a much milder retinal phenotype, demonstrating the impact of Crx expression level on disease severity. Both CRX[E168d2] and CRX[R90W] proteins fail to activate transcription in vitro, but CRX[E168d2] interferes more strongly with the function of wild type (WT) CRX, supporting an antimorphic mechanism. E168d2 and R90W are mechanistically distinct mouse models for CRX-associated disease that will allow the elucidation of molecular mechanisms and testing of novel therapeutic approaches for different forms of CRX-associated disease.
| The transcription factor Cone-Rod Homeobox (CRX) plays a central role in regulating gene expression of rod and cone photoreceptors, the primary light sensing cells of the retina. Mutations in the human CRX gene have been associated with the retinal degeneration diseases Retinitis Pigmentosa (RP), Cone-Rod Dystrophy (CoRD) and Leber Congential Amaurosis (LCA). These diseases cause progressive and permanent loss of vision, vary widely in age of onset and severity, and are currently untreatable. To understand how mutations in CRX cause distinct forms of retinal disease, we have genetically engineered mice to carry human disease-causing mutations in their Crx gene. These mouse lines accurately recapitulate distinct forms of CRX-associated disease, demonstrating that different classes of CRX mutations are responsible for phenotype variability in humans. We have characterized the pathology of these mice and identified critical mechanisms of disease. In addition, we have discovered that modifying the level of mutant protein had a dramatic effect on disease pathology in one mutant model, suggesting that targeted therapy against the mutant CRX could be an effective treatment strategy. These mouse models will allow for the testing of novel therapeutic strategies for retinal diseases caused by CRX mutations.
| CRX (Accession: AAH53672.1) is an Otd/OTX-like ‘paired’ homeodomain transcription factor that is preferentially expressed in vertebrate rod and cone photoreceptor cells in the retina and pinealocytes in the brain [1], [2]. CRX plays an essential role in the development and maintenance of functional mammalian rod and cone photoreceptors [3]. Previous studies show that CRX acts as a transcriptional activator [1][4]–[6] by interacting with co-activators, promoting histone acetylation at target gene promoters [7][8] and mediating enhancer/promoter intrachromosomal looping interactions [9] of target photoreceptor genes. Crx encodes a 299 amino acid protein that contains a homeodomain (HD) near its N-terminus that is responsible for DNA binding (Figure 1A) [1][10]. The HD is followed by glutamine rich (Gln), basic, WSP and OTX-tail motifs. The C terminal region of CRX (from the basic to the OTX-tail domains) is required for transactivation activity [4]. CRX interacts with transcription co-regulators including the rod-specific transcription factors NRL (Accession: NP_006168.1) [11][12], NR2E3 (Accession: AAH41421.1) [13][14], and general co-activator proteins GCN5, CBP and p300 (Accessions: AAC50641.1, AAC17736.1, NP_001420.2, respectively) [7] to coordinately control photoreceptor gene expression. In the homozygous Crx Knock-Out mouse (“−/−”), photoreceptors fail to form outer segments (OS), a highly specialized photoreceptor organelle which contains visual pigment opsins and other proteins required for phototransduction [15][16]. As a result, −/− photoreceptors do not function [3], form abnormal synapses [17], and undergo progressive degeneration [3]. Gene expression profile studies showed that −/− mice have severely reduced expression of many photoreceptor specific genes [18]–[20]. Most of these genes are direct CRX targets as detected by ChIP-seq analyses of the genomic CRX binding profile in the mouse retina [21].
Mutations in human CRX (NCBI Reference Sequence: NG_008605.1) have been associated with autosomal dominant forms of the retinal degenerative diseases Retinitis Pigmentosa (adRP), Cone-Rod Dystrophy (adCoRD) and Leber Congenital Amaurosis (adLCA), with different ages of onset and severity [12][22]–[45]. CRX is the only gene associated with all three diseases [22][23][26][43], demonstrating its central role in rod and cone integrity. However, null mutations in CRX may not be responsible for severe dominant disease. A null mutation in CRX, P9ins1, was associated with LCA in a heterozygous patient but the patient's father, a carrier of P9ins1, had a normal ocular phenotype suggesting either recessive or multigenic inheritance [44]. The heterozygous Knock-Out mouse (“+/−”), also shows only a slight delay in photoreceptor development and fails to model severe forms of dominant human disease [3]. The phenotypes of the human heterozygous null mutation and the +/− mouse phenotype suggest that haploinsufficiency is unlikely to underlie the severe forms of dominant CRX-associated disease.
Dominant disease-causing human CRX mutations primarily fall into two classes (Figure 1A): frameshift mutations (blue text) mostly in the transactivation domains and amino acid substitution mutations (black text) mostly within the DNA binding homeodomain. Both classes are expected to produce mutant forms of CRX protein that are pathogenic. Truncated CRX proteins resulting from the frameshift mutations E168d1, E168d2, A196d4 and G217d1 lost the ability to transactivate the promoter of Rhodopsin (Rho) in HEK293 cell transient transfection assays, but are expected to bind DNA normally since CRX 1–107, a complete activation domain truncation mutant, retained CRX target binding activity [4]. It was predicted that these truncated mutant proteins could interfere with the function of WT CRX by an antimorphic mechanism and cause a severe dominant retinal phenotype. Supporting this hypothesis, E168d1, E168d2, G217d1 and several other truncation mutations were linked to early onset (0–20 years) severe adCoRD/adLCA [22]–[33][36]–[40] and A196d4 was associated with adult onset adCoRD [42] . Furthermore, rescue experiments of the otduvi phenotype in Drosophila demonstrate the CRX truncation mutation I138fs48 possessed dominant-negative activity on target gene expression [46], providing experimental evidence for an antimorphic mechanism for this class of CRX mutations.
Four substitution mutations in the homeodomain: R41W, R41Q, R90W [11][45][47], and K88N [12], also reduced the ability of CRX to bind to and transactivate the Rhodopsin promoter. R41Q and R90W both reduced CRX:NRL protein interaction [11], while K88N additionally interfered with basal NRL-mediated transcription [12]. R41W, R41Q, and R90W were predicted to represent hypomorphic alleles associated with either recessive or less severe dominant forms of disease, while K88N was predicted to possess antimorphic activity on NRL function causing a stronger phenotype. Supporting this hypothesis, R41W, R41Q, R90W and several other substitution mutations were associated with late onset (∼40–60 years old) adCoRD [22][23][33][36][42][45], while K88N was associated with adLCA [12]. A patient homozygous for R90W was also diagnosed with autosomal recessive LCA [45]. In contrast, four other substitution mutations E80A [22][23][33][39], A56T [31], A158T and V242M [42] did not lose DNA binding or transactivating activity [47] and were associated with early onset adCoRD/LCA. In vivo rescue experiments in Drosophila also demonstrate that E80A but not R90W or K88N possesses some dominant-negative activity on Rh5 expression [46]. Collectively these experiments support our hypothesis that substitution mutations may cause disease through several distinct mechanisms.
Currently, there is no treatment strategy for CRX-associated diseases. Establishing animal models that accurately recapitulate different disease mechanisms is critical for developing and testing novel therapeutic approaches. Here we report the generation of two mechanistically distinct Knock-IN (K-IN) mouse models, each carrying a different class of CRX mutation, and present a detailed morphological, functional and biochemical characterization of these mouse models. The frameshift mutation E168d2 produces a severe dominant phenotype through an antimorphic mechanism, while the substitution mutation R90W produces a very mild late-onset ‘CoRD-like’ phenotype in heterozygotes and ‘LCA’-like disease in homozygotes. Furthermore, the expression level of a mutant allele can dramatically affect the disease phenotype, providing insight into potential treatment strategies.
In this study, we have generated two Crx K-IN mouse lines, each carrying a human disease-causing mutation in the mouse allele (Accession: NM_007770.4). CrxE168d2 (“E168d2”) mice carry a 2-bp deletion mutation, Glu168del2, which resulted in a codon frameshift and early truncation of the transactivation domains of CRX protein (Figure 1A–C). CrxR90W (“R90W”) mice carry Arg90Trp, an amino acid substitution mutation in the homeodomain of CRX (Figure 1A–C). An intermediate subline of each (“E168d2neo” and “R90Wneo”) carrying a neomycin (neo) cassette in intron 3–4 was also maintained (Figure 1B), since the neo cassette specifically reduced the expression of the mutant allele (Figure 2). The neo was removed from the germline by crossing E168d2neo and R90Wneo mice to the Sox2-Cre mouse [48] to generate the final E168d2 and R90W mouse lines (Figure 1B). Successful K-IN was confirmed by PCR amplification of neo (Primer set: Neo F/R) and the respective Crx allele (Table S1, Figure S1) and Sanger sequencing of homozygous mice (Figure 1C).
To determine if E168d2 and R90W K-IN mice properly express their respective CRX proteins, immunofluorescence (IF) staining for CRX was performed on paraffin-embedded retinal sagittal sections of P10 mice (Figure 2). The mouse monoclonal CRX antibody M02 (Abnova) used recognizes WT (Accession: NP_031796.1) and both mutant forms of CRX. Slides were immunostained in the same batch and imaged using a common exposure. As reported previously [13][19][49], CRX staining in WT retina (Figure 2A) was predominantly localized to the outer nuclear layer (ONL), comprised of the rod and cone photoreceptor cell bodies. Less intense CRX staining was also seen in the outer portion of the inner nuclear layer (INL), which is comprised of bipolar and horizontal cell bodies. E168d2 homozygous (“E168d2/d2”) and heterozygous (“E168d2/+”) mouse retinas showed higher intensity CRX staining than WT, especially in the ONL (Figure 2B&C). The heterozygous E168d2neo (“E168d2neo/+”) retina on the other hand showed similar intensity CRX staining as WT retina (Figure 2D vs 2A). In contrast, CRX staining in the ONL of R90W homozygous (“R90W/W”) and heterozygous (“R90W/+”) mouse retinas was reduced compared to WT retinas, although a few cells expressing high levels of CRX are scattered across the ONL (Figure 2E&F). This mosaic pattern of variable CRX expression was not seen in WT retinas. Crx Knock-Out (“−/−”) retinas didn't show CRX reactivity in the ONL and served as negative controls (Figure 2G). The positive CRX staining in E168d2/d2 and R90W/W retinas suggests that the CRX[E168d2] and CRX[R90W] mutant proteins were expressed in the appropriate cell layers.
The expression levels of WT CRX and mutant CRX[E168d2], CRX[R90W] proteins were compared and quantified in P10 E168d2 and R90W K-IN retinas using quantitative Western blots assayed with the polyclonal CRX 119b-1 antibody [7], which also recognized all forms of CRX proteins assayed. WT retina extracts showed a ∼37 kD band (Figure 2H, Lane 1). In contrast, a ∼27 kD dublet CRX band was detected in E168d2/d2 (Lane 2) and homozygous E168d2neo (“E168d2neo/d2neo”) (Lane 3) retinas, suggesting that the CRX[E168d2] protein was a truncated CRX protein as predicted by Sanger sequencing and genomic alignment (Figure 1C). Furthermore, the band intensities suggest that the amount of CRX[E168d2] protein in mutant retinas is higher than that of the full-length CRX in WT retinas (Figure 2H, Lanes 2&3 vs. Lane 1). Quantification of CRX protein levels (Figure 2I) revealed a significant genotype difference (p = 0.0002) overall. E168d2/d2 retinas made twice as much total CRX protein as WT retinas, while E168d2neo/d2neo retinas produce similar amounts of CRX protein as WT retinas.
Heterozygous E168d2/+ (Figure 2H, Lane 4) and E168d2neo/+ (Lane 5) mice expressed both full-length WT CRX and truncated CRX[E168d2] protein but in different ratios. Quantification of CRX protein in E168d2/+ retinal extracts (Figure 2I) revealed that the full-length WT CRX protein was present at approximately half of the level in WT retinas, but the level of CRX[E168d2] protein was more than twice that of the WT CRX. As a result, the total CRX protein level in these retinas was significantly increased by 2-fold compared to normal retinas. E168d2neo/+ retinal extracts also expressed WT CRX at approximately half WT levels but expressed less CRX[E168d2] protein than E168d2/+ retinas (Figure 2H, Lane 5 vs. 1&4, Figure 2I). As a result, the total CRX level in E168d2neo/+ was comparable to the WT control levels. These results are consistent with immunostaining results shown in Figure 2B–D and suggest that the E168d2 allele overproduces mutant protein, which was prevented by the presence of the neo cassette in E168d2neo.
CRX expression patterns in R90W mice differed from E168d2. In P10 R90W/W retinal extracts (Figure 2H, Lane 6; Figure 2I), CRX[R90W] was not significantly different from CRX in WT retinal extracts (Figure 2H, Lane 1; Figure 2I), while levels were reduced in R90Wneo/Wneo retinas (Figure 2H, Lane 7; Figure 2I). R90W/+ retinas (Figure 2H, Lane 8; Figure 2I) had normal total CRX protein levels compared to WT mice, although it was not possible to distinguish the quantity of WT CRX vs. CRX[R90W]. As seen with the E168d2 allele, the presence of the neo cassette reduced total CRX protein levels in R90Wneo/Wneo and R90Wneo/+ retinas, compared to corresponding R90W retinas (Figure 2H, Lane 7 vs. 6, Lane 9 vs. 8; Figure 2I). Thus, the presence of the neo cassette similarly affected the expression of both K-IN alleles.
To investigate whether the changes observed in CRX protein levels correlate with altered Crx mRNA transcription, Crx mRNA levels were determined by quantitative real-time reverse transcriptase PCR (qRT-PCR) (Figure 2J). Specific PCR primer pairs were used that selectively amplified sequences from either WT or total (WT+mutant) Crx cDNA (Primer sets: Crx E168WT F/R and Crx R90WT F/R; Table S1). Primer specificity was validated by amplification of WT, E168d2/d2 and R90W/W retinal cDNA preparations. The results show that E168d2/d2 retinas made twice as much total Crx mRNA as WT retinas, consistent with the elevated CRX protein levels in E168d2/d2. Total Crx mRNA levels in E168d2neo/d2neo retinas were lower than E168d2/d2 levels (FDR p = 0.07) but remained elevated relative to the WT (p<0.05) retinas, in contrast to the normal total CRX protein levels observed in these retinas.
E168d2/+ mice also showed moderately elevated total Crx mRNA levels (Figure 2J). Similar to protein levels, E168d2 mRNA levels (deduced from Total - WT) were much higher than WT levels (∼2∶1 ratio). By comparison, E168d2neo/+ mice expressed slightly elevated levels of total Crx mRNA that were lower than E168d2/+. WT and E168d2 alleles were evenly expressed in these retinas. These results are consistent with the differences in CRX protein levels, supporting an RNA-based mechanism for CRX[E168d2] overexpression, which was partially reversed in E168d2neo/+ mice.
R90W mice showed a distinct pattern of mRNA expression compared to E168d2. R90W/W retinas had normal Crx mRNA levels (Figure 2I), in contrast to their reduced CRX protein levels. This suggests a post-transcriptional mechanism either in the production or degradation of CRX[R90W] protein is likely responsible. Crx mRNA levels in R90Wneo/R90Wneo mice were substantially reduced in comparison to WT (p<0.05) and R90W/R90W mice (FDR p = 0.07). The R90W/+ and R90Wneo/+ mice showed essentially normal levels of total Crx mRNA, contributed either by both alleles equally (in R90W/+) or the WT allele predominantly (in R90Wneo/+). Together, our results suggest that E168d2 and R90W mRNA and corresponding proteins are produced in K-IN mouse retinas, but expression levels are differentially regulated. The mechanism of differential expression appears to be determined by features intrinsic to each mutant allele.
To determine the effect of E168d2 and R90W mutations on retinal morphology, paraffin embedded retinal sections from E168d2/d2 and R90W/W mice at P14, 1 month (mo) and 3 mo were stained with hematoxylin and eosin (H&E), imaged by light microscopy and compared to sections from WT and −/− mice [3][17]. Cell specification in WT retina is complete by P14 and three distinct neuronal layers are present: the ONL, INL and the ganglion cell layer (GCL) (Figure 3A). At P14 E168d2/d2, R90W/W and −/− retinas all had established normal cellular lamination (Figure 3B–D). Quantitative morphometric measures across the sagittal plane of the retina presented by ‘spider graphs’ (Figure 3M) did not show a genotype*distance interaction (the statistical threshold required to make individual comparisons when analyzing data with two-way ANOVA) (p = 0.15) at P14. These results support previous finding that CRX is not required for retinal cell fate specification [3], including rod photoreceptors, which constitute the majority of cells in the ONL. However, unlike WT retinas none of the mutant ONL cells had begun to form OS's at this age (Figure 3B, C, D vs. A). This OS defect persisted through 1–3 mo when OS's were fully formed in WT retina (Figure 3F, G, H vs. E; J, K, L vs. I). By 1 mo, loss of ONL nuclei was evident in all mutant retinas (Figure 3F–H). In comparison to the ∼12 rows of ONL nuclei seen in WT retinas, E168d2/d2 had only ∼3–4 rows, and R90W/W and −/− had ∼7–9 rows (Figure 3F, G, H vs. E). Quantification of ONL thickness shows photoreceptor degeneration occurred evenly across the sagittal plane of all mutant retinas (Figure 3N, red, green & blue lines vs. black). While R90W/W and −/− mice had similarly reduced ONL thickness (green and blue line, respectively), E168d2/d2 retinas showed greater ONL thinning at 1 mo (red line vs. green & blue), suggesting that degeneration was accelerated in these retinas. At 3 mo, all models exhibited greatly reduced ONL thickness (Figure 3O) with only ∼2–3 rows of ONL cells remaining (Figure 3J, K, L vs. I), suggesting ONL degeneration is progressive and extensive in all homozygous mutant mice.
To determine if ONL thinning is mediated by programmed cell death, “terminal deoxynucleotidyl transferase dUTP nick end labeling” (TUNEL) analysis was performed on P21 and P35 sagittal retinal sections (Figure S2). At P21 (Figure S2A–E), E168d2/d2, R90W/W and −/− mice all had significantly increased TUNEL+ cells present, almost exclusively in the ONL, E168d2/d2 exhibited the highest number of TUNEL+ cells (∼34 fold over WT). At P35 (Figure S2F–J), TUNEL+ cells remained elevated in the ONL of all mutant models but E168d2/d2 mice showed fewer TUNEL+ cells compared to R90W/W and −/− mice. There was no increase in TUNEL+ cells in other retinal layers of any of the mutant mice. These timecourse analyses suggest that the peak of ONL degeneration is earlier in E168d2/d2 mice compared to R90W/W and −/− mice, corresponding with the earlier ONL thinning observed in morphometric analyses.
To assess the consequence of these morphological changes on retinal function, electroretinograms (ERG) were performed under various light intensities on WT, E168d2/d2 and R90W/W mice at 1 month of age [50]. E168d2/d2 and R90W/W mice did not show any detectable dark-adapted or light-adapted responses (Figure S3). These results suggest E168d2/d2 and R90W/W mice are blind at young ages, similar to the phenotype reported for −/− mice [3]. The functional deficits of rod and cone photoreceptors in E168d2/d2 and R90W/W mice are consistent with the necessity of photoreceptor OS's for phototransduction [15][16] and suggest defective development of photoreceptor function in the homozygous mutant mice, similar to deficits in retinal function in LCA patients.
In spite of reduced Crx expression levels, homozygous mice from the sublines of each strain that carry a neo cassette (E168d2neo/d2neo, R90Wneo/Wneo) displayed retinal morphology and function (data not shown) that was indistinguishable from the respective neo-deleted line. Thus, in homozygous mice lacking WT alleles, the onset and rate of photoreceptor degeneration was not greatly affected by mutant protein expression level.
To determine the inheritance of E168d2 and R90W-associated phenotypes, retinal morphology of heterozygous E168d2/+, E168d2neo/+ and R90W/+ mice was assessed by histology and morphometry. Paraffin embedded sagittal retina sections of heterozygous mutant mice at P14, 1 mo, 3 mo and 6 mo were stained with H&E, imaged by light microscopy and compared to WT sections (Figure 4A–P). At P14, all retinas of heterozygous mutant mice displayed normal cellular lamination (Figure 4B–D vs. A). However, morphometric measurements of the ONL thickness showed that E168d2/+ had increased thickness at the two points most proximal to the optic nerve head (ON) (Figure 4Q, colored lines vs. black). E168d2/+ retinas also showed shortened rod OS's compared to WT (Figure 4B vs. A). The OS defect in E168d2/+ retinas remained at 1 mo (Figure 4F vs. E), 3 mo (Figure 4J vs. I) and 6 mo (Figure 4N vs. M). At 1 mo and 3 mo (Figure 4E–L, R–S), morphometric measurements of ONL thickness did not identify a significant genotype*distance interaction overall, therefore differences at each distance were not tested. However, at 3 mo, E168d2/+ had fewer rows of ONL cells ∼6–8 and had reduced mean ONL thickness at each distance. By 6 mo, most of E168d2/+ ONL cells had degenerated with only ∼2–3 rows of nuclei remaining (Figure 4N vs. M; Figure 4T, red vs. black line). By morphometric analyses, E168d2/+ exhibited reduced ONL thickness at all distances. These results suggest that E168d2/+ retinas undergo progressive rod photoreceptor degeneration through 6 mo of age. Consistent with this observation, TUNEL analysis showed at P35 E168d2/+ mice had 15-fold more TUNEL+ cells than WT all of which were located in the ONL (Figure S2L vs. K; Figure S2O), consistent with the observed photoreceptor degeneration phenotype. These results suggest that the E168d2 mutation causes dominant rod photoreceptor morphological defects and degeneration.
To determine if mice expressing lower levels of CRX[E168d2] protein have a less severe retinal phenotype, the morphology of E168d2neo/+ retinas was compared with that of E168d2/+ retinas. At P14, similar to E168d2/+ (Figure 4B), the OS's of E168d2neo/+ mice appeared shorter than in WT mice (Figure 4C vs. A). However, unlike E168d2/+, E168d2neo/+ formed fully elongated outer segments by 1 mo (Figure 4G vs. F), which were well maintained at 3 mo (Figure 4K vs. J) and 6 mo (Figure 4O vs. M). These results suggest that, despite a delay in maturation, E168d2neo/+ mice had less disrupted rod photoreceptor structure than E168d2/+. Furthermore, E168d2neo/+ did not show significant thinning of the ONL through 6 mo (Figure 4S&T, blue vs. black line) or elevated TUNEL+ cells compared to WT (Figure S2M vs. K; Figure S2O). Overall, the rod photoreceptor phenotype of E168d2neo/+ mice is mild compared to E168d2/+ mice, suggesting that E168d2 disease severity was influenced by the expression level of the mutant allele in heterozygous mice, consistent with E168d2 being an antimorphic mutation.
To further reveal morphological defects in E168d2 photoreceptors at the ultra-structural level, transmission electron microscopy (TEM) imaging analyses were performed on the retinas of P21 E168d2/+, E168d2neo/+ and WT mice (Figure 4U–W; Figure S4). Images were randomly coded for blinded data analysis. Compared to the morphology of WT OS's (Figure 4U), E168d2/+ mice (Figure 4V) exhibited severely shortened and disordered OS's including the presence of ‘wave-like’ disc patterns (white ‘*’s), ectopic vesicle formation (white ‘+’s), and improper stacking of OS discs including vertically oriented discs (white triangles). OS morphology was largely normal in E168d2neo/+ mice (Figure 4W); although minor ‘wave-like’ disc patterns and ectopic vesicle formation were occasionally seen.
Rod nuclei in P21 WT retina adopt a characteristic nuclear architecture with large areas of highly electron dense heterochromatin in the center and smaller regions of translucent euchromatin in the nuclear periphery [51] (Figure S4A&D). The chromatin pattern of E168d2/+ rods, however, appeared less condensed than WT (Figure S4B&E vs. A&D). This did not occur in E168d2neo/+ mice (Figure S4C&F vs. A&D). To quantify these changes, the percentage of the nuclear area comprised of condensed heterochromatin was measured in randomly selected WT, E168d2/+ and E168d2neo/+ rod nuclei. Figure S4G shows that the mean area of heterochromatin in E168d2/+ rods was significantly reduced by 8% compared to WT. This reduction in rod heterochromatin territory was not seen in E168d2neo/+ mice, suggesting more normal rod nuclear architecture. In addition, photoreceptor degeneration in E168d2/+ and E168d2neo/+ mice was evidenced by the presence of highly electron dense nuclei corresponding to pyknotic photoreceptor cells undergoing cell death, which were not observed in WT retinas (Figure S4E&F vs. G, white pentagon).
Unlike E168d2/+, R90W/+ mice had normal retinal morphology at all ages (Figure 4D, H, L&P), comparable to +/− mice [3]. They formed and maintained full-length OS's and normal ONL thickness (Figure 4H, L&P) through 6 mo of age. No increase in TUNEL+ cells over WT was detected (Figure S2N&O). These results suggest rod photoreceptor development and maintenance are normal in R90W/+ mice. This is consistent with clinical evaluations for heterozygous R90W carriers in human cases [22][23][33][42][45].
To determine how mutant forms of CRX protein affect target gene transcription, we assessed their ability to bind to DNA and transactivate transcription. First, electrophoretic mobility shift assays (EMSA) were used to measure DNA binding activity of CRX WT, CRX[E168d2] and CRX[R90W] protein expressed in HEK293 cells on the rhodopsin promoter target site BAT-1 [1] (Figure 10A). To compare relative binding affinity, the amount of CRX in each nuclear extract was quantified using Western blots and equalized between transfections (Figure 10B). EMSA was then performed on a 2-fold dilution series of nuclear extracts of each CRX protein. Following incubation with BAT-1 probe, WT CRX extract produced a single species of specific band shift (marked as ‘WT’) with a concentration-dependent intensity. This shifted band represented specific binding of the indicated CRX protein to BAT-1 CRX sites, as it is absent in the lane receiving the GFP control extract and when the probe contains mutated CRX binding sites (BAT-1 Mut AB). CRX[E168d2] nuclear extract also produced a specific band shift (marked ‘E168d2’), which migrated much faster than the full-length CRX band as expected for a truncated protein. The intensity of the E168d2 band was comparable to the WT full-length band at each corresponding concentration, suggesting that CRX[E168d2] binds target sites with similar efficiency as WT CRX, providing a basis for competition binding to common targets. In contrast, CRX[R90W] nuclear extract produced a faint band with the same mobility as WT (Figure 10A), but significantly reduced intensity (∼69% lower than WT). Reduced but not abolished DNA binding activity was also reported for bacterially expressed CRX homeodomain peptides carrying the R90W mutation [47]. These results support the hypothesis that CRX[E168d2] protein maintains normal DNA binding ability, while CRX[R90W] protein has reduced DNA binding ability.
To determine if in vitro DNA binding activity of each mutant reflected ability to associate with target chromatin in vivo, the association of WT CRX, CRX[E168d2] and CRX[R90W] protein with target gene promoter regions was examined using chromatin immunoprecipitation (ChIP) assays. ChIP was performed on P10 mouse retinas of WT, E168d2/d2, R90W/W and −/− mice using the CRX 119b-1 antibody [7]. As expected, enrichment of CRX[E168d2] protein was detected on the promoter of genes expressed in rods (Rho, Gnat1), cones (Arr3, Opn1mw, Opn1sw) and both rods/cones (Crx, Rbp3 (Accession: AJ294749.1)) (Figure 10C, red bars). Despite reduced DNA binding activity in vitro, CRX[R90W] protein was found on the promoter of all candidate genes tested (Figure 10C, green bars). The mechanism by which CRX[R90W], which has reduced DNA-binding ability, is recruited to target gene chromatin in vivo remains to be determined. However, these results are consistent with R90W's hypomorphic effect on target gene expression in the retina (Figure 8, Figure 9).
The ability of CRX[E168d2] and CRX[R90W] proteins to transactivate target promoters, either alone or in combination with WT CRX, was assessed by dual-luciferase reporter assays in transiently transfected HEK293 cells. Consistent with a previous report [47], WT CRX was able to cooperate with NRL to activate a Rhodopsin promoter-driven luciferase reporter, BR130 (Figure 10D). However, CRX[E168d2] failed to increase transactivation above NRL alone, suggesting that CRX[E168d2] was unable to form functional interactions with transcription co-activators despite its normal DNA binding ability. In contrast, CRX[R90W] weakly promoted NRL-mediated transactivation, consistent with CRX[R90W]'s weak ability to bind target DNA (Figure 10A) and interact with NRL [11] in vitro to promote low levels of gene expression in the retinas of homozygous R90W mice (Figure 8). To test the effect of mutant protein on WT CRX function, E168d2 and R90W expression vectors were each co-transfected at increasing concentrations with WT CRX. CRX[E168d2] protein significantly impaired WT CRX function when the ratio of E168d2:WT vector reached 2∶1 or higher, suggesting CRX[E168d2] actively interfered with WT CRX via an antimorphic mechanism, consistent with the dose-dependent toxicity observed in E168d2/+ and E168d2neo/+ mice. In contrast, at the same mutant:WT vector ratios, CRX[R90W] protein did not disrupt WT protein function, consistent with the hypomorphic effect of R90W in mice.
The Crx promoter is another known CRX direct target. It contains two CRX consensus binding sites within a 500-bp upstream region that is required for CRX auto-activation [57]. However, unlike Rhodopsin, which is downregulated, Crx was overexpressed in E168d2 mice (Figure 2). To determine if Crx overexpression resulted from the direct action of CRX[E168d2] protein on the Crx promoter, dual-luciferase reporter assays using the 0.5K Crx promoter were performed (Figure 10E). As expected, WT CRX protein transactivated this in a concentration-dependent manner (Figure 10E), while CRX[E168d2] and CRX[R90W] at the highest concentration did not transactivate. When both WT and mutant proteins were present, CRX[E168d2] interfered with the transactivation activity of WT CRX, even at a 1∶2 mutant:WT vector ratio. CRX[R90W] protein also reduced WT CRX transactivation activity, though less strongly, at the 1∶1 and 2∶1 mutant:WT vector ratios. These results suggest that both CRX[E168d2] and CRX[R90W] proteins 1) are less effective than WT CRX at activating target promoters, and 2) interfere with WT CRX autoactivation.
Taken together, functional analyses of CRX[E168d2] and CRX[R90W] proteins revealed that they affected target gene transcription via distinct mechanisms. While CRX[E168d2] bind DNA equally well as WT CRX, it fails to activate transcription and interferes with WT CRX function, resulting in a dose-dependent antimorphic effect. In contrast, CRX[R90W] has reduced ability to bind target DNA and regulate transcription, qualifying CRX[R90W] as a hypomorphic protein.
Several pieces of evidence support that CRX[E168d2] and CRX[R90W] protein cause disease via different mechanisms, as illustrated in Figure 11. CRX[E168d2] protein bound to DNA, interfered with the function of CRX WT and impaired the expression of CRX target genes, classifying it as an antimorphic protein with dominant-negative activity (Figure 11B). All of our results suggest that CRX[E168d2]'s activity was largely restricted to CRX target genes. Of the 82 uniquely downregulated genes identified in homozygous E168d2neo mice, most (76.8%) also exhibited direct CRX binding. The average fold change of these distinct genes was less dramatic than genes shared between E168d2 and −/−, suggesting they were likely to be similarly affected in −/− but failed to pass the significance threshold. Many shared genes including: Rho, Arr3, Ramp3, Drd4, Cpm, and Pde6c were more strongly downregulated in homozygous E168d2neo than −/− mice (Figure 8D). This suggests CRX[E168d2] protein had an antimorphic effect on the expression of these genes even in the complete absence of WT CRX, possibly by interfering with other co-factors like the homeodomain transcription factor OTX2. Supporting this hypothesis, removal of one allele of Otx2 from the −/− mouse produced a severe phenotype similar to homozygous E168d2 mice [63]. Since OTX2 and CRX have overlapping spatial and temporal roles in retinal development and share DNA binding domain homology, it is possible that CRX[E168d2] interfered with OTX2 activity, resulting in a stronger phenotype than −/−. This antimorphic effect is unlikely to involve interference with NRL function, since NRL expression was comparable in all homozygous models (Table S5), CRX[E168d2] did not interfere with NRL transactivation (Figure 10D) and a similar truncation mutation in bovine CRX C160 (1–160) maintained interaction with NRL [11]. qRT-PCR analysis of CRX target gene expression showed downregulation that correlated with mutant CRX expression level (Figure 9), supporting the conclusion that CRX[E168d2] is an antimorphic mutant protein with dominant negative activity. The E168d2 mouse model thus demonstrates the effects of an antimorphic truncated CRX protein associated with human disease.
The CRX[R90W] protein had reduced DNA binding and weakly promoted transcription in vitro, classifying CRX[R90W] as a hypomorphic protein (Figure 11C&F). Although binding of CRX[R90W] to the BAT-1 oligo in vitro was reduced (Figure 10A), CRX[R90W] associated with CRX target DNA in vivo (Figure 10C), suggesting co-factors may anchor CRX[R90W] to target DNA. CRX[R90W] weakly promoted NRL-mediated transactivation of the Rho promoter in vitro (Figure 10D), consistent with early findings that CRX[R90W] protein reduced the physical interaction with NRL [11]. Thus, even though CRX[R90W] was associated with target promoters in vivo, it may have lost specific interactions with co-factors, therefore reducing its function. Indeed, despite being present on target promoters, CRX[R90W] only weakly promoted target gene expression in vivo, as shown by reduced expression of many CRX target genes in homozygous R90W retinas as detected by microarray (Figure 8A–E) and qRT-PCR (Figure 8J–M). However, target gene expression in R90W retinas was less reduced compared to −/− retinas (Figure 8D, J–M), suggesting CRX[R90W] possessed some residual transcriptional activation activity. In Drosophila, human R90W was able to partially rescue the otduvi phenotype, consistent with a hypomorphic mechanism [46]. Taken together, our results show that CRX[R90W] is a predominantly hypomorphic mutant CRX protein, representative of substitution mutations associated with mild forms of CRX disease.
The molecular functions of several CRX mutations associated with human retinopathy have been investigated in vitro [12][45][47] and in vivo in Drosophila [46]. Such studies indicate that mutant CRX proteins have distinct molecular functions, which could in part explain the variation in CRX-disease phenotypes. The distinct phenotypes of mice carrying E168d2, an antimorphic frameshift mutation, and R90W, a hypomorphic substitution mutation, further expand our understanding of the impact of mutation type on disease pathology and closely match the functions and associated phenotypes of other similar type mutations. This suggests that E168d2 and R90W K-IN mice are representative animal models for two larger groups of disease causing mutations, increasing their utility as research tools for studying pathology and developing therapies. There are likely additional mechanisms of CRX-associated disease yet to be modeled in vivo, such as substitution mutations that do not affect DNA-binding but are nonetheless associated with dominant disease [12][45][47]. Collectively, these studies demonstrate the diversity of molecular defects mediating CRX-associated disease and highlight the value of having multiple small-animal models to understand them.
Currently, there are no treatment strategies for CRX-associated diseases. Since CRX influences many cellular processes, designing targeted therapy is exceptionally difficult. The availability of phenotypically and mechanistically distinct models for CRX-associated disease will greatly improve our ability to develop novel therapies. E168d2, E168d2neo and R90W present unique mechanistic challenges for therapy to address. Stem cell based therapies have previously been shown to restore function in the −/− mouse [64]. Like −/− mice, E168d2/+, E168d2/d2 and R90W/W mice all have highly abnormal photoreceptor morphology and undergo rapid degeneration, which may restrict the time course and effectiveness of treatment. The improved phenotype of E168d2neo/+ mice, compared to E168d2/+, provides evidence that gene replacement strategies that shift the ratio of WT to mutant CRX could be effective at improving vision and promoting rod and cone survival in cases were a mutant protein is toxic and/or overexpressed. Previous studies have shown this strategy to be effective in treating a dominant-negative adRP RHO animal model [65][66]. Lastly, the similarity of the E168d2/+ mouse and the Rdy/+ cat provide excellently matched small and large animal models. Therapies that are proven to be effective in the E168d2/+ mouse can immediately be tested in the Rdy/+ cat, which improves our ability to develop translational therapies.
In summary, Crx E168d2 and R90W are mechanistically distinct mouse models for CRX-associated disease, demonstrating how different classes of CRX mutations yield drastically different retinal phenotypes. E168d2 and R90W accurately recapitulate human diseases caused by distinct classes of human mutations and have greatly improved our understanding of disease pathobiology. The availability of these stratified mouse models for CRX-associated disease is an invaluable resource for developing effective mechanism based therapies.
All procedures involving mice were approved by the Animal Studies Committee of Washington University in St. Louis, and performed under Protocols # 20090359 and 20120246 (to SC). Experiments were carried out in strict accordance with recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health (Bethesda, MD), the Washington University Policy on the Use of Animals in Research; and the Guidelines for the Use of Animals in Visual Research of the Association for Research in Ophthalmology and Visual Science (http://www.arvo.org/animals/). Every effort was made to minimize the animals' suffering, anxiety, and discomfort.
Mice were housed in a barrier facility operated and maintained by the Division of Comparative Medicine of Washington University School of Medicine. All mice used for experiments were backcrossed to C57BL6/J mice obtained from Jackson Laboratories (Bar Harbor, ME, Stock number 000664) for at least 5 generations. Knock-IN of E168d2neo and R90Wneo were generated by the Mouse Genetics Core, Department of Ophthalmology and Visual Sciences, Washington University (Saint Louis, MO). E168d2neo and R90Wneo constructs were transfected into 129Sv/J SCC#10 (ATCC SCRC-1020) embryonic stem cells and Knock-IN was achieved by homologous recombination into the endogenous mCrx locus and selected by neomycin. The targeted ES cells were injected into C57BL6/J blastocysts to form chimeric Knock-IN E168d2neo and R90Wneo mice. Germline transmission of E168d2neo and R90Wneo was identified by PCR genotyping and Sanger sequencing of genomic DNA from F1 mice (Figure 1, Figure S1, Table S1). Crx−/− mice were provided by Dr. Constance Cepko, Harvard University (Boston, MA).
Genomic DNA was prepared from mouse tail tissue using the Gentra Puregene Tissue Kit (Qiagen). PCR amplification was performed using Jumpstart RedTaq (Sigma-Aldrich). Primer sets (Table S1) are as follows: For all mice: neo (Neo-F/R) and Crx (Total Crx-F/R); for E168d2 colony: WT Crx allele (E168d2 WT-F, E168d2-R), E168d2 allele (E168d2 Mut-F, E168d2-R); for R90W colony: WT Crx allele (R90W WT-R, R90W-R), R90W allele (R90W Mut-F, R90W-R).
Genomic DNA was prepared from mouse tail tissue using the Gentra Puregene Tissue Kit (Qiagen). mCrx DNA was amplified by PCR using the Genomic mCrx Int/Ex4-F/R primer pair (Table S1). Sanger sequencing was performed by the Protein and Nucleic Acid Chemistry Laboratory (Washington University, Saint Louis, MO) using the Sequencing primers E168 and R90W (Table S1) and Big Dye V3.1 (Advanced Biotechnologies).
At least 5 mice of each genotype were tested for ERG at 1 mo, 3 mo, or 6 mo of age. Bilateral flash ERG measurements were performed using a UTAS-E3000 Visual Electrodiagnostic System running EM for Windows (LKC Technologies, Inc., Gaithersburg, MD) and recordings from the higher amplitude eye were used for analysis. Mice were dark-adapted overnight, then anesthetized with 80 mg/kg ketamine and 15 mg/kg xylazine under dim red illumination for electrode placement and testing. Body temperature was maintained at 37±0.5°C with a heating pad controlled by a rectal temperature probe (FHC Inc., Bowdoin, ME). The mouse's head was positioned just inside the opening of the Ganzfeld dome and pupils were dilated with 1.0% atropine sulfate (Bausch & Lomb, Tampa, FL). The recording electrode was a platinum loop 2.0 mm in diameter, positioned in a drop of 1.25% hydroxypropyl methylcellulose (GONAK; Akorn Inc., Buffalo Grove, IL) on the corneal surface of each eye. The reference needle electrode was inserted under the skin at the vertex of the skull. The ground electrode was inserted under the skin of the mouse's back or tail. The stimulus (trial) consisted of a brief, full-field flash (10 µs) either in darkness, or in the presence of dim (29.2 cd/mm) background illumination after 10 minutes adaptation time to the background light. The initiation of the flash was taken as time zero. The response was recorded over 250 ms plus 25 ms of pre-trial baseline. Responses from several trials were averaged. For complete test parameters see Table S7. The log light intensity (log [cd*s/m2]) was calculated based on the manufacturer's calibrations. The mean amplitudes (in microvolts) of the averaged dark-adapted A and B-waves and light-adapted B-waves were measured and quantified for comparison. The between-group differences in peak amplitude were determined by testing genotype*flash intensity interactions (p<0.05, n≥5) at each age were compared using two-way ANOVA for repeated measurement data to account for potential correlations among readings from the same mice. If the overall genotype*flash intensity interaction was significant, post-hoc multiple comparisons for differences between each genotype and the control group at each light intensity level were performed. All the tests were two-tailed, significance: p<0.05. The statistical analysis was performed using SAS 9.3 (SAS Institutes, Cary, NC). p-values were adjusted for multiple comparisons by a permutation test using the default parameters provided in the LSMestimate statement in Proc Mixed. Average percent reductions for each wave form were calculated by normalizing the peak amplitude of the mutant to WT and results were averaged for the flashes listed in Table 1; ±STDEV.
For retinal sections: eyes were enucleated by removing the cornea and lens and fixed in 4% paraformaldehyde for 24 hrs at 4°C. A small corneal tag on the superior portion of the eye was used for orientation. Eyes were embedded in paraffin and 5 µM sagittal retinal sections were cut using a Leica RM 2255 microtome as previously described [67]. Hemotoxylin and eosin immunohistochemistry was performed on sections for histology. Fluorescent antibody immunostaining was performed using as previously described using 1% BSA/0.1% Triton X in 1× PBS for blocking and antigen retrieval for all samples [13][67].
For whole flat-mounted retinas: eyes were enucleated by removing the cornea and lens and fixed in 4% paraformaldehyde for 1 hr at 4°C. Retinas were then dissected from the eye cup and 4 evenly spaced relief lines were cut (Figure 6A). A scleral tag was left on the superior retina for orientation. Retinas were mounted on poly-D lysine coated slides (Thermo Scientific), blocked with 1% BSA/0.1% Triton X in 1× PBS and immunostained as previous.
Primary antibodies and dilutions used as follows: Mouse monoclonal anti-CRX M02 (1∶200, Abova), rabbit anti-CRX 261 (1∶200), rabbit anti-cone arrestin (CARR) (1∶1000, Millipore), Rabbit anti-Opsin Red/Green (MOP) (1∶1000, Millipore), Goat anti-OPN1SW (N-20) (SOP) (1∶500, Santa Cruz), Mouse anti-Rhodopsin RET-P1 (RHO)(1∶400, Chemicon), Peanut Agglutanin conjugated to Rhodamine (PNA)(1∶500, Vector Labs). Secondary Antibodies (1∶400): Goat anti-rabbit or mouse IgG antibodies coupled to Alexa Fluor A488, Rhodamine 568 or Cy2 647 (Molecular Probes) and Chicken anti-goat IgG (Molecular Probes). All slides were counterstained with hard set DAPI (Vectashield), except when using Cy2 secondary, which were counterstained with Slow Fade Gold DAPI (Invitrogen). All brightfield and fluorescent imaging was performed using an Olympus BX51 microscope and Spot RT3 Cooled Color Digital camera (Diagnostic instruments inc.).
TUNEL analysis was performed using the Apoptag Fluorescein in situ Apoptosis Detection Kit (Millipore) per kit instructions. TUNEL+ cells were counted in retinal sagittal sections of P21 and P35 mice. Significant differences from WT control (p<0.05) were determined by the Kruskal-Wallis rank order test, which was used to protect against departures from the normal distribution assumption.
For ONL morphometry, 20× retinal composites of hematoxylin and eosin (H&E) stained sagittal sections were analyzed using Image J software (http://rsb.info.nih.gov/ij/). The distance from the Optic Nerve (ON) was determined by drawing a curved line along the outer limiting membrane. The ONL thickness was measured at 100 µM, 500 µM, 1000 µM, and 1500 µM from the ON and 200 µM from the peripheral edge on both the superior and inferior retina. Results are presented by ‘spider graph’. The between-group differences in ONL thickness were determined by testing overall genotype*distance interactions (p<0.05, n≥3) at each age were tested using two-way ANOVA for repeated measurement data, followed by a post-hoc test to adjust p-value for multiple comparisons between each genotype and the WT control group at each distance using SAS 9.3 (SAS Institutes, Cary, NC), as above.
Cone nuclear localization was determined by immunostaining retinal sections with CARR. The ONL was divided into 3 equally sized zones (OONL, MONL, IONL; Figure 5A) on 20× retinal composite images using Image J software (http://rsb.info.nih.gov/ij/) and the cone nuclei within in each zone from three sections for each mouse were counted. Significant differences from WT for each zone were determined by Kruskal-Wallis rank order test (p<0.05, n≥3)
For cone density and opsin expression assessment, 10 images at 40× magnification of whole flat-mounted retinas were taken in the zones specified in Figure 6A. All peripheral images were taken ∼400 µM from the edge of the retina and the central image was taken ∼250 µM from the ON along the lateral axis. Cones were counted within a 200×200 µM square grid for each image using Image J software and the density of cones/(mm2*1000) was calculated. The between-group differences in cone density were determined by testing overall genotype*retinal region interactions (p<0.05, n≥3) at each age were tested using two-way ANOVA for repeated measurement data, followed by a post-hoc test to adjust p-value for multiple comparisons between each genotype and the WT control group in each retinal region using SAS 9.3 (SAS Institutes, Cary, NC), as above. For regional cone opsin expression analysis (Figure 6E–P), differences in the fraction of cones expressing SOP, MOP, SOP/MOP or no opsin was tested in each region using a Kruskal-Wallis rank order test (p<0.05).
For TEM studies, eyes were enucleated by removing the cornea and lens and fixed in 2% paraformaldehyde/3% gluteraldehyde in 0.1 M phosphate buffer (pH 7.35) for 24 hrs, post-fixed in 1% osmium tetroxide for 1 hr and stained en bloc with 1% uranyl acetate in 0.1 M acetate buffer for 1 hr. Blocks were then dehydrated in a graded series of acetones and embedded in Araldite 6005/EMbed 812 resin (Electron Microscopy Sciences). Semi-thin sections (0.5–1 µm) were cut through the entire retina at the level of the optic nerve and stained with toluidine blue, post-stained with uranyl acetate and lead citrate, viewed on a Hitachi H7500 electron microscope and documented in digital images. Three retinas for each genotype were sampled at P21 at 800–1200 µM from the optic nerve. ≥10 images of four key features were collected by random sampling: OS-RPE (10000×), OS-IS (12000×), ONL (5000×), OPL (10000×). Images were analyzed in a blinded manner using Image J software.
The nuclear percent area of heterochromatin was measured using Image J software in a randomized blinded analysis. For each genotype, 10 5000× images of the ONL were taken for three mouse retinas. For each image, 10 rod nuclei were randomly selected for analysis. The rod nucleus was outlined using the segmented polygon tool, electron dense regions of the nuclei associated with heterochromatin were thresholded and the percentage of the area above the threshold was measured. Thresholding was variably adjusted to accommodate for differences in brightness and contrast. The between-group differences were compared using one-way ANOVA for repeated measurement data, to account for potential correlations among photos from the same mouse. All the tests were two-tailed, significance: p<0.05 (n = 3). The statistical analysis was performed using SAS 9.3 (SAS Institutes, Cary, NC). The overall test for genotype difference was statistically significant (p = 0.02), therefore E168d2/+ and E168d2neo/+ were compared to WT (Figure S2).
HEK293 cells (ATCC CRL-11268) were cultured on 60 mm plates in Dulbucco's minimum essential media (DMEM) with 10% fetal bovine serum and Penicillin/Streptomycin. Cells in 60% confluence were transfected with pCAGIG-NRL and pCAGIG-hCRX WT, E168d2 and R90W either alone or in combination using CaCl (0.25 M) and Boric Acid Buffered Saline (1×) pH 6.75 as previously described [13]. Cells were harvested 48 hours post transfection for either RNA (PerfectPure RNA tissue kit, 5Prime), protein (NePER nuclear and cytoplasmic extractions reagents, Thermo Scientific), or Dual-luciferase assays. Dual-luciferase assays were performed as previously described [13]. Significant differences from pcDNA3.1hisc control were determined by Kruskall-Wallis rank order test (p<0.05; n = 3). Post-hoc comparisons (Figure 10 D&E; indicated by brackets) were tested using a less conservative FDR p-value method for multiple comparisons using PROC Multtest of SAS (V9.3). FDR p<0.09 was considered marginally significant.
Whole retina protein lysates were prepared by homogenization of four genotype-matched isolated whole retinas from P10 mice and lysis in 1× RIPA buffer (Sigma) for 10 min with protease inhibitors (Aprotinin, Leupeptin, peptistatin, 0.1 mM Phenylmethaneslfonyl fluoride). Nuclear lysates were prepared using NE-PER Nuclear and Cytoplasmic Extraction Reagants (Thermo Scientific) with protease inhibitors. Either 30 µg of whole protein lysate or 5 µg of nuclear protein lysate was boiled for 10 min. Samples were run on a 4–11% SDS-PAGE gel and transferred onto Transblot Turbo nitrocellulose membranes (Bio-Rad) using the Transblot Turbo system (Bio-Rad). Membranes were probed with Rabbit anti-CRX 119b1 (1∶750) and Mouse anti-β-Actin (Sigma)(1∶1000). Goat anti-Mouse IRDye 680LT and Goat anti-Rabbit IRDye 800CW (LI-COR) were used as secondary antibodies. Signal was detected and quantified using the Odyssey Infrared Imager (LI-COR) and associated manufactory software. Kruskal-Wallis rank order test (Proc Npar1way of SAS, V9.3) was used to test for an overall difference among genotypes (p = 0.0002), then each genotype was compared to WT control (p<0.05). Post-hoc analyses were performeded using FDR p methods for multiple comparisons using PROC Multtest of SAS (V9.3) (FDR p<0.09) (n≥3).
RNA was extracted from whole retinas of one male and one female mouse at either P10 or P21 for each biological replicate using the PerfectPure RNA tissue kit (5Prime). RNA was quantified using a NanoDrop ND-1000 spectrophotometer (NanoDrop Technologies, Wilmington, DE). cDNA was synthesized from 1 µg of RNA using the Transcriptor First Strand cDNA Synthesis kit (Roche Applied Science). A 10 µl QRT-PCR reaction mixture containing 1× EvaGreen with Low Rox reaction mix (BioRad), 1 µM primer mix, and diluted cDNA was prepared and run on a two-step 40 cycle amplification protocol with melt curve determination on a BioRad CFX thermocycler in triplicate. The Cq's of technical replicates were averaged and the results were analyzed using the Delta Cq method in QBase software (Biogazelle). Primer sets (Table S1) were designed using MacVector software and synthesized by IDT DNA technologies. For mCrx allele specific amplification the following primers were used: for E168d2 and E168d2neo mice: WT allele specific- Crx E168d2 WT RTF/R, total- Crx R90W WT-RTF/R; for R90W and R90Wneo mice: WT allele specific- Crx R90W WT-RTF/R, total Crx E168d2 WT RTF/R (Figure 2J), Relative gene expression was normalized to Ubb and Tuba1b. Kruskal-Wallis rank order test (Proc Npar1way of SAS, V9.3) was used to test for an overall difference among genotypes (p<0.05; n≥3). Post hoc analyses were adjusted for multiple comparisons using FDR p methods, as above (FDR p≤0.09).
Triplicate RNA samples were prepared from 4 pooled retinas from 1 male and 1 female mouse at P10 for WT and homozygous E168d2neo, R90Wneo and −/− mice. The RNA was fluorescent labeled and hybridized to MouseWG-6 v2.0 Expression Beadchips (Illumina) by Washington University Genome Technology Access Center (GTAC). The raw microarray datasets are available at the NCBI GRO website (http://www.ncbi.nlm.nih.gov/gds, access number: GSE51184). Microarray data were analyzed using significance analysis of microarrays (SAM) following background subtraction and quantile normalization in Illumina Genome Studio platform. Control probes and probes with detection p-value <0.05 across all samples were removed prior to any analysis. Candidate probes with 2.0-fold disregulation at false discovery rate ≤0.05 from each comparison were chosen for further analysis. Cellular processes associated with differentially expressed genes were assigned based on gene ontology provided by Mouse Genome Informatics (http://www.informatics.jax.org/).
BAT-1 and BAT-1 mutated AB probes 5′ end-labeled with 700 IRDye were synthesized by Integrated DNA Technologies (IDT). Nuclear protein extracts from HEK293 cells (∼1×108 cells) transfected with pCAGIG-hCRX, pCAGIG-hCRX E168d2, or pCAGIG-hCRX R90W were prepared following NE-PER kit instructions (Thermo Scientific). Nuclear extracts were tested for CRX expression by running on a Western Blot as above (Figure 10B). CRX levels were quantified by normalizing to β-Actin (Sigma) and a 2-fold dilution series of equivalent amounts of CRX WT, CRX[E168d2] and CRX[R90W] protein were used for binding reactions. Binding reactions were performed using the Odyssey Infrared EMSA kit (LI-COR), per kit instructions using 1 µg of nuclear protein extract and 50 nM IRDye labeled oligo. Samples were run on a native 5% polyacrylamide; 0.5× Tris/Borate/Ethylenediaminetetraacetic acid (EDTA) buffered gel and imaged on the Odyssey Infrared Imager (LI-COR).
ChIP was performed as previously described [7][13][68]. Basically, 6 retinas per sample were dissected and chromatin was cross-linked with 1% formaldehyde in PBS for one minute at room temperature. After cell lysis and chromatin fragmentation by sonication, chromatin fragments were immunoprecipitated with the CRX 119b-1 antibody [7] or normal rabbit IgG (Santa Cruz) bound to Protein A beads (GE Healthcare Life Sciences, Piscataway, NJ). After extensive washing, the immunoprecipitated chromatin was eluted with 50 mM NaHCO3 1% SDS, heated to 67°C to reverse the cross-links, the DNA purified by ethanol precipitation and analyzed by PCR with gene-specific primers (Table S1) (n≥3). Fold enrichment was determined by quantitative ChIP PCR. Critical threshold (Ct) values for CRX and IgG immunoprecipitation (IP) were normalized to input and mock subtracted. The fold enrichment of CRX:IgG was calculated based on the formula shown below. Significant enrichment was determined by testing overall promoter*genotype interactions by two-way ANOVA for repeated measures using SAS 9.3 (SAS Institutes, Cary, NC) (p<0.05, n = 3), as above.
ΔCt = (Ct[CRX or IgG]-Ct[Input])
ΔΔCt = ΔCt[CRX or IgG]- ΔCt[mock]
Fold enrichment = ((2−ΔΔCt CRX)/(2−ΔΔCt IgG)
|
10.1371/journal.ppat.1000565 | Trypanosoma brucei PUF9 Regulates mRNAs for Proteins Involved in Replicative Processes over the Cell Cycle | Many genes that are required at specific points in the cell cycle exhibit cell cycle–dependent expression. In the early-diverging model eukaryote and important human pathogen Trypanosoma brucei, regulation of gene expression in the cell cycle and other processes is almost entirely post-transcriptional. Here, we show that the T. brucei RNA-binding protein PUF9 stabilizes certain transcripts during S-phase. Target transcripts of PUF9—LIGKA, PNT1 and PNT2—were identified by affinity purification with TAP-tagged PUF9. RNAi against PUF9 caused an accumulation of cells in G2/M phase and unexpectedly destabilized the PUF9 target mRNAs, despite the fact that most known Puf-domain proteins promote degradation of their target mRNAs. The levels of the PUF9-regulated transcripts were cell cycle dependent, peaking in mid- to late- S-phase, and this effect was abolished when PUF9 was targeted by RNAi. The sequence UUGUACC was over-represented in the 3′ UTRs of PUF9 targets; a point mutation in this motif abolished PUF9-dependent stabilization of a reporter transcript carrying the PNT1 3′ UTR. LIGKA is involved in replication of the kinetoplast, and here we show that PNT1 is also kinetoplast-associated and its over-expression causes kinetoplast-related defects, while PNT2 is localized to the nucleus in G1 phase and redistributes to the mitotic spindle during mitosis. PUF9 targets may constitute a post-transcriptional regulon, encoding proteins involved in temporally coordinated replicative processes in early G2 phase.
| The unicellular protozoan Trypanosoma brucei is the causative agent of African sleeping sickness, responsible for over 100,000 deaths annually, and is related to other important pathogens (e.g. Leishmania major and Trypanosoma cruzi). Unusually, these organisms do not regulate their genes by changing the rate at which they are copied into RNA, but by changing the rate of RNA destruction or the rate of translation into protein. We identified an RNA-binding protein, PUF9, responsible for the accumulation of several RNA molecules at a specific time point in the cell division cycle, just after DNA replication. Correspondingly, the proteins encoded by these RNAs appear to function in the division of various cellular structures at this time point or shortly afterwards. Two of them facilitate replication of the kinetoplast (an organelle containing the mitochondrial DNA) while another was found in the mitotic spindle. Their temporal co-expression may stem from another unusual feature of trypanosomes: only one copy of the kinetoplast (and several other organelles) are present per cell, their replication being coordinated with cell division. Indeed, PUF9 may be important in the control of organelle copy-number because suppression of PUF9 resulted in cells with too many kinetoplasts, flagella, or nuclei.
| The eukaryotic cell cycle is an ordered program of coordinated processes that mediate the replication of key cellular structures and their subsequent distribution to daughter cells. The proteins involved are often expressed at specific points during the cell cycle, for example dihydrofolate reductase (DHFR) [1] and histones [2], which are required before and after DNA synthesis, respectively, and whose regulation has been intensively studied. In most eukaryotes, such regulation might be expected to occur through transcriptional regulation. However, post-transcriptional regulation of mRNA, particularly in terms of differential message stability, is also important in ensuring that mRNAs do not remain for long after their usefulness has expired. For example, differential RNA stability of both DHFR [3],[4] and histone [5] mRNAs during the cell cycle plays an important part in the cell-cycle dependent expression of their protein products.
Post-transcriptional regulation is especially important for the kinetoplastids, an early diverging branch of model unicellular eukaryotes that includes many important human pathogens. This is due to an unusual feature of kinetoplastids: their mature mRNA transcripts are produced through cleavage and processing of large polycistronic RNAs, which are synthesized via unidirectional, RNA pol II transcription across large stretches of chromatin [6]. This unusual mechanism of mRNA generation precludes individual regulation of gene expression at the level of transcription. Instead, regulation of gene expression generally occurs via differential mRNA decay or other post-transcriptional mechanisms [7]. Despite this, kinetoplastids are capable of complex developmental regulation. For example, the model kinetoplastid parasite Trypanosoma brucei differentiates into two distinct cell types in the mammalian host and at least five in the Tsetse fly vector [8], and the different types have distinct gene expression profiles [9]. Kinetoplastids therefore constitute a class of eukaryotes that are able to proliferate, differentiate and adapt by means of entirely post-transcriptional gene regulatory networks, raising interesting questions about the nature of the pathways involved.
Several genes are known to be strongly differentially expressed during the kinetoplastid cell cycle. One group of co-regulated genes includes DHFR, topoisomerase 2 (TOP2) and replication protein-A (RPA1), whose transcript levels peak during S-phase due to the presence of a cycling element in the 5′ UTR with a consensus sequence (C/A)AUAGAA(G/A) [10]. This element can be functional in the 3′ UTR [11] or even in the pre-mRNA only [12], indicating that regulation can occur prior to mRNA maturation; and it also seems to be conserved in other kinetoplastids [13]. The DHFR, TOP2 and RPA1 transcripts, while originating from different parts of the genome, appear to belong to a post-transcriptional regulon [14] since they are co-regulated via common cis-regulatory motifs, and possess related biological functions.
Another interesting feature of the kinetoplastids is that organelles such as the mitochondrion, Golgi apparatus and flagellum exist as single-copy structures whose replicative cycles are tightly linked to that of the cell (reviewed in [15]). Hence, the expression of proteins involved in duplication of these organelles may also be regulated in concert with the cell cycle. This is illustrated by the replication of the kinetoplast, which is a distinctive structure housing the mitochondrial DNA (kDNA), a network of concatenated open-circular DNA minicircles and maxicircles and associated proteins. The replication of the kinetoplast is coordinated with the cell cycle and involves several dedicated kDNA-processing proteins, some of which are members of the post-transcriptional regulon mentioned above. Kinetoplast DNA ligase alpha (LIGKA) is involved in kDNA replication and is regulated during the cell cycle in Crithidia fasciculata [16] and T. brucei [17], however its transcript levels peak some time after those of the DHFR co-regulated group. Whether LIGKA is unique or a member of a broader group of similarly regulated transcripts is unknown, as is the nature of the RNA elements responsible for its regulation in cis. Differential expression may also occur via mechanisms operating at the levels of protein translation, localization and/or stability, since the kDNA ligase proteins are unstable and differentially localized during the cell cycle [16]. Indeed, regulation at both the mRNA and protein levels could synergize to ensure that certain key players in DNA replication, organelle replication, and cell division are tightly regulated. This would especially apply to proteins whose ectopic expression at other points in the cell cycle could short-circuit the program of organellar and cellular duplication.
Since kinetoplastid protozoa rely on RNA-binding proteins, rather than transcription factors, to regulate gene expression, their genomes might be expected to contain a disproportionately large number of genes coding for proteins with RNA-binding domains. This is certainly true for proteins possessing a Puf (Pumilio/Fem-3) RNA-binding domain, of which we found at least 10 encoded in the T. brucei genome [18]. The structure of the Puf domain consists of multiple copies of a tri-helical Puf repeat. Each tri-helical repeat binds one nucleotide via three key amino acid residues that cooperatively determine the base preference for that repeat [19],[20]. The various Puf proteins found in yeast bind a significant proportion of all mRNAs and those mRNAs bound by the same Puf protein tend to encode proteins that function in similar locations and processes [21]. Thus, Puf proteins have found functions regulating several large post-transcriptional regulons in a single-celled eukaryote.
Here we describe a Puf protein of T. brucei, PUF9, which is conserved within kinetoplastids and possesses 6 copies of the Puf tri-helical repeat. Our results indicate that in mid-to-late S-phase, PUF9 neutralizes a specific destabilizing sequence motif present on its target mRNAs, thus stabilizing them. Consistent with their temporal expression profiles, some of the proteins encoded by PUF9 target mRNAs appear to play roles in maturation and segregation of the daughter kinetoplasts after division, a role supported by the protein localization and over-expression phenotype of an uncharacterized PUF9 target, PNT1, as well as the previously reported characteristics of another target, LIGKA. A third PUF9 target transcript, PNT2, encodes a nuclear protein that relocates to the mitotic spindle midzone during nuclear division. Hence, PUF9 could function in the temporal coordination of nuclear and kinetoplast replication.
All T. brucei cells used were derived from the Lister 427 line. To obtain stably transformed clonal lines, 1-2×107 cells were transfected by electroporation with ∼10 µg linearized DNA at 1.5 kV followed by cloning by limiting dilution in medium containing the appropriate selective drug. For tet-inducible expression constructs, expression was induced by including 100 ng/ml tetracycline in the culture medium. Plasmids created for transformation of T. brucei cells are summarized in Table 1. Primers used to generate PUF9 fragments by PCR for cloning were as described [18]. The PNT1 ORF was amplified using the following primers:
GATAAGCTTATGTTGTCCCGAGCCCCA / GATGGATCCGCCGTTCTCACTGCTCACG.
The PNT2 ORF was amplified using:
GATAAGCTTATGCAGTGGAAGAAAGATGACT / GATGGATCCGAAATGCAGAGGTAAACTTTCG.
The PNT1 3′ intergenic region was amplified using:
GATCGGATCCGCATAGATGGAGAGAGTTATACG / GATCACTAGTCTCCACCTTTGTCACTATCCTG.
Point mutations were introduced into the PNT1 3′ UTR sequence in CAT-reporter plasmid pHD1876 by site-directed ligase-independent mutagenesis (SLIM) [22] in a multiplex PCR reaction using a plasmid containing the wild-type UTR as template and the forward primers:
GTAATGTAACATTATACCATTTGTGTTGTTGTTTAG and ACCATTTGTGTTGTTGTTTAG
and reverse primers:
TAATGTTACATTACAACACCCGCTGCAGAATTTTTGTG and ACCCGCTGCAGAATTTTTGTG
(mutated residues underlined). The products of this reaction were heat-denatured, re-annealed and transformed into E. coli. Plasmids from the resulting transformants were isolated and sequenced to check for side-mutations. Due to a strain-specific G→T SNP in the Lister427 gDNA initially used as template, the G residue 9 nt downstream from the point mutation is actually a T in the wild-type PNT1 3′ UTR but was mutated back to a G in this point-mutant construct, because the SLIM primers were designed from the published genomic sequence from the TREU 927 strain.
The TAP (Tandem Affinity Purification) tag used here possesses Protein A and Calmodulin Binding Protein domains. The PUF9 ORF was cloned into plasmid pHD918, generating a construct encoding PUF9 linked to the TAP tag at the C-terminus via a peptide linker that contains a TEV protease cleave site. This was expressed from the PARP promoter, under the control of the Tet-repressor, in bloodstream form (BS) cells by induction with 100 ng/ml tetracycline for 24 hr. RNA co-purification was performed as described [23]. Approximately 3×109 cells were induced to express the fusion protein by the addition of tetracycline for 12–24 hr prior to harvesting. Cells were washed in cold PBS, crosslinked on ice by UV irradiation at 400 mJ/cm2 in a Stratalinker, then snap frozen. Cell pellets were broken in 6 ml breakage buffer (10 mM Tris-HCl pH 7.8, 10 mM NaCl, 0.1% IGEPAL CA 630 (Sigma; identical to the previously used detergent Nonidet P-40), 4 mM Vanadyl Ribonucleoside complexes (VRCs, Sigma), 4 U/ml RNAseIn (Promega), 1× Complete Inhibitor without EDTA (Roche)) by passing through a 21-gauge syringe 15 times at 4°C. Insoluble material was removed by ultracentrifugation (100,000 ×g, 45 min at 4°C) and the salt concentration of the supernatant was adjusted to 150 mM. 200 µl of IgG sepharose bead suspension (Fastflow – GE Healthcare) was washed in IPP150 and rotated with the lysate for 2 hours at 4°C. IPP150 contained 10 mM Tris-HCl, pH 7.8, 150 mM NaCl, 0.1% IGEPAL CA 630. The flow-though was collected, and beads washed three times in 10 ml of IPP150 and once in 10 ml of TEV cleavage buffer (IPP150 with 0.5 mM EDTA, 1 mM DTT, 2 mM VRCs, 4 U/ml RNAseIn (Promega)). The TAP tag was then cleaved by adding 1 ml of TEV cleavage buffer and 100 units of TEV protease (Invitrogen), and rotating the beads for two hours at 16°C followed by collecting the eluate. RNA was isolated from the eluate using the QIAgen RNAeasy kit or Trizol LS according to the manufacture's instructions. The entire procedure was scaled down if less RNA was required, e.g. for RT-PCR. Aliquots, equivalent to 4×106 cells, were taken at various points in the procedure for analysis by western blot. Calmodulin selection was not used for RNA isolation due to the requirement to add calcium to the buffer during binding.
Genomic T. brucei microarrays were generated containing 24,567 random shotgun clones from T. brucei brucei strain TREU927/4 genomic DNA [9]. Test and control samples of RNA were reverse-transcribed using SuperscriptII (Invitrogen) according to the manufacturer's instructions in the presence of either Cy5-dCTP or Cy3-dCTP and cDNA purified using the QIAquick PCR purification kit (QIAGEN), ethanol-precipitated and resuspended in 5 µl TE. Cy3- and Cy5- labelled cDNAs were mixed, denatured at 95°C for 5 min and snap-chilled, then added to 60 µl of hybridization buffer (50% formamide, 3× SSC, 1% SDS, 5× Denhardt's reagent and 5% dextran sulphate). This was added to the slide, a coverslip affixed and incubated at 62°C overnight in a humidified chamber. Slides were washed at RT for 10 min in 2× SSC, 0.2% SDS, 10 min in 2× SSC, and 10 min in 0.2× SSC, dipped in isopropanol and dried. Microarrays were scanned with ScanArray 5000 (Packard BioScience, Dreieich, Germany) and analyses of resulting images were performed using GenePix software (Axon Instruments, Union City, USA). The software package MCHIPS [24] was used for data quality assessment and normalization. Clones corresponding to positive hits were sequenced from one end and mapped onto the published T. brucei genome.
For RNA detection by Northern blot, RNA was size-separated by overnight agarose-gel electrophoresis on a 3.5% formaldehyde gel, transferred onto a nylon membrane by capillary transfer and fixed by UV irradiation as described [25]. The membrane was prehybridized in a hybridization bottle in 5× SSC, 0.5% SDS with salmon sperm DNA (200 µg/ml) and 1× Denhardt's solution for 2 hours at 65°C. Probe was generated by PCR in the presence of [32P]-labelled dCTP followed by purification using the QIAGEN nucleotide removal kit according to the manufacturer's instructions. Probe was added to the prehybridization solution and the bottle rotated at 67°C overnight. After rinsing the membrane in 1× SSC/0.5% SDS, probe was washed out with two 20 minute washes in 0.2× SSC/0.5% SDS at 67°C and the membrane exposed on a phosphorimaging screen for 2–48 hours. The screen was read on a recently calibrated Fugifilm FLA-3000 reader. Signal density from bands were quantified in Image Quant v3.45 and background signal density from a nearby region within the same lane of the gel was quantified and subtracted from the value for the band. Blot picture intensities were adjusted such that the darkest pixel was set to zero intensity, and 5% of the lightest pixels were clipped to 100% intensity.
TAP-copurified RNA, or RNA from the flow-through (derived from the equivalent of ∼2×108 or 4×107 cells respectively) was reverse-transcribed using a cocktail of gene-specific primers and Superscript III reverse transcriptase (Invitrogen) in a 20 µl reaction. This was used as template in a semi-quantitative PCR reaction to detect control and test genes (1 µl cDNA per 50 µl reaction). Samples (7 µl) were removed after 28, 32 and 36 cycles and analysed on an agarose gel. All primer pairs were designed using Primer3 [26] and had similar melting temperatures (∼60°C) and product lengths (250–400 nt).
We adapted the cell-starvation protocol previously described [27], which produces semi-synchronous cultures arrested predominantly at the G1 phase by starvation. Cells were seeded at 1×106/ml, and two days later the starved culture (∼2–4×107 cells/ml) was diluted in auto-conditioned MEM-pros medium [28] to induce resumption of the cell cycle. Aliquots of ∼106 cells were taken at regular intervals over the next 9–17 hours for flow cytometry. Cells were collected from these samples by centrifugation, resuspended in 100 µl of PBS, fixed by dropwise addition of 1 ml of 70% ethanol, 30% PBS while gently vortexing, and stored at 4°C. Cells were collected by centrifugation (2000 ×g, 10 min), resuspended in 500 µl of PBS with 10 µg/ml RNase A and 30 µg/ml propidium iodide, incubated at 37°C for 30 minutes and analysed by FACSSCAN. The proportion of cells in each phase of the cell cycle was estimated using the Watson algorithm [29] as implemented in the Flowjo software package.
Approximately 106 cells were collected by centrifugation and resuspended in 50 µl of PBS. Cells were fixed in 4% paraformaldehyde (or 4% formaldehyde/5% acetic acid if staining for mitotic spindles) in PBS for 20 min, washed twice in PBS and allowed to settle onto polylysine-coated slides. Cells were permeabilized in 0.1% Triton-X in PBS, washed twice and blocked in 1% BSA or gelatine for 1 hour. Primary antibody was added at the recommended dilution, incubated for 1 hour and the slides washed 3 times in PBS before addition of fluor-conjugated secondary antibody in the dark for 1 hour. Slides were washed, stained with DAPI for 10 minutes, and washed twice more before drying the slide. A drop of Vectorshield was added, the coverslip was affixed, and cells were viewed on a Leica DM fluorescence microscope. Where Mitotracker staining was used, Mitotracker CMXros (Invitrogen) was added to the cell culture medium at 250 nM for 30 min, after which the cells were pelleted and resuspended in fresh media. After incubating for 15 minutes, cells were fixed and stained using the protocol described above.
The 3′-UTR sequences of T. brucei genes were taken from the published genomic sequence of the TREU 927 strain, using start and end positions predicted previously [30] or, where the prediction was absent, taking the entire intergenic region to a maximum of 5 kb downstream of the stop codon. The sequences of interest were compared to 1000 randomly chosen 3′ UTRs as a background sample using Trawler [31]. Homologous genes in Trypanosoma congolense were also identified by TBLASTN searches and the 3′-UTR sequences predicted from the downstream intergenic regions by similar means.
Synthesis and maturation of mRNA were simultaneously inhibited by addition to the growth medium of 10 µg/mL actinomycin D and 2.5 µg/ml sinefungin. Sinefungin was added 5 minutes prior to actinomycin D [32]. Cells were collected at the indicated time points and RNA isolated by Trizol extraction. RNA levels were estimated by Northern blotting using [32P]-labelled probes, and quantitated by phosphorimaging. The stable SRP RNA was used as a loading control.
The Puf domain protein PUF9 (Tb927.1.2600) is conserved among Trypanosoma and Leishmania species. It contains 6 Puf repeats, as well as extended N- and C- terminal domains lacking any homology to characterized proteins (Figure 1A). We decided to investigate the biological function of PUF9 through RNAi and by co-purification of target RNAs.
The phenotype of bloodstream-form (BS) cells in which PUF9 was depleted by tet-inducible RNAi, or over-expressed using a tet-inducible VSG promoter, was examined. A Northern blot confirmed PUF9 mRNA knockdown after 24 hours of induction (Figure 1B). Although over-expression of PUF9 did not cause any noticeable phenotype, PUF9 RNAi reduced overall cell growth over the six days of RNAi induction (Figure 1C). We used freshly thawed clonal cell lines, which may account for the growth phenotype not seen previously [18]. Flow cytometry revealed an accumulation of cells with 2C DNA content (G2/M cells) in PUF9 RNAi cells relative to uninduced cells (Figure 1D), and there were also more polyploid (>2C) cells. Examination of trypanosomes by fluorescence microscopy after staining for DNA can also be used to score them for cell cycle stage: cells with a single nucleus and kinetoplast (1N1K) are in G1 or S phase, cells with two kinetoplasts and one nucleus (1N2K) are in G2, and cells with two kinetoplasts and two nuclei (2N2K) are mitotic or post-mitotic. The proportion of 1N1K cells was lower in PUF9 RNAi cells (Figure 1D). In addition, the PUF9 RNAi cells often possessed more than two flagellae, nuclei, or kinetoplasts (Figure 1E). Extra nuclei or kinetoplasts were seen after 24 hours of PUF9 RNAi (Figure 1D), suggesting a possible defect in control of organelle copy number. However, there was no obvious difference in the occurrence of annucleate “zoid” cells. Similar experiments in insect-form procyclic (PC) cells yielded no obvious phenotype (not shown).
In order to find out which mRNAs were targeted by PUF9, we expressed the protein with a C-terminal TAP tag in BS cells to allow affinity purification. Western blotting indicated that the tagged protein was stable in BS cell lysate and it was found in the cytoplasm of BS cells by immunofluorescence staining (Figure 1F). Protein-RNA complexes from BS lysates were selected on an IgG column, and eluted by cleaving the tag with TEV protease. Cells expressing the TAP tag alone served as a control. RNA that co-precipitated with PUF9::TAP or the TAP-tag alone was reverse transcribed with fluorescently labelled nucleotides and the cDNA hybridized to a microarray of shotgun genomic clones. Spots showing more than 2-fold higher intensity for the PUF9 channel were flagged and the corresponding genomic clones end-sequenced. Results from two biological replicates are summarized in Table 2. Several genomic loci were identified from multiple overlapping DNA clones, indicating that a gene of interest was present in the overlapping regions. Genes were considered as candidates for interactions with PUF9 if more than one spot was flagged that contained sequences overlapping that gene, or the same spot containing the gene was flagged from both biological replicates. rRNA and PUF9 itself were also identified in one replicate. While intriguing, it is possible that these hits may represent artefacts caused by the over-expression of the tagged PUF9, which is integrated into an RRNA locus, leading to higher background levels of these two RNAs. The number of transcripts that were identified was surprisingly small in comparison to similar experiments in yeast where hundreds of transcripts were coprecipitated with PUF proteins [21]; nonetheless, the fact that sequencing of several different flagged spots resulted in them repeatedly being mapped back to the same few transcripts indicates that significant coverage of high-affinity PUF9 targets was attained.
Since some microarray hits spanned several adjacent genes, another independent TAP-purification was performed and the co-purifying RNA was analysed for enrichment of mRNAs by semi-quantitative RT-PCR using primers specific for individual gene ORFs. Four candidate sequences were amplified more strongly from the PUF9-copurified RNA than the TAP-only copurified RNA, while amplification of the abundant TUBA transcript was approximately equal between the samples (Figure 2). This was not due to differences in transcript abundance in the lysates since RT-PCR on RNA isolated from flow-through fractions showed no detectable enrichment. The confirmed PUF9-associated transcripts are LIGKA (kDNA ligase α/Tb927.7.610), a histone H4 variant (H4V/Tb927.2.2670, possibly involved in transcription termination [33]), and two uncharacterized kinetoplastid-specific genes that we have named Puf Nine Target 1 (PNT1/Tb11.02.4400), and Puf Nine Target 2 (PNT2/Tb11.01.6470). Tb11.02.6460, an adjacent gene to PNT2, was also tested (not shown) because we could not delineate which of the two transcripts was responsible for the hits around this genomic locus using the microarray data alone (Table 2). However, only PNT2 mRNA was found to be a genuine target of PUF9 after analysis by RT-PCR. The PUF9 transcript itself could not be amplified from cDNA from the PUF9-copurified RNA. However, this transcript has an exceptionally long 3′ UTR that might be vulnerable to degradation during the procedure.
The effect of PUF9 on its target transcripts was examined by Northern blotting of RNA from the PUF9 over-expressing and PUF9 RNAi BS cells. Probing the Northern blots for the four target genes showed that RNAs from three of them were clearly more abundant when PUF9 was over-expressed and less abundant when PUF9 was depleted (Figure 3A and Figure S1), while the remaining target, H4V, was only slightly affected. Because H4V transcript is only weakly regulated by PUF9, it was excluded from further analysis here. However, its association with PUF9 could still indicate regulation at a different level, e.g. translation, that was not tested in this work. Tb11.01.6460, which is adjacent to PNT2, showed no dependence on PUF9 (not shown), consistent with the PUF9-mediated upregulation operating on individual mature transcripts rather than genomic loci or polycistronic precursor RNA.
To find out whether mRNA half-lives were influenced by PUF9, we inhibited mRNA synthesis using actinomycin D and sinefungin [32], and followed the abundance of the PUF9 target mRNA LIGKA. Actinomycin D binds to DNA and inhibits RNA transcription elongation, while sinefungin inhibits methylation of the spliced leader RNA leading to a blockage in mRNA maturation [34]. LIGKA mRNA half-life was approximately 30 minutes in normal cells, but was reduced four-fold when PUF9 was depleted by RNAi, while the half-life of the actin control transcript remained unchanged (Figure 3B). These data support a role for PUF9 in stabilizing its target mRNAs. The effects on target transcript abundance and half-life were not due to slower growth or the previously observed increase in the proportion of cells in G2 phase, because they also occurred when PUF9 was targeted by RNAi in PC cells, which exhibited wild-type growth and flow cytometry profiles (see below).
The PUF9 target gene LIGKA has a homologue in C. fasciculata (kinetoplast DNA ligase α), for which the mRNA was previously shown to vary in abundance with the cell cycle [16]. This, together with the PUF9 RNAi phenotype that hinted at a defect in cell cycle progression, led us to investigate whether PUF9 plays a role in cell-cycle-coupled differential expression of genes. PC cells are amenable to synchronization by starvation [27] and by hydroxyurea treatment [35] while BS cells have also recently been synchronized by hydroxyurea [36]. Although no hydroxyurea-generated artifacts have been observed during synchronization of T. brucei, drug-mediated synchronization has been observed to cause uncoupling of DNA replication status from cyclin levels in other systems [37], therefore we opted to follow the T. brucei starvation-induced synchronization protocol [27].
Starved PC trypanosomes accumulate in the G1 phase of the cell cycle (Figure 4). Upon release from starvation, about 70% of cells begin to progress through the cell cycle after a ∼4 hour lag (Figure 4). Analysis by Northern blot showed that transcript levels of housekeeping genes such as TUBA increased rapidly for the first hour, then increased at a steady rate throughout the assay, relative to structural RNAs such as SRP (data not shown), therefore TUBA was used as a “baseline” mRNA transcript to normalize for loading and the global effects of starvation upon mRNA levels. DHFR and HISH4 transcripts, which are known to be regulated during the eukaryotic cell cycle, peaked in early- and mid- S-phase respectively, consistent with good synchronization of the recovering cell culture (Figure 4). The timing of their transcript maxima fits with the fact that DHFR protein is needed prior to DNA synthesis while HISH4 is required during or immediately after synthesis. The three PUF9 target genes LIGKA, PNT1 and PNT2 were also regulated during the cell cycle, peaking in mid- to late- S-phase. The C. fasciculata homologue of LIGKA also cycles out-of-phase with the DHFR transcript in that organism [16].
To determine whether PUF9 plays a role in the S-phase-specific upregulation of its target genes, PUF9 was targeted by RNAi in procyclics during synchronization. Northern blotting showed that RNAi against PUF9 effectively repressed the PUF9 transcript and also lowered levels of PUF9 target transcripts as with BS cells (Figure S2); although as previously noted, PUF9 RNAi had no effect on PC growth, perhaps due to residual expression that is sufficient for function in PCs, or stage-specific differences between the cell-cycle checkpoint mechanisms. Flow cytometry also showed that there was no significant difference between synchronization efficiencies of induced and uninduced PUF9 RNAi PC cells, and this is supported by the nearly identical cyclical expression of the cell-cycle regulated HISH4 transcript between induced and uninduced cells (Figure 5A, triangles in 5B). However, probing for the PNT1 and PNT2 transcripts revealed that they no longer oscillated over the cell cycle when RNAi against PUF9 was induced (Figure 5A, circles in 5B), indicating that PUF9 is required for the peak in their transcript levels that occurs in S-phase in the control cells.
Interestingly, the PUF9 transcript itself also showed moderate cell cycle-coupled regulation, peaking at a similar time to its target mRNAs (Figure 4 bottom panel). However, it seems doubtful that this relatively moderate regulation at the mRNA level could fully account for the larger changes in expression of the target mRNAs, so it is likely that other regulatory mechanisms exist. Attempts to generate PUF9 antisera failed; therefore we tagged one PUF9 allele in situ with an N-terminal V5 epitope. However, the protein showed approximately the same degree of variation through the cell cycle as was previously observed for the mRNA levels, and no difference was seen in the electrophoretic mobility by western blot (not shown). It should be noted that N-terminal in situ tagging replaces the 5′ UTR, so 5′ UTR-dependent translational regulation would not be detected. We also attempted to co-purify interacting protein partners of PUF9 by tandem affinity purification of the TAP-tagged PUF9 protein, but mass-spectrometry of affinity-purified protein bands only revealed degradation products of PUF9 itself. Thus, the mechanism whereby PUF9 specifically stabilizes its target transcripts in late S-phase remains to be elucidated.
PUF9 targets appear to be regulated independently to most previously characterized cell-cycle-regulated transcripts (e.g. DHFR, TOP2, RPA1), which peak at a different time and possess known sequence motifs that act as cell-cycle regulatory elements (CCREs) [10],[11],[12],[13],[16]. To identify the RNA determinants responsible for PUF9-mediated cell-cycle-coupled transcript regulation, a combination of bioinformatics and experimental approaches was used. The 3′ UTRs of the three known mRNAs regulated by PUF9 were analyzed using Trawler, a program that identifies over-represented motifs in sets of sequences relative to a background set of genes [31]. This strategy is useful for identifying putative recognition sequences for Puf proteins because they tend to recognize primary RNA sequences rather than secondary RNA structures [38]. The most over-represented motif instance identified by Trawler was “UUGUACCW”, found 7 times in the 3 sequences. The best cluster within this family is summarized in a position weight matrix (Figure 6A). This motif is a promising candidate PUF9-recognition motif as it contains the Puf protein consensus core binding sequence “UGUA” [39]. In addition, the predicted key nucleotide-binding residues of the 4th, 5th and 6th Puf repeats in the T. brucei PUF9 protein are homologous, respectively, to “U”, “G” and “U”-binding repeats of characterized Puf proteins [40], indicating that at least the minimal conserved “UGU” trinucleotide forms part of the PUF9 recognition motif. A search of the preliminary T. congolense genome (the closest relative of T. brucei for which a largely complete genomic sequence is available) shows that it possesses clear homologues to all three T. brucei PUF9 targets. Despite being sufficiently distant from T. brucei to have no detectable similarity in most of the 3′ UTRs, the three PUF9 target homologues possessed several copies of the candidate motif in their 3′ UTRs.
To test experimentally whether the 3′ UTRs of PUF9 target genes contain CCREs, the 3′ intergenic region of PNT1 was cloned downstream of a CAT reporter gene and the reporter was expressed in procyclic cells. This 3′ UTR was chosen because its three “UGUA” motifs are clustered within a 20 nt region, the last one of which is contained within “UUGUACC” (the motif we identified as being over-represented in PUF9 targets). Cell synchronization experiments showed that the 3′ UTR could indeed confer cell cycle-coupled regulation upon the reporter transcript (Figure 6B, 6C), similar to the behavior of the native PNT1 transcript, although the magnitude of regulation was somewhat reduced and the peak in transcript level is delayed (see below). This is possibly due to increased stability conferred by the CAT ORF or the exogenous 5′ UTR. Alternatively, the 3′ UTR of PNT1 may contain only one component of a set of dispersed functional elements located over the entire transcript that cooperatively lead to efficient cell cycle-driven regulation. To locate this element, we further examined the PNT1 3′ UTR. All the “UGUA” motifs are clustered around ∼425 nt downstream of the stop codon, and initial mapping experiments showed that the region between +324 nt and +680 nt was critical for cycling of the CAT::PNT1-3′ UTR reporter transcripts (Figure 6B). We then mutated the central, conserved “G” of the last motif to an “A” and found that this also abolished its ability to confer cell-cycle regulation to the CAT reporter transcript (Figure 6B, 6C). Thus, this motif seems to act as a CCRE component. Despite this, preliminary expression-profiling experiments on synchronized cells (data not shown) showed no evidence for cell-cycle regulation in several hundred other transcripts containing at least one copy of this motif in the predicted 3′-UTR region, indicating that, although necessary, it is not sufficient to mediate cell-cycle regulation. Hence, the secondary structural context of the motif or the presence of other cooperating elements may heavily influence its effectiveness.
When the 3′-UTR reporter constructs were transformed into PUF9 RNAi PC or BS cells, the abundance of the reporter mRNA bearing the wild-type 3′ UTR was dependent upon the expression of PUF9 (Figure 6D). Subsequently, Western blotting indicated that CAT protein levels were also reduced in PUF9 knockdown cells, but as this seemed to be roughly proportional to the drop in mRNA levels (Figure S3), it is still uncertain whether PUF9 appreciably modulates translation. More elaborate kinetic studies and reporter transcripts also bearing the 5′ UTR of PNT1 would be required to thoroughly investigate potential translational regulation by PUF9.
Significantly, the same point mutation that abolished cell-cycle response also abolished the dependence of the transcript levels on PUF9 (Figure 6D). This may indicate that the CCRE is potentially a direct binding site for PUF9, and that the point mutation abolishes PUF9 binding. However, if PUF9 were the only player in regulation then the mutant transcript should be constitutively unstable, which is not the case: steady-state levels of the mutant reporter (normalized against the SRP RNA) were closer to those of the wild-type in the presence of PUF9.
The proteins encoded by PUF9 target transcripts might be expected to function in similar processes since they are co-regulated in the cell cycle. To determine if this is the case, PNT1 and PNT2 were expressed as C-terminally myc-tagged proteins in PC cells. LIGKA is already known to be associated with the kDNA [17]. Remarkably, PNT1::myc was also found either forming a doublet closely flanking the kDNA or overlapping it (Figure 7A). We cannot rule out that the apparently overlapping signals are actually doublets orientated in-line with the kDNA as seen from above. Interestingly, when PNT1::myc was over-expressed, very small and faintly staining additional kinetoplasts appeared in 33% of cells after 8 hours, relative to 5% in uninduced cells. These were similar in size and localization to the “ancillary kinetoplasts” observed at low frequency in some other kinetoplastids [41]. They stained for PNT1::myc (red arrowheads, Figure 7A and 7B), and unlike normal kinetoplasts, were often mislocated anterior to the nucleus. These extra structures could represent degenerating kinetoplasts retained from a sister cell during division, fragmented kinetoplasts, or semi-synthesized kinetoplasts formed by aborted, late re-replication. A minor proportion of cells (∼1-3%, only marginally higher than in uninduced cells) lacked a normal kinetoplast, possessing only ancillary kinetoplasts or no kinetoplast at all. Total myc signal was dramatically increased in these cells and was localized throughout the mitochondrion (Figure 7B). Consistent with this, SignalP 2.0 HMM predicts a potential peptidase cleavage site 25 amino acids from the N-terminus, hinting at the presence of an N-terminal signal peptide.
PNT2::myc was expressed in PC cells using the same system described above. In clonal transfected cell lines, PNT2::myc appeared to localize to the mitotic spindle during mitosis and by the 2N2K stage it was seen in an elongated structure mid-way between the two nuclei (Figure 7C). Co-staining cells with the KMX1 antibody, which reveals mitotic spindles, confirmed that this structure was the mitotic spindle midzone (Figure 7D). The slightly granular appearance of cells here is an artifact of acetic acid fixation. No obvious over-expression phenotype was seen for PNT2::myc. However, PNT-2::myc was only expressed at low levels as seen by western blots of multiple transfected clones (not shown), which could preclude generation of an over-expression phenotype. We cannot rule out the presence of PNT2::myc in the cytoplasm since there was some diffuse cytoplasmic signal, but this was similar to the background signal seen in untransfected controls.
The increased multiplicity of protein function in eukaryotes has been proposed to be due partially to the replacement of inflexible prokaryotic polycistronic regulatory operons with gene-specific combinations of cis-acting regulatory elements [14]. The kinetoplastids may at first appear to challenge this paradigm of individualized gene regulation in eukaryotes, since they transcribe virtually all genes in large polycistrons that are even larger than operons of prokaryotic species. However, the kinetoplastid genomes are approximately two-fold enriched in genes encoding Pumilio domain and CCCH zinc finger proteins, relative to unicellular organisms that show transcriptional control, suggesting that these RNA-binding proteins have stepped in to replace transcription factors in regulating gene expression. This idea is supported by the current study where a small group of mRNAs that are likely to function in coordinated biological processes is shown to be under the control of a common upstream regulator, PUF9.
The temporal expression and localization of the proteins encoded by the PUF9 target transcripts indicates that they function in certain organelles (the nucleus, mitotic spindle and kinetoplast) at a specific time in the cell cycle (late S-phase/G2 phase). In kinetoplastid cells, the copy numbers of kinetoplasts and other major organelles and cellular structures are stringently maintained at one per G1 phase cell, and their replication is coordinated with each other and coupled to that of the cell. We hypothesize that PUF9 switches on the expression of target genes in late S-phase of the cell cycle to ensure simultaneous performance of their respective functions in organelle replication or division. Three lines of evidence support a role for PUF9 in co-coordinating cell-cycle governed replicative processes: firstly, the extra nuclei and kinetoplasts seen in BS cells when PUF9 is knocked down indicate that organelle replication is de-coupled from cell division. Secondly, PUF9 drives the upregulation of its target transcripts specifically in mid- to late- S-phase. Thirdly, bypassing PUF9 regulation by directly interfering with the expression of the downstream LIGKA or PNT1 proteins results in aberrant kinetoplast DNA content [17] or copy-number (shown here).
The parallels between two PUF9 targets, LIGKA (characterized previously [16],[17]) and PNT1, are particularly striking. Both encode proteins localized to the kinetoplast and both probably function in kinetoplast replication as indicated by their over-expression or knockdown phenotypes. The kinetoplast DNA is unique in nature in that it is a disc-shaped network of open circular DNA molecules, concatenated together with a “chain mail” topology (reviewed in [42]). It also has a unique mechanism of replication: individual minicircles disassociate from the network, migrate to the posterior kinetoflagellar zone where they are replicated, and then migrate back to one of the two “antipodal” sites which flank the disk at its perimeter, and where topoisomerase-mediated minicircle reattachment to the kinetoplast occurs. It has been hypothesized that LIGKA seals the final nick in replicated kDNA minicircles, re-licensing them for replication [16]. If so, this process would require some temporal and spatial regulation to prevent re-licensing in the vicinity of active minicircle replication machinery in the same cell cycle. Similarly, the fact that PNT1 over-expression results in an observable defect suggests that a tight reign on its protein levels must also be maintained to avoid negative consequences for the cell. We speculate that these shared requirements for tightly controlled expression may explain the involvement of PUF9 in regulating these two genes. Interestingly, the doublet that PNT1 sometimes forms, flanking the kinetoplast disc, is similar to that seen for proteins belonging to the antipodal sites where newly replicated minicircles are reattached to the kinetoplast.
PNT2::myc displayed an interesting cell-cycle dependent dynamic localization, being present in the nucleus in pre-mitotic cells but localized to the spindle midzone in post-mitotic cells. Because its mRNA levels peak at around mid- to late- S-phase, the protein levels of endogenous PNT2 might be expected to peak shortly afterwards, probably co-inciding with this relocalization during mitosis. Similar localization patterns have been reported for certain other proteins such as TbNOP86, a nucleolar protein that localizes to the spindle during cytokinesis and whose RNAi phenotype resembles that of PUF9, i.e. an increase in G2 phase and polyploid cells [43]. Some chromosomal passenger proteins such as TbCPC1 and TbCPC2 also display similar localization during mitosis, although they additionally relocalize to a dot on the cleavage furrow during cytokinesis, which we could not detect for PNT2::myc. These proteins are involved in spindle function and cytokinesis as their suppression in PCs leads to mitotic spindle abnormalities and accumulation of G2-phase cells [44]. Given its similar localization and temporal expression, we speculate that PNT2 may have a similar function related to the timing of mitotic spindle assembly or disassembly.
In addition to being regulated at the mRNA level, LIGKA, PNT1 and PNT2 are probably regulated at the protein level. LIGKA protein is known to be relatively unstable [16], while levels of tagged PNT1 in the PNT1::myc over-expressing cells seem to be self-limiting, in light of the fact that many cells displaying the associated phenotype no longer expressed detectable protein at the time of fixation (e.g. the top cell in Figure 7A, which possesses an anterior kinetoplast fragment but does not express PNT1::myc). Further, the massive increase in PNT1::myc protein levels in the mitochondrion of cells lacking a normal kinetoplast may mean that the kinetoplast not only sequesters PNT1 but also suppresses PNT1 expression at the protein level, although here we have only demonstrated an associative, rather than causal, relationship between kinetoplast loss and increased PNT1 protein. Additionally, the ancillary kinetoplasts we observed, while capable of sequestering PNT1, were not by themselves associated with repressed PNT1 protein levels. In general, rapid protein degradation for cyclically regulated proteins should allow a sharper peak in protein levels to occur at the proper time.
Puf proteins in other eukaryotes generally act to destabilize their target transcripts. However, PUF9 was shown to stabilize its target transcripts, and similarly, to stabilize a reporter transcript carrying the 3′ UTR of PNT1. The presence of the UUGUACC motif (common to all PUF9 target 3′ UTRs) is essential for regulation by PUF9. The fact that mutations in this motif appear to stabilize the transcript, rather than destabilizing it, suggests that the motif recruits a destabilizing factor that is then inhibited by PUF9. Probable candidates for this destabilizing factor include the other Puf-domain proteins, since as mentioned, they generally act to destabilize targets, and are quite likely to bind the same motif as PUF9 because Puf proteins usually bind a conserved “UGUA”-containing motif [39]. This would potentiate a simple binding-site competition model of regulation whereby PUF9 and the destabilizing factor compete for binding at the same RNA motif. Other more complex mechanisms are also plausible, and it is worth noting that the 3′ UTR of PNT-1, while capable of conferring cell-cycle regulation onto a reporter transcript, did not confer exactly the same expression as the native transcript, which implicates the 5′ UTR and ORF, and perhaps even pre-mRNA sequences, in fine-tuning post-transcriptional regulation of PNT1.
The means by which PUF9 seems to be most active in late S-phase is of particular interest if PUF9 is to be placed within a cell-cycle regulatory network. Higher expression of PUF9 protein during S-phase may occur, and indeed there was an indication that PUF9 transcript levels oscillate with the cell cycle. However, the magnitude of PUF9 transcript upregulation in S-phase seemed to be insufficient to fully explain that of its downstream targets. While regulation may occur at the translational level, a suitable antibody against PUF9 could not be raised to test this. Post-translational modifications such as phosphorylation, ubiquitination etc. may also play a role, although a V5-tagged PUF9 protein showed no changes in electrophoretic mobility (which might indicate post-translational modification) over the cell cycle, and a recent analysis of the T. brucei phosphoproteome did not identify PUF9 among the set of phosphorylated proteins [45]. However, all these possible modes of regulation might not act only on PUF9 but rather on factors cooperating with PUF9 in regulating transcriptional stability, for instance the factor hypothesized above to act contrary to PUF9 to destabilize the target transcripts.
The characterization of PUF9 target transcripts in T. brucei appears to be revealing a post-transcriptional regulon of genes involved in coordinating the replication of subcellular structures in the cell cycle. This is regulated independently to the set of cell-cycle-regulated transcripts investigated by several other groups [10],[11],[12],[13] containing DHFR, LIGKB, TOP2 and RPA1, because the timing of peak expression of the PUF9 targets is significantly later in the cell cycle (as was first observed when the expression of LIGKB was compared to that of LIGKA in C. fasciculata [16]). The three PUF9 targets encode kinetoplast- or spindle- associated proteins, are co-regulated in the cell cycle, and perturbations in expression of at least two of them (PNT1 and LIGKA [17]) cause defects in kinetoplast replication. The main function of PUF9 is probably to ensure simultaneous function of its target genes during late S-phase in order to temporally coordinate certain processes in organellar and cellular replication. In particular, the putative roles for PUF9 target proteins in kinetoplast replication and the mitotic spindle implicates PUF9 in synchronizing kinetoplast maturation and mitotic spindle function. The further investigation of the upstream regulatory network and downstream effectors will lead to insights into how trypanosomes coordinate the replication of their organelles with that of the entire cell, and how they regulate gene expression in general without transcriptional control.
|
10.1371/journal.pcbi.1004042 | A Computational Approach to Identifying Gene-microRNA Modules in Cancer | MicroRNAs (miRNAs) play key roles in the initiation and progression of various cancers by regulating genes. Regulatory interactions between genes and miRNAs are complex, as multiple miRNAs can regulate multiple genes. In addtion, these interactions vary from patient to patient and even among patients with the same cancer type, as cancer development is a heterogeneous process. These relationships are more complicated because transcription factors and other regulatory molecules can also regulate miRNAs and genes. Hence, it is important to identify the complex relationships between genes and miRNAs in cancer.
In this study, we propose a computational approach to constructing modules that represent these relationships by integrating the expression data of genes and miRNAs with gene-gene interaction data. First, we used a biclustering algorithm to construct modules consisting of a subset of genes and a subset of samples to incorporate the heterogeneity of cancer cells. Second, we combined gene-gene interactions to include genes that play important roles in cancer-related pathways. Then, we selected miRNAs that are closely associated with genes in the modules based on a Gaussian Bayesian network and Bayesian Information Criteria. When we applied our approach to ovarian cancer and glioblastoma (GBM) data sets, 33 and 54 modules were constructed, respectively. In these modules, 91% and 94% of ovarian cancer and GBM modules, respectively, were explained either by direct regulation between genes and miRNAs or by indirect relationships via transcription factors. In addition, 48.4% and 74.0% of modules from ovarian cancer and GBM, respectively, were enriched with cancer-related pathways, and 51.7% and 71.7% of miRNAs in modules were ovarian cancer-related miRNAs and GBM-related miRNAs, respectively. Finally, we extensively analyzed significant modules and showed that most genes in these modules were related to ovarian cancer and GBM.
| A microRNA (miRNA) is a small RNA molecule that regulates the expression of mRNA genes. A miRNA can regulate multiple genes, and a gene can be regulated by multiple miRNAs. The regulation of genes by miRNAs may vary from patient to patient, even if they suffer from the same type of cancer. In this study, we identify the relationships between genes and miRNAs in cancer patients using expression data. Because these relationships are complicated by the involvement of transcription factors, which are among the most influential regulators of genes, we also attempt to explain the triple relationship among genes, miRNAs, and transcription factors. We constructed modules consisting of a set of genes and miRNAs, in which the expression levels are highly correlated. In most of these modules, genes and miRNAs are related to specific cancer types; their relationships are explained both by direct regulation of genes by miRNAs and by indirect relationships via transcription factors.
| Cancer is one of the leading causes of death worldwide. Although remarkable progress has been achieved in cancer therapies, the molecular mechanisms of cancer have not yet been fully identified. Among various regulations of cancer-related genes and pathways in several stages, the regulation of genes by microRNAs (miRNAs) in cancer cells has drawn particular attention, because many miRNAs are located in chromosomal regions that are frequently altered in cancer [1]. MiRNAs are small RNAs, known as important regulators of genes through binding to 3’ UTR regions of target genes [2]. In many cancer types, miRNAs have been studied as important biomarkers for diagnosis and prognosis of cancer, as many miRNAs function as oncogenes or tumor suppressors by regulating other oncogenes or tumor suppressor genes [1, 3].
Because miRNAs regulate genes by binding to the 3’ UTR regions of genes, many methods were developed to identify conserved sequence regions between miRNAs and mRNAs [4]. However, sequence-based approaches generate many false positive bindings sites and cannot identify functional changes of genes. Hence, the expressions of genes and miRNAs were also integrated to address possible negative correlations between the two sets of expression data [5, 6]. With the advances in high throughput technologies, large-scale mRNA expression and miRNA expression data sets from the same tumor samples have become available, due to collaborative efforts such as The Cancer Genome Atlas (TCGA) project. [7, 8]. These data sets enable researchers to apply computational approaches to identify relationships between mRNAs and miRNAs and help understand their effects in cancer.
Another important approach to understanding relationships between mRNAs and miRNAs is to analyze multiple genes and miRNAs simultaneously by constructing modules of them rather than analyzing each gene-miRNA pair separately [5, 9, 10]. It is widely known that a miRNA can regulate multiple genes [11], and a gene can be targeted by multiple miRNAs [12]. Changes in these numerous relationships can significantly alter the biological functions or signaling pathways associated with a specific cancer [13]. Although it is known that several pathways, such as the p53 and TGF-beta signaling pathways, are related to ovarian cancer [14, 15], the functions of miRNAs in these pathways have not yet been fully explained.
Although a few algorithms for finding gene-miRNA modules have been proposed, improvements are still needed. Peng et al. [5] proposed a bi-clique approach based on a gene-miRNA correlation matrix; however, most of the modules contained only one miRNA, and a few modules contained at most three miRNAs. Hence, it may be difficult to address multiple relationships between genes and miRNAs. Zhang et al. [6] integrated miRNAs, gene expression and gene-gene interactions based on a non-negative matrix factorization (NMF) framework [16]. The decomposed matrix components were considered as gene-miRNA regulatory modules. Although many modules were enriched with known pathways, the relationships between genes and miRNAs were not explained.
Relationships between genes and miRNAs become even more complicated because molecules such as transcription factors or signal transducers regulate genes and miRNAs. For example, p53, the most frequently mutated gene in cancer, regulates hundreds of genes and a set of miRNAs, including miR-24 family, miR-145, miR-107, and miR-192 [17, 18]. In [19], the authors constructed modules that contain highly correlated genes and miRNAs in their expression levels and found that miR-200a regulates the transcription factor ZEB1, which regulates genes contained in the same module as miR-200a.
To enhance the understanding of relationships between genes and miRNAs, we propose a framework that combines a biclustering approach and a Gaussian Bayesian network. Using the biclustering approach, gene-sample modules are first constructed based on gene expression and gene-gene interaction data sets. Here, a subset of genes that are correlated with each other in a subset of samples is clustered, because gene aberrations are different among patients, even if cancer occurs in the same organ or tissue type [20]. Next, using a Gaussian Bayesian network, gene-miRNA modules are constructed to identify miRNAs that regulate genes in gene-sample modules. Here, we use the expression data on genes and miRNA. When we applied our approach to ovarian cancer data sets and glioblastoma (GBM) data sets from TCGA, we identified several modules consisting of genes and miRNAs related to ovarian cancer and GBM. In many modules, relationships between genes and miRNAs were explained either by direct regulations of genes by miRNAs or by indirect relationships via transcription factors. In addition, functional pathway enrichment tests using several biological and signaling pathways demonstrated that these modules were biologically coherent. Based on ratios of cancer-related genes and cancer-related miRNAs, we extensively analyzed several significant modules and performed network analyses of these modules to demonstrate the regulation of genes by miRNAs.
Ovarian cancer. We collected mRNA expression and miRNA expression data sets for 587 tumor samples and 8 unmatched normal samples for ovarian cancer from TCGA [8]; mRNA and miRNA expression data were generated using an Affymetrix HG-U133A microarray and an Agilent H-miRNA_8X15K microarray, respectively. We normalized the expression levels of 12,042 genes using log2 ratios between tumor samples and the average of normal samples for each gene, and then selected 2,933 differentially expressed genes using a t-test (p-value < 0.001). Similarly, we normalized the expression levels of 479 miRNAs using the log2 ratios between tumor samples and the average of normal samples for each miRNA (Fig. 1 (A)).
Glioblastoma. We collected mRNA expression and miRNA expression data sets for 482 tumor samples and 10 unmatched normal samples for GBM [7]. These data sets were generated using the same microarray platforms used in the ovarian cancer study. After normalization, we selected 4,059 differentially expressed genes using a t-test (Bonferroni corrected p-value < 0.05). We used the expression levels of 423 miRNAs normalized using normal samples.
Selecting a p-value threshold for a t-test. The degree of expression changes depending on the cancer type. In this study, the number of differentially expressed genes was small in ovarian cancer compared to GBM. Hence, we used a less strict threshold for ovarian cancer.
Gene-gene interactions. We collected gene-gene interaction data from the HPRD database [21].
In this study, we first hypothesized that if a group of genes has similar expression tendencies in a subset of samples, and they are differentially expressed in these samples, then these genes might be related to similar functions or pathways in the development of cancer. We also hypothesized that a gene might have multiple functions and could function in several pathways. To incorporate these hypotheses, we use a biclustering algorithm to allow the duplication of genes and samples in multiple clusters. First, we construct a matrix of differentially expressed genes and samples, and then we normalize the expression values for each gene using a z-score to determine the tendency toward changes of gene expression in the samples. Next, we apply a SAMBA biclustering algorithm [22] to the normalized matrix to construct modules in which genes and samples are highly correlated (Fig. 1 (B)). The SAMBA biclustering algorithm models gene expression data in a bipartite graph G = (U,V,E), where genes in V are represented as nodes on one side and samples in U on the other side. There is an edge in E between a gene v in V and a sample u in U if the expression value of gene v changes significantly in sample u, having high absolute expression values. The biclustering algorithm generates subgraphs from the bipartite graph, in which most of the genes are connected to most of the samples as edges. These subgraphs represent highly correlated gene-sample clusters, where the tendency toward gene expression changes is similar for a subset of samples. Additional details are provided in Fig. S1. We calculate the statistical significance of each module based on a null hypothesis that the expression level of a gene is independent of the expression level of other genes for samples in a module, assessing that the average Pearson correlation coefficients (PCCs) of gene expression levels for genes in the module are higher than the ones from random modules for selected samples. For each module, we conduct the following test.
(Step 1) Construct a random module by randomly selecting the same numbers of genes and samples from the normalized matrix.
(Step 2) Calculate the PCC matrix of expression level values of genes in the module across a subset of samples. Then, calculate the average value of the PCC matrix, excluding diagonal elements.
(Step 3) Repeat Steps 1 and 2 N times, letting the average value from the i-th permutation serve as the randomavg(i).
(Step 4) Let the average PCC value of genes in the observed module be the moduleavg.
(Step 5) Calculate the p-value of the observed module using the following equation, where I is an indicator function.
p − v a l u e = ∑ i = 1 N I ( m o d u l e a v g < r a n d o m a v g ( i ) ) N
When we calculate the p-value, we try to take into account that observed modules are not independent of each other as genes overlap among modules. Hence, we construct random modules where genes in the modules share the same overlap ratio as the observed modules.
Recent research has shown that not all of the genes in cancer-related pathways undergo expression or genomic changes [23]. Consequently, certain genes that play important roles in cancer-related pathways might not be differentially expressed. To include functionally related genes in the gene-sample modules, we expand the gene-sample modules using a gene-gene interaction network. If a gene interacts directly with at least one gene in a module, then this gene can be regarded as a candidate gene for the module. For each module, we collect candidate genes and calculate the average PCC values of expressions between a candidate gene and the genes in the module. We add candidate genes to the module in descending order from the gene having the highest PCC value until the average PCC values of the expressions of genes in the module do not increase.
Because a set of genes with similar expression changes might be regulated by common miRNAs, we construct gene-miRNA modules by including regulating miRNAs in the gene-sample modules. For this task, we employ a Bayesian network model. Bayesian networks have been extensively used for analyzing gene expression patterns [24]. They are useful in modeling local dependencies and causal influences among variables. Hence, we estimate dependencies between expression values of genes and expression values of miRNAs based on a Bayesian network model. A joint distribution of genes X = {X1,X2, …, Xn} and miRNAs Y = {Y1,Y2, … Ym} is represented by a Gaussian Bayesian network. If Xi is normally distributed around a mean that linearly depends on its parents, then the conditional probability of Xi given its parents PaG(Xi) = {Yj, … Yk} can be represented by
P ( X i | P a G ( X i ) ) = P ( X i | Y j, …, Y k ) ∼ N ( a 0 + ∑ j ' a j ' · Y j ', σ 2 )
(1)
Then, the likelihood of X and Y can be represented by
L ( X, Y ) = P ( X 1, X 2, …, X n, Y 1, Y 2, … Y m ) = ∏ i = 1 n P ( X i | P a G ( X i ) )
(2)
To determine which sets of miRNAs explain the expression levels of genes in gene-sample modules, we use a Bayesian information criterion (BIC) as a measure for determining a Bayesian network structure between genes and miRNAs, which can be represented by
B I C = l o g ( L ) - l o g M 2 + O ( 1 ),
(3)
where M is the sum of the number of genes and miRNAs. To determine the parents PaG(Xi) of a gene Xi yielding the optimal BIC score, we should consider all combinations of miRNAs; however, this approach is highly time-consuming. To reduce the search space, we select candidate miRNAs whose average of absolute Spearman’s rank correlation coefficient (SCC) values for genes in a given module are within the top T% among all miRNAs. Note that we use SCC values for selecting candidate miRNAs to reduce the effects of possible outliers in the PCC. From candidate miRNAs, we first add a miRNA with the highest SCC value as a regulator and calculate the BIC score. Then, we add miRNAs with the next highest SCC values, until adding more miRNAs no longer improves the BIC score. After adding miRNAs to gene-sample modules using the above approach, modules with fewer than two miRNAs are filtered out because these modules cannot represent the combinatorial effects of genes and miRNAs. Finally, gene-miRNA modules are obtained.
To validate the relationships between genes and miRNAs in the modules, we consider four cases of gene regulations. In the first case, genes are directly bound and regulated by miRNAs. To validate this case, we select gene-miRNA pairs from miRTarbase [25] and MicroCosm (http://www.ebi.ac.uk/enright-srv/microcosm/htdocs/targets/v5/). Interacting pairs in miRTarbase are validated by various molecular experiments. Among them, reporter assays and western blot analysis confirm direct interactions. We compare the gene-miRNA pairs in our modules with these direct interactions in miRTarbase. MicroCosm provides computationally predicted binding sites for miRNAs in genomic sequences. Among these pairs, we select only gene-miRNA pairs with a negative correlation in expression values. From this process, we collect target genes for each miRNA, which we use for validation. Then, we perform a hypergeometric test for each miRNA in the modules to check for enrichment of genes in a module against the target genes of a miRNA.
However, certain genes in the modules are not directly regulated by miRNAs, even though the expressions of the genes and the miRNAs are highly correlated. To investigate this indirect relationship, we introduce transcription factors (TFs). We confirm relationships between miRNAs and TFs by manually searching the literature for evidence of cases where miRNAs are regulated by TFs or TFs are regulated by miRNAs. In the second case, we consider a relationship in which the miRNAs in a module regulate TFs, and these TFs regulate genes in the module. Here, it is not necessary that TFs be members of the module. We identify relationships between TFs and genes using the ChIP-X database [26]. For each TF in the database, we perform a hypergeometric test to determine if there is enrichment of genes in a module against the target genes of the TF. Here, the correlation of expression values between the miRNA and the TF must be negative, and the correlation values between the TF and the mRNA can be either positive or negative.
In the third case, genes and miRNAs are regulated by a common TF. In this case, correlations of expression values between gene-TF and miRNA-TF should be both positive or both negative.
In the fourth case, interacting pairs in miRTarbase [25], experimentally validated by the coexpression of miRNA and mRNA, are used to validate gene-miRNA pairs in our module. Molecular experiments for this case include quantitative real-time PCR (qPCR), microarrays, stable isotope labeling with amino acids in culture (SILAC) and pulsed SILAC.
To determine the functional relevance of the modules, we test whether the genes from the modules are enriched for specific biological functions or signaling pathways. We perform a pathway enrichment test using gene ontology (GO) biological process terms [27], KEGG pathways [28], and BioCarta pathways (http://www.biocarta.com). First, we download these pathways from GSEA (http://www.broadinstitute.org/gsea) and apply a hypergeometric test to each module, obtaining the p-values. We exclude biological functions or signaling pathways containing more than 300 genes, as such functions are too general. Supplementary Fig. S2 shows the distribution of GO biological functions as well as KEGG and BioCarta pathways. It can be seen that 51 of 825 GO terms contain more than 300 genes. To address any issues with multiple comparisons, we compute the q-values from the p-values based on a Benjamini & Hochberg correction. Then, we use a q-value < 0.05 for the enrichment threshold.
To validate that modules are related to the specific cancer, we first examine whether enriched pathways are related to the cancer being evaluated. For this task, we collect 2,032 cancer genes from the allOnco database (http://www.bushmanlab.org/links/genelists), which is a collection of list of cancer genes from several databases [29–32], 379 ovarian cancer genes from the Dragon Database for Exploration of Ovarian Cancer Genes (DDOC [33]), and 98 GBM genes from the literature ([34, 35]). Then, we calculate the ratios of these cancer genes in the modules. We also collect 100 ovarian cancer miRNAs and 92 GBM miRNAs from the Human miRNA & Disease Database (HMDD [36]). Then, we calculate the ratios of ovarian cancer-related miRNAs in the modules.
Genes involved in the development of cancer vary depending on cancer subtypes. In several papers [8, 37–39], the expression levels of marker genes are used to determine the subtype. For example, GBM samples were classified as a proneural subtype if marker genes DLL3, NKX2–2, SOX2, ERBB3, and OLIG2 were overexpressed [8]. Similarly, we check whether modules identified by our approach are related to a specific subtype of cancers using marker genes.
For this task, we perform the following two steps. In the first step, we cluster all samples into subtypes using hierarchical clustering with a dynamic tree cut [40]. For clustering, we use genes with high variability across the samples. Then, we assign each cluster to a subtype of cancers if known marker genes of cancer subtypes are overexpressed or underexpressed. If a cluster is not related to any subtype or is related to more than one subtype, that cluster is not assigned to any subtype. In the second step, for each module, we use marker genes of the subtype to compare the expression levels of the marker genes of samples in a module to the expression levels of samples in the other subtype clusters using the t-test. If the p-values of markers genes of the subtype are significant, we consider the module to be related to the given subtype.
To construct gene-sample modules, we applied the SAMBA biclustering algorithm to the gene expression matrix, allowing duplication of genes and samples in modules using an overlap factor of 0.5 in [0, 1], where 1 represents non-overlap. For ovarian cancer and GBM, we identified 90 and 135 modules, respectively, that represent similar tendencies of gene expression changes for a subset of samples. After performing 1000 permutation tests, we selected 58 and 88 modules with a q-value < 0.05 for ovarian cancer and GBM, respectively. Then, we enlarged these modules by adding genes using gene-gene interactions. On average, we added 15 and 33 genes to each module for ovarian cancer and GBM, respectively.
We constructed gene-miRNA modules from gene-sample modules by including miRNAs. As described in the Methods section, we pre-selected the candidate miRNAs based on the SCC values between the genes and the miRNAs and then added miRNAs to the module, which increased the BIC score. As shown in Fig. 2, we applied 20 different SCC thresholds (T% in [1%, 20%] of candidate miRNAs among all miRNAs) to reduce the search space. In Fig. 2 (A), the number of modules for ovarian cancer decreased as the thresholds decreased. We observed similar trends when the PCC was used instead of the SCC or when we did not integrate the gene-gene interaction data. Fig. 2 also shows that the ratios of cancer genes, ovarian cancer genes, and ovarian cancer miRNAs were similar for various SCC thresholds > 5%, and that these ratios increased when SCC thresholds decreased. Fig. S3 shows similar results for GBM. Note that we filtered out modules with fewer than two miRNAs, as such modules cannot represent the combinatorial effects of genes and miRNAs.
Among the various thresholds for candidate miRNAs, we selected a value of 3% (SCC value = 0.157 for ovarian cancer and 0.194 for GBM) for further analysis and constructed 33 and 54 modules for ovarian cancer and GBM, respectively. Tables S1, S2, S3, and S4 present lists of genes and miRNAs for the modules. For ovarian cancer, the average size of the modules was 34 genes and 10 miRNAs. On average, 19.1% of genes were cancer genes, 5.7% were ovarian cancer genes, and 51.7% of miRNAs were ovarian cancer-related miRNAs in the ovarian cancer modules. When combining genes and miRNAs from all modules, 18.6% (145 out of 777) of genes were cancer genes, 6.0% (47 out of 777 genes) were ovarian cancer genes, and 43.5% (47 out of 108) of miRNAs were ovarian cancer-related miRNAs. Based on the pathway enrichment test, 48.4% of the modules were enriched with biological functions or signaling pathways, and most of the modules contained at least one ovarian cancer gene. Table 1 shows ovarian cancer genes and miRNAs for the selected modules. Table S5 presents lists of cancer genes, ovarian cancer genes, and ovarian cancer miRNAs for all of the ovarian cancer modules. For GBM, the average numbers of genes and miRNAs for each module were 66 genes and 14 miRNAs. In the GBM modules, on average, 23.2% of the genes were cancer genes, 1.2% were GBM-related genes, and 71.7% of the miRNAs were GBM-related miRNAs. For all genes and miRNAs in the GBM modules, 20.6% (386 out of 1867) of the genes were cancer genes, 1.7% (32 out of 1867 genes) were GBM-related genes, and 48.4% (46 out of 95) of the miRNAs were GBM-related miRNAs. Table S6 presents lists of cancer genes, GBM genes, and GBM miRNAs for all of the GBM modules. Based on the pathway enrichment test, 74.0% of the modules were enriched in biological functions or signaling pathways.
Because our approach includes genes belonging to multiple modules, we calculated the overlap ratios of genes and miRNAs among the modules. The overlap ratio is defined as ∣m1 ∩ m2∣/∣m1 ∪ m2∣, where m1 and m2 are the number of genes or miRNAs in module 1 and module 2, respectively. Figs. S4 and S5 show the overlap ratios among the modules. The average overlap ratios of genes were 1.6% and 2.0% for ovarian cancer and GBM, respectively, and the average overlap ratios of miRNAs were 7.3% and 14.2% for ovarian cancer and GBM, respectively. The overlap ratios of miRNAs are higher than the overlap ratio of genes, indicating that a miRNA regulates many genes involved in several pathways.
As described in the Methods section, we examined the direct relationships between genes and miRNAs and their indirect relationships through TFs in the identified modules, as well as experimentally validated interactions between genes and miRNAs. For the ovarian cancer modules, we tested the direct relationship based on whether potential targets of a miRNA in the module were enriched for the genes in the same module using MicroCosm. Table 2 shows 8 miRNAs and their target genes in 12 ovarian cancer modules. For example, in Table 2, let-7b may directly regulate several genes (ESPL1, DEPDC1, BUB1B, AURKB and UBE2C) in module 33. Additionally, 19 gene-miRNA direct interaction pairs that were experimentally confirmed in miRTarbase are shown in Table 3. Previously, it was confirmed using a luciferase reporter assay and the western blot method that miR-93 targets E2F1. Also, it was confirmed using a luciferase reporter assay that miR-125b targets BCL3 in ovarian cancer cell [41]. All 156 gene-miRNA interaction pairs experimentally validated in miRTarbase are shown in Table S7, which includes both direct and coexpression based interactions.
Table S8 shows the indirect relationships in 19 ovarian cancer modules, where genes and miRNAs are co-regulated by the same TF. Note that some TFs are not members of the modules. Regulation of miRNAs by TFs is validated by literature evidence (PubMed IDs are shown in the table), and the significance of the regulations of the genes in the modules by TFs was demonstrated using p-values that were obtained based on the ChIP-X database [26]. In many modules, one TF regulates multiple miRNAs and multiple genes. For example, Fig. 3 (A) shows ovarian cancer module 22, in which the TF EGR1 positively regulates several genes (AEBP1, COL1A1, COL5A1, COL5A3, COL6A1, ITGA5, LOXL2, MMP11, MMP2 and THBS2) and miRNAs (miR-214 and miR-152). Fig. 3 (B) shows ovarian cancer module 8, in which EGR1 positively regulates several genes (AQP1, BGN, CALB2, CEND1, COL1A1, COMP, HNT, IRX5, ITGA5 and ITGB1) and miRNAs (miR-214, miR-152, miR-199a and miR-199b) in the module at the same time. In both cases, we can infer that the genes and miRNAs are indirectly related via EGR1.
Table S9 shows another type of indirect relationship in ovarian cancer, where miRNAs regulate TFs, and the TFs regulate genes in 14 ovarian cancer modules. Regulation of TFs by miRNAs was found in the literature, and is shown in the third column of the table. One example of this relationship is shown in Fig. 3 (C): let-7b directly regulates the TF BACH1, and BACH1 regulates several genes (BUB1, CCNA2, CENPF, MCM10, BIRC5, TK1, OIP5, KIF11, RRM2 and CENPA); miR-156b and let-7b regulate the TF E2F1, which regulates several genes and other miRNAs in the modules; and miR-101, miR-29a, miR-29b and miR-29c regulate the TF MYCN, which regulates genes in the module. This module is related to ovarian cancer-related pathways such as those involved in mitosis and the cell cycle.
Similarly, relationships among genes, miRNAs, and TFs in GBM modules are shown in Fig. 4 and in Tables S10, S11, and S12. Table S10 shows 8 miRNAs and their target genes in 12 GBM modules. Genes targeted by miRNAs were highly enriched in these modules. In addition, Tables S11 and S12 show indirect relationships between genes and miRNAs through TFs. Fig. 4 (A) shows one example of an indirect relationship in GBM module 11, where even though genes might not be directly regulated by miRNAs, they are indirectly related via two TFs: RUNX1 and TCF4. For ease of reference, the genes in module 11 were divided into three groups (GA, GB and GC): the TF RUNX1 positively regulates miR-221, miR-222, and genes in GA and GB; miR-155 negatively regulates the TF TCF4; and TCF4 positively regulates genes in GB and GC. Similarly, Fig. 4 (B) shows that miR-29a regulates the TF MYCN, which regulates several genes and miR-93 in GBM module 5. Experimentally validated 438 gene-miRNA interactions from the miRTarbase are shown in Table S13, including 112 direct interactions. In addition, we verified in the literature that miR-21 interacts with BMPR2 and miR-222 interacts with ICAM1 in GBM cell [42].
Fig. S6 summarizes these direct and indirect relationships in the ovarian cancer and GBM modules. These analyses show that, in total, 91% (30 out of 33) of ovarian cancer modules and 94% (51 out of 54) of GBM modules can be explained by direct regulations or indirect relationships, which allows us to understand how genes are regulated in modules.
To determine the functional relevance of modules identified in ovarian cancer, we performed pathway enrichment tests for GO biological processes, KEGG pathways, and BioCarta pathways. We found that 16 out of 33 modules (48.4%) were enriched in at least one function. Table 4 presents enriched functions or signaling pathways for selected modules. Several modules have many enriched functions or pathways related to ovarian cancer, such as the p53 signaling pathway [43], ECM receptor interactions [44], and cell cycles [45]. Tables S14, S15, and S16 present lists of all enriched pathways. As mentioned previously, on average, 19.1% of genes in our modules were cancer genes and 5.7% were ovarian cancer genes. Our further manual literature search revealed that most of the cancer genes in several modules are also ovarian cancer-related genes, suggesting that cancer genes in the modules have a high potential to be ovarian cancer-related genes. In addition, most of the enriched modules had at least one ovarian cancer gene, supporting the idea that all enriched modules might be related to ovarian cancer. Therefore, we extensively analyzed modules 22 and 8 because module 22 has a relatively high fraction of ovarian cancer genes (12.8%) and cancer genes (28.2%) and is enriched for important pathways in ovarian cancer, and module 8 also contains a high fraction of ovarian cancer genes (18.5%), cancer genes (33.3%), and three enriched pathways related to ovarian cancer.
Fig. 5 shows a network representation of module 22, where 25 genes (2 genes are not shown) and 6 miRNAs are presented as nodes. In this module, 5 genes (FN1, MMP2, MMP1, PLAU, and SPARC), colored in green, were identified as ovarian cancer-related genes in the DDOC database. Moreover, the literature showed that 14 genes (ITGA5, COL6A1, THBS2, COL1A1, MMP19, MMP11, CTSK, ECM1, GREM1, VCAN, LOXL2, ADAM12, FAP, and INHBA), colored in pink, are ovarian cancer genes (shown in Table S17) and that these genes have high-average SCC values with at least one miRNA colored in sky blue. Most of the genes enriched in ECM receptor interaction, focal adhesion and proteolysis pathways are green or pink nodes, suggesting that these pathways are closely related to ovarian cancer. The literature confirms that these pathways are related to ovarian cancer [44, 46, 47]. In this module, COL3A1 might be related to ovarian cancer, as it is a known cancer gene targeted by all ovarian cancer miRNAs and belongs to ECM receptor and focal adhesion pathways. COL5A1 and COL5A3 are also likely to be ovarian cancer genes: they are targeted by ovarian cancer miRNAs and enriched in the above pathways, although they are not known cancer genes. Similarly, DPT also might be an ovarian cancer gene, as it is a cancer gene and is targeted by all ovarian cancer miRNAs. Evidence in the literature shows that the previously known ovarian cancer-related miRNAs miR-152, miR-22, and miR-214 are also related to enriched pathways in this module: miR-152 is involved in ECM-receptor-interaction [48, 49], and miR-22 and miR-214 regulate the AKT/PTEN pathway and the p53 signaling pathway [50, 51], which are highly related to the ECM-receptor, focal adhesion and proteolysis pathways [52–55]. These observations support the idea that genes and miRNAs interact with each other and play critical roles at the pathway level.
Fig. 6 illustrates module 8, which contains 34 genes and 8 miRNAs (5 genes are not shown). Because several genes and miRNAs are duplicated in module 22, the same pathways (ECM receptor and focal adhesion) are enriched. However, other important pathways in ovarian cancer, such as the TGF-beta signaling pathway and the complement and coagulation cascades pathway, are also enriched [56]. From this module, COL16A1, COL3A1, and COL1A2 are likely to be ovarian cancer genes, as they are cancer genes and are enriched with at least one pathway containing ovarian cancer genes. For miRNAs, several articles support that miR-199a, miR-199b, miR-214, and miR-382 are involved in the TGF-beta signaling pathway [57–60], and that miR-22 regulates the AKT/PTEN pathway [50, 51], which is closely related to the TGF-beta signaling pathway in several cancers [50, 61].
We performed pathway enrichment tests for modules identified from the GBM data set. Of 54 modules tested, 40 (74%) were enriched with at least one function. Several modules had many enriched functions or pathways related to GBM, such as the p53 signaling pathway [62], the ERBB signaling pathway [63], and the MAPK signaling pathway [64]. Tables S18, S19, and S20 present lists of enriched pathways. As mentioned above, on average, 23.2% of genes in the modules were cancer genes, and 1.2% were GBM genes. A list of GBM genes was extracted from two articles [34, 35]. Similarly to ovarian cancer, the literature results demonstrated that most of the cancer genes in our modules were also GBM-related genes, suggesting that cancer genes in the modules are likely to be related to GBM. We extensively analyzed module 11 because this module contained many GBM-related genes and pathways.
Fig. 7 illustrates a network presentation of module 11, where 74 genes (15 genes are not shown) and 7 miRNAs are presented as nodes. In this module, 4 genes (MAPK1, CDKN1A, SHC1, and ERBB2), colored in green, are GBM genes that were validated by the literature. Most of the genes on the left side of Fig. 7 are cancer genes and are enriched with at least one pathway, including the p53, ERBB, and GRNH signaling pathways. CBLC might be involved in the development of GBM because it is a cancer gene and is contained in the ERBB signaling pathway, an important GBM-related pathway that includes four GBM genes in this module. Additionally, the literature shows that miRNAs in this module function together in the enriched pathways: miR-34a, miR-135, miR-21, mi-222, miR-221, miR-27a, and miR-34b are involved in the p53 signaling pathway [65–71] and the MAPK signaling pathway [71–75], and miR-34a, miR-135, miR-21, miR-222, and miR-221 are involved in the ERBB signaling pathway [76–79].
In Bell et al. [8], ovarian cancer was classified into four ovarian cancer subtypes depending on the expression levels of marker genes: “immunoreactive,” “proliferative,” “differentiated,” and “mesenchymal.” The immunoreactive subtype was identified by the chemokine receptor CXCR3 and its ligands CXCL11 and CXCL10, indicating that considerable expression changes of these genes are important markers for identifying the subtype. The proliferative subtype was identified by the overexpression of transcription factors HMGA2 and SOX11, proliferation marker genes such as MCM2 and PCNA, and underexpression of MUC1 and MUC16, which are known ovarian tumor marker genes. The differentiated subtype was identified by overexpression of MUC16, MUC1 and SLPI. Finally, the mesenchymal subtype was identified by overexpression of FAP and ANGPTL2.
In this study, we used the marker genes described above to determine which subtype was related to the majority of samples in the modules. First, we calculated the average expression level of the marker gene in the samples belonging to the module. Fig. 8 (A) represents the average expression levels of the 12 subtype marker genes across 33 ovarian cancer modules, showing that the expression levels of marker genes vary depending on the modules. As explained in the Methods section, we identified the cancer subtypes of samples by performing a hierarchical clustering with a dynamic tree cut (minModuleSize = 30) using gene expression data, and then we calculated the p-values of marker genes for the identified modules. As shown in Fig. 8 (B), among marker genes in the immunoreactive subtype, CXCL10 is underexpressed in module 5 (p-value: 0.08), and all of the marker genes (CXCL10, CXCL11 and CXCR3) are overexpressed in module 18 (p-values: 0.04, 0.02 and 0.67). Marker genes of the mesenchymal subtype are overexpressed in module 10 (p-values: 0.0003 and 0.0002), module 23 (p-values: 0.03 and 0.66), and module 32 (p-values: 0.02 and 0.09).
In Verhaak et al. [37], GBM was classified into four subtypes depending on the marker genes: “proneural,” “neural,” “classical,” and “mesenchymal.” It was observed that marker genes DLL3, NKX2–2, SOX2, ERBB3, and OLIG2 were overexpressed in the proneural subtype; marker genes FBXO3, GABRB2, SNCG and MBP were overexpressed in the neural subtype; FGFR3, PDGFA, EGFR, AKT2, and NES were overexpressed in the classical subtype; and CASP1, CASP4, CASP5, CASP8, ILR4, CHI3L1, TRADD, TLR2, TLR4, and RELB were overexpressed in the mesenchymal subtype. Note that marker genes of the GBM subtype were overexpressed in samples belonging to that subtype, while marker genes of other GBM subtypes were underexpressed in those samples.
For GBM, we first calculated the average expression levels of marker genes. Fig. 9 (A) presents the average expression levels of the 23 subtype marker genes across 54 GBM modules, and shows the distinct expression levels of marker genes depending on the modules. Fig. 9 (B) shows 6 modules related to GBM marker genes. Marker genes in the proneural subtype (DLL3, NKX2–2, SOX2, ERBB3 and OLIG2) are overexpressed in module 7 (p-values: 0.01, 0.001, 0.0002, 0.07 and 0.004) and module 15 (p-values: 0.001, 0.00003, 0.002, 0.017 and 0.007). All of the marker genes in the mesenchymal subtype (CASP1, CASP4, CASP5, CASP8, ILR4, CHI3L1, TRADD, TLR2 and RELB), except TLR4, are overexpressed in module 22 (p-values: 0.001, 0.001, 0.003, 0.022, 0.048, 0.001, 0.036 and 0.0004). Two marker genes (SNCG and MBP) in the neural subtype are overexpressed in module 32 (p-values: 0.07 and 0.0001), all of the marker genes in the neural subtype (FBXO3, GABRB2, SNCG and MBP) are overexpressed in module 45 (p-values: 0.02, 0.02, 0.11 and 0.02), and two marker genes in the neural subtype (FBXO and MBP) are overexpressed in module 51 (p-values: 0.05 and 0.03). In addition, we obtained the subtype classification of GBM samples from Carro et al. [80], which shares 162 samples in common with our study (proneural: 62, neural: 22, classical: 35 and mesenchymal: 53). When we used these subtypes of samples for the enrichment of a particular subtype in our modules through a hypergeometric test, we confirmed that modules 32 and 45 are closely related to the neural subtype (p-values: 0.053 and 0.018).
Zhang et al. [6] previously showed that their NMF approach outperformed the bi-clique algorithm proposed by Peng et al. [5]. Hence, we assessed the performance of our approach by comparing it with the NMF approach using TCGA ovarian cancer data. By applying our criteria to the modules generated from their approach, we selected modules having at least one gene and two human miRNAs. As a result, we removed 7 out of 50 modules. Fig. 10 shows that the ratio of modules containing enriched pathways in the NMF approach was slightly higher than the ratios of our modules. However, the average number of enriched pathways in our modules was larger than that in the NMF approach.
When we compared enriched pathways, two approaches had 43 common pathways, including ovarian cancer-related pathways such as the immune response, ECM-receptor, and TGF-Beta signaling pathways. In addition, 71 pathways were enriched only in our modules and 67 pathways only in the NMF modules, indicating that the two approaches most likely complement each other and capture different pathways related to ovarian cancer. Table S21 lists the common pathways and pathways enriched in each approach.
Additionally, modules identified by our approach contain more differentially expressed genes and cancer-related genes, because we primarily used differentially expressed genes, which provide more chances to incorporate cancer type-specific genes. In Zhang et al. [6], the modules contain a small fraction of differentially expressed genes and cancer-related genes, because 12,456 genes were used after filtering out genes with small absolute values and little variation. When we computed the overlap ratios of differentially expressed gene, most genes in our modules (79.4%, 617 out of 777 genes) were differentially expressed. However, modules generated by Zhang et al. [6] contained 28.3% (462 out of 1630 genes) differentially expressed genes on average. When we compared ratios of cancer genes, ovarian cancer genes, and ovarian cancer miRNAs in modules, our approach outperformed the NMF approach, as shown in Fig. 10.
The difference between the NMF approach and ours from a methodological viewpoint is that our approach can be more flexibly generalized to incorporate other regulatory components. In our approach, gene-sample modules are first constructed, and then miRNAs regulating genes are added to the modules (generating gene-miRNA modules). To demonstrate the range of our approach, we incorporated DNA copy number aberrations (CNAs) as another type of regulators in gene-sample modules. As a result, 23 out of 58 ovarian cancer gene-sample modules were explained by the regulation of CNAs, and 15 ovarian cancer gene-sample modules were explained by both miRNAs and CNAs. A detailed analysis regarding regulations by CNAs is provided in the Discussion section. By contrast, the NMF approach simultaneously incorporates gene-expression, miRNA expression, gene-gene interaction, and gene-miRNA sequence prediction information. Hence, when other regulators are included, they generate modules, where correlations between genes and regulators are simultaneously high. Indeed, in another paper from the same authors [81], they extended their NMF model to incorporate miRNAs, genes, and methylation of genes. In the generated modules, correlations of the expression levels of these three data sets were coordinately high due to a common basis matrix. Although it is a good approach, it omits modules representing the regulation of genes by a single type of regulators when incorporating multiple regulators.
Additionally, we compared our approach with the Context-Specific MicroRNA analysis (COSMIC) algorithm [82] using TCGA ovarian cancer data. COSMIC combines gene-miRNA target prediction information, mRNA expression, and miRNA expression data. The modules constructed by the COSMIC algorithm consisted of a single miRNA and genes, which indicated that several genes are regulated by the miRNA. When we applied a q-value threshold of < 0.05 to 479 identified modules, 102 modules were obtained. Since COSMIC generates modules consisting of a single miRNA, it is difficult to directly compare COSMIC with our approach. Hence, we applied pathway enrichment tests using GO biological processes and BioCarta and KEGG pathways with a q-value threshold of < 0.05 to these 102 modules, and observed that 25.5% (27 out of 102) of the modules were significantly enriched. This enrichment ratio is lower than the value obtained using our approach (48.4%). However, we need to consider that the higher enrichment ratio in our approach is partially because two studies developed algorithms using different data sets and different assumptions. We incorporated gene-gene interactions and indirect interactions among genes and miRNAs based on mRNA expressions and miRNA expressions, while COSMIC incorporated direct interactions using sequence information of genes and miRNAs, which might reduce false positive interactions. In spite of the differences, the two approaches had 26 common pathways, including ovarian cancer-related pathways such as the ECM-receptor, DNA replication, and the G2 pathway. In addition, 88 and 38 pathways were enriched only in our modules and only in the COSMIC algorithm, respectively. Table S22 lists the common pathways as well as the pathways enriched in each approach.
In this study, we developed an approach to constructing gene-miRNA modules by integrating genes and miRNAs. We applied our approach to ovarian cancer and GBM data sets from the TCGA project. Finally, we constructed 33 modules for ovarian cancer and 54 modules for GBM. We employed gene-gene interactions to include genes with high absolute correlations with genes in the modules, because some important cancer-related genes might not be clustered together by the biclustering algorithm or might not be differentially expressed. Fig. 2 shows that incorporating gene-gene interactions increased the performance in terms of the average number of enriched terms, the number of modules with at least one enriched pathway, and the ratios of cancer-related genes and cancer-related miRNAs. Although we used gene-gene interactions to add biologically relevant genes to modules in the proposed approach, gene-gene interactions can be used to filter out biologically irrelevant genes from modules to reduce false positives. However, because the currently available human gene-gene interactions are not complete, closely related but unidentified genes might also be filtered out. It is an important challenge to incorporate gene-gene interactions to reduce false positive genes in modules, while true relevant genes still remain. We will address this issue in our future work.
Because the identified modules might miss relevant interactions, we measured a potential false negative rate using miRTarbase. Let NG be the number of common genes in the modules and miRTarbase, and let NG_interaction be the number of common genes that interact with the same miRNAs in the modules and miRTarbase. Then, 1 - NG_interaction / NG might be a potential false negative rate. As a result, the rates of false negative were 0.789 (1–118/559) for ovarian cancer and 0.775 (1–316/1405) for GBM, respectively. However, the false negative rate should be adjusted when more accurate miRNA-gene interaction data become available, as this ratio is estimated based on all gene-miRNA interactions from miRTarbase and is not based on the specific cancer type and miRTarbase, which itself contains only a fraction of the gene-miRNA interactions.
In the Results section, we described a functional enrichment test of genes in modules using GO terms, KEGG, and BioCarta pathways. Although we employed a widely used approach in the enrichment test, a hypergeometric test followed by a Benjamini & Hochberg method for multiple comparison correction, several issues that require further improvement still remain. For the first issue, the Benjamini & Hochberg method hypothesizes independence of the terms, while the biological processes in various ontologies represent a hierarchical structure and inter-correlation. Thus, we performed an additional enrichment test for ovarian cancer and GBM modules using TANGO [83], which considers dependencies among biological pathways. It corrects p-values by computing the distribution of enrichment p-values in a large number of randomly generated gene sets of the same size. For ovarian cancer, 16 of 33 modules (48%) were enriched with at least one GO biological process term. For GBM, 28 out of 54 modules (48%) were enriched with at least one term. Tables S23 and S24 list all pathways enriched in each cancer. Further, Fig. S7 shows a comparison of the two approaches (a Benjamini & Hochberg method and TANGO) in terms of the ratio of enriched modules and the number of enriched terms. Although there are small differences in the two approaches, both approaches confirm that a large fraction of our identified modules were enriched with biologically relevant terms. For the second issue, because annotated pathways in GO terms, KEGG, and BioCarta pathways are still incomplete, validations on these pathways might miss biologically related sets of genes. An approach to reveal the pathways unannotated in GO, KEGG and BioCarta is to search for evidence about gene functions in the literature, and then to analyze them collectively. As part of such efforts, we manually searched scientific articles on ovarian cancer-related genes and GBM-related genes (Table S17), and relationships among genes, microRNAs, and TFs (Tables S8, S9, S11, and S12). However, this approach only solves the above problem partially so a more systematic approach is called for. Very few efforts, including LitVan (http://www.c2b2.columbia.edu/danapeerlab/html/software.html), have been developed to carry out an automatic literature search to connect genes with over-represented biological terms in millions of scientific articles. Although we attempted to analyze our modules using such tools, either there are no currently available tools or websites are not connected. Hence, we will further analyze modules for functional enrichments in the future.
Certain oncogenes and tumor-suppressor genes such as P53 and PTEN may play important roles in many cancer types rather than only in specific cancer type. Hence, we examined how many genes in the identified modules were specific to ovarian cancer or GBM. We collected 1393 genes from five cancer type specific databases: the DDOC [33], GBM genes from the literature [34, 35], the Cervical Cancer gene Database (CCDB) [84], the Dragon Database of Genes associated with Prostate Cancer (DDPC) [85], and Lung Cancer Gene Database (LUGEND). We refer to genes contained only in the DDOC as potentially ovarian cancer specific genes. Although these genes are not compared with genes from all types of cancers, it might helpful to remove common cancer genes. Among the 47 DDOC genes included in our ovarian cancer modules, 18 genes were potentially ovarian cancer specific genes. Similarly, among the 32 GBM genes included in our GBM modules, 7 genes were potentially GBM cancer specific genes. Lists of these cancer type specific genes are shown in Table S25.
The accuracy of the identified modules might be largely dependent on the quality of the data sets. In this study, we used TCGA microarray data sets, as in many previous reports they have been used to identify core genes and pathways significantly related to ovarian cancer and GBM. Additionally, when TCGA microarray data sets were compared to RNA-Seq data from the same samples, their expression values were highly correlated in most cases [86] confirming that these data sets are less dependent on a particular platform.
The proposed approach can be generalized to incorporate other regulatory components. To demonstrate the range of applicability of our approach and to provide additional support of biological relevance to the modules, we incorporated somatic DNA copy numbers from the paired patients of gene expression data. For this task, we downloaded TCGA level 3 data sets that provide segmented copy number ratio data compared to normal samples. We first recalculated the copy number aberration ratios for every 1 MB region and filtered out regions whose absolute copy number ratio values are less than 0.2, corresponding to 99.9% among all ratio values. Then, CNA regions were incorporated into gene-sample modules based on correlations between genes in modules and CNA regions. As a result, for the ovarian cancer modules, 23 out of 58 gene-sample modules were explained by the regulation of CNAs, and genes in 15 out of 33 gene-miRNA modules (45%) were also regulated by CNAs, as shown in Table S26. In particular, genes in several modules were located in the regulating CNA regions, indicating that the expression of genes in the modules might be directly affected by CNAs. DNA copy numbers in the chr 1: 32.1 MB - 53.4 MB region were highly correlated with genes in ovarian cancer module 9 with a PCC value of 0.301, and 13 out of 18 genes in the module (CDCA8, C1orf109, AK2, SNIP1, GNL2, RLF, TRIT1, YRDC, RRAGC, PPIE, PSMB2, MED8 and COL9A2) were located in this CNA region. Similarly, the DNA copy numbers in the chr 1: 180.6 MB - 247.9 MB region were highly correlated with genes in ovarian cancer module 23 with a PCC value of 0.319, and most of genes (14 out of 19 genes) in this module were located in this region. Additionally, for ovarian cancer module 29, DNA copy numbers in chr 1: 31.9 MB - 59.1 MB regions have a high correlation value (0.345) with gene in the module, and 78.3% of the genes are located in this region. For GBM, 26 out of 88 gene-sample modules were explained by regulation of the DNA copy numbers shown in Table S27, and 19 out of 54 gene-miRNA modules (35%) were commonly regulated by CNAs and miRNAs.
|
10.1371/journal.pntd.0000401 | Estimating Household and Community Transmission of Ocular Chlamydia trachomatis | Community-wide administration of antibiotics is one arm of a four-pronged strategy in the global initiative to eliminate blindness due to trachoma. The potential impact of more efficient, targeted treatment of infected households depends on the relative contribution of community and household transmission of infection, which have not previously been estimated.
A mathematical model of the household transmission of ocular Chlamydia trachomatis was fit to detailed demographic and prevalence data from four endemic populations in The Gambia and Tanzania. Maximum likelihood estimates of the household and community transmission coefficients were obtained.
The estimated household transmission coefficient exceeded both the community transmission coefficient and the rate of clearance of infection by individuals in three of the four populations, allowing persistent transmission of infection within households. In all populations, individuals in larger households contributed more to the incidence of infection than those in smaller households.
Transmission of ocular C. trachomatis infection within households is typically very efficient. Failure to treat all infected members of a household during mass administration of antibiotics is likely to result in rapid re-infection of that household, followed by more gradual spread across the community. The feasibility and effectiveness of household targeted strategies should be explored.
| Trachoma is a major cause of blindness worldwide and results from ocular infection with the bacterium Chlamydia trachomatis. Mass distribution of antibiotics in communities is part of the strategy to eliminate blindness due to trachoma. Targeted treatment of infected households could be more efficient, but the success of such a strategy will depend on the extent of transmission of infection between members of the same household and between members of the community. In this work, we estimated the magnitude of household and community transmission in four populations, two from The Gambia and two from Tanzania. We found that, in general, transmission of the bacteria within households is very efficient. In three of the four populations, persistent infection within households was predicted by the high level of household transmission (a phenomenon observed in longitudinal studies of trachoma). In all of the studied populations, individuals who live in households with more individuals contribute more to the number of new infections in the community than those who live with fewer individuals. Further studies are required to identify and examine household-targeted approaches to treatment.
| Trachoma is the leading cause of infectious blindness worldwide. Eight million people are visually impaired from the disease and a further 46 million people with active disease are in need of treatment to prevent blindness [1]. Mass drug administration (MDA) with antibiotics (predominantly azithromycin but also topical tetracycline) is one of the four arms of the SAFE strategy, advocated by the World Health Organization (WHO) to control trachoma with the aim of Global Elimination of Blinding Trachoma by 2020 (GET 2020). Large scale vertical control programmes currently operate, such as those through the partners of the International Trachoma Initiative, and control efforts are expected to expand when trachoma control is integrated with that of other neglected tropical diseases [2].
The presence of active disease is currently used to guide trachoma control programs and to evaluate the success of interventions. The WHO advises that if the prevalence in a district of trachomatous inflammation follicular (TF) in a district among 1–9 year-old children is ≥10%, annual treatment of the district along with face-washing and environmental improvement should occur for at least three years until the prevalence of active disease in that age group is reduced to less than 5% [3]. However there is a loose relationship between an individual showing signs of active disease and being infected with the causative bacterium of trachoma, Chlamydia trachomatis. There is typically a lag before the appearance of active disease after an individual has been infected and a persistence of active disease after infection resolves [4],[5]. Signs of conjunctival inflammation may also be the result of other bacterial infections or mechanical irritation [6] and even after infection is eliminated from a community, some individuals may still show signs of active disease [7]. Therefore the proportion of individuals with active disease may not correspond to the proportion of individuals with infection. This was recently illustrated by a study in The Gambia in which the overall prevalence of infection among children under 10 years of age in two regions was 0.3% based on qualitative PCR testing of conjunctival swabs, whereas the prevalence of active disease in this age group was 10.4% [8].
Control programmes that have used MDA as part of their control strategy have had some success [9], and people may also benefit from other bacterial infections being cleared by the antibiotic. Although most antibiotics are currently donated, donation is not universal and is likely to be time-limited. There are also many costs associated with delivering antibiotics in rural settings [10],[11]. Furthermore, MDA results in many uninfected individuals receiving treatment and could promote antibiotic resistance among other bacterial infections such as Streptococcus pneumoniae [12]. Targeted treatment to those infected would reduce the number of drug doses required, potentially reducing the cost of MDA. However, the loose relationship between infection and active disease makes targeted treatment of individuals with active disease ineffective at the population level. Targeting households with at least one member with active disease may be more effective since infection clusters by household [13] and so asymptomatic infections are more likely to be treated. In The Gambia, this strategy has been used as national policy in communities with less than 5% of TF among children aged 1–9 years old (Personal communication, Mr Ansumana Sillah, Manager, Gambian National Eye Care Programme).
Clustering of active trachoma disease by household has been shown to occur in a number of communities [13]–[17] and individuals living with people who have active trachoma are more likely to have active disease than individuals who live with individuals without active disease [15], [18]–[20]. Furthermore, in Jali village in The Gambia, the same serovar of C. trachomatis was predominantly found within a household even though three serovars were present in the community [21], suggesting that transmission between members of the same household is more common than between other members of the community with different serovars. However, the rates of transmission between individuals of the same household and between members of the same community have not been estimated and little is known about the likely impact of targeted treatment of households on transmission of C. trachomatis.
Here we examine the contribution of transmission between members of the same household and that between households of the same population to the incidence of ocular C. trachomatis infection using cross-sectional data on the prevalence of infection from four endemic communities, two in West Africa (The Gambia) and two in East Africa (Tanzania). We discuss the implications of our findings for the resurgence of infection after community-wide treatment and the potential for targeted treatment of households to reduce infection efficiently.
Individuals of all ages from four endemic populations (Upper Saloum District and Jali village in The Gambia and Kahe Mpya and Maindi villages in Tanzania) were examined and conjunctival swabs taken to test for the presence of chlamydial infection using PCR amplification of a target sequence in the common cryptic plasmid of the bacteria. In one community, Maindi village, the presence of infection was based on quantitative PCR amplification of the omp1 gene. Detailed information on the bedroom (Upper Saloum District, Kahe Mpya sub-village and Jali village only), household (Upper Saloum District, Kahe Mpya sub-village and Maindi village only), compound (Jali village and Upper Saloum district only), balozi (Kahe Mpya sub-village and Maindi village only) and village (Upper Saloum district) of the individuals examined was recorded; along with a number of other risk factor for trachoma and clinical signs of disease. Characteristics of these populations and detailed methods have been reported previously [15],[19],[22],[23].
The study in Upper Saloum district was approved by the Gambian Government/Medical Research Council Joint Ethics Committee (SCC 856) and the London School of Hygiene and Tropical Medicine Ethics Committee. Written informed consent was obtained from all individuals. The Kahe Mpya study was approved by the London School of Hygiene and Tropical Medicine committee and the Kilimanjaro Christian Medical Centre, Tanzania. Written consent was obtained. The study in Maindi village was approved by the Johns Hopkins Institute Review Board and the Tanzanian National Institute for Medical Research; all participants provided oral informed consent. Both IRBs approved oral informed consent because many of the rural villagers are illiterate and asking them to sign a document they cannot read is unethical; in the past, unscrupulous persons have had them sign official “documents” that were really signing away their land. Oral consent was witnessed and documented by a member of the team on a study document. These three studies were done in accordance with the Helsinki Declaration. The study in Jali received ethical approval from the joint Gambia Government and Medical Research Council Ethics Committee (SCC 508). All subjects gave oral informed consent that was witnessed and signed by the witness following the standard consent procedures at the time.
Trachoma is a disease in which a fully protective immune response against re-infection is not elicited and so individuals can be repeatedly infected [18],[24]. We therefore chose to describe transmission using a simple Susceptible→Infected→Susceptible (SIS) model, in which the population is categorised into two groups - individuals susceptible to infection (S) or infected individuals (I) - and infected individuals recover to become susceptible again. Household SIS models have been previously examined by Ball [25] and Neal [26].
The probability that a household of size has infected individuals (and susceptible individuals) at time is given by . A susceptible individual can be infected from either an infected member of the community (global transmission) at a rate: , in which is the global transmission coefficient and is the prevalence of infection in the community; or from an infected member of the same household (local transmission) at a rate: , in which is the local transmission coefficient. is multiplied by either the number of infected individuals in the household, , if transmission is assumed to be density dependent (the average number of contacts per individual increases with household size, corresponding to ), or the fraction of infected individuals in the household , representing that the average number of contacts per individual is constant, regardless of household size, and corresponding to . The parameter is therefore the coefficient for density dependence, which in the application described we allow to vary on a continuous scale with .
Individuals recover from infection at a rate , taken as the reciprocal of the average duration of infection. Births and deaths are not included in the model because the duration of infection is relatively short compared to the average human life expectancy.
We can write the difference-differential equation for ,(1)where and .
At endemic equilibrium, assuming the number of households is large (), solving , leads to the recursion:(2)where(3)
The prevalence of infection in the community described by equations (2) and (3) is(4)where is the fraction of households of size in the population. Solving equations (2) and (3) therefore requires the implicit equation to be satisfied at equilibrium.
An epidemic can occur when the household basic reproduction number is greater than 1 [25]. is defined as the mean number of households infected following the introduction of a single infected individual to a randomly chosen household. It is analogous to the basic reproduction number in a non-structured, randomly mixing population [27]. If a household of size is initially infected then is(5)where(6)and is the average across all individuals according to their probability of being in a household of a given size,(7)
Maximum likelihood was used to estimate , and simultaneously. The likelihood, , of a household of size , with individuals infected is given by and the total log-likelihood is the summation of across all households.
The duration of infection was assumed to be 17.2 weeks based on cohort studies of infection with frequent follow-up [4] and was taken to be the prevalence of infection in the cross-sectional survey (i.e. infection in the communities, prior to antibiotic intervention, is assumed to be at endemic equilibrium). The sensitivity of the estimates to the assumed duration of infection was examined for a range of plausible values (12–24 weeks) [4]. Confidence intervals (CI) for each parameter were calculated by assuming that is approximately (chi-squared) distributed [28]. We therefore tested the hypothesis of density dependence in the contact rate by estimating parameter and its confidence intervals; the null hypothesis of density dependence () was contrasted with the alternative hypothesis of frequency dependence (), by ascertaining whether the confidence intervals around the estimate included 0 or 1.
A small number of individuals were not tested for the presence of infection, due to refusal or because they were away travelling. The sensitivity of the estimates to the inclusion of these individuals as members of the household such that they may have contributed to transmission was examined (Text S1). If there were members of a household tested for infection and an additional individuals who were not tested for infection but who contribute to transmission, the probability that individuals were found positive in the sample, given that members of the overall household of size were actually infected (according to a hypergeometric distribution [29]) is:(8)
In this case the likelihood for each household can be modified such that(9)
This assumes that infected individuals are equally likely to be sampled as uninfected individuals. The sensitivity of this assumption was explored using the non-central hypergeometric distribution [29] (Text S1).
The impact of different definitions of a ‘household’ on the estimates of and was examined, from bedroom, household, compound and village for the Upper Saloum District; room and compound for Jali village; room, kaya and balozi for Kahe Mpya sub-village and kaya and balozi for Maindi village. (See below in the Results section for the definitions of ‘kaya’ and ‘balozi’).
The appropriateness of the household SIS model of C. trachomatis transmission was assessed by simulating the number of people infected at endemic equilibrium and the household to which they belong under the model using the estimated parameters and assuming a negative binomial distribution for the underlying household size distribution (with inverse overdispersion parameter equal to (95% CI denoting 95% confidence intervals): , and , for respectively Upper Saloum district and Jali village (The Gambia), and for both Kahe Mpya and Maindi village (Tanzanaia), where corresponds to a random or Poisson distribution; see Text S1). The probability mass function used for the negative binomial is [30]:(10)and when (11)where is the (arithmetic) mean household size (Table 1). Comparison of the model simulations with the observed data was based on the mean intraclass correlation coefficient for the prevalence of infection within households (ICC). The ICC provides a quantitative measure of similarity between individuals within groups and is based upon the comparison of within- and between-group sums of squares from an analysis of variance [31]. One thousand stochastic simulations were run for each setting using the numerical integration package Berkeley Madonna [32].
In The Gambia one household or a cluster of households forms a compound, a unit which is fenced off from the rest of a community. In Upper Saloum district the household unit ranges from 1–55 individuals and the compound ranges from 2–77 individuals. In Jali village the compound unit ranges from 4–70 individuals (household data unavailable). In Tanzania, the household unit is the ‘kaya’, (ranging from 1 to 14 individuals) and on average the unit is smaller than the household unit in The Gambia (Table 1). Kayas which are situated within the same geographical zone are grouped into a ‘balozi’ and share a balozi leader. The number of individuals examined in each community along with the prevalence of infection among households and among individuals is given in Table 1.
The estimates for the global and local transmission coefficients ( and ) along with the density-dependent coefficient, and the household reproduction number are given in Table 2 along with their 95% confidence interval. In Jali the compound unit was used because household data were unavailable. The estimates of and were sensitive to changes in the duration of infection, whereas the estimates of , and the ratio were not affected by changes in the duration of infection (Text S1). Estimates of were close to 1, and in all of the four populations the 95% CIs included 1, consistent with frequency-dependent transmission, such that the number of contacts made by an infected individual was not larger in bigger households. Estimates of the rate of household transmission were large and was greater than in three of the four populations. In all four, individuals from larger households were estimated to contribute more to incidence than those from smaller households (Figure 1). This effect reverses somewhat at very large household sizes in Upper Saloum District where the estimate of (>1) is consistent with a decline in the number of infectious contacts with increasing household size. An average of 71% of incident infections were the result of transmission within the household (with a range of 48%–91%) in the four populations.
The estimate of increased as the definition of the household unit became smaller in size (from village to compound; balozi to household; kaya to room) and the estimates of and decreased (except for in the Upper Saloum District) and remained approximately constant (Text S1). Exclusion of the individuals who were not examined at the moment of sampling but were members of households in the four populations does not change the parameter estimates significantly (Text S1). Assuming infected individuals to be more or less likely to be sampled did not alter the parameter estimates significantly either (Text S1).
The average ICCs from the model simulations were in agreement with the ICCs calculated from the data, suggesting that the simple SIS model of household transmission captures much of the dynamics of C. trachomatis infection in these communities (Table 3).
Clustering of infection by household is an important epidemiological feature of many communicable diseases and is thought to be a key characteristic of trachoma. However, the magnitude of transmission of C. trachomatis between individuals belonging to the same household and that between individuals living in different households but the same community have not, to our knowledge, previously been estimated. Here they are estimated in four different populations by fitting a household model of transmission to prevalence data using maximum likelihood estimation. In these communities an average of 71% of incident infections were the result of transmission within the household, indicating the important role of household transmission in the repeat infections with C. trachomatis that result in progression to trachomatous scarring and blindness. In all four populations, individuals who live in relatively large households (i.e. with many individuals) contribute more to incidence than those who live in households with fewer individuals. Further to this, in the two Gambian populations and in Maindi village, Tanzania, the household transmission coefficient was estimated to be greater than the rate of recovery from infection, such that sustained transmission within the household is possible (Table 2) In other words, the expected duration that a household is infected will be significantly longer than an individual's duration of infection, despite eventual stochastic extinction. The resulting persistence of infection within households permits epidemic spread on average following the introduction of infection into a household (i.e. ) even if the community transmission coefficient is low. For this reason, the dynamics of infection following community-wide treatment may be different from that expected based on a non-structured mathematical model of transmission [33],[34].
The persistence of infection within households due to efficient household transmission and repeated infection of household members has been observed during follow-up of endemic communities [35],[36]. Gradual spread across communities over the course of about one year has been observed following community-wide treatment in several studies [19], [37]–[39]. Such gradual spread is difficult to reconcile with the estimated, rather short duration of infection of individuals with ocular C. trachomatis unless the importance of household transmission is considered. In comparison to the other three populations, in Kahe Mpya sub-village, Tanzania, the estimated household transmission coefficient was lower than both the global transmission coefficient and the rate of recovery from infection , indicating that in this community, persistence within households does not occur. This may be the result of a difference in social behaviour of this community or perhaps a difference in the fly population that may act as a mechanic vector of trachoma. Interestingly, infection was successfully eliminated from this community after two mass treatments with azithromycin and multiple targeted treatment of active disease with topical tetracycline at follow-up time points [7].
The variation of the results within the two studies countries and the small number of populations studied in each country make inter-country comparisons difficult. Generally, the two populations studied in The Gambia were estimated to have higher local (household) and lower global (community) transmission compared to the two populations in Tanzania i.e. household transmission was estimated to be more efficient in The Gambia than Tanzania. The higher household transmission in The Gambia is not intuitive from the differences in geographical distances between households in the two countries. Households are further apart in the Tanzanian populations than those in The Gambia and from this one may think community transmission to be lower in Tanzania. However our work indicates community transmission to be higher in Tanzania. This may be explained by differences in their community structure: Individuals in The Gambia live in much larger households which cluster together to form large compounds. The larger size of the living unit may limit the number of contacts made with the rest of the community therefore sustaining transmission within the household. Moreover, the results of the sensitivity analysis of the household unit definition indicate that the smaller the unit, the higher the amount of community transmission required to sustain transmission. The estimates of the transmission coefficients are less certain for the Tanzanian populations than for the Gambian ones because there are fewer large households, which contribute most information to the estimate of household transmission.
Estimates of the density dependence of transmission found that was close to 1 in all communities, with the 95% confidence intervals containing 1 (Table 2). This indicates that individuals typically have a fixed number of contacts per household regardless of household size (i.e. the risk of infection is proportional to the fraction of infective individuals in a household, rather than the number). This phenomenon has also been shown for other infections, such as Streptococcus pneumoniae and influenza virus [40],[41]. The estimate of from the data from Upper Saloum district is slightly higher than the other estimates (), resulting in a slight decline in the number of contacts per individual with household size, although the confidence intervals include 1 (Figure 1).
The household model used in this work assumes that all individuals mix homogeneously outside their household at the same rate (specific to each setting), such that each household is at equal risk of infection. It ignores any protective immunity against re-infection, does not include infection with different serovars, and assumes that an individual's age does not affect their duration of infection or risk of acquiring infection. It also assumes that each infected individual is equally infectious and does not therefore take into account that some individuals harbor a much higher number of bacteria than others. These assumptions are simplifications of disease transmission and natural history, and in particular, neglect the differences between adults and children in their contribution to transmission. Children have a longer duration of infection and a higher prevalence of infection than adults. Children may also have a different within/between household contact pattern than adults. However, the correspondence between the model simulations and the data indicate that the model is a reasonable description of the household transmission of ocular chlamydial infection. Further work will examine in more detail the contribution of individuals of different ages to the transmission of ocular Chlamydia within households. We have assumed accurate testing of individuals for ocular chlamydial infection and that there was no contamination of the conjunctival swabs. Although, cross-contamination of samples is a risk when using PCR techniques, standard precautions were taken to prevent this [22],[42].
The strategy of mass antibiotic treatment to control trachoma can be costly [10], may result in antibiotic treatment of uninfected individuals and may increase the chance of antibiotic resistance developing, as observed for other bacterial infections [43]–[46]. A control approach which minimises the number of antibiotic doses given out in a community but still has similar effects in reducing prevalence in a community compared to mass distribution would therefore be advantageous. In this paper we have quantified the amount of household and community transmission for the first time and have shown that this leads to persistently infected households in 3 of the 4 study populations. Furthermore, in all four such populations, individuals living in larger households contributed more to transmission than those living in smaller households (Figure 1). This suggests a potential role for the targeted treatment of households more likely to harbor infection. Two field studies have explored the use of the household as the unit for targeting treatment and come to differing conclusions. In Nepal, the reduction in prevalence of active disease after community-wide treatment and after targeted treatment of households containing children showing active disease were not significantly different [47]. In Mali, treatment of those households where at least one child had active disease was significantly less effective at controlling active disease than mass treatment [48]. However, these two studies used active disease as an indicator for treatment and therefore may have missed some children who would have been infected but without showing signs of active disease [4],[5]. Other methods of targeted treatment could also be explored, such as the use of a dipstick assay for rapid diagnosis of the presence of infection, which is currently being developed [49].
The critical role of the household in the transmission and persistence of trachoma demonstrated by our study, along with the high cost of community-wide antibiotic treatment, highlight both the potential and the need for targeted approaches for the treatment of ocular chlamydial infection. Further studies are needed to identify efficient and effective methods to achieve this.
|
10.1371/journal.pbio.1002114 | Regulation of Protein Quality Control by UBE4B and LSD1 through p53-Mediated Transcription | Protein quality control is essential for clearing misfolded and aggregated proteins from the cell, and its failure is associated with many neurodegenerative disorders. Here, we identify two genes, ufd-2 and spr-5, that when inactivated, synergistically and robustly suppress neurotoxicity associated with misfolded proteins in Caenorhabditis elegans. Loss of human orthologs ubiquitination factor E4 B (UBE4B) and lysine-specific demethylase 1 (LSD1), respectively encoding a ubiquitin ligase and a lysine-specific demethylase, promotes the clearance of misfolded proteins in mammalian cells by activating both proteasomal and autophagic degradation machineries. An unbiased search in this pathway reveals a downstream effector as the transcription factor p53, a shared substrate of UBE4B and LSD1 that functions as a key regulator of protein quality control to protect against proteotoxicity. These studies identify a new protein quality control pathway via regulation of transcription factors and point to the augmentation of protein quality control as a wide-spectrum antiproteotoxicity strategy.
| To function properly, proteins must assume their correct three-dimensional shapes. There are numerous mechanisms within the cell, collectively referred to as protein quality control (PQC), that verify proper folding. If abnormal folding is detected, PQC can either help the protein to refold or target it for degradation. Failures in protein folding and PQC lead to the accumulation of misfolded proteins, which often self-associate into large aggregations that are thought to be the underlying cause of several neurodegenerative diseases. In this study, we use the roundworm Caenorhabditis elegans as a model to understand how cells handle disease-associated misfolded proteins. In a large-scale genetic screen, we discovered two suppressor genes, ufd-2 and spr-5, which encode a ubiquitin ligase and a lysine-specific demethylase, respectively. When these two proteins are inactivated, we observe a marked reduction in the toxicity of several misfolded proteins. ufd-2 and spr-5 are conserved in humans (UBE4B and LSD1, respectively), as are their effects on misfolded proteins. We show that UBE4B and LSD1 regulate the activity of protein degradation machineries including the proteasome and autophagosomes. Using microarrays and biochemical analyses, we identify p53 as a key downstream transcription factor that mediates the action of UBE4B and LSD1 on protein clearance. This work establishes p53 as a regulator of proteome integrity and uncovers a new protein quality control pathway that could potentially be exploited to increase the degradation of misfolded proteins in diseased cells.
| Living organisms endure environmental stress and metabolic errors that inflict damage on macromolecules, including DNA and protein, which are either repaired or removed by quality control programs in the cell. Toxicity resulting from protein misfolding and aggregation, known as proteotoxicity, underlies many degenerative diseases, including those affecting the nervous system, such as Creutzfeldt-Jakob disease, Alzheimer disease, Parkinson disease, Huntington disease, frontotemporal dementia (FTD), and amyotrophic lateral sclerosis (ALS) [1,2]. To guard against proteotoxicity, the cell coordinates several major quality control systems, including molecular chaperones, the ubiquitin-proteasome system, and autophagy [3–6]. The regulation of these systems occurs on different scales, from individual proteins to the whole organism [7]. It is tantalizing to envisage enhancing the protein quality control systems to defend against proteotoxicity associated with neurodegenerative diseases. However, how the protein quality control might be harnessed in the cell to alleviate proteotoxicity-associated neurodegeneration is not yet fully understood.
Mutant Cu/Zn superoxide dismutase (SOD1), linked to ~20% of familial ALS, represents a simple molecular model for protein misfolding and aggregation. The wild-type (WT) SOD1 protein has a stable β-barrel structure with a two-state folding process, whereas mutant SOD1 proteins show a heightened propensity to aggregate in vitro and in vivo [8–10]. There is increasing evidence that heightened propensity to misfold and aggregate is a common feature of ALS/FTD-associated proteins, including TAR DNA binding protein 43 (TDP-43) and fused in sarcoma (FUS) [11–13]. Identifying mechanisms that suppress the toxicity of protein misfolding and aggregation may help elucidate the pathogenesis of neurodegenerative diseases and provide potential targets for correction.
Here we took advantage of a Caenorhabditis elegans model that expresses neuronal ALS-linked SOD1 mutant proteins and develops robust movement defects to perform an unbiased genetic screen for potent suppressors of the behavioral defects. We identified mutations in two genes, ufd-2, encoding a ubiquitin ligase, and spr-5, encoding a lysine-specific demethylase, that synergistically attenuate the neurotoxicity of mutant human SOD1 proteins. The actions of the suppressor genes are conserved in Drosophila, and they protect against proteotoxicity initiated by diverse mutant proteins, including TDP-43, FUS, and the polyglutamine (polyQ) tract. Furthermore, we found human orthologs of these modifiers to be part of a pathway regulating protein quality control in mammalian cells. Further analysis showed that this pathway acts through the transcription factor p53, which mediates cellular stress responses. Together, these results identify a new mechanism involving previously unrecognized players, which the cell utilizes to augment protein quality control.
To better understand the regulatory mechanisms that mitigate an increased load of misfolded proteins, we conducted a forward genetic screen for cellular factors that alleviate such stress and relieve cells from proteotoxic insults. This screen took advantage of a C. elegans model of ALS, in which the neuron-directed expression of the ALS-linked, G85R mutant human SOD1 (SOD1G85R) protein leads to its aggregation into misfolded soluble oligomers and larger insoluble aggregates [14,15]. Misfolded SOD1G85R protein is highly toxic, leading to age-dependent synaptic dysfunction, neurodegeneration, and severely impaired movement in the worms [14]. This severe locomotor defect allowed us to perform a large-scale screen for genes that suppress neurodegeneration and improve worm locomotion. In these experiments, we treated homozygous transgenic SOD1G85R C. elegans with ethyl methanesulfonate (EMS) to induce genomic mutations, and the mutagenized P0 hermaphrodites were allowed to self-reproduce for two generations (Fig. 1A). Next, in the F2 offspring, which contain both heterozygous and homozygous suppressor mutations, we selected individual C. elegans based on a salient improvement in the locomotion on a background of poorly moving populations. The potential suppressor clones were bred through until 100% of progeny showed phenotypic improvements and were then subjected to further analysis (Fig. 1A).
After screening >105 haploid genomes, we isolated hundreds of independent strains with markedly improved locomotion. Most of these strains were dismissed upon closer examination because they showed a reduction in the expression of a green fluorescent protein (GFP) reporter gene that had been coinjected as an internal reference and expressed independently in the pharynx, suggesting silencing of the transgene cassette. Among the few remaining suppressor strains that survived this test, one designated M1 showed potent suppression of the locomotion defect when compared with the parental SOD1G85R line, reaching ~76% of the locomotion robustness of the SOD1-WT transgenic line (Fig. 1B and S1 Movie). Such strong recovery of locomotion was apparently not a consequence of diminished SOD1G85R transgene expression because SOD1G85R mRNA and protein levels were unchanged between the parental and M1 mutant strains (Fig. 1C). Further segregation analysis of M1 indicated that more than one genetic locus, in addition to the SOD1 transgene on chromosome IV, was linked to the suppressor phenotype, suggesting a rare multigenic suppressor underlying the suppressor phenotype.
To map and identify genes responsible for the suppression of the locomotor defect, we carried out single-nucleotide polymorphism (SNP) mapping [16]. SNP mapping localized the M1 suppressor mutations to two linkage regions: a 2.2-Mb interval on chromosome I and an 8-Mb interval on chromosome II (Fig. 1D). Next, we performed two rounds of deep sequencing on the M1 strain genomic DNA [17], attaining a 27-fold coverage. When the M1 genomic DNA sequencing data was aligned with the C. elegans reference genome, we found over 200 variants in the two linkage regions. Next, we performed deep sequencing of the parental strain carrying only the SOD1G85R transgene, with 7.5-fold coverage. Comparison of the parental and M1 genomic sequences indicated that most of the nonreference variants existed prior to the EMS mutagenesis and thus were not responsible for the suppressor phenotype. Our analysis pinpointed two variants as likely candidates for the suppressor mutations in M1: in the chromosome I linkage region, there is only one missense mutation, G1937A, resulting in a single amino acid change (R646Q) in the gene suppressor of presenilin 5 (spr-5); and on chromosome II, among the few remaining variants is one nonsense mutation, G2472A, which results in a premature stop (W824X) in the gene ubiquitin fusion degradation 2 (ufd-2) (Fig. 1D and 1E).
To examine the role of the double mutations ufd-2(W824X) and spr-5(R646Q) in the suppression of mutant SOD1-mediated neurotoxicity, we performed a series of genetic, biochemical, and behavioral analyses. ufd-2 encodes a U-box type ubiquitin ligase, and the W824X mutation results in a truncated protein lacking the C-terminal U-box (Fig. 1E). spr-5 encodes a lysine-specific demethylase, and the R646Q substitution occurs at a highly conserved residue in the C-terminal portion of an amine oxidase-like (AOL) domain (Fig. 1E). While either ufd-2(W824X) or spr-5(R646Q) alone did not lead to the strong locomotor defect-suppressing phenotype in the M1 strain, the double mutation ufd-2(W824X) and spr-5(R646Q) segregated perfectly with the M1 phenotype, recapitulating the full rescuing effect of the suppressor.
To confirm ufd-2 and spr-5 as the suppressor genes, we obtained independent null alleles of the two genes: a deletion mutation, ufd-2(tm1380), that lacks 80% of the protein at the C-terminus [18] and a nonsense mutation, spr-5(by134), that lacks the C-terminal half of the protein (Fig. 1E) [19]. When crossed to the mutant SOD1 strain, the single allele of ufd-2(tm1380) provided a moderate, 2-fold locomotor improvement, and less improvement was seen for the single allele of spr-5(by134) (Fig. 1F). However, combining the alleles of spr-5(by134) and ufd-2(tm1380) completely recapitulated the strong locomotor-defect-suppressing phenotype observed in the M1 strain (Fig. 1F). Total levels of SOD1G85R protein were similar among the WT, single-, and double-mutant strains (S1B Fig.). However, further analysis after fractionation by solubility revealed that the insoluble level of SOD1G85R, which accounts for less than 2% of total proteins, was decreased by the spr-5(by134);ufd-2(tm1380) mutations, while the soluble level of SOD1G85R remained unchanged (Fig. 1G). Finally, we found that restoring the function of either ufd-2 or spr-5 alone by expressing transgenic wild-type ufd-2 or spr-5 under a neuron-specific promoter from the synaptobrevin (snb-1) gene completely blocked the protection in the M1 strain (Fig. 1H), indicating that it was the loss of function in these two genes and not any other background mutation that was responsible for the suppressor phenotype. Taken together, these results establish that the synergistic loss of ufd-2 and spr-5 creates a potent novel suppressor of the neurodegenerative phenotypes in the SOD1 C. elegans model of ALS, which we have termed the spr-5– and ufd-2–dependent neurodegeneration suppressor (SUNS).
Next, we asked whether the SUNS suppressor genes affect the aggregation and toxicity of other misfolded proteins. In transparent C. elegans models, yellow fluorescent protein (YFP) fusions of several disease-relevant, aggregation-prone proteins, such as SOD1G85R-YFP [14], TDP-43c25-YFP [20], and PolyQ-YFP [21], have been used to facilitate the visualization of their protein aggregation. These fusion proteins form fluorescent protein aggregates readily visible in live C. elegans, and, when present in neurons, these protein aggregates correlate with the toxicity to the animals as manifested in their locomotor defects [14,20,21]. To determine whether the SUNS mutant reduces the toxicity associated with these protein aggregates, we introduced the double-null mutations ufd-2(tm1380);spr-5(by134) into the strains that pan-neuronally express SOD1G85R-YFP, TDP-43c25-YFP, or PolyQ-YFP. Indeed, loss of ufd-2 and spr-5 function resulted in a marked reduction in the neuronal protein aggregation when compared with controls. Reduction in the number and intensity of protein aggregates was evident in the change in the fluorescent inclusions in the head and ventral cord regions of the SUNS-mutant C. elegans (Fig. 2A–2C). Consistently, the locomotor phenotypes in these C. elegans strains were significantly improved by the introduction of the SUNS mutations, ufd-2(tm1380);spr-5(by134) (Fig. 2A–2C).
To investigate the biochemical states of the misfolded proteins in the C. elegans models, we performed a protein solubility assay by differentially extracting and sedimenting worm lysates into soluble supernatants and insoluble pellets. The worm pellet fraction is enriched with sedimentable large SOD1 protein aggregates, whereas the supernatant fraction contains smaller aggregates and oligomeric species (S1A Fig.) [14,15]. Western blot analysis of both supernatant and pellet fractions displayed a significant decrease in the levels of misfolded SOD1G85R-YFP, TDP-43c25-YFP, and PolyQ-YFP in the ufd-2(tm1380);spr-5(by134) double mutant when compared with controls (Fig. 2A–2C). Compared with the untagged SOD1G85R (the soluble level of which was unchanged, S1B Fig.), SOD1G85R-YFP was significantly reduced in its soluble fraction by ufd-2(tm1380);spr-5(by134). This suggests that soluble SOD1G85R-YFP proteins were degraded more rapidly than the untagged SOD1G85R, consistent with the notion that the YFP tag may increase the misfolding of fusion proteins and therefore decrease their total protein levels (S1C Fig.). Taken together, these data indicate that the ufd-2;spr-5 double mutations are capable of reducing the proteotoxicity of various misfolded proteins associated with neurodegeneration, suggesting a wide-spectrum effect of the suppressor.
To investigate whether this effect was evolutionarily conserved, we assessed the actions of ufd-2 and spr-5 in transgenic Drosophila models expressing known ALS disease-causing human TDP-43M337V or FUSR521C proteins. These models have been previously shown to cause photoreceptor degeneration and rough-eye phenotypes [22,23]. We found that knockdown of CG9934, a Drosophila homolog of C. elegans ufd-2 and human ubiquitination factor E4 B (UBE4B), rescues the TDP-43M337V-induced degenerative rough-eye phenotype, resulting in a smoother eye appearance (Fig. 2D, left) and restoration of pigmentation (Fig. 2E). Similarly, knockdown of the ufd-2 homolog CG9934 corrected the ommatidial defects (Fig. 2D, middle) and pigmentation loss induced by FUSR521C (Fig. 2E). Additionally, we observed that knockdown of Drosophila Su(Var)3-3, the homolog of C. elegans spr-5 and human lysine-specific demethylase 1 (LSD1), rescued the degenerating eye phenotypes induced by FUSR521C (Fig. 2D, right, and Fig. 2E). Neither of the suppressors changed the protein expression levels of TDP-43M337V or FUSR521C (S1D Fig.). Taken together, these results indicate that the loss of ufd-2 and spr-5 homologs also suppresses proteotoxicity-related phenotypes in diverse Drosophila models.
Homologs of ufd-2 and spr-5 are present in all eukaryotes. UBE4B and LSD1 are the human orthologs of ufd-2 and spr-5, respectively. UBE4B and LSD1 share 32% and 29% sequence identity with ufd-2 and spr-5, respectively, and all the major protein domains are conserved (Fig. 1E). To determine whether UBE4B and LSD1 affect protein aggregation in mammals, we used a protein solubility assay that was established to characterize the aggregation of mutant SOD1 in HEK293T cells (S2A Fig.) [9,24]. This assay utilizes the aggregation-prone SOD1G85R protein, which migrates faster on SDS-PAGE than its WT counterpart, as a reporter of protein aggregation, with WT SOD1 serving as an internal control. We knocked down UBE4B and/or LSD1 with multiple RNA interference (RNAi) oligonucleotides and analyzed the levels and solubility of SOD1G85R protein in HEK293T cells. Cell lysates were subjected to ultracentrifugation to separate insoluble pellets, which contain large and sedimentable SOD1G85R aggregates, from soluble supernatants, which contain correctly folded native proteins, misfolded proteins, and small oligomeric aggregates (S1A Fig.). The WT SOD1 protein remained in the supernatant (S) in all tested conditions, but a significant portion of SOD1G85R protein (25%–30%) was enriched in the insoluble pellet (P) fraction (Fig. 3A). Knockdown of both UBE4B and LSD1 significantly decreased levels of SOD1G85R in both supernatant and pellet fractions (Fig. 3B), consistent with reduction of total SOD1G85R proteins (S2B Fig. and S2C Fig.). The UBE4B/LSD1 knockdown did not affect the protein levels of WT SOD1. These results were consistent with the notion that the mutant SOD1G85R had a much larger fraction of misfolded and aggregated proteins that were sensitive to UBE4B- and LSD1-dependent clearance than the WT SOD1 protein. In agreement with the observations in C. elegans and Drosophila, the effects of UBE4B/LSD1 knockdown on enhanced protein clearance was not specific to mutant SOD1G85R but also occurred with other aggregation-prone proteins, including TDP-43Q331K (S2D Fig. and S2E Fig.), indicative of a general effect on protein quality control. Additionally, single knockdown of either LSD1 or UBE4B also resulted in decreased steady-state levels of SOD1G85R, with the UBE4B knockdown producing a stronger effect than LSD1 (Fig. 3B). However, the double knockdown produced an even more pronounced decrease in the SOD1G85R levels, indicating an additive effect between UBE4B and LSD1.
To determine whether this decrease in the levels of misfolded and aggregated proteins was a consequence of increased degradation of proteins, we performed cycloheximide chase experiments. Cells were transfected with SOD1G85R together with either nontargeting control small hairpin RNAs (shRNAs) or a mix of UBE4B and LSD1 shRNAs, and the clearance of the SOD1G85R protein was quantified. Cycloheximide was used to block de novo translation, and the amount of SOD1G85R protein remaining in the supernatant at the indicated time points after the translation block was determined by SDS-PAGE and western blotting (Fig. 3C). The UBE4B and LSD1 double knockdown decreased the half-life of SOD1G85R from 8.5 h to 5 h, indicating that increased clearance of the mutant protein underlies the reduction of the protein aggregates (Fig. 3D and 3E). Similarly, the reduction of TDP-43Q331K by knockdown of UBE4B and LSD1 was also attributable to increased protein clearance, as shown by cycloheximide chase experiments (S2F Fig. and S2G Fig.), further suggesting that a general enhancement of protein quality control is the consequence of the loss of function of the suppressor genes.
To identify the downstream effectors of UBE4B and LSD1 in the antiproteotoxicity pathway, we performed a comprehensive transcriptional analysis using the cell-based SOD1 misfolding model. We treated HEK293T cells with shRNAs targeting either UBE4B or LSD1 alone or UBE4B and LSD1 simultaneously in the presence of SOD1G85R proteins. Upon confirmation of reduction in UBE4B and LSD1 protein levels, total RNA was isolated and subjected to microarray profiling of the whole human transcriptome (S3A Fig.). In triplicate samples of the three knockdown conditions and nontargeting controls, differentially regulated genes and pathways were analyzed in unbiased approaches to identify those that convey the UBE4B/LSD1-mediated activation of protein quality control systems.
The most intriguing observation in our unbiased microarray analysis did not concern individually regulated genes but instead related to the upstream regulators that elicited characteristic pattern of mRNA level changes in a whole pathway or network. By employing the Ingenuity Pathway Analysis (IPA) algorithm to compare the predicted pattern of changes and the actual changes in these genes in our microarray profiles, we identified a number of upstream regulators whose downstream targets are significantly changed (z-score ≥ 2.0) in UBE4B and LSD1 single or double knockdowns (Fig. 4A). Among these upstream regulators, only a few were shared by more than one experimental condition, and remarkably, p53 was the only upstream regulator common to all three conditions (Fig. 4A and S1 Table). In the UBE4B and LSD1 double-knockdown condition, a large number of p53 target genes were affected, and importantly, a large fraction were changed in the directions that statistically suggest an activation of the p53 transcription factor (Fig. 4B, S3B Fig., S3C Fig., S2 Table). We examined a sample of 11 p53 target genes in HEK293T cells and confirmed by reverse transcription-quantitative polymerase chain reaction (RT-qPCR) that their expression levels were consistent with the microarray dataset (Fig. 4C). Since UBE4B is a ubiquitin ligase that decreases the stability of p53 [25–27], we examined the total protein levels of p53 in the UBE4B and LSD1 double-knockdown cells. We detected significantly increased p53 protein levels in the UBE4B and LSD1 double-knockdown cells when compared with the mock-knockdown control, indicating a stabilization of the p53 protein (Fig. 4D).
To further confirm that the single or double knockdowns of UBE4B and LSD1 were activating p53-mediated transcription, we used a firefly luciferase (p53RE-luc) reporter construct carrying p53-responsive elements in its promoter. Knockdown of either UBE4B or LSD1 increased p53 transcriptional activity. However, the simultaneous knockdown of both UBE4B and LSD1 resulted in an even higher p53 transcriptional activity (Fig. 4E), consistent with the synergistic antiproteotoxicity effects of knocking down both UBE4B and LSD1. To examine if the increased luciferase activity reflected p53-dependent transcriptional activation, we expressed MDM2, an E3 ubiquitin ligase and negative regulator of p53, or β-galactosidase as a control, together with the p53 activity reporter. The introduction of MDM2 significantly reduced p53-dependent transcriptional activation of the luciferase reporter under the UBE4B and LSD1 double-knockdown condition (S4A Fig.), confirming the specificity of the regulation of p53 by UBE4B and LSD1.
LSD1 demethylates the p53 protein, and loss of LSD1 increases K370-p53 dimethylation, an activating post-translational modification specifically recognized by 53BP1 [28]. To determine whether the activation of p53 resulted partially from its enhanced interaction with 53BP1, we co-immunoprecipitated p53 and 53BP1 from HEK293T cells in which LSD1 and UBE4B were previously knocked down. An increased amount of 53BP1 was pulled down by an equal amount of p53 protein in the double-knockdown cells when compared with the control, indicating an increased interaction between p53 and its coactivator, 53BP1 (Fig. 4F). To confirm directly that the dimethylation on the p53 protein is increased, we performed western blotting using a dimethyl K370-p53-specific antibody, and detected an increase in the level of dimethylated p53 in the double-knockdown samples compared with the control and compared with the total p53 (S4B Fig.). In sum, these data demonstrate that p53, as a transcription factor, is significantly elevated and activated by the knockdown of UBE4B and LSD1.
Among the genes that were up-regulated by the UBE4B and LSD1 double knockdown in our microarray data set, there were a few that had been reported to be important for protein quality control, including forkhead box O3 (FOXO3a), FOXO4, and proteasome 26S subunit non-ATPase11 (PSMD11) (S3C Fig. and S2 Table). FOXOs are a family of transcription factors invoked in protein quality control [20,29,30], and the 19S proteasome component PSMD11 is a critical regulator of proteasome activity [31,32]. It is notable that FOXO3a is transcriptionally up-regulated by p53 [33], and PSMD11 is transcriptionally induced by FOXOs [31,32]. We confirmed through RT-qPCR that FOXO3a, FOXO4, and PSMD11 were all transcriptionally up-regulated when UBE4B and LSD1 were knocked down (Fig. 4G), linking these positive regulators of protein quality control downstream of p53 to the UBE4B- and LSD1-dependent antiproteotoxicity activity.
Since FOXO3a is a transcription factor downstream of p53 and positively regulates protein quality control, we examined the FOXO3a-mediated transcriptional activity by employing a previously established luciferase reporter that is driven by the forkhead-responsive element (FHRE) [34]. A constitutively active form of FOXO3a, FOXO3a-TM, was co-expressed with the FHRE-luciferase reporter in HEK293T cells to measure the transcriptional activity of this particular FOXO member. Consistent with the p53 activation, we found that the FOXO3a activity is induced most strongly when both UBE4B and LSD1 are simultaneously knocked down (Fig. 4H). Similar results were observed with another FOXO-family member, FOXO1 (S4C Fig.). The daf-16 gene is the sole C. elegans ortholog of the mammalian FOXO family [35,36]. To examine a role of daf-16/FOXO in the SUNS pathway in C. elegans, we built a quadruple mutant strain carrying the SOD1G85R transgene, the SUNS mutations spr-5(by134) and ufd-2(tm1380), and loss-of-function mutation daf-16(mu86). In C. elegans locomotor assays, daf-16(mu86) partially blocked the rescuing effect of the SUNS mutations on SOD1G85R toxicity, suggesting that the daf-16 gene activity is partially required for the suppressor activity (S4D Fig.).
To determine whether the enhanced clearance of SOD1G85R upon the knockdown of UBE4B and LSD1 reflects an increase in proteasome-mediated degradation, we examined abundance of several proteasomal subunits. Consistent with our transcriptome analysis that demonstrated an increase in the level of PSMD11 RNA, the PSMD11 protein level was significantly increased in the double-knockdown cells (Fig. 5A). Both PSMD11 and PSMD4 (proteasome 26S subunit, non-ATPase, 4) are resident to the 19S regulatory proteasome particle, and we found that the PSMD4 protein level was also significantly increased. Moreover, components of the 20S core proteasome particle, including 20Sα3, 20Sβ1, and 20Sβ5, were also significantly increased.
To determine whether the increase in the quantities of proteasome subunits corresponds to augmented proteasomal activity, we measured the chymotrypsin-like proteasome activity in lysates derived from HEK293T cells with single or double knockdown of UBE4B and LSD1 using a luciferase assay (S5A Fig.). The proteasomal activity was increased in LSD1 or UBE4B single-knockdown cells, but the highest activity was observed in double-knockdown cells (Fig. 5B). This finding is consistent with the results of C. elegans suppressor studies and protein solubility assays in mammalian cells, in which both UBE4B and LSD1 are required for the maximal effects of the suppressors. Consistent with the activation of proteasomal activity by the double knockdown of UBE4B and LSD1, inhibition of the proteasome by treating cells with MG132 blocked the degradation of SOD1G85R conferred by the UBE4B and LSD1 knockdown (S5B Fig. and S5C Fig.). Thus, the knockdown of UBE4B and LSD1 significantly increases both the subunit quantity and the activity of proteasomes.
In addition to the activation of the proteasome, we asked whether autophagy is also up-regulated in the UBE4B and LSD1 double-knockdown cells. To measure autophagic activity, we employed a Gaussia luciferase (GLuc) release assay [37,38] that reports the autophagy-dependent ATG4B cleavage of an actin-tethered actin-LC3-GLuc-fusion protein and its subsequent release from the cell into the medium (S5D Fig. and S5E Fig.). HCT116 cells, which are amenable to this assay, were transfected with shRNA constructs to knock down UBE4B and LSD1, with the LC3-GLuc plasmid used to measure the cleavage of LC3 and the constitutively secreted control (cytomegalovirus-secreted embryonic alkaline phosphatase [CMV-SEAP]) for transfection/secretion normalization (S5D Fig. and S5E Fig.). The LC3-dependent GLuc activity, measured over a period of 72 h, showed a 2-fold increase in ATG4B proteolytic activity at the end of the time course demonstrating the activation of autophagy by the knockdown of UBE4B and LSD1 (Fig. 5C). The cells transfected with the noncleavable, LC3-less fusion, the Act-GLuc construct, showed only background levels of Gluc activity, similar to the levels observed in nontransfected cells.
To confirm that autophagy was activated by the UBE4B and LSD1 double knockdown, we measured LC3-II accumulation by western blotting in cells. Double knockdown samples showed increased levels of LC3-II compared with controls, even when LC3-II was elevated by treatment with 3-MA (Fig. 5D). Together, these data demonstrate that a UBE4B- and LSD1-dependent protein quality control pathway similar to that in C. elegans also operates in mammalian cells, since a reduction in these two enzymes promotes the removal of aggregating proteins through enhanced post-translational quality control systems involving the proteasome and autophagy.
Until now, p53 has not been associated with antiproteotoxicity activity. p53 has been shown to regulate autophagy, but in opposing directions [39]. Our microarray analysis and subsequent studies establish a correlation between the activation of p53-mediated transcription and enhanced protein quality control conferred by the knockdown of UBE4B and LSD1 (Fig. 5). It has been demonstrated that p53 is a target of polyubiquitination by UBE4B and demethylation by LSD1, and each of these functions decreases the p53 activity [25,28]. Thus, p53 has emerged as a potential effector that mediates the synergistic action of UBE4B and LSD1 in the antiproteotoxicity pathway.
To determine whether p53 directly protects against proteotoxicity, we first used small molecule activators of p53 in the cell-based SOD1G85R protein aggregation assay. Tenovin-1 activates p53 by inhibiting the SIRT1/2 deacetylase and therefore promoting p53 acetylation, thereby increasing its stability and activity [40]. CP-31398 is another drug that activates p53 by stabilizing the p53 DNA-binding domain in an active conformation and inhibiting its ubiquitination [41,42]. Both p53 activators significantly reduced the amount of SOD1G85R protein in the supernatant and pellet fractions (Fig. 6A, S6A Fig., and S6D Fig.) at various concentrations of Tenovin-1 (0.4–1.2 μM) and CP-31398 (2–4 μg/ml). Furthermore, to confirm that Tenovin-1 and CP-31398 acted through the activation of p53, we reduced p53 levels via shRNAs while testing the clearance of SOD1G85R protein with the drug treatment. The p53 reduction blocked the effect of either Tenovin-1 or CP-31398 on promoting SOD1G85R clearance, indicating that the drugs act through p53 (S6G Fig.).
To investigate the mechanism by which the activation of p53 by Tenovin-1 and CP-31398 reduced the levels of SOD1G85R proteins, we asked whether autophagy was activated by these drug treatments. In agreement with a previous report that CP-31398 activates autophagy [43], we observed that increasing concentrations of either Tenovin-1 or CP-31398 up-regulated LC-II protein levels (S6H Fig.), consistent with the activation of autophagy. To further validate that Tenovin-1 and CP-31398 act post-translationally to promote the clearance of SOD1G85R proteins, we performed cycloheximide chase experiments and confirmed the increase in the degradation of SOD1G85R proteins upon treatment of either of the p53 activator drugs (S6B Fig., S6C Fig., S6E Fig., and S6F Fig.).
Conversely, we asked whether reducing p53 activity would negatively affect the clearance of SOD1G85R proteins. First, we performed the SOD1G85R protein solubility assay in cells in which p53 was reduced by RNAi. We found that partial removal of p53 in HEK293T increased SOD1G85R protein levels in both the supernatant and pellet fractions (Fig. 6B). Next, using a human HCT116 cell line in which p53 was knocked out, we asked how the complete removal of p53 affected the clearance of misfolded SOD1G85R. Unlike the WT SOD1 protein, whose level was not affected by the absence of p53, the SOD1G85R mutant protein was significantly increased in the p53 knockout cells when compared with the controls, indicating that endogenous p53 promotes the clearance of misfolded proteins (Fig. 6C).
To determine whether p53 mediates the UBE4B- and LSD1-dependent clearance of the SOD1G85R proteins, we knocked down UBE4B and LSD1 with or without the removal of p53 and then examined the levels of SOD1G85R. We applied both transient and stable shRNA knockdown by creating an inducible, stable HEK293T cell line expressing tetracycline-regulated shRNAs against UBE4B, LSD1, and p53. Both transient and stable knockdown of p53 significantly reversed the SOD1G85R protein clearance conferred by the UBE4B and LSD1 knockdown (Fig. 7A and S7A Fig.). This result was confirmed with an independent set of shRNAs against UBE4B and LSD1 (S7B Fig.) and by knockdown of UBE4B and LSD1 in HCT116 cells with or without the p53 gene knocked out (S7C Fig.). Notably, unlike HEK293T cells, HCT116 cells do not express SV40 large T-antigen (LT-Ag), a regulator of p53 stability [44], suggesting that the action of UBE4B and LSD1 does not require LT-Ag. Taken together, these results demonstrate that p53 is required for the UBE4B- and LSD1-dependent clearance of the SOD1G85R proteins, and it acts downstream of UBE4B and LSD1 to positively regulate the clearance of misfolded proteins.
To confirm that p53 can modulate proteotoxicity in vivo, we used the Drosophila TDP-43M337V neurodegeneration model as described earlier (Fig. 2D) [22]. Either knockdown of p53 by RNAi or expression of a dominant negative form of Drosophila p53 (p53.R155H) [45] with the GMR-Gal4 driver exacerbated the TDP-43M337V-induced eye phenotype (Fig. 7B). This aggravation of the phenotype was evident in increased loss of pigmented ommatidia and, in p53 RNAi flies, the appearance of necrotic patches, which were observed at low penetrance (Fig. 7B). Expression of the dominant negative p53.R155H transgene on its own, in a wild-type background, did not cause any eye phenotype. Together, these results indicate that endogenous p53 plays a role in reducing TDP-43M337V proteotoxicity in the Drosophila eye.
We further tested whether p53 activation would alleviate SOD1G85R-induced neurotoxicity. We employed a previously characterized SOD1 neurotoxicity assay [46], in which spinal cord primary motor neurons were prepared from rat embryos, maintained on astrocyte monolayers supplemented with neurotrophic factors, and stained with a mature motor neuron marker, the neurofilament H (NF-H) antibody SMI-32. Expression of SOD1G85R via a neuron-specific herpes simplex virus (HSV) vector reduced the survival of motor neurons significantly by approximately 50% in contrast to the HSV-LacZ control over 5 days (Fig. 7C). When treated with the p53 activator Tenovin-1, the motor neurons showed protection from SOD1G85R-induced proteotoxicity, as compared with the vehicle control. Among the various concentrations tested, 0.8 μM of Tenovin-1 completely blocked the neurotoxicity of SOD1G85R, with minimal toxicity from the drug itself (Fig. 7C). These results confirm that the activation of p53 provides protection against the toxicity of misfolded proteins in neurons.
This study uncovers a previously unknown pathway that mitigates the toxicity of misfolded proteins by boosting protein quality control systems. Using a C. elegans genetic screen for suppressors of neurotoxicity induced by mutant SOD1, we have identified the SUNS pathway, which is mediated by two conserved genes, ufd-2/UBE4B and spr-5/LSD1. From C. elegans to human cells, inactivation of the highly conserved lysine-modifying enzymes ufd-2/UBE4B and spr-5/LSD1 is shown to enhance the clearance of misfolded proteins. In mammalian cells, the pathway mediated by UBE4B and LSD1 acts to improve the cellular protein quality control by increasing proteasomal and autophagic activities (Fig. 7D). Although it was initially surprising that loss of ubiquitin ligase UBE4B and lysine-specific demethylase LSD1 protects against proteotoxicity, further results reveal positive downstream effectors, including transcription factors, with a novel implication of p53 in antiproteotoxicity activities. Together, these results demonstrate the capacity of a cell to reprogram its protein quality control through transcriptional regulation to defend against proteotoxicity.
We isolated the SUNS C. elegans mutant based on the potent suppression of SOD1-induced neurotoxicity. The suppressor was found to significantly enhance the removal of misfolded proteins, underscoring the critical role of protein misfolding in SOD1-mediated neurodegeneration. The enhanced clearance also applies to other misfolded proteins, such as TDP-43, FUS, and polyglutamine-containing proteins, indicating a general improvement in protein quality control. This rare but strong suppressor requires modulation of only two genes, suggesting that it provides a major protein quality control program with a readily accessible switch. Furthermore, the synergistic cooperation of two genes, ufd-2 and spr-5, points to a common downstream pathway with integrative regulation.
Consistent with the observation that the loss of function of C. elegans ufd-2 and spr-5 promotes the clearance of misfolded proteins, inactivation of their Drosophila and mammalian orthologs reduces the toxicity of aggregation-prone proteins, indicating the existence of a protein quality control regulatory mechanism that is functionally conserved across species. Interestingly, both mammalian genes encode lysine-modifying enzymes: UBE4B is a U-box type ubiquitin ligase, and LSD1 is a lysine-specific protein demethylase. Both UBE4B and LSD1 are highly expressed in neurons and essential for early development in mammals [47–50].
In contrast to the conventional notion that ubiquitin ligase promotes protein degradation, our studies indicate that UBE4B negatively affects the clearance of misfolded proteins, and its down-regulation protects against severe proteotoxicity. In line with our observation that the down-regulation of UBE4B protects against proteotoxicity in the nervous systems of C. elegans and Drosophila, mice with elevated levels of UBE4B show autophagy defects with accumulation of ubiquitin- and p62-positive aggregates in the brain [51]. UBE4B forms a complex with an AAA-ATPase p97/VCP to ubiquitinate and degrade specific client proteins [47,52]. p97/VCP plays an essential role in handling unfolded proteins, such as in endoplasmic-reticulum-associated protein degradation [53], and it was recently linked to familial ALS [54]. Our findings thus provide a new link between p97/VCP and protein quality control, which is regulated by UBE4B.
The fact that both UBE4B and LSD1 are enzymes catalyzing post-translational modifications suggests that their effects on protein quality control can be timely, energy-efficient, and integrative. The synergistic interaction between the two lysine-modifying enzymes, UBE4B and LSD1, also suggests that their downstream pathways converge to influence protein quality control. Consistent with recent studies showing enhancement of protein quality control [20,55–57], the identification of the strong antiproteotoxic effects mediated by UBE4B and LSD1 demonstrates that plasticity of the cellular protein quality control programs can be substantially augmented to yield overall protection to an organism.
Unbiased transcriptome analysis points to p53 as a central regulator of the transcriptional reprograming that mediates the effects of UBE4B and LSD1 on protein quality control. Consistent with this observation, p53 has been found to have a number of direct transcriptional targets functioning in protein quality control and neuroprotection, and it also activates additional stress-response transcription factors such as FOXOs [33]. Interestingly, p53 is elevated in the central nervous system of patients with neurodegenerative conditions such as Alzheimer disease and ALS [58,59]. Our observation that the transcription factors mediate the effects of this strong suppressor is reminiscent of other signaling pathways governing protein homeostasis. For example, the heat shock response activates the expression of molecular chaperones and other protein quality control machinery via the master transcription regulators, the heat shock factors [60]. Also, the unfolded protein response promotes the endoplasmic reticulum (ER) quality control programs through the activation of a set of the transcription factors, including XBP1, ATF4, and ATF6 [61]. In recurring themes, the post-translational regulation by UBE4B and LSD1 activates the p53 transcription factor, which is then capable of eliciting a systematic protective program against proteotoxic stress.
p53 has a well-established role in regulating responses to DNA damage [62,63], and recently, a neuroprotective role of an activated DNA damage checkpoint has been demonstrated in a tau-dependent neurodegeneration model [64]. Here we propose that p53 is a versatile transcriptional switch that guards against both genotoxicity and proteotoxicity. The specific activity of p53 may be fine-tuned at the post-translational level by upstream regulators such as UBE4B and LSD1. In addition, it is known that p53 promotes apoptosis in cells with irreversible genotoxic damage [65]. p53 may also function as a dual regulator in proteotoxicity: it promotes the repair and survival of moderately damaged cells but turns on cell death pathways in cells whose damage is irreparable. Such duality has been observed for other protein quality control systems, such as the ER stress responses [61]. Thus, p53 could serve as a critical regulator of cellular responses to proteotoxicity by repairing or removing damaged cells.
Taken together, these findings reveal a previously unrecognized pathway that systematically antagonizes the proteotoxicity associated with neurodegenerative diseases, and they point to potential targets for harnessing the protective capacity of the cells’ reprogrammed protein quality control to develop a wide-spectrum antiproteotoxicity therapeutic strategy.
The pregnant rat dams were euthanized by overdosing with nembutal. The Children’s Hospital of Philadelphia IACUC approved these procedures (protocol #597).
For mammalian expression, SOD1 and TDP-43 were expressed in pEF-BOS and pRK5-Myc, respectively, as previously described [9,20]. The UBE4B (TF308519) and LSD1 shRNA (TF316984) plasmids and the scrambled control (TR30015) were from Origene. The p53 shRNA plasmid pLVTH-sip53 and control pLVTH were from D. Trono (Addgene #12239) [66]. The p53 transcriptional reporter PG13-Luc was a generous gift from B. Vogelstein [67]. The autophagy luciferase release plasmids Act-LC3-Gluc and Act-Gluc were kindly provided by B. Seed [37], and the control pCMV-SEAP was from A. Cochrane (Addgene #24595). For transgenic C. elegans, ufd-2 and spr-5 complementary DNAs (cDNAs) were cloned into a vector under the control of an snb-1 promoter, as previously described [20]. Additional information on the shRNA targeting sequences and vectors is given in the Supporting Information Materials and Methods (S1 Text).
The Bristol N2 C. elegans strain was used in all experiments unless otherwise specified. A list of C. elegans strains is given in the Supporting Information Materials and Methods (S1 Text). Transgenic lines were generated according to standard procedures by injecting 20 μg/ml of expression plasmid DNA into hermaphrodite gonads. For the suppressor screen, worms were mutagenized with 47 mM ethyl methanesulfonate, and a semiclonal strategy was used with five P0 worms in one plate. Suppressors were visually selected based on strong recovery in the movement phenotype in the F2 generation. The suppressor mutations were mapped by using single-nucleotide polymorphism markers in the Hawaiian strain and then identified by whole-genome deep sequencing, followed by Sanger sequencing validations (see Supporting Information Materials and Methods [S1 Text]).
The C. elegans strains were observed stereoscopically, and their motility was quantified by the thrashing assay [20]. Animals were transferred from the feeding plate into M9 buffer (3 mg/ml KH2PO4, 6 mg/ml Na2HPO4, 5 mg/ml NaCl and 1 mM MgSO4). After 1 min of adaptation, the number of body bends or thrashes was counted for 1 min as an index of the locomotor phenotype. A thrash was counted when both the head and the tail bent away from the anteroposterior axis by more than 45°. Videos of C. elegans locomotion were recorded using a Leica M165 fluorescence stereoscope. High-magnification imaging was performed on a Zeiss AxioObserver Z1 with Apotome, with C. elegans immobilized by 10 mM levamisole.
See Supporting Information Materials and Methods (S1 Text).
See Supporting Information Materials and Methods (S1 Text).
The protein solubility assay to measure aggregate proteins in C. elegans and mammalian cells was modified from a previously described protocol [9] (see Supporting Information Materials and Methods [S1 Text]).
After a 72-h knockdown, cells were detached and transfected with firefly luciferase p53 reporter plasmid (PG13-luc), together with a thymidine kinase promoter Renilla luciferase (tk-Rluc) reporter for normalization. Cells were lysed in passive lysis buffer (Promega) 24 h after transfection and analyzed with the Dual Luciferase Reporter System according to the manufacturer’s recommendations (Promega) using an injector-equipped Synergy H1 microplate reader (Bio-Tek).
Proteasome assays were performed as described previously [68], using the Suc-LLVY-Luciferin substrate for chymotrypsin-like activity of the proteasome (the Proteasome Glo kit, Promega). In brief, cells were detached and washed in DMEM/10, followed by several washes in cold PBS. Proteasome lysis buffer (50 mM Tris-HCl, pH 7.5, 0.025% digitonin, 250 mM sucrose, 5 mM MgCl2, 0.5 mM EDTA, 2 mM ATP, and 1 mM DTT) was added to the cells and incubated on ice for 5–10 min. The lysates were then centrifuged for 15 min at 20,000 g to isolate the cytoplasm containing the proteasomes. The supernatant was transferred to a fresh tube, and equal amounts of protein were used in each assay.
Autophagy was quantified with a Gaussia luciferase release assay [37,38], which is based on the ATG4B-induced proteolytic cleavage of an actin-anchored fusion LC3-Gluc fusion protein (S5D Fig. and S5E Fig.). ATG4B-induced proteolytic cleavage of LC3 releases the Gluc fragment and enables its secretion into the cell medium. The activity of the released Gluc in the medium (together with constitutively secreted SEAP) was measured by the Secrete-Pair Dual Luminescence Assay kit (GeneCopoeia). Cells, plated in 12-well dishes, were transfected with the Act-LC3-Gluc or control Act-Gluc plasmid together with the normalization control, CMV-SEAP. At 24 h after transfection, the DMEM/10 medium was replaced, and 100 μl of cell growth medium was withdrawn at 24 h, 48 h, and 72 h. The medium was centrifuged at 6,000 g for 5 min to remove detached cells, followed by the luciferase analysis according to the manufacturer’s recommendations (GeneCopoeia) using a microplate reader (Synergy H1, Bio-Tek).
For LC3 western blot analysis, cells were lysed in LC3 buffer (50 mM Tris-Cl, pH 8.0, with 1% SDS, 0.5% NP40, 150 mM NaCl, and 5 mM EDTA) and sonicated with a Diagenode Bioruptor device (set on high, 30-sec pulse, 30-sec pause, 7.5 min total).
Total RNA was isolated from HEK293T cells with the RNeasy Mini kit and analyzed using the Affymetrix human GENE 1.0ST array. The microarray data were managed using the Partek Genomic Suite (Partek, St. Louis) and Spotfire DecisionSite software (TIBCO Software, Palo Alto, California) and analyzed using Ingenuity Pathways Analysis software (IPA, Ingenuity Systems). In addition to RNA, total protein was isolated from the same samples by acetone precipitation and resolubilizing the flow-through lysates to verify the reduction of the UBE4B and LSD1 proteins. For quantitative RT-qPCR validations, cDNAs were synthesized with the QuantiTect reverse transcription kit (Qiagen). Primers for quantitative RT-qPCR were from PrimerBank (S3 Table) [69]. RT-qPCRs were performed on a BioRad thermal cycler with iQ SYBER Green PCR mix (BioRad).
Embryonic Sprague Dawley rat spinal cord cultures and neuronal survival assays were previously described [46] (see Supporting Information Materials and Methods [S1 Text]).
The p-values for all analyses were obtained using Student’s t tests performed in Excel or GraphPad Prism 6, unless otherwise indicated. For the microarray data, Student’s t test was used to analyze the gene expressions. For the Upstream Regulator Ingenuity Pathway Analysis, Fisher’s exact test was used. For the LC3-II western blotting analysis and the spinal cord motor neuron toxicity assay, a one-way ANOVA with multiple comparison test was used.
|
10.1371/journal.pcbi.1000399 | The Effect of Ongoing Exposure Dynamics in Dose Response
Relationships | Characterizing infectivity as a function of pathogen dose is integral to
microbial risk assessment. Dose-response experiments usually administer doses to
subjects at one time. Phenomenological models of the resulting data, such as the
exponential and the Beta-Poisson models, ignore dose timing and assume
independent risks from each pathogen. Real world exposure to pathogens, however,
is a sequence of discrete events where concurrent or prior pathogen arrival
affects the capacity of immune effectors to engage and kill newly arriving
pathogens. We model immune effector and pathogen interactions during the period
before infection becomes established in order to capture the dynamics generating
dose timing effects. Model analysis reveals an inverse relationship between the
time over which exposures accumulate and the risk of infection. Data from one
time dose experiments will thus overestimate per pathogen infection risks of
real world exposures. For instance, fitting our model to one time dosing data
reveals a risk of 0.66 from 313 Cryptosporidium parvum
pathogens. When the temporal exposure window is increased 100-fold using the
same parameters fitted by our model to the one time dose data, the risk of
infection is reduced to 0.09. Confirmation of this risk prediction requires data
from experiments administering doses with different timings. Our model
demonstrates that dose timing could markedly alter the risks generated by
airborne versus fomite transmitted pathogens.
| We model the relationship between the temporal patterns of pathogen exposure and
infection take off within people. Since different routes of transmission (e.g.,
airborne versus surface transfer routes) may result in different temporal
patterns of exposure, this model helps to better compare the risks of
transmission from one person to another through these different routes. Previous
models assumed that the risk of infection is the same whether pathogens are
inoculated all at once or over one day. Our model, in contrast, captures how one
pathogen affects the potential of immunity to keep concurrently or subsequently
arriving particles from initiating an infection. Since the pattern of timing of
airborne and surface spread pathogen arrivals differ, our model shows that each
airborne pathogen could carry less risk than each surface transmitted pathogen.
Unfortunately, data to fully fit our model are not currently available.
Therefore new experiments will have to be conducted where doses are given across
different temporal windows.
| Microbial risk assessment models are valuable tools for estimating the risks
associated with exposures to pathogens in the environment pathogens [1]. Central
to this estimate is a dose-response model that predicts the probability of infection
given a dose exposure magnitude. In current microbial risk assessment models dose
accumulates over time and the probability of infection is based on the total
accumulated dose over that period of time [2]–[4]. This
assumes that each pathogen particle carries a risk of infection that is independent
of when other pathogens have arrived to a host; i.e., three exposures to dose X
generate the same total risk as one exposure to a 3× dose. We put forth an
alternative dose response model that assumes the current capacity of immune
effectors to control an arriving pathogen should be affected by 1) how many
effectors are occupied fighting previously or simultaneously arriving pathogens, 2)
how many effectors have been depleted in fighting previously arriving pathogens, and
3) how many effector reinforcements have arrived due to usual effector turnover
rates or due to a stimulus from prior pathogen exposure.
If dose-timing effects arise from such immune effector dynamics, then infection-risk
calculations that do not take these dose-timing effects into account could lead to
errors. For example, errors could arise in models of influenza transmission as
follows. Pathogens arriving to a host via aerosols do so more frequently but at
lower doses than pathogens arriving via hand or fomite mediated inoculations. Models
of influenza transmission that do not account for dose-timing effects, such as the
model by Atkinson and Wien [4], might misdirect influenza control resources to
masks from hand hygiene. Models that assume independent single dose effects will
require more extreme cleaning to reduce risks to acceptable levels than models
capturing immune effects on dose timing.
Evaluating the potential importance of such dose-timing effects is difficult for two
reasons. First, immune control of pathogens is complex; not enough detailed
knowledge regarding that complexity is available to provide a high degree of
confidence in a-priori causal model predictions. Second, there is almost no direct
observational data documenting the presence or absence of dose-timing effects.
Although various studies have given pathogen exposure doses over time [5]–[10], only Brachman et al.
[11], has been conducted in a manner that allows one to
calculate risks for comparable doses administered over different temporal windows.
In this paper we have taken an approach intended to stimulate science that will
address both of these issues. We develop a simple model that illustrates the need to
generate new data that can describe dose-timing effects while at the same time
providing a base upon which to build more realistic models that incorporate more
data and theory on immunity. Our model addresses immune control of pathogens between
the time pathogens arrive at a host and the time they are either eliminated or have
multiplied enough so that an acquired immune response will be needed for control.
We make our model general enough to capture dynamics of pathogen control that might
arise from established antibodies and T-cells, macrophages, polymorphonuclear
leukocytes, plasma cells, dendritic cells, complement cascades, chemokines,
interleukins, interferons, toll like receptors, and other diverse elements affecting
immunity. But we lump all these mediators of pathogen control into a highly abstract
entity we label as immune effectors. We assume that the dynamic effects of limited
immune effector numbers are similar whether the limitation arises from immune
effectors being occupied with previously arrived pathogens or from prior consumption
of immune effectors in their process of killing pathogens. Therefore we only model
the latter source of immune effector limitations. The resulting model is one where
any single pathogen always has some chance of initiating an infection but the risk
of infection associated with each additional pathogen exposure can markedly increase
at higher pathogen doses given over short temporal windows. The exact dynamics of
our model will vary as realistic details are added to it. Our goal here is simply to
illustrate the importance and inevitability of immune mediated dose-timing effects
so as to stimulate further empirical and theoretical work.
The structure of the paper is as follows: in the methods section we describe the
Cumulative Dose model and analyze its dynamics. In the results section we use the
Cumulative Dose model to fit experimental data assuming a fixed temporal exposure
window to simulate the archetypical single dose experiment of dose-response trials.
Using the estimated model we show the effect of changing the length of the temporal
exposure window. Finally, the conclusions and future research are presented in the
discussion section.
The model is based on a stochastic population of individual pathogens and immune
effectors. Since the focus of our analysis is how small populations of pathogens
either die out or lead to infection initiation, we cannot rely on the mean-field
solution provided by the deterministic framework [12]–[14].
The state of the system is defined by the pair () representing the number of immune effectors and the number of
pathogens, in any single host, respectively. The system is defined by the
following set of state transitions:(1)(2)(3)(4)
The number of immune effectors can increase at: 1) a rate , which models the constant arrival of immune effectors
regardless of the current state of the immunological system; and 2) a rate , which models the recruitment of immune effectors in the
presence of pathogens. This term is intended to reflect cytokine induced
recruitment of remote immune effectors to a pathogen invasion site and not
acquired immunity. We assume that the relative endpoints of infection takeoff or
pathogen elimination are reached before an acquired immune effect comes into
play. Immune effectors decrease either at a natural death rate , or at a mass-action deactivation rate due to the encounter
with pathogens .
The number of pathogens can increase by reproduction at a rate or by arrival during the inoculation period at a rate . Here represents the net reproduction rate that aggregates birth and
death rates. Pathogen numbers decrease due to interaction with immune effectors
as a mass-action deactivation process at the rate .
The initial state of the system is set to . No chronic low-level exposures or remaining pathogens from
prior exposures are considered. The system starts from the clean state: no
pathogens and the stationary number of immune effectors in the absence of
pathogens. The inoculation process is characterized by the dose of exposure and the temporal exposure length ; i.e., the dose that is composed by pathogens is inoculated into the host during a period of time units. Therefore, the arrival of external pathogens is
modeled as the rate during the inoculation period. Once inoculation has finished
the pathogen arrival rate becomes zero. Thus, the rate depends on time and is defined as
During , the pathogens, arrive over a continuous time in the presence of the
immunological response to those pathogens. Once the inoculation has finished,
only the immunological response remains. We set the unit of time to an hour.
That keeps us in the range where we think exposure fluctuations are making a
difference and out of the range where adaptive immune system feedbacks come into
play.
Due to stochastic effects and the fate of a relatively small population of
pathogens and immune effectors, the same inoculation dose administered in the same time frame does not necessarily have the same outcome. Each replication
(i.e. run) of the model corresponds to a dose trial on a new subject. All the
numerical results are the average of 104 runs of the Cumulative Dose
model implemented with the Gillespie algorithm [15] using
C. The criteria to stop the simulation is either extinction of
pathogens after the inoculation period () or pathogens diverging to a very large number, , corresponding to no infection and infection respectively. The
probability of infection for a pair is the proportion of simulations that diverge to a large
number as opposed to equilibrating to the state of no pathogens.
Figure 1 illustrates the
stochastic process effects on pathogen dynamics given a fixed time of exposure
for different inoculation doses. The main plot in this figure is the time course
of the number of pathogens for 100 independent dose trials given a dose of 60
pathogens administered over one unit of time. The number of pathogens steadily
grow during the inoculation period, from 0 to 1, since the rate of arrival of
pathogens () is much faster than immunological killing of pathogens. Once
the entire dose has been inoculated at
time = 1, the external arrival of pathogens
stop () and the immunological response dominates the rest of the
dynamics. In this particular case, the population of pathogens becomes extinct
in 33 cases out of 100, thus, the probability of infection given a dose of 60
pathogens over 1 unit of time is 0.67. Analogously, for a dose of 25 the
probability of infection is 0.02 and for a dose of 90 the probability of
infection is 0.98 (insets of Figure
1).
Figure 1 illustrates how the
Cumulative Dose model yields higher probability of infection when the inoculated
dose is increased. The length of time over which the dose is administered, , also plays a crucial role in the probability of infection. At
one extreme where all the pathogens were inoculated at once (), the immune system has no time to react, and the initial
state of the system is . From this initial state, the immunological response dynamics
determines the fate of the pathogens: either extinction or unbounded growth of
pathogens diverging towards infinity.
For , however, the initial state after all pathogens have been
inoculated () is not the expected , but rather a distribution of probabilities over the space of
possible states. Figure 2
shows the stochastically determined distribution of system states at the point
in Figure 1 where the
exposure time has just ended. It illustrates the effect of different temporal
exposure lengths, ranging from (six minutes) to
Te = 50 hours.
Panel B shows this point of time for the settings in Figure 1. The longer the exposure length, the
larger will be the variance in the distribution of probabilities. Furthermore, a
longer exposure length also affects the average state after inoculation. Both
the pathogen levels and the immune effector levels decrease from the
instantaneous inoculation values as the exposure window length increases. But
the balance between these increasingly favors the immune effectors. Longer
temporal exposure lengths dilute the arrival rate of external pathogens, . Consequently the immunological response has more time to
neutralize the existing pathogens before the arrival of new pathogens. On the
other hand, as the temporal exposure lengths decrease, an increased number of
immune effectors are consumed in killing pathogens, leading to a higher
probability of unbounded growth of pathogens, and thus infection.
For and the average state after inoculation is very close to the ideal
instantaneous inoculation, . To better understand the dynamics once inoculation is over,
we included the numerically calculated separatrix as if the system were
deterministic (red-dashed line in Figure 2). Although this separatrix is only truly valid for the
analogous deterministic model, it indicates the probable fate of different
initial states. For the deterministic system, the separatrix separates those
states that go to infection from those that do not (see subsection on
Deterministic Analysis). As temporal exposure length increases, the distribution
of probabilities gravitates towards the space of states that go to no-infection
(below the separatrix).
Further understanding of the stochastic dynamics of the Cumulative Dose model can
come from a deterministic description of the system that assumes a continuous
large number of immune effectors and pathogens. We focus our analysis on the
dynamics after the inoculation period, so is set to 0 and removed from the equations. This analysis on
the deterministic version helps illustrate the interactions between pathogens
and immune effectors that result either in infection or extinction of pathogens.
The stochastic system is fully described by a multivariate master equation [16],
which can be expanded in a deterministic formulation known as
macroscopic law. The deterministic version of the
cumulative dose model is as follows,(5)(6)where and are continuous variables of the population of pathogens and
immune effectors respectively. The fixed points of the deterministic version of
the cumulative dose model are where the pathogen has been eliminated and immune effectors are
in equilibrium and where the forces of pathogen growth are balanced by immune
dynamics affecting pathogen death. Note that in the stochastic analyses of this
model as in Figure 1, this
point is never reached. Instead simulations are terminated when growth takes off
toward this point. A simple analysis of the stability of the fixed points
reveals the space of parameters in which the solution is well-defined.
The point is the equilibrium of no infection—the equilibrium
of the system in the absence of pathogens. When the system gravitates towards the immunological system prevents pathogens from growing,
resulting in pathogen extinction and therefore no infection.
To evaluate the stability of the fixed point, we formulate the Jacobian matrix of
the system of equations on .(7)
For a stable equilibrium, both Eigenvalues of the Jacobian matrix need to be
negative, or equivalently, the matrix must have a negative trace and a positive
determinant. For the trace of the Jacobian to be negative the condition must be true. Since the positive determinant condition, , is more restrictive it subsumes the condition for a negative
trace.
The second fixed point is only well-defined when both and are positive, since negative number of pathogens and immune
effectors are impossible. The number of pathogens is only positive when . Given the condition of a positive determinant, , the sign can only be negative, consequently . Therefore, the system is well defined — i.e. has a
stable equilibrium at no infection and with both fixed points in the positive
quadrant — only when the following condition 8 is met(8)
Once we determine the stability of we need to characterize the second fixed point . After some basic algebra, the determinant of the Jacobian
matrix for can be expressed as follows: . Given condition 8, both terms are positive, which makes the
determinant negative. As a result the Eigenvalues of the Jacobian are real with
different signs. Therefore, is a saddle point as shown in Figure 3.
The vector field in Figure 3
illustrates the dynamics of the cumulative dose after the inoculation period.
The probability of being in a given state after inoculation is shown in Figure 2. If the system were
deterministic then we could anticipate the probability of infection by summing
the probability of those states below the separatrix. This does not hold for the
stochastic Cumulative Dose model. Nonetheless, the deterministic vector field,
shown in Figure 3, serves as
an approximate description of what happens in the stochastic model.
For instance, let us take the probability distribution of states when centered at , i.e., and . The typical dynamic results in the decrease in number of
pathogens and immune effectors, gravitating towards the saddle point , from which it will bifurcate to the stable point of
no-infection , or an unbounded growth of pathogens. In the case of and , most of the states are already very low in pathogens, and
consequently the number of immune effectors will eradicate the few pathogens
still existing and go to the stable equilibrium of no infection. However, there
is a non-zero probability, albeit small, of being in a state with a large number
of pathogens and a small number of immune effectors. In this case, stochastic
perturbations aside, the pathogens will keep multiplying producing infection in
the host.
In this section, we fit empirical data on multiple pathogens for the single event
inoculation scenario. Next, we extend our analysis to incorporate different
temporal exposure windows and patterns of inoculation.
The first empirical dataset to which we apply the Cumulative Dose model is
Poliovirus type 1 [17]. The cohort for this experiment was 32
2-month-old infants. Inoculation was oral. Figure 4 and Table 1 show the fit alongside a fit to the
Exponential model () according to [18].
The cohort for the Cryptosporidium parvum study [18] was
35 healthy subjects (12 men and 17 women, age range between 20 and 45 years).
The strain was an isolate from a calf and the inoculums were orally administered
via capsules. Figure 5 and
Table 2 show the fit
alongside a fit to the Exponential model () according to [20].
Finally, we tested the Cumulative Dose model against a dataset for Rotavirus
[19].
The cohort for rotavirus was 62 adult males, 18 to 45 years old. The inoculation
was oral. Unlike the previous dose-response empirical datasets, neither the
Cumulative Dose model nor the Exponential model produce a good fit. The
Beta-Poisson model () was statistically a better fit than the Exponential model
[20].
Both the Exponential and the Cumulative Dose model increase too rapidly in
relation to the probability of infection of 1; i.e. these models cannot maintain
a non-zero or non-one probability of infection for a dose range of several
orders of magnitude. Conversely, the Beta-Poisson model does not suffer from
this limitation since its convergence to 1 is slower, providing a wider range of
variance (Figure 6 and Table 3).
A possible explanation of the poor fit of the Cumulative Dose model is the high
degree of acquired immunity to Rotavirus and the changing serotype profile
circulating within populations [23]. Unlike the polio virus study, the rotavirus
cohort consisting of adults (18–45 years old), is likely to have been
exposed multiple times to various rotavirus serotypes [24]. Such
heterogeneity in susceptibility flattens out dose response curves beyond what
can be captured by exponential dose response models or this Cumulative Dose
response model.
In the previous subsections we fixed temporal exposure length, , to 1 hour, and assume that this is the time corresponding to
the single shot inoculation, analogous to existing experimental dose-response
trials. In this section, we present simulations for a range of different
temporal exposure lengths, illustrating how longer times affect the dose
response curve. The model is set to the parameters that provided an optimal fit
for a temporal exposure length of .
Figure 7 shows the
dose-response curves for Poliovirus type 1 for different lengths of exposure for
the estimated parameters used in Figure 4 to fit the experimental data for the condition
Te = 1.0: . As the exposure length increases, the probability of
infection decreases dramatically. Therefore, assuming that the unit of time is
one hour, and this is the equivalent for a dose that is administered in a single
shot, the probability of infection generated by the Cumulative Dose model for a
dose of of 90 pathogens administered in one hour is 0.82. If the dose
were administered not in one hour, but uniformly over ten hours the probability
of infection would be 0.18. If the dose were administered over fifty hours the
probability of infection would be reduced to 0.0001. To obtain the same
probability of infection for a ten hours inoculation period instead of one, we
would require a dose of 139 pathogens instead of 90.
Because data on the impact of temporal patterns of inoculation are currently not
available, a model with dose-time dependence such as ours is not identifiable
[25]; i.e., the model can be fit to existing single
dose empirical data with many different parameters sets. For example, in Figure 8 we show model
simulation results for Cryptosporidium parvum for two different
parameter sets. Both parameters sets have a similar fit to the
Cryptosporidium parvum dataset when (mean square error using and is 3.5×10−3 and
9.7×10−3 respectively). For values of , however, the dose response relationships of the two parameter
sets diverge. Parameter set is much less sensitive to exposure time than due its slower dynamics. Using parameter set R, pathogens
proliferate faster, are being eliminated by each immune effector more quickly,
are recruiting fewer immune effectors, and are eliminating immune effectors at a
slower rate. On the other hand, using parameter set R, the natural rate of
turnover of immune effectors is more rapid. We cannot argue at this point which
is the most plausible configuration since identifiability cannot be resolved
without data from dosing trials for different exposure lengths.
In this section we relax the assumption that pathogens are inoculated at a fixed
rate. We allow variation both in dose magnitude and length of exposure time, in
order to capture a more realistic exposure scenario.
The temporal pattern of inoculation of pathogens within a host depends both on
the behavior of the host and the contamination of the environment the host
interacts with. For instance, a susceptible host in a venue contaminated with
influenza will be exposed to pathogens from air and fomites. However, the
temporal patterns of exposure for these two modes of transmission are different.
The host is likely to receive a small dose with every breath when breathing
contaminated air. In fomite mediated transmission, however, the touching of a
mucous membrane with contaminated fingers, for example, is likely to transmit a
larger but less frequent dose.
To illustrate this effect we devised an experiment where both the total
inoculated dose and the exposure time length are fixed. The only parameter that varies is the number of
inoculation events, , which ranges from 1 to the total dose . Consequently, once the number of inoculations events is
determined, the dose inoculated in each event is and the rate at which inoculation occur is .
Figure 9 shows the results of
this experiment where the same parameter sets are used as in Figure 8. The pathogen is
Cryptosporidium parvum, and the same two different
parameters sets, S and R, are used to inform
the cumulative dose model. The total dose inoculated is set to and the temporal exposure length is set to
Te = 120.0
hours.
For both parameter sets S and R we observe the same behavior: infectivity
decreases as the frequency or number of inoculations events increases. The
temporal pattern more likely to be associated with fomite transmission (low
frequency and high dose, Figure
9.B) is more likely to produce infection than the patterns associated
with airborne transmission (high-frequency and low dose, Figure 9.C) .
For parameter set R, the probability of infection if the dose is
inoculated with a single exposure (Figure 9.A) is 0.752. The same dose inoculated over 4 events, where
each event is one fourth of the total dose (Figure 9.B), reduces the probability of
infection to 0.443. In addition, if the dose is inoculated over 50 events (Figure 9.C) the probability
decreases to 0.111. For parameter set S, the reduction of the
infection probability is less pronounced: 0.740, 0.676 and 0.601 for 1, 4 and 50
inoculation events respectively.
In previous sections we showed that longer temporal exposure lengths decrease
infectivity due to the action of the immune system. In this section, we show
that not only the duration of the exposure matters, but also the way in which
pathogens arrive within that interval can decrease infectivity. These results
suggest that risk assessments based on current dose-response data might be
over-estimating risk of infection. An important corollary is that risk of
infection for a given exposure dose may depend on the route of transmission
based on their differences in the pattern of exposure.
We examined a dynamic mechanistic model where immune system effects generated dose
response dependence on the timing of doses. The specific aspects of our model that
generate these dose-timing effects are: 1) decreases in available immune effectors
because they are being eliminated as they kill pathogens; and 2) increases in
available immune effectors due to both pathogen dependent and independent
recruitment. An additional mechanism resulting in decreases in available immune
effectors that is not included in our model could be the time of immune effector
engagement with pathogens in the killing process. The dose-timing effects we
illustrate would be absent in a model where some effector like a T-cell
instantaneously kills pathogens or pathogen generating cells, where no killing
capacity is lost with each kill, and where effector dynamics are not otherwise
altered by encounters with pathogens. Any such model, however, is highly
unrealistic, and therefore we conclude that the dose-timing effects presented in our
model could be important and warrant further study.
Dose-timing effects have implications for microbial risk assessment, for infection
transmission system modeling, and for the evolution of emerging pathogens.
Considering a microbial risk assessment example, the implications of our findings
suggest that exposure routes with different dose-timing dynamics could have
different risks and therefore result in different clean up protocols for
contamination events such as a norovirus outbreak or a Katrina-like disaster. Dose
timing could, therefore, affect decisions on which venues to close or what the total
dose that workers would be permitted to accrue during a cleanup operation.
Considering modeling infection transmission, the standard approach is to define a
contact and a transmission probability per contact while the physical route of
transmission is ignored. Modeling the physical route of transmission is important
when it is necessary to specify how much transmission is taking place in particular
public venues and when specifying which control actions in these venues will reduce
transmission. When different routes have different temporal exposure patterns, we
demonstrate here that there is considerable potential for immune system effects to
alter the ratio by which airborne transmitted and hand-fomite transmitted pathogens
generate new infections. If we had data on infection risks under different
dose-timing patterns, we could say more precisely how much difference in risk there
might be from an airborne and a hand-fomite mediated pathogen. Unfortunately such
data is lacking.
The evolution of emerging infection implications derive from the route of
transmission effects just discussed. When pathogens first jump species, they are
likely to encounter strong innate immune responses to which they must evolve some
escape strategy. That means very high transmission doses will be required to sustain
transmission and that low dose exposure over longer times such as occurs with
airborne transmission will be the most unlikely to be effective in transmitting
infection. But, as escape from innate immune responses evolves, the balance could
begin to favor airborne transmission which might be more effective in disseminating
infection.
We do not have enough dose timing data for any infection to evaluate either the
microbial risk assessment implications, the infection transmission system
implications, or the emerging infection evolution implications. Any data providing
indications of the magnitude of dose-timing effects generated by any type of
immunity to any agent would provide an important first step that would at least
indicate what range of effects might be expected. Animal studies could compare the
risks associated with a single instantaneously delivered dose with the same dose
magnitude delivered over extended periods of time. Measurements of specific immune
effector dynamics, such as interferon gamma [26] would improve our
mechanistic understanding of a cumulative dose effect and indicate how to refine our
models for different animal/pathogen systems.
The issue of dose-response trial design is crucial for advancing both quantitative
microbial risk assessment and analysis of population infection transmission systems.
Due to the absence of a prior theoretical framework, there has been no motivation to
conduct dosing trials that take multiple doses and multiple dosing times into
account. Now that the potential effects of dose timing have been demonstrated and
the practical significance of such measurements for microbial risk assessment and
transmission system analyses is more evident, we hope to see such experiments.
|
10.1371/journal.pcbi.1006989 | Sequential exploration in the Iowa gambling task: Validation of a new computational model in a large dataset of young and old healthy participants | The Iowa Gambling Task (IGT) is one of the most common paradigms used to assess decision-making and executive functioning in neurological and psychiatric disorders. Several reinforcement-learning (RL) models were recently proposed to refine the qualitative and quantitative inferences that can be made about these processes based on IGT data. Yet, these models do not account for the complex exploratory patterns which characterize participants’ behavior in the task. Using a dataset of more than 500 subjects, we demonstrate the existence of sequential exploration in the IGT and we describe a new computational architecture disentangling exploitation, random exploration and sequential exploration in this large population of participants. The new Value plus Sequential Exploration (VSE) architecture provided a better fit than previous models. Parameter recovery, model recovery and simulation analyses confirmed the superiority of the VSE scheme. Furthermore, using the VSE model, we confirmed the existence of a significant reduction in directed exploration across lifespan in the IGT, as previously reported with other paradigms. Finally, we provide a user-friendly toolbox enabling researchers to easily and flexibly fit computational models on the IGT data, hence promoting reanalysis of the numerous datasets acquired in various populations of patients and contributing to the development of computational psychiatry.
| The ability to perform decisions and learn from their outcomes is a fundamental function of the central nervous system. In order to maintain their homeostasis and maximize their biological fitness, organisms must maximize rewards and minimize punishments. Yet, pure exploitation often leads to suboptimal solutions. In order to discover the best course of action, organisms must also explore their environment, especially when this environment is complex or volatile. Here, we dissected exploratory strategies in one of the most classic decision-making paradigms of cognitive neuroscience. First, we found that humans tend to sample sequentially the space of possible actions. Second, we developed a new mathematical model better able to predict trial-by-trial choices, by articulating this sequential exploration mechanism with random exploration and exploitation. Third, we showed that sequential exploration reduces across lifespan, a result which might be explained by specific neuroanatomical or neurochemical changes associated with normal aging. Together, these findings may contribute to a better understanding of exploratory behaviors and a better assessment of their disruption in a wide range of neuropsychiatric conditions.
| Many neuropsychiatric disorders are associated with alterations of learning and decision-making. Standardized cognitive paradigms are thus increasingly used to improve diagnosis and evaluate the response to treatments. Developed 25 years ago [1], the Iowa Gambling Task (IGT) remains one of the most popular tools used for this purpose in clinical settings (Fig 1A). Over the years, it has been applied more or less successfully to many populations such as patients suffering from brain lesions, Parkinson disease, behavioral or substance addictions, mood disorders, personality disorders, etc. Although its reinforcement schedule confounding risk and punishment processing can be criticized, the IGT thus remains of considerable importance for the development of scalable methods in cognitive science and in the emerging field of computational psychiatry.
The classical analysis strategy for IGT data results in a crude estimate of decision-making deficits. Based on the relative preferences for “advantageous” decks (typically offering small gains but even smaller losses) over “disadvantageous” decks (typically offering big gains but even bigger losses), this approach does not leverage the full potential of the IGT. From a computational viewpoint, the IGT is indeed a highly complex task engaging value-based learning, risky decision-making, working memory and—as we shall see—different types of exploration. A series of computational models have been developed to better isolate these components, thereby offering clinicians and clinical neuroscientists more precise analytical tools to assess the cognitive profile of their patients. So far, computational neuroscientists interested in the IGT have mainly focused their efforts on the value-based learning and decision-making components of the task. Accordingly, the Expected Value (EV) or the Prospect Valence Learning (PVL, PVL-Delta) algorithms aim at capturing the non-linear and/or asymmetric decision weights associated with gains and losses [2,3]. More recently, the Value Plus Perseveration (VPP) model was developed to capture systematic perseveration or alternation tendencies across successive decisions [4]. However, the VPP model can be criticized for its high number of free parameters (8) relative to the number of trials (100) as well as for the uncertain cognitive validity of its perseveration module [5]. Finally, the last model proposed to date—termed Outcome Representation Learning (ORL)—also encompasses a perseveration module but tracks and weights separately (at the time of decision) the magnitude and the probability of positive and negative outcomes [6]. While the ORL was shown to perform better than alternative models, this latter feature is rather unlikely from the perspective of behavioral economics, as it implies that the decision-maker never combines reward magnitude and reward probability into an estimate of reward expectancy.
Here, we adopted another modeling strategy leveraging the existence of directed exploration (DE) in the class of multi-armed bandit tasks to which the IGT belongs [7–9]. Wilson and colleagues defined directed exploration as “a strategy in which choices are explicitly biased toward information”, as opposed to undirected (or random) exploration corresponding to a “strategy in which decision noise leads to exploration by chance” [10]. Thus, DE constitutes an “umbrella term” as it can refer to any regular choice pattern which: (i) maximize information about available options, (ii) cannot be readily explained by participants’ sensitivity to gains and losses. In the context of the IGT, a straightforward DE strategy is to allocate an “exploration bonus” to the behavioral options which have been sampled less often or less recently than others. This mechanism entails the sequential selection of all available option irrespective of their value or uncertainty, hence resulting into a specific choice pattern: namely, the tendency to select the four available decks over 4 consecutive trials (hereafter referred to as the Sequential Exploration (SE) index). Exploration bonuses and sequential exploration have a long history in the reinforcement learning literature, as they were already proposed in the seminal book of Sutton and Barto (Dyna-Q+ algorithm) [11]. They are also conceptually related to the optimistic initialization techniques used to ensure that all available options available to a decision-maker are sampled before settling on one of them [12].
Thus, we designed a compact computational architecture termed Value and Sequential Exploration in order to simultaneously capture exploitation, random exploration and sequential exploration in the IGT using 5 parameters. The core innovation of this new model is to articulate two types of choice strategies: a reward-seeking strategy shaped by reinforcement history and an information-seeking strategy shaped by choice history (Fig 1B). Governed by a value sensitivity and a decay parameter, the former module reacts to the gains and losses delivered during the task. As such, it resembles to the PVL model except that it includes no loss aversion parameter. The latter module relies on an exploration bonus specific to each participant—which can be either positive or negative depending on whether a given participant tends to explore or avoid options which have not been sampled recently—as well as on a learning rate—which determines how fast the exploration weight of a given deck goes back to the initial value of the bonus after having been sampled. Unlike other forms of directed exploration (such as uncertainty-dependent exploration), the exploration weights of the VSE model are totally independent of gains and losses. Different parameter combinations are thus able to reproduce the full range of possible SE indexes, from 0 to 100% frequency (see Methods for more details).
In order to demonstrate the superiority of the VSE model over the alternatives mentioned above (EV, PVL, PVL-delta, VPP, ORL), we reanalyzed a multi-study dataset of 504 participants who passed the 100 trials version of the IGT [13] (Fig 1A). State-of-the-art model comparison, simulation, as well as model and parameter recovery analyses were performed [14]. Second, in order to evaluate the cognitive validity to our model and illustrate its heuristic value, we focused on the data corresponding to the study of Wood and colleagues testing IQ-matched groups of old and young adults [15]. Indeed, it was recently shown that directed exploration diminishes across lifespan [16,17], so that the exploration bonus of older participants should be smaller than that of young participants. Third, we provide an open-source, user-friendly Matlab toolbox which has been developed to obtain the current results and which shall enable researchers who are not experts in computational models to re-analyze IGT data using both our new model and previous ones (https://github.com/romainligneul/igt-toolbox).
First, we evaluated whether sequential exploration (SE) occurred in the Iowa Gambling Task. To this end, we computed the “SE index” probing situations in which participants selected each of the four different decks over four successive trials using 25 independent consecutive quadruplets: e.g. 1–4, 5–8, etc. In the 504 subjects dataset, we observed such pattern 1400 times (11.1%) while only 1182 occurrences would be expected under random exploration (i.e. 9.38%, binomial test: p<10−10). Note that this test is highly conservative, as reward-maximization strategies bias choices towards the most valuable decks. Accordingly, a permutation approach in which trials were shuffled in time for each subject independently (total number of permutations: 5000) showed that the actual chance level was at 6.0%.
The target pattern was much more frequent in the first 20–30 trials of the task and it continuously declined as subjects formed more precise representation of each desk value and learned to exploit the reward structure of the task (Fig 1C). Interestingly, SE had a complex but strong relationship with decision-making performances in the IGT. A general linear model (GLM) analysis indicated that subjects with the highest overall performance had lesser SE indexes (linear effect: t(1,501) = -3.40, p<0.001), presumably due to the fact that these subjects needed less exploratory trials to figure out the reinforcement structure of the task. However, we also observed low SE indexes in the worst subjects, presumably due to maladaptive perseveration, which translated into a significant quadratic relationship between SE and performance (t(1,501) = 2.13, p = 0.034). Overall, the analysis of the SE index justified the development of a computational model capturing this important and previously overlooked exploration strategy in the IGT.
Like all previous models, the Value and Sequential Exploration (VSE) architecture updates “exploitation weights”, which keep track of the recent trends in gains and losses associated with each deck. However, the VSE model also updates on each trial the “exploration weights” associated with each deck. Depending solely upon the choice history, this exploration module was designed to capture the dynamics of sequential exploration observed in the IGT. On each trial, exploitation and exploration weights are simply summed into a composite value before being transformed by a conventional softmax step into choice probabilities. Hereafter, we describe the equations and the parameters which fully characterize VSE.
The exploitation module is directly inspired by the PVL model (Steingroever et al., 2013) although it includes no “loss aversion” parameter. A value sensitivity parameter controlled by θ (bound between 0 and 1) is instead applied separately to wins and losses.
v(t)=Gain(t)θ−Loss(t)θ
(1)
On each trial, the exploitation weight of each desk d is updated according to the following equations:
Exploitd(t+1)=Exploitd(t)*Δ+v(t)
(2.A)
Exploitd(t+1)=Exploitd(t)*Δ
(2.B)
Eq (2.A) controls the update of the deck chosen, by adding the feedback just experienced to the (decayed) value of this deck. Eq (2.B) controls the update of unchosen decks, whose exploitation weight progressively returns to 0 at a rate controlled by the decay parameter Δ (bound between 0 and 1). Note that a decay of 1 indicate that exploitation weights are integrated over all previous trials, while a decay parameter of 0 indicate that subjects’ decisions rely mostly on the most recent outcomes.
The main innovation provided by VSE consists in modeling sequential exploration in the IGT. Exploration weights reflect the attractiveness of each deck as a function of the number of trials for which the deck has not been selected. Exploration weights are agnostic regarding the monetary feedbacks experienced in the task. As such, they capture a pure information-seeking process, hence contrasting with Bayes-based uncertainty-minimization algorithms as well as the exploration modeled by the softmax or e-greedy rules [8]. Exploration weights are controlled by the following equations:
Explored(t+1)=0
(3.A)
Explored(t+1)=Explored(t)+α*(φ−Explored(t))
(3.B)
Eq (3.A) controls the update of exploration weights for the selected deck, which fall to zero as soon as the outcome of that deck is sampled. Eq (3.B) controls the update of unselected decks, which is governed by a simple delta-rule. The learning rate α (bound between 0 and 1) determines at which speed the exploration weights return to their initial value, defined by a free parameter termed “exploration bonus” or φ (unbounded). A positive exploration bonus implies that the agent is attracted by decks which have not been explored recently, whereas a negative exploration bonus implies that the agent tends to favor familiar decks. Therefore, the exploration bonus φ directly reflects the strength of sequential exploration, so that a more positive value will translate into a higher probability of reproducing the aforementioned pattern of 4 different choices over 4 consecutive trials.
Finally, Eq (4) models decision-making as a stochastic process controlled by the consistency parameter C: a higher C value indicates that choices are strongly driven by the composite values derived from Eqs 1–3, whereas a C value of zero indicate random selection of each deck. Note that C results from the transformation of an inverse temperature β (bound between 0 and 5), in order to match PVL, PVL-Delta, VPP and ORL models (where C = 3β -1 as well).
Crucially, the architecture of VSE can account for purely random exploration (β = 0), for purely value-based exploitation (β>>0, θ>0 and φ = 0), for purely directed exploration (β>>0, θ = 0, and φ>0) and for a mixture of value-based exploitation, directed exploration and random exploration. Note that under purely directed exploration, the model predicts that the 4 decks should be successively selected in a cyclical manner, during the whole task (e.g. 3,2,4,1,3,2,4,1,3,etc.), hence reflecting exactly the definition of the SE index. Indeed, in such case, the deck with the highest exploration weight is always the deck which has not been selected for the longest period of time. Finally, a variant of the VSE model including a loss aversion parameter was also tested (VSE+LA). In this model, the update of exploitation weights is therefore identical to that of the PVL model, augmented with the sequential exploration module of the VSE (see S2 Fig for model comparison analyses including VSE+LA).
The comparison of the VSE architecture with the 5 alternative models (described in Methods) was performed using the 504 subjects dataset. First, a fixed-effect analysis comparing summed Bayesian Information Criterion (BIC), Akaike Information Criterion (AIC) and Free Energy (F) metrics over the whole cohort demonstrated decisive evidence in favor of VSE. In order to compare Free Energy (F) with the other metrics, it was transformed to -2*F for this analysis [18]. The difference between the VSE model and other models was everywhere superior to 512 (Fig 2A; the least difference being observed with the VPP model based on the Free Energy estimator). Note that a difference superior to 100 is generally considered as decisive evidence indicating that choosing the second-best fitting model would incur unacceptable information loss [19]. Going further, we performed a Bayesian Group Comparison (Stephan et al., 2009) based on the log-evidence of each model and treating model attribution as a random effect. In order to obtain log-evidences, we transformed AIC and BIC values to -AIC/2 and -BIC/2, respectively (Free Energy natively represents that quantity and takes into account the uncertainty over parameters when penalizing for model complexity). Performed using all available metrics (BIC, AIC, F), this analysis showed that the estimated frequency of the VSE model was in every case superior to 40% and that its approximate exceedance probability (Ep, probability that a given model is the best candidate model to explain the data) was always superior to 0.99 (Fig 2B). Overall, both approaches to model comparison provided overwhelming evidence in favor of the VSE architecture.
Regarding the relationship of model parameters with performance (defined as the number of advantageous minus disadvantageous deck selection), it appeared that value sensitivity was the strongest predictor (ρ = 0.39, p<0.001), followed by φ (ρ = -0.24, p<0.001), decay (ρ = 0.23, p<0.001) and temperature (ρ = -0.12, p = 0.006)(S1A Fig). Moreover, although a substantial interindividual variability was observed in SE, the parameter φ corresponding to the exploration bonus was significantly superior to 0 (z = 4.78, p<0.001), in line with the existence of directed exploration. A correlation approach then confirmed that this key parameter reflected to which extent participants engaged sequential exploration as assessed by the SE index (Spearman ρ = 0.76, p<0.001). Finally, since the likelihood of observing SE depends on which extent participants exploited the reward structure of the IGT, other parameters also predicted the SE index, but to a much lesser extent (update of exploration weights: ρ = -0.26, p<0.001; value sensitivity: ρ = -0.20, p<0.001; consistency parameter: ρ = 0.13, p = 0.003)(S1B Fig).
In order to make sure that the advantage of the VSE model reflected a better ability to predict choice data, both qualitatively and quantitatively, we used the 504 sets of parameters associated with each model to simulate 504 in silico agents playing the IGT. For each deck, feedbacks were drawn randomly from their corresponding empirical distributions, hence keeping reward contingencies similar across actual and simulated datasets. Then, we applied the exact same fitting procedure to this simulated dataset.
First, we evaluated to which extent each model was able to reproduce participants’ choices using both first-pass and second-pass (i.e simulated) predictions (chance level: 25%, Fig 2C). Again, the VSE model was the most performant model. The first-pass (i.e fit on actual data) reproduced 59.2+/-16% of the choices, whereas the second best model in this respect (VPP) reproduced 58.7+/-17% of the choices. While the difference between VSE and VPP was not significant (z = 0.67, p = 0.50), it must be noted that the VPP has 3 more free parameters which results in a greater chance of overfitting. In this respect, it is interesting to note the difference observed when comparing how the choices derived from simulated data reproduced participants’ choices. Here, the advantage of VSE was clear, with 42.5+/-17% of successful predictions against 39+/-16% for VPP (z = 3.63, p<0.001). Importantly, the VSE model was also the model which predicted the highest number of sequential exploration events (Fig 3A; 1993 against 1389 for VPP, the second model in this respect; raw data: 5247 events) and it was also the most sensitive model according to a d-prime analysis computed over all participants (Fig 3B; 1.15 against 1.11 for the VPP). Note that dependent quadruplets were used for this analysis (ie. 1–4, 2–5, etc.).
Second, we evaluated to which extent each model could be recovered. Low models recovery rates implies that the choice data generated by a given model can be better explained by other models, hence suggesting that there is no specific behavioral signature associated with this model and that the interpretations of the best-fitting parameters should be taken with caution. This computationally intensive analysis (see Methods for details) showed good recovery performance for three models: EV, PVL and VSE (Fig 3C). The advantage for the model actually used to generate the data was decisive in all cases for all estimators, except for the BIC metric, when fitting data generated with the VSE model with the PVL model. This latter exception is very likely due to the fact that the BIC over-penalized model complexity and thus favored the simpler PVL architecture. Moreover, the fact that the highest confusion occurred between the VSE and the PVL models is sensical, given that the “exploitation module” of the VSE model is based on the same accumulation mechanism used by the PVL model. By contrast, the analysis of the three remaining models (ORL, PVL-Delta and VPP) showed that the confusion spanned several models on at least one estimator (e.g. VSE, PVL and VPP performed better to fit ORL-generated data, all models performed better to fit VPP- and PVLDelta-generated data).
Third, we investigated how the parameters estimated from individual choice data could be recovered for each model. Indeed, methodological studies in the field of computational modeling have demonstrated that different combinations of parameters can account for the same sequence of decisions, and that small deviations in parameters values can conversely result in significantly different sequences of decisions, hence impeding the interpretability of best-fitting parameters in some cases. Parameter recovery was assessed by examining how the best-fitting parameters from the second-pass correlated with the best-fitting parameters from the first-pass (i.e that based the actual data). Overall, the recoverability of parameters of VSE was superior to that of other models (mean R = 0.81, range: 0.67–0.95). EV, PVL and ORL also showed good recoverability (EV, mean = 0.76, range: 0.66–0.83; PVL: mean = 0.79, range 0.51–0.94; ORL: mean = 0.77, range = 0.65–0.89), while PVL-delta and VPP were less stable (PVL-delta: 0.71, range: 0.5–0.86; VPP, mean = 0.70, range: 0.41–0.94). In particular, it must be noted that the parameter φ reflecting the exploration bonus of the VSE had the highest recoverability (0.95), hence making it a relevant target for the study of inter-individual differences (Fig 3D).
In their study (included in the 504 participants dataset analyzed above), Woods and colleagues reported that old and young adults performed equally well on the Iowa Gambling Task but resorted to different strategies. More precisely, old adults appeared to forget more rapidly about outcomes than healthy participants but compensated this forgetting by a better ability to translate what they learned into consistent choice patterns.
Thus, we used this subset of the data to evaluate how well the VSE model could capture heterogeneities in IGT strategies and to validate our modeling approach (Fig 4A). In particular, based on the existing literature, we hypothesized that the exploration bonus should be lower in old as compared to young participants. First, we confirmed that old participants indeed forgot more rapidly than young participants according to the VSE model, as indicated by a lower decay parameter (young: 0.55+/-0.27; old: 0.44+/-0.23; z = 2.75, p = 0.006). Second, the consistency parameter of old participants was indeed higher than that of young participants (young: 0.75+/-0.42; old: 0.92+/-0.41; z = 2.71, p = 0.007). Third and most importantly, young and old participants differed significantly in their φ parameter controlling the intensity of directed exploration in the IGT (young:0.94+:-2.14; old: 0.54+/-2.25; z = 2.10, p = 0.036). This latter result paralleled the model-free analysis of SE indexes which also revealed a reduction in directed exploration in the aging group (pattern frequency: young = 16.5+/-14/6%, old = 10.5+/-12.8%; t(151) = 2.61, p = 0.01; Fig 4B) Overall, these results demonstrate the ability of the VSE model to capture age-related changes in directed exploration.
In this study, we uncovered a new choice pattern reflecting the presence of directed exploration within the standard version of the IGT. Indeed, the selection of 4 different decks over 4 consecutive trials—a phenomenon captured by the “sequential exploration index”—largely exceeded chance levels in a group composed of 504 participants, especially in the initial phase of the task. This discovery implies that the IGT can be used to study information-seeking behaviors within risky decision-making contexts. In order to better characterize and quantify this cognitive process within single individuals, we developed a new computational model (VSE, for Value and Sequential Exploration) able to articulate directed exploration with the motivation to optimize gains and losses, using only 5 parameters. The VSE model outperformed the 5 most prevalent models previously used to fractionate the cognitive processes engaged by the IGT, in terms of log-likelihood (penalized for model complexity), prediction accuracy, as well as model and parameter recovery rates. We further demonstrated the potential of this architecture to capture fine-grained differences in IGT behavior between young and old participants. Last but not least, we published the scripts used to generate our results under the form of a user-friendly Matlab toolbox which shall enable the community of researchers and clinicians relying on the IGT as a routine assessment of risky decision-making to report more informative and detailed results with minimal programming and mathematical skills.
In the field of reinforcement-learning, most algorithms are oriented towards normative utility-maximization goals. To do so, they rely heavily upon reward prediction errors, a quantity widely used as a teaching signal enabling step-by-step convergence towards a utility maximum. Yet, likewise all gradient ascent methods, reinforcement-learning algorithms face the risk of reaching only local, rather than global, utility maxima. Directed exploration aims at solving this problem by expanding knowledge about the environment, despite immediate opportunity costs. Thus, directed exploration is particularly valuable when an agent is required to perform numerous decisions within complex or volatile contexts, in which the optimal policy may not be immediately obvious. The IGT is a canonical example of such environment. Indeed, with 4 possible actions leading to highly variable outcomes, the IGT is typically characterized by several successive phases: a “pre-punishment” period during which all decks only produce gains but no losses, a “pre-hunch” period during which punishments start occurring, a “hunch” period during which most healthy participants start feeling that the decks offering the highest average gains actually entails even higher average losses (i.e. deck A and B) and a “conceptual” period during which these participants are able to verbalize that the decks offering small gains (i.e. C and D) are actually the most advantageous ones [20].
Although the term exploration is often applied to choices which are not maximizing utility with respect to a given learning model, the recent rise of predictive coding has completed this conceptualization [21–23]. Indeed, this framework postulates that uncertainty-minimization constitutes a driving principle of our cognitive system alongside utility-maximization, such that modeling the dynamics of exploratory decisions became an important endeavor in the field. Accordingly, recent studies have investigated uncertainty-minimizing strategies in multi-armed bandit tasks using Bayesian methods. In this formalism, options whose mean value is the least precise (or, equivalently, associated with the largest variance) are the best candidates for exploration. The existence of uncertainty-driven exploration was confirmed by some of these studies [9,24,25]. Yet, the type of directed exploration described here does not involve uncertainty computations. Instead, it relies on a simpler recency approach which promotes the exploration of options which have not been selected for a while, independently of the objective uncertainty bound to their pay-offs. This logics can be justified in three ways. First, despite its mathematical elegance, uncertainty-driven exploration does not provide a fully normative solution to the exploration-exploitation trade-off in multi-armed bandits (the process being heavily dependent upon higher-order priors regarding the structure of tasks). Second, the recency method implemented here might still reflect an uncertainty-based mechanism if the subjective uncertainty associated with a given option increases with the duration elapsed since that option was tested for the last time. Third, the seminal study of Daw and colleagues had shown that uncertainty-driven exploration was not useful to describe exploratory patterns in a 4-armed bandit task sharing many commonalities with the IGT [8]. The computational costs associated with uncertainty-tracking may thus become too high within tasks involving more than 3 options. Accordingly, this number recently appeared as an upper limit on the number of stimulus-response mappings whose reliability can be simultaneously monitored by the human executive system [7,26].
The exploration module of our VSE architecture helped going beyond existing models used to account for healthy participants’ decisions in the IGT. Combined with the model-free analysis of the SE index, it expands the heuristic value of the IGT beyond the study of exploitation and reward-seeking behaviors. Moreover, the fact that VSE parameters were on average more recoverable than parameters of previous models will facilitate the interpretation of inter-individual and inter-group differences. Numerous studies which had used the IGT to characterize clinical populations may thus benefit from re-analyzing their data using the toolbox associated with the current paper. Once the trial-by-trial IGT data is converted to the appropriate Matlab format, this toolbox make such re-analysis extremely simple and intuitive, thanks to its compact but informative documentation and its densely commented scripts. With minimal programming knowledge, the six models described hereinabove can be fitted to any standard IGT dataset, compared and evaluated with respect to recoverability and prediction accuracy. These variables as well as other model-free measures (net scores, sequential exploration indices, choice entropy, etc.) can also be calculated, plotted and compared across different groups. While our analyses indicate that the IGT can be suitable to study sequential exploration in conjunction with the VSE model, it must however be emphasized that no modeling approach can circumscribe the non-orthogonal relationship of risk and valence inherent to the structure of the task. In particular, the IGT is not well suited to estimate loss aversion, as implemented by the VPP, PVL and PVL-Delta models. Accordingly, our data showed that the loss aversion parameter had very low recovery rates in the PVL-Delta and VPP models (0.55 and 0.48, respectively) and likely interfered with the value sensitivity parameter in the PVL model (recovery rate: 0.52). Moreover, adding a loss aversion parameter to the VSE model did not improve model fits (see S2 Fig for details). Researchers specifically interested in loss aversion should thus use other paradigms designed to do so [27].
At the neurobiological level, directed exploration likely depends on the prefrontal cortex (PFC), and more particularly on its rostrolateral portion (rlPFC). Indeed, several neuroimaging studies of directed exploration found that the rlPFC is more active during exploratory decisions (Badre et al., 2012; Boorman et al., 2009; Daw et al., 2006). Brain stimulation studies further showed that disrupting or facilitating rlPFC activity can significantly diminish or increase directed exploration, respectively [9,28]. Importantly, disrupting rlPFC activity using continuous theta burst TMS similarly lowered directed exploration and exploration bonuses as assessed by the VSE model in the IGT [29]. This involvement of the rlPFC might also explain the decrease in directed exploration seen in aging individuals, as grey matter density in this area is significantly reduced in old as compared to young adults [30]. Yet, other neural systems certainly interact with the rlPFC to orchestrate information-seeking in reinforcement-learning tasks, including the dmPFC which may control the switch from exploitation to exploration [31]. The prefrontal turn-over of dopamine might also play a pivotal role in regulating directed exploration [32,33], whereas noradrenaline might be involved in the control of random but not directed exploration [34].
In order to further validate our model and illustrate the utility of VSE for the analysis of group differences, we investigated how aging influenced its parameters and more particularly the exploration bonus parameter. The results of this analysis were well aligned with those reported in the study of Wood and colleagues [15], in that the VSE model still evidenced the exacerbated forgetting of previous outcomes in older adults, as well as the reduction in random exploration (i.e increased choice consistency) thought to compensate faster forgetting rates in these participants. More importantly, old adults also displayed a lower exploration bonus than young adults. This effect paralleled the reduction in directed exploration observed when computing directly the frequency of choosing 4 different decks over 4 consecutive trials (SE index). It is also highly consistent with recent papers showing that directed exploration reduces across lifespan [16,17]. Since directed exploration requires the retention of the last few choices made in the task, the phenomenon may be related to the decline of working memory performances sometimes observed in aging cohorts [35].
Beyond the study of cognitive aging, the good recoverability of the VSE model itself and the excellent recoverability of its free parameters constitute two useful features with respect to the development of computational psychiatry. Indeed, the Variational Bayes approach adopted here can be readily combined with the advanced clustering techniques underlying this growing field of research (https://mbb-team.github.io/VBA-toolbox/), which aims at redefining the dimensionality of behavioral impairments across clinical labels in the hope of promoting drug discovery and personalized medicine [36]. To which extent the decomposition of IGT-related behaviors will contribute to this effort remain uncertain, but it has a high potential which could be realized if more clinical teams subscribe to the open-science philosophy by sharing their raw data. Following the initiative of Ahn and colleagues who provided data of stimulant and opiate users [37], addiction research appears as a timely candidate: indeed, large datasets exist for alcohol use [38,39], cannabis use [40], as well as for behavioral addictions such as gambling and eating disorders [41,42].
To conclude, our study leveraged the power of an open “many labs” dataset in order to demonstrate the existence—and characterize the influence—of an overlooked behavior in the IGT. Building on previous work and more particularly on the Prospect Valence Learning (PVL) model [43], the VSE architecture represents not only a quantitative but also a qualitative improvement upon alternative models by shedding light on directed exploration. Besides enabling any experimenter to fit the VSE and its ancestors (EV, PVL, PVL-Delta, VPP, ORL) on IGT data, the toolbox accompanying this paper might be used as an environment to develop even better models in the future. It must be acknowledged that this tool relies heavily on two other open-source packages for Matlab: modeling analyses largely depend on the VBA toolbox by Daunizeau and colleagues [18] whereas visualizations take advantage on the Gramm toolbox by Morel [44]. Last but not least, this study is fully aligned with the ideals of reproducibility and transparency in science: the dataset used is both large and freely available, while the scripts used to generate figures and statistics are available online alongside a clear documentation (https://github.com/romainligneul/igt-toolbox).
The dataset comes from a ‘many labs’ initiative grouping 10 studies and containing data from 617 healthy participants [13]. Here, we restricted the analysis to the subset of 7 studies which used the classical 100 trials version of the IGT, resulting in 504 participants (age range: 18–88 years; for the 5 studies with available information about sex: 54% of females). Within this dataset, 153 participants come from a single study on aging [15]. Among these participants, 63 are older adults (61–88 years old; 17 males) and 90 are younger adults (18–35 years old; 22 males) matched in terms of education level and intelligence (WASI vocabulary).
In order to quantify directed exploration in the IGT, we computed the probability of choosing the 4 different decks during series of 4 consecutive trials. We refer to the frequency of such choice pattern as “SE index”. We used this metrics because the occurrence of such events has a probability of only 9.38% under purely random exploration (note that exploitation makes this probability even smaller by introducing an imbalance in the choice probability of different decks). Although directed exploration is certainty governed by more complex heuristics (resulting in more complex choice patterns), this index was used to ascertain its presence and provide an estimation of its intensity. Inferences about the presence of sequential exploration used independent quadruplets of successive trials (i.e: 1–4, 4–8, etc.), whereas inferences about interindividual differences used dependent quadruplets to maximize sensitivity (i.e: 1–4, 2–5, 3–6, etc.).
Four previous models have been exhaustively and excellently described in a previous publication by Steingroever and colleagues [45]. The ORL model is described in Haines et al. [6]. Therefore, we will only provide a brief overview of their characteristics and then focus mainly on describing the features of the new VSE model.
All the models described above have in addition a consistency parameter determining to which extend choices are driven by learned values (or any type). This consistency parameter c is allowed to fluctuate in the [0,5] interval and is transformed before being used as an inverse temperature parameter β (β = 3c-1), except for the EV model where c is allowed to fluctuate in the [–2,2] interval and is transformed differently (β = (t/10)c with t corresponding to current trial number). In sum, the EV model has 3 parameters, the PVL and PVL-delta models have 4 parameters, the VPP model has 8 parameters and the ORL has 5 parameters. Note that the consistency parameter capture the opposite of “random exploration” (i.e. decision temperature).
A validated toolbox (http://mbb-team.github.io/VBA-toolbox) was used to optimize model parameters [18]. This toolbox relies on a Variational Bayesian (VB) scheme. Compared to non-Bayesian methods, this approach has the advantage of accounting for the uncertainty related to estimated model parameters and of informing the optimization algorithm about prior distributions of parameters’ values. All priors were natively defined as Gaussian distributions of mean 0 and variance 3, which approximates the uniform distribution over the [0–1] interval after a sigmoid transformation. Depending on the range of values in which each parameter was allowed to vary, the sigmoid-transformed parameters were further stretched or shifted to cover different intervals while preserving the flatness of their prior distribution (e.g. “multiplied by 2, minus 1”, to obtain the interval [–1,1]). Model comparison results were replicated using a non-Bayesian model fitting procedure which relied on the standard fminunc function of Matlab (line-search algorithm).
All hidden states (i.e values) were initialized at 0 except for exploration weights which were initialized at 1 (since no deck has been sampled at the beginning of the task). The VB algorithm was not allowed to update the initial values for hidden states.
Comparison of VSE model with the 5 alternatives was first based on a classical fixed-effect analysis comparing summed Bayesian Information Criterion (BIC), Akaike Information Criterion (AIC) and Free energy (F) metrics over the whole group. In this approach, it is classically considered that a difference of 10 units between the models with the lowest and the second lowest criterion value reflects very strong evidence in favor of the model with lowest value (corresponding to a Bayes Factor of 150).
Then, a Bayesian Group Comparison was performed which treated model attribution as a random-effect varying from subject to subject. Also based on BIC, AIC and F, this type of analysis produces an exceedance probability corresponding to the probability that a given model is more likely than any other candidate model (Stephan et al., 2009).
There is a growing consensus among computational neuroscientists that evaluating models only based on estimators such as the AIC or BIC is not sufficient [14,47]. The problem is particularly salient when one aims at drawing inferences about cognitive processes from estimated parameters (which is most often the case), because the same choice pattern can sometimes be explained by very different combinations of parameters and because models associated with lower information losses do not always better reproduce qualitative choice patterns. To address these issues and ensure that the VSE model performed equivalently or better than the VPP model in this respect, we performed the simulation and parameter recovery analyses detailed below.
We used the best-fitting parameters of each subject to simulate an artificial decision-maker confronted to the IGT. Simulated choices were generated stochastically according to the consistency parameter, and feedbacks (gains/losses) were drawn from the distributions of feedbacks actually encountered by the participants. Then, we reran model estimations based on these simulated choices, which resulted in a new set of parameters. The quality of parameter recovery for the VSE and VPP models could then be assessed by examining the correlation of this second set of parameters with the parameters initially obtained by fitting real choices. We examined to which extent the initial choices predicted by the model and the choices performed by the simulated participants matched the actual choices of the participants, across models. In this latter analysis, we restricted our statistical inference and compare the VSE model with the second-best fitting model only.
The model recovery analysis consisted in: (i) using each model to simulate 504 series of 100 trials using the parameters distribution obtained after fitting the model on the real dataset; (ii) fitting the 6 candidate models on each of these 6 simulated datasets, hence requiring in 18144 individual fits; (iii) performing a separate model comparison for each of the 6 simulated datasets.
|
10.1371/journal.pgen.1003038 | Blockade of Pachytene piRNA Biogenesis Reveals a Novel Requirement for Maintaining Post-Meiotic Germline Genome Integrity | Piwi-interacting RNAs are a diverse class of small non-coding RNAs implicated in the silencing of transposable elements and the safeguarding of genome integrity. In mammals, male germ cells express two genetically and developmentally distinct populations of piRNAs at the pre-pachytene and pachytene stages of meiosis, respectively. Pre-pachytene piRNAs are mostly derived from retrotransposons and required for their silencing. In contrast, pachytene piRNAs originate from ∼3,000 genomic clusters, and their biogenesis and function remain enigmatic. Here, we report that conditional inactivation of the putative RNA helicase MOV10L1 in mouse spermatocytes produces a specific loss of pachytene piRNAs, significant accumulation of pachytene piRNA precursor transcripts, and unusual polar conglomeration of Piwi proteins with mitochondria. Pachytene piRNA–deficient spermatocytes progress through meiosis without derepression of LINE1 retrotransposons, but become arrested at the post-meiotic round spermatid stage with massive DNA damage. Our results demonstrate that MOV10L1 acts upstream of Piwi proteins in the primary processing of pachytene piRNAs and suggest that, distinct from pre-pachytene piRNAs, pachytene piRNAs fulfill a unique function in maintaining post-meiotic genome integrity.
| Small non-coding RNAs play critical roles during development and in disease. The integrity of the germline genome is of paramount importance to the wellbeing of offspring and the survival of species. Piwi-interacting RNAs (piRNAs) are a class of small non-coding RNAs abundantly expressed in the gonad. Compared to microRNAs and small-interfering RNAs (siRNAs), the biogenesis and function of piRNAs remain poorly understood. Here we have identified MOV10L1, a putative RNA helicase, as a master regulator of piRNA biogenesis in mouse. We find that production of pachytene piRNAs requires MOV10L1. Blockade of pachytene piRNAs disrupts germ cell development and results in defects in post-meiotic genome integrity. Therefore, mutations in MOV10L1 and other piRNA pathway components may contribute to male infertility in humans.
| Piwi-interacting RNAs (piRNAs) are a diverse class of gonad-specific small interfering RNAs that bind to members of the Piwi subfamily of Argonaute proteins. One common function of piRNAs in all species studied so far is the silencing of transposable elements, which is essential for the protection of genome integrity during germ cell development [1]–[3]. Distinct from miRNAs and siRNAs in origin, length, structure, and biogenesis, piRNAs are generated by dicer-independent processing of long precursor transcripts, however, the precise mechanisms of their biogenesis remain largely unclear [4], [5]. In mice, the Piwi family has three members: Miwi (Piwil1), Mili (Piwil2), and Miwi2 (Piwil4). These Piwi genes exhibit different developmental expression patterns in testis. While Miwi2 is expressed in fetal and perinatal germ cells [6], the expression of Miwi is restricted to pachytene spermatocytes and round spermatids in adult testes [7]. Mili is expressed from the fetal germ cell stage onwards through the round spermatid stage [8]. Two developmentally distinct populations of piRNAs are expressed in mouse male germ cells at the pre-pachytene and pachytene stages. Pre-pachytene piRNAs are mostly derived from transposable elements and are associated with MILI and MIWI2 in fetal and perinatal male germ cells [6], [9], [10]. Pachytene piRNAs originate from ∼3000 genomic clusters [11] and bind to both MILI and MIWI [12]–[17]. Interestingly, more than 90% of MILI- and MIWI-bound pachytene piRNAs shared identical 5′end sequences [18]. As a result, most MILI- and MIWI-bound pachytene piRNAs map to the same genomic clusters [18].
The biogenesis of piRNAs involves primary and secondary processing mechanisms [1], [2]. Pre-pachytene piRNAs derive from precursor transcripts that are cleaved into putative primary piRNA intermediate molecules by a yet unknown primary processing mechanism, followed by loading onto MILI for further processing. In embryonic germ cells, the endonuclease (slicer) activity of MILI is required for the secondary piRNA processing mechanism, which amplifies MILI-bound piRNAs through an intra-MILI ping-pong loop and generates all MIWI2-bound secondary piRNAs [19]. In this feed-forward ping-pong model, Piwi proteins with piRNAs complimentary to retroelement-derived transcripts drive transcript cleavage and piRNA amplification [6], [9], [10], [19]. In contrast, the biogenesis of pachytene piRNAs only engages the primary processing mechanism, i.e. the presumptive cleavage by an unknown nuclease and eventual processing of the precursor transcript into mature piRNAs [5], [17], [20], [21]. Therefore, pachytene piRNAs provide a simple and ideal system for dissecting the mysterious primary processing mechanism in mammals [11], [13]–[16].
We and others previously demonstrated that MOV10L1, a putative RNA helicase, interacts with all mouse Piwi proteins and is required for biogenesis of pre-pachytene piRNAs [22], [23]. MOV10L1 homologues are evolutionarily conserved among insects (Armi in Drosophila melanogaster), plants (SDE3 in Arabidopsis thaliana), and vertebrates (MOV10 and MOV10L1). Arabidopsis SDE3 is required for post-transcriptional gene silencing [24]. Drosophila Armi is essential for the maturation of RISC (RNA-induced silencing complex) and miRNA-mediated silencing [25], [26]. Armi is also relevant to the piRNA pathway, evident from the loss of specific piRNAs and the activation of retrotransposons in armi mutants [27], [28]. Specifically, Armi plays an essential role in the primary piRNA processing pathway [29]. In contrast to Drosophila and Arabidopsis with a single Mov10l1 homologue, the vertebrate genome encodes two genes (Mov10 and Mov10l1), which apparently arose by gene duplication. MOV10 is ubiquitously expressed and associates with Ago proteins, forming part of the purified human RISC [30], [31]. Depletion of MOV10 in cultured cells leads to reduced miRNA-mediated silencing [30]. We initially identified MOV10L1 as a putative RNA helicase that is specifically expressed in mouse germ cells [32], [33]. Disruption of Mov10l1 leads to meiotic arrest, de-repression of transposable elements, and depletion of both MILI- and MIWI2-associated perinatal piRNAs [22], [23]. Apparently, MOV10 and MOV10L1 function in the miRNA and the piRNA pathway, respectively, due to specialization after gene duplication during vertebrate evolution.
The existing piRNA pathway mouse mutants either fail to deplete all pachytene piRNAs or exhibit meiotic arrest prior to the pachytene stage, leaving the biogenesis and role of pachytene piRNAs largely unexplored. Inactivation of either Mili or Miwi2 causes postnatal meiotic arrest at the leptotene/zygotene stage in the male germline [8], [34]. Similarly, other piRNA pathway mutants, such as Ddx4 (Vasa), Mael, Gasz, Tdrd9, Mov10l1, and Mitopld, also exhibit early meiotic arrest in males [22], [35]–[40]. Inactivation of Miwi leads to spermiogenic arrest at the round spermatid stage [7]. However, MILI-associated pachytene piRNAs are abundant in Miwi-deficient testes [17], [18]. Therefore, a mouse mutant containing pachytene spermatocytes, but lacking all pachytene piRNAs (both MILI- and MIWI-bound piRNAs) has not been available to specifically study the function of pachytene piRNAs. In this study, we have specifically and completely depleted the pachytene piRNA population in the male germline of Mov10l1 mutant mice, uncovering a novel function for pachytene piRNAs in maintaining post-meiotic genome integrity.
MOV10L1, a putative RNA helicase, interacts with all three mouse Piwi proteins, and is an essential component of the piRNA pathway [22]. To explore the biogenesis and function of pachytene piRNAs, we disrupted MOV10L1 function specifically during and after male meiosis using Cre-mediated inactivation of a conditional Mov10l1 allele (Mov10l1fl) (Figure 1A and Figure S1) at the following stages: after postnatal day 7 (Neurog3-Cre) [41], at the pachytene stage (Hspa2-Cre) [42], and in post-meiotic spermatids (Prm-Cre) [43].
Cre-mediated recombination of the conditional Mov10l1 allele deletes the RNA helicase domain, producing a truncated protein termed MOV10L1Δ. In male Mov10l1fl/- Neurog3-Cre mice resulting from intercrosses of Mov10l1fl/fl mice with Neurog3-Cre mice [41], Cre-mediated disruption of Mov10l1 was first detected in testes at postnatal day 9 (leptotene/zygotene spermatocytes), with a decrease in the abundance of the full-length MOV10L1 protein in the mutant testes compared with the wild type (Figure S2A). Mov10l1fl/- Neurog3-Cre males were sterile, with substantially smaller testes (140±10.5 mg/pair at 2–4 months of age) compared to age-matched wild-type mice (189±18.4 mg/pair) (Student's t test, p<0.0008). In contrast to seminiferous tubules from wild-type mice (Figure 1B), tubules from Mov10l1fl/- Neurog3-Cre mutant mice lacked elongated spermatids, while earlier germ cell stages including pachytene spermatocytes and round spermatids were present (Figure 1C). Acrosome staining with the anti-ACRV1 antibody identified spermiogenic arrest at the step 4 spermatid stage. Therefore, very different to the meiotic arrest observed in male germ cells with ubiquitous deletion of Mov10l1 [22], [23], postnatal disruption of Mov10l1 mediated by Neurog3-Cre causes post-meiotic spermiogenic arrest (Figure 1C), revealing that MOV10L1 plays an essential role during the post-meiotic stages of spermatogenesis.
To distinguish consequences of inactivation of MOV10L1 during the pachytene stage from those resulting from disruption at earlier stages such as in differentiating spermatogonia, we generated Mov10l1fl/- Hspa2-Cre mice in which Cre is expressed specifically in spermatocytes, particularly pachytene cells (Figure 1A) [42]. Deletion of MOV10L1 in Mov10l1fl/- Hspa2-Cre mice occurred by postnatal day 14, apparent from a decrease in the abundance of the full-length MOV10L1 protein in the mutant testes (Figure S2B). Notably, Mov10l1fl/- Hspa2-Cre males were also sterile. Although testes (159±24 mg/pair) from 2–3 month old Mov10l1fl/- Hspa2-Cre mice were slightly smaller than those from Mov10l1+/− males (182±26 mg/pair) (Student's t test, p<0.2), histological analysis revealed spermiogenic arrest at the round spermatid stage (Figure 1D). The most advanced spermatids in Mov10l1fl/- Hspa2-Cre males were late round spermatids at step 8. The arrest of spermiogenesis at early and late round spermatid stages in Mov10l1fl/- Neurog3-Cre and Mov10l1fl/- Hspa2-Cre mutant mice, respectively, demonstrates that MOV10L1 is required for the differentiation of post-meiotic germ cells. The temporal delay in the spermiogenic arrest in Mov10l1fl/- Hspa2-Cre testes is likely due to the late onset of Hspa2-Cre expression, which may allow residual MOV10L1 to persist longer.
The round spermatid arrest in Mov10l1fl/- Neurog3-Cre and Mov10l1fl/- Hspa2-Cre testes could be due to disruption of MOV10L1 function during the pachytene stage of meiosis, or at early spermatid stages. To define the requirement for MOV10L1 more precisely, we disrupted Mov10l1 with Cre recombinase under the control of the protamine 1 (Prm) promoter, which is only expressed in post-meiotic spermatids [43]. Mov10l1fl/- Prm-Cre males exhibited normal fertility but a slight reduction in testis weight (Table S1). Histological analysis of testes from Mov10l1fl/- Prm-Cre males revealed normal spermiogenesis (Figure S3). These genetic studies demonstrate that disruption of MOV10L1 function at the pachytene stage causes spermiogenic arrest.
Isolation and radiolabeling of total testicular small RNAs from adult Mov10l1fl/- Neurog3-Cre testes showed that mutant testes were devoid of pachytene piRNAs (Figure 2A). Immunoprecipitation experiments further revealed that both MILI- and MIWI-associated pachytene piRNAs were absent in the mutant (Figure 2B, 2C). As Mov10l1 mutant testes contained less MIWI protein than wild-type testes, we performed serial dilutions of immunoprecipitated complexes to rule out the possibility that the observed loss of MIWI-bound piRNAs was due to the detection limit of the assay. However, MIWI-associated piRNAs were detectable in wild-type testes even when MIWI protein was not detectable (Figure 2C, lane 4), indicating a specific depletion of pachytene piRNAs in the testes from Mov10l1fl/- Neurog3-Cre mice. Moreover, the abundance of pachytene piRNAs was sharply reduced in Mov10l1fl/- Hspa2-Cre testes (Figure 2D). In addition, Northern blotting showed that individual pachytene piRNAs (piR1, piR2, and piR3) were absent in testes from Mov10l1fl/- Neurog3-Cre mice and dramatically reduced in abundance in testes from Mov10l1fl/- Hspa2-Cre mice (Figure 3). As expected, the abundance of individual pachytene piRNAs was not affected in the testes from Mov10l1fl/- Prm-Cre mice (Figure 3). Therefore, MOV10L1 function is essential for the biogenesis of all pachytene piRNAs.
Pre-pachytene piRNAs are present in mitotic germ cells such as spermatogonia (Figure 1A). Because Neurog3-Cre initiated the disruption of Mov10l1 at post-natal day 9, we anticipated that the production of pre-pachytene piRNAs would not be affected in Mov10l1fl/- Neurog3-Cre testes. To test this hypothesis, we performed immunoprecipitation of postnatal day 10 testis lysates with anti-MILI antibodies. Postnatal day 10 testes do not contain pachytene spermatocytes, and express only MILI but not other Piwi proteins (Figure 1A). Consequently, all MILI-bound piRNAs in postnatal day 10 testes are pre-pachytene piRNAs [10]. We found that pre-pachytene piRNAs were present in postnatal day 10 Mov10l1fl/- Neurog3-Cre testes (Figure 2E). Furthermore, Northern blot analysis showed that abundance of a specific pre-pachytene piRNA was not reduced in adult testes from Mov10l1 mutant mice, regardless of whether deletion had been mediated by Neurog3-Cre, Hspa2-Cre, or Prm-Cre (Figure 3). These data demonstrate that pre-pachytene piRNA production is not affected in the Mov10l1 conditional mutant testes.
We next examined the consequences of the loss of pachytene piRNAs on the localization of piRNA pathway components such as MILI, MIWI, TDRD1, and GASZ. In wild-type pachytene spermatocytes, these proteins localize to cytoplasmic nuage granules (also called inter-mitochondrial cement) (Figure 4A, 4C, 4E, 4G) [7], [8], [37], [44]. Strikingly, in Mov10l1-deficient pachytene spermatocytes, these four proteins congregated to one extremely large novel perinuclear polar “granule” (Figure 4B, 4D, 4F, and 4H). Further analyses revealed immunoreactivity of the polar granule to a cocktail of antibodies against mitochondrial proteins (OXPHOS), demonstrating co-localization of mitochondria with MILI in polar granules of Mov10l1-deficient pachytene spermatocytes (Figure S4). Electron microscopy (EM) analysis confirmed that in Mov10l1-deficient pachytene spermatocytes, mitochondria form a single cluster (Figure 4J), in contrast to their random distribution in wild-type pachytene cells (Figure 4I). Consistent with a recently described role for MitoPLD, a mitochondrial surface protein, in the piRNA pathway [39], [40], these data strongly suggest a novel but yet unknown role for mitochondria in the biogenesis of pachytene piRNAs and/or a function for pachytene piRNAs in the cytoplasmic organization and distribution of mitochondria and piRNA pathway protein components.
We next examined the status of chromatoid bodies, which are large and dynamic ribonucleoprotein aggregates prominent in haploid spermatids. Chromatoid bodies contain various RNA regulatory proteins as well as piRNA pathway components, but their precise function remains unclear [45]. Wild-type round spermatids contained one prominent chromatoid body, visualized by EM as a multi-lobular electron-dense nuage (Figure S5). In Mov10l1-deficient spermatids, however, the chromatoid body was fragmented (Figure S5). A similar fragmentation of chromatoid bodies has been observed in other mouse mutants of RNA processing pathway proteins with male infertility phenotype, implying importance of their structural integrity (Miwi, Tdrd5, and Tdrd6) [7], [18], [46], [47].
The introduction of DNA double strand breaks (DSBs) into the germ cell genome takes place as part of the chromatin remodeling process occurring at the elongating spermatid stage (Figure 5A). This chromatin remodeling process is initiated by the replacement of canonical histones first with transition proteins and eventually by protamines. Concurrently, nucleosomal DNA supercoils must be resolved, presumably by topoisomerase IIB (TOP2B). TOP2B generates DNA double-strand breaks (DSBs), relaxes supercoils, and subsequently religates DNA ends [48]. DSBs trigger a DNA damage response, resulting in the phosphorylation of histone H2AX (γH2AX). In wild-type testis, histone H2AX phosphorylation is therefore detectable in several germ cell stages that undergo changes in their chromatin configuration, including elongating spermatids (Figure 5B), but it is absent from round spermatids. Intriguingly, round spermatids from Mov10l1fl/- Neurog3-Cre testes exhibit a high degree of DNA damage visualized by γH2AX (Figure 5C). This could be due to a developmental progression of piRNA-deficient round spermatids to the “elongating” spermatid stage without apparent morphological change. However, the absence of both TOP2B and PRM2 (protamine 2) in γH2AX-positive round spermatids from Mov10l1 mutant testes indicated that these round spermatids were not undergoing chromatin remodeling, excluding that γH2AX-positivity was due to TOP2B activity (Figure 5E, 5G). Secondly, DNA damage might be induced by de-repressed transposable elements active in piRNA-deficient round spermatids. Genetic studies have shown that the piRNA pathway is required for silencing of retrotransposons such as LINE1 and IAP in pre-pachytene germ cells [19]. However, quantitative RT-PCR analysis revealed no de-repression of LINE1 (Figure 6B, 6C) or IAP in Mov10l1fl/- Neurog3-Cre testes, confirmed by immunofluorescent analyses of testis sections with anti-LINE1 and anti-IAP antibodies (Figure S6). Therefore, pachytene piRNAs are not required for silencing of LINE1 and IAP retrotransposons, although we cannot rule out the possibility that other transposable elements might be de-repressed in Mov10l1-deficient round spermatids. Notably, we did not observe γH2AX foci in round spermatids from Rnf17-deficient mice, in which spermatogenesis is also arrested at the round spermatid stage [49] but piRNA biogenesis does not appear to be severely affected (data not shown). These results suggest that the DSBs observed in round spermatids from Mov10l1fl/- Neurog3-Cre testes and Mov10l1fl/- Hspa2-Cre testes are not a direct consequence of their developmental arrest. Rather, these observations suggest that the piRNA pathway, specifically MOV10L1 and pachytene piRNAs, play a yet undefined role in maintaining genome integrity in post-meiotic round spermatids.
We previously found that MOV10L1 interacts with all three Piwi proteins (MILI, MIWI, and MIWI2) [22]. The low abundance of MOV10L1Δ in both Mov10l1−/− (ubiquitous null mutant) testes [22] and adult Mov10l1 conditional mutant testes (Figure S2A) precluded co-immunoprecipitation (IP) experiments to ascertain if deletion of the helicase domain affected interaction of the truncated protein with Piwi proteins in vivo. However, a peak expression of MOV10L1Δ protein in post-natal day 20 Mov10l1fl/- Neurog3-Cre testes, with a level highly exceeding that of the remaining wild-type MOV10L1 (Figure S2A), allowed us to perform co-immunoprecipitation of day 20 testicular extracts with anti-MILI and anti- MOV10L1 antibodies (Figure 6A). While wild-type MOV10L1 (due to the lack of Neurog3-Cre expression in spermatogonia) could be detected as a very faint band in the MILI immunoprecipitate as expected, the much more abundant MOV10L1Δ was absent (Figure 6A, Lane 6). Notably, MILI was not detectable in the MOV10L1/MOV10L1Δ immunoprecipitate, and the level of MIWI was extremely low (Figure 6A, Lane 4). These results suggest that, apart from its putative enzymatic activity, the RNA helicase domain of MOV10L1 is also essential for its association with MILI and MIWI, and that piRNA production could be affected by disruption of the MOV10L1-Piwi interactions.
Pachytene piRNAs are derived from only one strand of genomic clusters [11], [13]–[16], prompting the hypothesis that a single long primary piRNA transcript is made from each cluster and is cleaved into intermediate RNAs by an unknown Dicer-independent mechanism [5], [20], [21]. Due to their large size and low abundance, detection of these precursor transcripts requires RT-PCR analysis, with the exception of a ∼10 kb piLR (piRNA like small RNA) transcript that can be visualized on Northern blots of testicular extracts [50]. As the depletion of pachytene piRNAs in Mov10l1fl/- Neurog3-Cre testes may be due to a blockade of pachytene piRNA precursor processing, we examined the abundance of precursors of four pachytene piRNAs (piR1, piR2, piR3, and piLR) by RT-PCR assays (Figure S7). All four precursors accumulated substantially in Mov10l1fl/- Neurog3-Cre testes, at 8 to 20 fold increased levels (Figure 6B, 6C). As expected, abundance of the pre-pachytene piRNA precursor (cluster 10) [10] and the miRNA precursor Pri-let7g remained constant (Figure 6B, 6C). These data suggest that MOV10L1 is required for the primary processing of precursor transcripts and thus plays an essential role in the early steps of the piRNA biogenesis pathway, i.e. primary processing and loading onto Piwi proteins (Figure 6D).
We have identified MOV10L1 as the only factor known to date that is required for the production of all pachytene piRNAs in mouse. As the biogenesis of pachytene piRNAs only involves the primary processing pathway, our conditional Mov10l1 mutants provide a unique opportunity to delineate this enigmatic component of piRNA biogenesis in mammalian species. Presumably, long piRNA precursor transcripts are first cleaved into intermediate molecules, and then processed into mature piRNAs (Figure 6D). Observations that the Drosophila Armi-Piwi-Yb complex is associated with a population of 25–70 nt piRNA intermediate-like (piR-IL) molecules support this hypothesis [20]. Furthermore, recent biochemical studies using silkworm ovarian cell lysate have shown that intermediate piRNA molecules with 5′ U are specifically loaded onto Piwi proteins and then trimmed from the 3′end to generate mature piRNAs [21]. Here, we show that, in the mouse male germline, postnatal disruption of Mov10l1 does not affect the expression of Piwi proteins (MILI and MIWI) but causes a complete loss of pachytene piRNAs, demonstrating that MOV10L1 functions upstream of Piwi proteins in the piRNA biogenesis pathway. Consistent with its homology to Drosophila Armi [25], [26], MOV10L1 is therefore a master regulator of piRNA biogenesis in mouse. This notion is further supported by the dramatic accumulation of pachytene piRNA precursors in the Mov10l1 mutant testes. As MOV10L1 interacts with Piwi proteins, we postulate that MOV10L1 may facilitate the loading of intermediate piRNA molecules onto the Piwi proteins in mouse (Figure 6D).
In spermatocytes, proteins of the piRNA pathway such as MILI, MIWI, TDRD1, MAEL, and GASZ, localize to the nuage - inter-mitochondrial cement [7], [8], [36], [37], [44], [51]. However, the functional significance of the physical association of nuage with mitochondria in germ cells is poorly understood. MitoPLD, a mitochondrial signaling protein, is essential for nuage formation and piRNA production, suggesting an important role for mitochondria in these mechanisms [39], [40]. In this study, we find an unusual polar congregation of piRNA pathway proteins (such as MILI, MIWI, TDRD1, and GASZ). Similar to wild-type MOV10L1, truncated MOV10L1Δ is distributed diffusely through the cytoplasm of pachytene spermatocytes; therefore the polar coalescence of the other piRNA pathway components in MOV10L1-deficient pachytene cells is likely caused by the absence of pachytene piRNAs. However, as the association of Piwi-MOV10L1 is disrupted in the Mov10l1 mutant, it is also possible that the localization of Piwi proteins and their interacting partners has become perturbed as a consequence of this disruption. The unusual polar congregation of piRNA pathway proteins with mitochondria in Mov10l1 mutant spermatocytes suggests that MOV10L1 and/or pachytene piRNAs are essential for nuage formation and proper mitochondria distribution. Consistent with such a role, we find that the chromatoid body, a prominent nuage in spermatids, is fragmented in pachytene piRNA-deficient mutant cells. This previously unknown role in organelle distribution shows that pachytene piRNAs are intricately integrated in the inter-dependent relationships among piRNA production, nuage formation, and mitochondria organization that are essential for male germ cell maturation.
A recent study has shown that MIWI is an RNA-guided RNase with slicer activity that directly cleaves transcripts of the LINE1 retrotransposon [18]. Miwi-deficient and MiwiADH (slicer inactive) mutant testes, in which MIWI is either absent or lacks slicer activity, exhibit substantial accumulation of LINE1 transcripts and protein. In Mov10l1fl/- Neurog3-Cre testes, however, LINE1 RNA levels are not affected. One possible explanation for these differential effects on LINE1 abundance could be that, in Mov10l1fl/- Neurog3-Cre testes, MIWI is catalytically intact and may function as a slicer through pachytene piRNA-independent mechanisms. Moreover, MIWI directly binds to spermiogenic mRNAs, independent of piRNAs [17].
Although previous genetic studies of piRNA pathway mutants show that perturbation of pre-pachytene piRNAs causes meiotic arrest and de-repression of LINE1 and IAP retrotransposons, the functions of pachytene piRNAs have remained elusive. Our study on the role of Mov10l1 and the piRNA pathway during later stages of meiosis and spermiogenesis demonstrates that pachytene piRNAs fulfill distinct and essential functions during post-meiotic stages of male germ cell development. Most importantly, the massive DNA damage observed in piRNA-deficient round spermatids in the absence of de-repression of LINE1 and IAP transposable elements suggests that the integrity of the post-meiotic germ cell genome remains highly prone to damage, and that pachytene piRNAs fulfill a protective role at this stage by yet undefined mechanisms.
Mice were maintained and used for experimentation according to the guidelines of the Institutional Animal Care and Use Committee of the University of Pennsylvania.
Neurog3-Cre, Hspa2-Cre, and Prm-Cre mice were purchased from the Jackson Laboratory (Stock numbers: Neurog3-Cre, 006333; Hspa2-Cre, 008870; Prm-Cre, 003328). Mov10l1fl/fl mice were generated previously [22]. Genotyping for Mov10l1 and Cre alleles was performed separately on genomic DNA isolated from tails. The anti-MOV10L1 antibody was generated previously [22]. Other antibodies used were MILI (Abcam), MIWI (Abcam, or gifts from R. Pillai), GASZ (M. M. Matzuk), LINE1 ORF1p (S. L. Martin), IAP (B. R. Cullen), TDRD1 (S. Chuma), TOP2B (Santa Cruz Biotechnology), PRM2 (SHAL), and ACTB (Sigma-Aldrich).
Mouse testicular extract preparation, immunoprecipitation, and 5′ end-labeling of piRNAs were performed as described previously [22]. Antibodies were described previously [22].
Northern blot analyses were performed as previously described with modifications [14]. Total RNAs were isolated from mouse testes using Trizol reagent, separated by 15% denaturing polyacrylamide gel, and electro-blotted onto GeneScreen Plus hybridization membrane. Membranes were UV crosslinked and hybridized with 32P end-labeled oligonucleotide probes in Ultrahyb Oligo Buffer (Ambion Cat#8663) at 42°C. Probes for detecting pachytene piRNAs, a pre-pachytene piRNA, or microRNA were perfectly complementary to their sequences: probe-piR1: AAAGCTATCTGAGCACCTGTGTTCATGTCA; probe-piR2: ACCAGCAGACACCGTCGTATGCATCACACA; probe-piR3: ACCACTAAACATTTAGATGCCACTCTCA; probe-let7g: TACTGTACAAACTACTACCTCA; pre-pachytene piRNA probe (derived from sense SINE B1): 5′-TGGCTGTCCTGGAACTCACTYTGT [10]. After hybridization, membranes were washed three times at 42°C in 2×SSC buffer containing 0.5% SDS, or stripped by boiling in 0.1×SSC containing 0.1% SDS. Membranes were exposed to a phosphor imager screen for autoradiography.
For histology, testes were fixed in Bouin's solution overnight, processed, sectioned, and stained with hematoxylin and eosin. Immunofluorescence was performed on frozen sections of testes fixed in 4% paraformaldehyde as previously described [52]. EM followed a standard protocol used at the Electron Microscopy Resource Laboratory of the University of Pennsylvania.
PCR primers for piRNA precursor transcripts were chosen from genomic clusters to which each piRNA was mapped [10], [14], [50]. PCR primers and PCR product sizes are listed in Table S2.
|
10.1371/journal.pgen.1007141 | Sem1 links proteasome stability and specificity to multicellular development | The transition from vegetative growth to multicellular development represents an evolutionary hallmark linked to an oxidative stress signal and controlled protein degradation. We identified the Sem1 proteasome subunit, which connects stress response and cellular differentiation. The sem1 gene encodes the fungal counterpart of the human Sem1 proteasome lid subunit and is essential for fungal cell differentiation and development. A sem1 deletion strain of the filamentous fungus Aspergillus nidulans is able to grow vegetatively and expresses an elevated degree of 20S proteasomes with multiplied ATP-independent catalytic activity compared to wildtype. Oxidative stress induces increased transcription of the genes sem1 and rpn11 for the proteasomal deubiquitinating enzyme. Sem1 is required for stabilization of the Rpn11 deubiquitinating enzyme, incorporation of the ubiquitin receptor Rpn10 into the 19S regulatory particle and efficient 26S proteasome assembly. Sem1 maintains high cellular NADH levels, controls mitochondria integrity during stress and developmental transition.
| The cellular ubiquitin-proteasome pathway is essential to control cell cycle, gene expression or the response to oxidative stress. Sem1 is conserved in eukaryotes from single cell yeasts to humans as intrinsically disordered and multifunctional protein. Sem1 supports the assembly of several multiprotein complexes but becomes eventually exclusively a subunit of the lid of the 26S proteasome, a cellular machine with a molecular mass of about two megadalton. Defects in the function of the proteasome, which degrades a large fraction of intracellular proteins, result in cancer or neurodegenerative diseases. We showed that Sem1 from a multicellular fungus is required for accurate 26S proteasome assembly and specific activity as prerequisites for mitochondria integrity, oxidative stress response and cell differentiation. Our findings of the complex and dynamic interplay between multiple cellular processes mediated by a small conserved intrinsically unordered protein sheds light and supports current efforts to understand and explore in more details potential therapies to eventually treat age-related human diseases.
| The 26S proteasome represents the major cytoplasmic and nuclear ubiquitin-dependent protein degradation machinery and is composed of a barrel-like 2.5 MDa 20S proteolytic core particle (CP) capped with one or two 19S regulatory particles (RP). Proteins destined for degradation are unfolded, de-ubiquitinated and translocated by the RP to the CP where proteolytic activity takes place. Each regulatory particle consists of a lid and a base. The lid is composed of a nine-subunit protein composed of six PCI (Proteasome/COP9/Initiation factor) domain proteins (Rpn3, 5, 7, 9, 12), two MPN (Mpr1p and Pad1p N-terminal) domain proteins (Rpn 8, 11) and the Sem1 (Suppressor of Exocytosis Mutation 1) / Dss1 (deletion of split hand/split foot 1) protein. Sem1 was originally identified in Saccharomyces cerevisiae in genetic suppression studies of the exocyst [1]. It is a multifunctional and intrinsically disordered protein, associating with several functionally diverse protein complexes where it is also supporting assembly without being a subunit of the final active complex itself [2]. This includes the human BRCA2 complex (breast cancer complex) required for homologous recombination, the TREX2 complex (transcription export complex 2) needed for mRNA elongation and nuclear export, or the yeast Csn12-Thp3 complex involved in RNA splicing [2].
Deletion of the Sem1 encoding gene in yeasts resulted in viable cells with phenotypes including temperature sensitivity, enhanced filamentation and cell cycle delay [3]. Dss1 represents the human homolog and was able to rescue the growth defect of both S. cerevisiae and S. pombe mutant strains, suggesting a conserved molecular function from yeast to humans [4, 5].
The in vivo function of Sem1 in multicellular metazoans is difficult to address, as deletion of the corresponding gene in C. elegans revealed essential functions in oogenesis resulting in embryonic lethality and larval arrest [6]. The fungus Aspergillus nidulans can be used as model system as it not only grows by forming polar reiterated cellular units but can also differentiate and produce fruiting bodies including specialized cell types, which are surrounded by a tissue of auxiliary cells for nursing [7, 8].
The function of the fungal Sem1 counterpart was analysed in A. nidulans. The encoding gene was named here for convenience sem1 and corresponds to semA according to A. nidulans nomenclature. Fungal Sem1 protein is associated with the 19S RP and links an appropriate oxidative stress response to cellular differentiation and coordinated fungal development. The majority of the proteasomes in the Δsem1 mutant strain were 20S core particles, which provides an increased ATP-independent protease activity. The small proportion of cellular Δsem1 deficient 19S and 26S proteasomes lacked any interaction with the chaperon Ecm29, which facilitates the association of the 19S RP with the 20S CP. The Δsem1 19S regulatory particles were also deprived from detectable incorporation of the ubiquitin receptor Rpn10, which facilitates the association of the lid with the base. The cellular redox state in A. nidulans is linked to Sem1-dependent transition from vegetative growth to differentiation. Sem1 links increased 26S proteasome stability to mitochondria integrity and is a prerequisite for an appropriate oxidative stress response required for multicellular development.
Sem1 proteins are conserved throughout eukaryotes from fungi to humans (S1 Fig). The impact of Sem1 on developmental programmes of a multicellular fungus was analysed. Filamentous fungi perform the transition from vegetative growth as hyphae to multicellular development by forming fruiting bodies consisting of tissues with distinct specialized cells. A. nidulans exhibits an asexual and a sexual life cycle propagated by spores. Light promotes asexual development and reduces sexual development, whereas darkness and oxygen limitation promote the sexual life cycle [9].
Deletion of the gene encoding fungal Sem1 resulted in a viable Δsem1 deletion strain, which exhibited a reduced growth rate with a smaller colony size compared to wildtype (Fig 1A and 1B). The mutant strain accumulated a reddish pigment similar to a fungal strain defective in the COP9 signalosome, required for specific ubiquitination of substrates (Fig 1A and 1B) [10, 11]. Conidia are the asexual spores of A. nidulans and are formed at conidiophores. Conidia formation, which is normally favoured in light, is delayed in the absence of Sem1 (Fig 1C). Conidia formation was examined during 6 days of asexual growth. In the absence of Sem1, significant reduction of conidia was observed compared to wildtype and complementation strains. The wildtype and complementation strain produced ≈ 150x106 spores/ml after only 2 days, whereas the mutant strain was able to produce less than 1% after 2 days (5000 spores/ml). The number of spores in the mutant strain increased over time, reaching a maximum after 5 days with approximately 64% of the spores produced by wildtype or complementation strains (Fig 1C). To determine whether the delay in conidiophore formation can explain the reduced conidia in the absence of Sem1, the morphology of conidiophores was examined after 20h, 26h and 48h (Fig 1D). No conidiation was observed in the absence of Sem1 after 20h. After 26h, most of the conidiophores of wildtype and complementation strains showed several lines of conidia, whereas the deletion strain only produced metulae and phialides. After 48h, all strains were able to produce conidiophores (Fig 1D). A gene encoding endogenously tagged Sem1-GFP complements these developmental phenotypes resulting in a functional gene that can be used to investigate in vivo Sem1 localization and interaction partners.
Growth of the deletion strain was examined in the dark, which promotes the formation of sexual fruiting bodies named cleistothecia, containing meiotic ascospores (Fig 1E and 1G). Cleistothecia maturation includes the formation of hyphal nests and primordia of fruiting bodies, a development that requires one week for wildtype or the sem1-gfp complemented strain. In contrast, the Δsem1 strain showed after one week mostly white air mycelium and only nests with primordia (Fig 1F). Cleistothecia from wildtype and complementation strains contained meiotic ascospores, which were able to germinate, suggesting the ascospores are viable (Fig 1F, squeezed panel). Cultivation of Δsem1 strain for 10 days still resulted in only white air mycelium and nests with primordia (Fig 1F, 10 days panel). Even after 10 days, cleistothecia from Δsem1 strain contained only an empty cleistothecia-envelop without any ascospores, indicating a specific blockage in sexual fruiting body formation (Fig 1F, squeezed panel). The average size of these empty cleistothecia-envelops was 5850±823 μm2. In contrast, cleistothecia smaller than 5000 μm2 from wildtype or complementation strains (Fig 1G, 15%) were pigmented and contained ascospores (Fig 1F, squeezed panel).
These data demonstrate that the Δsem1 mutant strain can grow vegetatively but is delayed in asexual development and blocked in sexual development with a misregulated secondary metabolism, as indicated by the accumulation of an orange dye. The finding that sem1-gfp can complement all these developmental phenotypes makes A. nidulans an attractive system to compare cellular 26S proteasome composition and assembly with or without Sem1 in a multicellular organism.
Cellular proteasome fractions from Δsem1, wildtype and sem1-gfp complementation strain were isolated and the ratios of intact 26S proteasomes versus 20S CP were compared. Negative staining electron microscopy revealed three forms of proteasome complexes in the cellular fractions, including a 20S CP, composed of a single capped (20S+19S) and a double-capped (19S+20S+19S) proteasomes, respectively (Fig 2A).
Cellular proteasomes with functional Sem1 comprise approximately half of the proteasomes as composed particles, ranging from 45% for Sem1 to 59% for Sem1-GFP. This includes 18% and 41% double-capped proteasomes for wildtype and Sem1-GFP complemented strain, respectively. Single capped proteasomes vary between 18% for the complemented and 27% for the wildtype strain in the analysed fungal extracts.
The ratio between assembled 26S and 20S proteasomes was significantly shifted within Δsem1 mutant strain, which comprised only 6% composite proteasomes including only 2% double-capped proteasomes. Accordingly, the Δsem1 mutant strain produced primarily 20S proteasome complexes (94%), whereas the 20S proteasome complexes of wildtype (55%) and Sem1-GFP complementation (41%) strains represent approximately half of the total cell extract derived proteasome fraction.
This increase in the percentage of 20S proteasomes from 50% to more than 90% of the total cellular proteasomes corroborates that Sem1 is required for an efficient in vivo assembly of 26S proteasomes and that Sem1-GFP can fulfil this function. The function of Sem1 might either be to accelerate the assembly or to stabilize functional 26S proteasomes or a combination of both. 20S CP might represent, even in the presence of Sem1, a substantial part of the cellular proteasome complexes during vegetative growth of the fungus.
Activities of purified proteasomes were measured by monitoring the hydrolysis of a fluorogenic peptide in the presence of ATP and KCl. Δsem1 strain proteasomes showed 2.3 time higher rates of peptidase activities compared to the wildtype strain in the presence of ATP and KCl (Fig 2B, blue curves). Proteasome specificity was further demonstrated by the addition of the proteasome inhibitor MG132, which inhibited these peptidase activities almost completely to 3% or less (Fig 2B, green curves).
The overall proteasome activities in the presence of ATP were compared to the basal ATP-independent peptidase activities derived primarily from 20S CP in the absence of ATP and potassium ions (Fig 2B, red curves). These ATP-independent peptidase activities were lower than the overall activities of proteasomes in all strains, as the absence of ATP and potassium ions retard spontaneous activation of the 20S core particle [12]. In the presence of Sem1, the ATP-independent overall peptidase activity was 26% compared to ATP-dependent activity. This value increased to 75% in the Δsem1 proteasome fraction. The ratio of 26S/20S peptidase activity in the wildtype and complementation strain was 3.8 times and 2.7 times higher, respectively, indicating that 26S proteasomes are more active than the 20S proteasomes. This ratio of peptidase activity was only 1.3 in Δsem1 mutant strain, indicating similar peptidase activities in Δsem1 regardless to the presence or absence of ATP and 19S RP. These data further support the finding of the electron microscopy analysis that Sem1 is required for the efficient assembly of functional 26S proteasomes. The high ATP-independent relative to ATP-dependent peptidase activity of Δsem1 cells in comparison to wildtype could considerably contribute to the observed mutant phenotypes including the failure to establish developmental programmes.
Significant changes in the ubiquitination pattern were detected in the Δsem1 mutant strain with only 6% of conjugated substrates in Δsem1 compared to wildtype cells (Fig 3A). This loss of ubiquitin conjugated proteins in the Δsem1 mutant strain suggests a direct or indirect Sem1 effect on cellular ubiquitination process or on the control of components of the ubiquitin conjugation pathway.
Sem1-dependent transcription of six genes providing cellular ubiquitin was examined to analyse whether Sem1 affects cellular ubiquitin homeostasis. The ubi1 gene encodes a protein where ubiquitin is fused to the small ribosome subunit; thereby synthesis of fusion protein will yield ribosomal protein and ubiquitin. The ubi4 gene product contains four head to tail repeats of ubiquitin and supplies monoubiquitin to the cell. RT-PCR revealed that the transcription levels of both de novo synthesis genes of ubiquitin were similar in Δsem1 mutant strain compared to wildtype (Fig 3B). The genes doa4, ubp6 and ubp14 encode deubiquitinating enzymes (DUBs), which recycle polyubiquitin chains. The rfu1 gene encodes an inhibitor of Doa4 and balances the amount of monoubiquitin and polyubiquitin chains [13]. Transcript levels of these four genes involved in recycling of polyubiquitin were comparable in strains with or without functional Sem1 (Fig 3B). These data suggest that cellular ubiquitin synthesis and recycling functions were independent of Sem1.
A possible function of Sem1 on the ubiquitination pathway was analysed. The last step of the ubiquitination cascade is the attachment of ubiquitin to target substrates by neddylated E3 ubiquitin cullin RING ligases (CRLs). CRLs are under the control of the COP9 signalosome and its deneddylase Csn5/CsnE. COP9 is required for sexual fungal development and physically interacts with Den1/DenA, a second conserved deneddylase, which promotes asexual development [14–16]. The Δsem1 strain showed reduced amounts of neddylated (65%) and unneddylated Cul1/CulA (40%) compared to wildtype (Fig 3C). The majority of the CulA proteins were in their unneddylated and inactive form. RT-PCRs revealed that the expression levels of csnE, culC or culD genes were decreased in the Δsem1 mutant strain compared to wildtype, whereas denA transcripts were unchanged (S2A Fig). This could be due to Sem1’s chaperon functions for the assembly of various protein complexes involved in transcription, RNA splicing or nuclear export [2].
Reduced levels of neddylated cullins and subsequently a limited ubiquitin conjugation of substrates in the absence of Sem1 suggest an important function of Sem1 in the ubiquitination pathway of target proteins due to its impact on transcription and on proteasome assembly and function.
Accelerated proteolysis due to increased numbers of proteasomes can cause decreased levels of ubiquitin conjugates [17, 18]. Transcript levels of genes encoding proteasomal subunits or ubiquitin receptors were compared between strains with or without Sem1 to examine whether increased transcription contributes to decreased ubiquitination observed in the Δsem1 strain (Figs 3B and S2B).
The rpn3 transcription of wildtype cells was significantly higher than in the Δsem1 mutant strain, where transcription was reduced. The rpn3 gene encodes a protein which is tethered by Sem1 to the proteasome lid during biogenesis and interacts in the mature lid with Rpn7 [19]. The reduced transcription of rpn3 transcripts suggest limited incorporation of Rpn3 protein into the 26S proteasome.
The rpn10 gene encodes one of the intrinsic ubiquitin receptors of the proteasome and is located at the interface of the regulatory particle between the base and the lid. Expression of rpn10 as well as of other rpn transcripts for the lid subunits were similar in strains with or without Sem1, except for rpn11 mRNA (Figs 3B and S2B).
The transcription level of rpn11 encoding the intrinsic ubiquitin isopeptidase of the 26S proteasome was doubled in the mutant strain compared to wildtype. Controlled rpn11 expression using an inducible PTetOn-rpn11 fusion gene was applied to examine whether increased amounts of rpn11 transcripts result in higher deubiquitinase activity and reduce the overall population of ubiquitinated proteins. The PTetOn-rpn11 strain was only able to grow in the presence of a threshold concentration of at least 5μg/ml doxycycline, indicating that Rpn11 is essential for growth (S3A Fig). Eight-fold increase in the transcription of rpn11 did not affect the expression of the control genes sem1 or csn5, but led to an overall decrease in ub-conjugated proteins compared to wildtype (S3B and S3C Fig). This supports that significantly increased Rpn11 isopeptidase activity can contribute to the reduction in ubiquitin conjugates as observed in the Δsem1 mutant strain (Fig 3A). Cellular Rpn11 protein levels for the deubiquitinating enzyme were compared to ubiquitin receptor Rpn10 levels in cells with or without Sem1. The genes encoding Rpn11 or Rpn10 were replaced by functional Rpn-GFP fusions. Rpn10-GFP derived from Δsem1 or wildtype strain resulted in stable Rpn10 protein levels. In contrast, Rpn11-GFP was instable and resulted in only 35% of full-length protein in the Δsem1 strain compared to wildtype (Fig 3D).
These data imply a possible effect of Sem1 on the transcription of specific proteasomal genes. Increased rpn11 transcripts result in less full-length Rpn11 protein in a Δsem1 mutant strain lacking the conserved zinc–binding site in the MPN+ domain (S3D Fig), suggesting that Sem1 supports cellular Rpn11 stability. Reduced amounts of an intact 26S proteasomes observed by electron microscopy correlate with the reduced Rpn11 protein levels in Δsem1. Incorporation of the deubiquitination enzyme into the 26S proteasomes presumably provides a Sem1-mediated Rpn11 stabilization in fungal wildtype cells.
The cellular localization of functional GFP fusions of Sem1 and the four RP subunits ubiquitin receptor Rpn10, deubiquitinating protein Rpn11, its inhibitor Rpn5, and Rpn3 that is tethered by Sem1 were compared (Fig 4A). Identical microscopy settings and the same number of spores were used for cultivation to obtain relative concentrations of 19S RP subunits in the hyphae, reflected by GFP intensities.
Significant nuclear staining was observed in the sem1-gfp strain including a minor cytoplasmic and a major nuclear Sem1 subpopulation (Fig 4B). The weakest monitored GFP signal in the cytoplasm and the nucleus was observed for Rpn3, Rpn10 and Rpn11, indicating similar cytoplasmic and nuclear abundance. Rpn5-GFP and Sem1-GFP had similar intensities, which were significantly higher than the Rpn3, Rpn10 or Rpn11 levels. Rpn5 inhibits the Rpn11 deubiquitinase and the increased intensities of Rpn5 might reflect its importance to reduce false DUB activity. Increased cellular Sem1 levels might be required, because it is not only part of the RP of the proteasome but also functions as chaperone in the assembly of several other complexes for cellular processes including transcription.
Affinity purifications of endogenously GFP-tagged Sem1, Rpn3, Rpn5, Rpn11 and Rpn10 combined with subsequent protein identification by mass spectrometry resulted in 34 putative interaction partners for Sem1 and 29 for Rpn GFP-fused subunits (Fig 5).
Sem1-GFP recruited two proteins of the transcription export complex 2 (TREX2) and two proteins homologous to subunits of the yeast transcription regulator complex, Csn12-Thp3, in agreement with previous approaches [2]. These associations were not observed with the other lid subunits and were Sem1-specific. An additional Sem1-specific association was found with the hypothetical protein HypoP2 (encoded by the AN4931 gene), which is conserved among 21 Aspergillus species but not in the unicellular yeasts S. cerevisiae or S. pombe.
29 proteins were identified associating both with Sem1 and the other GFP- tagged 19S RPs. The 29 interaction partners were grouped into five clusters: (1) protein degradation by the proteasome, (2) proteins involved in mitochondria-related activities, (3) proteins associated with ribosomes, (4) tubulin of the cytoskeleton and (5) conserved hypothetical protein HypoP1 (encoded by AN2234) with orthologs only in Aspergillus-related species and no conserved domain (Fig 5A lower panel). These associations point to a link between Sem1 as part of the regulatory particle and protein homeostasis, transport and mitochondria-related activities.
GFP pull-downs corroborated that A. nidulans Sem1 associates as part of the lid, with the complete 19S RP consisting of all 19S RP subunits (Fig 5A). Four identified in vivo interactions, which were also identified with the other analysed 19S RP subunits, support an important contribution of Sem1 to proteasome assembly, enabling the lid to associate with the base and the 19S RP to associate with the 20S CP. The Sem1-Rpn10 interaction might stabilize the connection between the proteasome lid and the base [20, 21]. The Sem1 interactions with base and lid associated chaperons, namely Nas6 (PSMD10 in human) and Hsp90, corroborates a Sem1 assembly function [22–25]. Sem1 also interacted with Ecm29, which stabilizes the 26S proteasome by tethering the 20S CP to the 19S RP [23].
Neither the ubiquitin receptor Rpn10 nor the tethering protein Ecm29 could be identified with any of the rpn-gfp strains when Sem1 was absent (Fig 5B). Consistently, in the absence of Sem1, Rpn10-GFP failed to pull any of the base and lid related proteasome subunits (Fig 5B). This supports an in vivo function of Sem1 through the ubiquitin receptor Rpn10 and Ecm29 for the interaction of base, lid and CP, which is essential for the assembly of intact 26S proteasomes.
The domain architecture of the lid of the proteasome is conserved in eukaryotic cells (Fig 6A). The Δsem1 mutant strain possesses a lid where the Rpn10 ubiquitin receptor is missing and the Rpn11 deubiquitinase protein levels are reduced, although rpn11 transcript levels are increased. This suggests that Sem1 supports the assembly of stable functional capped 26S proteasomes by a molecular mechanism, which includes the physical interaction between Sem1 and Rpn10 to assemble lid to base and that Sem1 protects Rpn11 protein integrity, which is required for the specific ATP/ubiquitin-dependent 26S proteasome activity.
Bimolecular fluorescence complementation studies (BiFC) were performed to determine whether Sem1 interaction with Rpn10 and Rpn11 can be monitored in fungal cells in vivo. Fungal strains expressing functional Sem1 fused through its C-terminus to the C-terminal half of YFP and C-terminal Rpn10 and Rpn11 fusions to the N-terminal half of YFP were examined (strains sem1-yfpc+rpn10-yfpn and sem1-yfpc+rpn11-yfpn, respectively). The fluorescence observed in strains containing fused Sem1-Rpn10 and fused Sem1-Rpn11 was significantly higher compared to the respective control strains (Fig 6B right panel). These cellular interactions of Sem1 could be due to an escorting of these proteins for lid assembly.
A homology model of the A. nidulans 19RP based on the cryoEM structure of human proteasome was generated (EMDB-4002, PDBs: 5L4K and 5L46 [26]) to examine the possibility of interactions between Sem1-Rpn10 and/or Sem1-Rpn11 upon assembly of the lid (Fig 6C). The modeled C-terminal fragment of Sem1 is bound in a structurally conserved cleft between the lid subunits Rpn3 and Rpn7. This structural conservation results in a very similar binding mode of Sem1 observed for yeast and human proteasomes (Fig 6A). In that binding mode, the extension of Sem1 towards the N-terminus reaches to the other side of the lid (opposite) due to an opening in the center of the lid. The missing (not modeled) N-terminal tail, comprising approximately 30 amino acids, could be responsible for direct interactions of Sem1 with both Rpn10 and Rpn11 forming the opposite surface of the lid (Fig 6C).
These data corroborate direct interactions between Sem1-Rpn10 as well as Sem1-Rpn11 in the fungal cell. Sem1 might escort Rpn10 and Rpn11 proteins to the lid and support the assembly and positioning of both proteins into a stable capped 26S proteasome through the assistance of its flexible N-terminal tail.
In contrast to wildtype, the 19S regulatory particle lacking Sem1 associated to proteins related to NADH or ATP production (Figs 5 and 7A). The dihydrolipoamide acetyltransferase Pdh1 (AN6708) is part of the pyruvate dehydrogenase complex for the oxidative decarboxylation of pyruvate to acetyl-CoA. Cytoplasmic Pcy1 (AN4462) converts pyruvate to oxaloacetate. Both enzyme products are used by the mitochondrial TCA cycle. The Sem1-interacting protein Rpn3 associated only in the absence of Sem1 with the ADP/ATP carrier Pet9 (AN4064) of the mitochondrial inner membrane, which exchanges cytosolic ADP for mitochondrial synthesized ATP. Rpn3, Rpn5 or Rpn10 interacted in the absence of Sem1 with the mitochondrial porin Por1 (AN4402). This outer membrane protein is required for maintenance of mitochondrial osmotic stability and membrane permeability. The 19S RP without Sem1 associates with the β-subunit of the mitochondrial processing protease Mpp (AN0747), which cleaves the N-terminal targeting sequence from mitochondrial-imported proteins. Subunit II of complex III (AN8373) and NADH-ubiquinone oxidoreductase, complex I (AN4288) are two components of the mitochondrial inner membrane electron transport chain which interact with RPs without Sem1. These findings suggest a specific physical interaction of RP subunits in the absence of Sem1 with mitochondria, which is not found when Sem1 is present.
The morphology of the mitochondria in Δsem1 and wildtype strains were compared to examine the impact of the association of lid subunits with the TCA cycle and respiratory chain related proteins, which were exclusively found in the absence of Sem1 (Fig 7A). The mitochondria of Δsem1 cells showed dots of disrupted filaments and differed significantly from the wildtype (Fig 7B). This phenotype suggests a defect in the dynamic equilibrium between mitochondria fusion and fission processes, that could be caused by the physical interactions of Δsem1 RPs with the mitochondrial machinery, which is suppressed in wildtype where Sem1 is present.
The total cellular NADH production was determined in Δsem1 and compared to wildtype (Fig 7C). Strains expressing Sem1 or Sem1-GFP showed similar high concentrations of NADH produced per gram mycelium after 20h of vegetative growth (63.0±15.2 and 53.3±2.4 ΔOD460nm/g mycelium, respectively). Deleting sem1 resulted in only 26% of NADH compared to wildtype (13.84±3.7 ΔOD460nm/g mycelium). The fragmented mitochondria observed in the mutant strain might be defective and less active compared to mitochondria in the wildtype strain.
A Δsem1 mutant strain showed fragmented mitochondria, produced less NADH and accumulated orange/red pigments (Fig 7B–7D). Mutant strains with a deficient COP9 signalosome or CAND-proteins controlling cellular cullin E3 ubiquitin ligase activities also accumulated red dyes and were linked to a misregulated secondary metabolism and an inappropriate oxidative stress response [27, 28]. Consistently, Δsem1 mutant strain was not able to grow on hydrogen peroxide and could hardly grow in the presence of menadione, whereas strains with functional Sem1 germinated and produced normal looking colonies (Fig 7D). The oxidative stress response was monitored at the transcriptional level of three superoxide dismutase encoding genes sodA, sodB and sodM and the catalase encoding gene catA (Figs 7E and S2B). Deletion of sem1 resulted in significant 2.5-fold up-regulation of transcripts for catalase A and 1.6-fold increase for superoxide dismutase B compared to wildtype (Figs 7E and S2B). These results underline a critical Sem1 function in the oxidative stress response. The mutant strain presumably tries to minimize the damaging effects of ROS caused by the damaged mitochondria, thereby inducing an antioxidative defence system.
Wildtype mitochondria were not damaged when exposed to moderate oxidative stress. Transcription levels of mitochondrial genes fzo1 for fusion or fis1 and dnm1 for fission were similar in fungal wildtype cells with an intact Sem1 in absence or presence of oxidative stress. The encountered oxidative stress was reflected in a response of increased expression of catalase encoding catA, sodB for a superoxide dismustase or the regulatory gene for oxidative stress nap1 corresponding to yeast yap1 (Figs 7F and S6).
It was analysed, whether Sem1 expression levels varied in response to oxidative stress. The inflicted oxidative stress resulted in a significantly increased transcription of sem1 and rpn11 genes (Fig 7F). This increasing level of sem1 and rpn11 transcripts represents a yet undescribed physiological cellular oxidative stress response, which might protect the cell from increased 20S proteasome levels. Increased transcription to produce more Sem1 and Rpn11 proteins might counteract the damaging interaction of aberrant 19S regulatory particles with the mitochondria, which is detrimental for vegetative cells and for fungal differentiation, requiring an oxidative stress signal as developmental trigger [29, 30].
The conserved Sem1 protein supports the assembly of multiple cellular complexes and represents the ninth bona fide subunit of the 19S regulatory particle of the 26S proteasome. A novel cellular function was detected, which connects the proteasome function and the cellular redox state at the molecular level. Oxidative imbalances in the multicellular ascomycete Aspergillus nidulans resulted not only in increased transcription of genes for detoxification enzymes such as catalases, but also in increased transcription of sem1 and rpn11 encoding the proteasomal deubiquitinating enzyme. Sufficient amounts of Sem1 and Rpn11 proteins are necessary during oxidative stress to provide higher amounts of correctly assembled 26S proteasomes. A lack of Sem1 resulted in increased oxidation-driven 20S proteasomes and instable capped proteasomes lacking the Rpn10 ubiquitin receptor, a functional Rpn11 deubiquitinating enzyme and the chaperone Ecm29 that connects the CP to the RP. Decreased amounts of Sem1 compromise multicellular fungal development, which requires internal reactive oxygen signals as triggers (Fig 8).
Sem1 is required for morphological integrity and functionality of the mitochondria, evident by the structural defects caused by the absence of Sem1. A physiological link between dysfunctional mitochondria due to mistransferred proteins and a proteostatic response had been described [31]. Lack of Sem1 results in a five-fold decrease in cellular NADH production compared to a wildtype strain. Consistently, Sem1 is required to allow fungal vegetative growth in the presence of oxidative-stress inducing compounds such as H2O2 or menadione. An appropriate oxidative stress response therefore includes in a wildtype fungus not only increased transcription of genes for detoxificating enzymes such as catA or sodB genes, but also increased transcripts for subunits of the 19S RP of the proteasome such as sem1 or rpn11. This corroborates that increased protein levels for these subunits are part of the cellular answer to stress. Mutant strains without Sem1 protein are hypersensitive towards stress, although they constitutively induce transcription of genes for detoxification enzymes. The Sem1-dependent stress response is linked to coordinated fungal secondary metabolism. This link is reminiscent to genetic studies with mutant strains defective in COP9 signalosome or CAND proteins, which control the activity of cullin E3 ubiquitin ligases. Impaired function of COP9 signalosome, CAND or Sem1 results in a redox imbalance and in accumulation of red orcinol derived secondary metabolites visible in the fungal colony as red dye [27, 28].
Multicellular development specifically requires protection against oxidative stress, as internal cellular stress signals are required for the progression of differentiation. This includes transient increase in reactive oxygen species (ROS) for developmental programmes in animals [32] or fungi [29, 30]. In fungi, increased ROS production interferes with hyphal fusion as one of the initial steps from vegetative hyphal growth to multicellular development [33]. In humans, increased ROS production is associated with mitochondria disorders, aging or neurodegenerative diseases, where unfolding of oxidized proteins promotes accumulation of protein aggregates [34, 35].
Sem1 of the unicellular yeast had been proposed to stabilize the interactions between the lid and the base [36]. In multicellular A. nidulans, Sem1 is not only required for correct 26S assembly, but represents a lid subunit which is mandatory for 26S proteasome composition, stability and specificity [36–38]. A. nidulans can form a lid without Sem1, as it had been described for S. cerevisiae, where a comparison between wildtype and Δsem1 lids by single particle cryo-EM analyses revealed significant structural differences with rearrangements of Rpn3 and Rpn7 in the Δsem1 lid [5, 39]. Sem1-dependent functions on assembly and stabilization of A. nidulans 26S proteasome were visualized by negative staining electron microscopy, where Δsem1-deficient proteasomes from mutant strains consist mostly of 20S proteasomes with only low abundance of 26S proteasome complexes. Sem1 fulfils its stabilization function by the recruitment of Rpn10, which is mandatory to stabilize the interaction between lid and base [19, 21]. In addition to the incorporation of Rpn10, Sem1 is required for the recruitment of Ecm29 as facilitator, which associates the 19S regulatory particle to the 20S core particle. Direct interactions between Sem1 and the base were not reported, but Rpn10 makes extensive connections with different lid and base subunits, namely the Rpn11/Rpn8 heterodimer, Rpn9 and Rpn12 or the base subunits Rpn1 and the Rpt4/Rpt5 heterodimer [40–43]. A. nidulans Rpn10 is a stable protein, which is unable to associate with any lid or base subunit without Sem1. The molecular function of Sem1 could be to escort Rpn10 to the lid and to stabilize it during the assembly. The homology model of the 19S regulatory particle of A. nidulans positions the C- terminal fragment of Sem1 between Rpn3 and Rpn7 and the last modeled N-terminal residue of Sem1 in the cleft formed by Rpn3. This cleft is located in a close proximity to an opening in the centre of the lid, which could be a structural feature allowing the N-terminal tail of Sem1 to pass through and interact with Rpn10 and Rpn11. Thereby, it is conceivable that this cleft accommodates the N-terminal tail of Sem1 and stabilizes the interaction of Rpn10 and Rpn11 within the lid. This is in agreement with recent cross-linking experiments between the N-terminal part of Sem1 and C-terminal part of Rpn11 in the fully assembled lid but not in LP2 (lid particle 2), the lid intermediated consisting all eight lid subunits except of Rpn12 [44]. A cross-linking between Sem1 and Rpn10 was not yet described. Modelling the corresponding regions of Sem1 from human or S. cerevisiae into the A. nidulans model leads to similar results. The in vivo BiFC study show direct physical interaction between Sem1 and Rpn10 and supports an escorting and assembly function of Sem1 for Rpn10 prior and during 26S proteasome assembly. Conclusively, the data suggests that the presence of Sem1 is a prerequisite to Rpn10 and is essential for accurate and efficient assembly of a stable capped 26S proteasome.
Oxidation drives 26S disassociation presumably by posttranslational modifications of α5, α6, α7 rings of the 20S CP [45, 46]. It was demonstrated that S-glutathionylation through redox-regulation promotes gate opening of the 20S CP, which is otherwise closed unless 19S RP is bound to it [47]. Cellular proteolysis of the oxidation-driven 20S proteasomes derived from the Δsem1 mutant strain impaired in redox regulation resulted in higher degradation rates compared to the wildytype 20S proteasomes. As ATP/ubiquitin-independent degradation requires 20S proteasome complexes but no ATP and no polyubiquitinated proteins [48], this presumably allows the Δsem1 strain to maintain efficient degradation. In the mutant strain the 20S proteasomes are kinetically favoured regardless of the absence of 19S RP or ATP. The Δsem1 mutant strain produces less NADH, which potentially reduces oxidative respiration and ATP production.
Mutant analysis revealed that the presence of the sem1 gene correlated with increased ubiquitin-conjugates and reduced ATP/ubiquitin-independent degradation in comparison to the mutant strain lacking sem1. A ubiquitin receptor function has been described for the human counterpart Dss1 with a ubiquitin binding site overlapping the Rpn3-Rpn7 binding sites [49]. This suggests that Sem1 can associate with the proteasome leaving the two acidic patches available for ubiquitinated substrates and/or can even dynamically associate with the proteasome to escort ubiquitinated substrates to the proximity of proteasomes [2]. The observed decreased levels of ubiquitinated substrates in the A. nidulans mutant strain lacking sem1 could be related to CRLs and the attachment of ubiquitin to substrates. In yeast, the deletion of Sem1 resulted in an accumulation of ubiquitinated substrates presumably due to non-functional proteasomes [3]. Decreased cellular levels of ubiquitin-conjugates were also observed in mammalian epithelial cells exposed to H2O2, as a consequence of the oxidation of cysteine residues in the active sites of E1-E3 ubiquitin-conjugating enzymes [50–52]. In A. nidulans, oxidative stress not only reduces the amount of the E3 Cul1 scaffold subunit, but also influenced the transcription of culC for Cul3, culD for Cul4 and csnE for the COP9 signalosome subunit Csn5. This suggests that ubiquitinated proteins are regulated in A. nidulans in response to oxidative stress where ATP/ubiquitin-independent degradation takes place.
Sem1 represents a novel molecular link between proteasome assembly and specificity, mitochondrial integrity and cellular development. The viable Δsem1 mutant strain from A. nidulans is a valuable tool to investigate ATP/ubiquitin-independent proteolysis to elucidate the cross talk between cullins, COP9 signalosome and proteasome in response to oxidative stress. Understanding Sem1 function during mitochondrial stress will provide new insights for our understanding of mitochondrial-associated pathogenesis. Increased Sem1 activity might delay mitochondria dysfunction and can be used for further therapies. Elucidating the mechanisms by which Sem1 affects and regulates oxidative stress is beneficial in the efforts to understand and treat age-related human diseases and explore potential therapies.
The study did not involve human participants, specimens or tissue samples, or vertebrate animals, embryos or tissues.
A. nidulans strains used in this study are listed in Table A (S1 File-supporting information). The A. nidulans sem1 gene corresponds to semA in the fungal nomenclature. Spore concentration was determined by Z2 Coulter particle count and size analyser (Beckmann counter). Vegetative growth was performed in flasks with indentations containing supplemented liquid media and 5x105 spores/ml at 37°C for 20 hours. For asexual and sexual development 10 μl of 10000 spores/ml were spotted on supplemented MM and incubated at 37°C. For the time point experiment, asexual spores were counted with Thoma chamber. Incubating the plates in constant white light triggered asexual sporulation, whereas sexual fruiting body formation was induced by oxygen limitation (the plates were sealed) and darkness. Asexual spores were counted using Thoma chamber. For sexual growth, 100 μl of 1x106 spores/ml were spared on supplemented MM and incubated at 37°C and incubated in the dark with limited oxygen for 7 and 10 days.
Samples were bound to a glow discharged carbon foil covered grid. After staining with 1% uranyl acetate, the samples were evaluated with a CM 120 transmission electron microscope (FEI, Eindhoven, and The Netherlands). Images were taken with a TemCam F416 CMOS camera (TVIPS, Gauting, Germany).
The total NADH production was measured as described earlier [53]. 8x 5ml mycelia of each strain were used for the measurements. Activity was measured at 460nm and was normalized to DCW (g mycelium) after drying for 3 days at 60°C.
Proteasomes were purified from 8 g grained mycelium using rapid 26S proteasome purification kit from UBPBio. Concentrations were determined by Nanodrop and activity assays were performed according to manufacturer recommendations. The activity was measured in an assay buffer containing 12.5 mM Tris HCl pH = 7.5+10 mM KCl+1.25 mM MgCl2+0.125 mM ATP+0.25 mM DTT+0.0125 mg/ml BSA. The release of AMC from 100 μM fluorgenic peptide, Suc-LLVY-AMC, was measured over 30–60 min and background fluorescence was subtracted from all measurements.
t-test was used to determine significance of the results (http://www.quantitativeskills.com/sisa/statistics/oneway.htm). The intensity of western blot bands was determined with ImageJ v1.48 analysis software. Mean intensities from biological replicates (n) were relative to glyceraldehyde-3-phosphate dehydrogenase (GAPDH) serving as loading control and were normalized to wildtype (%). Expression levels assayed by RT-PCR are shown as relative expression compared to wilidtype and represents mean value and standard error of the indicated independent experiments (n).
|
10.1371/journal.pbio.2004752 | Agent-specific learning signals for self–other distinction during mentalising | Humans have a remarkable ability to simulate the minds of others. How the brain distinguishes between mental states attributed to self and mental states attributed to someone else is unknown. Here, we investigated how fundamental neural learning signals are selectively attributed to different agents. Specifically, we asked whether learning signals are encoded in agent-specific neural patterns or whether a self–other distinction depends on encoding agent identity separately from this learning signal. To examine this, we tasked subjects to learn continuously 2 models of the same environment, such that one was selectively attributed to self and the other was selectively attributed to another agent. Combining computational modelling with magnetoencephalography (MEG) enabled us to track neural representations of prediction errors (PEs) and beliefs attributed to self, and of simulated PEs and beliefs attributed to another agent. We found that the representational pattern of a PE reliably predicts the identity of the agent to whom the signal is attributed, consistent with a neural self–other distinction implemented via agent-specific learning signals. Strikingly, subjects exhibiting a weaker neural self–other distinction also had a reduced behavioural capacity for self–other distinction and displayed more marked subclinical psychopathological traits. The neural self–other distinction was also modulated by social context, evidenced in a significantly reduced decoding of agent identity in a nonsocial control task. Thus, we show that self–other distinction is realised through an encoding of agent identity intrinsic to fundamental learning signals. The observation that the fidelity of this encoding predicts psychopathological traits is of interest as a potential neurocomputational psychiatric biomarker.
| In order for people to have meaningful social interactions, they need to infer each other’s beliefs. Converging evidence from humans and nonhuman primates suggests that a person’s brain can represent a second person’s beliefs by simulating that second person’s brain activity. However, it is not known how the outputs of those simulations are identified as ‘yours and not mine’. This ability to distinguish self from other is required for social cognition, and it may be impaired in mental health disorders with social cognitive deficits. We investigated self–other distinction in healthy adults learning about an environment both from their own point of view and the point of view of another person. We used computationally identified learning variables and then detected how these variables are represented by measuring magnetic fields in the brain. We found that the human brain can distinguish self from other by expressing these signals in dissociable activity patterns. Subjects who showed the largest difference between self signals and other signals were better at distinguishing self from other in the task and also showed fewer traits of mental health disorders.
| Social interactions are underpinned by an ability to infer the mental states of self and others, referred to as mentalising [1]. The discovery of mirror neurons in the macaque premotor cortex [2] introduced the notion that in mentalising, the primate brain might directly simulate another agent’s cognitive process. More recently, functional magnetic resonance imaging (fMRI) studies [3–8] and intracranial recordings [9] in humans, as well as single-cell recordings in monkeys [10], have shown that when a subject observes another agent interact with its environment, the subject’s brain encodes not only the other agent’s motor activity but also their reward prediction errors (RPEs). In other words, subjects appear to simulate the reinforcement learning processes of other agents.
These simulated learning signals localise to specific cortical regions, such as the anterior cingulate gyrus [9–11]. A functional segregation of learning signals can allow the brain to encode information about whether learning is arising out of the individual’s own behavioural interactions with the environment or whether learning is taking place vicariously through observing the behaviour of another agent. In a similar vein, it has been suggested that the medial prefrontal cortex (mPFC) supports a functional axis that encodes whether behaviour is executed or imagined [12, 13].
For simulation to be useful in social interactions, the brain must discriminate signals attributed to self from simulated signals attributed to other agents [14–17]. An impairment in this self–other distinction is a defining feature of autism spectrum disorder [18–21]. Similar impairments have also been reported in conditions such as schizophrenia [22, 23], psychopathy [17], and borderline personality disorder [24]. An aberrant self–other distinction might also underpin the social dysfunction seen in psychopathologies, including depression [25, 26] and addiction [27–29].
A prefrontal coding scheme that discriminates between instrumental and observational learning, or executed and imagined behaviour, could provide a useful heuristic for a self–other distinction but would be insufficient for discriminating amongst signals attributed to different agents as a general-purpose computation. For instance, the false belief task [30], a standard test of mentalising ability, requires that subjects make inferences about an environment and then selectively attribute one belief-state to self and a different belief-state to another agent for whom the environment is only partially observable. These belief-states are not informed by the behaviour of the subject or the other agent but arise through passively observing the environment. In this case, neural coding schemes that discriminate between executed behaviour and observed or imagined behaviour cannot facilitate a self–other distinction, and a more fundamental computation for selectively attributing signals to different agents is required.
Thus, an open question for simulation theory is how self–other distinction is achieved [16]. If inferring another agent’s mental state requires the brain to simulate that agent’s computations, how are the outputs of those computations identified as ‘belonging to other’? One possibility is that variables for learning and decision-making are encoded in distinct neural activity patterns, depending on the agent to whom these signals are attributed. Such architecture would entail an encoding of agent identity intrinsic to representations of these low-level signals. A second possibility is that a learning signal is always encoded in an agent-independent pattern. In this case, the learning signal and the identity of the agent to whom the signal is attributed would need to be encoded in 2 separate activity patterns.
Here, we test whether learning signals are encoded in agent-specific activity patterns, and thus whether self–other distinction requires agent identity to be encoded separately from a low-level learning signal. We used a novel paradigm inspired by false belief tasks, in which subjects learned about a fluctuating state in the environment. In so doing, they were also required to intermittently switch their frame of reference between self and other. The 2 agents (self and other) received different information such that their belief trajectories were uncorrelated, enabling us to measure self-attributed and other-attributed learning signals independently. Unlike previous paradigms eliciting simulated signals, subjects did not observe the agent’s behaviour, and there was no reinforcement of learning by either reward or punishment. Learning for self and learning for other thus recruited the same input channels and were identically salient and identically motivating. The task design rules out a potential confound of simulated reward learning [3–8] wherein self-attributed reward-related decision signals (such as RPEs) pertain to rewards expected to be received by the participant, while other-attributed reward-related decision signals do not. We measured the neural encoding of learning signals using magnetoencephalography (MEG) in order to measure whether the representations of these signals are agent-specific and also how agent-specificity evolves over the time course of a single trial. Of note, our task design required sparse probe trials and therefore a larger total number of trials than would be possible to acquire in a single fMRI session.
We present data from 38 healthy adults (see Materials and methods for participant details). During MEG scanning, they observed a sequence of samples from a Bernoulli distribution, with a drifting Bernoulli parameter P. This is the probability, on each trial, of seeing 1 of 2 possible outcomes. Another participant, who sat outside the scanner in a different room, played the exact same game (see Materials and methods). This other subject was able to observe some of the samples seen by the scanned participant (‘shared’ trials) but not all of them (‘privileged’ trials). Additionally, the nonscanned participant was occasionally presented with misleading samples (‘decoy’ trials). Therefore, the nonscanned participant sampled evidence that induced a false belief about P (Pfb). These 3 types of trial (‘privileged’, ‘shared’, and ‘decoy’) were balanced in frequency and distributed evenly throughout the task in a pseudorandom order.
In a version of the game we refer to as the social version (SV), ‘privileged’ and ‘decoy’ trials were signalled to the scanned participant, who thus had access to information about both P and Pfb. On ‘self’ probe trials, the scanned participant was required to report their estimate of P, by positioning an arrow along a virtual continuous scale that ranged from a probability of 0 (certain to see one outcome) to 1 (certain to see the other outcome). On ‘other’ probe trials, the scanned participant had to put themselves in the shoes of the nonscanned participant and report their estimate of Pfb. Crucially, the information that the scanned subject used to compute Pfb was sampled at the same rate as the information used to compute P. We refer to the subject’s belief about P as B, and we refer to their belief about Pfb as Bfb. The structure of sampling trials and probe trials is outlined in Fig 1A.
All subjects also played a nonsocial version (NSV) of this game, which did not involve another participant. Here, the scanned participant had to keep track of the belief-state of a fictional ‘computer’ that received limited and misleading information and stored a false estimate of P (Pfb). On ‘other’ probe trials in the NSV, the scanned participant was asked to imagine themselves in a counterfactual situation, wherein they acted using the false information provided by the computer. Thus, in the SV, participants switched their frame of reference between self and other, whilst in the NSV, participants switched their frame of reference between self and a counterfactual self. The only structural differences between the SV and NSV pertained to the cover stories, the images used for stimuli, and the wording of the ‘other’ probe trials (see Materials and methods). Note that for both the SV and NSV, we pregenerated trial sequences that minimised correlation between the belief trajectories of the 2 agents (Fig 1D, Fig 1E; also see Materials and methods).
For each subject, we assessed behavioural accuracy independently for the 2 versions of the task (SV and NSV) as well as on the 2 types of probe trial (‘self’ and ‘other’). This resulted in 4 conditions overall, for which accuracy was defined relative to chance performance (see Materials and methods). Where an accuracy of 0 is equivalent to chance-level performance, mean accuracies (with SDs) for the SV were 0.11 (0.04) in ‘self’ probe trials and 0.11 (0.04) in ‘other’ probe trials. For the NSV, these were 0.11 (0.04) for ‘self’ probe trials and 0.10 (0.05) for ‘other’ probe trials (see S1 Fig). The group performed significantly better than chance in all 4 conditions as assessed with 4 separate one-sample t tests on the mean accuracies per subject (P < 0.0001 in all 4 conditions). There were no differences in accuracy between the 2 probe trial types, or between the 2 versions of the game (ANOVA: main effect of probe trial type: F[1, 148] = 1.54, P = 0.22; main effect of game version: F[1, 148] = 0, P = 0.96; interaction: F[1, 148] = 0.06, P = 0.81). Thus, all 4 conditions were similar in difficulty.
We fitted 21 models to the probe trial behaviour of each subject, separately for the SV and NSV. There were 3 principal groups of model (see Table 1 for a summary and Materials and methods for details). Group A models assumed that subjects’ beliefs were constructed from an average over recently sampled information. Group B models were based on an assumption of Rescorla-Wagner (RW) updating [31], in which the models derive prediction errors (PEs) on each trial from the difference between the actual and expected outcomes. PEs updated the beliefs of self, while PEo was a simulation of the other agent’s PE, for updating the beliefs of other in the SV or ‘counterfactual self’ in the NSV. A subset of group B models also included ‘leak’ parameters that allowed PE signals to erroneously update the wrong agent’s belief, thus capturing an inability to maintain separate belief updates for the 2 agents. All group B models also assumed that the PEs had a value of 0 on ‘decoy’ trials whilst PEo had a value of 0 on ‘privileged’ trials. Group C models were like group B models except that they did not make this assumption; instead, they allowed PEs and PEo to update the beliefs of self and other, respectively, in all 3 trial types.
We compared models separately for the SV and NSV using the Bayesian Information Criterion (BIC). For both the SV and NSV, model 8 had the lowest mean BIC value (Fig 2A). This model incorporated 2 separate PE signals and included 4 free parameters: a learning rate (α) regulated the update of the beliefs of the 2 agents, a memory decay parameter (δ) controlled the rate of ‘forgetting’ for the beliefs of the 2 agents, and 2 temperature parameters (τs, τo) governed choice stochasticity on ‘self’ probe trials and ‘other’ probe trials, respectively (see S3 Fig for parameter recovery). This model generated synthetic choice data qualitatively similar to subjects’ real choice data (Fig 2B). Noting large intersubject variability in BIC values, we also employed a random-effects Bayesian model selection [32] to compare the winning model with the second best model (S4 Fig) and found, for both the SV and NSV, an exceedance probability in excess of 0.99. This is the probability that the winning model better explains a randomly chosen subject’s data.
We also assessed the correlation between parameter estimates fitted to the SV and parameter estimates fitted to the NSV (Fig 2C). For each model, we obtained a correlation coefficient for each parameter and then took the mean of those coefficients as a summary statistic for the between-game consistency of the model. Because parameter values were not normally distributed, we computed the nonparametric Spearman’s rank correlation coefficient. We found that model 8 also had the highest between-game consistency. Thus, this model captured consistent dispositions in subjects’ choice behaviour across the 2 games.
After fitting models to the behavioural data, we then had parameter estimates for each model and each subject. We used model 8 along with each subject’s parameters for this model to generate trial-wise estimates of latent PEs and beliefs, which we then used in subsequent analyses on the MEG data. Note that PEs and belief values generated by other models were very similar, and consequently, our findings were not sensitive to the selection of a particular model.
We next asked whether |PEs| and |PEo| were encoded in the MEG signal, recorded during task performance, using a mass-univariate analysis (Fig 3). For each subject, we fit 2 separate linear regression models at each sensor and each peristimulus time point. We obtained trial-wise estimates of |PEs| and |PEo| using the winning model’s estimated free parameters fitted to the choice data. The first model regressed |PEs| against the event-related field (ERF) on ‘privileged’ and ‘shared’ sampling trials (i.e., trials in which PEs was nonzero). The second regression model regressed |PEo| against ERFs on ‘decoy’ and ‘shared’ trials (i.e., trials in which PEo was nonzero). This resulted in 4 statistical maps over sensors and time, 2 for the SV and 2 for the NSV.
We converted each of these maps into a 3D image (2 spatial dimensions and 1 temporal dimension) of baseline-corrected effect sizes (see Materials and methods). To make group-level inferences, we conducted a one-sided Wilcoxon signed-rank test at each pixel to determine whether the group median was significantly greater than 0. We thresholded the resulting 3D image with a cluster-forming threshold (P < 0.001) and identified clusters of contiguous suprathreshold pixels, which could extend through space and time.
We determined whether any clusters were significantly larger than chance with a nonparametric permutation test to generate null distributions of cluster extent. In each of the 4 regression models, we found clusters significantly larger than chance at a 0.05 family-wise error (FWE) level. In the SV, the clusters extended through parietal and occipital sensors, whilst in the NSV, the clusters extended through frontal and parietal sensors. For the SV PEs, the largest cluster extended from 330 ms to 390 ms and comprised 2,628 pixels (threshold 612). For the SV PEo, the largest cluster extended from 340 ms to 420 ms and comprised 2,032 pixels (threshold 624). For the NSV PEs, the largest cluster extended from 370 ms to 440 ms and comprised 1,621 pixels (threshold 554). For the NSV PEo, the largest cluster extended from 310 ms to 370 ms and comprised 847 pixels (threshold 569). Despite finding significant clusters at the group level, we also noted large intersubject differences in these spatiotemporal patterns (e.g., S5 Fig).
We wanted to test whether we could distinguish a neural pattern encoding a PEs from a pattern encoding a PEo and thus determine whether a self–other distinction can be achieved on the basis of these signals. A typical way to identify the neural pattern encoding a PE is to regress the magnitude of the PE (derived from our learning model) against the brain activity, across trials. This would yield a single beta estimate at each sensor, capturing the slope of the relationship between PE and brain activity at that sensor. However, in order to use powerful multivariable methods like support vector machine (SVM) classification to look for differences in the spatial patterns of PEs and PEo, it was necessary to obtain multiple samples of each pattern. One way to achieve this is to divide the data into multiple partitions (without replacement) and repeat the analysis in each partition to obtain multiple independent samples of the spatial pattern for each type of PE. This is the approach we opted for, using the smallest possible partitions: pairs of trials (Fig 4A).
To maximise power without introducing bias, we randomly partitioned trials into pairs under the constraint that each pair contained 1 trial above the median |PE| and 1 trial below the median |PE|. Thus, the difference in brain activity between the 2 trials within a pair corresponded to a representation of |PE|. We performed this random partitioning independently for PEs and PEo. This resulted in 2 sets of difference images, corresponding to neural representations of |PEs| and |PEo|. Finally, we could then apply multivariable methods to classify whether each difference image was a representation of |PEs| or |PEo|.
It should be noted that this method differs slightly from typical pattern-based neuroimaging analyses described in, for example, [33]. Usually, such an analysis looks for a neural representation of some variable. This is achieved by training a classification or regression model to distinguish patterns of neural activity corresponding to different values of that variable. Above-chance accuracy of the model indicates that the brain activity contains information about the variable. However, in our case, we were interested in a difference in the representation of a variable between 2 conditions. Because the representation itself is defined by a difference in neural activity between a large PE and a small PE, we were looking for a difference of differences. Thus, it was necessary to train classifiers on patterns of subtracted activity rather than activity patterns from individual trials.
We started with N trials in total. First, we partitioned all ‘privileged’ and ‘shared’ trials (2N/3 trials) by median split on |PEs|. We then randomly sampled 2 trials, one from either side of this partition, and subtracted the ERF on the low |PEs| trial from the ERF on the high |PEs| trial, at every sensor and time point. For ease of reference, we call this contrast image a ‘pseudotrial’. We continued randomly sampling pairs of trials without replacement to obtain a total of N/3 pseudotrials. Each of these pseudotrials describes the difference in activity between a trial with a high |PEs| and a trial with a low |PEs|. The brain activity in the difference image thus constituted a representation of |PEs|.
Second, we partitioned all ‘decoy’ and ‘shared’ trials (2N/3 trials) by median split on |PEo|. We carried out the same procedure as for PEs, resulting in a second set of N/3 pseudotrials, each of which constituted a representation of |PEo|. At each time point, we trained a classifier to distinguish PEs pseudotrials from PEo pseudotrials.
We tested classifiers in crossvalidation, yielding a time course of classification accuracies (CAs). The absolute difference in CA underlying reliable effects was, in some cases, as small as 1%. In observing this, we note that effect sizes cannot be inferred from absolute CAs [34–36]. Therefore, to make statistical inferences, we adopted a permutation-based method to determine whether any CA was significantly better than chance. This procedure has been recommended for making inferences on ‘information-based’ neural measures such as CA [35].
To derive a threshold for statistical significance, we repeated the whole pseudotrial analysis many times, each time using data generated from a permuted trial sequence. For every permutation, we took the maximal CA (or maximal difference in CA between the SV and NSV) across all time points. We thus generated a null distribution of maximal CAs (or maximal CA differences). The 95th percentile of the distribution was taken as our threshold for statistical significance. This procedure allowed us to make statistical inferences without making assumptions about how CAs (or CA differences) are distributed, whilst also correcting for multiple comparisons across time points, at a 0.05 FWE level.
We found that self and other could be classified significantly above chance level from the spatial patterns of activity that represented |PEs| and |PEo| approximately 300 ms after stimulus onset (Fig 4B). However, CA did not exceed chance level when we conducted this same analysis on the NSV data. Moreover, at approximately 300 ms, there was a significant difference between CA in the SV and CA in the NSV. Thus, distinct spatial activity patterns for |PEs| and |PEo| were evident in the SV but not in the NSV. This implies information about self and other is intrinsic to the representations of low-level learning signals, whilst information about self and counterfactual self is not.
To test the robustness of this finding, we performed 2 additional variants of the analysis, by constructing pseudotrials from subjects’ trial-wise ‘signed beliefs’ (B and Bfb) and ‘unsigned beliefs’ (|B − 0.5| and |Bfb − 0.5|). The former are the subject’s trial-wise estimates of the underlying Bernoulli parameter from the perspective of each agent. The latter are the absolute distances of these estimates from 0.5, which represents an equal probability of either outcome. The ‘unsigned belief’ is thus a measure of confidence in what the next outcome will be. It should be noted that here, we can use all N trials to generate pseudotrials. Thus, we end up with N/2 pseudotrials for each class and N pseudotrials in total. We found that classifiers trained on pseudotrial data, generated from either of these latent variables, could predict agent identity (self or other) significantly above chance in the SV. However, in the NSV, the classifiers could only predict agent identity (self or counterfactual self) for pseudotrials generated from ‘signed beliefs’, and in this instance, the signal was weaker and occurred later in time than was the case for the SV (Fig 4B). Furthermore, we found that CAs for the SV were significantly larger than CAs for the NSV at multiple time points for both of these pseudotrials. Finally, for comparison, in a separate analysis classifying between the visual stimuli, we obtained similar decoding accuracies in the SV and NSV (S6 Fig).
An important question is whether the neural distinction in learning signals is related to a behavioural measure of self–other distinction. A subset of our behavioural models (models 11 to 19) included a ‘leak’ (λ) parameter that governed the extent to which PEs was erroneously used to update Bfb and/or PEo was erroneously used to update B, thus indexing an inability to discriminate between 2 different agents’ learning processes. We estimated λ values by selecting the best-fitting λ-containing model for each individual subject. If the best model contained 2 λ parameters, we took the mean of the 2 values. We derived 2 estimates of λ for each subject, one for the SV and one for the NSV.
We then computed, for each subject, a metric describing overall neural self–other distinction. In order to do this, we took the maximal CA from each of the time courses from the 3 types of pseudotrial (Fig 4B) and summed these 3 numbers. This provided one number for neural agent decoding in the SV and another number for neural agent decoding in the NSV.
Because λ in the SV and λ in the NSV were strongly correlated across subjects, we examined the difference between the SV and NSV. Due to the non-normally distributed parameter estimates, we computed a nonparametric Spearman’s rank correlation coefficient. We found a strong negative correlation (Fig 5) between the neural decoding contrast (SV − NSV) and the estimated λ contrast (SV − NSV): Spearman’s rho: −0.43, P < 0.01. We also tested the accuracy of a linear regression model that used neural decoding contrasts to predict the estimated λ contrasts. Here, we used crossvalidation with random subsampling (train on half, test on half) and recorded the correlation between predicted and observed values on every fold. The median Pearson coefficient across 10,000 folds was 0.31, which was significantly greater than chance as determined by a nonparametric permutation test (P = 0.039).
We also found a significant positive correlation (Spearman’s rho: 0.32, P < 0.05) when, instead of using raw parameter estimates, we used the relative model evidence (BIC) of a subject’s best lambda-containing model and best nonlambda model. In other words, subjects whose SV behaviour is better explained by a model with lambda parameters than their NSV behaviour show less neural self–other distinction in the SV than in the NSV.
These findings show that, in subjects for whom agent identity could be more accurately decoded in the SV than in the NSV, there was also more behavioural evidence for segregating the beliefs of self and other in the SV than in the NSV. This suggests that the distinctiveness of neural patterns encoding learning signals attributed to 2 different agents is predictive of how well a subject behaviourally succeeds in distinguishing between these 2 agents’ beliefs.
Finally, we asked whether neural agent decoding relates to intersubject differences in subclinical personality traits. All subjects filled out 5 questionnaires of interest: Beck Depression Inventory (BDI), Empathy Quotient (EQ), Interpersonal Reactivity Index (IRI), Inventory of Callous-Unemotional traits (ICU), and the Community Assessment of Psychic Experience (CAPE). These questionnaires were specifically chosen to assess the presence of psychopathological traits previously proposed to relate to a dysfunctional self–other distinction or more general social cognitive deficits [17, 21, 22, 25]. These questionnaires assessed autistic (EQ), schizotypal (CAPE), antisocial (ICU), and depressive (BDI) traits as well as general capacities for empathy and sympathy (EQ, IRI). We also obtained measures of response bias using an additional questionnaire, the Balanced Inventory of Desirable Responding (BIDR) [37]. None of the subjects were considered to have an unacceptably high response bias (see Materials and methods).
We performed dimensionality reduction on age and gender-controlled personality questionnaire data (see Materials and methods) using a principal components analysis (PCA), having included all subscales of the 5 questionnaires of interest, giving 9 dimensions in total (Fig 6A). The first principal component (PC1) explained 32% of the variance in the questionnaire data and loaded negatively with both subscales of the CAPE questionnaire (schizophrenia), BDI (depression), ICU (antisocial behaviour), and 1 subscale of the IRI (personal distress in social situations); it also loaded positively with EQ and other subscales of the IRI (Fig 6A). Thus, the PC1 negatively captured psychopathological features in our personality data in a nonspecific manner. We projected the personality data into the space of this principal component to obtain a score for each subject.
First, we correlated the neural self–other distinction metric, as described in the previous section, with the PC1 scores. When using the contrast of (SV − NSV), this yielded a significant correlation: R = 0.39, P = 0.017 (Fig 6B). We also tested the accuracy of a linear regression model that used neural decoding contrasts to predict the PC1 scores, using the same method as described for Fig 5. The median Pearson coefficient across 10,000 folds was 0.34, which was significantly greater than chance as determined by a nonparametric permutation test (P = 0.04). Therefore, subjects for whom we obtained higher CAs in the SV than in the NSV scored higher on the PC1. In other words, subjects for whom it was easier to neurally decode self from other than to decode self from counterfactual self scored higher on a nonspecific anti-psychopathological component. When looking at the SV and NSV neural self–other distinction metrics separately, we found a significant positive correlation for the SV (R = 0.43, P < 0.01) but no significant correlation for the NSV (R = 0.01, P = 0.94).
We then investigated the temporal evolution of this relationship for each of the 3 types of pseudotrial. At each peristimulus time sample, we correlated the subjects’ PC1 scores with (CA SV − CA NSV) to generate a time course of Pearson coefficients (Fig 6C). Using permutation-based thresholding, to correct for multiple comparisons in time and between the 3 types of pseudotrial, we found a significant positive correlation (P < 0.05 FWE) approximately 110 ms after stimulus onset (Fig 6C) when using the ‘signed belief’ pseudotrials. This falls within the window of significant self–other distinction in signed beliefs (100–340 ms) as shown in Fig 4B.
We show that a representation of a learning signal (PE or belief) is encoded with a different neural spatial pattern when the signal is attributed to self as compared to when it is attributed to another agent. Intersubject variability in this difference correlated between subjects with a behavioural measure of self–other distinction and with subclinical psychopathological traits. This suggests that self–other distinction is realised by an encoding of agent identity that is intrinsic to low-level learning signals, and the fidelity with which this occurs is an important dimension of variation between individuals.
In our experiment, subjects had to solve 2 simultaneous computational problems. The first problem was predicting what the next outcome would be. The second problem was identifying whether this belief-state about the next outcome should be attributed to one agent or another, a computation that requires a self–other distinction. We found a spatial segregation between self-attributed and other-attributed learning signals. This means that the neural representations of beliefs and PEs in this task also contained information about the agent to whom these signals belong, and consequently, the neural resources that compute the next outcome also inevitably contribute to computing a self–other distinction. It is of interest, therefore, that the degree of spatial segregation was correlated with a behavioural measure of self–other distinction derived from our learning models.
Previous work has shown that neuronal populations in the macaque anterior cingulate cortex preferentially encode simulated RPEs [10] and the future decisions [38] of another monkey. Conversely, human fMRI data has identified common activations in the mPFC that represent RPEs [6] or subjective preferences [12] for both self and other in an agent-independent manner. Likewise, mirror neurons recorded from the macaque premotor cortex are also agent independent [2]. A self–other distinction in the affective domain has been reported in terms of dissociable networks for experienced versus vicarious pain [39, 40], though other reports suggest that these are both subserved by the same structures [41, 42].
The above accounts are conflicting with respect to whether self- and other- attributed signals share common or distinct neural activations. One possible reason for this is that previous studies eliciting simulated signals were not indexing a self–other contrast per se but rather a contrast of executed behaviour versus observed behaviour, in which the subject receives feedback from the observee’s behaviour. In these cases, learning or decision variables are discriminated not by virtue of the agent to whom they are attributed but instead by virtue of distinct input modalities and cognitive demands required for instrumental learning and observational learning, respectively. Because these factors are heavily task specific—for instance, dependent on the way in which the subject accrues information about the other agent’s behaviour—it is unsurprising that observational learning paradigms have produced inconsistent accounts of the neural encoding of agent identity.
In the present study, we devised a self–other contrast per se by allowing subjects to observe what the other agent observed but not the other agent’s behaviour. By requiring subjects to switch between attributing a signal to self and attributing a signal to other, with fixed sensory input modalities and cognitive demands, we could show that learning signals do contain information about the identity of the agent to whom they are attributed. Consistent with this are our findings that the neural self–other distinction is modulated by individual personality traits, as well as by the precise contextual relationship between the agents in question, as assessed with the SV and NSV of the paradigm. Our findings do not rule out a possibility that the brain uses additional mechanisms to distinguish self from other, for instance, with an explicit encoding of agent identity that is separate from low-level learning signals. However, our results support the theory that agent-specific learning signals are sufficient for the brain to achieve a self–other distinction during mentalising.
Our results support the idea that the brain updates simulated beliefs of another agent using PEs calculated within the frame of reference of that agent. Previous work on other-referenced processing has shown that humans and other primates simulate another agent’s experience of unexpected reward [3, 6, 10]. Our work extends these findings to the domain of updating beliefs about non–reward-related quantities. A simulated sensory PE such as what we observed, combined with information about a preceding state, could also be used to simulate how another agent learns transition models of complex environments with multiple states, a requisite for goal-directed behaviour [43–47].
Differences between neural representations of signals attributed to self and other, which we attribute to agent identity, might relate to some other features of the task frame [48–50]. However, if this were true, we would expect representations to be similarly distinct in the NSV as in the SV because both versions shared a shift in task frame (between self and other in the SV and between self and a counterfactual self in the NSV). The finding that representations were significantly less distinct in the NSV supports our conclusion that features intrinsic to agent identity are fundamental to the distinct representations observed in the SV. Although representations were overall less distinct in the NSV, there was nevertheless a detectable difference in ‘signed belief’ representation between self and counterfactual self, occurring approximately 550 ms after stimulus onset. It will be important in future work to clarify why this long latency separation occurs in a nonsocial context.
The differences we found between the SV and NSV do not necessarily mean that subjects were not engaging in social computations in the NSV. Despite evidence for a so-called Theory of Mind network [51–53] recruited during mentalising, there is evidence that the brain might also rely on domain general computations for social cognition [54, 55]. It has been suggested that ‘social’ computations like mental state inference and ‘nonsocial’ computations, such as mental time travel and metacognition, are underpinned by the same general capacity for metarepresentation [25, 56–58]. If self–other distinction is a special case of a domain-general computation, future work should seek to understand why this computation is executed differently in social and nonsocial contexts.
We observed substantial intersubject heterogeneity in the spatiotemporal pattern of PE signals in our task. Although anatomical inferences are limited for data acquired in sensor space [59, 60], the heterogeneity would suggest a diversity of cortical regions encoding PEs. fMRI studies, employing both learning and nonlearning paradigms, have reported unsigned sensory PE activity or activity corresponding to unexpected neutral stimuli in a range of cortical and subcortical regions, including the anterior insula and inferior frontal gyrus [61], primary sensory cortices [62–64], superior temporal sulcus [65], hippocampus [66], cerebellum [67], striatum [64, 68, 69] and midbrain [70].
With regards to timing, previous studies in a nonsocial context using electroencephalography (EEG) [71–73] and MEG [74] have identified signed PE signals 200 to 350 ms after stimulus onset. However, unsigned PE signals are less well characterised and appear to be encoded across a much broader time window, ranging from 145 ms to 640 ms after stimulus onset [71]. Finally, previous false belief experiments using EEG [75, 76] and MEG [77] have reported latencies ranging from 100 ms to 800 ms after stimulus onset, at which time signals have differentiated true beliefs from false beliefs. Here, we show that in a social setting, these PE signals are agent specific as soon as they are detectable, approximately 300 ms after stimulus onset. Conversely, in the NSV, these PE signals were not agent specific at any time point.
The contrast in agent decoding accuracy between the SV and NSV correlated with subjects’ behavioural ability to differentiate between agents and with the PC1 of subclinical personality traits. Specifically, subjects for whom self and other brain representations were more distinct than self and counterfactual self scored higher on this principal component. An inability to differentiate between self and other is a feature of psychopathology [17–24]. Our measures of agent decoding might be useful as a sensitive gauge of a self–other distinction in the context of phenotypic markers for psychopathology.
Computational phenotyping in psychiatry has recently been mooted [78] and posits individualised diagnostic and therapeutic tools in mental health based on computational models of behaviour and brain function. Recent efforts to develop computational models of Theory of Mind [79, 80] do not address how representations are attributed to different agents. Here, we present the foundations for a model of Theory of Mind that specifically addresses computations that contribute to a self–other distinction, a quantifiable characteristic necessary in both social and nonsocial contexts.
Ethical approval for this study was obtained from the University College London Research Ethics Committee, application number 9929/002.
Forty-one healthy adults (23 female) aged 18 to 42 participated in the experiment. They were recruited from the UCL Institute of Cognitive Neuroscience subject pool. All participants had normal or corrected-to-normal vision and had no history of psychiatric or neurological disorders. One participant was excluded from the analysis due to excessive head movements in the scanner and an additional 2 were excluded due to technical faults with MEG data acquisition, leaving 38 subjects (21 female) for the analysis, with a mean age of 26.6 (SD 6.9). All participants provided written informed consent.
We followed a stringent pipeline to minimise the correlation between our variables of interest. First, we generated 2 random walks to represent the time courses of P and Pfb. One walk started with a value randomly selected from a uniform distribution bound between 0.1 and 0.3, and the other walk started with a value randomly selected from a uniform distribution bound between 0.7 and 0.9. The walks proceeded with step sizes of 0.025. The sign of this step was random in most instances, but because P and Pfb are probabilities, the 2 walks were bound between 0 and 1. To achieve this, the walks were always reflected by these boundaries, which sometimes required a reversal of the randomly selected step sign. The walks were terminated after 408 steps. This resulted in 2 pseudorandom walks, each with 408 data points. If these 2 datasets had a nonsignificant Pearson correlation coefficient, they were saved. This process was repeated iteratively until 300 pairs of uncorrelated pseudorandom walks had been generated. Each pair of walks was then used to generate a trial sequence.
To generate the trial sequence, we first generated a sequence of 408 trial types (‘privileged’, ‘shared’, or ‘decoy’). This sequence consisted of 34 concatenated blocks of 12 trials, such that each block was a random sequence of 4 of each trial type. Thus, trial type frequencies were balanced throughout the experiment. We then used this sequence of trial types along with one of the pairs of random walks to simulate a sequence of Bernoulli trials. One of the random walks represented P and the other represented Pfb. For a trial ti, if the trial type was ‘privileged’, we drew from a Bernoulli distribution with P equal to the ith data point in random walk P. If the trial type was ‘decoy’, we drew from a Bernoulli distribution with P equal to the ith data point in random walk Pfb. If the trial type was ‘shared’, we drew from a Bernoulli distribution with P equal to 0.5. This resulted in 1 sequence of 408 ‘heads’ and ‘tails’. The complete trial sequence consisted of a sequence of Bernoulli outcomes and a corresponding sequence of trial types.
We then simulated an agent that observed this trial sequence to generate trial-wise estimates of our variables of interest: B, Bfb, PEs, and PEo. In order to do this, we used the 2-parameter RW model (model 3 in Table 1), which was the simplest RW model in our model space. We selected parameters by taking the mean of parameter estimates across 18 subjects in a separate pilot study; we used a learning rate (α) of 0.1. We tested for correlations between B and Bfb and between |PEs| and |PEo|. If these correlations were nonsignificant, we saved the trial sequence and moved on to the next pair of random walks. It should be noted that on ‘decoy’ trials, PEs was coded as 0 while PEo was usually nonzero, but on ‘privileged’ trials, PEo was coded as 0 while PEs was usually nonzero (Eq. 1). This meant that there was always a negative correlation between |PEs| and |PEo| (Fig 4A). Therefore, we tested for correlations between these regressors only on ‘shared’ trials, in which both values were nonzero. If a significant Pearson coefficient was discovered, the process was attempted again with a different pseudorandom sequence of trial types. If after 300 iterations no trial sequence was generated that provided uncorrelated variables, we moved on to the next pair of random walks. At the end of this process, we ended up with 158 suitable trial sequences.
Every subject played 2 versions of this game while inside the MEG scanner—an SV and an NSV—one after the other. We created 2 cover stories that could interchangeably be applied to either the SV or NSV, allowing for 4 possible games to be played (SV1, SV2, NSV1, or NSV2). Our cover stories were designed to be immersive and to make the underlying structure of the task, particularly the drifting Bernoulli parameter (P), as intuitive as possible. Each subject played 2 games with different cover stories to make the SV and NSV feel as different as possible. Each subject was allocated to 1 of 4 groups. These groups were defined by the order in which the SV and NSV were played and the cover stories that were applied to both of them. Group 1 played SV1 → NSV2. Group 2 played NSV2→ SV1. Group 3 played SV2 →NSV1. Group 4 played NSV1 → SV2. This 2 × 2 factorial design meant that game order and cover story mappings were counterbalanced across subjects. For each subject, 2 of the 158 pregenerated trial sequences were selected, 1 for their first game and 1 for their second game.
In cover story 1, subjects played the role of a ‘shop assistant’ working in a shop on a tropical island selling only pink umbrellas and yellow sun-shades (i.e., ‘heads’ or ‘tails’). On every trial, a ‘customer’ would come to the shop and buy an umbrella or a sun-shade. The weather on the island was unknown, but it was always changing, and this was reflected in the items that the customers chose to buy. It was the job of the shop assistant to observe and remember the sequence of sales to infer the gradual changes in the weather in order to make predictions about what the next customers would buy. In cover story 2, subjects played the role of an assistant in a shop in the centre of a city selling only red cans of cherry-cola and blue cans of diet-cola. Outside the shop there were large digital advertisements, which were always showing images of these products. The advertisements could be biased towards one of the products, but they were always changing. The assistant could not see the advertisements. On every trial, a customer would come to the shop and buy one of the drinks, as determined by which product was currently favoured by the advertisements outside. It was the job of the shop assistant to observe and remember the sequence of sales to infer the gradual changes in the advertisements in order to make predictions about what the next customers would buy. In both cover stories, subjects were instructed that the hidden states (weather or advertisements) change constantly and slowly and that they must consider every sale in order to keep track of the fluctuations. These environmental fluctuations provided cover stories to justify the random drifting of P.
The differences between the SV and NSV applied to both cover stories. In the SV, there was an accompanying ‘manager’. The manager represented a real person. This was another participant, outside the scanner in a different room. Subjects fully understood the perspective of this other person because they had previously participated in that role (see section Full protocol). Subjects were told that the manager spends some time in a ‘back room’ and therefore does not get to observe all of the sales. Therefore, some of the sales contained ‘privileged’ information that only the shop assistant could see. However, on some trials, the manager came out of the ‘back room’ and did get to see the sale. These were the trials with ‘shared’ information. Finally, subjects were told that the manager was watching ‘CCTV footage’ from the ‘back room’ to keep track of the sales but that the manager was unaware that this was actually last week’s footage, so the information was misleading. The shop assistant was provided with a video link of what the manager was watching on closed-circuit television (CCTV) so they could see all the misleading information that the manager was receiving (‘decoy’ trials). Subjects had to try and keep track of the ‘manager’s’ beliefs.
In the NSV, there was a ‘hi-tech cash register’ instead of a manager. Instead of a real person’s beliefs, subjects had to keep track of a computer’s belief-state. Subjects were told that the cash register keeps a record of every sale that’s been entered to it and computes a prediction of what the next customer will buy. Subjects were also told that some customers buy with coupons instead of cash and that these sales would not be entered into the cash register. These were the ‘privileged’ trials. Subjects were told that some customers pay with cash and in these instances, the cash register would be used and it would update its prediction. These were the ‘shared’ trials. Finally, subjects were told that the cash register had an internet connection to a partner shop and received updates from the sales happening there. However, the partner shop was far away from this shop, on another island with different weather for cover story 1, or in another city with different digital advertisements for cover story 2. When these sales occurred, the ‘cash register’ was updated with misleading information, and the subject could also see this information (‘decoy’ trials).
All 4 games had identical designs and trial structures (Fig 1A). A sampling trial was one in which the subject samples the environment on behalf of themselves (‘privileged’), on behalf of the manager/cash register (‘decoy’), or on behalf of themselves and the manager/cash register at the same time (‘shared’). Sampling trials started with cue image (Fig 1B) on the screen to indicate whether it was going to be a ‘privileged’, ‘decoy’, or ‘shared’ trial. The cue was presented in the centre of the screen with a grey background for 1,100 ms. Then, the cue disappeared, and an outcome was presented (Fig 1C) at the centre of the screen for 900 ms. The outcome represented what the current customer had chosen to buy (i.e., the outcome of the Bernoulli trial: heads or tails). Then, the outcome disappeared and was followed by an intertrial interval (ITI) with a central fixation cross that lasted 750 to 1,250 ms.
After 4 to 9 sampling trials, there was a probe trial (Fig 1A) in which subjects reported their estimate of either P or Pfb. Subjects were told that the 2 products of their shop were kept in 2 separate boxes. On probe trials, a horizontal scale appeared on the screen with one of the boxes on the left and the other box on the right. The positions of the 2 boxes were randomly generated on every probe trial, to avoid any directional biases. A probe trial could be a ‘self’ trial (report P) or an ‘other’ trial (report Pfb). On ‘self’ trials, subjects were probed with the question ‘Which box would YOU reach into now?’ as if anticipating what the next customer would buy. Subjects had 7 seconds to give their response by moving a virtual arrow left or right along the scale and then pressing an ‘enter’ button once they were happy with the position. The arrow was initially invisible but appeared in a random location along the scale as soon as subjects pressed left or right. This was to avoid any systematic biases that the starting location of the arrow might have induced. ‘Other’ trials were different between the SV and NSV. In the SV ‘other’ trials, subjects were asked ‘Which box would the MANAGER reach into now?’ and subjects had to respond by putting themselves into the shoes of the manager and reporting their estimate of Pfb. In the NSV, this question was rephrased as ‘Which box would you reach into now IF you used the readout on the cash register?’ Subjects never received any feedback from their choices on probe trials, so they never knew how accurate their responses were. Probe trials were randomly selected to be ‘self’ or ‘other’, so subjects never knew which question was going to come next and they had to try to keep track of P and Pfb at all times. However, subjects were never probed with the same question more than 3 times in a row.
Subjects were incentivised to perform well because their final payment at the end of the experiment depended on their accuracy scores on probe trials. We calculated scores based on how much subjects’ responses deviated from the ‘optimal’ response on any given probe trial. In the SV ‘other’ trials, the ‘optimal’ response was whatever the other participant selected for that trial as the ‘manager’. For all other probe trials, the ‘optimal’ response was taken from the random walks P or Pfb. Once again, participants never got to see these scores nor received any kind of feedback until they were paid at the end of the experiment.
Before playing the games, participants were carefully instructed and well practiced to make sure they understood that after a probe trial, the hidden state of the environment would continue from the state it was in before the probe trial. Specifically, they were instructed that after a probe trial, they should not treat the upcoming information as independent of what they had already seen but rather treat the entire stream of sampling trials as a continuous sequence while the hidden state of the environment (weather or advertisements) changed gradually.
Deception was never used in this experiment; the setup was exactly as it was described to participants. All participants came to the lab on 2 occasions, T1 and T2. T2 occurred no more than 4 days after T1. We will explain both sessions with the fictional participants ‘Sally’ and ‘Anne’. Sally and Anne were both in group 1, which meant that while they were in the scanner, they played SV1 followed by NSV2, but they never played SV2 or NSV1. Anne arrived at the lab at time T1 for a behavioural session with no scanning. Anne was instructed how to play a simplified version of SV1 called simpleSV1. In simpleSV1, Anne played the role of the ‘manager’. The game proceeded exactly as SV1 (described above) but with the following differences: (1) there was no mention of a shop assistant; (2) ‘privileged’ trials were excluded; (3) Anne was instructed to ignore the cue images and told that they were not relevant to the game; and (4) there was only one type of probe trial, which asked ‘Which box would YOU reach into now?’ The presence of cue images was justified with instructions that said that the images simply showed whether Anne (as the ‘manager’) was seeing the sale directly or via CCTV from the back room but that this was irrelevant for Anne’s task.
After this, Anne was instructed how to play a simplified version of NSV2 called simpleNSV2. In this version of the game, Anne played the role of a ‘shop assistant using the information from a hi-tech cash register’. The same differences applied as in simpleSV1. The presence of cue images was justified with instructions that said that the images simply showed whether the cash register was getting information from a sale in this shop or in a partner shop next door but that this was irrelevant for Anne’s task. We did not analyse the behavioural data from simpleSV1 or simpleNSV2.
Finally, Anne was given the following set of personality questionnaires to complete: BIDR, EQ, IRI, CAPE, ICU, and BDI. Anne and all other participants filled out the BIDR first in order to provide us with a measure of response bias. The subsequent 5 questionnaires were completed in a random order.
The next day, Anne returned for session T2. She was then informed that there was a social element to the experiment and that another participant, Sally, was also there but that Sally was doing a T1 session—i.e., Sally was doing what Anne had done the previous day. We walked Anne past a testing room so that she would briefly see Sally to convince her that there was indeed another participant. Anne was then taken to a different testing room where she played a short working memory (n-back) task and then learned how to play SV1 and NSV2 with extensive practice. We explained to Anne that in SV1, she would be trying to predict Sally’s choices as the ‘manager’ and that in NSV2, she would be trying to make choices as if she were using the predictions of a computer (cash register). For SV1, Anne was told that the CCTV footage that Sally was receiving was actually uninformative, and for NSV2, Anne was told that the partner shop down the road was closed and so by default the cash register was now receiving misleading information from another partner shop in a completely different place (different island or different city). Finally, Anne was taken to the MEG scanner where she played SV1 followed by NSV2. For various logistical reasons, Sally and Anne did not play their games as the ‘manager’ and ‘shop assistant’ simultaneously, which we explained to Anne. Sally’s responses as the ‘manager’ were saved to a network drive, and they were used to calculate Anne’s score on ‘other’ probe trials in SV1. The same pregenerated trial sequence was given to Sally and Anne except that ‘privileged’ trials were excluded from Sally’s simpleSV1 game—which consisted of only 272 sampling trials—while Anne’s SV1 game consisted of the full 408 sampling trials.
Anne had to play simpleSV1 the previous day for 2 reasons. Firstly, it provided ‘manager’ responses for a previous participant who was playing as the ‘shop assistant’ in the scanner (just like Sally provided ‘manager’ responses while Anne was in the scanner). Secondly, it was to make it more intuitive for Anne to put herself in Sally’s shoes and to understand exactly what information was and wasn’t available to Sally. The reason why Anne had to play simpleNSV2 was simply so that the images and cover stories for the 2 games that Anne played at T2 (SV1 and NSV2) were equally familiar.
All tasks were implemented in MATLAB (MathWorks, Natick, MA) using Cogent (Wellcome Trust Centre for Neuroimaging, University College London, London, England). All images used were edited in Inkscape and processed in MATLAB to ensure equal luminance (average luminance per pixel) with each other and with the plain grey background.
Our primary hypothesis was that subjects would use PEs to update their beliefs and that they would also simulate the other person’s PEs to solve the belief inference problem. In order to test this hypothesis, we developed a series of computational models to try and explain subjects’ choice behaviours. All the models described here were fitted to choice behaviour from the SV and NSV using identical procedures. We tested 3 different groups of models.
The models in group A assumed that subjects did not use PEs to update their beliefs on each trial but rather used an ‘averaging’ technique. Model 1 predicted a response on each probe trial by taking the average of information sampled since the last probe trial. This model predicted choices on ‘self’ probe trials by averaging information from ‘privileged’ and ‘shared’ trials and predicted choices on ‘other’ probe trials by averaging information from ‘shared’ and ‘decoy’ trials. Model 2 was the same as model 1, but instead of taking the average over the sampling trials since the last probe trial, this model took the average over the last 10 sampling trials.
The models in group B assumed that subjects used PEs to form trial-wise belief updates. All of these models used 2 different PE signals—a PEs (Eq 1) and PEo (Eq 2)—and assumed 2 update equations on each sampling trial, 1 for updating the beliefs of the ‘self’ agent (Eq 3) and 1 for updating the beliefs of the ‘other’ agent (Eq 4). All of the group B models assumed that a nonzero PEs was generated on ‘privileged’ and ‘shared’ trials and that a nonzero PEo was generated on ‘shared’ and ‘decoy’ trials. However, PEs was equal to 0 on ‘decoy’ trials (Eq 1), and PEo was equal to 0 on ‘privileged’ trials (Eq 2).
PEs(t)={0(Decoy)Outcome(t)−B(t−1)(Otherwise)
(1)
PEo(t)={0(Privileged)Outcome(t)−Bfb(t−1)(Otherwise)
(2)
where B(t) is the subject’s belief about P on trial t and Bfb(t) is subject’s belief about Pfb on trial t. All group B models assumed that these estimates were both initialised at 0.5. Outcome was coded as 1 or 0. The simplest models used the following pair of update equations:
B(t)=B(t−1)+α.PEs(t)
(3)
Bfb(t)=Bfb(t−1)+α.PEo(t)
(4)
where α is the learning rate. This is a free parameter that was fitted to each subject and remained constant throughout the task. It should be noted that when PEs and PEo are zero, B and Bfb remain stationary. To account for any information lost since the last update, some models included an additional memory-decay parameter δ, which governs a decay of B and/or Bfb back to the initial value of 0.5:
B(t)=B(t−1)+α.PEs(t)+δ(0.5−B(t−1))
(5)
Bfb(t)=Bfb(t−1)+α.PEo(t)+δ(0.5−Bfb(t−1))
(6)
Finally, some models allowed for the possibility that the PEs might be used to erroneously update Bfb or that PEo might be used to erroneously update B. Some models assumed a degree of this PE ‘leakage’ with the parameter λ:
B(t)=B(t−1)+α.PEs(t)+λ.PEo(t)
(7)
Bfb(t)=Bfb(t−1)+α.PEo(t)+λ.PEs(t)
(8)
These 3 free parameters α, δ and λ could take any value between 0 and 1. Different models that incorporate various combinations of these parameters can be generated. α can be shared between the 2 update equations, or alternatively, each equation can have its own α with different values. The same can be said for δ and λ, which can also be excluded from either equation entirely. For example, some models included a λ parameter for one update equation but not the other, to allow for a unidirectional PE leak.
The final group of models, group C, had the same update equations as in Eq 5 and Eq 6 (incorporating a shared α and a shared δ), but these models tested the possibility that subjects did not selectively update their beliefs depending on the cues. Model 20 assumed that a nonzero PEs was generated on every sampling trial despite the fact that the information on ‘decoy’ trials should not have been relevant for updating ‘self’. Model 21 assumed that a nonzero PEo was generated on every sampling trial despite the fact that the information on ‘privileged’ trials should not have been relevant for updating ‘other’.
MEG was recorded continuously at 600 samples per second using a whole-head 275-channel axial gradiometer system (CTF Omega, VSM MedTech, Coquitlam, Canada) while participants sat upright inside the scanner. Two gradiometers (ML042 and MRC12) were out of service throughout data collection; we preprocessed and analysed data from the remaining 273 channels. We recorded 4 runs of data for each subject (2 runs for the SV and 2 runs for the NSV). Participants made responses on 4 buttons with a button box using the fingers they found the most comfortable. The buttons had the following functions: move arrow left, move arrow right, move arrow faster, and enter choice.
All MEG preprocessing was carried out using the FieldTrip data analysis toolbox [81] on each run of data independently. First, we epoched the data around the relevant triggers and scanned the data for any jump artifacts. In the whole experiment, 1 jump artifact was found, and the relevant trial was excluded from the analysis. Then we downsampled the data from 600 Hz to 100 Hz and filtered the data using a bandpass of 0.5 to 150 Hz and a stopband of 48 to 52 Hz to remove line noise. We then ran an independent components analysis using the built-in ft_componentanalysis function in FieldTrip and manually inspected the components for obvious eye artifacts and cardiac ECG artifacts. The relevant components were removed, and the data were reconstructed. All analyses were conducted on the epoched, filtered, resampled and cleaned data in units of femtotesla.
For each subject, on each probe trial, we computed the absolute difference between the subject’s report and the true underlying probability from random walk P or random walk Pfb. We then computed this difference again for a simulated agent that positioned the arrow at random locations along the response scale. We then subtracted the real difference from the simulated difference on each probe trial and took the mean across probe trials. The resulting value describes how much better than a random subject the real subject performed. We took the mean across subjects to assess group-level performance.
All 21 models were fitted to the choice behaviour (on probe trials) of each subject with MATLAB’s nonlinear optimisation function fmincon. We used this function to find the optimal model parameters for each subject as defined by the minimum negative log-likelihood of the subject’s choices conditioned on a set of estimated parameter values. To start the optimisation, every parameter was initialised with a value randomly drawn from a uniform distribution bound by the relevant upper and lower bounds for that parameter. The optimisation procedure was iterated at least 20 times for each model fit, with different initial parameter values each time, to avoid local minima. We selected the iteration with the best-fitting optimised parameters and discarded the rest.
In order to obtain likelihood values, we derived an action likelihood function (ALF) for each trial, which specified the likelihood of any choice that could have been made by the subject on that trial. The subject’s choice could take any value along a continuous scale. Because subjects were technically reporting a probability, this scale was bound between 0 and 1, and we used a Beta distribution to approximate this ALF. Beta distributions are conventionally parameterised by 2 shape parameters, α and β (Eq 9). It should be noted that this is not the same α as the learning rate used in our learning models.
f(x;α,β)=x(α−1)(1−x)(β−1)Beta(α,β)
(9)
where x is the observed data and f(x; α, β) is the likelihood of that data given the shape parameters of the Beta distribution. The Beta function is a normalisation constant to ensure that the total probability integrates to 1. We wanted to parameterise the ALF with more meaningful parameters: (1) the most likely choice on that trial and (2) the variability or ‘temperature’ of the subject’s actions. We assumed that the most likely choice on a ‘self’ probe trial was the current estimate B(t) and the most likely choice on an ‘other’ probe trial was the current estimate Bfb(t), and the mode was assigned to one of these variables (Eq 10). The variance of the Beta distribution was assigned to the value of a free parameter called τ (Eq 12). This parameter was fitted to each subject separately and remained constant throughout the task. It was bound between 0.0001 and 0.08. τ is a temperature parameter that captures how noisy each subject was in his or her mapping from belief to action. A model could have a shared τ parameter for ‘self’ and ‘other’ probe trials or have 2 separate parameters that can vary independently. In order to draw the ALF on each trial, we derived the Beta distribution shape parameters, α and β, which can both be expressed in terms of mode (Eq 11) and variance (Eq 13). These 2 equations could then be solved simultaneously to obtain the shape parameters and then the Beta distribution itself (Eq 9).
Mode≔{B(t)(SelfProbe)Bfb(t)(OtherProbe)
(10)
α−1α+β−2=Mode
(11)
σ2≔τ
(12)
αβ(α+β)2(α+β+1)=σ2
(13)
The real choice that was made by the subject could then be read off the ALF, and the corresponding likelihood would contribute to the joint likelihood of all choices made conditioned on the current model parameter estimates. We computed the BIC for each model and each subject separately (Eq 14) and compared the mean BIC value across subjects for each model to assess relative model evidence. It should be noted that the ALF is a probability density function, and so the likelihoods were often larger than 1. Therefore, the log-likelihoods were often positive and their corresponding BIC values negative.
BIC=ln(N).k−2ln(L^)
(14)
where N is the number of data points that the model was fitted to (in this case, number of trials), k is the number of free parameters in the model, and L^ is the maximised value of the likelihood function of the model.
Using the BIC as an approximation for log model evidence, we also compared the winning model and the second-best model using a hierarchical Bayesian model to estimate the posterior probability that any randomly chosen subject in the sample had data generated by one of those models and not the other. This random-effects Bayesian model selection approach enables us to estimate the parameters α of a Dirichlet distribution of the probabilities r of the models being compared. These probabilities inform a multinomial distribution over the model space. The hierarchical model is inverted using variational Bayesian approximation [32]. Then, from the Dirichlet parameters, one can compute the expected multinomial parameters 〈rk〉 for each model k as follows:
〈rk〉=αk/(α1+…+αk)
(15)
One can also compute an exceedance probability φk, i.e., the belief that a particular model k is more likely than any other model tested, given the group data y:
φk=p(rk>0.5|y;α)
(16)
All analyses were performed on data time locked to the onset of the outcome of the sampling trial because this is the information that subjects would require to generate a PE and update their beliefs. In order to regress PE against recorded brain activity, we performed a mass univariate regression analysis in which we took the absolute PE magnitude as our regressor to ensure we were not capturing brain activity that correlated with the visual attributes of, or other associations with, the image that represented the outcome. We had to take into account the fact that PEs and PEo were coded as 0 on some trials. Therefore, we conducted 2 separate regressions. One regression for PEs excluded ‘decoy’ trials, and 1 regression for PEo excluded ‘privileged’ trials. Therefore, in both regressions, the regressor would contain no systematic pattern of zero-valued elements. If we hadn’t excluded the zero-valued trials, the regression would have been confounded by trial type and cue images (PEs was 0 only after a particular cue image on a ‘decoy’ trial, and PEo was 0 only after a particular cue image on a ‘privileged’ trial).
We regressed PE magnitude against the ERF at each sensor and each time point, to produce a spatiotemporal map of unsigned regression weights. At each sensor, we subtracted the median prestimulus value. We then upsampled the data to create a 95 × 95 2D pixel map of these baseline-corrected effect sizes. Including the time dimension, we ended up with a 3D image for each subject. At the group level, we performed a one-sided Wilcoxon signed-rank test at each pixel in these 3D images to ask whether the group median was significantly greater than 0. We used a nonparametric test here because baseline-corrected unsigned regression weights are not normally distributed. Here, we are testing the null hypothesis that the effect size (unsigned regression weight) is no larger than the prestimulus effect size. We identified points with activations above a cluster-forming threshold (P < 0.001). We then identified clusters of contiguous suprathreshold points in this 3D image, which could extend through space and time. We made cluster-level inference by repeating this analysis 300 times, using permuted trial sequences, to generate null distributions of cluster extent, from which we derived significance thresholds (P = 0.05 FWE-corrected).
We used a nonlinear SVM—LIBSVM for MATLAB [82]—with a radial basis function (RBF) kernel, which was trained on pseudotrial data with 273 features, one for each sensor. Crossvalidation was performed with repeated random subsampling. We performed 200 iterations, and on each iteration, 2 pseudotrials from each class (‘self’ and ‘other’) were randomly sampled and constituted a testing set, while the remainder constituted a training set. On each fold of crossvalidation, the training set and testing set were both centred and normalised with respect to the means and SDs of the data in each feature (MEG sensor) in the training set. Accuracy was measured as the percentage of correct predictions, averaged over the 200 folds of crossvalidation. This value indicated how much information was available in the |PE| signals to discriminate between PEs and PEo. We performed this analysis at each point in peristimulus time to generate a time course of CAs. This procedure was conducted once for the SV and again for the NSV.
We tested all classifiers on a range of hyperparameter combinations (regularisation constant C and RBF parameter γ). We optimised each classifier by selecting the combination of hyperparameters that produced the highest CA averaged across subjects. For the |PE| pseudotrial analysis, C = 103 and γ = 10−6 in the SV, and C = 103 and γ = 10−9 in the NSV. For the ‘signed belief’ pseudotrial analysis, C = 103 and γ = 10−6 in the SV, and C = 103 and γ = 10−6 in the NSV. For the ‘unsigned belief’ pseudotrial analysis, C = 10−9 and γ = 10−9 in the SV, and C = 10−6 and γ = 10−9 in the NSV. The time course of CAs was then smoothed with a moving average filter (moving average span of 10 data points).
We computed significance thresholds by running 150 simulations of these exact analyses, each time with data generated from permuting the trial order. For each permutation, we obtained a maximal CA and used these values to generate a null distribution. If the real CA exceeded the 95th percentile of the null distribution, it was deemed significantly above chance. This method corrects for multiple comparisons in the time domain. We also used this method to determine whether the difference in CA between SV and NSV, at each time point, was significant. For this, we generated a null distribution of maximal (SV − NSV) CA differences and another null distribution of maximal (NSV − SV) CA differences.
We used scores from the BIDR questionnaire to determine whether any subjects were likely to have a large response bias. This questionnaire allocates a point whenever a subject gives an extreme response (6 or 7 on a 7-point Likert scale) to a question indicating that they might be answering in such a way as to preserve their reputation. No subject had a score more than 2.5 SDs greater than the sample mean, so we had no reason to believe that any subject had an unusually large response bias.
For each subscale of our 5 questionnaires of interest, we set up a regression model with gender and age as predictor variables and the subscale score as a dependent variable. We then took the residuals from these regression models as age- and gender-controlled scores for each subscale. We then z-scored each of these 9 age- and gender-controlled subscales and entered them into a PCA. We investigated the principal component that explained the most variance in the data.
When conducting the correlational analysis at each time point between the PC1 score and the neural agent decoding value, we computed a significance threshold with a permutation-based null distribution. For each of the 150 simulations of the classification analysis, we simulated the correlation analysis to generate a time course of −ln(p) values for each of the 3 types of pseudotrial and then concatenated these 3 time courses together. For each permutation, we took the maximal value of this triple-length time course, resulting in a null distribution of maximal −ln(p) values. The 95th percentile of this distribution represented a significance threshold correcting for multiple comparisons across time and across the 3 types of pseudotrial.
|
10.1371/journal.pcbi.1002178 | Large-Scale Conformational Changes of Trypanosoma cruzi Proline Racemase Predicted by Accelerated Molecular Dynamics Simulation | Chagas' disease, caused by the protozoan parasite Trypanosoma cruzi (T. cruzi), is a life-threatening illness affecting 11–18 million people. Currently available treatments are limited, with unacceptable efficacy and safety profiles. Recent studies have revealed an essential T. cruzi proline racemase enzyme (TcPR) as an attractive candidate for improved chemotherapeutic intervention. Conformational changes associated with substrate binding to TcPR are believed to expose critical residues that elicit a host mitogenic B-cell response, a process contributing to parasite persistence and immune system evasion. Characterization of the conformational states of TcPR requires access to long-time-scale motions that are currently inaccessible by standard molecular dynamics simulations. Here we describe advanced accelerated molecular dynamics that extend the effective simulation time and capture large-scale motions of functional relevance. Conservation and fragment mapping analyses identified potential conformational epitopes located in the vicinity of newly identified transient binding pockets. The newly identified open TcPR conformations revealed by this study along with knowledge of the closed to open interconversion mechanism advances our understanding of TcPR function. The results and the strategy adopted in this work constitute an important step toward the rationalization of the molecular basis behind the mitogenic B-cell response of TcPR and provide new insights for future structure-based drug discovery.
| There is an urgent need for the development of better drug therapies for tropical diseases, including Chagas' disease, sleeping sickness and leishmaniasis. Known collectively as the human trypanosomiases, these traditionally neglected diseases are responsible for substantial human suffering and death in Latin America and sub-Saharan Africa. Current chemotherapy for Chagas' disease is particularly unsatisfactory, with available drug treatments displaying poor efficacy and undesirable toxic side effects. Recent developments in the study of the basic biochemistry of the causative Trypanosoma cruzi parasite and its host infection mechanism have identified an essential proline racemase enzyme (TcPR) as a novel target for chemotherapeutic intervention for Chagas' disease. Conformational changes associated with substrate binding to TcPR are believed to expose critical residues that elicit a host mitogenic B-cell response, a process contributing to parasite persistence and the undermining of host immunity against T. cruzi. Here we describe advanced accelerated molecular dynamics simulations that capture previously uncharacterized large-scale motions of TcPR. These motions reveal new conformational epitopes of potential importance for the mitogenic B-cell response. Furthermore, knowledge of the conformational interconversion mechanism and corresponding transient binding pockets will greatly aid future structure-based drug discovery efforts.
| The protozoan diseases African sleeping sickness, leishmaniasis and Chagas' disease are responsible for substantial human suffering and death. Caused by parasites from the genus Trypanosoma these insect spread diseases mainly affect the underprivileged in tropical regions [1]. Limited drug therapies, human migration and environmental changes have contributed to the increasing spread of these traditionally neglected diseases. Chagas' disease, caused by the Trypanosoma cruzi parasite (T. cruzi), threatens the lives of millions of people from southern USA to southern Argentina [1], [2]. The need for new drugs is urgent with current treatments having poor efficacy and safety profiles, particularly in the late stage of the disease when the parasite has infected critical organs.
Recent studies have revealed an essential T. cruzi proline racemase enzyme (TcPR) as an attractive new candidate for chemotherapeutic intervention [3]. TcPR catalyzes the reversible stereoinversion of L- and D-proline [4]. Tonelli et al. showed that L-proline is essential for the intracellular differentiation of T. cruzi. [5]. Later, Chamond et al. demonstrated that over-expression of TcPR increases parasite differentiation into infective forms and its subsequent penetration into host cells [6]. In another study, Coatnoan et al. observed that, in addition to free D-amino acids, parasite extracts contain peptides composed of D-proline; indicating a possible mechanism used by the parasite to confer resistance against host proteolytic mechanisms [7].
TcPR has also been characterized as a potent host B-cell mitogen that sustains parasite evasion of specific host immune responses [3], [8]. B-cell proliferation and polyclonal antibody activation constitute a widespread mechanism of immune evasion shared by many pathogens. This process compromises immune response activation through generation of non-pathogen-specific B-cells that effectively mask specific reactions against the invading pathogen. In Chagas' disease, B-cell proliferation has also been linked with resistance to infection, disease progression and the pathology associated with its chronic form [3]. Taken together, these data provide strong evidence that TcPR represents a promising target for therapies that may more efficiently combat Chagas' disease.
Emerging crystallographic and mutagenesis data indicate that ligand-induced conformational changes in TcPR modulate the exposure of critical residues that elicit a host mitogenic B-cell response [3], [9]. Two crystal structures of TcPR are currently available (Figure 1). Each structure was solved with the transition state analog pyrrole-2-carboxylic acid (PYC) bound to either one or both monomers. In the presence of PYC monomers display a common closed conformation. In contrast a semi-open conformation is apparent for the ligand-free monomer (Figure 1). Together with calorimetric studies, these results indicate that a large inter- and intra- domain closure movement is coincident with ligand binding [9]. Although enzymatic inhibition by PYC abolishes mitogenic activity, point mutations of the catalytic cysteine residues (C130S and C300S) have little or no effect. Therefore, these studies not only showed that the mitogenic and enzymatic activity of TcPR are decoupled, but also strongly indicate that ligand-induced conformational changes upon binding prevent the interaction of TcPR with B-cell receptors.
Here we show for the first time a model for the open form of TcPR. Characterization of the opening transition required the application of state of art accelerated molecular dynamics simulations, which extends the effective simulation time scale of conventional molecular dynamics. When combined with sequence conservation and small molecule fragment mapping analyses our results indicate that the mitogenic properties of TcPR are likely associated with the exposure of conserved conformational epitopes located around previously unidentified binding pockets. This work represents an important step toward the rationalization of the molecular basis of TcPR initiated B-cell response and provides a basis for future structure-based drug discovery.
Characterization of the opening movement of TcPR requires access to long-time scale, inter-domain motions that are currently inaccessible by conventional molecular dynamics (MD) simulations [10]. To overcome this limitation, we applied an enhanced sampling technique developed in our group, called accelerated molecular dynamics (aMD), which extends the effective simulation time scale. In aMD, a continuous non-negative boost potential function, ΔV(r), is added the original potential surface (V(r)) such that regions around the energy minima are raised and those near the barriers or saddle points are left unaffected. ΔV(r) is defined according to ΔV(r) = (E−V(r))2/(α+E−V(r)). Whenever V(r) is below a chosen threshold boost energy, E, the simulation is performed on the modified potential V*(r) = V(r)+ΔV(r), otherwise sampling is performed on the original potential V*(r) = V(r). The parameter α modulates roughness and the depth of the energy minima on the modified surface, as previously shown (see materials and methods for details) [11]–[17].
The closed crystal structure of TcPR in complex with two transition-state analog inhibitors (PDB code: 1W61) was used to build our initial model. Atomic coordinates of bound PYC were removed from the active site of each monomer resulting in a ligand-free closed system that underwent 100 ns of aMD simulation. To characterize dominant conformational states, along with inter- and intra-domain opening motions, the final aMD trajectory was subjected to principal component analysis (PCA) [14], [18]. Figure 2 displays the two-dimensional representation of the structural dataset as a projection of the distribution onto the subspace defined by the first and second principal components. Large-scale opening motions of TcPR were well characterized and captured by PC1 and PC2 (which together with PC3 accounted for over 70% of the variance in the original distribution: see Figure S1). Clustering of trajectory conformers was used to visualize the dominant conformations sampled by the simulation (Figures 2 and Figure S2). Two major clusters, encompassing the closed and open conformational states, are clearly identified in the ensemble of conformers. Six representative structures, which include closed and open cluster representatives of TcPR, are displayed in Figure 2. The TcPR structures are shown in molecular surface representation colored according to the level of residue conservation within the proline racemase family (with blue and red representing low and highly conserved residues respectively, see materials and methods for details).
Projection of the two available crystal structures onto the PCs obtained from the aMD trajectory reveals that both closed and semi-open forms of TcPR are well characterized by the conformers sampled in the vicinity of state 1 (Figure 2), indicating that significantly larger opening motions are observed in the trajectory. Projection of TcPR aMD trajectory onto PC sub-space characterized two dominant global motions: (a) A large-scale inter-domain motion that exposes several conserved residues located in the monomer-monomer interface, and that is observed when the conformer population shifts from state 1 to states 15, 16 or 17 (moving along PC1) and (b) a large-scale intra-domain opening movement that exposes highly conserved segments around the active site region of TcPR, and that is observed when the system shifts population from state 1 to 7 (moving along PC2). Combination of these two global motions leads TcPR to regions around states 18, 19 and 20. These states represent some of the most open structures accessed by aMD and display the newly identified and highly solvent exposed regions at both monomer-monomer interface and in the surrounds of the active site (see Video S1). As shown in Figure 3, the surface area exposed by states 1 to 20 is dramatically larger than the one presented by the semi-open crystallographic structure 1W62. For instance, the large conformational change involved in the formation of the bound complex from open states 6, 17, 18 or 20 buries an area of ∼6000 Å2; corresponding to approximately four times the buried area observed in crystal structures 1W61 and 1W62 (∼1500 Å2) [9]. To better visualize the magnitude of the long-time large-scale conformational changes, states 1 (closed) and 19 (open) were compared to the superimposed crystal structures (See Videos S2 and S3). It is worth noting that the amplitude of the motion associated with the experimentally observed conformational change is significantly smaller than the ones associated with the inter- and intra-domain motions obtained in our aMD simulation.
To further understand the physical basis of the observed opening motions, we analyzed the available structures with a simplified elastic-network normal mode method [19]. In the elastic network approach, a single model (expressed in terms of Cα coordinates) leads to an objective expression of possible protein dynamics in terms of a superposition of collective normal mode coordinates [20]. The structural mobility predicted by Normal Modes Analysis (NMA) performed on the semi-open structure (PDB ID 1W62) revealed a high overlap between the lowest three modes and the eigenvectors obtained from aMD simulations (0.6 for mode 1 to PC1, see Figure S3). This result indicates that the dominant collective motions during aMD simulation, that capture the TcPR opening movement, are indeed low-frequency motions intrinsic to the structure.
As previously noted, in vitro assays of B-cell proliferation together with structural information strongly indicate that the closure movement induced by ligand binding prevents the interaction of TcPR with B-cell receptor molecules [3], [9]. Activation of B-cell polyclonal response is likely to be associated with the occurrence of transient binding pockets, along with conformational epitopes, in the open ligand-free form of TcPR. In order to identify potential B cell binding sites in the newly identified open states, we used a fragment-based approach (FTMAP) to map binding hot spots on each of the twenty dominant trajectory conformers [21]. Based on the ideas behind screening small organic fragments by NMR and X-ray crystallography, FTMAP correlates pocket druggability with their propensity to bind clusters of small organic compounds. Figure 4 displays the mapping results for states 1 to 20. To further characterize the location of each hot spot, probe occupancy was calculated and assigned to each residue of TcPR (see materials and methods for details). Figure 5a displays the final probe occupancy values obtained after combining and normalizing the results from all twenty conformational states.
As expected, high probe occupancy values were obtained for sites around the catalytic cysteines (residues 130 and 300), consistent with the existence of this binding site in all states, 1 to 20. Several additional pockets, displaying low and high occupancies, were also identified. It is worth mentioning that the large variation in probe occupancies reveals the intrinsic dynamic nature of these binding pockets. Nevertheless, residues showing low probe occupancy values identify regions on the protein surface where potential interaction sites are exposed only in the open form of TcPR, (this includes interaction sites in the vicinity of residues 186–191, 217–218 and 288–291). In order to quantify the exposed surface area of each residue associated with the opening movement, the percentage exposure was calculated based on the per residue solvent accessible surface area of each state and the closed form of the TcPR (1W61). As can be seen in Figure 3c, the newly identified binding pockets are indeed sites that become considerably more exposed in the open states (Figure 3c, light grey). Moreover, sequence conservation analysis shows that these binding pockets are also highly conserved in all proline racemases (Figure 3b).
X-ray crystallography and mutagenesis studies indicate that interaction with B-cell receptors is likely to be associated with the presence of transient binding pockets that are fully formed only in the ligand-free open TcPR. In this work, we show for the first time a model for the open form of this important drug target obtained through the application of state of the art molecular dynamic simulation. Additionally, our results indicate that the mitogenic properties of TcPR may be associated with the exposure of conformational epitopes located around the newly identified binding pockets. Experimental mutagenesis studies of these sites is required to verify their potential role in eliciting host B-cell responses. In summary, the strategy adopted in this work allowed the characterization of large-scale conformational changes associated with the dynamic formation of potential interaction sites coupled with the exposure of highly conserved regions of the protein surface (Figure 5). Furthermore, the results presented in this work represent the first attempt to rationalize the molecular basis of the mitogenic B-cell response to TcPR and provide a basis for future structure-based drug discovery.
All simulations were performed with the AMBER10 package [22] and corresponding all- atom potential function ff99SB [23]. Unless otherwise noted, all additional analyses were performed with the Bio3D package (available from http://mccammon.ucsd.edu/~bgrant/bio3d/) [18].
The crystal structure of Trypanosoma cruzi proline racemase (TcPR) in complex with 2 molecules of pyrrole-carboxilic acid (PDB code 1W61) was used to build our model. Initial atomic coordinates build of the apo form of TcPR was obtained by removing both inhibitors molecules from 1W61. In our model, basic residues like Arg and Lys are protonated, and acidic residues like Asp and Glu are deprotonated. Due to its normal pKa, the His residues were assumed to be neutral at physiological pH.
Initial energy minimization was performed by applying 500 steps of steepest descent followed by 500 steps of conjugate gradient minimization. The relaxed structures were then solvated in a truncated cubic box of pre-equilibrated TIP3P water molecules, which extended 10 Å further than the protein atoms. To neutralize the systems, sodium counterions (Na+) were added to balance the charge of the protein. The system was then energy minimized for 500 steps of steepest descent followed by 500 steps of conjugate gradient minimization using constant volume periodic boundaries. We kept the protein atoms and the ions fixed throughout the whole preparation process. In order to relax the protein in the solvent environment, all coordinates were optimized by employing 1000 steps of steepest descent followed by 1500 steps of conjugate gradient. After that, a 1 ns molecular dynamics (MD) simulation was preformed to heat the system from 0 K to 300 K, for which we applied the NVT ensemble (T = 298 K). To bring the systems to the correct density, we carried out a 100 ps MD simulation on which NPT ensemble (T = 298 K, P = 1 bar) was applied. For the production runs, we performed an additional Accelerated MD simulation (aMD) of 100 ns. The equations of motion were integrated every 2.0 fs using the Verlet Leapfrog algorithm. For analysis, the trajectory was sampled every 1.0 ps. During the MD runs, temperature and pressure were controlled via a weak coupling to external bath with a coupling constants of 0.5 and 1.0 ps, respectively. The center-of-mass motion was removed at regular intervals of 500 fs. The PME summation method was used to treat the long-range electrostatic interactions in the minimization and simulation steps of the solvated systems. The short-range nonbonded interactions were truncated using an 8 Å cutoff and the nonbonded pair list was updated every 20 steps. All calculations, conventional and accelerated MD simulations, were performed using an in-house modified version of AMBER10 package.
Accelerated MD approach modifies the energy landscape by adding a boost potential, ΔV(r), to the original potential surface every time V(r) is below a pre-defined energy level E [16], as(1)where α modulates the depth and the local roughness of the energy basins in the modified potential. In principle, this approach also allows the correct canonical averages of an observable, calculated from configurations sampled on the modified potential energy surface, to be fully recovered from the accelerated MD simulations. In order to simultaneously enhance the sampling of internal and diffusive degrees of freedom a dual boosting approach was employed, based on separate torsional and total boost potentials as [15](2)where Vt(r) is the total potential of the torsional terms, ΔVt(r) and ΔVT(r) are the boost potentials applied to the torsional terms Vt(r) and the total potential energy VT(r), respectively. The parameters were set as follows. Et = , i.e. 30% higher than the ensemble-averaged torsional potential energy from conventional MD simulation. αt≈500 kcal mol−1 chosen based on previous work by de Oliveira [11]. ET = 0.2 kcal mol−1 (nr. atoms)−1 plus the ensemble-averaged total potential energy from conventional MD simulation. αT≈0.2 kcal mol−1 (nr. atoms)−1 [11]. These ET and αT values allow to reproduce the most relevant structural and energetic properties of liquid water while increasing the water self-diffusion coefficient by ∼15% [11], [15].
Prior to trajectory superposition and Principal component analysis (PCA), iterated rounds of structural superposition were used to identify the most structurally invariant region. This procedure entailed excluding those residues with the largest positional differences (measured as an ellipsoid of variance determined from the Cartesian coordinates for equivalent Cα atoms of each frame), before each round of superposition, until only the invariant “core” residues remained [18]. This structurally invariant core consists predominantly of residues within secondary structure elements and was used as the reference frame for superposition of both crystal structures and subsequent MD trajectory snapshots. PCA was then employed to further examine inter-conformer relationships. The application of PCA to MD trajectories, along with its ability to provide considerable insight into the nature of conformational differences in a range of protein families has been previously discussed [14]. Briefly, PCA is based on the diagonalization of the covariance matrix, C, with elements Cij, built from the Cα atom Cartesian coordinates, r, of the superposed trajectory frames:(3)where i and j represent all possible pairs of 3N Cartesian coordinates, where N is the number of atoms being considered. The highly mobile N and C-terminal residues (positions 42–43 and 380–398) were excluded from analysis as their high intrinsic mobility was found to mask the separation of the more pertinent open-to-closed domain displacements. The eigenvectors of the covariance matrix correspond to a linear basis set of the distribution of structures, referred to as principal components (PCs), whereas the eigenvalues provide the variance of the distribution along the corresponding eigenvectors. Projection of the distribution onto the subspace defined by the largest principal components results in a lower dimensional representation of the structural dataset. The percentage of the total mean-square displacement (or variance) of atom positional fluctuations captured in each dimension is characterized by their corresponding eigenvalue.
Clustering of trajectory conformers was used to visualize the dominant conformations sampled by each simulation. Structures from aMD simulations underwent average-linkage hierarchical clustering according to the pairwise distances obtained from their projection onto the first 3 principal components. Clustering based on pairwise RMSD yielded similar major clusters. However, a significantly larger number of small clusters were returned due the influence of TcPRAC's highly flexible termini that do not contribute to the dominant principal components. Note that PCs 1–3 account for ∼70% of the variance in the original distribution (Figure 6) and produce a more succinct distance measure than the examination of average all-atom distances. This metric aids interpretation of an otherwise noisy signal as it is derived primarily from the concerted displacement of subdomains relative to one-another.
Inspection of the resulting dendogram was used to partition structures into 20 dominant groups (ranked according to their populations). The closest structure to the average structure from each cluster, in terms of RMSD, was chosen as a representative for further fragment mapping and virtual screening analysis described below.
To assess the level of sequence conservation at each position in the alignment, the similarity, identity, class identity and entropy per position were calculated. The “similarity” was defined as the average of the similarity scores of all pairwise residue comparisons for that position in the alignment (where the similarity score between any two residues is the score value between those residues in the BLOSSUM 62 substitution matrix [24]). The “identity” (i.e. the preference for a specific amino acid to be found at a certain position) was assessed by averaging the identity scores resulting from all possible pairwise comparisons at that position in the alignment (where all identical residue comparisons are given a score of 1 and all other comparisons are given a value of 0). The “class identity” was calculated in a similar manner to the “identity”. The exception being that amino acids were considered class identical (i.e. assigned a score of 1) if they possessed similar physicochemical properties. For this analysis residues were partitioned into three classes based on their relative hydrophobicity and the extent to which they are distributed between the surface and interior of known globular aqueously soluble protein structures (see [25], [26], and references therein). The first class contains hydrophobic residues (C, V, L, I, M, F and W) that have a high probability of residing within protein interiors. The second class contains hydrophilic residues (R, K, E, D, Q and N) that are most likely to be found on the surface of proteins. Finally, the third class is comprised of neutral residues (P, H, Y, G, A, S and T) that have a roughly equal chance of being on the surface or in the interior. “Entropy” is based on Shannon's information entropy for both a 21-letter alphabet (20 amino acids and a gap character) and a 7-letter alphabet (6 groups of amino acids and a gap character) [27], [28] (Equation 4):(4)where S is Shannon's entropy, pi is the frequency of each alphabet's letter at position i and N is the alphabet's size (7 or 21). The six groups chosen were aliphatic (A, V, L, I, M and C), aromatic (F, W, Y and H), polar (S, T, N and Q), positive (K and R), negative (D and E), and finally special conformations (G and P). Entropy scores plotted in Figure 3 are normalized so that conserved (low entropy) columns score 1 and diverse (high entropy) columns score 0 (Equation 5):(5)where, C is the normalized entropy, pi is the frequency of each alphabet's letter at position i, N is the alphabet's size and Nseq is the length of the sequence. We define a position to be conserved if the similarity, identity, class identity entropy 21 or entropy 7 at a position is >0.6. Positions in which more than 30% of the sequences had gaps were excluded from all sequence conservation analysis.
Percent solvent exposure per position was calculated with the NACCESS program available at http://www.bioinf.manchester.ac.uk/naccess/. A residue was considered exposed when the accessible surface area (ASA) of the residue was more than 40% of the measured ASA of that residue in an extended G-X-G tripeptide context.
We employed the coarse-grained AD-ENM normal mode analysis approach developed by Zheng et al. [19]. AD-ENM implements a single-parameter Hookean potential, which has previously been shown to yield low-frequency normal modes that are in good agreement with those obtained from more detailed, empirical, force fields. For further details see [19], [20]
We used the FTMap method of Brenke and co-works to highlight regions on the TcPR surface that have the potential to bind the highest number of small molecular probes [21]. Both crystal structures and each cluster representative form aMD were subject to fragment mapping. Hot-spot residues (those that comprise prominent fragment binding sites) were analyzed across all structures. A residue was assumed to be in contact with a probe molecule if any two heavy atoms from the probe and residue were closer than 5.0 Å.
|
10.1371/journal.pgen.1002334 | Relative Burden of Large CNVs on a Range of Neurodevelopmental Phenotypes | While numerous studies have implicated copy number variants (CNVs) in a range of neurological phenotypes, the impact relative to disease severity has been difficult to ascertain due to small sample sizes, lack of phenotypic details, and heterogeneity in platforms used for discovery. Using a customized microarray enriched for genomic hotspots, we assayed for large CNVs among 1,227 individuals with various neurological deficits including dyslexia (376), sporadic autism (350), and intellectual disability (ID) (501), as well as 337 controls. We show that the frequency of large CNVs (>1 Mbp) is significantly greater for ID–associated phenotypes compared to autism (p = 9.58×10−11, odds ratio = 4.59), dyslexia (p = 3.81×10−18, odds ratio = 14.45), or controls (p = 2.75×10−17, odds ratio = 13.71). There is a striking difference in the frequency of rare CNVs (>50 kbp) in autism (10%, p = 2.4×10−6, odds ratio = 6) or ID (16%, p = 3.55×10−12, odds ratio = 10) compared to dyslexia (2%) with essentially no difference in large CNV burden among dyslexia patients compared to controls. Rare CNVs were more likely to arise de novo (64%) in ID when compared to autism (40%) or dyslexia (0%). We observed a significantly increased large CNV burden in individuals with ID and multiple congenital anomalies (MCA) compared to ID alone (p = 0.001, odds ratio = 2.54). Our data suggest that large CNV burden positively correlates with the severity of childhood disability: ID with MCA being most severely affected and dyslexics being indistinguishable from controls. When autism without ID was considered separately, the increase in CNV burden was modest compared to controls (p = 0.07, odds ratio = 2.33).
| Deletions and duplications, termed copy number variants (CNVs), have been implicated in a variety of neurodevelopmental disorders including intellectual disability (ID), autism, and schizophrenia. Our understanding of the relevance of large, rare CNVs in a range of neurodevelopmental phenotypes, varying in severity and prevalence, has been difficult because these studies were restricted to the analysis of one disorder at a time using different CNV detection platforms, insufficient sample sizes, and a lack of detailed clinical information. We tested 1,227 individuals with different neurological diseases including dyslexia, autism, and ID using the same CNV detection platform. We observed striking differences in CNV burden and inheritance characteristics among these cohorts and show that ID is the primary correlate of large CNV burden. This correlation is well illustrated by a comparison of autism patients with and without ID—where the latter show only modest increases in large CNV burden compared to controls. We also find significant depletion in the frequency of large CNVs in dyslexia compared to the other cohorts. Further studies on larger sets of individuals using high-resolution arrays and next-generation sequencing are warranted for a detailed understanding of the relative contribution of genetic variants to neurodevelopmental disorders.
| Recent studies have implicated large, rare CNVs in a range of neurodevelopmental disorders including intellectual disability (ID) [1], [2], autism [3], [4], schizophrenia [5], [6], bipolar disorder [7], [8], epilepsy [9], [10], and attention deficit hyperactivity disorder (ADHD) [11], [12]. Several themes have emerged from these studies: first, a significant enrichment for rare CNVs in individuals with the disease compared to unaffected controls was observed, independently, for each of these disorders; second, the same recurrent CNVs are associated with different neuropsychiatric phenotypes; and third, locus heterogeneity is substantial as many distinct variants can lead to similar phenotypes.
Our understanding of the relevance of rare CNVs across a broad spectrum of neurodevelopmental disorders, varying in severity and prevalence, is limited as previous studies were restricted to the analysis of one phenotype at a time and each of such studies was performed using different CNV genotyping methodologies with distinct platform-specific biases, making comparisons difficult. We undertook a systematic analysis of 1,227 cases and 337 controls to assess the relative contribution of CNVs in three phenotypically distinct neurodevelopmental disorders. We designed a whole-genome custom microarray targeted to genomic hotspots for comparative genomic hybridization (CGH) to identify potentially pathogenic CNVs that contribute to ID, autism, and dyslexia.
We analyzed 1,227 individuals ascertained for three neurodevelopmental disorders: 376 dyslexic children with a verbal IQ (VIQ) ≥90 on the Wechsler Intelligence Scale for Children [13] and dyslexia defined as poor performance and IQ-performance discrepancy in one or more of a set of standardized reading measures, 350 cases with sporadic autism from the Simons Simplex Collection (SSC), and 501 cases with ID. We used 337 NIMH control individuals for comparison. Further, based on the presence or absence of ID (full-scale IQ score cutoff of 70), autism cases were divided into those with ID (n = 97) or without ID (n = 253) (see Materials and Methods). Based on the presence of multiple congenital anomalies (MCA), individuals with ID were divided into those with ID only—i.e. idiopathic ID (n = 428)—and those with ID and MCA (n = 73).
All copy number variation analyses were performed using a custom microarray with a high probe density (∼2.6 kbp) targeted to 107 genomic hotspot regions [14] (∼251 Mbp) and a median probe spacing of ∼36 kbp in the genomic backbone (see Materials and Methods, Table S1). We used a Hidden Markov Model (HMM)-based algorithm to identify deletions and duplications. We restricted our analysis to CNVs >50 kbp in size to reduce false positive calls and validated all relevant CNVs using a second custom designed high-density array. To empirically determine the validation rate of the array at different genomic regions, we examined 118 CNVs detected in 24 samples and confirmed 117 events (>99% accuracy, see Table S2). While we were easily able to detect smaller events in the hotspot regions, the specificity of the array restricted our CNV discovery to >50 kbp in hotspot-associated regions and to >300 kbp in regions not associated with genomic hotspots (Figure S1).
After quality control (QC) filtering and manual curation, we obtained 5,086 CNVs in 1,395 out of 1,564 individuals (89.2%) with high-quality array CGH data (Table 1; Datasets S1, S2, S3, S4). Using these data, we compared the CNV enrichment between the multiple cohorts tested. We found a significant excess of large CNVs (>1 Mbp) in individuals with ID (p = 2.75×10−17, odds ratio = 13.71) or autism (p = 0.012, odds ratio = 2.99) when compared to controls analyzed on the same microarray platform (Figure 1). The frequency of large CNVs among children with dyslexia was similar to controls (p = 0.64, odds ratio = 0.94), although this might indicate a lack of statistical power in our study to detect any subtle enrichment (power >0.8 to detect 4.2% increase in burden) for large CNVs in dyslexia.
Within the neurodevelopmental disorder cohorts, a comparison showed a significantly greater large CNV burden in individuals with ID compared to autism (p = 9.58×10−11, odds ratio = 4.58) or dyslexia (p = 3.81×10−18, odds ratio = 14.45). When we partitioned the ID cohort into subsets with and without MCA, we observed a significantly increased large CNV burden in individuals with ID/MCA compared to ID alone (p = 0.001, odds ratio = 2.54). This trend was also observed when individuals with autism were separated into those with ID and without ID (Figure 1), although not statistically significant (p = 0.102, odds ratio = 2.1). When compared to controls, we noted a trend for increase in large CNV burden for autism without ID (p = 0.07, odds ratio = 2.33) as well as autism with ID (p = 0.0048, odds ratio = 4.85). In addition, a gene-based analysis showed an incremental increase in the proportion of disrupted genes and average gene density per CNV with higher estimates for the ID/MCA cohort as compared to ID alone or autism (Table 1). We also note that within the cohorts no bias towards deletions or duplications was observed in relation to phenotypic severity or variability (Tables S3, S4, S5, S6, S7). Overall, our results suggest a positive correlation of the severity of the phenotype to the size and gene density of CNVs.
To identify rare CNVs of likely pathogenic significance, we compared the pattern of CNVs from dyslexia, autism, ID, and NIMH control cohorts to a map developed from an expanded set of 8,329 normal individuals genotyped with Illumina microarrays and to the publicly available Database of Genomic Variants [15] (see Materials and Methods). We eliminated common copy number polymorphisms and CNVs from our cases if they had a reciprocal overlap of 50% or more of their length with CNVs found in these 8,329 controls. After filtering, we compared the groups. We found a significant increase of rare CNVs in individuals with autism (35/336, 10%; p = 2.4×10−6, odds ratio = 6) or ID (69/431, 16%; p = 3.55×10−12, odds ratio = 10) compared to individuals with dyslexia (6/322, 2%) (Figure 2A, Table 2). In fact, when analyzed separately, the frequency of rare CNVs in NIMH controls (6/306, 2%) was not different compared to dyslexia (p = 0.57, odds ratio = 0.94) (Table S8).
Given the high population prevalence of dyslexia [16], we then relaxed our selection to include events present at an allele frequency of <0.1% in controls (8/8,635) and identified four additional CNVs—i.e., a total of 10 CNV events (Table 3). The analysis of hotspot regions identified only one individual with dyslexia who carried a 15q11.2 BP1–BP2 deletion, which has previously been associated with ID [17], schizophrenia [18], [19], and epilepsy [20]; however, this deletion was also observed in 25/8,635 of our total control individuals. None of the seven deletions and three duplications detected in our dyslexia cohort mapped to candidate loci known to be associated with dyslexia [21].
Analysis of 336 individuals from the SSC autism cohort showed that 35 individuals (10%) carried 36 rare CNVs (680 RefSeq genes, median size = 662 kbp) and about 58% (21/36) of these CNVs mapped to genomic hotspots (Table 2). Only eight of the events (all hotspot sites) associated with genomic disorders, including 22q11.2 deletion (TBX1, DiGeorge syndrome), 17p12 duplication (PMP22, Charcot-Marie-Tooth disease), and 15q11.2q13.1 duplication (UBE3A and SNRPN) (Table S9). In addition, as reported previously [3], [22], [23], the autism-associated proximal 16p11.2 deletion (TBX6) was observed in approximately 1% (3/336) of all autism cases analyzed. Interestingly, one case with a de novo 16p11.2 deletion also inherited a 2 Mbp duplication 22q11.2 (TBX1) from the mother.
Among 431 cases with ID (358 cases with ID only and 73 cases with ID plus MCA), 69 individuals carried 77 rare CNVs (2,215 RefSeq genes, median size = 1.5 Mbp) that were either of known pathogenic significance or not observed in a total set of 8,635 controls, and 32% (25/77, median size = 1.42 Mbp) of these variants localized to genomic hotspot regions (Table 2). This is a significant enrichment for rare CNVs in the ID cohort compared to autism (p = 0.019, odds ratio = 1.6) or dyslexia cohorts (p = 3.55×10−12, odds ratio = 10). Interestingly, 20/77 CNVs (16 hotspot and four non-hotspot sites) mapped to a known genomic disorder site, including those associated with variable phenotypes such as 15q13.1q13.3 (CHRNA7), 16p11.2 proximal (TBX6; two cases) and distal (SH2B1) hotspots, 16p13.11 (MYH11; three cases), 17q12 (TCF2), and 3q29 (DLG1) as well as syndromic regions such as 7q11.23 (Williams syndrome), 17q21.31 (MAPT), 5q35 (Sotos syndrome), 8p23.1, 22q13 (Phelan-McDermid syndrome) [24] and 1p36 [25].
We next sought to determine whether these rare CNVs were inherited or if they arose de novo in the probands. Parental DNA samples were available to investigate inheritance for 90 out of 123 rare CNVs detected in all three disease cohorts (Table S8). In four cases, only maternal DNA was available. We find that 44/90 CNVs arose de novo and a majority (77%, 34/44) of these de novo CNVs were large (>1 Mbp). Overall, we find a greater proportion of de novo events in ID (64%, 30/47) compared to autism (40%, 14/35; p = 0.027, odds ratio = 2.6) or dyslexia (0/8; p = 0.0009, odds ratio = infinity) (Figure 2B). These data are suggestive of a general trend of increased de novo rates and CNV size with increased severity of the disorder.
We then focused on rare CNVs involving single genes or regions of potential interest. In the dyslexia cohort, two unrelated families carried CNVs on chromosome 7q11.23 that involved the autism susceptibility candidate 2 (AUTS2, MIM# 607270). A 669 kbp duplication that included AUTS2 and WBSCR17 was transmitted from an affected father to the daughter and an approximately 84 kbp deletion was transmitted from the affected paternal grandmother through the unaffected father to the proband (Figure 3). In addition, we also identified a 354 kbp deletion encompassing AUTS2 in one individual with idiopathic ID, pervasive developmental delay, partial epilepsy, and left hemihypertrophy. An approximately 1.2 Mbp deletion encompassing the eyes shut drosophila homolog gene (EYS, MIM# 612424) on chromosome 6q12 was detected in an affected proband and several unaffected family members. Although autosomal recessive single-nucleotide mutations in EYS have been reported in patients with retinitis pigmentosa [26], [27], the role of heterozygous microdeletions involving this gene is unknown.
We also identified a 471 kbp deletion involving IMMP2L inherited by the proband from the affected mother (Figure 3). Deletions involving IMMP2L have been associated with ADHD [12], autism [28], and Tourette syndrome [29]. Recently, Pagnamenta and colleagues also reported a 594 kbp IMMP2L-DOCK4 deletion resulting in a fusion transcript and an intragenic DOCK4 deletion segregating with dyslexia [30]. Our results are best interpreted within the context of candidate gene identification in dyslexia. Although at least nine chromosomal loci are associated with dyslexia, for two of these loci the candidate genes were identified on the basis of a rare balanced chromosomal translocations disrupting ROBO1 [31] and DYX1C1/EKN1 [32], [33]. More recently, a Danish Cytogenetic Registry study of all cases with chromosomal translocations identified additional novel dyslexia candidate genes affirming the value of rare structural variants in understanding the genetics of dyslexia [34]. Our study is the first to systematically characterize rare CNVs in dyslexia and thus evaluate the contribution of rare deletions and duplications to this common genetic disorder.
Within the autism cohort, several novel deletions and duplications involving neurologically-relevant genes were identified. A 5 Mbp de novo deletion involving FOXP1 on chromosome 3p14.1 was identified in an individual with features of idiopathic autism (full-scale IQ = 75). An additional 6.6 Mbp de novo deletion overlapping FOXP1 was also identified in an individual with idiopathic ID (Figure 4). A review of the DECIPHER database revealed a similar-sized deletion disrupting FOXP1 in an individual with developmental delay, sensorineural deafness, hypotonia, club foot, and dislocation of hip. Recently, FOXP1 was implicated in autism, ID, and language impairment [35], [36], [37]. It is believed that FOXP1interacts with FOXP2 and CNTNAP2, both implicated in speech disorders and autism [38], [39]. The overlapping 1.16 Mbp region of the deletion common to both autism and ID indicates a potential involvement of FOXP1 in pathways related to both of these disorders. Other variants involving functionally relevant genes include 7q36.2 deletion and duplication (DPP6), 17q23.3 duplication (SCN4A), and 17q21.32 duplication (WNT3 and WNT9B).
Our analysis of the ID cohort was enriched for singleton events often involving genes related to developmental or neurological functions including SYNPR, GABRA, AUTS2, FOXP1, FKBP6, COBL, and FMR1. However, within the same cohort we were also able to detect novel overlapping deletions (3.5 Mbp and 9 Mbp) on chromosome 9p24 in two unrelated individuals (Figure 5A). Both cases exhibited clinical features of ID and Pervasive Developmental Delay-Not Otherwise Specified. The distal breakpoints of these deletions map to segmental duplications while the proximal end maps within a high density of repeat elements. A survey of this region in the DECIPHER database [40] revealed about 15 cases with overlapping deletions. Variable clinical presentations and heterogeneity of deletion breakpoints preclude further genotype-phenotype correlation studies for this region (Figure S2). We also identified a nonrecurrent 6q16 deletion (chr6: 100,383,567-103,310,184) that potentially narrows the critical region for this recently described Prader-Willi-like syndrome [41] to approximately 2.9 Mbp. The refined critical region contains only five genes including the obesity-associated SIM1 [42] and the autism-associated GRIK2 [43] (Figure 5B). About 70% of children with 6q16 deletion manifest obesity [41]; however, our case with the smaller deletion, encompassing SIM1, showed no evidence of obesity at 10 years of age (Table S10).
We find that 8/69 (11.6%) cases in the ID cohort carried more than one large, rare CNV and all of these individuals presented with severe clinical features (Tables S6, S7, S8). A striking difference (p = 0.008, odds ratio = 11.2) in multiple CNV rates was also observed when the ID cohort was divided into those with severe MCA (44%) and those with idiopathic ID (6.7%). Notable examples are co-occurrences of a 3.4 Mbp 16p13.11 duplication and a 3.3 Mbp deletion on chromosome 4q25 involving PITX2 in a case with features of Rieger syndrome [44] (Figure 6) and a 17p13.3 deletion (YWHAE, Miller-Dieker syndrome) and 3q29 duplication (DLG1) in a child with cryptorchidism, ventricular septal defect, and seizures. The 3q29 duplication is a recurrent interstitial rearrangement [45], [46] potentially mediated by flanking segmental duplications of high sequence identity (27 kbp size, 96% identity). The 17p13.3 deletion is a previously reported nonrecurrent rearrangement associated with Miller-Dieker syndrome [47], [48]. In contrast, only 1/35 (2.9%) autism cases and none of dyslexia individuals carried another large CNV. This observation suggests that the severity of the phenotypes can be influenced by more than one large, rare CNV co-occurring in the same individual. During this analysis we considered the possibility of a derivative chromosome representing an unbalanced translocation possibly creating the impression of multiple CNVs in our cases. We carefully reviewed available chromosomal analysis data (G-banded karyotyping or FISH) for each of the individuals with two hits reported in our study. We did find one case with two hits where apparent CNVs represent a derivative chromosome inherited from a balanced translocation carrier parent (Table S8D).
Initial discoveries of significant enrichment of rare CNVs for ID and autism led to testing the CNV basis for other behavioral and neurodevelopmental disorders of varying population frequency and severity, such as schizophrenia, ADHD, epilepsy, bipolar disorder, and Tourette syndrome. However, comparisons between these studies have been difficult due to differences in study design, insufficient sample sizes, and lack of detailed phenotype information. In this study, we compared 1,564 individuals (cases and controls) on a single platform of relatively modest density with the same type of detection bias. We utilized the duplication architecture of the human genome to custom design a DNA oligonucleotide microarray enriched for genomic hotspots, i.e., regions flanked by high-identity segmental duplications. This array has an advantage over several other commercial arrays in that there is a 25-fold enrichment for recurrent events in the genomic hotspots compared to the rest of the genome [49]. Therefore, fewer samples are required to identify several unrelated individuals with the same pathogenic mutation. We find that our array has a comparable diagnostic yield of 16% for the ID cohort compared to other clinical chromosomal microarray studies reported in the literature (Figure S3).
In strong agreement with previous studies, our data suggest that multiple, rare CNVs contribute to the etiology of autism and ID. In contrast, we find no increase in large pathogenic CNVs in individuals with dyslexia compared to controls. Notwithstanding, our analysis revealed novel regions of potential relevance to the etiology of dyslexia. Two unrelated children (2/322, 0.6%) with dyslexia carried CNVs encompassing AUTS2, both inherited from a parent. While the phenotype of dyslexia segregated with the AUTS2 duplication in the first family (Figure 3A), in the second family the deletion was inherited from affected grandmother through unaffected father (Figure 3B). This could be due to a phenomenon described as “compensation”, where some adults that reported difficulties with reading in childhood no longer evidence signs of dyslexia [50]. Previous studies of de novo chromosomal translocations and inversions identified breakpoints within AUTS2 in individuals with autism and/or ID phenotypes [51], [52], [53], [54], [55]. More recently, unique AUTS2 deletions and duplications were observed in Juvenile Myoclonic Epilepsy [10] and ADHD [11], [12]. It is interesting to note that ADHD and dyslexia are frequently comorbid and may have shared genetic risk factors [56], [57], [58]. When our study is taken together with recent CNV studies of ADHD [11], [12], AUTS2 CNVs were observed in 5/2,306 combined cases and 3/46,947 unscreened controls (p = 1.12×10−5, odds ratio = 33.9), indicating that AUTS2 might have an important role in pathways related to cognition. While the function of AUTS2 is still unclear, it is strongly expressed in fetal and adult brains, particularly in the frontal, parietal, and temporal lobes [59]. Interestingly, AUTS2 and the 7q11.2 region were identified as having the strongest statistical signal for positive selection in early modern humans as compared to the Neanderthal genome [60], indicating that AUTS2 might be important for a specialized human function such as cognition.
The CNV profile we observed in individuals with dyslexia was essentially the same as that in control individuals. This is not surprising if we take into consideration that all the subjects in our dyslexia sample had a VIQ above the 25%ile and the mean VIQ of the cohort was 110 (2/3 standard deviations above the general population mean), and given that we have shown that the CNV profile correlates with the severity of ID. The genes involved in dyslexia are likely to affect more specialized cognitive functions, may not adversely affect general intelligence, and may be more amenable to discovery with high-density arrays capable of detecting single gene or single exon CNVs or SNP microarrays that can leverage SNP allele frequency information in addition to signal intensity. In addition, all hybridization-based platforms fail to detect copy number neutral changes, such as balanced chromosomal rearrangements and inversions. This is particularly germane to dyslexia where a large number of candidate genes have been identified through mapping of translocation breakpoints [21], [34].
A comparison of rare de novo CNV rates for autism shows that our estimates (4%, 14/336) fall within a range of 4–10% reported previously by other large-scale, high-density array studies [4], [61], [62], [63]. This suggests that no platform-specific bias exists for large variants and also that the contribution of large CNVs is consistent across all studies for autism. We find a significantly greater enrichment for large CNVs, higher de novo rates, and a higher frequency of two rare CNV hits in individuals with ID-associated phenotypes compared to autism or dyslexia. This observation is exemplified by the fact that individuals with autism with ID have more large CNVs than those with autism only. We also find a significant difference between individuals with autism versus those with dyslexia. Sanders and colleagues recently analyzed 1,124 SSC families affected with autism spectrum disorder (ASD). Using stepwise linear models, they evaluated the relationship between intellectual functioning, sex, and the number of genes within rare, de novo CNVs. While the number of genes affected correlated with the size of the de novo CNV, the authors did not find a strong correlation of the Autism Diagnostic Observation Schedule (ADOS) combined severity score (p = 0.25, R2 = 0.005) or full-scale IQ (p = 0.02, R2 = 0.08) with the size of the CNV. In contrast, we considered all large CNVs (common and rare, de novo and transmitted) identified in a relatively smaller sample size and essentially bifurcated the autism cohort using a full-scale IQ score cutoff of 70. There was also a greater enrichment of two hits in the ID cohort (11.6%) compared to the autism cohort (2.8%). In fact, one individual carrying a 16p11.2 deletion with autism and features of ID also has a maternally inherited 22q11.2 duplication (TBX1) providing further evidence for the two-hit hypothesis we previously proposed for severe developmental delay [64]. Further, the frequency of two hits was even more striking when only individuals with ID/MCA were considered (44%), albeit the number of cases is few. We believe these data provide support for an incremental effect of CNV size and number on the severity of phenotypic outcome.
Our experimental design is biased towards interrogating hotspot regions in the human genome. A comparison to recently reported studies [61], [63] suggests that the majority of false-negative calls will reside within non-hotspot regions due to a lack of probe coverage (<10 probes). While the detection power of our array increases with the size of the variant, we would certainly miss smaller and intragenic CNVs, for example in autism candidate genes such as NRXN1 [65], [66], CACNA1C, SLC4A10, MAGI1 [63], SYNGAP1, DLGAP2 [62], NLGN1, ASTN2 [67], and exonic copy number variants in ASPM, DPP10, CNTNAP2, A2BP1, PCDH9 [68], and PTCHD1 [69]. While we find no excess of large CNVs in dyslexia, there is still the possibility that large CNVs are relevant in some familial cases of the disease as well as occasional sporadic cases. Further studies are warranted for a more detailed analysis of all the three neurodevelopmental cohorts using high-resolution arrays and next-generation exome and/or whole-genome sequencing. While it can be difficult to compare data derived from different microarrays, there is value to multiple array platforms and cross-platform validation. The depositing of the resulting data into publicly available databases will facilitate the continued elucidation of recurring clinically significant CNV and genotype-phenotype correlations.
Patients from each of study cohort were recruited after appropriate human subjects approval and informed consent. Informed consent was also obtained to publish photographs.
DNA samples were obtained from cases ascertained for three neurodevelopmental disorders of varying severity: (1) ID/developmental delay and MCA, (2) dyslexia or reading impairment, and (3) idiopathic autism. We defined severity of clinical features based on presence or absence of ID (IQ<70) for the autism group and congenital malformation for the ID group. Our dyslexia cohort had no ID or congenital malformation cases; as an IQ≤90 and the presence of congenital malformations were exclusion criteria. Individuals with idiopathic autism were partitioned into those with autism and ID (IQ<70) and those without ID (IQ>70). For the ID cohort, those individuals with brain malformations, gross craniofacial dysmorphology, cardiac defects, and neurological deficits were separated into an ID plus MCA (ID/MCA) group. Thus, in the order of severity, the ID/MCA cohort is considered the most severe, followed by ID only, autism with ID, autism without ID, dyslexia, and normal controls. However, we note that although the individuals with dyslexia do not have ID, they have severe impairments in core phonological measures leading to significantly reduced reading abilities despite normal IQ (IQ≥90). Detailed descriptions of each of the cohorts are given below.
For the dyslexia subject set, children were considered eligible for the study if they met researcher-defined criteria based on test scores from a standardized battery of tests. DNA samples were obtained from two cohorts. The first cohort included probands aged 6 to 16 from 198 families who were initially ascertained at the University of Washington (UW) multidisciplinary Learning Disability Center (UWLDC) under protocols approved by the UW Institutional Review Board. For the UWLDC cohort, probands were required to have a prorated VIQ at or above 90 (≥25%ile) on the Wechsler Intelligence Scale for Children – 3rd edition [13], with performance below the age-specific population mean and at least one standard deviation below the VIQ on one or more out of 10 research measures of reading, writing, or spelling. As a group, on average, probands met the impairment criteria between 6 and 7 measures. As expected by ascertainment requirements, the average VIQ of probands in this cohort was 110 (≥75%ile) [70], [71]. Siblings older than 6.5 years were invited to participate, and additional family members were added using a sequential sampling strategy to extend pedigrees through family members with the most extreme impairment values on the same 10 research measures. Detailed recruitment and evaluation procedures for the UWLDC cohort were described earlier [50], [72] (see Table S4A).
For the second cohort, 178 children aged 5 to 12 were recruited from a special K-6 school for students with dyslexia or via their direct relatives in the Atlanta area (The Schenck School, Atlanta, GA). For this cohort, children were required to have a psychological battery of tests completed by a licensed psychologist and usually have a diagnosis of a reading disability. Based on strong verbal comprehension score, perceptual reasoning score, Peabody picture vocabulary test, or other cognitive tests that measure intelligence, these children have average to above-average intelligence. Both cohorts were composed of individuals with >90% Caucasian ethnicity with an approximately equal number of males and females. Except for ADHD, children with other psychiatric and neurological disorders, moderate to severe receptive language disorders, developmental disabilities, or other conditions known to affect cognition were excluded based on parental questionnaire. Clinical details are shown in Table S4B.
For the autism cohort, families were identified through the SSC (www.sfari.org) [73]. The Simons Foundation-funded SSC includes families with no more than one child with autism ascertained through 12 data collection sites across North America. Of the 350 individuals included in this study 297 (85%) are of Caucasian ethnicity. Inclusion criteria in the collection requires that the child with autism meet ASD criteria on the ADOS [74], on the Autism Diagnostic Interview, Revised (ADI-R) [75], and meet expert clinical judgment. Nonverbal IQ estimate must also be greater than 35. Children with significant hearing, vision, or motor problems, significant birth complications (e.g. extended NICU stay), or with a diagnosis of ASD-related disorders, such as Fragile X, were excluded. Children with a relative (up to third degree) with ASD or sibling who showed ASD-related symptoms were also excluded. Diagnostic evaluations, cognitive assessment, and phenotypic characterization were conducted at each site with data collection, data entry, and data validation methods standardized across sites to ensure reliability of sample collection. We further partitioned the autism cohort into those associated with ID (average full scale IQ = 49) consisting of 97 cases (73 males, 24 females; median age, 12 years) and those without ID (average full scale IQ = 98.9) comprising 253 cases (228 males, 25 females; median age, 11 years and 11 months). Clinical details are shown in Table S5.
The idiopathic ID cohort was selected from individuals admitted to the IRCCS Associazione Oasi Maria Santissima and screened for ID according to the Diagnostic and Statistical Manual of Mental Disorders-IV-Text Revision (DSM-IV-TR) criteria. This cohort consists of 428 cases (153 females, 275 males; median age, 15 years) of Caucasian ethnicity with idiopathic ID and previously excluded for common causes of ID, including Fragile X syndrome, trisomies 21 and 13. In addition, classical genomic syndromes such as Smith-Magenis, DiGeorge, Prader-Willi/Angelman, and Williams syndromes, if recognized by clinical evaluation, were followed up for confirmation using targeted multiplex ligation-dependent probe amplification and excluded. We note that cases with phenotypic variability that escape clinical detection might not have been excluded. Typically, idiopathic cases of ID with no classical constellation of clinical features suggestive of a known disorder or those with mild to moderate ID without significant congenital malformation were included in this cohort. Clinical details are shown in Table S6.
Individuals with features of ID with MCA not necessarily assigned to a specific syndrome were evaluated and recruited at the University of Torino. This cohort consists of 73 individuals (32 females and 41 males) of Caucasian ethnicity with a median age of 2 years at diagnosis. Clinical features of these individuals included brain malformations, craniofacial dysmorphology, and neurological deficits along with variable ID (Table S7). Informed consent was obtained from all the subjects included in both the studies.
The control cohort consisted of 337 DNA samples obtained from the Rutgers University Cell and DNA Repository (www.rucdr.org). These individuals were ascertained by the NIMH Genetics Initiative [76] through an online self-report based on the Composite International Diagnostic Instrument Short-Form (CIDI-SF) [77] and screened specifically for eight mental health disorders, including major depression, bipolar disorder, and psychosis, but were not screened for dyslexia and therefore not ideal for such comparisons. Those who did not meet DSM-IV criteria for major depression, denied a history of bipolar disorder or psychosis, and reported exclusively European origins were included [78], [79].
Additionally, CNV data from 8,329 additional cell line and blood-derived controls were used to assess the frequency of our putative pathogenic CNVs in a larger population of neurologically normal individuals. These data were derived primarily from genome-wide association studies of non-neurological phenotypes. Although these data were not ascertained specifically for neurological disorders, they consist of adult individuals providing informed consent. Specifically, datasets from the following sources were included in our analysis: Human Genome Diversity Project [49], [80]; National Institute of Neurological Disorders and Stroke (NINDS) (dbGaP accession no. phs000089) [49], [81]; Pharmacogenomics and Risk of Cardiovascular Disease (PARC/PARC2) [82], [83]; parents of asthmatic children courtesy of Stephanie London [49]; Fred Hutchinson Cancer Research Center (prerelease data provided courtesy of Aaron Aragaki, Charles Kooperberg, and Rebecca Jackson as part of an ongoing genome-wide association study to identify genetic components of hip fracture in the Women's Health Initiative); InCHIANTI (data provided by InCHIANTI study of aging, www.inchiantistudy.net) [49], [84]; and the Wellcome Trust Case Control Consortium phase 2 (National Blood Service) [7]. All samples were genotyped on Illumina arrays using methodology described previously [49] [85] and either natively processed in hg18 or re-mapped after CNV calling (NINDS and PARC) to hg18 using the UCSC LiftOver tool (http://genome.ucsc.edu).
We designed custom targeted hotspot v1.0 arrays comprised of 135,000 probes (by Roche NimbleGen) with higher density probe coverage (median probe spacing 2.6 kbp) in the genomic hotspots (regions flanked by segmental duplications) and a lower probe density in the genomic backbone (median probe spacing 36 kbp). All microarray hybridization experiments were performed as described previously [86], using a single unaffected male (GM15724 from Coriell) as reference. All validation experiments were performed using two custom array designs: (1) a custom targeted 4×180 K Agilent chip with median probe spacing of 2 kbp in the genomic hotspots and whole-genome backbone coverage of one probe every 36 kbp (Agilent Technologies) and (2) a custom targeted 3×720 K NimbleGen or 2×400 K Agilent chip with median probe spacing of 500 bp in the genomic hotspots and probe spacing of 14 kbp in the genomic backbone.
All arrays were analyzed by mapping probe coordinates to the human genome assembly Build 36 (hg18). Using chromosome-specific means and standard deviations, normalized log intensity ratios for each sample were transformed into z-scores. These z-scores were then classified as “increased”, “normal”, or “decreased” in copy number using a three-state HMM. The HMM was applied using HMMSeg [87]. For each sample, HMM state assignments of probes were merged into segments if consecutive probes of the same state less than 50 kbp apart. If two segments of the same state were separated by an intervening sequence of ≤5 probes and ≤10 kbp, both segments and intervening sequence were called as a single variant. Further, we employed stringent QC measures and empirically estimated post-HMM filtering thresholds (absolute z-score >1.5 and >10 probes) to increase the specificity of our experiments. With these filtering criteria, we were able to thoroughly scan HMM outputs for CNV events and manually check the validity of each call by examining the normalized log intensity ratios across a chromosome. For the Agilent arrays, data analysis was performed following feature extraction using DNA analytics with ADM-2 setting according to the manufacturer's instructions. All CNVs calls were visually inspected in the UCSC genome browser.
First, we carried out validation on 24 samples from the developmental delay cohort and confirmed 117/118 HMM-inferred calls with a validation rate of 99.15%. Next, we validated 84 CNVs from an independent set of cases both validated using fluorescence in situ hybridization (FISH) and locus-specific custom high-density arrays [64] (Girirajan and Eichler, unpublished). We also validated all 44 calls from the autism cohort, 22 calls from the dyslexia cohort, and 78 calls from the developmental delay cohort. In addition, in an analysis of 517 individuals with epilepsy using this array design, 61/63 CNVs were validated on a different array platform [10].
|
10.1371/journal.pgen.1002831 | S Phase–Coupled E2f1 Destruction Ensures Homeostasis in Proliferating Tissues | Precise control of cell cycle regulators is critical for normal development and tissue homeostasis. E2F transcription factors are activated during G1 to drive the G1-S transition and are then inhibited during S phase by a variety of mechanisms. Here, we genetically manipulate the single Drosophila activator E2F (E2f1) to explore the developmental requirement for S phase–coupled E2F down-regulation. Expression of an E2f1 mutant that is not destroyed during S phase drives cell cycle progression and causes apoptosis. Interestingly, this apoptosis is not exclusively the result of inappropriate cell cycle progression, because a stable E2f1 mutant that cannot function as a transcription factor or drive cell cycle progression also triggers apoptosis. This observation suggests that the inappropriate presence of E2f1 protein during S phase can trigger apoptosis by mechanisms that are independent of E2F acting directly at target genes. The ability of S phase-stabilized E2f1 to trigger apoptosis requires an interaction between E2f1 and the Drosophila pRb homolog, Rbf1, and involves induction of the pro-apoptotic gene, hid. Simultaneously blocking E2f1 destruction during S phase and inhibiting the induction of apoptosis results in tissue overgrowth and lethality. We propose that inappropriate accumulation of E2f1 protein during S phase triggers the elimination of potentially hyperplastic cells via apoptosis in order to ensure normal development of rapidly proliferating tissues.
| Rapidly growing tissues provide an excellent opportunity to study the careful balance between cell proliferation and apoptosis needed for normal organ structure and function in developing organisms. We present evidence that a transcription factor critical for regulating progression of the Drosophila melanogaster cell cycle, E2f1, serves also as an indicator of normal tissue development. E2f1 activation during G1 phase of the cell cycle triggers entry into S phase. E2f1 activity is then rapidly inhibited during S phase by a mechanism that couples E2f1 proteolysis directly to DNA synthesis. Expression during larval development of an S phase-stabilized form of E2f1 results in apoptosis in rapidly proliferating adult wing precursor cells, even when this stabilized E2f1 protein is mutated such that it cannot induce transcription or cell cycle progression. Preventing the ability of S phase-stabilized E2f1 to induce apoptosis results in massive tissue overgrowth. We propose that aberrant E2f1 accumulation during S phase triggers apoptosis in order to remove potentially hyper-proliferative cells and to maintain homeostasis during tissue growth.
| During development, cells continually integrate extrinsic and intrinsic signals that control cell growth, proliferation and apoptosis. Mitogenic signals that drive growth and cell proliferation are balanced with apoptotic signals that eliminate damaged or unneeded cells. Genetic changes that inappropriately stimulate cell proliferation, reduce apoptosis, or both disrupt this homeostasis and result in aberrant development or neoplastic diseases like cancer [1]. Understanding the mechanisms that exist to maintain such homeostasis is thus an important area of investigation.
The balance between cell proliferation and cell death in growing tissues must ultimately function through key regulators of the cell cycle. These regulators include the E2F family of transcription factors, which control the expression of many genes responsible for cell proliferation, differentiation and apoptosis [2]. E2Fs are highly conserved proteins that act as either activators or repressors of transcription based on protein partners and structural features. As key mediators of cell proliferation and apoptosis, tight regulation of E2F activity is essential for normal development in mammals, flies, worms, and plants [2], [3]. The best-characterized mode of regulation involves members of the retinoblastoma (pRb) tumor suppressor protein family, which bind to and inhibit those members of the E2F family that dimerize with DP proteins [2]. In addition, pRb family/E2F complexes function as transcriptional repressors [4]. Loss of pRb function causes ectopic proliferation and apoptosis that is partially repressed by reducing E2F activity [5].
pRb family regulation of E2F occurs in quiescent cells and during G1 phase. Several pRb-independent mechanisms have been described that regulate activator E2Fs outside of G1, including Cyclin A/Cdk2-dependent phosphorylation of the DP subunit [6], [7], [8], SCFSkp2-directed proteolysis [9], 10, antagonism by the atypical E2F7 and E2F8 proteins [4], [11], [12], and binding to DP-4 [13]. These mechanisms are thought to down-regulate transcriptional activation by E2Fs during S phase or after DNA damage. In particular, disruption of Cyclin A/Cdk2 phosphorylation of E2F1 causes S phase defects and apoptosis in mouse cells, as does simultaneous loss of E2F7 and E2F8 [7], [8], [11]. In addition, E2F7/8 mutation in mice results in lethality, indicating that E2F7/8 play an essential role in the E2F regulatory network during development [11]. Mouse mutant genotypes that would specifically determine the contribution to development of Cyclin A/Cdk2 phosphorylation or the other modes of pRb-independent E2F inhibition have not been developed.
Here we examine the function of pRb-independent E2F regulation in developing Drosophila tissues, where E2F regulatory pathways are simpler than in mammals. While eight mammalian E2F genes encode nine distinct proteins (5 activators and 4 repressors), Drosophila encodes a single E2F activator (E2f1) and a single E2F repressor (E2f2), both of which bind the single Dp protein [2]. The primary cell cycle regulator is E2f1/Dp, which activates the transcription of replication factor genes and is negatively regulated by Rbf1, one of the two Drosophila pRb family members [14]. E2f1 mutant cells proliferate poorly [15], [16], [17], in part because of E2f2-mediated repression [18], [19]. Conversely, over-expression of E2f1 can drive cells into S phase [20], [21], [22]. E2f1 over-expression also induces apoptosis [17], [20], [21], and this may reflect the positive role E2f1 plays in developmentally controlled and DNA damage induced apoptosis [23], [24], [25], [26]. While many S phase and apoptotic transcriptional targets of E2f1 have been described [27], [28], the aspects of E2f1 regulation that coordinate the expression of these targets in rapidly growing tissues to achieve the proper balance of cell proliferation and apoptosis are not well understood.
In addition to the evolutionarily conserved pRb mode of activator E2F regulation, Drosophila E2f1 is inhibited by rapid destruction during early S phase [20], [29], [30], [31]. We recently determined that this S phase destruction is mediated by a “PIP degron" in E2f1 [32]. PIP degrons promote direct binding to DNA-loaded PCNA and the subsequent recruitment of the CRL4Cdt2 ubiquitin E3 ligase, thereby coupling proteolysis with DNA synthesis that occurs during S phase or after DNA damage [33], [34]. Drosophila E2f1 thus joined a small but growing number of proteins involved in genome duplication and maintenance that are regulated by CRL4Cdt2 [33], [34].
We previously demonstrated that expression of an S phase-stabilized E2f1 causes cell cycle acceleration, apoptosis, and developmental defects [32]. Because similar levels of wild type E2f1 expression, which is degraded during S phase, do not induce these phenotypes, we concluded that accumulation of E2f1 during S phase is poorly tolerated during development. However, we did not determine whether apoptosis and the developmental defects were a consequence of changes to the cell cycle in response to hyperactive E2f1 transcriptional activity, or to some other consequence of E2f1 accumulation during S phase. To explore this issue, we used assays in larval imaginal discs to understand the in vivo consequences of stabilizing E2f1 during S phase in developing tissues, focusing specifically on which activities of the E2f1 protein (e.g. DNA binding or Rbf1 binding) were responsible for the deleterious phenotypes resulting from stabilization during S phase.
We demonstrate here that the apoptosis and developmental defects caused by accumulation of E2f1 protein during S phase do not require E2f1's ability to induce transcription and cell cycle progression. Instead, apoptosis may occur via alleviation of Rbf1-dependent repression of the pro-apoptotic gene hid. We also show that simultaneously stabilizing E2f1 in S phase and blocking apoptosis results in extensive tissue overgrowth. We propose that inappropriate S phase accumulation of E2f1 protein in proliferating Drosophila cells triggers a form of proliferative stress, and that the cells experiencing this stress are consequently eliminated via apoptosis in order to prevent hyper-proliferation and maintain homeostasis during rapid tissue growth.
In order to examine the biological functions of CRL4Cdt2-mediated destruction of E2f1 during tissue growth and development, we examined larval wing imaginal discs, which grow from a ∼50 cell primordium to a ∼50,000 cell epithelial monolayer via canonical G1-S-G2-M cell division cycles and then differentiate into the adult wing during pupal development [17], [35]. Imaginal disc growth is highly tuned to modulate the balance between proliferation and apoptosis in response to particular stimuli. A dramatic example is the ability of wing discs to utilize “compensatory proliferation" in order to achieve normal wing development when as many as 50% of the disc cells have been killed via apoptosis following ionizing radiation [36]. This is possible because Drosophila apoptotic cells release mitogens such as Dpp and Wg that signal neighboring cells to begin proliferating and replace the dying cells [37], [38], [39]. We utilized this well characterized, rapidly proliferating tissue to examine the consequences of disrupting the normal S phase-coupled destruction of E2f1 (Figure 1A). We sought to determine the extent to which this destruction contributes to the balance between proliferation and apoptosis.
We previously established an assay for E2f1 destruction during S phase using flow cytometry of cultured Drosophila S2 cells expressing GFP-E2f1 fusion proteins [32]. In this assay, a mutation of E2f1 predicted to disrupt interaction with PCNA (GFP-E2f1PIP-3A) or a mutation predicted to abrogate CRL4Cdt2 binding (GFP-E2f1R161A) blocks S phase destruction (Figure S1A, S1B) [40]. We adapted this assay to wing imaginal discs in order to establish a quantifiable assay for measuring E2f1 destruction in vivo. We used engrailed-Gal4 (en-Gal4) to induce GFP or GFP-E2f1 fusion protein expression (e.g. “en-Gal4>GFP") in all cells of the posterior compartment of the disc (Figure S1C). Wing discs were dissected from third instar larvae, dissociated into individual cells by trypsin digestion, and subjected to flow cytometry after staining cells with a DNA binding dye [41]. We were able to directly compare the cell cycle profile of GFP-expressing posterior compartment cells to GFP-negative, anterior compartment control cells from the same tissue (Figure S1D–S1F). Because GFP is stable throughout the cell cycle, all posterior compartment S phase cells from en-Gal4>GFP discs were also GFP-positive (Figure S1D, S1G). In contrast, en-Gal4>GFP-E2f1 posterior compartment cells with an S phase DNA content were unlikely to be GFP-positive, because GFP-E2f1 is destroyed during S phase (Figure S1E, S1G). Only ∼12% of all GFP-E2f1 expressing cells in the posterior compartment were also in S phase, whereas ∼27% of GFP-expressing cells were in S phase (Figure S1G). This S phase destruction requires an intact PIP degron, as expression of GFP-E2f1PIP-3A resulted in an amount of GFP-positive posterior compartment S phase cells similar to GFP controls (Figure S1F, S1G). (For the rest of this manuscript we will refer to stabilized E2f1PIP-3A as E2f1Stable). These data extend our previously published wing disc experiments, in which we measured the effects of E2f1Stable expression on cell cycle progression by flow cytometry, but not directly on E2f1 destruction [32].
We previously showed that E2f1Stable expression accelerates cell cycle progression by using en-Gal4 to drive expression of GFP or GFP+GFP-E2f1 fusion proteins together in the posterior compartment of wing imaginal discs [32]. To measure such cell cycle effects for this study, we switched to co-expressing RFP with GFP or GFP-E2f1 fusion proteins (Figure S1C). By determining the number of RFP-positive cells in each phase of the cell cycle via DNA content, we can obtain a cell cycle profile of all posterior compartment cells. E2f1 stimulates cell cycle progression in wing imaginal disc cells by reducing the duration of G1 phase [17]. Therefore, by comparing the number of RFP-positive cells with G1 DNA content after expression of GFP or GFP-E2f1, we are able to quantify the extent to which E2f1 expression affects the cell cycle. For example, expression of either GFP-E2f1 or GFP-E2f1Stable caused a decrease in the percentage of cells in the population with a G1 DNA content compared to GFP expression alone (∼11% versus ∼28%, respectively; Figure S1H), indicating that both wild type and S phase-stabilized E2f1 proteins are equally able to increase the rate of wing disc cell cycle progression by reducing G1 length, as we previously described [32].
We previously demonstrated that in addition to an increase in the rate of cell proliferation, ectopic expression of E2f1Stable in wing imaginal discs caused an increase in apoptosis [32]. Interestingly, under the conditions of these experiments, expression of wild type E2f1 did not induce apoptosis although it did increase the rate of proliferation. We therefore hypothesize that E2f1Stable -induced apoptosis is not merely a consequence of increased cell proliferation resulting from excess E2f1 activity, but that the stabilization of E2f1 specifically in S phase triggers cell death.
To explore this phenomenon further, we constructed variant forms of E2f1Stable in which key E2f1 activities–DNA binding, Rbf1 binding, and transactivation–were disrupted in order to determine those aspects of E2f1 function that are necessary for E2f1Stable -induced phenotypes (Figure 1B). To disrupt DNA binding, we mutated to alanines four amino acids in the highly conserved RRXYD motif (R292, R293, Y295 and D296) that make direct contact with bases in the E2F recognition sequence (E2f1DBD Mut) [42]. Mutation of the E2F RRXYD motif was previously demonstrated to block DNA binding [43]. To disrupt interaction with Rbf1, we engineered into our constructs a previously characterized missense mutation (L786Q) within the COOH-terminal Rbf1-binding domain of E2f1 that disrupts normal Rbf1-E2f1 interaction but leaves E2f1 transactivation intact (E2f1Rb Mut) [44]. Because this single amino acid change does not completely eliminate Rbf1-E2f1 interaction (see Figure 2E), we also engineered into our constructs a previously described mutation (E2f1i2) that inserts a stop codon at amino acid Q527 [45]. This allele produces a truncated protein lacking the COOH terminal 1/3 of E2f1, thereby eliminating both transactivation function and Rbf1 binding. We will refer to this allele as E2f1Trunc.
We first determined whether the mutations we engineered affected GFP-E2f1 and GFP-E2f1Stable activity as predicted. We generated UAS-transgenic lines and selected for analysis those that expressed equivalent amounts of GFP-E2f1 mRNA when driven with en-Gal4 (Figure 1C). Each GFP-E2f1Stable mutant protein accumulated to a similar level that was 30–40% higher than either GFP-E2f1 or endogenous E2f1 (Figure 1D). This increase in protein level is consistent with stabilization only during S phase, which represents about one third of the total cell cycle length (Figure 1F, GFP only).
We next assessed the ability of the E2f1 mutant proteins to drive cell cycle progression and to activate E2f1 target gene expression. The GFP-E2f1Rb Mut and GFP-E2f1Stable/Rb Mut Rbf1 binding mutants with intact transactivation domains were able to promote cell cycle progression (Figure 1E). In contrast, expression of either GFP-E2f1 or GFP-E2f1Stable proteins with mutations that disrupt the transcriptional activity of E2f1, either by blocking DNA binding (GFP-E2f1DBD Mut) or removing the transactivation domain (GFP-E2f1Trunc), failed to shorten G1 (Figure 1E). Identical results were obtained using S2 cells (Figure S1I).
Mutations that disrupt DNA binding or transactivation should prevent E2f1 from activating expression of replication factor genes. To test this prediction, we measured the level of RnrS mRNA, a well-known E2f1-regulated transcript [15]. While expression of GFP did not change the level of RnrS mRNA, both GFP-E2f1 and GFP-E2f1Stable expression resulted in a ∼3 fold increase in RnrS mRNA in wing imaginal discs (Figure 1F). Similar to the cell cycle progression results, those GFP-E2f1 or GFP-E2f1Stable mutant derivatives that are predicted to be deficient for E2f1 transcriptional activity (GFP-E2f1DBD Mut and GFP-E2f1Trunc) were unable to induce RnrS expression, while the Rbf1 binding point mutant (GFP-E2f1Rb Mut) induced RnrS expression similarly to GFP-E2f1 (Figure 1F). Thus, the introduction of the S phase stabilizing mutation did not alter the transcriptional activity of E2f1.
E2f1 requires dimerization with Dp for transcriptional activity and Rbf1 binding for normal regulation in G1 phase [46]. To determine whether our mutations affected Dp or Rbf1 binding, we transiently transfected Myc-E2f1 with either HA-Dp or HA-Rbf1 into S2 cells and performed co-immunoprecipitation assays. All of the E2f1Stable mutant proteins bound Dp equivalently to wild type E2f1 (Figure 1G). Likewise, we found that E2f1, E2f1Stable, and E2f1Stable/DBD Mut precipitated similar amounts of Rbf1 (Figure 1H). In contrast, E2f1Stable/Rb Mut precipitated a reduced amount of Rbf1 relative to E2f1, and the truncated E2f1Stable/Trunc showed no ability to precipitate Rbf1 (Figure 1H). These data indicate that we have successfully created PIP degron mutant derivatives of E2f1 that have the predicted effects on the ability to activate transcription and drive cell cycle progression (GFP-E2f1Stable/DBD Mut), bind Rbf1 (E2f1Stable/Rb Mut), or both (E2f1Stable/Trunc).
We next asked whether any of these mutations affected S phase-coupled E2f1 destruction. Using either the wing disc or S2 cell flow cytometric assays, we found that E2f1DBD Mut, E2f1Rb Mut, and E2f1Trunc are each destroyed during S phase in a PIP degron-dependent manner (Figure 1I, Figure S1J) [32]. These data indicate that neither the DNA binding, Rbf1 interaction, or transactivation domains of E2f1 are required for S phase-coupled destruction. We previously demonstrated that E2f1 destruction during S phase requires Dp [32], a result that could be interpreted as a requirement for E2f1/Dp DNA binding [34]. However, an alternative interpretation from our observations that the E2f1DBD Mut protein binds Dp and is destroyed normally during S phase is that E2f1/Dp heterodimers are the preferred substrate of CRL4Cdt2. In addition, these data suggest that a nuclear pool of E2f1/Dp that is not bound to DNA can interact with PCNA at replication forks and recruit the ubiquitylation machinery.
As we showed previously [32], GFP-E2f1Stable induces apoptosis in wing imaginal discs although expression of GFP-E2f1 or GFP does not (Figure 2A–2C). We hypothesized that some activity of E2f1 is necessary to cause cell death only when the protein is inappropriately stabilized in S phase. To determine which functional domains of GFP-E2f1Stable were required to induce apoptosis, we expressed GFP-E2f1Stable variants containing each of the three mutations described above and stained wing imaginal discs with anti-cleaved caspase 3 antibodies (CC3). We first examined the E2f1 DNA binding domain mutant. As expected, GFP-E2f1DBD Mut, which does not function as a transcription factor or cell cycle regulator, did not induce apoptosis (Figure 2D). Very surprisingly, however, we detected robust CC3 staining when this protein was stabilized during S phase with the PIP3A mutation (GFP-E2f1Stable/DBD Mut) (Figure 2E). This result indicates that apoptosis in response to stabilizing E2f1 in S phase is neither a consequence of aberrant cell cycle progression or E2f1 target gene expression, nor is it solely due to gross over-expression as the normally degradable E2f1DBD Mut did not cause this phenotype.
We next addressed whether GFP-E2f1Stable-induced apoptosis requires an interaction with Rbf1. Expression of GFP-E2f1Rb Mut did not induce apoptosis (Figure 2F). The S phase-stabilized Rbf1 binding mutant GFP-E2f1Stable/Rb Mut caused an increase in CC3 staining compared to controls, but less than we observed with either GFP-E2f1Stable or GFP-E2f1Stable/DBD Mut expression (Figure 2G). Intriguingly, this effect suggested that the ability of S phase-stabilized E2f1 to induce apoptosis requires an interaction with Rbf1 but not the ability of E2f to bind to E2F response elements at target genes or to shorten G1 phase. To test the role of the E2f1-Rbf1 interaction further, we examined the C-terminally truncated GFP-E2f1Stable/Trunc protein, which is devoid of Rbf1 binding. Neither expression of the GFP-E2f1Trunc nor the GFP-E2f1Stable/Trunc protein resulted in an increase in CC3 staining (Figure 2H, 2I). Importantly, both the GFP-E2f1Stable/Rb Mut and the GFP-E2f1Stable/Trunc proteins were expressed at levels equivalent to the GFP-E2f1Stable and GFP-E2f1Stable/DBD Mut proteins that induce apoptosis (Figure 1D).
To quantify the apoptosis induced by different GFP-E2f1 proteins, we measured the number of cells within a specific range of sub-G1 DNA content via flow cytometry of dissociated wing discs. By this assay, we detected ∼5% apoptotic cells in GFP-expressing control discs, which likely reflects both the normal low levels of apoptosis present in unperturbed discs (e.g. arrow Figure 2D) and the consequences of the extensive trypsinization required for dissociation (Figure 2J). GFP-E2f1 caused only a slight increase in sub-G1 cells relative to GFP controls, as did the transcriptionally inactive GFP-E2f1DBD Mut (Figure 2J). In contrast, and in correspondence with the CC3 staining of intact discs, expression of GFP-E2f1Stable or GFP-E2f1Stable/DBD Mut, which lacks a functional DNA binding domain, caused a significant increase in the apoptotic population of cells relative to controls (Figure 2J). The E2f1Stable/Rb Mut Rbf1-binding mutant triggered apoptosis, but less so than GFP-E2f1 proteins with a wild type Rbf1 binding domain, and the GFP-E2f1Stable/Trunc Rbf1-binding deficient mutant did not significantly increase apoptosis above controls (Figure 2J). These data indicate that interaction with Rbf1 is required for S phase-stabilized E2f1 to induce apoptosis. They also suggest that cells have a mechanism to detect aberrant E2f1 protein levels during S phase that is independent of E2f1's role as a transcription factor.
Our experiments thus far utilize en-Gal4 to drive GFP-E2f1 expression continuously in the posterior compartment during growth of the wing imaginal disc. Because this expression initiates very early during development, we cannot determine whether phenotypes arise in the very first cell cycle after stabilizing E2f1 during S phase, or result from E2f1Stable expression over many cell cycles. To address this issue, we took advantage of the distinct cell cycle program of eye imaginal discs. During third instar larval development, a wave of differentiation associated with a coordinated cell shape change called the morphogenetic furrow (MF) sweeps across the eye disc from posterior to anterior over a period of two days [47]. Cells anterior to the MF are undifferentiated and undergo asynchronous cell proliferation, while cells posterior to the MF differentiate into the neurons and other specialized cell types of the compound eye. Cells within the MF arrest in G1 phase, and as they exit the MF some cells remain in G1 and differentiate while others synchronously reenter a final cell division cycle prior to terminal differentiation called the “second mitotic wave" (SMW) (Figure 3A) [48].
The GMR-Gal4 driver is activated in late G1 cells of the MF and remains on in all cells posterior to the MF (Figure S2A). By using GMR-Gal4 we could examine the very first cell cycle after expression of the E2f1 transgenes. Normal eye discs have a very organized and stereotypical pattern of S phase in the SMW, and very few cells enter apoptosis immediately posterior to the MF (Figure 3A). Expression of GFP-E2f1 resulted in minimal changes to S phase of the SMW and no significant increase in apoptosis posterior to the MF (Figure 3B, 3F). (Note that others have demonstrated previously that co-expression of E2f1 and Dp with GMR results in ectopic S phase in the MF and apoptosis [21].) In contrast, expression of E2f1Stable disrupted the normal S phase pattern: we observed an increase in the number of cells entering S phase as well as an expansion of the zone of EdU labeling posterior to the MF, suggesting an increase in the length of S phase (Figure 3C). The changes in the S phase pattern caused by GFP-E2f1Stable were accompanied by an increase in DNA damage, as measured by anti-phospho-H2Av staining (Figure 4A–4C, 4F), and apoptosis posterior to the MF (Figure 3C, 3F). There was no change in the number of cells entering mitosis posterior to the MF, as measured by anti-phospho-histone H3 staining (Figure S2B), suggesting that cells die before entering mitosis. In addition, E2f1Stable did not induce apoptosis when expressed in G1-arrested epidermal cells in the embryo (Figure S3), suggesting that apoptosis may be S phase specific. These data suggest that the presence of stabilized E2f1 in even a single S phase can disrupt cell cycle progression, induce DNA damage, and result in apoptosis. Importantly, however, DNA damage and apoptosis does not occur in all of the cells expressing E2f1Stable, much like we observed by flow cytometry in the wing discs (Figure 2J).
We next asked whether the DNA damage and apoptosis observed after S phase stabilization of E2f1 resulted from aberrant cell cycle progression. Expression of GFP-E2f1Stable/DBD Mut did not perturb the organization of S phase in the SMW (Figure 3D) or result in an increased number of phospho-H2Av foci (Figure 4D, 4F), likely because this protein does not alter cell cycle progression. Thus, the DNA damage observed with E2f1Stable was most likely due to proliferation defects, because mutants that failed to shorten G1 did not induce phospho-H2Av. On the other hand, when compared to controls, GFP-E2f1Stable/DBD Mut expression did not cause an increase in phospho-H2Av foci (Figure 4D, 4F), although it still resulted in an increase in apoptosis posterior to the MF Figure 3D, 3F). These data suggest that stabilizing E2f1 in S phase can trigger apoptosis independently of cell cycle effects. The level of apoptosis in GMR>GFP-E2f1Stable/DBD Mut discs was somewhat less than in GMR>GFP-E2f1Stable discs, suggesting a contribution from proliferative stress that is dependent on E2f1 DNA binding (Figure 3F). As in wing discs, apoptosis required an interaction with Rbf1 because GFP-E2f1Stable/Rb Mut expression resulted in reduced apoptosis compared to GFP-E2f1Stable (Figure 3E, 3F). Taken together, these data suggest that two factors contribute to apoptosis when E2f1 is stabilized in S phase in the SMW: proliferative stress caused by aberrant E2f1 activity that leads to DNA damage, and a mechanism independent of E2f1 DNA binding activity that relies on interaction with Rbf1.
We previously reported that E2f1Stable expression in the posterior compartment of the wing discs did not increase the amount of detectable DNA damage [32]. Our eye discs results prompted us to reexamine this issue. Using a different source of anti-phospho-H2Av antibody, we detected an increase in phospho-H2Av foci in wing imaginal discs following expression of GFP-E2f1Stable, and as in the eye discs this amount was more than with GFP-E2f1 expression (Figure S4).
Our data are consistent with the hypothesis that an interaction between S phase-stabilized E2f1 and Rbf1 triggers apoptosis, even when E2f1Stable cannot bind DNA and is functionally inactive as a transcription factor. This result suggests that cells can specifically detect and respond to E2f1/Rbf1 complexes that inappropriately assemble in S phase. However, another possibility is that over-expression of any Rbf1 binding protein would trigger apoptosis. To distinguish between these possibilities, we utilized a NH2-terminally truncated allele of E2f1 (E2f1336–805) that we previously characterized [49]. E2f1336–805 contains only the C-terminal half of the E2f1 protein, and thus lacks the PIP degron and DNA binding domain but retains the Rbf1 binding and transactivation domains (Figure 5A). We hypothesized that this protein would interact with Rbf1 during S phase, but not trigger apoptosis because of the absence of a domain necessary for cells to detect the E2f1Stable/Rbf1 complex. Indeed, en-Gal4 expression of E2f1336–805 failed to induce apoptosis (Figure 5B), even though this protein accumulated to levels similar to GFP-E2f1Stable (Figure 5C) and efficiently interacted with Rbf1 in co-immunoprecipitation assays (Figure 5D). These data indicate that interaction with Rbf1 is not by itself sufficient to induce apoptosis, and suggest that full-length E2f1Stable is specifically recognized by cells to induce apoptosis.
What mechanism could explain the induction of apoptosis upon stabilization of a transcriptionally inert, but Rbf1 binding-proficient, E2f1 protein during S phase? Recent work from several laboratories showed that loss of Rbf1 function causes apoptosis in several developmental contexts by triggering expression of the pro-apoptotic gene, hid [18], [24], [25], [50], [51]. Hid is homologous to SMAC/Diablo family proteins that function to antagonize IAPs, which act to block activator caspases. hid expression triggers an apoptotic cascade by antagonizing DIAP1, thus releasing inhibition of the initiator caspase Dronc and activating the effector caspase Drice [36], [52].
We hypothesized that GFP-E2f1Stable or GFP-E2f1Stable/DBD Mut binds to Rbf1 and disrupts its function, resulting in activation of hid expression. This hypothesis predicts that GFP-E2f1Stable or GFP-E2f1Stable/DBD Mut expression will increase hid expression, while E2f1 mutants that cannot bind Rbf1 will fail to increase expression. To test this prediction, we used qRT-PCR to measure the levels of hid mRNA in wing imaginal discs expressing the various GFP-E2f1 transgenes with en-Gal4. Consistent with our hypothesis, there was a two-fold increase in hid mRNA in GFP-E2f1Stable- or GFP-E2f1Stable/DBD Mut -expressing discs relative to those expressing GFP-E2f1 or GFP-E2f1DBD Mut (Figure 6A). Similar levels of hid induction were previously observed following ionizing radiation treatments that trigger apoptosis [24]. hid expression was not significantly increased by the GFP-E2f1 mutants lacking normal Rbf1 binding activity (Figure 6A). To test whether the hid de-repression was a specific response to stabilizing E2f1 in S phase, we measured expression of another pro-apoptotic gene, reaper, which is not de-repressed by Rbf1 mutation [24]. We detected no increase in reaper mRNA in discs expressing any GFP-E2f1 transgene (Figure 6B).
To test if hid expression contributed to E2f1Stable-induced apoptosis, we determined whether reducing hid gene dose would result in a decrease in apoptosis. We utilized two different hid alleles: hid05014, containing a transposable element insertion between amino acids 105 and 106 in the open reading frame that effectively reduces hid expression [53], and Df(3L)H99, which deletes the entire hid gene as well as the neighboring pro-apoptotic genes, reaper and grim [53], [54]. Wing discs heterozygous for either hid allele contained a significantly reduced amount of apoptosis after GFP-E2f1Stable or GFP-E2f1Stable/DBD expression compared to controls (Figure 6C–6H, 6I). Quantification of CC3 staining revealed no significant difference between the results obtained with hid05014 (Figure 6D, 6G, 6I) and Df(3L)H99 (Figure 6E, 6H, 6I). This result suggests that grim and reaper do not contribute as substantially as hid to E2f1Stable-induced apoptosis, consistent with our failure to detect an increase in reaper expression by E2f1Stable and its derivatives (Figure 6B). These data support the idea that stabilizing E2f1 during S phase results in disruption of Rbf1 function leading to de-repression of hid expression and apoptosis.
Why would Drosophila cells induce a potent activator of apoptosis in response to elevated E2f1 protein levels during S phase? We considered the possibility that a small number of individual cells in a growing population of adult precursor cells, like those in wing imaginal discs, might stochastically experience hyper-expression of E2F that would manifest as the presence of excess E2f1 protein in S phase. Such cells would be eliminated by apoptosis, thereby helping to maintain growth homeostasis by suppressing the appearance of potentially hyperplastic cells that could lead to aberrant overgrowth. If this was a developmentally important event, then blocking the ability of tissues to eliminate such cells by apoptosis should disrupt normal development.
To test this idea, we used en-Gal4 to co-express GFP-E2f1 transgenes in wing imaginal discs together with baculovirus p35, which efficiently blocks apoptosis in Drosophila cells [55]. Expressing p35 together with GFP had no deleterious effects on wing disc development (Figure 7A). In contrast, GFP-E2f1/p35 co-expression resulted in a range of morphological defects caused by hyperplastic growth. While some GFP-E2f1/p35-expressing discs appeared normal, most displayed various degrees of overgrowth in the posterior portion of the disc (Figure 7B). We quantified this overgrowth by microscopically measuring posterior compartment “thickness", which we defined as the sum of the number of confocal sections one micron apart required to image through the entire posterior compartment. Using this measurement we binned the discs into four phenotypic categories: normal (<9 µm), mild (9–11 µm), moderate (12–15 µm), and severe (>15 µm) (Figure 7D). GFP-E2f1Stable/p35 expression caused a more severe phenotype than did GFP-E2f1/p35 expression. None of the discs were normal, and a larger percentage of the discs fell into the severe overgrowth category (Figure 7C, 7D). In addition, GFP-E2f1Stable/p35 expression caused the appearance of a unique fifth phenotype in ∼1/3 of the discs, which we called “arrest" (Figure 7C, 7D). In these discs the posterior compartment was almost absent, as confirmed by co-expression of RFP. We speculate on the origin of this class of discs in the Discussion. Expression of p35 together with either GFP-E2f1 or GFP-E2f1Stable caused 100% lethality. Importantly, the hyperplastic growth induced by GFP-E2f1 or GFP-E2f1Stable required the normal transcriptional and cell cycle-promoting activity of E2f1, as co-expression of p35 with GFP-E2f1DBD Mut or GFP-E2f1Stable/DBD Mut resulted primarily in normal wing discs and did not cause lethality (Figure 7D). These data indicate that the developmental effects of E2f1 hyper-activity during tissue growth are exacerbated by simultaneously blocking apoptosis and E2f1 destruction in S phase.
We show here that stabilizing the single Drosophila activator E2f1 in S phase results in apoptosis that is necessary to prevent hypertrophy of wing imaginal discs. We conclude from these data that hyper-accumulation of E2f1 during S phase represents a form of proliferative stress during development that is sensed by the apoptotic machinery and results in the elimination of cells with excess E2f1 activity to maintain homeostasis during tissue growth.
What might be the function of a Drosophila cell's ability to detect abnormal accumulation of E2f1 protein during S phase and subsequently trigger apoptosis? One possibility is that accumulation of E2f1 during S phase resembles instances of abnormally high E2f1 activity that might occur sporadically during rapid growth of a population of precursor cells such as those in the wing imaginal disc. These events could be caused by stochastic or even genetic changes that affect either E2f1 gene transcription or the ability of the CRL4Cdt2/PCNA pathway to destroy E2f1 after replication factor genes are activated in late G1. The cell's ability to detect E2f1 accumulation in S phase clears these potentially hyperplastic cells from developing tissues via apoptosis, consequently contributing to the balance between cell proliferation and cell death that is necessary for normal tissue growth.
Growing Drosophila imaginal discs possess another mechanism of homeostasis in which a process of compensatory proliferation is activated in order to achieve normal tissue development when as many as 50% of cells are killed by external stimuli like radiation-induced DNA damage [56]. Indeed, in spite of high levels of apoptosis (15% of the cells), 50% of en-Gal4>E2f1Stable progeny survive until adulthood with about 2/3 of these surviving flies containing wings with somewhat mild morphological defects [32]. Blocking apoptosis with baculovirus p35 when E2f1Stable is expressed shifts the cell proliferation/apoptosis balance too strongly in favor of cell proliferation, resulting in massive hypertrophy and 100% lethality.
p35 is a broad caspase inhibitor that blocks effector caspase activity at a step downstream of their proteolytic activation [55]. Therefore, cells expressing p35 can initiate apoptosis, but lack the capacity to complete it and are referred to as “undead cells." These undead cells produce signals that stimulate unaffected neighboring cells to proliferate [36]. Thus, the dramatic hypertrophy we see in E2f1Stable/p35 wing discs might be the result of two synergizing growth signals: hyper-active E2f1 and compensatory proliferation from undead cells. Our experiments cannot precisely discern the relative contribution of these two inputs, but E2f1 activity appears to make a larger contribution because E2f1Stable/DBD Mut expression does not cause dramatic overgrowth.
What might explain the 32% of en-Gal4>E2f1Stable discs that displayed a reduced posterior compartment rather than an overgrown one: the “arrest" phenotype in Figure 7? The DNA damage we observed in our eye discs experiments provides a possible answer. Perhaps early in development the “arrest" class of wing discs sustained enough genomic damage to prevent proliferation, resulting in too small a pool of cells that could respond to the hyper-active E2f1 and undead cell signals to support disc overgrowth. Thus, the wide range of phenotypes that we observed in E2f1Stable/p35 wing discs may result from multiple influences that act stochastically within the population (Figure 7E).
Because endogenous E2f1 is quantitatively destroyed only in S phase, the relative amount of hyper-accumulation of E2f1Stable is greater during S phase than during any other stage of the cell cycle. Therefore, one possibility is that E2f1Stable-induced phenotypes result from the stability of E2f1 protein in S phase, and not from general over-expression throughout the cell cycle. Our failure to detect E2f1Stable induced apoptosis in G1-arrested embryonic cells is consistent with this possibility. However, another difference between these embryonic cells and wing discs cells is that the former are cell cycle arrested and the latter are continuingly proliferating during larval development. Thus, another possibility is that S phase-destruction of E2f1 modulates the levels of E2f1 in proliferating cells, and cells that fail to destroy E2f1 during S phase have an increased chance of activating apoptosis at any point in the cell cycle. In either model, S phase E2f1 destruction is not essential for proliferation per se. In marked contrast, E2f1Stable expression blocks endocycle progression [57], suggesting that knocking in E2f1Stable to the endogenous locus would be lethal during development, perhaps even dominant lethal.
E2f1Stable induces apoptosis at least in part through expression of the pro-apoptotic gene hid. Surprisingly, these events still occur after expression of an E2f1Stable variant that cannot bind DNA and therefore lacks the ability to stimulate transcription of replication factor genes or cell cycle progression. Instead, E2f1Stable requires the ability to bind Rbf1 to induce hid gene expression and apoptosis. Genetic disruption of Rbf1 is well known to result in increased hid expression [18], [25], [51]. We therefore propose that the inappropriate accumulation of E2f1 in S phase disrupts some aspect of Rbf1 function leading to hid expression and apoptosis.
Our data do not discern either the function of Rbf1 that is disrupted by E2f1Stable or the mechanism of hid induction. While the mechanism connecting Rbf1/E2f1 function and hid may be indirect, some studies suggest that Rbf1 and/or E2f1 could regulate hid directly. Su and colleagues recently demonstrated that Drosophila wing disc cells undergo apoptosis in response to ionizing radiation independently of p53 and that this response requires E2f1 and is triggered by hid expression [26]. In eye discs, loss of Rbf1 function in the MF results in apoptosis that requires E2f1 transactivation function and is accompanied by hid expression [18], [50]. However, whether these effects represent a direct induction of hid by E2f1 is not clear. E2f1 binding at the hid locus has been observed, but the binding site is located ∼1.4 kb upstream of the of the start of hid transcription, which is more distal than in well characterized E2F-regulated promoters [58]. When located this far upstream the hid E2f1 binding site fails to activate gene expression in S2 cell reporter assays [25]. hid is also a target of p53 [59], and so any DNA damage resulting from stabilizing E2f1 during S phase, as we observed in eye discs, may also contribute to the activation of hid expression via p53-mediated DNA damage response pathways.
Another possibility is that E2f1, in combination with Rbf1, plays primarily a repressive role at the hid locus. In this model, our result that E2f1Stable or E2f1Stable/DBD Mut both induce apoptosis would be explained by disruption of Rbf1/E2f1 repressive complexes at the hid locus causing de-repression of hid expression. This model has interesting caveats: what protects the Rbf1/E2f1 complex at the hid locus from being disrupted by Cyclin E/Cdk2, which is active during S phase and inactivates Rbf1-mediated repression of E2f1 [60], or by CRL4Cdt2-mediate destruction of E2f1? Recent data indicate that the dREAM/MMB complex is required for the stability of E2F/Rbf1 repressive complexes in S phase, and acts to protect these complexes from CDK-mediated phosphorylation at non-cell cycle-regulated genes [61]. While there is yet no evidence that dREAM/MMB regulates hid [62], this work provides precedent for gene specific Rbf1 regulation during S phase.
Finally, while hid might be a critical player in the response to E2f1Stable, there are likely other mechanisms responsible for sensing and modulating the apoptotic response to E2f1 levels. For instance, Frolov and colleagues recently demonstrated that a micro-RNA, mir-11, which is located within the last intron of the Drosophila E2f1 gene, acts to dampen expression of pro-apoptotic E2f1 target genes following DNA damage [28]. In this way, the normal controls of E2f1 gene expression modulate apoptosis. In addition, our transgenic constructs lack the normal E2f1 3′ UTR, which serves as a site for suppression of E2f1 expression by pumilio translational repressor complexes [63]. Thus, we have bypassed several modes of E2f1 regulation via transgenic expression of E2f1Stable.
Our finding that stabilized Drosophila E2f1 can induce apoptosis independently of transcription and cell cycle progression parallels previous observations made in mammalian cells, albeit with important differences. In mammalian cells, E2F1 can induce apoptosis independently of transcription and cell cycle progression, but apoptosis required E2F1 DNA binding activity, unlike in our experiments [64], [65]. These studies suggested that DNA binding by E2F1 prevented pro-apoptotic promoters from binding repressor E2F family members.
This comparison of results highlights the way similar phenotypic outcomes in different species can arise from different mechanisms. While mammalian activator E2Fs are also inhibited during S phase, they are not subject to CRL4Cdt2-mediated, S phase-coupled destruction like Drosophila E2f1. Instead, mammalian activator E2Fs are inhibited by direct Cyclin A/Cdk2 phosphorylation [6], [7], [8], targeted for destruction by SCFSkp2 [9], [10], and functionally antagonized by E2F7 and E2F8 [11], . The regulation provided by E2F7 and E2F8 is of particular note, as it is essential for mouse development [11]. These atypical E2Fs homo and hetero-dimerize and act redundantly to repress E2F1 target genes independently of pRb family proteins, thus blocking E2F1 from inducing apoptosis [11]. Moreover, the E2F7 and E2F8 genes are E2F1 targets [11], consequently creating a negative feedback loop that limits E2F1 activity after the G1/S transition. A similar negative feedback loop among factors that regulate G1/S transcription exists in yeast [66]. The analogous Drosophila negative feedback loop involves E2f1 inducing its own destruction by stimulating Cyclin E transcription, which triggers S phase [60]. Therefore, the evolution of eukaryotes has resulted in the use of different molecular mechanism to achieve negative feedback regulation of G1/S-regulated transcription, and in the case of activator E2Fs this regulation is essential for normal development.
E2f1 constructs were generated and expressed using pENTR TOPO (Invitrogen) and Gateway-compatible P element vectors (http://www.ciwemb.edu/labs/murphy/Gateway%20vectors.html).
For S phase-coupled protein destruction analysis, S2 cells stably transfected with hsp70 constructs were heat shocked for 30 minutes at 37°C, which results in GFP or GFP-E2f1 expression in all cells of the population, and allowed to recover at room temperature for 200 minutes prior to analysis by flow cytometry. During the 200 min chase period GFP-E2f1 is destroyed in S phase cells while GFP is not, as measured by the percentage of GFP(+) cells in each phase of the cell cycle. For cell cycle analysis, S2 cells were transfected with plasmid DNA expressing GFP or GFP-E2f1 encoding mRNA from the Actin5C promoter and analyzed by flow cytometry 48 hours later.
For flow analysis of wing imaginal discs, at least 15 third instar larvae of the appropriate genotype were dissected in PBS. 30 imaginal discs were collected and immediately dissociated in PBS containing 0.05% Trypsin- EDTA (Gibco), and 1X Hoechst 33342 DNA binding dye (Sigma) and rocked for 3 hours at room temperature. The dissociated tissue was then immediately analyzed using a LSRII Flow Cytometer and Diva software (Becton Dickinson). Cell cycle profiles were calculated using FlowJo™ Software. Percentages of G1, S, and G2 cells were calculated using Modfit LT software (Verity Software House). P values for all experiments were calculated by student's T test.
S2 cells stained with propidium iodide were analyzed by flow analysis as previously described [32] using the Cyan flow cytometer with Summit 4.3 software (Deko).
Total RNA was extracted from 30 third instar wing imaginal discs using Trizol reagent (Invitrogen) and tissue was sheared with eight passes through a 25-gauge needle. 0.75 µg of total RNA was used for reverse transcription with RevertAid Reverse transcription kit (Fermentas). The resulting cDNA was used for qRT-PCR performed using an ABI prism 7700 Sequence Detection system. Relative levels of specific mRNAs were determined by detection of Maxima Sybr Green (Fermentas). Primers are listed in Table 1. Comparative CT methods were used to quantify levels versus control Rp49 mRNA using the manufacturer's protocol.
Transgenic flies were generated by injecting UASp-E2f1 plasmids into w1118 (Best Gene Drosophila Injection Services, Chino Hills, CA). UAS-GFP, Engrailed-Gal4, UAS-RFP and UAS-p35 stocks were obtained from the Bloomington Stock Center. For antibody staining, imaginal discs were dissected from third instar larvae in PBS, fixed in 6% paraformaldehyde for 20 minutes at room temperature, then permeabilized for 20 minutes in PBS-1.0% Triton-X. Wing discs were incubated overnight with mouse anti-GFP (1∶1000, Abcam) and rabbit anti-cleaved Caspase-3 (Asp175) (1∶200, Cell Signaling Technology) at 4°C. Secondary antibodies were goat anti-mouse Oregon Green 488 (1∶2000 Invitrogen) and goat anti-rabbit Rhodamine (1∶2000 Invitrogen) for 1 hour at room temperature. Eye discs were dissected, incubated in 10 µg/mL EdU (Click-iT™ EdU Alexa Fluor 594, Invitrogen) for 30 minutes, fixed and permeabilized as described above. EdU was detected according to manufacture protocol. To detect mitosis, eye discs were incubated overnight at 4°C with rabbit anti-PH3 (1∶1000, Abcam) and then with goat anti-rabbit Rhodamine (1∶1000 Invitrogen) for 1 hour at room temperature. For DNA damage detection, rabbit anti-p-H2Av antibody from Kim McKim's lab was incubated over night at 4°C at 1∶1000 and then goat anti-rabbit Rhodamine (1∶1000 Invitrogen). DAPI was added for DNA detection (1∶1000 Invitrogen) for 2 minutes. Tissue samples were analyzed with a Zeiss LSM 510 scanning confocal microscope. Quantification of CC3 and p-H2Av foci was collected by projecting confocal images that were one micron apart through the eye disc of 6 discs per genotype and using ImageJ software to count all foci above threshold detection posterior to morphogenetic furrow. 7 images per disc projected for p-H2Av, 6 images per disc for CC3. Graph shown represents the average number of foci of those 6 discs.
30 third instar larvae wing imaginal discs were dissected in PBS then dissociated by eight passes through a 25 gauge needle after addition of ice-cold NP40 buffer with protease inhibitors aprotinin (1∶1000), leupeptin (1∶1000) and PMSF (1∶100). E2f1 protein levels were measured with affinity-purified rabbit anti-E2f1 raised against full-length Drosophila E2f1 (1∶1000) [32] overnight at 4°C and anti-rabbit HRP secondary (1∶10,000 GE Healthcare) for 1 hour at room temperature. B-tubulin was used as loading control (1∶1000, Abcam) with anti-rabbit HRP secondary (1∶10,000 GE Healthcare). Co-immunoprecipitation was performed by co-transfecting S2 cells with 2 µg Myc-E2f1 and 1 µg HA-Dp or HA-Rbf1 using the Amaxa transfection system (Lonza) and incubating the cells for 24 hours at 28°C. S2 cells were lysed on ice using NP40 buffer with the protease inhibitor cocktail described above. 10% of each total extract was subjected to western blot analysis with mouse anti-Myc (1∶2000 UNC Hybridoma) or mouse anti-HA (1∶2000, UNC Hybridoma). Secondary antibodies were ECL donkey anti-mouse HRP (1∶10,000, GE Healthcare) and ECL donkey anti-rabbit HRP (1∶10,000, GE Healthcare). The remainder of the extract was incubated overnight at 4°C with 0.5 µg mouse anti-Myc antibody (UNC Hybridoma) and 1/10 volume Protein G Sepharose 4 Fast-Flow beads (GE Healthcare).
|
10.1371/journal.pntd.0007449 | Disruption of the NlpD lipoprotein of the plague pathogen Yersinia pestis affects iron acquisition and the activity of the twin-arginine translocation system | We have previously shown that the cell morphogenesis NlpD lipoprotein is essential for virulence of the plague bacteria, Yersinia pestis. To elucidate the role of NlpD in Y. pestis pathogenicity, we conducted a whole-genome comparative transcriptome analysis of the wild-type Y. pestis strain and an nlpD mutant under conditions mimicking early stages of infection. The analysis suggested that NlpD is involved in three phenomena: (i) Envelope stability/integrity evidenced by compensatory up-regulation of the Cpx and Psp membrane stress-response systems in the mutant; (ii) iron acquisition, supported by modulation of iron metabolism genes and by limited growth in iron-deprived medium; (iii) activity of the twin-arginine (Tat) system, which translocates folded proteins across the cytoplasmic membrane. Virulence studies of Y. pestis strains mutated in individual Tat components clearly indicated that the Tat system is central in Y. pestis pathogenicity and substantiated the assumption that NlpD essentiality in iron utilization involves the activity of the Tat system. This study reveals a new role for NlpD in Tat system activity and iron assimilation suggesting a modality by which this lipoprotein is involved in Y. pestis pathogenesis.
| We have previously shown that the NlpD lipoprotein, which is involved in the regulation of cell morphogenesis, is essential for virulence of the plague bacteria, Yersinia pestis. To uncover the role of NlpD in Y. pestis pathogenicity, we conducted a whole-genome comparative transcriptome analysis as well as phenotypic and virulence evaluation analyses of the nlpD and related mutants. The study reveals a new role for the Y. pestis NlpD lipoprotein in iron assimilation and Tat system activity.
| Bacteria and in particular bacterial pathogens have successfully evolved sophisticated mechanisms that allow them to sense, cope and adapt to varying conditions in their immediate surroundings. The rapid detection of extracellular signals such as the concentrations of nutrients, ion sources, temperature, stress conditions and the presence of host immune cells, influence transcriptional regulatory systems that in turn lead to physiological and morphological changes that enable the organism to survive within hostile environments such as those encountered in the host during infection [1].
The Gram-negative pathogen Yersinia pestis is the causative agent of plague, a disease that has inflicted millions of deaths in three world pandemics [2]. Plague still persists in Africa, Asia and the Americas and as of today, it is categorized as a re-emerging disease [3]. The prevalent form of the disease is bubonic plague, which develops following transmission of the pathogen from rodent reservoirs to humans via infected fleas and has high mortality rate if untreated [4,5]. Primary pneumonic plague, which represents one of the most severe forms of the disease, is less abundant in nature and results from the inhalation of Y. pestis-containing droplets or aerosols. Pneumonic plague is a contagious rapidly progressing disease that leads to 100% mortality in untreated patients [2]. These characteristics as well as the inhalational nature of infection led to the recognition of plague as a bioterror threat agent [6].
The ability of Y. pestis to invade the mammalian host, colonize internal organs and overcome the immune response is attributed to the combined activities of multiple virulence pathways that are activated in a timely manner during infection in response to the host milieu signals [7,8]. Some of these pathways have been demonstrated to be absolutely required for the bacterial pathogenesis in animal models. These include molecular systems that enable the acquisition of essential nutrients during infection as well as those required for evading the host immune response such as the type 3 secretion system (T3SS) [9,10,11,12,13,14,15,16,17,18,19].
We have previously documented that the NlpD lipoprotein is essential for Y. pestis virulence in animal models of plague infection and that the nlpD mutant is impaired in its ability to colonize internal organs [20]. NlpD is conserved among Gram-negative bacteria and exhibits characteristic LysM and LytM domains found in enzymes involved in bacterial cell wall remodeling [21,22]. Consistent with this known biological role of LytM-containing proteins, the phenotype of the Y. pestis nlpD-disrupted mutant is characterized by altered chain-forming cell morphology [20]. Interestingly, despite its substantial virulence attenuation, the mutant was not affected in terms of in vitro growth or in the activity of the T3SS [20], which is essential for the pathogenicity of the bacteria. To gain further insights into the role of NlpD in the manifestation of Y. pestis virulence, we extended the characterization of the nlpD mutant in the present study by performing transcriptomic, phenotypic, and molecular genetic analyses. Integration of the results led to the unexpected finding that in Y. pestis, NlpD is required for iron assimilation and for the activity of the twin-arginine system (Tat) which translocates folded proteins across the bacterial cytoplasmic membrane in a wide range of bacteria [23]. Systematic deletion mutagenesis of Tat genes in the virulent Y. pestis Kim53 strain indicated that the Tat system is required for Y. pestis iron assimilation as well as virulence manifestation in the mouse plague infection models.
This study was carried out in strict accordance with the recommendation in the Guide for the Care and Use of Laboratory Animals of the National Institute of Health. All animals experiments were performed in accordance with Israeli law and were approved by the Ethics Committee for animal experiments at the Israel Institute for Biological Research (Protocols M-03-16, M-81-16, M-57-11).
The Y. pestis strains and plasmids used in this study are listed in Table 1. Construction of the Kim53 deletion mutants was performed by replacing part of the gene coding sequence with a linear fragment containing a resistance cassette by homologous recombination using previously established methodologies [16]. The sequence that was deleted from the Y. pestis genome in each mutant is indicated in Table 1.
Deletion of the Y. pestis nlpD gene has been described previously [20]. Deletion of the Y. pestis tatA, tatB and tatC and amiC genes was performed using a linear fragment containing a kanamycin resistance gene that was amplified from pKD4 plasmid and used for homologous recombination as described in [27]. Deletion of the tatA, tatB, tatC and amiC genes from the Y. pestis genome was verified by PCR analysis. TatC expression was monitored by Western blot analysis in the tatA and tatB mutants to test for possible polar effect on TatC expression (S1 Fig). The analysis indicated that tatC expression was not affected by the tatA mutation yet it was affected in the tatB mutant, and therefore the tatB mutant was excluded from the analysis documented in this report (S1 Fig).
The pTorAsignal-GFP [25] and pCA24N-napG [26] plasmids were used to test the functionality of the Tat system in Y. pestis.
For complementation of the tatA and tatC deletion mutants, each of the respective tat genes was cloned into the low copy plasmid pWKS30 [28]. Primers pWKS-tatA-For and pWKS-tatA-Rev for the tatA gene and primers pWKS-tatC-For and pWKS-tatC-Rev for tatC gene were used for PCR amplification from Y. pestis Kim53 DNA (S1 Table). The PCR products were digested with XbaI and HindIII, and then cloned into the pWKS30 vector to generate the pWKS-tatA and pWKS-tatC ampr plasmids. The plasmids were transformed into the compatible Y. pestis deletion mutant. PCR analysis verified that all the newly constructed mutants carried the pMT1, pCD1 and pPCP1 plasmids and the chromosomal pgm locus.
Y. pestis strains were routinely grown on brain heart infusion agar (BHIA, BD, MD USA) for 48 h at 28°C. The nlpD and tat mutants were grown on BHIA supplemented with 50 μg/ml kanamycin (Sigma-Aldrich, Israel), and all complemented Y. pestis mutant were grown on BHIA supplemented with 100 μg/ml ampicillin.
For bacterial total RNA preparation, bacterial colonies were grown on BHIA plates for 48 h at 28°C, harvested and diluted in heart infusion broth (HIB) (BD, USA) supplemented with 0.2% xylose and 2.5 mM CaCl2 (Sigma-Aldrich, Israel) to an OD660 of 0.01 and grown over night (o.n.) at 28°C in a shaker (200 rpm). The resulting cultures were diluted in fresh broth to an OD660 of 0.05 and allowed to grow for 5 h at 37°C. Aliquots of ~5×108 colony forming units (cfu) were collected by centrifugation and the cells were immediately frozen in liquid nitrogen and stored at -70°C until use.
To assess growth under iron limiting conditions, we followed protocols established at the Perry and Fetherston laboratory [29]. Several isolated colonies grown for 48 h at 28°C on a BHIA plate were collected for o.n. growth at 28°C in PMH2 medium [30] [31]. The next day, 0.1 OD660 of the o.n. cultures was inoculate into fresh PMH2 grown for 6–7 h at 37°C and then diluted to 0.1 OD660 with fresh PMH2 and grown o.n. at 37°C. The next morning, the cultures were diluted again to 0.1 OD660 with PMH2 and 10μl drops containing ~106 bacilli were plated on iron-depleted, gradient plates and incubated for ~50 h at 37°C. Gradient plates were prepared using square plates (USA scientific, 5668–8102) to which was added a total of 70 ml of medium was added, 35 ml for the bottom layer (poured on a slope) and 35 ml for the top layer, which was poured 24 h before performing the growth assay. The bottom layer contained 1% agarose, 1× PMH2, 20 μM MgCl2 and 100 μM 2,2’-dipyridyl (DIP) as a chelator. The top layer contains 1% agarose, 1× PMH2 and 20 uM MgCl2 (no chelator). In this manner, a DIP gradient was established ranging between 0 and 100 mM. In addition, plates containing 1% agarose, 1× PMH2, 20 μM MgCl2 and 80 or 100 μM DIP were prepared [32] and 10μl drops containing ~106 bacilli were plated and incubated for ~50 h at 37°C as described above. To rescue bacterial growth on plates containing 100 μM DIP, iron dextran (d8517, Sigma-Aldrich) was added to the medium at a concentration of 0.5 mg/ml.
To evaluate the functionality of the Tat system, bacteria containing pTorAsignal-GFP or pCA24N constructs (Table 1) were grown o.n. at 28°C in HIB supplemented with 0.2% xylose and 2.5 mM CaCl2 (Sigma-Aldrich, Israel) containing 100 mg/ml ampicillin, and the next-day cultures were diluted to OD600 of 0.05 into 15 ml culture and grown at 37°C. After incubation for 5–7 h, 5 ml of each culture was centrifuged (10,000 g), and the cell pellets were washed with 5 ml of double-distilled water (DDW). The cells were centrifuged and then resuspended to an OD660 of 0.01 with DDW. For morphological analysis, bacterial cells were washed with DDW twice and diluted to an OD660 of 0.01.
Total RNA was purified using the RNeasy-Mini Kit (QIAGEN) according to the manufacturer’s instructions. Seven micrograms of total RNA was used for microarray analysis of each sample using the FairPlay III microarray labeling kit (Stratagene) according to the manufacturer’s instructions. To examine changes in expression of Y. pestis genes, a custom array was used [33]. The array contains 4196 coding regions and pseudogenes out of the 4321 predicted genes of the Y. pestis CO92 strain (Acc no.: NC_003143, NCBI). Hybridization and scanning were performed as suggested by Agilent. The slides were scanned in an Agilent DNA Microarray Scanner G2505B. Images were analyzed and data were extracted using Agilent Feature Extraction software version 9.5.1.1 (FE), with linear and lowess normalization performed for each array. A reference design with two biological replicates was applied to compare the wild-type Kim53 strain and the ΔnlpD mutant. Statistical analysis was performed using the Limma (Linear Models for Microarray Data) package from the Bioconductor project (http://www.bioconductor.org). The processed signal resulting from the FE was read into Limma using the "read.maimages" function. Background subtraction and lowess normalization were performed within each array. A quantile normalization between arrays was applied. Standard quality control was performed using the plot functions of Limma [34]. Differential expression was assessed using linear models for designed microarray experiments. The fold changes (FC) and standard errors were estimated by fitting a linear model for each gene and applying empirical Bayes smoothing to the standard errors [34]. The FDR (false discovery rate) was controlled using the method of Benjamini and Hochberg for multiple comparisons [35,36].
The P value is the result of a one-sample Student’s t test, which was applied to the natural log of the mean of each normalized value against the baseline value of 0. Genes with differences corresponding to P<0.05 in either the high or the low photomultiplicator scans and that had signal-to-control or control-to-signal ratios ≥2.0 were considered to be differentially regulated. The results were submitted to the GEO depository and are available online (http://www.ncbi.nlm.nih.gov/geo/, record GSE101490).
For qRT-PCR analysis, 2μg of total RNA was reverse-transcribed using the Reverse Transcription (RT) System kit (Promega) with random primers according to the manufacturer’s instructions. The cDNA was used as a template for qRT-PCR with specific primers (S1 Table) using an ABI 7500 instrument (Applied Biosystems, USA) with SYBR green PCR master mix (Applied Biosystems, USA). Relative quantification was determined using an average of 2 genes: YPO1045 (tsf gene) and YPO1415 (pyrD gene), for standardization of all qRT-PCR results using the comparative (-2ΔΔCt) method. Forty cycles of PCR amplification were performed in duplicate for each primer set.
For Western blot analysis, bacterial colonies were grown on BHIA plates for 48 h at 28°C, harvested and diluted in heart infusion broth (HIB) (BD, USA) supplemented with 0.2% xylose and 2.5 mM CaCl2 (Sigma-Aldrich, Israel) to an OD660 of 0.05 and grown at 37°C in a shaker (200 rpm). Bacteria (OD660 = 0.1) were lysed in Laemmli Sample buffer (Bio-Rad) and protein concentrations were determined using bicinchoninic acid (BCA Protein Assay Reagent, Pierce) according to the manufacturer's instructions. Equal amounts of protein were loaded and separated by SDS-PAGE. After transfer to nitrocellulose membranes, duplicate membranes were developed with anti-peptide antibodies against NlpD, Pcm [20] and TatC (see below). Probing with the primary antibody was followed by incubation of the membranes with HRP-conjugated secondary antibody (A6154, Sigma-Aldrich) visualized using the LAS-3000 imaging system (Fuji) or by IRDye 680RD-conjugated secondary antibodies (LIC-92668071 and LIC-92668070 LI-COR) visualized using the Odyssey CLx imaging system from LI-COR. The TatC anti-peptide antibodies were raised by immunizing rabbits with maleimide-activated KLH (Pierce) conjugated to the synthetic peptides CYNLVSAPLIKQLPAGAS (amino acids 41–59 out of 258aa of TatC).
For the morphological analysis, 5μl bacterial aliquots were placed in an 8-well slide (#5638–01940, Bellco Glass) to dry. Cells were fixed with cold methanol (−20°C) for 15 minutes and Gram stained according to the manufacturer’s instructions (HT90A kit, Sigma-Aldrich). Images were captured under a Nikon Eclipse E200 microscope connected to a Nikon DS-Fi-1 camera at a magnification of ×400 and ×1000.
For the Tat and Sec functionality analysis, bacterial cells were mounted on poly L-lysine-treated microscope slides with Fluoromount-G (Southern Biotechnology, Birmingham, Al) and covered with a glass coverslip. The slides were examined by phase-contrast and fluorescence (fluorescein isothiocyanate filter set) microscopy. The images shown were analyzed using Zeiss LSM 710 Confocal Microscope (Zeiss, Oberkochen, Germany). Fluorescence intensity quantification of the above-mentioned markers was performed using Zen1 software, Zeiss.
For DAPI and TatC labeling of bacterial cells, approximately 106 cfu where placed in a well on a DoubleCytoslide (Thermo). The cells where dried for 30 minutes and fixed with cold methanol (-20°C) for 10 minutes. The slides were then transferred to 80% cold acetone (-20°C) for 30 seconds and allowed to dry. Blocking was performed for 15 minutes with 2% BSA (in PBS). Slides where washed with DDW 3 times and then labeled with a primary αTatC antibody for 1 hour (2% BSA, 0.05% Tween 20 suspended in PBS), washed 5 times with PBS and then labeled for 15 minutes with a secondary anti-rabbit antibody labeled with Alexa 594 succinimidyl ester.
Labeling of Y. pestis cells was performed with αF1 antibodies generated against the F1 capsular protein [37] and linked to FITC or with antibodies generated against the bacterium [38] and linked to Alexa fluor 647. After labeling, the slides where washed 5 times in DDW, labeled by DAPI staining for 2 minutes, washed two times with DDW, covered and mounted with cover glasses and allowed to dry in the dark. For fluorescence microscopy, the cells where captured using a Zeiss LSM 710 confocal microscope (Zeiss, Oberkochen, Germany).
The assay was preformed according to Alcock et al., [39]. Briefly, overnight HIB culture of Y. pestis strains bearing the pTorAsignal-GFP plasmid were grown at 28°C. Cultures were diluted the next morning to 0.1 OD660 and grown for 6 hours at 37°C. Cells were harvested and washed twice in 10 mM Tris.Cl, 150 mM NaCl, pH 7.3.
Equal volumes of the cell suspensions (10 OD660) were then centrifuged, and the cell pellets resuspended in 1ml SET buffer (17.12% sucrose (w/v), 3 mM EDTA, 10 mM Tris.Cl, pH 7.3), and left at room temperature for 20 min. Cell were concentrated in the 2 ml Eppendorf tubes at maximum speed for 10 min (20,000x g). The cell pellet was resuspended in 250 μl lysozyme (3 mg/ml in water) and 750 μl ice-cold water and incubated for 20 min at 37°C. Spheroplasts were released from the periplasm by centrifugation at maximum speed for 10 min (20,000x g). Samples were analyzed by immunoblotting for GFP (αGFP antibodies, G1546, sigma-aldrich, Israel) or the cytoplasmic marker protein Pcm [20]. The data presented is representative of at least two independent experiments.
Assessment of Sec functionality in Y. pestis strains was performed following transformation with two plasmids from the ASKA collection [26] encoding the Sec substrates BtuC and GadC fused to GFP. These plasmids were a kind gift from professor Eitan Bibi [40]. Y. pestis wild-type and ΔnlpD strains were grown on BHI agar plates with chloramphenicol (25 mg/ml) for 48 h. The cells were resuspended to a final concentration of 0.2 OD660/ml in PBS. Fluorescence of the ΔnlpD was quantified using a Victor3 (PerkinElmer) instrument with wavelength of 485nm (excitation) and 535nm (emission) and presented relatively to the wild type strain [40].
Female CD1 mice (5–6 weeks old) were purchased from Charles River (UK) and maintained under defined flora conditions at the animal facilities of the Israel Institute for Biological Research. The subcutaneous infections were performed as described previously [41]. Briefly, bacterial colonies grown on BHIA plates for 48 h at 28°C were harvested and suspended in saline solution to the required infection dose and quantified by counting cfu after plating and incubating on BHIA plates (48 h at 28°C). The intranasal (i.n.) infections were performed as described previously [20]. Briefly, bacterial colonies were grown in HIB supplemented with 0.2% (+) xylose and 2.5 mM CaCl2 and incubated overnight at 28°C. Cultures were diluted in saline solution to the required infection dose and quantified by cfu counting. Prior to infection, mice were anaesthetized with a mixture of 0.5% ketamine HCl and 0.1% xylazine was injected intraperitoneally followed by i.n infection with 35 μl/mouse of the bacterial suspension. The LD50 experiments were performed using groups of five mice. The LD50 values were calculated according to the method described by Reed and Muench [42],[15].
To evaluate the complementation of virulence by iron supplementation, mice (n = 6) received 5 mg iron-dextran (d8517, Sigma-Aldrich) intraperitoneally (i.p) 2 to 3 h before s.c. inoculation of 1×107 cfu of the Y. pestis strains. Beginning on the second day post-infection, iron dextran (5 mg/mouse) was administered once daily i.p. during the course of the experiment.
Statistical significance was measured using the Student’s unpaired t test. Survival curves were compared using the log-rank test. In all analyses, p values equal to 0.05 served as the limit of significance. Calculations were performed using GraphPad Prism 5 statistical pack.
To elucidate pathways by which NlpD is required during plague infection, we compared the transcriptional profiles of the parental virulent Y. pestis Kim53 strain and its isogenic nlpD mutant grown under conditions reminiscent of early stages of infection of the mammalian host.
Total RNA was prepared from Y. pestis cultures grown at 28°C, transferred to 37°C for five hours and then used as template for cDNA synthesis. Fluorescently labeled cDNA served as a probe for hybridization to a custom Y. pestis array [33]. A total of 220 genes were differentially expressed in the nlpD mutant (≥2-fold change in two experimental repeats, P<0.05) compared with the Y. pestis Kim53 strain. Among these genes, 113 were up-regulated (S2 and S4 Tables, S2A Fig) and 107 were down-regulated (S3 and S5 Tables, S2B Fig) in the mutant strain. The microarray data were validated by qRT-PCR analysis of the mRNA levels of selected genes. Plotting the data obtained using two complementary approaches revealed a strong positive correlation, confirming the microarray results (R2 = 0.846, see Supplementary S3 Fig, S6 Table).
Inspection of the functional annotation of the 220 differentially expressed genes (inferred from the NCBI and KEGG databases (https://www.ncbi.nlm.nih.gov/gene/ and http://www.genome.jp/kegg, respectively) indicated that membrane stress response was induced in the nlpD mutant (S2 Table) and iron-related metabolic pathways as well as transport systems for nutrients such as sulfate, arginine and sugar were differentially regulated (S2 Table and S3 Table). Two of the genes that were most significantly up-regulated in the mutant strain were cpxP and pspA (15.82-fold and 3.5-fold, respectively, S2 Table), belonging to the membrane stress response pathways Cpx (conjugative plasmid expression) and Psp (phage shock protein). These two pathways are involved in maintaining the homeostasis of the cytoplasmic membrane and preventing damage resulting from the accumulation of misfolded proteins in the periplasm [43,44]. These results suggest that in the absence of the NlpD lipoprotein, misfolded protein accumulation is increased in the periplasm. Activation of the Cpx and Psp stress response pathways may therefore represent a compensatory modality for retaining the integrity of the nlpD mutant membranes.
As indicated, many of the differentially modulated genes in the nlpD mutant were related to iron metabolism (~ 20%). Iron is an essential nutrient for most pathogenic bacteria and for Y. pestis in particular [45,46]. Some iron uptake systems involve an outer membrane receptor, a periplasmic binding protein and an inner membrane ATP-binding cassette (ABC) transporter. Coupling of the proton motive force of the cytoplasmic membrane to the outer membrane by the TonB, ExbB, and ExbD proteins provides the energy required for transport. Interestingly, genes that are up-regulated in the nlpD mutant (S2 Table) included the exbBD-tonB genes as well as the ybtA transcriptional regulator of the major iron acquisition system Yersiniabactin (Ybt) and the irp5 gene of this system required for synthesis of the Yersiniabactin siderophore [47]. Genes encoding additional iron uptake and storage systems such as, Yiu, Fit and the ferrichrome and ferrisiderophore receptor proteins were also up regulated in the nlpD mutant. These results, which indicate a compensatory up-regulation of iron uptake systems, strongly suggest that the nlpD mutant has a limited ability to acquire iron from the environment. Inspection of the down-regulated genes in the nlpD mutant suggest that in response to the apparent shortage of iron, deletion of the nlpD gene results in the decreased expression of several iron-containing proteins (fumarate reductase, dimethyl sulfoxide reductase and nitrate reductase) as well as the expression of the IscR transcription factor, which has been shown to modulate the expression of iron-sulfur protein clusters in Escherichia coli [48]. Similarly, the expression of proteins that are active in metabolic processes involving Fe-S-containing proteins was down regulated probably to preserve the small amount of intracellular iron for more essential metabolic pathways (S3 Table). The suggested paucity of iron in the nlpD mutant is also supported by the observed decreased expression of the ccmA-ccmH gene cluster (S3 Table) encoding a heme export system that functions in E. coli in cytochrome c maturation [49]. Iron is required for the activity of many enzymes of the tricarboxylic acid cycle, the cytochromes, non-heme iron electron carriers of the electron transport chain and nitrogen assimilation [50]. Indeed, metabolic pathways that utilize enzymes that are co-factored by iron were down-regulated in the nlpD mutant, leading to a metabolic shift in comparison to the wild-type strain (S3 Table). Taken together, these results strongly suggest that the nlpD mutant has a limited ability to acquire iron from the environment, a characteristic which could not have been anticipated solely from the documented function of NlpD in cell wall remodeling.
To confirm the hypothesis drawn from the transcriptomic analysis that the nlpD mutant has limited iron assimilation, the nlpD mutant and the parental Kim53 strains were analyzed using an in vitro growth assay under iron-limiting conditions [29]. Indeed, growth of the nlpD mutant was impaired in comparison to the wild-type strain under the iron-deficient conditions induced by addition of the iron-chelating agent 2,2’-dipyridyl (DIP) to the PMH2 defined medium (Fig 1A). Consequently, the limitation of the nlpD mutant to acquire iron during growth could also be manifested in vivo during infection and therefore may represent the cause for its avirulent phenotype in the mouse model of plague. To directly address this possibility, reversion of the avirulent phenotype of the nlpD mutant was attempted by exogenous administration of iron to the Y. pestis nlpD-infected mice. Such an approach has been previously described for studying the phenotype of an EV76 attenuated Y. pestis strain lacking the Ybt iron acquisition system carried in the pgm genomic region [51]. Thus, mice were subcutaneously infected with a single dose of 1×107 cfu of EV76 or with a similar dose of the attenuated nlpD mutant. As shown in Fig 1B, the virulence of the EV76 strain, but not of the nlpD mutant, was restored by daily injection of iron dextran.
A possible explanation for the inability to revert the attenuated phenotype of the nlpD mutant by exogenous addition of iron is the existence of additional defects that prevent establishment of infection by the nlpD mutant. Apart from attenuation of virulence, the Y. pestis nlpD phenotype has thus far been characterized to include, impaired cell septation, slight sensitivity to acidic stress conditions, and the above deficiency in iron acquisition (Fig 1A) [20]. This set of phenotypes was reminiscent of the phenotypes characterizing the twin-arginine translocation system (Tat) mutants of Gram-negative bacteria [52]. The Tat system typically consists of three cytoplasmic membrane proteins: TatA, TatB, and TatC, which are encoded by a single operon and are responsible for the transport of folded proteins across the cytoplasmic membrane [53,54]. Proteins that are translocated by the Tat system include cofactor-containing enzymes, multimeric proteins that require assembly prior to export as well as integral membrane proteins [53,55,56,57]. Tat substrates function in energy metabolism, cellular division, motility and adaptation to environmental stress [52,58]. The system has been shown to be important for virulence in many bacterial pathogens including Yersinia species [25,52,59,60,61].
To examine the functionality of the Tat system in the nlpD mutant, an established method based on the GFP reporter protein fused to the TorA Tat signal was used [62]. In the wild type Y. pestis background, GFP was localized to the periplasm and enriched at the cell poles whereas in Y. pestis TatA and TatC mutants, localization to the periplasm and poles was lost and the GFP reporter protein was diffused completely throughout the cytoplasm (Fig 2A). This observation is consistent with previous observations for other bacterial pathogens [25,60]. In the nlpD mutant, the GFP reporter was completely diffused throughout the cytoplasm, as observed for the Tat mutants, suggesting loss of Tat activity (Fig 2A). Of note, mutation of the tatB gene resulted in a polar phenotype abrogating expression of both TatB and TatC subunits (see Materials and Methods), therefore the tatB mutant was excluded from the current analysis. To further confirm the loss of Tat functionality in the nlpD mutant strain, the cellular localization of NapG, an additional Tat-substrate protein was investigated [60]. Accordingly, a hybrid protein consisting of the full length NapG protein and a C-terminal fused GFP reporter [26], was expressed in the parental wild-type, ΔnlpD mutant, as well as in the NlpD trans-complemented strain. The data (S4 Fig) clearly establish that the nlpD mutation is accompanied by the loss of fluorescence polarity that is characteristic for the Tat dysfunctionality. Furthermore this disturbed localization was fully alleviated upon trans-complementation of the nlpD mutant with an NlpD-expressing construct (S4 Fig) that was also shown to restore the wild-type cell morphology and virulence phenotype [20].
The direct assessment of Tat-substrate localization in nlpD mutant cells by microscopy described above was further substantiated by inspection of subcellular localization of a Tat substrate reporter in Tat mutant strains which were engineered by specifically disrupting expression of each of the Tat subunits (see Table 1). Thus, the functionality of the Tat system was interrogated implementing the molecular approach [39,63] consisting of Western blot analysis of fractionated material obtained from the various mutants as well as the parental strain. The Western-blot analysis of the subcellular fractions of the Y. pestis strains clearly confirmed that Tat transport was substantially affected in the nlpD mutant as seen for the tatA and tatC mutants (Fig 2B, S5 Fig). These results indicated, as anticipated, that the Tat system is not functional in the nlpD mutant and that the Y. pestis NlpD lipoprotein is required, directly or indirectly, for Tat system functionality.
Interestingly, the mRNA levels of all Tat genes (TatA, TatB and TatC) quantified in the nlpD mutant were indistinguishable to those of the wild-type strain, according to the microarray transcriptome analysis (http://www.ncbi.nlm.nih.gov/geo/, record GSE101490) and unambiguously confirmed by the qRT-PCR quantification of tatC mRNA (Fig 3A). However, Western blot analysis of TatC protein (used as a representative component for the Tat system) indicated that the protein level was reduced in the nlpD mutant in comparison to the wild-type strain (Fig 3B, S6 Fig). A dramatic difference between the wild type strain and the nlpD mutant was also observed in the confocal microscopy images of bacterial cells labeled with anti-TatC antibodies (Fig 3C). In these images, regions of fluorescence identified by anti-TatC antibodies were clearly visible throughout the cytoplasmic membrane of DAPI-stained wild-type Y. pestis bacteria (Fig 3C, upper panel), whereas no signal could be detected in DAPI-stained cells of the nlpD mutant (Fig 3C, lower panel).
Overall, these observations suggest that whereas the transcript levels of tat genes were not altered due to NlpD deletion, the nlpD mutant strain exhibited decreased levels of the Tat protein as indicated by monitoring TatC. These data raise the possibility that inactivation of the Tat system in the nlpD mutant could result from post-transcriptional events that affect proteins of the Tat machinery.
To verify that the loss of Tat functionality did not result from a general destabilization of the cytoplasmic membrane, the activity of Sec machinery, an additional inner-membrane imbedded transport system, was assessed. The Sec machinery is essential for the transport to the periplasm of the Y. pestis F1 capsular protein [64,65]. As depicted in Fig 4A, a similar distribution of the F1 protein was observed in the outer membrane of the wild-type strain and the nlpD mutant. In addition, the Sec pathway substrates BtuC and GadC were effectively expressed and transported to the periplasm of both wild-type and nlpD strains (Fig 4B and 4C), indicating that the Sec translocon is operational in the nlpD mutant in a similar manner to the wild-type strain. These results strongly suggest that the loss of Tat translocation activity in the nlpD mutant did not result from a general destabilization of the cytoplasmic membrane.
As mentioned above, one of the major phenotypic characteristics of the nlpD mutant strain is its virulence attenuation. Therefore, we addressed the possibility that the NlpD associated effect on pathogenicity maybe attributed to the dysfunctionality of the Tat system. Indeed, it has been shown that the TatA protein is important for Y. pestis virulence in mouse models of plague [61]. Thus, we further addressed the attenuation of virulence and other phenotypic characteristics associated with the deletion of Tat proteins, in mutant strains exhibiting disruption of Tat genes in comparison to the phenotype exhibited by the nlpD mutated strain. Accordingly, the various mutants were characterized with respect to their morphology, iron acquisition capability and virulence in mouse models of plague.
Microscope analyses revealed a defect in cell segmentation of the tatA deletion mutant (Fig 5A). In addition, growth of the tatA and tatC mutants was severely inhibited under iron-limiting conditions in comparison to the wild-type strain, similar to the nlpD mutant (Fig 5Bi, Fig 1 and Table 2). Trans-complementation of the nlpD, tatA and tatC mutants with each of the corresponding genes (nlpD or tatA or tatC, respectively) was efficient in alleviating the growth under these conditions (Fig 5Bi and Table 2). Increasing the amounts of the iron chelator (DIP) to 100 μM resulted in growth inhibition of all Y. pestis strain (Fig 5Bii and Table 2). Addition of iron dextran (0.5mg/ml) to the plates containing the high concentration of DIP restored growth of the parental Kim53 strain as well as the trans-complemented nlpD/pnlpD, tatA/ptatA and tatC/ptatC, but not of tatA, tatC and nlpD mutants (Fig 5Biii and Table 2). Growth of the tatA, tatC and nlpD mutants was inhibited under iron limiting conditions and was not restored by addition of iron but only upon trans-complementation with the relevant gene (nlpD or tatA and tatC respectively, Fig 5Biii, Table 2).
Virulence of the Y. pestis tat mutants was evaluated in the well-established mouse models of bubonic and pneumonic plague [17,66]. As shown in Table 3, infection of CD-1 mice with high doses of the tatA mutant via the subcutaneous and intranasal routes was non-lethal. Furthermore, the tatA mutated bacteria could not be detected in the draining lymph node or the spleen of mice 48 hours after s.c. inoculation with high doses (107 cfu/mouse), similar to the observations reported with respect to the nlpD mutated bacteria [20]. The results obtained with the s.c. administration of the TatA mutant are consistent with those of Bozue and colleagues [61], who showed that a Y. pestis CO92 tatA mutant was severely attenuated upon subcutaneous infection of Swiss Webster mice. As with the nlpD mutant (Fig 1B), the avirulent phenotype of the Kim53tatA mutant could not be reversed by exogenous administration of iron to the infected mice in line with the in-vitro growth results (Fig 5Biii). Yet, complementation by episomal expression of the tatA gene which restored the parental cell morphology and growth on iron-depleted medium (Fig 5 and Table 2) also reverted the virulent phenotype (Table 3), confirming that the observed phenotype is attributed to abrogation of TatA expression.
Of note, similar results to those pertaining to the TatA mutation could be obtained upon deletion of the tatC gene, including full restauration of the wild-type phenotype upon trans-complementation (morphology, Fig 5A, growth under iron deprivation, Fig 5B and virulence in the plague murine models, Table 3).
Taken together the data support the notion that the virulence attenuation characterizing the NlpD mutant phenotype maybe attributed to the dysfunctionality of the Tat system via a possible involvement in iron exploitation. This general possible explanation is compatible both with the micro-array transcriptomic data and the direct inspection of the Tat mutants.
The lipoprotein NlpD emerged in a previous functional screen of Y. pestis genome as an important factor for the manifestation of Y. pestis virulence. The screen evidenced that NlpD gene disruption is incompatible with the survival of the bacteria in the host during infection [15,20]. In the current report, to gain further insight into the mechanisms underlying the role of NlpD in Y. pestis pathogenicity, the transcriptomes of the wild type Kim53 strain and the nlpD mutant were compared. Considering that the NlpD-mutated bacteria were rapidly cleared from inoculated animals [20], RNA for the transcriptome analysis was collected from bacteria cultured under conditions reminiscent of the initial stages of infection (growth of Y. pestis cultures at 37°C for several hours).
Examination of the differential transcriptome data clearly indicated that a pronounced membrane stress response is specifically induced in the nlpD mutant, as reflected by up-regulation of the cpxP and pspA genes (S2 Table). The Cpx and Psp systems are membrane stress-response pathways of Gram-negative bacteria that are involved in maintaining the homeostasis of the cytoplasmic membrane [43,44,67]. These systems sense and respond to periplasmic or cytoplasmic protein misfolding that disturb the integrity of the cytoplasmic membrane and could reduce the energy status of the cell [67,68,69,70]. In Y. enterocolitica, the Psp system has been shown to be essential for protecting bacterial cells against membrane damage due to miss-localization of the T3SS component secretin (YscC) that is induced at the mammalian body temperature of 37°C [71,72,73]. In the present study, Y. pestis cultures were grown at 37°C for five hours, resulting in induction of the T3SS [16]. However, inspection of the transcriptome data indicated that the expression of secretin was not differentially regulated in the nlpD mutant. The expression of only two T3SS components was altered in the nlpD mutant (S2 Table), the YscB chaperon that is required for the calcium-dependent regulation of Yop secretion [74] and YopQ (also known as YopK), which plays an important role in the regulation of Yops translocation [75].
The robust induction of Cpx and Psp stress response systems in the nlpD mutant suggests that in the absence of NlpD the integrity and/or stability of the membranes are affected and there is an increase in the accumulation of misfolded proteins in the periplasm. Interestingly, NlpE, which is another outer-membrane lipoprotein, is an accessory protein of the Cpx pathway in E. coli [76]. However, nlpE expression is not altered in the nlpD mutant compared with the wild-type strain. One may speculate that NlpD plays a similar role in responding to harmful changes that occur in bacterial membranes following exposure to environmental stress. Similarly, it has recently been demonstrated that changes in the peptidoglycan structure are part of the Cpx-mediated adaptation to envelope stress [77]. The proposed involvement of NlpD in the response to extracytoplasmic stress conditions is compatible with the genomic localization of the nlpD gene within a genomic stress response locus from which the SurE, Pcm and RpoS proteins are expressed [20].
Many of the differentially modulated pathways in the nlpD mutant were related to iron metabolism (~20%, S2 and S3 Tables), suggesting that the acquisition and consumption of iron may have been perturbed by NlpD deletion. The hypothesis that the mutated cells have an impaired ability to exploit iron was further confirmed by the observation that the mutated bacterial cells failed to grow under iron-limiting conditions (Fig 1A). Lowering the levels of free iron is an important host defense strategy that restricts the proliferation of infectious bacteria [78,79], and many pathogens have evolved sophisticated mechanisms to overcome this restriction and maximize the host iron supply [78,80]. Accordingly, we have recently shown that the Y. pestis EV76 live vaccine protected mice against an immediate lethal challenge with a virulent Y. pestis strain and that protection was associated with induction of the host heme- and iron-binding proteins hemopexin and transferrin [81].
Close inspection and integration of all the observed phenotypic characteristics of the Y. pestis nlpD mutant, namely, chain-forming morphology, attenuation of virulence, reduced tolerance to acidic stress, defective iron acquisition and envelope stresses, suggested a striking resemblance to the phenotypic characteristics of bacterial Tat mutants in other pathogens [52,53]. Importantly, similar to the situation of the Y. pestis nlpD mutant, the loss of virulence of Yersinia Tat mutated strains could not be explained by the dysfunction of the T3SS [60,61].
The assumption that the Tat system is impaired in the nlpD mutant was supported by the observed decrease in the expression of several known iron-sulfur protein substrates of this system (S2 Table). In addition, the modulation of type VI secretion system genes observed in the nlpD mutant (S2 Table) was also observed in Tat mutants of the phylogenetically related strain Y. pseudotuberculosis [82], and the fish pathogen E. tarda [83]. Furthermore, although Y. pestis is a non-motile bacterium, a flagellar operon was modulated in the nlpD mutant (S2 Table), suggesting a possible control on motility that characterize Tat mutants in many bacterial pathogens including Y. pseudotuberculosis [52,59,60]. The assumption that the phenotype of the NlpD mutant is related to deregulation of the Tat system was directly interrogated. Assessment of the functionality of the Tat system in the nlpD mutant was performed by visualization (using microscope analysis and confirmed by subcellular fractionation analysis) of the localization of two different Tat reporter substrates: a GFP-reporter protein fused to the signal peptide of the Tat-substrate TorA or the Tat-substrate NapG [25,60,62]. As hypothesized, the Tat system was not functional in the Y. pestis nlpD mutant (Fig 2), whereas other translocation systems including the inner-membrane embedded Sec pathway (Fig 4), and the T3SS that traverse the inner and outer membrane of the cell [20], were operational in the nlpD mutant indistinguishably from the wild-type strain. These observations suggested that inactivation of the Tat system in the nlpD mutant did not result from a general destabilization of the bacterial membrane and substantiate the specificity of the phenotype exhibited by the nlpD mutant strain.
TatC protein levels (Fig 3B) as well as the membrane localization of this protein (Fig 3C) confirmed that the Tat system is affected in the nlpD mutant. The TatC protein level in cell lysates of the nlpD mutant was decreased in comparison to the wild-type strain, and this protein was not detected in the mutant cytoplasmic membrane. Since the RNA levels of the Tat genes were similar in wild-type and the nlpD mutant, one may speculate that dysfunction of the Tat system in the nlpD mutant could have resulted from post-transcriptional molecular events. In Y. pseudotuberculosis, a tatC mutant was highly attenuated for virulence following oral or intraperitoneal infections [60]. The system was important for different virulence-related stress responses as well as for iron uptake [60]. Additional studies have indicated that the loss of virulence is related to the SufI Tat-substrate that was found to be required for establishment of systemic infection [84].
The Y. pestis tatA and tatC mutants were avirulent in mice. In Y. pestis CO92, the tatA mutant was highly attenuated in the bubonic and aerosol-infection mouse model but to a lesser extent in the intranasal infection model [61]. Attenuation of virulence in the bubonic plague model is therefore similar in both Y. pestis tatA mutant strains. The differences between the virulence characteristics of the Kim53 and CO92 tatA mutants in the i.n. infection model may reflect variations in the mouse strain used for the evaluation of virulence or the genetic diversity between the two Y. pestis strains.
The known functions of the NlpD lipoprotein, which belong to the M23-LytM endopeptidase family, involves the regulation of peptidoglycan hydrolysis and cell morphogenesis [20,21,85,86]. In E. coli, the NlpD protein is not catalytically active but controls the activity and recruitment to the septum of the cell wall amidase—AmiC, which is a known Tat system substrate [85,87,88,89]. Interestingly, the Y. pestis amiC mutant retained the wild-type single cell morphology (S7 Fig) and virulence characteristics (Table 3). These observations suggest that in Y. pestis the mode of interactions between NlpD and AmiC maybe different than in E. coli.
The present study shows that the nlpD mutant exhibited impaired Tat activity as well as limited iron acquisition. Both of these characteristics may have represented the cause for the severe virulence attenuation exhibited by the nlpD mutant. Considering the similarity between the phenotypic characteristics of the nlpD mutant and the tat mutants including chain morphology, iron assimilation defect and loss of virulence, the present data suggest a novel link between NlpD and the Tat system affecting Y. pestis pathogenesis.
The molecular mechanisms underlying the possible relationships between NlpD, the Tat system components and iron assimilation remain to be deciphered and raise several questions including the possible interactions between the outer membrane-predicted NlpD lipoprotein, and iron assimilation systems or the inner membrane Tat proteins. Studies addressing some of these issues are currently underway in our laboratory.
|
10.1371/journal.ppat.1005893 | DNA Polymerase κ Is a Key Cellular Factor for the Formation of Covalently Closed Circular DNA of Hepatitis B Virus | Hepatitis B virus (HBV) infection of hepatocytes begins by binding to its cellular receptor sodium taurocholate cotransporting polypeptide (NTCP), followed by the internalization of viral nucleocapsid into the cytoplasm. The viral relaxed circular (rc) DNA genome in nucleocapsid is transported into the nucleus and converted into covalently closed circular (ccc) DNA to serve as a viral persistence reservoir that is refractory to current antiviral therapies. Host DNA repair enzymes have been speculated to catalyze the conversion of rcDNA to cccDNA, however, the DNA polymerase(s) that fills the gap in the plus strand of rcDNA remains to be determined. Here we conducted targeted genetic screening in combination with chemical inhibition to identify the cellular DNA polymerase(s) responsible for cccDNA formation, and exploited recombinant HBV with capsid coding deficiency which infects HepG2-NTCP cells with similar efficiency of wild-type HBV to assure cccDNA synthesis is exclusively from de novo HBV infection. We found that DNA polymerase κ (POLK), a Y-family DNA polymerase with maximum activity in non-dividing cells, substantially contributes to cccDNA formation during de novo HBV infection. Depleting gene expression of POLK in HepG2-NTCP cells by either siRNA knockdown or CRISPR/Cas9 knockout inhibited the conversion of rcDNA into cccDNA, while the diminished cccDNA formation in, and hence the viral infection of, the knockout cells could be effectively rescued by ectopic expression of POLK. These studies revealed that POLK is a crucial host factor required for cccDNA formation during a de novo HBV infection and suggest that POLK may be a potential target for developing antivirals against HBV.
| HBV chronically infects 240 million people worldwide. Persistent HBV infection relies on stable maintenance of the nuclear form of viral genome, the covalently closed circular (ccc) DNA. However, the molecular mechanism underlying the conversion of HBV genomic relaxed circular (rc) DNA into cccDNA remains elusive. Our studies reported herein provide unambiguous evidence suggesting that host DNA polymerase κ (POLK) is required for repairing the gap of rcDNA and formation of cccDNA in a de novo HBV infection. POLK is thus a potential therapeutic target for treatment of chronic hepatitis B.
| Despite the availability of effective vaccines for more than three decades, hepatitis B virus (HBV) still persistently infects 240 million people worldwide [1, 2]. Antiviral therapies with nucleos(t)ide analog inhibitors of HBV reverse transcriptase dramatically reduce the viral load, significantly improve the liver function and lower the incidence of liver failure and hepatocellular carcinoma, but fail to cure the viral infection [3, 4], due to the persistence of covalently closed circular (ccc) DNA in the nuclei of infected hepatocytes [5–8]. Hence, better understanding the molecular mechanisms underlying the formation and maintenance of cccDNA is critical for development of novel therapeutics to cure chronic HBV infection.
HBV is the prototype member of Hepadnaviridae family and contains a relaxed circular (rc) partially double-stranded DNA genome with its DNA polymerase covalently linked to the 5’ terminus of minus strand [9]. While the minus strand of the rcDNA is completely synthesized with a redundant overhang at the 5’ end, the plus strand is incompletely synthesized, leaving a 3’ terminal gap of variable length [9]. HBV replicates its DNA genome via reverse transcription of an RNA intermediate, the pregenomic (pg) RNA [10]. Briefly, HBV entry into hepatocytes is mediated by its host cellular receptor human sodium taurocholate cotransporting polypeptide (NTCP) [11–14]. Upon entry into the cytoplasm of hepatocytes, rcDNA in the nucleocapsid is transported into the nucleus and converted into an episomal cccDNA, which is assembled into a minichromosome to serve as the template for the transcription of viral mRNAs [15, 16]. Following the synthesis of viral proteins, viral DNA polymerase binds to a stem-loop structure (termed epsilon) within the 5’ region of pgRNA to initiate its packaging into nucleocapsids where the pgRNA is reverse transcribed to progeny rcDNA [17]. The progeny “mature” rcDNA-containing nucleocapsids can be either enveloped and secreted out of the cell as virion particles or might be redirected into the nucleus to amplify the cccDNA pool [18–20] [21]. Thus, the formation and intracellular amplification of cccDNA plays a central role in the establishment and maintenance of persistent infection. Biochemically, conversion of rcDNA to cccDNA requires the removal of the viral DNA polymerase and RNA primer from the 5’-terminus of minus strand and plus strand DNA, respectively; filling in the gap in plus strand DNA, trimming and ligating the ends of both strands. Although it is speculated that all those reactions are most probably catalyzed by host cellular DNA repair enzymes, identification of the cellular proteins responsible for cccDNA formation has thus far only achieved limited success. For instance, tyrosyl-DNA phosphodiesterase-2 (Tdp2), a cellular enzyme responsible for cleavage of tyrosyl-5' DNA linkages formed between topoisomerase II and cellular DNA [22], can release covalently linked RT from the 5’ end of minus-strand DNA in vitro [23, 24], and has recently been shown to cleave the tyrosyl-minus strand DNA linkage of HBV. However, Tdp2 gene knockout only slows down the formation of duck hepatitis B virus (DHBV) cccDNA from intracellular amplification pathway, but does not inhibit HBV cccDNA formation in HBV infection of HepG2-NTCP cells [25, 26]. In addition to rcDNA, cccDNA can also be formed from double stranded linear DNA (dslDNA) [27], which is derived from in situ priming of plus strand DNA synthesis [28]. Interestingly, Ku80, a component of non-homologous end joining DNA repair pathway, has been reported to play an essential role in the synthesis of DHBV cccDNA from dslDNA, but not rcDNA [29].
Completion of plus strand DNA synthesis, or “filling the gap” in the plus strand of rcDNA, is essential for cccDNA synthesis. Studies of DHBV and woodchuck hepatitis virus (WHV), two hepadnaviruses distinct from human HBV but readily cultivable in vitro, showed that viral DNA polymerase inhibitors did not prevent cccDNA formation in the infection of primary hepatocytes of ducks and woodchucks, implying that viral DNA polymerase may be dispensable, while cellular DNA polymerase activity is required for the completion of plus strand DNA synthesis [30–33]. Moreover, continued treatment of primary tupaia hepatocytes and HepaRG cells with viral polymerase inhibitors during HBV infection did not inhibit HBV cccDNA formation [34, 35], further suggesting that cellular factor(s) play an important role. Here we set out to identify the DNA polymerase(s) that complete(s) the plus strand DNA synthesis required for HBV cccDNA formation. In HepG2-NTCP cells infected with HBV that is deficient in core protein production, we unambiguously demonstrated that viral DNA polymerase activity is not required for cccDNA formation in a de novo infection. Instead, a focused RNA interference loss-of-function screening identified POLK as a crucial cellular polymerase supporting HBV infection. Both knockdown and knockout of POLK impaired the conversion of rcDNA to cccDNA, which could be rescued by ectopic expression of POLK. Our findings thus suggest that POLK is a key host factor required for cccDNA formation during a de novo HBV infection, and therefore, a potential target for therapeutic intervention of chronic hepatitis B.
To investigate the molecular mechanism of HBV cccDNA formation, we first determined the kinetics of cccDNA synthesis in wild-type HBV infected HepG2-NTCP cells. Using Southern blot analysis, we examined HBV cccDNA in Hirt extracts of infected cells at various time points post infection. The identity of the cccDNA, which migrated at the position of 2.1kb linear DNA, was confirmed by the band shift to a 3.2kb DNA species corresponding to the size of unit-length linear HBV genomic DNA upon digestion by EcoRI, but not HindIII (S1 Fig). As shown in Fig 1A, the protein-free rcDNA species accumulated at 12 and 24 h post infection and cccDNA became detectable at day 2 post-infection, followed by a modest increase in the next 5 days. The appearance of cccDNA was coincident with reduction of the protein-free rcDNA in the infected cells. Consistently, based on the quantitative analysis of cccDNA using a more sensitive real-time PCR assay, cccDNA was detectable at 24 h post-infection, markedly increased in the first 2 days post-infection, followed by a slower increase to approximately 3 copies per infected cell in the next 5 days (Fig 1B). Similar to the kinetics of cccDNA synthesis, the levels of intracellular HBV 3.5kb vRNA (Fig 1C) as well as secreted HBeAg and HBsAg (Fig 1D) gradually increased following HBV infection. Immunostaining revealed that over 60% of HepG2-NTCP cells were HBcAg positive at day 7 post infection (Fig 1E). As expected, HBV preS1 myr-47 lipopeptide (myr-47) completely blocked the viral infection. Together, these results demonstrated that HepG2-NTCP cells support an efficient HBV infection, resulting in readily detectable cccDNA formation, gene expression and secretion of viral proteins. Hence, the HBV infection cell culture system is suitable for identification of viral and host cellular factors required for cccDNA formation during a de novo HBV infection.
As stated above, nuclear HBV cccDNA pool is established by direct conversion of rcDNA from incoming virions and supposedly via intracellular amplification pathway from rcDNA in the cytoplasmic progeny nucleocapsids. It is clear that viral DNA polymerase activity is essential for the intracellular amplification of cccDNA, but its role in cccDNA formation from incoming virions, specifically, in filling the gap in the plus strand of rcDNA, remains to be rigorously examined. Providing an unambiguous answer to this question is challenging due to the apparently indistinguishable nature of cccDNA synthesized from the two different pathways. In order to thoroughly determine the role of HBV DNA polymerase in cccDNA formation during a de nove infection, we first produced HBV virion particles containing genomic DNA with a stop codon at the 38th codon (Y) of the core gene open reading frame, designated as HBV-ΔHBc. As depicted in S2 Fig, this was achieved by co-transfection of Huh-7 cells with plasmid containing 1.05-mer HBV DNA with the desired mutation and plasmid expressing HBV core protein. HBV-ΔHBc virions were harvested from the culture fluid and purified by ultracentrifugation. Due to its inability to produce capsid protein, HBV-ΔHBc infection of cells will not be able to support progeny viral DNA synthesis and formation of cccDNA through the intracellular amplification pathway. Hence, HBV-ΔHBc infection provides a unique opportunity to investigate the role of DNA polymerases in cccDNA formation from incoming virions. HepG2-NTCP cells were infected with a multiplicity of 100 genome equivalents (mge) of wild-type HBV and HBV-ΔHBc virions, respectively. Notably, HBV-ΔHBc successfully infected HepG2-NTCP cells and formed cccDNA at a level comparable to that of wild-type HBV on day 7 post infection and persisted during the following two weeks of extended culturing (Fig 2A). Similarly, immunostaining analysis showed that the levels of intracellular HBsAg were similar in both wild-type HBV and HBV-ΔHBc infected cells during the prolonged 21 days of culture (Fig 2B). As expected, HBV-ΔHBc infected cells did not produce HBeAg in the culture supernatant due to the stop codon mutation. Meanwhile, HBV-ΔHBc infected cells secreted a slightly higher level of HBsAg than wild-type HBV infected cells at the indicated time points post infection (Fig 2C). Together, the results indicate that cccDNA in HBV infected HepG2-NTCP cells were mainly synthesized from input viral rcDNA and that the intracellular amplification pathway did not significantly contribute to the establishment of cccDNA pool under this experimental condition.
To directly examine whether HBV polymerase activity is required for conversion of viral rcDNA from the input viruses to cccDNA, HepG2-NTCP cells were infected with HBV-ΔHBc in the presence or absence of adefovir (ADV) or entecavir (ETV). As shown in Fig 2D and 2E, neither of the compounds prevented cccDNA accumulation in HBV-ΔHBc infected cells. Quantitative analysis indicated that the viral DNA polymerase inhibitor treatment at relatively high concentration reduced cccDNA accumulation by less than 27%. Interestingly, similar ADV and ETV treatment of wild-type HBV infected cells reduced cccDNA accumulation by 28 to 42% (S3 Fig). These results indicate that the activity of HBV DNA polymerase is dispensable in cccDNA formation from incoming virion rcDNA, and re-enforce the notion that intracellular amplification of cccDNA does not play an important role in establishment of cccDNA pool in the HBV-infected hepatoma cells.
There are fifteen different DNA polymerases in mammalian cells. They function in genome replication, DNA repair, and translesion DNA synthesis (TLS) of damaged DNA. To identify which host DNA polymerase contributes to HBV infection, we conducted a focused siRNA screening by targeting cellular DNA polymerase genes. We assessed HBV infection upon silencing the expression of individual DNA polymerase gene in HepG2-NTCP cells. All siRNA sequences targeting cellular DNA polymerase genes were obtained from previous studies; these individual siRNAs could unequivocally discriminate between the mRNAs of the 15 different polymerases [36]. An siRNA targeting NTCP and a scrambled sequence (siRNA NC) were used as positive and negative controls, respectively. The levels of intracellular 3.5kb viral RNA as well as secreted HBeAg in culture supernatants were measured by qRT-PCR and ELISA. As shown in Fig 3, among 15 cellular DNA polymerases, knocking down POLK, POLH or POLL gene expression significantly decreased the levels of intracellular 3.5kb viral RNA as well as secreted HBeAg, POLB and POLD2 silencing only led to a modest decrease of intracellular 3.5kb viral RNA production. Of note, no significant cytotoxic effect was observed upon silencing any of the host DNA polymerases (S4 Fig). To confirm the specificity of the siRNA in the targeted RNAi screen, we evaluated the efficacy of single and combined dual siRNA-mediated knockdown of POLK, POLL and POLH genes (S5 Fig and Fig 4), respectively. Western blot analysis or qRT-PCR assay showed that each of the siRNAs specifically reduced the expression of its targeted DNA polymerase gene (S5A Fig and Fig 4). Among the three cellular polymerases, silencing of POLK exhibited the most dramatic inhibitory effect on HBV infection, with approximately 74% reduction of HBV infection as judged by HBeAg level, which was comparable to the efficiency of silencing NTCP expression that reduced HBV infection by 80% (S5C Fig). Similar extents of reduction were also observed for intracellular 3.5kb vRNA (S5B Fig) and HBcAg levels (S5D Fig) in POLK siRNA transfected cultures. Knock-down of POLL gene also led to a notable decrease of intracellular 3.5kb viral RNA and HBcAg. Interestingly, at the same total concentrations of siRNAs, compared with only knock-down POLK gene (by siRNAs of NC and POLK), dual siRNAs-mediated knockdown of POLK and POLL, or, POLK and POLH demonstrated enhanced inhibition of HBV infection, but did not completely abolish HBV infection (S5E–S5G Fig). These results thus suggest that while POLK, POLL and POLH each individually could contribute to de novo HBV infection, POLK clearly plays a critical role.
POLK belongs to the Y family of DNA polymerases, which functions in translesion synthesis and nucleotide excision DNA repair. Its enzymatic activity is resistant to aphidicolin (APH) and dideoxynucleotides [37]. In line with this, APH treatment did not inhibit HBV cccDNA synthesis and secretion of HBeAg in HBV infection of HepG2-NTCP cells (S6 Fig). The results also imply that APH-sensitive DNA polymerases (e.g. POLA, POLE) are not required for HBV infection of HepG2-NTCP cells, which is consistent with the results obtained from siRNA knockdown experiments.
To confirm the role of POLK in HBV cccDNA formation, two individual siRNAs targeting different region of POLK gene were transiently transfected into HepG2-NTCP cells. The siRNAs significantly reduced the expression levels of endogenous POLK at day 3 and 5 post transfection (Fig 4A). We found that POLK knockdown led to 70% reduction of cccDNA levels compared to that in control siRNA transfected cells on day 7 post infection (Fig 4B). In line with this observation, the 3.5kb vRNA level was also reduced by approximately 60% in POLK knockdown cells (Fig 4C). HBeAg and HBsAg levels in culture supernatants were reduced to less than 30% of that from control siRNA transfected cells (Fig 4D). Immunostaining of intracellular HBcAg also showed significant decrease in POLK knockdown cultures (Fig 4E).
To investigate whether POLK plays a similar role in other in vitro HBV infection experimental models, we conducted siRNA-mediated knockdown of POLK in HepaRG cells (Fig 4F–4H) and primary tupaia hepatocytes (PTH) (Fig 4I and 4J), respectively. Compared to a negative control (NC), differentiated HepaRG cells treated with two individual siRNAs targeting POLK reduced the expression level of POLK, decreased HBV cccDNA synthesis and reduced secretion of HBeAg and HBsAg. No cytotoxic effect was observed at 5 days post siRNA transfection (Fig 4G). Consistent with the results for HBV infected HepG2-NTCP and HepaRG cells, silencing of tupaia POLK by siRNA (sitsPOLK) led to a marked decrease of cccDNA synthesis and production of HBeAg and HBsAg on day 7 post infection of PTH. Moreover, similar to infection by wild-type HBV, transfection of sitsPOLK also significantly reduced the levels of intracellular HBV cccDNA and HBsAg in HBV-ΔHBc infected PTHs (Fig 4J). Importantly, siRNA knockdown of POLK did not reduce the infection efficiency of HDV (S7A Fig) or an EGFP-encoding VSV-G pseudotyped lentivirus (VSV-EGFP) (S7B Fig), demonstrating that POLK has a specific role in HBV infection. Together, these data suggest that POLK is required for de novo HBV infection and depletion of POLK diminishes HBV cccDNA synthesis.
In order to further confirm that POLK is responsible for formation of HBV cccDNA, we intended to restore POLK expression in the siRNA transfected cells. To achieve this goal, an expression plasmid of POLK fused with GFP at N-terminus (GFP-POLK-wt) was constructed. Silent mutations were introduced to siPOLK1-targeting sequence for expression of the siRNA-resistant POLK mRNA (GFP-POLK-res). HepG2-NTCP cells were transduced with VSV-G protein pseudotyped lentiviruses expressing GFP-POLK-res, GFP-POLK-wt or GFP alone, respectively. The cells were then transfected with siPOLK-1. Fluorescence microscopic analysis showed that POLK localized in the nuclei, and siPOLK-1 dramatically reduced the expression of GFP-POLK-wt, but not GFP-POLK-res, suggesting that the expression of GFP-POLK-res is indeed resistant to siPOLK-1 (S8A Fig). We next performed HBV infection assay. As shown in S8B–S8D Fig, while siPOLK-1 transfection efficiently reduced HBV cccDNA formation as well as 3.5 kb vRNA expression and HBeAg secretion in HepG2-NTCP cells transduced with control lentivirus expressing GFP, the effects of siPOLK-1 transfection on cccDNA formation and function in HepG2-NTCP cells expressing POLK, in particular GFP-POLK-res were significantly attenuated. The results thus suggest that ectopic expression of POLK partially rescued the suppression of HBV cccDNA formation caused by POLK-targeting siRNA.
To rigorously determine the function of POLK in HBV cccDNA synthesis, we took advantage of the CRISPR/Cas9 system to generate POLK knockout in HepG2-NTCP cells (S9 Fig). We first created a stable HepG2-NTCP/Cas9 cell line that constitutively expresses Cas9. The cell line was then infected with lentivirus encoding an EGFP protein and sgRNA targeting exon 2 of polk gene. By monitoring the expression of EGFP, we could assess the sgRNA transduction efficiency. On day 3 post-transduction, EGFP positive cells were sorted and expanded by culturing for additional 10 days. Independent HepG2-NTCP clones with successful polk gene editing (HepG2-NTCPpolk+/-and HepG2-NTCPpolk-/-) were identified and used for further studies. Sequencing analysis revealed that the clones have a frame shift in the coding region owing to nucleotide deletions, which resulted in the disruption of intact POLK protein expression. Western blotting analysis confirmed that the expression of POLK protein was reduced in HepG2-NTCPpolk+/- and abolished in HepG2-NTCPpolk-/- clones (Fig 5A). Knockout POLK in HepG2-NTCP cells did not affect the viability of HepG2-NTCP cells. NTCP level remained unchanged as compared to that in parental HepG2-NTCP cells (S10A Fig). Consistently, a functional assay showed that HepG2-NTCPpolk-/- cells were able to uptake [3H]-labeled taurocholate (S10B Fig) and supported HDV infection (S10C Fig) at an efficiency similar to that of the parental HepG2-NTCP cells.
We next carried out HBV infection assay with these stable cell lines described above. The levels of cccDNA were examined by Southern blotting analysis at day 7 post-infection. No visible band of cccDNA could be detected in HepG2-NTCPpolk-/- cells, indicating that lack of POLK impaired HBV cccDNA synthesis (Fig 5B). Quantitative PCR also showed that the amounts of cccDNA decreased by approximately 3- or 4-fold in the cells with reduced (HepG2-NTCPpolk+/-) and abolished (HepG2-NTCPpolk-/-) POLK expression, respectively (Fig 5C). ELISA analysis showed that secreted HBeAg and HBsAg levels were markedly decreased from HepG2-NTCPpolk-/- cells (Fig 5E). Similarly, depletion of POLK also dramatically reduced intracellular HBcAg, which was closely related to intracellular 3.5kb vRNA levels (Fig 5D and 5F). Importantly, a time course analysis using Southern blot assay showed that cccDNA was readily detectable in parental HepG2-NTCP cells at day 2 post infection and modestly increased in the next 4 days. In contrast, cccDNA only became detectable in HepG2-NTCPpolk+/ -cells containing one intact polk allele at day 4 and day 6 post infection in much reduced amounts, compared to that in the parental HepG2-NTCP cells at the same time points. Only very faint cccDNA bands could be detected in HepG2-NTCPpolk-/- cells with both polk alleles disrupted during the same time period (Fig 5G). In agreement with the cccDNA formation results, POLK knockout cells produced low level HBeAg and HBsAg.
To confirm the function of POLK in HBV cccDNA synthesis, we restored POLK expression in HepG2-NTCPpolk-/-cells by stable transduction of lentivirus expressing POLK. HepG2-NTCPpolk-/- cell clones with defective endogenous polk but expressing exogenous POLK were established. Production of POLK protein was largely restored in two independent cell clones as demonstrated by Western blot analysis (Fig 6A). The amount of cccDNA upon HBV infection was assessed by Southern blot. Remarkably, HBV infection in these two clones was rescued, and the cccDNA level correlated with the expression levels of POLK in these cells (Fig 6B). In contrast, transduction of HepG2-NTCPpolk-/- cells with a control lentiviral vector did not rescue cccDNA synthesis. Consistently, ELISA analysis for HBeAg in culture supernatant and immunofluorescence staining of intracellular HBcAg also confirmed that restoration of POLK expression in the HepG2-NTCPpolk-/- cells efficiently rescued not only cccDNA formation, but also viral gene transcription and protein expression (Fig 6C and 6D).
Considering that knockout of POLK did not completely abolish HBV infection and cccDNA formation, and knockdown of POLL and POLH with siRNA also reduced HBV infection despite at a lesser extent, we further investigated the role of POLL in HBV infection with more rigorous experimental conditions. We accordingly generated POLL knockout cell lines with CRISPR/Cas9 technology. Infection of poll gene edited HepG2-NTCPpoll+/- and HepG2-NTCPpoll-/- cell clones (Fig 7A) with HBV demonstrated reduced levels of intracellular cccDNA (Fig 7B and 7C), 3.5kb vRNA (Fig 7D) and HBcAg (Fig 7F) as well as secreted HBeAg (Fig 7E) at day 7 post-infection, as compared to that in the parental HepG2-NTCP cells. However, the extent of POLL depletion on cccDNA formation and viral RNA and protein expression was less than that of POLK depletion. Taken together, our results strongly suggest that while POLK, POLL and POLH are each capable of supporting cccDNA synthesis at a different efficiency during a de novo HBV infection, POLK plays a more dominant role under the infection conditions examined in this study.
Synthesis of cccDNA is a critical, but not well-understood step in the life cycle of hepadnaviruses. Our current study characterized the kinetics of cccDNA formation in HBV infected cells and obtained strong evidence suggesting that cellular POLK plays a crucial role in cccDNA synthesis in de novo HBV infection. In addition, our findings reported herein also provide important clues for further investigation of viral and cellular factors in cccDNA biosynthesis and regulation.
We demonstrated in this study that cccDNA is formed from incoming virion DNA in HepG2-NTCP cells at as early as 24 h post infection and establishes the pool size of approximately 3 copies of cccDNA per infected cell within a few days of infection. The kinetics of cccDNA accumulation as well as two lines of independent evidence obtained from HBV-ΔHBc infection of HepG2-NTCP cells and viral DNA polymerase inhibitor treatment of wild-type HBV-infected cells strongly support the notion that intracellular amplification does not play a significant role in the establishment of cccDNA pool in the HBV-infected hepatoma cells. This observation is consistent with the findings from HBV infection of primary human hepatocytes and HepaRG cells [38, 39], but distinct from DHBV infection of primary duck hepatocytes where significant intracellular cccDNA amplification occurs in a manner regulated by the level of its large envelope protein expression [19, 20]. However, intracellular amplification of cccDNA has been observed in HepG2-derived cell lines supporting constitutive or inducible HBV replication [40–42]. Developing therapeutics against chronic HBV infection requires better understanding the contribution of intracellular cccDNA amplification in the maintenance of persistent infection, and further investigation of the activity and regulation of this pathway in HBV-infected hepatocytes in vivo is thus warranted.
An elegant study by Chisari and colleagues demonstrated that although DHBV deficient for capsid protein expression (DHBVΔcp) infected primary duck hepatocytes and produced similar amounts of cccDNA from the incoming virions as did wild-type DHBV, the cccDNA in DHBVΔcp-infected hepatocytes was significantly less efficiently transcribed into viral RNAs, suggesting an important role of capsid protein in DHBV cccDNA transcription [43]. However, our results showed that HBV-ΔHBc infected HepG2-NTCP cells and expressed viral genes at a similar efficiency as wild-type HBV did. The results therefore suggest that the synthesis of HBV capsid protein may not significantly modify HBV cccDNA transcription activity. Interestingly, DHBV capsid protein is structurally distinct from the capsid proteins of mammalian hepadnaviruses [44] and may have HBx-like function in regulation of DHBV cccDNA transcription. However, the possibility that HBV capsid proteins from in-coming virions interact with cccDNA and promote its transcriptional activity cannot be completely ruled out. It has been shown that HBV capsid protein is a structural component of viral cccDNA minichromosome and its binding reduces the nucleosomal spacing of the minichromosome [45]. In addition, it has also been suggested that capsid protein promotes an epigenetic permissive state of HBV cccDNA by binding on CpG islands of cccDNA [46]. Of note, some non-nucleoside analogue compounds targeting capsid protein can dysregulate functional HBV capsid assembly [47–52]. Those capsid assembly effectors may alter the amounts and/or structure of cccDNA-bound capsid protein and consequentially interfere with cccDNA metabolism and function [53]. Intriguingly, interferon-stimulated gene (ISG) APOBEC3A seems to have a role in the destruction of cccDNA by direct interaction with HBV core protein [54]. Further investigation on the differential roles of capsid proteins in regulation of cccDNA function should shed light on this aspect of hepadnaviral pathobiology.
We showed herein that, similarly to DHBV and WHV, completion of plus strand DNA synthesis during de novo HBV infection of HepG2-NTCP cells is not sensitive to viral DNA polymerase inhibitors, suggesting the reaction is most likely catalyzed by a host DNA polymerase. In support of this hypothesis, by following the fate of viral DNA sequence during conversion of rcDNA into cccDNA, Seeger and colleagues demonstrated that independent of a viral enzymatic activity, a cellular DNA polymerase may fill in the 3’ end of both DNA strands [55]. The observed slight reduction of cccDNA amounts in ADV or ETV treated cells in this study could indicate either a minor contribution of viral DNA polymerase to cccDNA formation or an off-target inhibition of cellular functions required for cccDNA formation. Interestingly, studies of HBV cccDNA biosynthesis via intracellular amplification pathway in HepG2-derived stable cell lines, such as HepAD38 or HepDES19 cells, suggested that deproteinization and uncoating of progeny rcDNA require the completion of plus strand DNA synthesis, which requires viral DNA polymerase activity [41, 42]. Hence, different from rcDNA in virion particles with various length of incompletely synthesized plus-stranded DNA, the precursor rcDNA for cccDNA synthesis from the intracellular amplification pathway may have a very short gap in plus strand DNA and thus distinct DNA repair enzymes may be recruited to convert the rcDNA to cccDNA. Moreover, it had been shown that a small fraction of cccDNA can be formed from dslDNA via NHEJ DNA repair pathway. The cellular DNA polymerases required for cccDNA synthesis through intracellular amplification pathway and from dslDNA remain to be determined in future studies.
While our results demonstrated that POLK plays a critical role in cccDNA formation during de novo HBV infection of cultured HepG2-NTCP, HepaRG and PTHs, we also showed that POLL and POLH play a role in cccDNA formation, albeit at a lesser extent. It is currently not clear whether each of the cellular DNA polymerases plays a redundant or distinct role in de novo cccDNA synthesis. As mentioned above, plus-strand DNA in rcDNA in virions has gaps of heterogeneous lengths. It is possible that depending on the length of gaps, distinct DNA repair complexes containing different repairing DNA polymerases are recruited to fill the gaps with different length. If this is the case, the three DNA polymerases may play non-redundant roles and be involved in conversion of distinct rcDNA precursors into cccDNA. Alternatively, each of the three cellular DNA polymerases may participate in a different DNA repair complex to fill the plus strand gaps, irrespective of their length, but in a different efficiency. These two possibilities will need to be further investigated.
POLK plays a functional role in nucleotide excision repair (NER) pathway by filling the gap produced upon excision of damaged nucleotides [56, 57]. The activity of POLK is partially dependent on the growth state of the cells, and reaches maximum activity under conditions of low deoxynucleotide concentration such as in non-dividing cells [56]. A previous study showed treatment of HBV-transfected HepG2 cells with aphidicolin arrested cells in the G1 phase could result in enhancement of cccDNA synthesis [58]. Consistent with this observation, we found that the efficiency of HBV infection closely correlates with the number of G0/G1 phase cells in HepG2-NTCP cultures. Therefore, HBV cccDNA formation may preferentially occur at G0/G1 phase of cell cycle, supporting the notion that HBV infects non-dividing cells, so that cccDNA is formed and stably exists in quiescent hepatocytes [59]. It is thus conceivable that cell cycle-dependent factor(s) or protein post translational modification affecting the physiologic state of hepatocytes may regulate the formation of HBV cccDNA.
Moreover, because cellular DNA polymerases must work in concert with other DNA repair proteins to restore the structure of damaged DNA, other DNA repair proteins in NER pathway may also play a role in HBV cccDNA formation. For example, it is possible that HBV hijacks cellular endonuclease (e.g. XPG) or exonuclease (e.g. Exo1) to cleave the capped RNA primer to leave a free 5’ end of plus strand DNA of rcDNA, and followed by POLK or other cellular DNA polymerase, such as POLL and POLH, to fill the gap using minus strand DNA as a template. Additionally, POLK has been shown to work together with POLD to fill in single stranded DNA gaps [56, 60] and XRCC1-Lig3 is required for ligation of NER-induced breaks in quiescent cells [61]; hence it will be interesting to test whether those host enzymes are involved in cccDNA formation.
In conclusion, taking advantage of highly efficient genetic manipulation of HepG2-NTCP HBV infection system, and in combination with studies using recombinant HBV virus and chemical inhibitors, we rigorously demonstrated that cellular DNA polymerase κ substantially contributes to HBV cccDNA formation in HepG2-NTCP cells. Our findings shed new light on the molecular mechanism of cccDNA formation and may facilitate the development of novel therapeutics to cure chronic hepatitis B.
Human embryonic kidney cell 293T and human hepatoblastoma cell HepG2 were obtained from American Type Culture Collection (ATCC). Human hepatocellular carcinoma cell (Huh-7) was obtained from the Cell Bank of Type Culture Collection, Chinese Academy of Sciences. All the cell lines were maintained in Dulbecco’s Modification of Eagle’s Medium (DMEM; Life technologies) supplemented with 10% fetal bovine serum, 100 U/ml penicillin, and 100 μg/ml streptomycin at 37°C in a 5% CO2 incubator unless otherwise indicated. All experiments with HepG2 cells were carried out with cells grown on collagen-coated plates. HepaRG cells were obtained from Biopredic International (Rennes, France) and cultured according to the product manual. Differentiated HepaRG cells were obtained following a two-step procedure as previously described [38]. Primary Tupaia hepatocytes (PTHs) were obtained and cultured following a method as described previously [11].
Antibodies for human POLK (A-9) and C9 tag (1D4) were obtained from Santa Cruz Biotechnology. Antibody for human POLL (EPR7519(2)) was purchased from Abcam. Antibody against GAPDH, HRP-conjugated anti-mouse IgG and HRP-conjugated anti-rabbit IgG were purchased from Sigma. Alexa Fluor 488- and 546-conjugated anti-mouse IgG antibodies were purchased from Life Technologies. Other mouse antibodies to detect the Core protein (1C10) and surface antigen (17B9) of HBV, HDV delta antigen (4G5) were described previously [11, 62]. N-terminally myristoylated peptide of the N-terminal 47 amino acid residues of pre-S1 domain of HBV strain S472 (Accession number: EU554535.1), myr-47, was synthesized by SunLight peptides Inc.. Adefovir (ADV), entecavir (ETV), aphidicolin (APH) and other general chemicals were purchased Sigma unless otherwise stated.
To construct human POLK expression plasmids, a full-length human POLK cDNA was cloned from HepG2 mRNA into a modified pLKO.1-puro lentiviral vector under the control of a CMV promoter for stable expression in POLK knockout HepG2-NTCP cells. An N-terminal GFP-tagged POLK expression plasmid (GFP-POLK-wt) was constructed for investigating POLK subcellular localization and RNAi rescue experiment. For the single siRNA-mediated gene knockdown experiments, siRNA transfection was performed using Lipofectamine RNAiMAX (Life technologies) according to the manufacturer’s instruction. HepG2-NTCP cells in 48-well plates were transfected with 5 pmol siRNA and 0.5 μl Lipofectamine RNAiMAX in 25 μl Opti-MEM and were challenged with the HBV (or HDV, VSV as indicated) at 72 h post transfection. For combined dual RNAi experiments, HepG2-NTCP cells in 48-well plates were transfected with individual siRNAs for indicated two target genes (each 2.5 pmol, total 5pmol) and 0.5 μl Lipofectamine RNAiMAX in 25 μl Opti-MEM and were challenged with the HBV (or HDV, VSV as indicated) at 72 h post transfection. The siRNA target sequences used in this study are shown in S1 Table, which have been confirmed to be functional in other previous studies [36]. siRNAs targeting Tupaia polk gene were designed by RNAi designer (https://rnaidesigner.thermofisher.com/rnaiexpress/). The specificities of siRNAs were confirmed using a BLAST search. The efficiency of gene knockdown was determined by qPCR or Western blot assays. The qPCR primers used for quantification of human POLK, POLH and GAPDH mRNA expression are listed below. Human POLK: qPOLK-F: 5'-CCAGACATCACAACCATTCC and qPOLK-R: 5'-TCAAGGCTTCCAGACTGATG; human POLH: qPOLH-F: 5'-GTGCCAGTTACCAGCTCAGA and qPOLH-R: 5'-AGGTAATGAGGGCTTGGATG; GAPDH: 5’-GAAGGTGAAGGTCGGAGTCA (forward) and 5’-TGGAATCATATTGGAACATGT (reverse). Both human POLK and POLH mRNA levels are normalized to the expression level of GAPDH, respectively. For Tupaia POLK mRNA measurements, qtsPOLK-F: 5'-TCACTAGCCAGCAGCTAAGGAAAGC and qtsPOLK-R: 5'-CATGCTCATTGATCCTACAGCAATG were used as qPCR primers. 5’-GTGAAGGTCGGAGTAAACG (forward) and 5’-CCATGGGTGGAGTCATACT (reverse) were used for Tupaia GAPDH, the mRNA expression level of tsPOLK is normalized by tsGAPDH. All siRNA oligos were synthesized by NIBS biological resources facility. All transfections were conducted in duplicates. Cytotoxic effects of siRNA were examined using alamarBlue reagents (Life technologies).
HBV genotype D virus and HDV genotype I virus were produced by transient transfection of Huh-7 cells with the corresponding plasmids as described previously [11]. HBc protein deficient virus (HBV-ΔHBc) was generated by co-transfection of Huh-7 cells with a plasmid harboring 1.05 copies of HBV genome with a stop codon at the 38th codon (Y) of core gene open reading frame, and an intact HBc protein expression vector. Virus stocks were aliquoted and stored at -80°C. HBV and HDV infection assays have been described previously [11]. Briefly, HepG2-NTCP cells were firstly cultured in 48-well plates with DMEM complete medium for 3–4 h, then the cells were cultured with PMM medium for another 20 hrs. The cells were then infected with a multiplicity of 100 genome equivalents (mge) of HBV or 500 mge of HDV in the presence of 4% PEG8000 for 24 hrs at 37°C. Cells were maintained in PMM with regular medium changing every other day. For viral inhibition assay, 100 nM myr-47 lipopeptide was pre-incubated with HBV viruses before adding to HepG2-NTCP cells; chemicals were pre-incubated with HepG2-NTCP cells at 37°C for 12 h before virus inoculation.
To generate VSV-G protein pseudotyped lentiviral particles expressing EGFP, PLKO.3G plasmid together with pSPAX2 and pMD2G were co-transfected into 293T cells at the ratio of 4:3:1 (PLKO.3G:pSPAX2:pMD2G) using Lipofectamine 2000 (Life technologies). Forty-eight hours after transfection, supernatants were harvested, cleared by centrifugation and stored at -80°C. Human POLK or gRNA expressing lentivirus was generated using a similar protocol as that described above but replacing the pLKO.3G plasmid with the pLKO.1-CMV-POLK-puro or pLKO.1-gRNA-EGFP plasmid, respectively. Infection of cells with lentiviral pseudovirus was performed as described previously [63]. Virus-related experiments were conducted in a BSL-2 facility at the National Institute of Biological Sciences, Beijing.
HBV viral particles were purified using Nycodenz gradient ultracentrifugation of the culture supernatants of Huh-7 cells transfected with plasmid for HBV production. The HBV DNA levels of the virus fractions were quantified using specific primers: 5’-GAGTGTGGATTCGCACTCC (forward) and 5’-GAGGCGAGGGAGTTCTTCT (reverse) by real-time PCR using SYBR Premix Ex Taq kit (TaKaRa) on an ABI 7500 Fast Real-Time system instrument (Applied Biosystems, United States). The viral genome equivalent copies were calculated based on a standard curve generated with known copy numbers.
For analysis of HBV 3.5kb viral RNA, total RNA from infected cells was extracted by TRIzol reagent (Life technologies), 0.5 μg of total RNA was digested with DNase I (Life technologies) and reverse-transcribed into cDNA using PrimerScript RT Reagent Kit (TaKaRa) in a 20 μl reaction. Real time-PCR analysis was performed using the SYBR Premix Ex Taq on ABI 7500 Fast Real-Time PCR System. HBV 3.5kb viral RNA copy numbers were deduced from a standard curve generated from known nucleic acid quantities. The mRNA level of HBV 3.5kb viral RNA was normalized to that of GAPDH mRNA. The primers used for each gene examined are listed below. HBV 3.5kb viral RNA: 5'-GAGTGTGGATTCGCACTCC (forward) and 5'-GAGGCGAGGGAGTTCTTCT (reverse); GAPDH mRNA: 5’-GAAGGTGAAGGTCGGAGTCA (forward) and 5’-TGGAATCATATTGGAACATGT (reverse).
For quantification of HBV cccDNA, infected cells were lysed in a lysis buffer (20 mM Tris, 0.4 M NaCl, 5 mM EDTA, 1% SDS, pH = 8.0) in the presence of proteinase K (QIAGEN), total DNA was extracted according to a standard phenol-chloroform extraction protocol. 500 ng of total DNA was digested with 0.5 μl plasmid-safe adenosine triphosphate (ATP)-dependent deoxyribonuclease DNase (PSAD) (Epicentre Technologies) in 25 μl reaction for 8 h at 37°C to allow removal of linear genomic DNA and HBV replication intermediates (rcDNAs, single-stand DNAs, linear double-strand DNAs). DNase was inactivated by incubating the reactions for 30 min at 70°C. 20 ng of digested DNA was used for quantification of HBV cccDNA, 5'-TGCACTTCGCTTCACCT (forward) and 5'-AGGGGCATTTGGTGGTC (reverse) were used as HBV cccDNA specific primers, the real-time PCR was performed using the SYBR Premix Ex Taq on ABI 7500 Fast Real-Time PCR System as the following reaction procedure: 95°C for 5 min then 45 cycles of 95°C for 30 s, 62°C for 25 s, and 72°C for 30 s. The amount of HBV cccDNA in a DNA preparation was determined by real-time PCR using a plasmid containing HBV-D genome as the standard. The pool size of HBV cccDNA per infected cell was calculated by quantification of cccDNA copies using digital PCR in the whole cell population and estimation of the number of infected cells by immunostaining of intracellular HBcAg.
Selective extraction of HBV cccDNA from HBV infected cells was achieved by a modified Hirt method as previously described [64, 65]. Briefly, infected cells from one well of 6-well plates were lysed in Hirt lysis buffer (10 mM Tris-HCl, 10 mM EDTA, 0.6% SDS, pH = 7.4) for 30 min at room temperature. After adding 5 M NaCl, the cell lysate was vigorously mixed and incubated at 4°C overnight. After centrifugation at 10,000 rpm for 30 min at 4°C, the supernatant was extracted twice with saturated Tris-phenol (pH = 8.0) and once with phenol:chloroform. The extracted DNA was precipitated with equal volumes of isopropanol at -20°C overnight. The DNA pellet was washed with 70% ethanol and dissolved in TE buffer (10 mM Tris-HCl, 1mM EDTA, pH = 8.0), and digested with HindIII or EcoRI restriction enzyme (NEB) before being analyzed.
For detection of cccDNA by Southern blot, the extracted HBV cccDNA sample was subjected to 1.2% agarose gel electrophoresis and transferred onto Amersham Hybond-N+ membrane (GE Healthcare). The Hybond-N+ membrane was crosslinked in a UV crosslinker chamber with UV energy dosage at 1200 mJ and followed by being probed with [α-32P]dCTP (250 μCi, Perkin Elmer)-labeled HBV genotype D (Accession number: U95551.1) linear full-length genomic DNA. Hybridization was performed in Perfecthyb plus hybridization buffer (Sigma) with prehybridization for 1 h at 65°C and overnight hybridization at 65°C, followed by two washes in wash buffer (0.1×SSC, 0.1% SDS) at 65°C. The membrane was exposed to Carestream X-OMAT BT Film (XBT-1, Carestream). 100 pg each of 3.2kb, 2.1kb and 1.7kb HBV DNA fragments prepared by PCR amplification of a plasmid containing 1.0 copies linear HBV genotype D genome was used as DNA marker.
HBeAg and HBsAg from supernatants of HBV infected cells were measured using ELISA kits (Wantai Pharm Inc. Beijing, China) by following the manufacturer’s instructions. Supernatants from HBV infected cells were harvested at each time point examined in the various assays and were diluted 2-fold with PMM before ELISA. All experiments were performed in duplicates and repeated at least two times independently.
Virus infected cells in 48-well plates were washed three times with pre-cooled PBS and fixed by 4% paraformaldehyde for 10 min, followed by permeablization for 10 min at room temperature with 0.5% Triton X-100. After incubation for 1 h with 3% BSA for blockade of nonspecific binding, primary antibodies were added for incubation for 1 h at 37°C. The bound antibodies were visualized by incubation with secondary antibodies (Alexa Fluor 488 donkey anti-mouse IgG or Alexa Fluor 546 anti-mouse IgG). Images were acquired using a Nikon A1-R confocal microscope or a Nikon Eclipse Ti Fluorescence Microscopy.
The POLK rescue cell lines were generated by infection of POLK stable knockout cells with recombinant lentivirus expressing POLK, and cells were selected with puromycin (Sigma). Expression of POLK in the established stable cell lines were verified by Western blot assay.
To generate POLK deficient HepG2-NTCP cells, genomic engineering of polk gene was achieved with the CRISPR/Cas9 system as described with the following single-guide (sg) RNA target sequences: CTTCTCCTTTGTGCTATCCA (sgPOLK-1), GATGATCTTCTGCTTAGGAT (sgPOLK-2). Firstly, stably expressing Cas9 cell line (HepG2-NTCP/Cas9) was generated by transfection of human codon-optimized Cas9 (hCas9) expressing vector (Addgene) using Lipofectamine 2000 and selection with Blasticidin (Calbiochem). Then the cells were infected with gRNA expressing lentivirus. After culture for 3 d in vitro, EGFP positive cells were sorted by flow cytometry FACSAria II, and further analyzed by T7 endonuclease I (NEB) assay. Genomic DNA sequence of POLK around the gRNA targeting site was amplified using the following primers: cas9-POLK-F: 5'-GTGTCGAACCCCTGAGCTCAGTCAATCT and cas9-POLK-R: 5'-AGGTGAACAGGAACATATACATTATTT. Single clones of sorted cells were obtained by serial dilutions and amplified, verified by sequencing of PCR fragments, and confirmed by Western blot using anti-POLK antibody. POLL deficient HepG2-NTCP cell lines were constructed by following the above mentioned procedure, a 20-bp single-guide sequence (sgPOLL: CGGGCCCATGTTGTGCGCAC) targeting DNA within the third exon of poll gene was selected. cas9-POLL-F: 5’-GCTATATGTAGAAGGAAAGCTGTC and cas9-POLL-R: 5’-ACTGGGATCAGCCCACCTACTGG were used as primers for amplification of a region around the gRNA targeting site and the PCR products were further analyzed by T7E1 assay. Individual clones were validated by sequencing of PCR fragments, and confirmed by Western blot using anti-POLL antibody.
|
10.1371/journal.pcbi.1004341 | Conformational Dynamics and Binding Free Energies of Inhibitors of BACE-1: From the Perspective of Protonation Equilibria | BACE-1 is the β-secretase responsible for the initial amyloidogenesis in Alzheimer’s disease, catalyzing hydrolytic cleavage of substrate in a pH-sensitive manner. The catalytic mechanism of BACE-1 requires water-mediated proton transfer from aspartyl dyad to the substrate, as well as structural flexibility in the flap region. Thus, the coupling of protonation and conformational equilibria is essential to a full in silico characterization of BACE-1. In this work, we perform constant pH replica exchange molecular dynamics simulations on both apo BACE-1 and five BACE-1-inhibitor complexes to examine the effect of pH on dynamics and inhibitor binding properties of BACE-1. In our simulations, we find that solution pH controls the conformational flexibility of apo BACE-1, whereas bound inhibitors largely limit the motions of the holo enzyme at all levels of pH. The microscopic pKa values of titratable residues in BACE-1 including its aspartyl dyad are computed and compared between apo and inhibitor-bound states. Changes in protonation between the apo and holo forms suggest a thermodynamic linkage between binding of inhibitors and protons localized at the dyad. Utilizing our recently developed computational protocol applying the binding polynomial formalism to the constant pH molecular dynamics (CpHMD) framework, we are able to obtain the pH-dependent binding free energy profiles for various BACE-1-inhibitor complexes. Our results highlight the importance of correctly addressing the binding-induced protonation changes in protein-ligand systems where binding accompanies a net proton transfer. This work comprises the first application of our CpHMD-based free energy computational method to protein-ligand complexes and illustrates the value of CpHMD as an all-purpose tool for obtaining pH-dependent dynamics and binding free energies of biological systems.
| Formation of insoluble amyloid plaques in the vascular and hippocampal areas of the brain characterizes Alzheimer’s disease, a devastating neurodegenerative disorder causing dementia. Site-specific hydrolytic catalysis of β-secretase, or BACE-1, is responsible for production of oligomerative amyloid β-peptide. As the catalytic activity of BACE-1 is pH-dependent and its structural dynamics are intrinsic to the catalysis, we examine the dependence of dynamics of BACE-1 on solution pH and its implications on the catalytic mechanism of BACE-1. Also, we highlight the importance of accurate description of protonation states of the titratable groups in computer-aided drug discovery targeting BACE-1. We hope the understanding of pH dependence of the dynamics and inhibitor binding properties of BACE-1 will aid the structure-based inhibitor design efforts against Alzheimer’s disease.
| Alzheimer’s disease is a neurodegenerative disorder characterized by loss of memory and failure in cognitive abilities, resulting from synaptic dysfunction and neuronal death in the brain [1–5]. Major damages found in the brains of Alzheimer’s patients include cerebral and vascular deposits of insoluble amyloid plaques, consisting of aggregates of amyloid β-peptide (Aβ) [6–8]. Aβ occurs in two different forms, Aβ40 and Aβ42, and the overproduction and oligomerization of Aβ42 is associated with the early onset of Alzheimer’s disease [9–12]. Aβ is produced by sequential proteolytic cleavage of the type 1 transmembrane protein amyloid precursor protein (APP) by β- and γ-secretases [13,14]. While γ-secretase generates several Aβ peptides varying in the length of C-termini, β-secretase, or β-site APP cleaving enzyme 1 (BACE-1), cleavage precisely gives the fibrillogenic Aβ42 [13–15]. Therefore, as it catalyzes the initial site-specific hydrolysis step of Aβ production, BACE-1 is an attractive therapeutic target for the treatment of Alzheimer’s disease [1–3,16,17].
As an aspartyl protease, the catalytic mechanism of BACE-1 involves two highly conserved aspartyl residues, Asp32 and Asp228, which form a symmetric dyad at the base of the catalytic cleft of the enzyme (Fig 1) [16]. Analogous aspartyl dyads are found in the aspartyl protease family including pepsin, cathepsin D, renin, and HIV-1 protease [18–21]. The dyad is central to the hydrolytic cleavage of the substrate through a nucleophilic attack of water bound to the dyad [19–23]. Due to the general acid-base catalytic nature of the mechanism, the enzymatic activity of BACE-1 is maximal at pH 4.5 and strongly depends on solution pH [24,25].
The active site of BACE-1 is covered by an antiparallel β hairpin (henceforth referred to as the flap region; residues 67 to 77 shown in green in Fig 1) that is characteristic of aspartyl proteases [16,26–29]. The X-ray crystal structures of other aspartyl proteases indicate that the flap is inherently flexible [26–29]. The flexibility of the flap region is likely utilized in catalysis, with transitions between open and closed conformations facilitating the entrance of substrates into the active site and release of hydrolytic products [21,29–31]. The conserved Tyr71 [20] located at the tip of the flap region is particularly essential for the conformational transitions of the flap. Observations from X-ray crystallographic structures and molecular dynamics (MD) simulations suggest that variation in hydrogen bond patterns between Tyr71 and surrounding residues such as Lys107, Lys75, Gly74, Glu77, and Trp76 enables the flexible motions of the flap [21,29,31–33]. In the presence of inhibitors, Tyr71 can directly interact with bound inhibitors and lock the flap in the closed state [31,33,34].
Given that the enzymatic activity of BACE-1 depends on solution pH and that the structural flexibility is intrinsic to catalysis, a comprehensive understanding of the pH dependence of BACE-1 dynamics would greatly benefit drug design efforts. A detailed description of the protonation state of the aspartyl dyad is also important as all known bound inhibitors directly contact the dyad. Several computational efforts have attempted to determine the protonation state of the dyad, employing methods such as molecular mechanics (MM) [35], quantum mechanics (QM) [36], QM/MM [37], molecular docking [38,39], and continuum electrostatics calculations [40]. However, the conformational flexibility of BACE-1 was not rigorously addressed in these computations.
The importance of accounting for conformational flexibility in pKa computations has been well established [41–46]. The instantaneous pKa of a titratable group is determined by its electrostatic environment, which is affected by the given conformation of protein and protonation states of other titratable residues [47]. Changes to the conformation of the protein can alter the electrostatics, which may, in turn, induce a shift in the pKa of titratable groups. The prevalence of such coupling of protonation and conformational equilibria has been observed in various systems both computationally and experimentally [44,48–59]. Furthermore, complex formation between protein and small molecules or other proteins can also induce changes in the pKa values of titratable groups on either binding partner [51,60–70].
Consequently, several computational methods have been developed to explicitly account for conformational changes in pKa computations [49,71–76]. Among these, various flavors of constant pH molecular dynamics (CpHMD) methodologies have emerged to incorporate pH as an additional external thermodynamic variable to the conventional MD framework [48,77–81]. CpHMD simulations have been successfully applied to predict pKa values of titratable groups in proteins [48,77–83] and nucleic acids [84–86], as well as to explain the acid-base catalysis by RNase A [87] and to understand the mechanisms behind the pH-dependent conformational changes [59,88].
Conventional molecular simulations or free energy computations typically employ fixed protonation states that are identical for both free and bound states, set prior to the computations. This assumption ignores the possibility of protonation states changing upon binding and can lead to significant errors when protein-ligand binding is a pH-dependent process [60]. Furthermore, the pH-dependent conformational dynamics cannot be appropriately addressed if the protonation states are fixed while conformational fluctuations propagate.
Recognizing the lack of a standard protocol to rigorously account for proton-linked ligand binding to protein, we recently developed a protocol utilizing CpHMD to compute pH-dependent binding free energies [89]. In our computational method, the binding polynomial formalism devised by Wyman [90] is applied with the CpHMD framework to obtain a pH-dependent correction to a reference free energy of binding obtained at a given level of pH (ΔG°ref,pH). The proton-linked binding free energy then can be expressed according to Eq 8 provided in Methods, using the notation used by Tanford [91], where the total charges of the protein-ligand complex, protein, and ligand in Eq 8 are obtained from the CpHMD simulations. The integral in the second term in Eq 8 provides a thermodynamic relation that holds for pH-dependent ligand binding in cases where proton binding to different titratable sites may be cooperative (i.e., no assumptions are made about sites titrating independently). When applied to binding of small molecules to the cucurbit [7]uril host, this CpHMD-based free energy method accurately obtained the pH-dependent binding free energy profiles. Also, the method demonstrated that the traditional use of fixed protonation states for both free and bound states predicted based on pH 7 in free energy computations could give errors larger than 2 kcal/mol in the host-guest systems with a single titratable site [89]. Given the complexity of protein environment where multiple titratable groups exist, the corresponding error in free energy may be even larger in protein-ligand binding, highlighting the significance of accurate description of the binding-induced pKa shifts in free energy computations.
In this work, we have performed constant pH replica exchange molecular dynamics (pH-REMD) simulations to study the proton-linked conformational dynamics and binding free energies of inhibitors to BACE-1. The conformational changes of the flap region of BACE-1 in the absence and presence of inhibitors shown in Fig 2 are analyzed with respect to solution pH, which is found to act as a conformational switch. The microscopic pKa values of ten titratable residues including the aspartyl dyad in BACE-1-inhibitor complexes are obtained from the pH-REMD simulations and compared with those computed for free enzyme. The results show significant binding-induced shifts in the pKa values. We further apply our CpHMD-based computational protocol to these results, computing the pH-dependent binding free energy profiles of various inhibitors to BACE-1. The results demonstrate that incorrect assignment of protonation state to the titratable groups can result in errors of over 8 kcal/mol in free energy computations for the systems considered here, highlighting the significance of correctly addressing the binding-induced protonation changes. To the best of our knowledge, this work presents the first application of CpHMD simulations to quantify binding in protein-ligand systems and shows high utility for addressing pH effects in computer-aided drug discovery workflows.
As conformational transitions of BACE-1 are suggested to play a role in catalysis [29–31], we first examine the dynamics intrinsic to apo BACE-1 before exploring the effect of inhibitor binding in the following section. Conventional molecular dynamics (cMD) simulations of duration 100 ns are carried out prior to constant pH molecular dynamics (pH-REMD) simulations in order to provide equilibration phase for apo and three inhibitor-bound systems prepared by homology modeling (see Methods). Using the protonation states assigned using the PROPKA program [75,92–94], the cMD simulations provide a benchmark for comparing pH-REMD simulations.
In order to quantify the extent to which each residue fluctuates, we compute the root-mean-square fluctuation (RMSF) of each residue in apo BACE-1 from the cMD trajectory. As shown in Fig 3A, higher RMSF values are noted for the flap region (residues 67 to 77), consistent with the suggestion by others [29,31–34]. Taking a closer look at the flap region, we measure the distance between the center of mass of the aspartyl dyad and Tyr71, which is located at the tip of the flap region. From the change of the dyad-flap distance plotted in Fig 3B, it is evident that the flap region undergoes transitions between open and closed conformations, within the distance range of 10 Å (closed) to 30 Å (open). In the closed conformation of the flap, we observe a water-mediated hydrogen bond network that includes the dyad, Ser35, Tyr71, Arg128, Thr231, and Thr329 (Fig 3C), agreeing with the findings from previous studies [23,31,34]. On the other hand, when the flap is open, the dyad forms contacts mediated by water with Ser35, Thr231, and Arg235, while Tyr71 is entirely exposed to solvent (Fig 3C).
While the cMD simulation is performed with fixed protonation states, we choose to examine the effect of protonation equilibria on the conformational flexibility of apo BACE-1 as the enzymatic activity of BACE-1 is shown to be pH-dependent [24,25]. We focus this investigation on a comparison of the dynamics of the flap region at acidic (pH 1 to 3) and basic (pH 9 to 11) pH levels. The conformational space of the flap region sampled at these differing levels of pH is quantified by measuring the distances between the center of mass of the dyad and Tyr71 (Fig 4A). Since the pH-based trajectories reconstructed from the pH-REMD simulations are not time-dependent, distributions of the measured distances are presented here. The distance between the dyad and Tyr71 exhibits a bimodal distribution at acidic pH, with the flap sampling both open and closed states. When the flap is closed, an average distance between the dyad and flap is about 8 Å, while open conformations are also populated, having an average dyad-flap distance of ~ 17 Å. In order to visualize structural characteristics typical at varying solvent environments, we carry out a clustering analysis based on pairwise RMSDs of Cα atoms of the conformations sampled at acidic and basic pH. We find three dominant conformers that encompass 86% of total conformations sampled at acidic pH. The flap regions from representative structures of these three clusters are shown in red in Fig 4B, further illustrating the flap region sampling both open and closed conformations at acidic pH.
At basic pH, the flap exhibits noticeably different dynamics compared to acidic pH. The flap region remains over 10 Å away from the dyad and most frequently found in the open conformation with a distance of about 17 Å from the dyad (Fig 4A). From the clustering analysis, a single conformer is found to represent 82% of total conformations sampled at basic pH. In this typical conformer, the flap is in widely open state with a distance of 17.6 Å from the dyad and completely exposed to the solvent area, as shown in blue in Fig 4B.
We continue to probe the changes in dynamics of BACE-1 that accompany inhibitor binding. While the shifts between open and closed conformations of the flap region are observed in the cMD simulations of apo BACE-1, the flap remains in a closed state in the inhibitor-bound cMD simulations. The distances between the center of mass of the dyad and Tyr71 observed in the cMD simulations of BACE-1 in complex with the inhibitors 2B8L and 2FDP, respectively, are shown in Fig 5A. In comparison to the distribution of distances observed in the cMD simulation of apo BACE-1 (Fig 3B), we observe significantly less flexibility in the flap region in the 2B8L and 2FDP systems. Fluctuation of the dyad-flap distance in the 2B8L system during the early stage of the simulation is likely due to structural instability arising from the homology modeling. However, after 40 ns, the flap region in the 2B8L complex achieves a stable state and remains closed at a distance of about 15 Å from the dyad. It is worth noting that the measured dyad-flap distances in the inhibitor-bound systems are inherently longer than that in apo enzyme; for instance, the open conformation of the flap in apo BACE-1 has the dyad-flap distance of ~ 15 Å whereas the closed conformation of the flap in the 2B8L system has the distance of 15 Å as well. This is because the flap is unable to penetrate into the active site as deep as in apo BACE-1, due to the presence of the bound inhibitors in the binding site. Similarly, the flap region in the 2FDP system maintains an average dyad-flap distance of ~ 11 Å. The bound inhibitors have hydrophobic interactions with Tyr71 and Phe108 while forming hydrogen bonds with the polar residues including the dyad, Tyr71, Thr231, and Ser325. These hydrogen bond networks effectively lock the flap in the closed state, as shown in Fig 5B for the 2B8L system. Similar trends in the dyad-flap distance are observed in other inhibitor-bound systems (S1 Fig).
As we observe the pH-controlled dynamics of the flap region in apo BACE-1, we also carry out similar analyses on the holo systems to examine the effect of pH on conformations of the flap in the presence of inhibitors. Unlike the apo system, a clear distinction in the dynamics of BACE-1 contingent on pH levels is not found in the inhibitor-bound systems. The distances between the center of mass of the dyad and Tyr71 indicate that a closed state of the flap is dominant in the 2B8L system at both pH, with average distances of ~ 15 Å, as well as slightly open conformation of the flap with the dyad-flap distance of ~ 19 Å at basic pH (Fig 6A). From the clustering analysis, the flap is in closed state in the typical structure representing 67 to 77% of the total conformations sampled at both pH conditions, as shown in Fig 6C, forming similar contacts with the bound inhibitor and dyad as observed in the cMD simulations (Fig 5B). Other cluster representative structures also exhibit similarly closed conformations the dyad-flap distances varying in the range of 15 to 19 Å. In the 2FDP system, the flap region exhibits essentially identical dynamics; the flap is in closed state at acidic pH and has both slightly open and closed conformation at basic pH. On average, the flap is found primarily in a closed conformation where the dyad-flap distance is 15.2 Å (Fig 6B). Consistent with these findings, the cluster representative structures of the 2FDP system at both acidic and basic pH have the flap in closed state, similar to the 2B8L complex (Fig 6C). Similar trends in the distributions of the dyad-flap distance in varying pH conditions are observed in other inhibitor-bound systems (S2 Fig).
We further investigate the acidic properties of the aspartyl dyad and surrounding titratable sites in both apo and holo BACE-1. The pKa values of the different titratable residues considered are obtained by fitting the Hill equation (Eq 3; Methods) to titration data obtained from the pH-REMD simulations conducted at different levels of pH (Table 1). We first analyze the computed pKa values of the titratable groups in apo BACE-1 and compare the values obtained for inhibitor-bound systems in the following section.
The predicted pKa values of the aspartyl dyad, Asp32 and Asp228, are 5.0 ± 0.2 and 5.9 ± 0.5, respectively. Titration curves for the dyad are shown in Fig 7, plotted as fraction of deprotonated species of each residue as a function of pH. From the titration curves, both aspartates are completely protonated at acidic pH levels (pH < 3), while fully deprotonated at basic pH (pH > 8). Between pH 4 and 8, Asp32 and Asp228 exist in an ensemble of protonated and deprotonated forms. To illustrate this, Asp32 and Asp228 are approximately 20% and 40% deprotonated, respectively, at pH 4.5, the pH at which BACE-1 is most active. These observed shifts from the typical pKa of Asp residues (4.0) [95,96] may aid in the proton transfer step required in BACE-1 catalysis. The Hill coefficients of the dyad deviate from one, suggesting that titration of Asp32 and Asp228 is coupled. The coupled titration observed for these residues also contributes to greater noise in their respective titration curves, leading to larger errors in in their pKa (Table 1).
The pKa values of the remaining titratable residues are also reported in Table 1. All ionizable residues besides the dyad appear to titrate independently of each other, as suggested by their Hill coefficients that are approximately one. Furthermore, the statistical errors from fitting procedure to obtain the pKa’s of these residues are minimal. Similar to those of the dyad, the computed pKa values of Asp138, Asp223, Glu116, and Glu339 are shifted higher than the canonical pKa values for Asp (4.0) and Glu (4.4) [95,96], whereas Asp106 exhibits titration behavior in line with model Asp. As these residues are distant from the active site, these deviations in the computed pKa values arise mainly from the microenvironments surrounding them. For instance, Glu116 and Glu339 are buried in the protein interior (Fig 1), where the microscopic dielectric constants can be different from that of the bulk solvent [43,44,97,98]. In order to compensate for the desolvation energy, the neutral, protonated forms of the glutamates are favored in the protein interior at the solution pH where they would normally be charged if they were not buried. Titration curves for the remaining titratable residues are provided in S3 Fig.
Having observed the pKa values of various titratable groups in apo BACE-1, we shift our attention to the inhibitor-bound systems. First considering the dyad, the pKa values of Asp32 and Asp228 are shifted toward more basic values of pH when various inhibitors are present, with computed pKa values greater than 8.1. Representative titration curves for the dyad in 2B8L and 2FDP systems (purple curves) are compared with those in apo BACE-1 (green curves) in Fig 8. Similar curves corresponding to other inhibitor-bound systems are shown in S4–S6 Figs.
In the 2B8L system, Asp32 and Asp228 have pKa values of 9.7 ± 0.2 and 8.4 ± 0.1, respectively (shifted + 4.7 and + 2.5 pK units relative to their values in apo BACE-1). From the titration curves in Fig 8A and 8B, it is apparent that both Asp32 are Asp228 are significantly protonated under pH 8–9 in the 2B8L system. Above this pH, the dyad exists in an ensemble of different protonation states between pH 8 and 10. Similar binding-induced pKa shifts are observed for the dyad in 2FDP system. In this case, the computed pKa’s of Asp32 and Asp228 are 8.9 ± 0.7 and 9.7 ± 0.6, respectively. As shown in Fig 8C and 8D, in the 2FDP complex, both aspartates are predominately protonated at the pH levels below 7, and exist as an ensemble of different protonated forms at basic pH. In both cases, as in the apo enzyme, we observe the coupled titration behavior of the dyad and relatively large errors during the fitting of titration data to the Hill equation.
The remaining titratable groups examined, all of which are distant from the binding site, do not undergo pKa shifts larger than 1.3 pK units upon inhibitor binding (S4–S8 Figs). The narrow pKa shifts of these titratable groups upon complex formation suggest that binding of inhibitors to BACE-1 is thermodynamically linked to a proton transfer that is primarily localized at the catalytic dyad.
The large shifts in pKa’s of the aspartyl dyad upon binding of inhibitors to BACE-1 indicate that proton binding is linked to complex formation in BACE-1 systems. Utilizing our recently developed computational protocol [89], we present the pH-dependent changes in free energies of binding of inhibitors to BACE-1. Application of the binding polynomial formalism to the results obtained from the pH-REMD simulations provides pH-dependent corrections to the reference binding free energies obtained for a given pH, ΔG°ref,pH (Eq 8). The reference free energies of binding of inhibitors for BACE-1-inhibitor systems are obtained from experimental association constants measured at pH 4.5 [99–102].
Binding free energy profiles as functions of pH of the 2B8L and 2FDP complexes are shown in Fig 9. Significant changes in binding free energies from the reference free energies at pH 4.5 are observed as solution pH increases. Considering the 2FDP complex, we see that binding is most favorable at acidic pH, where the maximum affinity is -11.2 kcal/mol. As pH increases, binding becomes less favorable. This is most pronounced in the pH range of 5 to 10, where the aspartyl dyad that interacts directly with the bound inhibitor begins to populate deprotonated states, leading to an ensemble of protonated and deprotonated species of the dyad at the pH levels between 5 and 10 (Fig 8B). As the deprotonated forms of the dyad develop, hydrogen bonds made between the diprotonated dyad and bound inhibitors at low pH are lost, and consequently, the binding affinity becomes weaker in this pH range and is least favorable with an affinity of 2.2 kcal/mol at pH 12. Similar binding free energy profiles as functions of pH for other inhibitor-bound systems can be found in S9 Fig.
In Table 2, the binding free energies are compared in the pH ranges that are most relevant to biological conditions. The binding free energy for the 2FDP system changes by 5.1 kcal/mol between pH 4.5 and 7. Comparison to the binding free energy at pH 10 leads to more dramatic changes; for instance, the binding affinity differs by 10.8 kcal/mol between pH 4.5 and pH 10 for the 2FDP complex. As mentioned above, such large changes are due to the shifts in protonation state of the dyad at these pH levels from the diprotonated state observed at pH 4.5, and further highlight the significance of correctly accounting for the protonation states of the titratable groups which accompany a net proton transfer upon inhibitor binding.
The dependence of BACE-1 enzymatic activity on solution pH and the need for conformational change to accompany the catalysis are both well established [24,25,30,31,34]. As a promising therapeutic target for the treatment of Alzheimer’s disease, understanding the detailed mechanism underlying the pH dependence of BACE-1 dynamics and enzymatic activity is imperative for structure-based drug design. In this work, we have performed constant pH replica exchange molecular dynamics (pH-REMD) simulations to examine the proton-linked conformational dynamics and inhibitor binding properties of BACE-1.
Significant flexibility of the flap region (residues 67 to 77), resulting in transitions between open and closed conformations, is noted for apo BACE-1 during conventional molecular dynamics (cMD) simulations (Fig 3). As we further probe the effect of pH on conformational flexibility of apo BACE-1, distinctive conformations that are characteristic of different pH environments are captured from the pH-REMD simulations. At acidic pH, both open and closed conformations of apo BACE-1 are significantly populated, whereas a single conformer with the flap closed predominates in basic conditions (Fig 4).
While we observe the flexible nature of apo BACE-1, the presence of inhibitors at the active site of BACE-1 greatly reduces conformational mobility of the enzyme (Fig 5). The bound inhibitors form various hydrophobic and polar interactions with the surrounding residues, holding the flap in a tightly closed state. Similarly closed conformations of the flap are observed regardless of the varying pH conditions, indicating that the structural flexibility of BACE-1 is largely limited by the presence of bound inhibitors.
We determine the microscopic protonation states of the dyad residues and surrounding titratable residues to further probe the mechanism underlying the pH dependence of the catalytic activity of BACE-1. First, the pKa values of ten titratable residues in various BACE-1 systems are obtained from the pH-REMD simulations (Table 1). The computed pKa values of Asp32 and Asp228 in apo BACE-1 are 5.0 ± 0.2 and 5.9 ± 0.5, respectively, shifted from the typical pKa of Asp (Fig 7). At acidic pH, protonated states are dominant for both aspartates of the dyad. This allows the dyad to form hydrogen bonds with the flap residues, i.e., Tyr71 and Thr72 and allows for closed conformations to be sampled at low pH. On the other hand, open conformations of the flap are also populated at acidic pH; these open conformations likely aid in substrate binding and product release in the course of catalysis [29–31]. When we solvate the open and closed conformers from the pH-REMD simulations, which are performed in implicit solvent, similar water occupancies to those in the cMD simulations are observed (Fig 3C). When the flap is open, the active site becomes largely accessible to water, which is needed to act as nucleophile for the hydrolytic catalysis by BACE-1. Also, water molecules entering the active site help to compensate for the breaking of hydrogen bonds between the dyad and flap by forming alternative hydrogen bonds and mediating the hydrogen bond networks with surrounding charged residues such as Ser35 and Arg235. Therefore, conformational transitions between open and closed states of the flap at acidic pH allow for channeling of solvent, substrates, and hydrolytic products to and from the active site in catalysis. With the pKa values near the pH of optimal enzymatic activity, i.e., pH 4.5, the dyad is also able to easily gain and release proton(s) during the catalytic cycle. Hence the pKa values of the dyad of BACE-1 shifted from the typical value may be an evolutionary result to achieve the maximal activity at pH 4.5.
As pH increases, deprotonated species of both the dyad and Tyr71 start to emerge. Consequently, the hydrogen bond networks observed in acidic conditions no longer persist and are instead replaced by water molecules, leading to the primarily open conformation of the flap at basic pH. While this open state of the flap is stabilized energetically through water-dyad interaction, it is likely that the persistence of the open conformation at basic pH disables the enzyme’s ability to corral the substrate into the binding site for catalysis. Therefore, observation of the invariably open state of the flap at basic pH is consistent with the suggested role of flexibility of flap in the catalysis at acidic pH [29–31]. Our results thus indicate that the conformational dynamics intrinsic to the enzymatic catalysis of BACE-1 are modulated by solution pH, further suggesting the enzyme’s structural adaptation during the evolution for its maximal activity.
The Hill coefficients for fitting the titration data to the Hill equation suggest the titration for the aspartyl dyad is cooperative in both free and complexed BACE-1, while independent, uncoupled titrations are observed for other residues considered. Such coupled titration behavior makes computing the microscopic pKa values difficult. Hence, the statistical errors associated with computing the pKa’s of the dyad are higher than those of other titratable residues.
At pH 4.5, the aspartyl dyad exists in an ensemble of protonated and deprotonated species in apo BACE-1 (Fig 7). Upon binding of inhibitors, however, significant shifts in the pKa values are observed for the dyad, with both Asp32 and Asp228 having pKa values between 8 and 10 (Table 1). Inhibitor binding effectively alters the protonation state of the dyad at pH 4.5 to its diprotonated form in all cases studied here. The protonated forms are preferred for both aspartates in the presence of inhibitor. This diprotonated state likely compensates for the unfavorable energetics associated with desolvation upon inhibitor binding, allowing for hydrogen bonds between the dyad and the bound inhibitor. The minimal pKa shifts observed for the remaining titratable residues of BACE-1 upon complex formation imply a thermodynamic linkage between inhibitor binding and proton transfer primarily localized at the dyad.
Among several computational efforts to determine protonation state of the aspartyl dyad in BACE-1, a recent work by Domínguez et al. also examined the pKa’s of buried titratable groups in the 2B8L and 2IRZ systems [40], whose computed values are compared with ours in Table 3. At first glance, the predicted pKa values differ the most between the two works for Asp138 and Asp228. In Domínguez et al., the pKa of Asp228 does not deviate much from pKa,ref of Asp (4.0) in both systems compared, while Asp32 was predicted to undergo more significant pKa shift relative to its dyad partner. On the other hand, pKa’s of both aspartates in the dyad shifted to more basic values of pH in our calculations.
When comparing the results in Table 3, several distinctions in the methods used in two studies for pKa prediction should be noted. First, the GB-OBC implicit solvent model [103] was used in Domínguez et al. [104] while the GB-Neck 2 model, in which improved results have been obtained with the added parameters to the GB-OBC [105], was employed in the pH-REMD method used here. Also, the internal dielectric constant of 10 [104] was used in their work while the GB-Neck 2 implementation in Amber 14 employs 1 [105]. In addition to the difference in the force fields utilized, perhaps more importantly, the conformational changes upon inhibitor binding were not rigorously accounted for in their study. Although the detailed comparison of the algorithms used for pKa calculation is beyond the scope of this work, addressing the conformational aspect is particularly important for studying BACE-1 due to its flexible dynamic nature. This is especially crucial when computing the pKa values of the titratable groups in the absence of bound inhibitors, as the conformational fluctuations in the flap also imply the change in solvent accessibility. Hence, the dielectric response of the aspartyl dyad can be different in the presence and absence of bound inhibitors, which in turn can affect the computed pKa values. Consequently, it is evident that conformational transitions accompanying binding of inhibitors should be accounted for in calculation of pKa values of BACE-1. As all previous attempts to compute the pKa’s of the titratable groups in BACE-1 have been largely limited to the use of static X-ray crystallographic structures, our results obtained from concurrent sampling of protonation and conformational spaces by pH-REMD provide a new insight into the microscopic pKa values of BACE-1.
Application of our recently developed constant pH molecular dynamics (CpHMD)-based computational protocol [89], which applies the binding polynomial formalism to address the pH dependence of binding free energies, enables us to obtain proton-linked binding free energy profiles of various inhibitors. As shown in Fig 9 and S9 Fig, all inhibitors bind most strongly at acidic pH. The changes in binding free energies are most pronounced in the pH range of 4 to 10, which essentially encompasses most biological reactions. The deviations in binding free energies within this pH range from the reference binding free energies at pH 4.5 arise from the shift in populations of major protonated species of the titratable residues, primarily those of the dyad. Between pH 5 and 10, the dyad starts to populate the deprotonated species (Fig 8), and as the deprotonated forms of the dyad develop, hydrogen bonds made between the diprotonated state of the dyad and bound inhibitors at low pH break. Subsequently, the binding free energies of the inhibitors become very unfavorable as pH increases. Such observations are impossible with cMD simulations where the protonation states are fixed and fractional protonation is not allowed. This highlights the benefit of using CpHMD method in order to address cases in which changes in protonation states are critical.
Furthermore, our results emphasize the importance of correctly addressing the binding-induced changes in protonation states in protein-ligand systems where binding accompanies a net proton transfer. In conventional molecular modeling or free energy computations, the protonation states of the titratable groups, which are set ahead of time, are fixed and assumed to be identical for both free and bound states. Consistent with this convention, consider a hypothetical scenario in which both Asp32 and Asp228 are assumed to be completely protonated in both apo and holo states. In the case of 2B8L system, such protonation state of the dyad will result in binding free energy of the inhibitor of -9.3 kcal/mol (Fig 9A). On the other hand, when both aspartates are considered fully deprotonated, the binding free energy of the inhibitor is -1.3 kcal/mol. In these two extreme scenarios where the identical, discrete protonation state of the dyad is assumed for both free and bound states, the binding free energies deviate from the true free energy in which the protonation states are considered separately for apo and holo states and fractional protonation is allowed. The errors are as large as 8 kcal/mol for the 2B8L system and similar deviations are noted for other inhibitor-bound systems considered here, ranging between 8 and 12.6 kcal/mol. Such errors are nontrivial and the magnitude is in great excess of typical errors from free energy computations [106,107].
In addition, we note the lack of binding free energies of inhibitors to BACE-1 that are experimentally measured at pH levels other than pH 4.5. For BACE-1, the inhibition assays are traditionally carried out at pH 4.5 where the catalytic activity of BACE-1 is maximal. However, from a free energy computational standpoint, it would be greatly beneficial if binding free energies were measured at other pH levels to incorporate the effect of pH into free energy computations. Availability of experimental reference binding energies at various pH will be of great importance to pushing the free energy computation field forward.
The results presented here demonstrate the dynamics of BACE-1 controlled by solvent pH. The flexible motions of the flap region at low pH, assisted by the diprotonated state of the aspartyl dyad, enable the enzyme’s optimal catalytic mechanism at acidic environment, implying a linkage between the protonation equilibria, conformational dynamics, and catalytic activity of BACE-1. In addition, we show the thermodynamic relation between binding of inhibitors and protons at the active site of BACE-1. Our results highlight the importance of accurately accounting for the protonation states of the titratable groups in protein-ligand systems where ligand binding is pH-dependent. Furthermore, we show that the CpHMD method can be used as an all-purpose tool to assess the pH-dependent dynamics and to quantify the binding free energies for protein-ligand systems where the protonation equilibria play an important role. To the best of our knowledge, this work presents the first application of our CpHMD-based free energy method to protein-ligand systems. In using the method, absolute binding free energies obtained by computational free energy calculations such as thermodynamic integration can be used in cases where experimental association constants are not available. Our results highlight high utility of CpHMD method to address the effect of pH on conformational dynamics and inhibitor binding in computer-aided drug discovery workflows.
Baptista et al. developed the constant pH molecular dynamics (CpHMD) method to enable concurrent sampling of discrete protonation states and conformational space according to the semi-grand canonical ensemble [78]. In this work, we apply the flavor of CpHMD coupled with replica exchange (pH-REMD) [108] implemented in the AMBER 14 suite of programs [95] with generalized Born (GB) electrostatics. In the CpHMD simulations, a conventional molecular dynamics (MD) simulation is periodically interrupted by a Monte Carlo (MC) step, in which a change in the protonation state of a random titratable residue is considered [81]. Acceptance of the new protonation state is contingent on the computed transition free energy, ΔGtrans:
ΔGtrans=kBT(pH−pKa,ref)ln10+ΔGelec−ΔGelec,ref,
(1)
where pH enters as an external thermodynamic parameter and kBT is the Boltzmann constant multiplied by the temperature of the system. For a given value of pH, the difference in electrostatic free energy that accompanies the change in protonation being considered, ΔGelec, is computed with respect to the difference in electrostatic free energy that accompanies analogous change in protonation for a model compound, ΔGelec,ref, which has a known pKa value (pKa,ref). As all titratable residues in this study are protein residues, the model compounds referenced in Eq 1 are individual amino acids in GB solvent. The respective pKa,ref values for CpHMD in AMBER 14 are 4.0 for Asp, 4.4 for Glu, 6.5 and 7.1 for His, 9.6 for Tyr, and 10.4 for Lys [95]. Computing ΔGtrans with respect to these model compounds enables cancellation of non-classical terms. The Metropolis criterion is then applied to ΔGtrans to determine whether to accept the proposed change in protonation, and the MD simulation is resumed. Repeated application of these steps builds an ensemble of protonation states along the MD trajectory.
In pH-REMD, the exchange between adjacent replicas is achieved in the pH dimension at a fixed conformation, whose acceptance ratio is dependent on the MC exchange probability for replicas i and j:
Pi→j=min{ 1,exp[ ln10(Ni−Nj)(pHi−pHj) ] },
(2)
where Ni is the number of titratable protons in replica i and pHi is the pH of replica i prior to the exchange attempt [108,109]. By enhancing the sampling through an application of replica exchange scheme, the method has been shown to achieve faster convergence and better sampling in both conformational and protonation spaces compared to original CpHMD [83,108].
From the pH-REMD simulations, the pKa of a given residue is computed as the midpoint of titration by fitting titration data to the Hill equation:
s = 11+10n(pKa-pH),
(3)
where s is the fraction of deprotonated species for a given residue and n is the Hill coefficient. The fraction of deprotonated species (s) for a titratable group is obtained at each value of pH from the pH-REMD simulations. In using Eq 8 shown below, the fractions of protonated species (1-s) of the protein-inhibitor complex and protein can be translated into ZPL and ZP, respectively.
The binding polynomial formalism devised by Wyman [90] was used by Tanford to study protein denaturation [91] and by several groups to examine the pH dependence of protein-protein binding [68,110]. Following their theoretical foundations, we recently applied it to binding of a general receptor to a ligand with a single titratable site, and the detailed derivation of the formalism can be found therein [89]. Here we briefly outline the theoretical basis of the method and show its usage for protein-ligand binding with multiple titratable sites.
First, consider complex formation between a protein (P) with a single titratable site and a ligand (L) that does not titrate in the biological range of pH levels. The association can be expressed as a general equation with the apparent equilibrium constant, Kapp:
{P}+L⇄Kapp{PL},
(4)
where the curly braces indicate that the ensembles of protein and protein-ligand complex (PL) may contain different protonated forms of the titratable species. Kapp can be expressed in terms of binding polynomials through an application of the thermodynamic cycle for proton-linked ligand binding shown in Fig 10:
Kapp=[PL]+[HPL+]([P]+[HP+])[L]=Kb∘(1+[HPL+][PL])(1+[HP+][P]),
(5)
where the concentrations of the given species, instead of activities, are shown assuming ideal dilute solutions and K°b is the equilibrium constant of binding for a reference reaction in which net proton transfer is ignored.
The overall free energy of binding (ΔG°) can then be expressed by using logarithmic representations of the acid dissociation constants for the free protein (pKaF) and protein-ligand complex (pKaC):
ΔG∘(pH)=ΔGref∘−kBTln(1+10pKaC−pH1+10pKaF−pH),
(6)
where ΔG°ref is the free energy of binding for the reference reaction.
In cases where proton binding to different titratable sites may be cooperative, Wyman [111] derived a relation between Kapp and pH such that
∂lnKapp∂ln[H+]=ΔνH+=ZPL−(ZP+ZL),
(7)
where, using the notation used by Tanford [91], ΔνH+ is the change in the number of bound protons in the protein-ligand complex, relative to the number of protons bound to the protein and ligand individually. With ΔZ = ZPL−(ZP + ZL), integration of Eq 7 provides a thermodynamic relation for proton-linked ligand binding where titratable sites may interact:
ΔG∘(pH)=ΔGref,pH∘−kBTln(10)∫pHrefpH{ ZPL(pH)−ZP(pH)−ZL(pH) }dpH,
(8)
where ZPL, ZP, and ZL are the total charges for protein-ligand complex, protein, and ligand, respectively, obtained from the pH-REMD simulations. Eq 8 provides framework for computing the pH-dependent binding free energy through a correction term to the reference free energy of binding. In this work, ZL is omitted from consideration since the inhibitors considered here do not titrate in the physiological pH range.
The X-ray crystallographic structures of BACE-1 in complex with inhibitors N-[(1S,1R)-benzyl-3-(cyclopropylamino)-2-hydroxypropyl]-5-[methyl(methylsulfonyl)amino-N’-[(1R)-1-phenylethyl]isophthalamide (PDB ID 2B8L) [112]; N-[(1S,2S,4R)-2-hydroxy-1-isobutyl-5-({(1S)-1-[(isopropylamino)carbonyl]-2-methylpropyl}amino-4-methyl-5-oxopentyl]-5-methyl(methylsulfonyl)amino]-N’-[(1R)-1-phenylethyl]isophthalamide (PDB ID 2P4J) [113]; N~2~-[(2R,4S,5S)-5-{[N-{[3,5-dimethyl-1H-pyrazol-1-yl)methoxy]carbonyl}-3-(methylsulfonyl)-L-alanyl]amino}-4-hydroxy-2,7-dimethyloctanoyl]-N-isobutyl-L-valinamide (PDB ID 2G94) [100]; 3,{5-[(1R)-1-amino-1-methyl-2-phenylethyl]-1,3,4-oxadiazol-2-yl}-N-[(1R)-1-(4-fluorophenyl)ethyl]-5-[methyl(methylsulfonyl)amino]benzamide (PDB ID 2IRZ) [114]; and N1-((2S,3S,5R)-3-amino-6-(4-fluorophenylamino)-5-methyl-6-oxo-1-phenylhexan-2-yl)-N3,N3-dipropylisophthalamide (PDB ID 2FDP) [102] were used to study the pH-dependent inhibitor binding to BACE-1. The chemical structures of the inhibitors are shown in Fig 2. A segment between residues Gly158 and Ser169 is not solved in 2B8L, 2IRZ, and 2FDP crystal structures and hence this loop was built by homology modeling using the Structure Prediction Wizard module of Schrödinger’s Prime program [115–117]. The FASTA sequence of protein including the missing loop region for each X-ray structure was obtained from UniProt [118]. The mutation AWA that exists in 2B8L and 2IRZ structures made for crystallography were corrected to the original sequence of KWE [112,114]. Using the homologs found by BLAST search algorithm [119], a chimera model containing the missing loop region was built for each structure. The homology-modeled loop region was energy-refined for relaxation using the Refine Loops panel of the Prime program [115]. Apo structure of BACE-1 was generated by removing the bound inhibitor from the refined 2B8L structure to even out any effects that may arise from homology modeling.
The geometries of the inhibitors were optimized at the B3LYP/6-31G(d) level of theory [120–123] using the Gaussian 09 suite of programs [124]. The electronic potentials (ESP) for the optimized geometries of the inhibitors were computed using MK radii [125] at the HF/6-31G(d) level of theory. Subsequently, the atomic point charges were computed from the ESPs by applying the RESP procedure using the antechamber module [126] in the AmberTools 14 suite of programs [95]. All other force field terms including Lennard-Jones parameters for use in molecular dynamics (MD) simulations were taken from the general AMBER force field (GAFF) [127].
Prior to the pH-REMD simulations, each system was subject to short conventional molecular dynamics (cMD) simulations for equilibration purposes. Protonation states of the titratable groups are assigned using the PROPKA web server [75,92–94]. All protein force field parameters are taken from the AMBER ff14SB force field [95], while the ligand parameters are taken from the AMBER GAFF force field [127]. Each system was solvated with TIP3P water [128] and counterions were added to neutralize the system by tleap program [129]. Water molecules were first minimized and simulated for 150 ps in the NPT ensemble with a harmonic restraint of 2.0 kcal/mol Å2 on the protein and ligand heavy atoms to relax the water. The entire system was then minimized and heated to 300 K over 500 ps. Two equilibrations with respective duration of 200 ps were performed. First, the system was equilibrated at constant volume and temperature (NVT) using a Langevin thermostat [130]. Following this, the second equilibration was carried out at constant pressure and temperature (NPT) using a Berendsen barostat (ntp = 1) [131] with isotropic position scaling to bring the system to a stable density. A 100 ns cMD production was then performed in the NVT ensemble. The Particle Mesh Ewald summation method was used to compute long-range electrostatic interactions [132,133], and short-range non-bonded interactions are truncated at 8 Å in the periodic boundary conditions. All dynamics are conducted using the pmemd.cuda module of AMBER 14 suite of programs [95,130,134]. The RMSD of the apo structure indicated a convergence to the starting conformation after first 20 ns, ensuring the stability of the system (S10 Fig).
Preliminary investigation of the pKa shift of the titratable groups of BACE-1 upon inhibitor binding was carried out using the PROPKA web server by submitting the apo and holo structures [75,92–94]. The results indicated binding-induced pKa shifts for a number of ionizable residues, with the most pronounced shift for the aspartyl dyad. A total of ten ionizable residues within 12 Å of the active site were chosen for titration, Asp32, Asp106, Asp138, Asp223, Asp228, Glu116, Glu265, Glu339, His45, and Tyr71. As the titratable groups in the inhibitors considered in this study have pKa’s above 12, the titration was carried out on the chosen protein side chains only.
pH-REMD simulations were performed using the pmemd.cuda.MPI module of AMBER 14 suite of programs for the pH range between 1 and 12 stepped by 1 pH unit [95,108]. All simulations employed the generalized Born (GB)-Neck 2 implicit solvent model (igb = 8) [105] with a salt concentration of 0.1 M. To ensure equilibration in the implicit solvent, a 5,000 step minimization was carried out for each system starting from the conformation obtained from the cMD simulations, with positional restraints on all heavy atoms with a force constant of 20 kcal/mol Å2. The system was then heated to 300 K over 500 ps using a Langevin thermostat while maintaining the positional restraints applied to all heavy atoms with a force constant of 5 kcal/mol Å2, followed by a 1 ns equilibration. The Monte Carlo (MC) moves for titration were performed during the production stage only, where the MC steps taken every 10 fs and exchange between replicas at adjacent pH attempted every 100 steps, i.e., 200 fs with a 2 fs time step. The production simulations were carried out for 60 ns and data from last 50 ns were used for analyses. In the equilibration and production steps, the bonds involving hydrogen atoms were constrained using the SHAKE algorithm [135].
RMSF and clustering analyses, and reconstruction of the pH-based trajectories from the pH-REMD simulations were performed using cpptraj program in the AmberTools 14 suite of programs [95]. Clustering analyses for the pH-based trajectories used pairwise RMSDs computed for Cα atoms between frames to divide the trajectories into five clusters using the average-linkage algorithm [136]. Fitting of titration data to the Hill equation (Eq 3) to obtain the pKa values was carried out using Matlab [137].
|
10.1371/journal.ppat.1004324 | RC1339/APRc from Rickettsia conorii Is a Novel Aspartic Protease with Properties of Retropepsin-Like Enzymes | Members of the species Rickettsia are obligate intracellular, gram-negative, arthropod-borne pathogens of humans and other mammals. The life-threatening character of diseases caused by many Rickettsia species and the lack of reliable protective vaccine against rickettsioses strengthens the importance of identifying new protein factors for the potential development of innovative therapeutic tools. Herein, we report the identification and characterization of a novel membrane-embedded retropepsin-like homologue, highly conserved in 55 Rickettsia genomes. Using R. conorii gene homologue RC1339 as our working model, we demonstrate that, despite the low overall sequence similarity to retropepsins, the gene product of rc1339 APRc (for Aspartic Protease from Rickettsia conorii) is an active enzyme with features highly reminiscent of this family of aspartic proteases, such as autolytic activity impaired by mutation of the catalytic aspartate, accumulation in the dimeric form, optimal activity at pH 6, and inhibition by specific HIV-1 protease inhibitors. Moreover, specificity preferences determined by a high-throughput profiling approach confirmed common preferences between this novel rickettsial enzyme and other aspartic proteases, both retropepsins and pepsin-like. This is the first report on a retropepsin-like protease in gram-negative intracellular bacteria such as Rickettsia, contributing to the analysis of the evolutionary relationships between the two types of aspartic proteases. Additionally, we have also shown that APRc is transcribed and translated in R. conorii and R. rickettsii and is integrated into the outer membrane of both species. Finally, we demonstrated that APRc is sufficient to catalyze the in vitro processing of two conserved high molecular weight autotransporter adhesin/invasion proteins, Sca5/OmpB and Sca0/OmpA, thereby suggesting the participation of this enzyme in a relevant proteolytic pathway in rickettsial life-cycle. As a novel bona fide member of the retropepsin family of aspartic proteases, APRc emerges as an intriguing target for therapeutic intervention against fatal rickettsioses.
| Several rickettsiae are pathogenic to humans by causing severe infections, including epidemic typhus (Rickettsia prowazekii), Rocky Mountain spotted fever (Rickettsia rickettsii), and Mediterranean spotted fever (Rickettsia conorii). Progress in correlating rickettsial genes and gene functions has been greatly hampered by the intrinsic difficulty in working with these obligate intracellular bacteria, despite the increasing insights into the mechanisms of pathogenesis of and the immune response to rickettsioses. Therefore, comparison of the multiple available genomes of Rickettsia is proving to be the most practical method to identify new factors that may play a role in pathogenicity. Here, we identified and characterized a novel retropepsin-like enzyme, APRc, that is expressed by at least two pathogenic rickettsial species, R. conorii and R. rickettsii. We have also established that APRc acts to process two major surface antigen/virulence determinants (OmpB/Sca5, OmpA/Sca0) in vitro and we suggest that this processing event is important for protein function. We demonstrate that APRc is specifically inhibited by drugs clinically used to treat HIV infections, providing the exciting possibility of targeting this enzyme for therapeutic intervention. With this work, we demonstrate that retropepsin-type aspartic proteases are indeed present in prokaryotes, suggesting that these enzymes may represent an ancestral form of these proteases.
| The genus Rickettsia represents a group of gram-negative obligate intracellular bacteria that exist as pathogens and symbionts of eukaryotic cells. These bacteria are transmitted to mammals by arthropod vectors such as ticks, lice, and fleas. With the advent of new molecular biology tools and whole genome sequence analysis Rickettsia species have been classified into several distinct genetic groups including the ancestral group (AG), spotted fever group (SFG), typhus group (TG), and transitional group (TRG) [1]–[4]. Many rickettsial species belonging to the TG and SFG are pathogenic to humans causing serious illnesses, including epidemic typhus (Rickettsia prowazekii), Rocky Mountain spotted fever (RMSF) (Rickettsia rickettsii), and Mediterranean spotted fever (MSF) (Rickettsia conorii) [5]–[7]. The life-threatening character of many rickettsial species is the consequence of their highly virulent properties and unique biological characteristics including aerosol transmission, persistence in infected hosts, and low infectious dose. There is growing concern about rickettsial diseases and their impact on global health, with members of the genus Rickettsia being identified, together with other bacteria, as emerging/re-emerging pathogens, responsible for the majority of emerging infectious diseases events between 1940 and 2004 [8]. Although in the U.S. the case fatality rate for RMSF has declined over the years (to less than 0.5% in 2010), the Brazilian spotted fever (also caused by R. rickettsii) has case fatality rates ranging from 30% to 80% [9]. MSF is also associated with high morbidity and mortality, with case fatality rates varying from 21% to 33% in Portugal [9]. In fact, from 1989 to 2000 the incidence rate of MSF in Portugal was one of the highest in the Mediterranean area (9.8 cases per million) [10]. From the typhus group, R. prowazekii infection is still recognized as one of the most severe rickettsioses with case fatality rates as high as 12%. The emergent and severe character of rickettsioses with their associated high morbidity and mortality rates, together with the lack of protective vaccines, strengthen the importance of identifying new protein factors that may work as potential targets for the development of more efficacious therapies against these diseases [6], [11].
In line with what has been described for other obligate intracellular bacteria, rickettsial species have highly conserved and reduced genome sizes, which derive from reduction of originally larger genomes accompanying the adaptation to strict intracellular lifestyles [12]–[15]. Significant progress has been made with regards to the analysis of genetic composition of a number of rickettsial species (now 55 sequenced genomes); however, the genetic intractability of these bacteria has severely limited molecular dissection of virulence factors associated with their intracellular parasitism and pathogenic mechanisms. Comparative genomics has resulted in identification of several genes encoding secreted proteins that are potential virulence factors involved in pathogenesis [16]–[18]. However, many more putative rickettsial virulence factors and hypothetical proteins remain to be functionally defined.
Bacterial pathogenicity generally results from a combination of factors and there are different bacterial components and strategies contributing to virulence [19]. Among these components emerges a diverse array of proteolytic enzymes (mainly localized to the bacterial surface or secreted), which have been recognized as virulence factors in several pathogenic bacteria. Such enzymes play critical functions related to colonization and evasion of host immune defenses, acquisition of nutrients for growth and proliferation, and facilitation of dissemination or tissue damage during infection [19]–[21]. The relevance of proteolytic events for bacterial pathogenicity and the progressive increase in antibiotic resistance among pathogenic bacteria contribute to positioning proteases as potential candidate targets for the development of alternative antibacterial strategies [20]. However, thus far only a few proteases have been identified and characterized in Rickettsia [18], [22], [23].
In this work, we have identified a gene coding for a putative membrane-embedded aspartic protease (AP) of the retropepsin type, conserved in all 55 sequenced Rickettsia genomes. The retropepsins (also known as family A2 of aspartic proteases) were first identified with the discovery of the Human Immunodeficiency Virus 1 (HIV-1) protease (PR) in the late 1980's [24] and the recognition of its essential role in the maturation of HIV-1. These proteases require homodimerization of two monomeric units in order to form a functional enzyme, structurally related to the pepsin family (A1) of bi-lobal APs [25]–[27]. Interestingly, the existence of retropepsin-like enzymes in prokaryotes has always been a matter of debate [28], [29] and never unequivocally demonstrated. Herein, we describe the characterization of the retropepsin-like homologue from Rickettsia conorii (RC1339/APRc – for Aspartic Protease from Rickettsia conorii) and demonstrate that this protease shares several enzymatic properties (e.g. autolytic activity, optimum pH, sensitivity to specific HIV-1 PR inhibitors) with other APs. Moreover, we demonstrate that this novel protease is expressed in vivo in two pathogenic species of Rickettsia and provide experimental evidence for its potential role as a modulator of rickettsial surface cell antigen (Sca) proteins OmpB (Sca5/rOmpB) and OmpA (Sca0/rOmpA).
This work provides the first evidence for a retropepsin-like protease in gram-negative intracellular bacteria such as Rickettsia, and the contribution of these results to change the paradigm on the evolutionary relationships between pepsin-like and retroviral APs is also discussed.
In silico analysis of the genome sequence of R. conorii str. Malish 7, the etiologic agent of MSF, revealed a gene (RC1339) with 696 bp, encoding a putative retropepsin-like aspartic protease. This gene is highly conserved among 55 sequenced Rickettsia genomes, with deduced amino acid sequences sharing more than 84% sequence identity among each other. This striking pattern of conservation is illustrated in Figure 1A, which shows the alignment of the deduced amino acid sequence of R. conorii RC1339/APRc with eight other homologues from representative species of each rickettsial group (SFG, TG, TRG and AG). Protein sequences are highly conserved, with amino acid identities ranging from 84.0% (R. bellii str. OSU 85–389 APRc homologue) to 99.6% (R. parkeri str. Portsmouth APRc homologue) (Table 1). Strikingly, this novel type of rickettsial AP showed no detectable sequence homology when compared with APs from other organisms, except for the presence of the hallmark sequence motifs of family A2 members. Although the overall sequence identity with other retropepsins was found to be lower than 14% (with only 6% for the HIV-1 PR which is considered the archetypal member of this family of APs), it was possible to identify the active site consensus motif Asp-Thr-Gly (contained in the sequence Xaa-Xaa-Asp-Xbb-Gly-Xcc, where a Xaa is hydrophobic, Xbb is Thr or Ser, and Xcc is Ser, Thr or Ala) corresponding to the sequence Met-Val-Asp-Thr-Gly-Ala (amino acids 138–143), followed downstream by a hydrophobic-hydrophobic-Gly sequence (Leu-Leu-Gly, amino acids 208–210). These features are characteristic of retropepsin-like APs which are obligate homodimers, with each monomer contributing one catalytic residue and one hydrophobic-hydrophobic-Gly motif to form the structural feature known as psi loop [25], [26]. A distinguishing feature of APRcs compared to retroviral enzymes is their predicted membrane-embedded nature, with different algorithms predicting three putative transmembrane (TM) α-helix segments in the N-terminal domain of APRc (Figure 1A). The presence of cysteine residues in these predicted transmembrane regions, which may be linking the three α-helical chains together through interchain disulfide bonds, may likely contribute to structural stability. Additionally, an inside orientation for the N terminus and an outside orientation for the C-terminal soluble protease domain of APRc (Arg87-Tyr231) relative to the membrane was predicted by the HMMTOP2 [30]. Interestingly, a similar domain organization - putatively membrane embedded with a soluble catalytic domain - with variations in the number of predicted TM helices has been also described for eukaryotic retropepsin-like proteases such as human and mouse SASPase [31], [32], as well as for one putative retroviral-type AP (SpoIIGA) [33] identified in Bacillus subtilis, for which no enzymatic characterization is yet available.
Despite the high divergence at the sequence level, a structure-based alignment of the putative soluble catalytic domain of RC1339/APRc with HIV, Equine Infectious Anemia Virus (EIAV), and Xenotropic Murine Leukemia Virus-related Virus (XMRV) retropepsins, as well as with Ddi1 putative protease domain (Figure 1B), further suggested an overall retention of structural similarity through conservation of secondary structure, thereby anticipating an evolutionary relationship between APRc and retropepsins.
Using R. conorii RC1339 as our working model, we started assessing, by producing the soluble catalytic domain fused to GST in E. coli, whether this gene would indeed encode an active aspartic protease. Assuming the predicted boundary between the transmembrane and soluble catalytic domains at Phe86-Arg87, the synthetic codon optimized sequence coding for the whole soluble domain was cloned into pGEX-4T2 (pGST-APRc87–231) and the fusion construct was expressed in E. coli (BL21 Star (DE3) strain). The soluble fraction of the cell lysates was subjected to a GSTrap HP affinity chromatography and the eluted fractions were pooled and further purified by size-exclusion chromatography on a Superdex 200 HiLoad 26/60. Purified fractions analyzed by SDS-PAGE confirmed the presence of the fusion protein with the molecular mass of approximately 42 kDa, as well as free GST (25 kDa), which likely results from proteolytic degradation by the host. One of the features shared by the retropepsins is their autoprocessing activity which promotes their own release from a larger polyprotein precursor [25]. As shown in Figure 2A, our results demonstrate that recombinant rGST-APRc87–231 also undergoes a multi-step processing in vitro upon incubation at pH 6.0, resulting in the generation of different cleavage products over activation time. Edman sequencing of these APRc fragments allowed the identification of the three autolytic cleavage sites: Tyr92-Ala93, Met98-Ser99, and Ser104-Tyr105 (Figure 2B).
As a first approach to assess enzyme activity, we used oxidized insulin B chain as a substrate as this polypeptide is usually cleaved by aspartic proteases, and tested its cleavage over activation time for purified rGST-APRc87–231. As illustrated in Figure 2C, samples from each time point (0, 12, 24, 36 and 48 h) were tested and the reaction products separated by RP-HPLC. Interestingly, the presence of several insulin cleavage products was concomitant with the appearance of the activation product APRc105–231, suggesting that autoprocessing may be an essential step for the activation of recombinant APRc.
In order to evaluate the role of the putative catalytic aspartate for this autoprocessing activity, an active site mutant of rGST-APRc87–231, where the putative catalytic aspartate residue was mutated to an alanine [rGST-APRc(D140A)87–231] was produced, purified, and activated under the same conditions as for the WT fusion protein. As predicted, the mutation significantly affected the activation process (Figure 2A, right panel), suggesting that APRc is dependent on the conserved catalytic aspartate residue for triggering autolytic activity. To confirm that the previously observed insulin degradation resulted from APRc activity, parallel tests were performed with the active site mutant rGST-APRc(D140A)87–231 incubated under similar conditions (Figure S1). The autoprocessing ability of APRc and the importance of the catalytic aspartate were further confirmed by expressing the constructs harboring the soluble domain (rGST-APRc87–231) and its active site mutant (rGST-APRc(D140A)87–231) in E. coli and by analyzing total soluble fractions for the presence of APRc-activated forms with a specific APRc polyclonal antibody (raised towards amino acids 165–178). As shown in Figure 2D, and consistent with the results obtained in our in vitro assays, the activation products were only detected when the WT sequence was expressed, further corroborating the role of the catalytic aspartate in autoprocessing. This intrinsic autoprocessing observed during expression in E. coli is in line with what has been documented for other retropepsins (e.g. HIV-1 PR and XMRV PR [34], [35]).
Since the expression of the soluble domain of APRc fused to GST resulted in a high degree of contamination with free GST, an alternative strategy was undertaken to streamline the production of APRc activation product with higher yield and purity. For this we designed two new constructs where the sequences encoding the intermediate of activation APRc99–231, as well as the final product APRc105–231, were cloned into pET23a expression vector (Invitrogen) in frame with a C-terminal 6×His-tag. Both constructs were readily expressed in the soluble form in E. coli. A purification protocol consisting of a Ni-IMAC step, followed by dialysis of APRc-enriched polled fractions, and further purification through a cation exchange chromatography with a Mono S column was optimized. As shown in Figure 3A, the His-tagged intermediate rAPRc99–231 was also able to undergo auto-activation into the mature form at pH 6.0. To further characterize the enzymatic activity of APRc we designed a specific fluorogenic substrate, which mimics the identified auto-cleavage site between Ser104-Tyr105 residues (Rick14 peptide: MCA-Lys-Ala-Leu-Ile-Pro-Ser-Tyr-Lys-Trp-Ser-Lys-DNP), and tested this substrate during rAPRc99–231 activation. As previously observed with the GST-fusion precursor, activity towards this substrate was shown to be also dependent on the conversion step and the highest activity was observed upon accumulation of the conversion product (Figure 3A–B), further strengthening the importance of enzyme activation. Interestingly, when the final product APRc105–231 was directly produced in E. coli, no proteolytic activity was observed towards the same substrate (data not shown). This result suggests that protease autoprocessing may indeed be accompanied by some conformational change that is not observed when the activation product is directly expressed in E. coli. Wan and co-workers have reported a similar result for HIV-1 PR by showing that a recombinant protein corresponding to the mature form of the protease (99 amino acids) with two additional amino-acids at the N-terminus (Met and Gly) displayed no proteolytic activity [36]. Based on this result, we have focused on the construct of the precursor form APRc99–231 for further analysis.
Given the observed impact of mutating the catalytic Asp residue on the autoprocessing ability of APRc we decided to evaluate the effect of pepstatin (a classical inhibitor of aspartic proteases) and indinavir (an HIV-1 PR inhibitor) in rAPRc99–231 autoprocessing. Our results (Figure S2A) show that in the presence of pepstatin the auto-activation step was slightly slowed, whereas indinavir had no apparent inhibitory effect on this auto-processing activity. Surprisingly, EDTA inhibited rAPRc99–231 auto-activation, suggesting that a metal ion may be involved in proper folding and/or enzyme activity.
Given the homodimeric nature of retropepsins, crosslinking studies using DSS as the crosslinking agent were also conducted with purified rAPRc99–231, as well as with the derived activation product, in order to provide an evidence for APRc dimer formation. Reaction products and control samples were analyzed by immunoblotting (Figure 3D) and, as expected, the results revealed a significant amount of APRc associated as dimer, although monomeric and larger aggregate species were also visible. Similar results were observed when glutaraldehyde, which differs from DSS in the length of connecting backbone (11.4 Å for DSS and 7 Å for glutaraldehyde) was used (Figure S2B). These results were consistent with the analysis of both forms by analytical size-exclusion chromatography under native conditions (Figure S2C) where the presence of oligomeric species was observed, although the large majority of the protein was shown to accumulate as a monomer.
Based on the observed enzymatic activity upon conversion of the precursor form rAPRc99–231, all characterization studies were focused exclusively on this derived activation product (for simplification denoted APRc). The effect of pH was determined using the same fluorogenic substrate - MCA-Lys-Ala-Leu-Ile-Pro-Ser-Tyr-Lys-Trp-Ser-Lys-DNP - in a range of pH values from 4 to 9. From this analysis an optimal activity at pH 6.0 was observed (Figure 4A), with no appreciable hydrolytic activity below pH 5.0. This higher optimal pH value is consistent with optimum pH values reported for other retropepsins [37], [38].
When investigating the susceptibility of APRc to classical protease inhibitors (Figure 4B), this protease was shown to be insensitive to pepstatin A, even though a slightly inhibitory effect was observed during autolytic processing. In contrast, APRc activity was strongly inhibited by EDTA, retaining only 25% activity, and a small inhibitory effect was also observed with Pefabloc. No significant effect was observed after incubation with E-64, whereas incubation with Zn2+ (Figure 4B) slightly affected enzyme activity.
In order to provide additional evidence that APRc is indeed a retropepsin-like enzyme we analyzed the effects of different clinical inhibitors of HIV-1 PR. Strikingly, incubation with indinavir resulted in a near complete inhibition of APRc, even when tested at a final concentration of 0.25 mM in the assay. Additionally, nelfinavir, saquinavir, amprenavir and atazanavir also had a significant inhibitory effect, ranging between approximately 30–50% of inhibition (Figure 4C). This inhibitory effect of specific HIV-1 protease inhibitors against a prokaryotic retropepsin-like enzyme has not been previously described.
To determine the APRc sequence specificity we employed Proteomic Identification of Protease Cleavage Sites (PICS) [39], [40], a high-throughput profiling approach based on the use of database-searchable proteome-derived oligopeptide libraries representing the natural biological sequence diversity as test substrates [39]. A unique advantage of PICS is the simultaneous determination of sequence specificity on both sides of the scissile bond, the prime-side (P′) and non-prime side (P) [41]. PICS substrate peptide libraries are prepared by digestion of a complex proteome from a sequenced model organism with highly specific endoproteases such as trypsin (cleavage after Arg and Lys) or GluC (cleavage after Glu and Asp), followed by blocking of primary amines at peptide N termini and in Lys side chains. Cleavage of library peptides by the test protease of interest generates C-terminal cleavage products with free α- amines that are exploited for selective enrichment and identification by LC-MS/MS. The identified cleavage products constitute the prime side sequences (P′) of the cleaved library peptides. Due to the non-random nature of the PICS libraries the non-prime side sequences can be inferred by database matching to allow reconstruction of the complete cleavage sites.
In this work, active APRc was incubated with PICS libraries generated by digestion of total human THP1 (human monocytic leukemia cell line) cell proteins by either trypsin or GluC. These PICS experiments resulted in the identification of 830 and 327 C-terminal cleavage products from tryptic and GluC libraries, respectively (Tables S1 and S2). The corresponding N-terminal cleavage products and complete cleavage sites were obtained and summarized using the WebPICS tool [40]. The PICS-based APRc specificity profiles are shown in Figure 5 and a good agreement was observed between the two complementary peptide libraries. APRc displays only moderate specificity and accepts multiple amino acids at each position. At P1, directly preceding the scissile bond, APRc showed a preference for large hydrophobic residues such as phenylalanine, tyrosine, methionine, leucine, and carboxyamidomethylated cysteine (modified during library preparation). In addition, APRc also preferred the neutral amino acids threonine and asparagine at this site. A similar preference was observed for P1′, although this further included small amino acids alanine, serine, and glycine, as well as aspartate. Notably, cleavage sites were almost devoid of Pro at P1 and P1′. Furthermore, PICS revealed distinct preferences for selected amino acids at other positions, likely reflecting structural constraints imposed by the substrate recognition and binding to the pocket site. In P2, APRc preferences include valine, isoleucine, proline, and threonine, whereas a predominant preference for small and branched aliphatic amino acids alanine, valine, and isoleucine was observed at P2′. More distant from the cleavage site, small preferences for valine and isoleucine at P3 and for alanine and glycine in P3′ and a strong preference for leucine or isoleucine at P4′ were observed. Interestingly, basic and acidic residues were significantly underrepresented throughout. The large number of APRc cleavage sites identified from the tryptic PICS library further allowed investigation of subsite cooperativity. When comparing two of the strongest cleavage site determinants, phenylalanine at P1 and proline at P2, we observed apparent mutual exclusion. Of the 103 unique cleavage sites that contained proline in P2, only 4 had phenylalanine in P1 (3.7% compared to 10.5% occurrence for all identified cleavage sites), which was compensated by more frequent occurrence of P1 methionine (10.3% compared to 5.8% total occurrence) and P1 asparagine (14% compared to 8.3% total occurrence). Correspondingly, peptides with phenylalanine in P1 yielded 4.6% P2 proline (compared with 12.9% total occurrence), whereas peptides with methionine or asparagine in P1 yielded 22.9% or 21.7% P2 proline, respectively. A similar trend was observed in identified cleavage sites from GluC libraries, indicating subsite cooperativity between P2 and P1.
These results clearly show that, although displaying a unique profile, APRc shares some specificity requirements with retropepsins as well as with pepsin-like enzymes (particularly BACE), further supporting APRc has being a member of the aspartic protease family.
As previously mentioned, full-length RC1339/APRc was predicted to be membrane-embedded with an extracytoplasmic orientation of the C-terminal domain. In order to provide experimental validation of these theoretical observations we used E. coli as our working model. An untagged construct in pET28a comprising RC1339/APRc full-length coding sequence was generated and protein expression carried out as described under Experimental Procedures. Protease insertion into the membrane was first assessed by subcellular fractionation studies followed by Western blot analysis with a specific APRc antibody. A band of approximately 21 kDa, whose nature was confirmed by peptide competition assays, was detected in the total membrane fraction and shown to accumulate in the outer membrane (Figure 6A). The purity of the outer membrane (OM) fraction was confirmed by Western blotting against E. coli Lep and OmpA proteins, as inner and outer membrane markers [42], respectively, and compared to the total membrane fraction (Figure 6B). As expected, OmpA was detected in the outer membrane fraction and the absence of cross-contamination with inner membrane proteins was confirmed through loss of signal for Lep, when compared with total membrane fraction. Interestingly, APRc displayed a molecular weight lower than expected (∼21 kDa instead of the predicted 26.4 kDa), and parallel experiments with a C-terminal His-tagged construct confirmed the presence of the tag in the membrane fractions (data not shown), clearly suggesting that the protease may be processed at the N terminus during translocation to the membrane.
In an attempt to expand our knowledge about the membrane topology of APRc, further studies were performed in order to determine the overall in/out orientation of this protein relative to the outer membrane of E. coli. To investigate this, PFA-fixed E. coli cells expressing untagged full-length APRc were subjected to flow cytometry with both anti-APRc and anti-α-subunit of RNA polymerase (mAb 4RA2) antibodies. The staining of E. coli cells with the 4RA2 mAb was primarily used to restrict the analysis to the non-permeable cells. As shown in Figure 6C, after gating out all the cells that stained positive for RNAPαprotein (permeable cells), bacterial surface staining with anti-APRc was observed, confirming the integration of RC1339/APRc into the outer membrane of E. coli and the orientation of the soluble catalytic domain to the extracellular milieu.
To determine whether RC1339/APRc and the R. rickettsii homologue are expressed in the context of the intact bacterium, we isolated total RNA from R. conorii and R. rickettsii grown in Vero cells and performed reverse transcriptase PCR (RT-PCR). As shown in Figure 7A, both R. conorii and R. rickettsii produce transcripts for rc1339 and A1G_07330, respectively, when grown in culture. To confirm expression of these transcripts, protein lysates from each rickettsial species were separated by SDS-PAGE and immunoblotting analyses were carried out with the specific APRc antibody. As depicted in Figure 7B, a major reactive species with an apparent molecular mass of 21 kDa was detected in R. rickettsii whole cell lysate and in the insoluble fraction of the R. conorii extract. These results clearly confirmed that rc1339 gene and its R. rickettsii homologue are indeed translated in both rickettsial species. Interestingly, and as previously observed in E. coli, a molecular weight of around 21 kDa was also detected for APRc in rickettsial extracts. Although we cannot exclude abnormal migration of the protease in the gel, the observed lower molecular weight may also be correlated with APRc processing at the N terminus, as anticipated by our results in E. coli.
To provide additional insights on the localization of APRc in these rickettsial species, fractionation studies were performed on purified bacteria. Whole cell lysates as well as isolated inner and outer membrane fractions were separated by SDS-PAGE and analyzed by Western blots. For both species tested, our results were consistent with localization of the protease at the outer membrane, as confirmed by the immunodetection of rickettsial OmpB, which was used as an internal marker for the outer membrane in these assays (Figure 7C). We further confirmed the presence of APRc on the surface of intact R. conorii by flow cytometry analysis and also showed that the enzyme's catalytic domain is presented to the extracellular milieu (Figure 7D). Together, these results further illustrate that a novel retropepsin-like enzyme is expressed in two pathogenic rickettsial species and that the APRc catalytic domain is oriented towards the extracellular environment when present at the outer membrane of these bacteria.
The evidence that a proportion of APRc is associated with the outer membrane led us to hypothesize that rickettsial surface proteins might be potential substrates for this newly characterized enzyme. As has been shown for other autotransporter proteins, rickettsial OmpB, OmpA, Sca1, and Sca2 are involved in mediating important interactions with mammalian cells and undergo processing events at the outer membrane [43]–[49]. As an example, R. conorii OmpB (rOmpB) is expressed as a preprotein of 168 kDa and is subsequently cleaved to release the passenger domain (120 kDa) from the β-barrel translocation domain (32 kDa) [46]. Interestingly, R. conorii and R. japonica OmpB do not undergo proteolytic cleavage when expressed at the outer membrane of E. coli, suggesting that the processing event is not autocatalytic [43]. However, the identity of the enzyme responsible for Sca protein maturation still remains elusive. Therefore, and based on the observed APRc outer membrane localization, we sought to determine whether APRc might participate in the processing of rOmpB (Figure 8A). In order to do this, we performed transactivation assays using E. coli outer membrane fractions enriched in recombinant rOmpB (C-terminally His-tagged) and purified active APRc (soluble catalytic domain). Reaction products were then separated by SDS-PAGE and analyzed by Western blot. As shown in Figure 8B, the detection of an anti-His immune reactive product with ∼35 kDa in the presence of APRc was correlated with the disappearance of rOmpB preprotein, suggesting that this enzyme may be indeed capable of promoting cleavage of recombinant rOmpB. Moreover, the generated reactive protein product has approximately the same molecular weight as that expected for rOmpB β-barrel (32 kDa), further suggesting that this proteolytic cleavage may likely be occurring somewhere between the passenger and the β-barrel domains, in agreement with what has been described for native rOmpB [46]. To further validate these results, parallel assays were performed in the presence of APRc active site mutant and the integrity of rOmpB proprotein evaluated by immunoblotting with a specific antibody to this outer membrane protein. As expected, the disappearance of rOmpB proprotein was observed in the presence of active APRc but not when the cell extract was incubated with the active site mutant protein (APRc(D140A)99–231). Interestingly, we observed a similar phenomenon using as a substrate another conserved rickettsial antigen, Sca0/OmpA, demonstrating that a protein other than OmpB can be processed by APRc in vitro (Figure 9). Altogether, these results suggest that APRc is sufficient to mediate rOmpB maturation and rOmpA maturation in vitro, thereby raising an exciting hypothesis regarding possible functional significance of APRc as being able to process these and possibly other autotransporter proteins in the context of intact R. conorii cells.
The intrinsic difficulty in working with obligate intracellular parasites such as rickettsiae greatly hampers the correlation of rickettsial gene products with their function. Therefore, valuable information on the nature of conserved genes as well as on the identification of new bacterial factors that may play a role in rickettsiae pathogenesis is mostly being provided by comparative genomics. Using this approach, we identified a gene encoding a putative membrane embedded aspartic protease with a retroviral-type signature, highly conserved in 55 Rickettsia genomes. Using the R. conorii gene homologue RC1339 as our working model we demonstrate that the gene product (APRc) displays a high degree of identity among Rickettsia spp., although no significant homology is observed when compared to other aspartic proteases, except for the conservation of the motif around the catalytic aspartate as well as the hydrophobic-hydrophobic-glycine motif required for the formation of the psi loop. These features resemble the retroviral APs comprising family A2, which are characterized by being active only as symmetric dimers with a single active site, where each monomer contributes one aspartate [25], [27]. Despite the observed low overall sequence similarity with retropepsins, our results on the enzymatic characterization of the soluble catalytic domain of RC1339/APRc further revealed that this novel rickettsial enzyme indeed shares several properties with this family of APs. The common properties include autolytic activity impaired by mutation of the catalytic aspartate, accumulation in the dimeric form, optimal activity at pH 6, inhibition by specific HIV-1 PR inhibitors, and specificity preferences resembling those of both AP families. The presence of retroviral-type APs in bacteria has been previously demonstrated (SpoIIGA in Bacillus subtilis [33] and PerP in Caulobacter crescentus [50]). However, no enzymatic characterization is yet available for these enzymes and their inclusion as retropepsin-type protease members has not been universally accepted [29]. Therefore, the results described here provide experimental substantiation that RC1339/APRc is a novel retropepsin-like enzyme expressed in bacteria.
Most viral retropepsins are strictly required for the processing of Gag and Gag-Pol polyproteins into mature structural and functional proteins (including themselves) and are, therefore, indispensable for viral maturation [51]. Because of this, retropepsin-type APs are generally characterized by their inherent autolytic function. Interestingly, our results with APRc soluble catalytic domain fused to GST also demonstrated the ability of this protein to undergo a multi-step autocatalytic conversion in vitro into APRc105–231 mature form, and this autolytic activity was again confirmed when the last intermediate of activation was produced in E. coli. As expected for a retropepsin-like enzyme, mutation of the catalytic aspartate impaired this process. The enzymatic activity assays performed during these autoactivation studies (either using oxidized insulin B chain or the fluorogenic peptide mimicking the final cleavage site between the Ser104-Tyr105 residues) clearly indicated that APRc activity appears to be dependent on the presence of the final activation product. These results suggest that the processing at the N terminus must be the determining step for the regulation of enzymatic activity, presumably through a conformational change occurring upon conversion from rAPRc99–231 to APRc105–231 form. This is in line with what has been already described for recombinant HIV-1 PR, where the increase in catalytic activity upon protease autolytic conversion has been correlated with a conformational rearrangement between the precursor/inactive vs. mature/active forms of the enzyme [51], [52]. However, further studies are required to better understand the maturation of APRc precursor forms in vitro and how this is accomplished and controlled in vivo. In fact, we have shown that APRc accumulates in the outer membrane in R. conorii and R. rickettsii and, therefore, we cannot rule out that the presence of the transmembrane domain may play an important role in this maturation process in vivo.
Another interesting observation was that APRc autolytic activity, as well as cleavage of the fluorogenic substrate, occurred at a pH optimum of 6.0. This is again in good agreement with the optimal pH of other retropepsin-like [37], [38] enzymes as well as of the pepsin-like renin [53], [54] and, actually, it is consistent with the presence in APRc of an alanine residue downstream from the catalytic motif (Asp-Thr-Gly-Ala), instead of the common threonine residue found in most pepsin-like APs [27]. Together with the observed inhibitory effect of specific HIV-1 PR inhibitors, these results strengthen the striking resemblance between the enzymatic properties of APRc and those of viral retropepsins. Unexpectedly, we observed a drastic inhibitory effect of EDTA on both APRc maturation and hydrolysis of the fluorogenic substrate, suggesting that this protease may depend on a metal ion for folding and/or activity. A similar effect has not been reported for other retropepsins and no homology to a metalloprotease consensus motif was identified in APRc that could justify this inhibition. Therefore, further structural studies will be required to help in understanding this effect.
To provide additional evidence on the nature of APRc as a retropepsin-like AP we determined both the prime and nonprime side specificity using PICS [39]. Although HIV-1 PR is the only AP for which a PICS analysis has been reported [40], there are several studies for many different APs on specificity towards individual substrates (compiled, at least partially, by MEROPS [55]) providing a collection of cleavage patterns for these enzymes. A comparison of the substrate specificity of APs with our PICS results confirmed common preferences between APRc and both retropepsin and pepsin-type APs. The amino acid preference of APRc for P1 position is in good agreement with the canonical specificity of APs for large hydrophobic amino acids, such as phenylalanine, methionine, carboxyamidomethylated cysteine (which results from the modification during generation of peptide libraries), or leucine. Despite the observed lower selectivity, a similar trend for accommodating hydrophobic amino acids is also observed in P1′. As observed in both tryptic and GluC libraries, APRc appears to display broader specificity for P1 and P1′, while a more constrained amino acid preference is observed for P3, P2, and P2′ positions. This observation may account for an important role on substrate recognition and binding to the active pocket site and may ultimately influence hydrolytic efficiency. Strikingly, a high degree of similarity is found with more specialized pepsin-like proteases such as BACE for P3 (with a preference for valine and isoleucine) and P2′ (alanine and valine) positions, as well as with cathepsin D (also for P2′). Interestingly, APRc also displays unique amino acid preferences such as proline at P2 (although the preference observed for valine and threonine in this position has also been described for feline immunodeficiency virus retropepsin [55], [56]), and leucine and isoleucine in P4′ position. When compared with the two major types of cleavage sites proposed for HIV-1 PR and other retropepsins, APRc specificity profile suggests a preference for type 2-like substrates with hydrophobic amino acids in P1 and P1′, whereas type 1-like substrates with the typical combination of tyrosine(phenylalanine)-proline at P1-P1′ appear disfavored [51], [57]. Moreover, our results suggest a cooperative effect between P2 and P1 positions by revealing that a P2 proline co-occurs more frequently with P1 methionine or asparagine residues and that proline is not favored at this position when P1 is occupied by phenylalanine. Curiously, APRc autolytic cleavage sites do not perfectly match the observed specificity preferences of the activated form used in PICS, suggesting either a different conformational arrangement of the protease or a dependence on the sequence context and/or conformation of the substrate. This is not totally unexpected, as for HIV-1 PR it has also been reported that specificity towards nonviral protein substrates significantly differed from viral polyprotein cleavage sites [51].
Aspartic proteases were assumed for a long time to be restricted to viruses and eukaryotes. However, more recently proteins bearing the characteristic hallmark features of the pepsin family have been identified in seven bacterial genomes [29] and the detailed biochemical characterization of the pepsin-like homologue from the Shewanella amazonensis, shewasin A [58], has clearly demonstrated that this bacterial AP is strongly reminiscent of its eukaryotic counterparts. These observations have raised a discussion on the evolutionary relationships between bacterial and eukaryotic pepsin-like APs, by suggesting that bi-lobal pepsin-like proteases may have evolved from primordial homodimeric aspartic proteases before divergence between eukaryotes and prokaryotes (through the proposed gene duplication and fusion event [59]). Our current results on RC1339/APRc further support this hypothesis by providing the first experimental evidence that a gene for a single-lobed AP is indeed present in prokaryotes, coding for an active enzyme with properties resembling those of retropepsins. Moreover, these results offer additional clues on the relationships between retropepsin-like and pepsin-like APs. The presence of single-lobed AP genes in prokaryotes suggests that enzymes such as APRc may actually represent the most ancestral forms of these proteases, whereas retroviral retropepsins may instead correspond to a derived state.
Besides demonstrating that RC1339 encodes an active enzyme, we have also shown that this rickettsial protease is expressed in both R. conorii and R. rickettsii. Unlike the large majority of α-helical type of integral membrane proteins, sub-cellular localization studies revealed an outer membrane accumulation for APRc which was also confirmed by expression of the full-length protease in E. coli and in the context of intact R. conorii. So far, only three transmembrane proteins with α-helical architecture have been reported to be embedded in the outer membrane of gram-negative bacteria [60]–[62]. Therefore, our results provide additional evidence that the bacterial surface is not restricted to proteins with β-barrel structures known to play essential roles in energetics, metabolism, signal transduction, and transport [63], [64], further suggesting that the repertoire of proteins with α-helices localized to the OM may be higher than anticipated. Nevertheless, transport and insertion of APRc into the OM definitely requires further studies in order to clarify whether the detected 21 kDa band is an intermediate processed form or the result of different gel mobility.
In line with our evidence for the native expression of APRc in R. conorii and R. rickettsii, a multiomics study performed in Rickettsia prowazekii to identify potential virulence factors has also confirmed transcription of RC1339 gene homologue (RP867) [17]. Importantly, these studies also showed differential regulation of RP867 expression in different strains of R. prowazekii with a fold change of 1.77 between the virulent strain Rp22 and the avirulent strain Erus. This evidence for an up-regulation of APRc's gene expression in R. prowazekii Rp22, combined with our results confirming protease expression and accumulation into the OM in R. conorii and R. rickettsii, strongly support a relevant role of this highly conserved protease in rickettsial life cycle. Serine-, cysteine-, and metalloproteases are widely spread in many pathogenic bacteria, where they play critical functions related to pathogenesis and virulence [19], [20]. However, much less is known about the role of aspartic proteases since the presence of this class of enzymes in pathogenic bacteria has not been previously reported. Taking under consideration the unique biochemical and enzymatic features of APRc presented in this work: i) the apparent non-stringent sequence requirement; ii) outer membrane localization and extracellular orientation of recombinant APRc catalytic domain and iii) autolytic activity suggesting that the soluble biological unit may be released from the surface of rickettsial cells by an ectodomain shedding-like process, we anticipate a potential multi-functional role for this rickettsial protease. One of the proposed functions concerns APRc contribution for the degradation of host tissues for supplying bacteria with nutrients, similar to that described for other extracellular proteases secreted by many pathogens [19]. Second, this protease may also support the spread of the infection and dissemination of bacteria into deeper tissue through the shedding of cell surface adhesion molecules or the inactivation of the components of the host immune system such as proteins from the complement system [21], [65], [66]. And third, APRc may participate in the degradation and/or maturation of other rickettsial proteins, in particular those located at the OM, such as Sca proteins [45], [67], exemplified by Sca5/rOmpB and Sca0/OmpA. In contrast to other autotransporter proteins from gram-negative bacteria with auto-proteolytic activity such as SPATEs (Serine Protease Autotransporters) [68], rOmpB processing is thought to implicate a protease as previous expression studies in E. coli have failed to demonstrate autocatalytic activity [43], [47]. In this work, we have started addressing this last hypothesis and we showed that APRc is indeed sufficient to catalyze the processing of Sca0/OmpA and Sca5/rOmpB in vitro and that, for the latter, the generated product is consistent with the cleavage between the passenger and the β-peptide regions. The N-terminal sequence of the β-peptide has been experimentally determined for R. typhi and R. prowazekii [46] rOmpB and the region spanning the cleavage site (/) corresponds to the sequence Ala-Ala-Val-Ala-Ala/Gly-Asp-Glu-Ala-Val. Although we cannot exclude that in R. conorii the cleavage of rOmpB may occur slightly upstream from this region, if considering a similar cleavage site the amino acids present in P4, P3, P1, P1′ and P4′ are in good agreement with the observed specificity preferences for APRc, while the differences observed for the remaining positions may reflect again the importance of sequence context/substrate conformation for APRc cleavage. Nevertheless, additional experiments are required to determine the cleavage site and its relevance in the context of intact rickettsiae as well as APRc role in the degradation of other rickettsial and/or host proteins.
In summary, the findings described herein show that this newly characterized aspartic protease from Rickettsia is an active enzyme with features highly reminiscent of retropepsin-type proteases and we anticipate its participation in a relevant proteolytic pathway in rickettsial life-cycle, likely as a modulator of activity of other rickettsial membrane-localized proteins. Determination of APRc three-dimensional structure and dissection of its contribution to rickettsial pathogenesis will be critical to start unveiling the significance of this novel protease as a potential target for therapeutic intervention.
With this work we expect to contribute to start changing the currently accepted evolutionary paradigm of aspartic proteases, by positioning what we denominate as “prokaryopepsins” as the new archetypes of modern APs.
Oligonucleotide primers were purchased from Integrated DNA Technologies, Leuven, Belgium. Synthetic genes encoding the full-length RC1339 and the predicted soluble catalytic domain, the fluorogenic peptide PepRick14 (MCA-Lys-Ala-Leu-Ile-Pro-Ser-Tyr-Lys-Trp-Ser-Lys-DNP) and the rabbit polyclonal antibody raised towards the sequence Cys-Tyr-Thr-Arg-Thr-Tyr-Leu-Thr-Ala-Asn-Gly-Glu-Asn-Lys-Ala (anti-APRc) were produced by GenScript (Piscataway, NJ, USA). N-terminal amino acid sequence analyses were performed in the Analytical Services Unit - Protein Sequencing Service, ITQB (Oeiras, Portugal). Rabbit polyclonal antibody against the purified APRc99–231His construct (see below) was generated by standard immunization schemes approved by the LSU School of Veterinary Medicine Institutional Animal Care and Use Committee (IACUC).
Gene and protein sequences for R. conorii str. Malish 7 RC1339 were obtained from the genome sequence at NCBI (NC_003103) (AAL03877). Amino acid sequence alignment and the degree of identity between RC1339/APRc homologues from Rickettsia (genus) (TaxID 780) were obtained with ClustalW [69], by comparing the sequences deposited in NCBI database with the following accession numbers: NP_360976 (R. conorii str. Malish 7), YP_005393543 (R. parkeri str. Portsmouth), YP_001495413 (R. rickettsii str. Sheila Smith), YP_005364747 (R. amblyomii str. GAT-30V), YP_005391701 (R. montanensis str. OSU 85–930), NP_221215 (R. prowazekii str. Madrid E), YP_067793 (R. typhi str. Wilmington), YP_247382 (R. felis URRWXCal2) and YP_001495500 (R. bellii OSU 85–389). The protein family, domain, and functional sites were searched using the InterProScan program [70]. Topology structure was predicted with HMMTOP2 algorithm [30]. A structure-based alignment of RC1339/APRc soluble catalytic domain with HIV-1 (PDB 3hvp), EIAV (PDB 2fmb), and XMRV (PDB 3nr6) retropepsins and with DdI1 putative protease domain (PDB 2i1a) was performed with PROMALS3D [71].
The sequences encoding the full-length and the predicted soluble domain of RC1339/APRc were chemically synthetized with OptimumGene codon optimization technology for E. coli codon usage and cloned into pUC57 vector. The gene encoding the full-length APRc (construct coding amino acids 1–231) was then amplified to include restriction sites for NcoI and NotI at 5′- and 3′-ends, respectively, using the forward primer 5′-CCATGGGAATGAACAAAAAACTGATCAAACTG-3′ and the reverse primer 5′-CTCGAGATAATTCAGAATCAGCAGATCTTT-3′; the resulting PCR product was cloned into pGEM-T Easy plasmid (Promega). After digestion with NcoI and NotI, APRc1–231 insert was subcloned into pET28a expression vector (Invitrogen) in frame with a C-terminal His-tag (pET-APRc1–231His). In order to generate the untagged construct, an insertion mutagenesis was performed to include the TGA stop codon at the end of the full-length sequence using the Quick Change site-directed mutagenesis kit (Stratagene) and the primers 5′-ATTCTGAATTATTGACTCGAGCACCAC-3′ (forward) and 5′-GTGGTGCTCGAGTCAATAATTCAGAAT-3′ (reverse) (pET-APRc1–231).
The optimized sequence encoding the predicted soluble catalytic domain of APRc (construct coding amino acids 87–231) flanked by restriction sites for BamHI (5′)/EcoRI (3′) was inserted in frame to the C terminus of GST in pGEX-4T2 expression vector (Amersham) using the same pair of restriction enzymes (pGST-APRc87–231).
For generating the expression construct bearing the sequence encoding the intermediate activation form APRc99–231 (construct coding amino acids 99–231), the sequence was firstly amplified using the construct pETAPRc1–231 as the template and the forward primer containing a NdeI restriction site (5′-CATATGTATAAATGGAGTACCGAAGTT-3′) and the same reverse primer used for amplification of APRc1–231 (5′-CTCGAGATAATTCAGAATCAGCAGATCTTT-3′), and cloned into pGEM-T Easy (Promega). The insert was then digested with NdeI/NotI and subcloned into pET23a expression vector (Invitrogen) in frame with a C-terminal His-tag (pETAPRc99–231His).
The active site mutant of APRc (both in pGEX4T2 and pET23a constructs) was generated by replacing the putative active site aspartic acid residue by alanine (D140A) using the Quick Change site-directed mutagenesis kit (Stratagene) and the primers 5′-AAAATCAAATTCATGGTGAATACCGGCGCCTCTGATATTGCA-3′ (forward) and 5′-TGCAATATCAGAGGCGCCGGTATTCACCATGAATTTGATTTT-3′ (reverse) (mutation underlined). All positive clones were selected by restriction analysis and confirmed by DNA sequencing.
The construct for expression of rOmpB in E. coli was generated as described elsewhere [43].
rGST-APRc87–231 and the corresponding active site mutant protein were expressed by standard procedures. Briefly, E. coli BL21 Star (DE3) cells transformed with each plasmid construct, pGST-APRc87–231 and pGST-APRc87–231(D140A), were grown at 37°C until an OD600 nm of 0.7. Protein expression was then induced with 0.1 mM IPTG for 3 hours, after which cells were harvested by centrifugation at 9000 g for 20 minutes at 4°C, and resuspended in PBS buffer. Lysozyme (100 µg/ml) was added and the harvested cells were frozen at −20°C. After freezing and thawing, bacterial cell lysates were incubated with DNase (1 µg/ml) and MgCl2 (5 mM) for 1 hour at 4°C. The total cell lysate was then centrifuged at 27216 g for 20 minutes at 4°C and the resulting supernatant filtered (0.2 µm) before loading onto a GSTrap HP 5 ml column (GE Healthcare Life Sciences) previously equilibrated in PBS buffer. After extensive washing, the protein of interest was eluted in 50 mM Tris-HCl pH 8 with 10 mM glutathione and immediately loaded onto a Superdex 200 HiLoad 26/60 (GE Healthcare Life Sciences) equilibrated in PBS buffer for further purification and glutathione removal.
Expression of E. coli BL21 Star (DE3) cells transformed with pETAPRc99–231His plasmid as well as isolation of total soluble protein were performed under the same conditions as described for rGST-APRc87–231, except that in this case the cell pellet was resupended in 20 mM phosphate buffer pH 7.5, 500 mM NaCl and 10 mM imidazole. The resultant supernatant was then loaded onto a Histrap HP 5 ml column (GE Healthcare Life Sciences) pre-equilibrated in the same buffer. Protein elution was performed by a three-step gradient of imidazole (50 mM, 100 mM and 500 mM) and fractions containing the protein of interest (100 mM imidazole gradient step) were pooled and buffer exchanged into 20 mM phosphate buffer pH 7.5 by an overnight dialysis step. Dialyzed protein was further purified by cation-exchange chromatography with a MonoS column (GE Healthcare Life Sciences) equilibrated in the same buffer and elution was carried out by a linear gradient of NaCl (0–1 M).
Time-course studies of APRc activation were undertaken with two recombinant forms of the soluble catalytic domain of APRc (rGST-APRc87–231 and rAPRc99–231). Purified samples of rAPRc were first diluted to 0.1 mg/mL with PBS and then diluted 1∶1 with 0.1 M sodium acetate buffer pH 6. Diluted samples were incubated up to 48 h at 37°C and aliquots were taken every 12 h for SDS-PAGE analysis and proteolytic activity assays. To evaluate the effect of inhibitors on rAPRc auto-activation processing, a time-course analysis was carried out in the presence of 20 µM pepstatin, 1 mM indinavir or 5 mM EDTA and protein samples were analyzed by SDS-PAGE.
Crosslinking reactions with disuccinimidyl suberate (DSS) (Pierce) were performed in 20 mM phosphate buffer pH 7.5 containing 150 mM NaCl. A solution of 0.2 mg/ml of purified rAPRc (rAPRc99–231 and activated product rAPRc105–231) was treated with a 50-fold molar excess of DSS in a total volume of 50 µl and allowed to react for 30 min at room temperature. For glutaraldehyde treatment, a solution of 0.5 mg/ml of purified rAPRc99–231 was treated with 5 µl of 1.15% freshly prepared solution of glutaraldehyde for 4 minutes at 37°C, in a total volume of 50 µl, under similar buffer conditions. To terminate the reactions, 5 µl of the quenching buffer 1 M Tris-HCl pH 8.0 were added. Crosslinked proteins were separated by SDS-PAGE and analyzed by Western blot with anti-APRc antibody.
Precursor rAPRc99–231 and activated rAPRc110–231 forms were analyzed under nondenaturing conditions by analytical size-exclusion chromatography on a Superdex 200 5/150 GL (GE Healthcare Life Sciences) column connected to a Prominence HPLC system (Shimadzu Corporation, Tokyo, Japan). The column was equilibrated in 20 mM phosphate buffer pH 7.5 containing 150 mM NaCl, and calibrated with Gel Filtration LMW and HMW calibration kits (GE Healthcare Life Sciences), according to the manufacturer's instructions. The molecular mass markers used for calibration were conalbumin (75 kDa), ovalbumin (43 kDa), carbonic anhydrase (29 kDa), and ribonuclease A (13.7 kDa).
The effect of pH on activity and inhibitory profile of purified recombinant rAPRc (rAPRc105-231) was determined by fluorescence assays in 96-well plates in in a Gemini EM Fluorescence Microplate Reader, using the fluorogenic substrate PepRick14 (MCA-Lys-Ala-Leu-Ile-Pro-Ser-Tyr-Lys-Trp-Ser-Lys-DNP) (final concentration of 2.5 µM). For determination of the pH profile, rAPRc105-231 was assayed for activity at 37°C in buffers ranging between pH 4 and 9 (50 mM sodium acetate pH 4.0, 5.0, 5.5 and 6.0; 50 mM Tris-HCl pH 7.0, 8.0 and 9.0) containing 100 mM NaCl. To test the effect of classical inhibitors, the protease was pre-incubated in the presence of each inhibitor, 20 µM pepstatin, 5 mM EDTA, 1 mM ZnCl2, 1 mM Pefabloc, or 10 µM E-64, for 10 minutes at room temperature in 50 mM sodium acetate pH 6.0 containing 100 mM NaCl before determination of proteolytic activity. The effect of the HIV inhibitors on rAPRc proteolytic activity was also evaluated. The following reagents were obtained through the NIH AIDS Research and Reference Reagent Program, Division of AIDS, NIAID, NIH: indinavir sulfate, nelfinavir, ritonavir, saquinavir, amprenavir, atazanavir sulfate, sarunavir, and lopinavir. Each inhibitor was again incubated with rAPRc for 10 minutes at room temperature in 50 mM sodium acetate pH 6.0 containing 100 mM NaCl and 5% DMSO, except for indinavir and darunavir that were assayed without DMSO. Indinavir, nelfinavir, ritonavir, saquinavir, amprenavir, atazanavir and lopinavir were tested in the range of 0.25 mM–1 mM and the inhibitor darunavir in the range of 2.5 µM–10 µM (final concentration). The rate of substrate hydrolysis was monitored for 3 hours by the increase in fluorescence intensity with excitation/emission wavelengths of 328/393 nm and the relative activity normalized by setting rAPRc activity as 100%.
Evaluation of proteolytic activity during auto-processing time-course analysis was performed towards oxidized insulin β chain. Substrate (1 mg/ml) was incubated with purified recombinant APRc enzyme: substrate mass ratio of 1∶15) in 0.1 mM sodium acetate buffer pH 6.0. After an overnight incubation at 37°C the reaction mixture was centrifuged at 20000 g during 6 min and the digestion fragments were separated by RP-HPLC on a C18 column (KROMASIL 100 C18 250, 4.6 mm), using a Prominence system (Shimadzu Corporation, Tokyo, Japan). Elution was carried out with a linear gradient (0–80%) of acetonitrile in 0.1% v/v trifluoroacetic acid for 30 min at a flow rate of 1 ml/min. Absorbance was monitored at 220 nm.
rAPRc (APRc105-231) specificity profiling was determined according to the Proteomic Identification of Protease Cleavage Sites (PICS) methodology as described elsewhere [40], with minor changes. Tryptic and GluC peptide libraries were generated from THP1 cells. rAPRc-to-library ratios were 1∶50 (wt/wt) and incubation was performed at 37°C for 16 h in 50 mM sodium acetate buffer pH 6.0, 150 mM NaCl. The separation of carboxy-peptide cleavage products was carried out on a column packed with PepMap C18 resin (Dionex) connected to a LC-MS/MS QSTAR Pulsar (AB SCIEX) operated by the UBC Michael Smith Laboratory/Laboratory for Molecular Biophysics Proteomics Core Facility, and on HALO C18 column (Eksigent) connected to a LC-MS/MS TripleTOF 5600, AB SCIEX (Center for Neuroscience and Cell Biology Proteomics Unit). Peptide sequences were identified with both Mascot [72] and X!Tandem [73], in conjunction with PeptideProphet [74] at a confidence level >95%. Mass tolerance was 10 ppm for parent ions and 0.6 Da for fragment ions. Search parameters were set to identify static modifications as carboxyamidomethylation of cysteine residues (+57.02 Da), dimethylation of lysines (+28.03 Da) and thioacylation of peptide amino termini (+88.00 Da). Semi-style cleavage searches were applied with no constraints for the orientation of the specific terminus. The Web-based PICS service [40] was used to derive nonprime sequences. Sequence logos were generated with IceLogo with a p-value of 5% [75].
For cDNA synthesis, total RNA was isolated from an aliquot of frozen R. conorii Malish 7 and R. rickettsii “Sheila Smith”-infected Vero cells using the SurePrep TrueTotal RNA Purification Kit (Fisher Scientific), according to the manufacturer's instructions. After extraction, RNA samples were treated with DNase I, RNase-free set (Thermo Scientific) for 30 minutes at 37°C. The reaction was inactivated by adding 50 mM EDTA and heating the mixture at 65°C for 10 minutes. Next, approximately 1 ug of total RNA was used as the template for reverse transcription using the iScript cDNA Synthesis Kit (BioRad), according to the manufacturer's instructions. For all extracted samples, negative RT-PCR controls were processed in the absence of reverse transcriptase. APRc gene expression was assessed by PCR reaction with the specific primers RC1339_RT-Fwd (5′-AAAGCCGCCCCTATAACCTT-3′) and RC1339_RT-Rev (5′-TCCTGAAACCTTTGAAACGCTC-3′) which were designed for the amplification of a segment with 136 bp. The PCRs were performed in a 50 µl volume, with 1 µl of cDNA as the DNA template, 0.1 µM of each primer, 1× PCR buffer (100 mM Tris-HCl (pH 9.0), 15 mM MgCl2, 500 mM KCl), 200 µM of dNTP mix, and 1 U of Taq DNA polymerase (GE Healthcare). The PCR mixtures were incubated at 95°C for 3 min, followed by 35 cycles of 95°C (30 sec), 55°C (30 sec), and 72°C (30 sec). The gene hrtA (17 kDa surface antigen) was used as the positive control using the primers Rc_htrA_RT-Fwd (5′-GGACAGCTTGTTGGAGTAGG-3′) and Rc_htrA_RT-Rev (5′-TCCGGATTACGCCATTCTAC-3′). An aliquot of 20 µl of each PCR product was electrophoresed on a 1.7% agarose gel and stained with ethidium bromide. The size of the PCR product was determined by comparison with GeneRuler 1 kb Plus DNA Ladder (Thermo Scientific).
Cell fractionation studies with Rickettsia spp. were performed as previously described [76]. Briefly, approximately 5×106 plaque forming units (pfu) of purified R. conorii Malish 7 or R. rickettsii “Sheila Smith” was fixed in 4% paraformaldehyde (PFA) in PBS, washed in PBS and then removed from BSL3 containment after verification that viable rickettsiae were no longer present according to standard operating procedures (SOPs). For whole-cell lysates, cells were resuspended in SDS-PAGE loading buffer and boiled. Total outer membrane proteins were extracted essentially as described in [77]. The sample was resuspended in 1.5 ml of 20 mM Tris pH 8.0 containing 1× protease inhibitor cocktail and then subjected to three rounds of French press treatment for cell lysis. The resulting lysate was centrifuged at 10,000 g for 3 minutes to remove unbroken cells and then incubated in 0.5% sarkosyl at room temperature for 5 minutes. The sarkosyl-treated lysate was centrifuged at >16,000 g for 30 minutes at 40°C. The sarkosyl soluble protein fraction was removed and the remaining insoluble pellet representing the outer-membrane protein fraction was washed in 20 mM Tris pH 8.0 and then boiled in 0.5 ml of 20 mM Tris pH 8.0 containing 0.5 ml of 2× SDS sample buffer. Protein samples were aliquoted and frozen at −20°C until use.
For isolation of total and outer membrane fractions of E. coli BL21 Star (DE3) cells expressing full-length rAPRc1–231 the protocol used was essentially as described in [77]. BL21 (DE3) cells were transformed with pETAPRc1–231 construct and grown at 37°C until an OD600 nm of 0.6–0.7. Expression of rAPRc1–231 was induced with 0.1 mM IPTG for 3 hours, after which cells were pelleted by centrifugation at 9000 g for 20 minutes at 4°C and resuspended in PBS buffer. Cells were then mechanically disrupted on a FrenchPress following the manufacturer's instructions (3×, 1500 psi), and cleared by centrifugation at 20000 g for 20 minutes. Total membrane were directly pelleted by ultracentrifugation at 144028 g for 1 hour at 4°C and resuspended in PBS buffer. For enrichment and purification of OM proteins, inner membrane proteins were extracted by incubating the supernatant of lysate clearance with sarkosyl (final concentration of 0.5%) at room temperature for 5 minutes. Outer membranes were pelleted by ultracentrifugation at 144028 g for 1 hour at 4°C and resuspended in PBS buffer. Total membrane, soluble/sarkosyl-solubilized and outer membrane proteins were resolved by SDS-PAGE and analyzed by immunoblotting using antibodies against APRc, Lep and OmpA, the last two used as internal markers for inner and outer membranes of E. coli, respectively.
E. coli BL21 (DE3) were transformed with pETAPRc1–231 construct and grown at 37°C until an OD600 nm of 0.6–0.7. Protein expression was induced with 0.1 mM IPTG for 3 hours. Cells were then fixed for 20 minutes in 4% PFA and subsequently washed in cold PBS. Fixed cells were incubated with anti-APRc rabbit polyclonal (2 µg/ml) and anti-RNAPα mouse monoclonal antibodies (mAb 4RA2, 50 ng/ml), and then labeled with both goat anti-rabbit IgG Alexa Fluor 488 (Life Technologies) and goat anti-mouse IgG R-PE-Cy5.5 conjugated secondary antibodies (SouthernBiotech) at the specified concentration (4 µg/ml). Bacteria were analyzed by flow cytometry using a BD FACSCalibur (BD Biosciences) instrument and FlowJo software. For analysis of non-permeable E. coli cells, positive anti-RNAPα staining cells were gated out, and intact cells analyzed for surface expression of APRc1–231 with anti-APRc antibody. Detection of APRc and OmpB on the surface of PFA-fixed, intact R. conorii cells was determined using rabbit polyclonal anti-APRc99–231 antibody, anti-OmpB mAb 5C7.27 [44] and the appropriate AlexaFluor 488 conjugated secondary antibody as described above for the E. coli samples.
The expression of rOmpB and rOmpA independently in E. coli BL21 (DE3) cells was performed as previously described [43], [78]. To assess for in vitro proteolytic cleavage of this outer membrane protein by rAPRc, the total membrane fraction of E. coli cells overexpressing rOmpB and rOmpA was isolated as described under the section R. conorii and E. coli fractionation and then incubated for 16 hours at 37°C in 50 mM sodium acetate pH 4.0, 100 mM NaCl with 25 µg of purified active rAPRc105–231. A parallel incubation was performed under similar conditions with the active site mutant rAPRc(D140A)99–231 as a negative control. The reaction products were separated by SDS-PAGE and analyzed by Western blot with anti-APRc, anti-rOmpB35–1334 and anti-rOmpA [78] rabbit polyclonal antibodies.
SDS-PAGE analysis was performed in a Bio-Rad Mini Protean III electrophoresis apparatus using 4–20% or 12.5% polyacrylamide gels. Samples were treated with loading buffer (0.35 M Tris-HCl, 0.28% SDS buffer pH 6.8, 30% glycerol, 10% SDS, 0.6 M DTT and 0.012% Bromophenol Blue) and boiled for 5 minutes before loading. Gels were stained with Coomassie Brilliant Blue R-250 (Sigma). For Western blot analysis, protein samples were resolved by SDS-PAGE and electrotransferred onto PVDF or nitrocellulose membranes by standard wet (using the buffer 25 mM Tris, 192 mM glycine and 20% methanol) or semi-dry (in buffer 25 mM Tris, 192 mM glycine, 20% methanol and 0.025% SDS) transfer apparatus. Membranes were then blocked for one hour in standard TBS containing 1% (v/v) Tween-20 supplemented with 5% (w/v) skim milk or 2% (w/v) BSA and then incubated with the antibodies, anti-APRc rabbit polyclonal (GenScript, 2 µl/ml), anti-rOmpB35–1334 and anti-rOmpA [78] rabbit polyclonal, anti-OmpA (E. coli) serum (1∶20000), anti-Lep (E. coli) serum (1∶1000). Anti-OmpA and anti-Lep antibodies were kindly provided by Professor Gunnar von Heijne (Stockholm University, Sweden). Membranes were washed in TBS, containing 0.1% (v/v) Tween-20 and incubated with secondary anti-mouse or anti-rabbit alkaline phosphatase-conjugated (GE Healthcare) and IRDye-conjugated (LI-COR Biotechnology) antibodies and revealed using ECF chemiluminescence detection kit (GE Healthcare) in a Molecular Imager FX (Bio-Rad) or by infrared detection using an Odyssey infrared dual-laser scanning unit (LI-COR Biotechnology), respectively.
To confirm the specific band reactivity of anti-APRc antibody, a peptide competition assay was performed. The primary antibody was pre-incubated with 100-fold (mass) excess of immunizing peptide (CYTRTYLTANGENKA) for 20 minutes at room temperature prior immunoblotting analysis and parallel experiments were performed with pre-incubated and non-incubated antibody.
|
10.1371/journal.pcbi.1006110 | Bayesian comparison of explicit and implicit causal inference strategies in multisensory heading perception | The precision of multisensory perception improves when cues arising from the same cause are integrated, such as visual and vestibular heading cues for an observer moving through a stationary environment. In order to determine how the cues should be processed, the brain must infer the causal relationship underlying the multisensory cues. In heading perception, however, it is unclear whether observers follow the Bayesian strategy, a simpler non-Bayesian heuristic, or even perform causal inference at all. We developed an efficient and robust computational framework to perform Bayesian model comparison of causal inference strategies, which incorporates a number of alternative assumptions about the observers. With this framework, we investigated whether human observers’ performance in an explicit cause attribution and an implicit heading discrimination task can be modeled as a causal inference process. In the explicit causal inference task, all subjects accounted for cue disparity when reporting judgments of common cause, although not necessarily all in a Bayesian fashion. By contrast, but in agreement with previous findings, data from the heading discrimination task only could not rule out that several of the same observers were adopting a forced-fusion strategy, whereby cues are integrated regardless of disparity. Only when we combined evidence from both tasks we were able to rule out forced-fusion in the heading discrimination task. Crucially, findings were robust across a number of variants of models and analyses. Our results demonstrate that our proposed computational framework allows researchers to ask complex questions within a rigorous Bayesian framework that accounts for parameter and model uncertainty.
| As we interact with objects and people in the environment, we are constantly exposed to numerous sensory stimuli. For safe navigation and meaningful interaction with entities in the environment, our brain must determine if the sensory inputs arose from a common or different causes in order to determine whether they should be integrated into a unified percept. However, how our brain performs such a causal inference process is not well understood, partly due to the lack of computational tools that can address the complex repertoire of assumptions required for modeling human perception. We have developed a set of computational algorithms that characterize the causal inference process within a quantitative model based framework. We have tested the efficacy of our methods in predicting how human observers judge visual-vestibular heading. Specifically, our algorithms perform rigorous comparison of alternative models of causal inference that encompass a wide repertoire of assumptions observers may have about their internal noise or stimulus statistics. Importantly, our tools are widely applicable to modeling other processes that characterize perception.
| We constantly interact with people and objects around us. As a consequence, our brain receives information from multiple senses as well as multiple inputs from the same sense. Cues from the same sense (e.g., texture and disparity cues to an object shape) are generally congruent as they usually reflect identical properties of a common external entity. Thus, the brain eventually learns to mandatorily integrate inputs from the same modality as a unified percept, which provides more precise information than either cue alone [1, 2]. Similarly, integration of cues represented in different modalities but associated with a common stimulus also improves perceptual behavior. There is a wealth of evidence that demonstrates increased precision [3–12], greater accuracy [13, 14] and faster speed [15, 16] of perceptual performance due to multimodal integration.
However, multimodal cues present a complex problem. Cues from different modalities are not necessarily congruent as different stimuli can simultaneously impinge on our senses, giving rise to coincident yet conflicting information. For example, in a classic ventriloquist illusion, even though the sound originates from the puppeteer’s mouth, we perceive that it is the puppet which is talking [17]. Mandatory integration of multimodal cues arising from different stimuli can induce errors in perceptual estimates [6, 14]. Thus, for efficient interaction with the world, the brain must assess whether the multimodal cues originated from the same cause, and should be integrated into a single percept, or instead the cues should be interpreted in isolation as they arose from different causes (segregation). Despite the often overwhelming amount of sensory inputs, we are typically able to integrate relevant cues while ignoring irrelevant sensory input. It is thus plausible that our brain infers the causal relationship between multisensory cues to determine if and how the cues should be integrated.
Bayesian causal inference—inference of the causal relationship between observed cues, based on the inversion of the statistical model of the task—has been proposed as the decision strategy adopted by the brain to address the problem of integration vs. segregation of sensory cues [18, 19]. Such a decision strategy has described human performance in spatial localization [18–27], orientation judgment [28], oddity detection [29], speech perception [30], time-interval perception [31], simple perceptual organization [32], and heading perception [33, 34]. In recent years, interest in the Bayesian approach to causal inference has further increased as neural imaging has identified a hierarchy of brain areas involved in neural processing while observers implemented a Bayesian strategy to perform a causal inference task [20]. At the same time, Bayesian models have become more complex as they include more precise descriptions of the sensory noise [22, 33, 34] and alternative Bayesian decision strategies [21, 24]. However, it is still unknown whether observers fully implement Bayesian causal inference, or merely an approximation that does not take into account the full statistical structure of the task. For example, the Bayes-optimal inference strategy ought to incorporate sensory uncertainty into its decision rule. On the other hand, a suboptimal heuristic decision rule may disregard sensory uncertainty [32, 35, 36]. Thus, the growing complexity of models and the need to consider alternative hypotheses require an efficient computational framework to address these open questions while avoiding trappings such as overfitting or lack of model identifiability [37]. For a more detailed overview of open issues in multisensory perception and causal inference at the intersection of behavior, neurophysiology and computational modeling, we refer the reader to [38–40].
Visuo-vestibular integration in heading perception presents an ideal case to characterize the details of the causal inference strategy in multisensory perception. While a wealth of published studies have shown that integration of visual and vestibular self-motion cues increases perceptual precision [9–12, 14, 41–43], and accuracy [14], such an integration only makes sense if the two cues arise from the same cause—that is optic flow and inertial motion signal heading in the same direction. Despite the putative relevance of causal inference in heading perception, the inference strategies that characterize visuo-vestibular integration in the presence of sensory conflict remain poorly understood. For example, a recent study has found that observers predominantly integrated visual and vestibular cues even in the presence of large spatial discrepancies [33]—whereas a subsequent work has presented evidence in favor of causal inference [34]. Furthermore, these studies did not vary cue reliability—a manipulation that is critical to test whether a Bayes-optimal inference strategy or a suboptimal approximation was used [35].
Another aspect that can influence the choice of inference strategy is the type of inference performed by the observer. In particular, de Winkel and colleagues [33, 34] asked subjects to indicate the perceived direction of inertial heading—an ‘implicit’ causal inference task as subjects implicitly assessed the causal relationship between visual and vestibular cues on their way to indicate the final (integrated or segregated) heading percept. Even in the presence of spatial disparities as high as 90°, one study found that several subjects were best described by a model which fully integrated visual and vestibular cues [33] (possibly influenced by the experimental design; see also [34]). It is plausible that performing an explicit causal inference task, which forces subjects to indicate whether visual and vestibular cues arose from the same or different events, may elicit different inference strategies, as previously reported in category-based induction [44], multi-cue judgment [45], and sensorimotor decision-making [46]. While some studies have tested both explicit and implicit causal inference [18, 21, 47], to our knowledge only one previous study contemplated the possibility of different strategies between implicit and explicit causal inference tasks [21], and a systematic comparison of inference strategies in the two tasks has never been carried out within a larger computational framework.
Thus, the goal of this work is two-fold. First, we introduce a set of techniques to perform robust, efficient Bayesian factorial model comparison of a variety of Bayesian and non-Bayesian models of causal inference in multisensory perception. Factorial comparison is a way to simultaneously test different orthogonal hypotheses about the observers [21, 48–50]. Our approach is fully Bayesian in that we consider both parameter and model uncertainty, improving over previous analyses which used point estimates for the parameters and compared individual models. A full account of uncertainty in both parameter and model space, by marginalizing over parameters and model components, is particularly prudent when dealing with internal processes, such as decision strategies, which may have different latent explanations. An analysis that disregards such uncertainty might produce unwarranted conclusions about the internal processes that generated the observed behavior [37]. Second, we demonstrate our methods by quantitatively comparing the decision strategies underlying explicit and implicit causal inference in visuo-vestibular heading perception within the framework of Bayesian model comparison. We found that even though the study of explicit and implicit causal inference in isolation might suggest different inference rules, a joint analysis that combines all available evidence points to no difference between tasks, with subjects performing some form of causal inference in both the explicit and implicit tasks that used identical experimental setups.
In sum, we demonstrate how state-of-the-art techniques for model building, fitting, and comparison, combined with advanced analysis tools, allow us to ask nuanced questions about the observer’s decision strategies in causal inference. Importantly, these methods come with a number of diagnostics, sanity checks and a rigorous quantification of uncertainty that allow the experimenter to be explicit about the weight of evidence.
We compiled a diverse set of computational techniques to perform robust Bayesian comparison of models of causal inference in multisensory perception, which we dub the ‘Bayesian cookbook for causal inference in multisensory perception’, or herein simply ‘the cookbook’. The main goal of the cookbook is to characterize observers’ decision strategies underlying causal inference, and possibly other details thereof, within a rigorous Bayesian framework that accounts for both parameter uncertainty and model uncertainty. The cookbook is ‘doubly-Bayesian’ in that it affords a fully Bayesian analysis of observers who may or may not be performing Bayesian inference themselves [51]. Fully Bayesian model comparison is computationally intensive, hence the cookbook is concerned with efficient algorithmic solutions.
The cookbook comprises of: (a) a fairly general recipe for building observer models for causal inference in multisensory perception (see Methods and Section 1 of S1 Appendix), which lends itself to a factorial model comparison; (b) techniques for fast evaluation of a large number of causal inference observer models; (c) procedures for model fitting via maximum likelihood, and approximating the Bayesian posterior of the parameters via Markov Chain Monte Carlo (MCMC); (d) state-of-the-art methods to compute model comparison metrics and perform factorial model selection. It is noteworthy that, while the current work focuses on the example of visuo-vestibular heading perception, this cookbook is general and can be applied with minor modifications to multisensory perception across sensory domains. Computational details are described in the Methods section and S1 Appendix. Here we present an application of our framework to causal inference in multisensory heading perception. For ease of reference, we summarize relevant abbreviations used in the paper and their meaning in Table 1.
We demonstrate our framework taking as a case study the comparison of explicit vs. implicit causal inference strategies in heading perception. In this section we briefly summarize our methods. Extended details and description of the cookbook can be found in the Methods and S1 Appendix.
We examined how subjects perceived the causal relationship of synchronous visual and vestibular headings as a function of disparity (svest − svis, nine levels) and visual reliability level (high, medium, low; Fig 3A). Common cause reports were more frequent near zero disparities than for well-separated stimuli (Repeated-measures ANOVA with Greenhouse-Geisser correction; F(1.82,18.17) = 76.0, ϵ = 0.23, p < 10−4, η p 2 = 0 . 88). This means that observers neither performed complete integration (always reporting a common cause) nor complete segregation (never reporting a common cause). Common-cause reports were not affected by visual cue reliability alone (F(1.23,12.33) = 1.84, ϵ = 0.62, p = .2, η p 2 = 0 . 16), but were modulated by an interaction of visual reliability and disparity (F(7.44,74.44) = 7.38, ϵ = 0.47, p < 10−4, η p 2 = 0 . 42). Thus, observers’ performance was affected by both cue disparity as well as visual cue reliability when explicitly reporting about the causal relationship between visual and vestibular cues. However, this does not necessarily mean that the subjects’ causal inference strategy took visual cue reliability into account. Changes in sensory noise may affect measured behavior even if the observer’s decision rule ignores such changes [35]; a quantitative model comparison is needed to probe this question.
We compared a subset of models from the full factorial comparison (Fig 2A), since some models are equivalent when restricted to the explicit causal inference task. In particular, here fixed-criterion models are not influenced by the ‘prior’ factor, and the (stochastic) fusion model is not affected by sensory noise or prior, thus reducing the list of models to seven: Bay-C-E, Bay-C-I, Bay-X-E, Bay-X-I, Fix-C, Fix-X, SFu.
To assess the evidence for distinct determinants of subjects’ behavior, we combined LOO scores from individual subjects and models with a hierarchical Bayesian approach [54] (Fig 3B). Since we are investigating model factors that comprise of an unequal number of models, we reweighted the prior over models such that distinct components within each model factor had equal prior probability (Fix models had 2× weight, and SFu 4×). In Fig 3B we report the protected exceedance probabilities φ ˜ and, for reference, the posterior model frequencies they are based on, and the Bayesian omnibus risk (BOR), which is the estimated probability that the observed differences in factor frequencies may be due to chance [55]. We found that the most likely factor of causal inference was the Bayesian model (φ ˜ = 0 . 78), followed by fixed-criterion (φ ˜ = 0 . 18) and probabilistic fusion (φ ˜ = 0 . 04). That is, fusion was ∼ 24 times less likely to be the most representative model than any form of causal inference combined, which is strong evidence against fusion, and in agreement with our model-free analysis. The Bayesian strategy was ∼ 3.5 times more likely than the others, which is positive but not strong evidence [57]. Conversely, the explicit causal inference data do not allow us to draw conclusions about noise models (constant vs. eccentric) or priors (empirical vs. independent), as we found that all factor components are about equally likely (φ ˜ ∼ 0 . 5).
At the level of specific models—as opposed to aggregate model factors –, we found that the probability of being the most likely model was almost equally divided between fixed-criterion (C-I) and Bayesian (either X-E or C-I). All these models yielded reasonable fits (Fig 3C), which captured a large fraction of the noise in the data (absolute goodness of fit ≈ 76% ± 3%; see Methods); a large improvement over a constant-probability model, which had a goodness of fit of 14 ± 5%. For comparison, we also show in Fig 3C the stochastic fusion model, which had a goodness of fit of 17 %± 5%. Visually, the Fix model in Fig 3C seems to fit better the group data, but we found that this is an artifact of projecting the data on the disparity axis. Disparity is the only relevant dimension for the Fix model; whereas Bay models fits the data along all dimensions. The visual superiority of the Fix model wanes when the data are visualized in their entirety (see S1 Fig).
We verified robustness of our findings by performing the same hierarchical analysis with different model comparison metrics. All metrics were in agreement with respect to the Bayesian causal inference strategy as the most likely, and the same three models being most probable (although possibly with different ranking). BIC and marginal likelihood differed from LOO and AICc mainly in that they reported a larger probability for the constant vs. eccentricity-dependent noise (probability ratio ∼4.6, which is positive but not strong evidence).
These results combined provide strong evidence that subjects in the explicit causal inference task took into account some elements of the statistical structure of the trial (disparity, and possibly cue reliability) to report unity judgments, consistent with causal inference, potentially in a Bayesian manner. From these data, it is unclear whether observers took into account the empirical distribution of stimuli, and whether their behavior was affected by eccentricity-dependence in the sensory noise.
We examined the bias in the reported direction of inertial heading computed as (minus) the point of subjective equality for left/rightward heading choices (L/R PSE), for each visual heading and visual cue reliability (Fig 4A). Specifically, for a given value of visual heading svis (or small range thereof), we constructed a psychometric function as a function of svest (see Methods for details). If subjects were influenced by svis and took visual heading into account while computing inertial heading, this would manifest as bias in the psychometric function (that is, a shifted point of subjective equality). If subjects were able instead to discount the distracting influence of svis, there should be negligible bias. As per causal inference, we qualitatively expected that there would be bias for smaller |svis|, but the bias would either decrease or saturate as |svis| increases. However, note that a nonlinear pattern of bias may also emerge due to eccentricity-dependence of the noise, even in the absence of causal inference.
The bias was significantly affected by visual heading (Repeated-measures ANOVA; F(0.71,7.08) = 19.67, ϵ = 0.07, p = .004, η p 2 = 0 . 66). We found no main effect of visual cue reliability alone (F(0.85,8.54) = 0.51, ϵ = 0.43, p = .47, η p 2 = 0 . 05), but there was a significant interaction of visual cue reliability and heading (F(2.93,29.26) = 7.36, ϵ = 0.15, p < 10−3, η p 2 = 0 . 42). These data suggest that subjects’ perception of vestibular headings was modulated by visual cue reliability and visual stimulus, in agreement with previous work in visual-auditory localization [21]. However, quantitative model comparison is required to understand the mechanism in detail since distinct processes, such as different causal inference strategies and noise models, could lead to similar patterns of observed behavior.
We performed a factorial comparison with all models in Fig 2A. In this case, factorial model comparison via LOO was unable to uniquely identify the causal inference strategy adopted by observers (Fig 4B). Forced fusion was slightly favored (φ ˜ ∼ 0 . 48), followed by Bayes (φ ˜ ∼ 0 . 27) and fixed-criterion (φ ˜ ∼ 0 . 25), suggesting that all strategies were similar to forced fusion. Conversely, eccentricity-dependent noise was found to be more likely than constant noise (ratio ∼ 5.7), which is positive but not strong evidence, and empirical priors were marginally more likely than independent priors (∼ 2.1). The estimated Bayesian omnibus risk was high (BOR ≥ 0.29), hinting at a large degree of similarity within all model factors such that observed differences could have arisen by chance.
All metrics generally agreed on the lack of evidence in favor of any specific inference strategy (with AICc and BIC tending to marginally favor fixed-criterion instead of fusion), and on empirical priors being more likely. As a notable difference, marginal likelihood and BIC reversed the result about noise models, favoring constant noise models over eccentricity-dependent ones.
In terms of individual models, the most likely models according to LOO were, in order, forced fusion (X-E), Bayesian (X-E), and fixed-criterion (C-E). However, other metrics also favored other models; for example, Bayesian (C-E) was most likely according to the marginal likelihood. All these models obtained similarly good fits to individual data (Fig 4C; absolute goodness of fit ≈ 97%). For reference, a model that responds ‘rightward motion’ with constant probability performed about at chance (goodness of fit ≈ 0.3 ± 0.1%).
In sum, our analysis shows that the implicit causal inference data alone are largely inconclusive, possibly because almost all models behave similarly to forced fusion. To further explore our results, we examined the posterior distribution of the prior probability of common cause parameter pc across Bayesian models, and of the criterion κc for fixed-criterion models (Fig 5, bottom left panels). In both cases we found a broad distribution of parameters, with only a mild accumulation towards ‘forced fusion’ values (pc = 1 or κ c ≳ 90 °), suggesting that subjects were not completely performing forced fusion. Thus, it is possible that by constraining the inference with additional data we would be able to draw more defined conclusions.
Data from the explicit and implicit causal inference tasks, when analyzed separately, afforded only weak conclusions about subjects’ behavior. The natural next step is to combine datasets from the two tasks along with the data from the unisensory heading discrimination task in order to better constrain the model fits.
Before performing such joint fit, we verified whether there was evidence that model parameters changed substantially across tasks, in which case we might have had to change the structure of the models (e.g., by introducing a subset of distinct parameters for different tasks [49]). For each model parameter, we computed the across-tasks compatibility probability Cp (Fig 5), which is the (posterior) probability that subjects were most likely to have the same parameter values across tasks, as opposed to different parameters, above and beyond chance (see Methods for details). We found at most mild evidence towards difference of parameters across the three tasks, but no strong evidence (all Cp > .05). Therefore, we proceeded in jointly fitting the data with the default assumption that parameters were shared across tasks.
For the joint fits there are nine possible models for the causal inference strategy (three explicit causal inference × three implicit causal inference strategies). However, we considered only a subset of plausible combinations, to avoid ‘model overfitting’ (see Discussion). First, we disregarded the stochastic fusion strategy for the explicit task, since this strategy was strongly rejected by the explicit task data alone. Second, if subjects performed some form of causal inference (Bayesian or fixed-criterion) in both tasks, we forced it to be the same. This reduces the model space for the causal inference strategy to four components: Bay/Bay, Fix/Fix, Bay/FFu, Fix/FFu (explicit/implicit task). Combined with the prior and sensory noise factors as per Fig 2A, this leads to sixteen models.
Factorial model comparison via LOO found that the most likely causal inference strategy was fixed-criterion (φ ˜ = 0 . 79), followed by Bayesian (φ ˜ = 0 . 13), and then by forced fusion in the implicit task (φ ˜ = 0 . 05 paired with Bayesian explicit causal inference, φ ˜ = 0 . 03 paired with fixed-criterion explicit causal inference; Fig 6A). This is positive evidence that subjects were performing some form of causal inference also in the implicit task, as opposed to mere forced fusion (ratio ∼ 11.4). Moreover, we found strong evidence for eccentricity-dependent over constant noise (φ ˜ > 0 . 99, ratio ∼ 132.7). Instead, the joint data were still inconclusive about the prior adopted by the subjects, with only marginal evidence for the empirical prior over the independent prior (∼ 2.9).
In terms of specific models, the most likely model was fixed-criterion (X-E), followed by Bayesian (X-E), and explicit Bayesian / implicit forced fusion (both X-I and X-E). The best models gave a good description of the individual joint data, with an absolute goodness of fit of ≈ 91% ± 1% (Fig 6B).
Examination of the subjects’ posteriors over parameters for the joint fits (Table 2 and Fig 5, black lines) showed reasonable results. The base visual noise parameters were generally monotonically increasing with decreasing visual cue reliability; the vestibular base noise was roughly of the same magnitude as the medium visual cue noise (as per experiment design); both visual and vestibular noise increased mildly with the distance from straight ahead; subjects had a small lapse probability. For Bayesian models, pc was substantially larger than the true value, 0.20 (t-test t(10) = 10.8, p < 10−4, d = 3.25), suggesting that observers generally thought that heading directions had a higher a priori chance to be the same. Nonetheless, for all but one subject pc was far from 1, suggesting that subjects were not performing forced fusion either. An analogous result holds for the fixed criterion κc, which was smaller than the largest disparity between heading directions. We found that prior parameters σprior and Δprior had a lesser impact on the models, and their exact values were less crucial, with generally wide posteriors.
Finally, we verified that our results did not depend on the chosen comparison metric. Remarkably, the findings regarding causal inference factors were quantitatively the same for all metrics, demonstrating robustness of our main result. Marginal likelihood and BIC differed from LOO and AICc in that they only marginally favored eccentricity-dependent noise models, showing that conclusions over the noise model may depend on the specific choice of metric. All metrics agreed in marginally preferring the empirical prior over the independent prior.
In conclusion, when combining evidence from all available data, our model comparison shows that subjects were most likely performing some form of causal inference instead of forced fusion, for both the explicit and the implicit causal inference tasks. In particular, we find that a fixed-criterion, non-probabilistic decision rule (i.e., one that does not take uncertainty into account) describes the joint data better than the Bayesian strategy, although with some caveats (see Discussion).
Performing a factorial comparison, like any other statistical analysis, requires a number of somewhat arbitrary choices, loosely motivated by previous studies, theoretical considerations, or a preliminary investigation of the data (being aware of the ‘garden of forking paths’ [58]). As good practice, we want to check that our main findings are robust to changes in the setup of the analysis, or be able to report discrepancies.
We take as our main result the protected exceedance probabilties φ ˜ of the model factors in the joint analysis (Fig 6A, reproduced in Fig 7, top row). In the following, we examine whether this finding holds up to several manipulations of the analysis framework.
A first check consists of testing different model comparison metrics. In the previous sections, we have reported results for different metrics, finding in general only minor differences from our results obtained with LOO. As an example, we show here the model comparison using as metric an estimate of the marginal likelihood—the probability of the data under the model (Fig 7, 2nd row). We see that the marginal likelihood results agree with our results with LOO except for the sensory noise factor (see Discussion). Therefore, our conclusions about the causal inference strategy are not affected.
Second, the hierarchical Bayesian Model Selection method requires to specify a prior over frequencies of models in the population [54]. This (hyper)prior is specified via the concentration parameter vector α0 of a Dirichlet distribution over model frequencies. For our analysis, since we focused on the factorial aspect, we chose an approximately ‘flat’ prior across model factors (see Methods for details), instead of the default flat prior over individual models (α0 = 1). We found that performing the group analysis with α0 = 1 did not change our results (Fig 7, 3rd row).
Another potential source of variation is specific model choices, or inclusion of model factors. For example, a common successful variant of the Bayesian causal inference strategy is ‘probability matching’, according to which the observer chooses the causal scenario (C = 1 or C = 2) randomly, proportionally to its posterior probability [24]. As a first check, we performed the model comparison again using a ‘probability matching’ Bayesian observer instead of our main ‘model averaging’ observer (Fig 7, 4th row). Results are similar to the main analysis. If anything, the fixed-criterion causal inference strategy gains additional evidence here, suggesting that probability matching is a worse description of the data than our original Bayesian causal inference model (as confirmed by looking at differences in LOO scores of individual subjects, e.g. for the Bay-X-E model; mean ± SEM: ΔLOO = −17.3 ± 5.7). A recent study in audio-visual causal inference perception has similarly found that probability matching provided a poor explanation of the data [21].
In the factorial framework we could also have performed the previous analysis in a different way, by considering ‘probability matching’ as a sub-factor of the Bayesian strategy, together with ‘model averaging’. As we have done before for the explicit causal inference task, we reassign prior probabilities to the models so that they are constant for each factor (in this case, the two Bayesian strategies get a × 1 2 multiplier). Results of this alternative approach show an increase of evidence for the Bayesian causal inference family (Fig 7, bottom row). The values of φ ˜ for the fusion models are also slightly higher, which is due to an increase of the Bayesian omnibus risk (the probability that the observed differences in factor frequencies are due to chance, a warning sign that there are too many models for the available data). This result and other lines of reasoning suggest caution when model factors contain an uneven number of models (see Discussion). Nonetheless, the main conclusion does not qualitatively change, in that observers performed some form of causal inference as opposed to forced fusion.
Finally, we performed several sanity checks, including a model recovery analysis to ensure the integrity of our analysis pipeline and that models of interest were meaningfully distinguishable (see Methods and S1 Appendix for details).
In conclusion, we have shown how the computational framework of Bayesian factorial model comparison, which is made possible by a combination of methods described in the cookbook, allows to explore multiple questions about aspects of subjects’ behavior in multisensory perception, and to account for uncertainty at different levels of the analysis in a principled, robust manner.
We presented a ‘cookbook’ of algorithmic recipes for robust Bayesian evaluation of observer models of causal inference that have widespread applications to multisensory perception and modeling perceptual behavior in general. We applied these techniques to investigate the decision strategies that characterize explicit and implicit causal inference in multisensory heading perception. Examination of observers’ behavior in the explicit and implicit causal inference tasks provided evidence that observers did not simply fuse visual and vestibular cues. Instead, observers integrated the multisensory cues based on their relative disparity, a signature of causal inference. Importantly, our framework affords investigation of whether humans adopt a statistically optimal Bayesian strategy or instead implement a heuristic decision rule which does not fully consider the uncertainty associated with the stimuli.
Our findings in the explicit causal inference task demonstrate that subjects used information about the discrepancy between the visual and vestibular cues to infer the causal relationship between them. Results in the implicit causal inference task alone were mixed, in that we could not clearly distinguish between alternative strategies, including forced fusion—in agreement with a previous finding [33]. However, when we combined evidence from all tasks, we found that some form of causal inference was more likely than mere forced fusion, in agreement with a more recent study [34]. Our findings suggest that multiple sources of evidence (e.g., different tasks) can help disambiguate causal inference strategies which might otherwise produce similar patterns of behavioral responses.
Our Bayesian analysis allowed us to examine the distribution of model parameters, in particular the causal inference parameters pc and κc, which govern the tendency to bind or separate cues for, respectively, a Bayesian and a heuristic fixed-criterion strategy. Evidence from all tasks strongly constrained these parameters for each subject. Interestingly, for the Bayesian models we found an average pc much higher than the true experimental value (inferred pc ∼ 0.5 vs. experimental pc = 0.2). This suggests that subjects had a tendency to integrate sensory cues substantially more than what the statistics of the task would require. Note that, instead, a Bayesian observer would be able to learn the correct value of pc from noisy observations, provided some knowledge of the structure of the task. Our finding is in agreement with previous studies which demonstrated an increased tendency to combine discrepant visual and vestibular cues [10, 33, 43, 59, 60] and also a large inter-subject variability in pc, and not obviously related to the statistics of the task [23]. We note that, in all studies so far, the ‘binding tendency’ (pc or κc) is a descriptive parameter of causal inference models that lacks an independent empirical correlate (as opposed to, for example, noise parameters, which can be independently measured). Understanding the origin of the binding tendency, and which experimental manipulations it is sensitive to, is venue for future work [23, 61]. For example, de Winkel and colleagues found that the binding tendency depends on the duration of the motion stimuli; decreasing for motions of longer duration [34].
Previous work has performed a factorial comparison of only causal inference strategies [21]. Our analysis extends that work by including as latent factors the shape of sensory noise (and, thus, likelihoods) and type of priors [48, 49]. Models in our set include a full computation of the observers’ posterior beliefs based on eccentricity-dependent likelihoods, which was only approximated in previous studies that considered eccentricity-dependence [22, 33, 34]. Indeed, in agreement with a recent finding, we found an important role of eccentricity-dependent noise [22]. Conversely, our analysis of priors was inconclusive, as our datasets were unable to tell whether people learnt the empirical (correlated) prior, or made an assumption of independence.
Our main finding, relative to the causal inference strategy, is that subjects performed causal inference both in the explicit and implicit tasks. Interestingly, from our analyses the most likely causal inference strategy is a fixed-criterion strategy, which crucially differs from the Bayesian strategy in that it does not take cue reliability into account—let alone optimally. This finding is seemingly at odds with a long list of results in multisensory perception, in which people are shown to take cue uncertainty into account [9, 10, 42, 62]. We note that this is not necessarily in contrast with existing literature, for several reasons. First, this result pertains specifically to the causal inference part of the observer model, and not how cues are combined once a common cause has been inferred [21]. To our knowledge, no study of multisensory perception has tested Bayesian models of causal inference against heuristic models that take into account disparity but not reliability, as it has been done for example in visual search [56, 63] and visual categorization [36, 64]. A quantitative modeling approach is needed—qualitatively analyzing the differences in behavior at different levels of reliability is not sufficient to establish that observers take uncertainty into account; patterns of observed differences may be due to a change in sensory noise even if the observer’s decision rule disregards cue reliability. Second, our results are not definitive—the evidence for fixed-criterion vs. Bayesian is positive but not decisive. Our interpretation of this result is that subjects are following some suboptimal decision rule which happens to be closer to fixed-criterion than to the Bayesian strategy for the presented stimuli and range of tested reliability levels. It is possible that with a wider range of stimuli and reliabilities, and possibly with different ways of reporting (e.g., estimation instead of discrimination), we would be able to distinguish the Bayesian strategy from a fixed-criterion heuristic.
Finally, we note that model predictions of our Bayesian models are good but still show systematic discrepancies from the data for the explicit causal inference task (Figs 3C and 6B). Previous work has found similar discrepancies in model fits of unity judgments data across multiple sensory reliabilities (e.g., see Fig 2A in [21]). This suggests that there is some element of model mismatch in current Bayesian causal inference models, possibly due to difference in noise models or to other processes that affect causal inference across cue reliabilities, which deserves further investigation.
We performed our analysis within a factorial model comparison framework [50]. Even though we were mainly interested in a single factor (causal inference strategy), previous work has shown that the inferred observer’s decision strategy might depend on other aspects of the observer model, such as sensory noise or prior, due to nontrivial interactions of all these model components [37]. Our method, therefore, consisted of performing inference across a family of observer models that explicitly instantiated plausible model variants. We then marginalized over details of specific observer models, looking at posterior probabilities of model factors, according to a hierarchical Bayesian Model Selection approach [54, 55]. We applied a few tweaks to the Bayesian Model Selection method to account for our focus on factors as opposed to individual models (see Methods).
Our approach was fully Bayesian in that we took into account parameter uncertainty (by computing a metric, LOO, based on the full posterior distribution) and model uncertainty (by marginalizing over model components). A fully Bayesian approach has the advantages of explicitly representing uncertainty in the results (e.g., credible intervals over parameters), and of reducing the risk of overfitting, although it is not immune to it [65].
In our case, we marginalized over models to reduce the risk of model overfitting, which is a complementary problem to parameter overfitting. Model overfitting is likely to happen when model selection is performed within a large number of discrete models. In fact, some authors recommend to skip discrete model selection altogether, preferring instead inference and Bayesian parameter estimation in a single overarching or ‘complete’ model [66]. We additionally tried to reduce the risk of model overfitting by balancing prior probabilities across factors, although we noted that this may not be enough to counterbalance the additional flexibility that a model factor gains by having more sub-models than a competitor. Our practical recommendation, until more sophisticated comparison methods are available, is to ensure that all model components within a factor have the same number of models, and to limit the overall number of models.
Our approach was also factorial in the treatment of different tasks, in that first we analyzed each bisensory task in isolation, and then combined trials from all data in a joint fit. The fully Bayesian approach allowed us to compute posterior distributions for the parameters, marginalized over models (see Fig 5), which in turn made it possible to test whether model parameters were compatibile across tasks, via the ‘compatibility probability’ metric. The compatibility probability is an approximation of a full model comparison to test whether a given parameter is the same or should differ across different datasets (in this case, tasks), where we consider ‘sameness’ to be the default (simplyfing) hypothesis. We note that if the identity or not of a parameter across datasets is a main question of the study, its resolution should be addressed via a proper model comparison.
With the joint fits, we found that almost all parameters were well constrained by the data (except possibly for the parameters governing the observers’ priors, σprior and Δprior). An alternative option to better constrain the inference for scarce data or poorly identified parameters is to use informative priors (as opposed to non-informative priors), or a hierarchical approach that assumes a common (hyper)prior to model parameters across subjects [67].
The general goal of a model comparison metric is to score a model for goodness of fit and somehow penalize for model flexibility. In our analysis we have used Pareto-smoothed importance sampling leave-one-out cross-validation (PSIS-LOO [53]) as the main metric to compare models (simply called LOO in the other sections for simplicity). In fact, there is a large number of commonly used metrics, such as (corrected) Akaike’s information criterion (AIC(c)) [68], Bayesian information criterion (BIC) [68], deviance information criterion (DIC) [69], widely applicable information criterion (WAIC) [70], and marginal likelihood [71]. The literature on model comparison is vast and with different schools of thought—by necessity here we only summarize some remarks. The first broad distinction between these metrics is between predictive metrics (AIC(c), DIC, WAIC, and PSIS-LOO) [72], that try to approximate out-of-sample predictive error (that is, model performance on unseen data), and BIC and marginal likelihood, which try to establish the true model generating the data [71]. Another orthogonal distinction is between metrics based on point estimates (AIC(c) and BIC) vs. metrics that use partial to full information about the model’s uncertainty landscape (DIC, WAIC, PSIS-LOO, based on the posterior, and the marginal likelihood, based on the likelihood integrated over the prior).
First, when computationally feasible we prefer uncertainty-based metrics to point estimates, since the latter are only crude asymptotic approximations that do not take the model and the data into account, besides simple summary statistics (number of free parameters and possibly number of data points). Due to their lack of knowledge of the actual structure of the model, AIC(c) and BIC can grossly misestimate model complexity [72].
Second, we have an ordered preference among predictive metrics, that is PSIS-LOO ≻ WAIC ≻ DIC ≻ AIC(c) [72]. The reason is that all of these metrics more or less asymptotically approximate full leave-one-out cross validation, with increasing degree of accuracy from right to left [53, 72]. As mentioned before, AIC(c) works only in the regime of a large amount of data. DIC, albeit commonly used, has several issues and requires the posterior to be multivariate normal, or at least symmetric and unimodal—gross failures can happen when this is not the case, since DIC bases its estimate of model complexity on the mean (or some other measure of central tendency) of the posterior [72]. WAIC is a great improvement over DIC and does not require normality of the posterior, but its approximation is generally superseded by PSIS-LOO [53]. Moreover, PSIS-LOO has a natural diagnostic, the exponents of the tails of the fitted Pareto distribution, which allows the user to know when the method may be in trouble [53]. Full leave-one-out cross validation is extremely expensive, but PSIS-LOO only requires the user to compute the posterior via MCMC sampling, with no additional cost with respect to DIC or WAIC. Similarly to WAIC, PSIS-LOO requires the user to store for each posterior sample the log likelihood per trial, which with modern computers represent a negligible storage cost.
The marginal likelihood, or Bayes factor (of which BIC is a poor approximation), is an alternative approach to quantify model evidence, related to computing the posterior probability of the models [71]. While this is a principled approach, it entails several practical and theoretical issues. First, the marginal likelihood is generally hard to compute, since it usually involves a complicated, high-dimensional integral of the likelihood over the prior (although this computation can be simplified for nested models [73]). Here, we have applied a novel approximation method for the marginal likelihood following ideas delineated in [74, 75], obtaining generally sensible values. However, more work is needed to establish the precision and applicability of such technique. Besides practical computational issues, the marginal likelihood, unlike other metrics, is sensitive to the choice of prior over parameters, in particular its range [66]. Crucially, and against common intuition, this sensitivity does not reduce with increasing amounts of data. A badly chosen (e.g., excessively wide) prior for a non-shared parameter might change the marginal likelihood of a model by several points, thus affecting model ranking. The open issue of prior sensitivity has led some authors to largely discard model selection based on the marginal likelihood [66].
For these reasons, we chose (PSIS-)LOO as the main model comparison metric. As a test of robustness, we also computed other metrics and verified that our results were largely independent of the chosen metric, or investigated the reasons when it was not the case.
As a specific example, in our analysis we found that LOO and marginal likelihood (or BIC) generally agreed on all comparisons, except for the sensory noise factor. Unlike LOO, the marginal likelihood tended to prefer constant noise models as opposed to eccentricity-dependent models. Our explanation of this discrepancy is that for our tasks eccentricity-dependence provides a consistent but small improvement to the goodness of fit of the models, which can be overrided by a large penalty due to model complexity (BIC), or to the chosen prior over the eccentricity-dependent parameters (wvis, wvest), whose range was possibly wider than needed (see Fig 5). The issue of prior sensitivity (specifically, dependence of results on an arbitrarily chosen range) can be attenuated by adopting a Bayesian hierarchical approach over parameters (or a more computationally feasibile approximation, known as empirical Bayes), which is venue for future work.
Model evaluation, especially from a Bayesian perspective, is a time-consuming business. For this reason, we have compiled several state-of-the-art methods for model building, fitting and comparison, and made our code available.
The main issue of many common observer models in perception is that the expression for the (log) likelihood is not analytical, requiring numerical integration or simulation. To date, this limits the applicability of modern model specification and analysis tools, such as probabilistic programming languages, that exploit auto-differentiation and gradient-based sampling methods (e.g., Stan [76] or PyMC3 [77]). The goal of such computational frameworks is to remove the burden and technical details of evaluating the models from the shoulders of the modeler, who only needs to provide a model specification.
In our case, we strive towards a more modest goal of providing black-box algorithms for optimization and MCMC sampling that exhibit a larger degree of robustness than standard methods. In particular, for optimization (maximum likelihood estimation) we recommend Bayesian adaptive direct search (BADS [78]), a technique based on Bayesian optimization [79, 80], which exhibits robustness to noise and jagged likelihood landscapes, unlike common optimization methods such as fminsearch (Nelder-Mead) and fmincon in MATLAB. Similarly, for MCMC sampling we propose a sampling method that combines the robustness and self-adaptation of slice sampling [81] and ensemble-based methods [82]. Crucially, our proposed method almost completely removes the need of expensive trial-and-error tuning on the part of the modeler, possibly one of the main reasons why MCMC methods and full evaluation of the posterior are relatively uncommon in the field (to our knowledge, this is the first study of causal inference in multisensory perception to adopt a fully Bayesian approach).
Our framework is similar to the concept behind the VBA toolbox, a MATLAB toolbox for probabilistic treatment of nonlinear models for neurobiological and behavioral data [83]. The VBA toolbox tackles the problem of model fitting via a variational approximation that assumes factorized, Gaussian posteriors over the parameters (mean field/Laplace approximation), and provides the variational free energy as an approximation (lower bound) of the marginal likelihood. Our approach, instead, does not make any strong assumption, using MCMC to recover the full shape of the posterior, and state-of-the-art techniques to assess model performance.
Detailed, rigorous modeling of behavior is a necessary step to constrain the search for neural mechanisms implementing decision strategies [84] We have provided a set of computational tools and demonstrated how they can be applied to answer specific questions about internal representation and decision strategies of the observer in multisensory perception, with the goal of increasing the set of models that can be investigated, and the robustness of such analyses. Thus, our tools can be of profound use not only to the field of multisensory perception, but to biological modeling in general.
The Institutional Review Board at the Baylor College of Medicine approved the experimental procedures (protocol number H-29411, “Psychophysics of spatial orientation and vestibular influences on spatial constancy and movement planning”) and all subjects gave written informed consent.
We build upon standard causal inference models of multisensory perception [18]. For concreteness, in the following description of causal inference models we refer to the visuo-vestibular example with binary responses (‘left/right’ for discrimination, and ‘yes/no’ for unity judgements). The basic component of any observer model is the trial response probability, that is the probability of observing a given response for a given trial condition (e.g., stimulus pair, uncertainty level, task). In the following we briefly review how these probabilities are computed.
All analysis code was written in MATLAB (Mathworks, Inc.), with core computations in C for increased performance (via mex files in MATLAB). Code is available at https://github.com/lacerbi/visvest-causinf.
For a given model, we denote its set of parameters by a vector θ. For a given model and dataset, we define the parameter log likelihood function as
LL ( θ , model ) = log p ( data | θ , model ) = log ∏ i = 1 N trials p ( r ( i ) | s vis ( i ) , s vest ( i ) , c vis ( i ) , θ , model ) = ∑ i = 1 N trials log p ( r ( i ) | s vis ( i ) , s vest ( i ) , c vis ( i ) , θ , model ) (10)
where we assumed conditional independence between trials; r(i) denotes the subject’s response (‘right’ or ‘left’ for the discrimination trials; ‘common’ or ‘separate’ causes in unity judgment trials); s vis ( i ) and s vest ( i ) are, respectively, the direction of motion of the visual (resp. vestibular) stimulus (if present), and c vis ( i ) is the visual coherence level (that is, reliability: low, medium, or high), in the i-th trial.
We built different observer models by factorially combining three factors: causal inference strategy (Bayesian, fixed-criterion, or fusion); shape of sensory noise (constant or eccentricity-dependent); and type of prior over heading directions (empirical or independent); see Fig 2A and ‘Causal inference models’ section of the Methods for a description of the different factors.
For each subject, we fitted the different observer models, first separately to different tasks (unity judgment and bisensory inertial discrimination), and then performed a joint fit by combining datasets from all tasks (including the unisensory discrimination task). We evaluated the fits with a number of model comparison metrics and via an objective goodness of fit metric. Finally, we combined evidence for different model factors across subjects with a hierarchical Bayesian approach.
We verified our ability to distinguish different models with a model recovery analysis, described in S1 Appendix.
The Bayesian cookbook for causal inference in multisensory perception, or simply ‘the cookbook’, consists of a recipe to build causal inference observer models for multisensory perception, and a number of algorithms and computational techniques to perform efficient and robust Bayesian comparison of such models. We applied and demonstrated these methods at different points in the main text; further details can be found here in the Methods and S1 Appendix. For reference, we summarize the main techniques of interest in Table 3.
|
10.1371/journal.pgen.1000602 | Aging and Environmental Exposures Alter Tissue-Specific DNA Methylation Dependent upon CpG Island Context | Epigenetic control of gene transcription is critical for normal human development and cellular differentiation. While alterations of epigenetic marks such as DNA methylation have been linked to cancers and many other human diseases, interindividual epigenetic variations in normal tissues due to aging, environmental factors, or innate susceptibility are poorly characterized. The plasticity, tissue-specific nature, and variability of gene expression are related to epigenomic states that vary across individuals. Thus, population-based investigations are needed to further our understanding of the fundamental dynamics of normal individual epigenomes. We analyzed 217 non-pathologic human tissues from 10 anatomic sites at 1,413 autosomal CpG loci associated with 773 genes to investigate tissue-specific differences in DNA methylation and to discern how aging and exposures contribute to normal variation in methylation. Methylation profile classes derived from unsupervised modeling were significantly associated with age (P<0.0001) and were significant predictors of tissue origin (P<0.0001). In solid tissues (n = 119) we found striking, highly significant CpG island–dependent correlations between age and methylation; loci in CpG islands gained methylation with age, loci not in CpG islands lost methylation with age (P<0.001), and this pattern was consistent across tissues and in an analysis of blood-derived DNA. Our data clearly demonstrate age- and exposure-related differences in tissue-specific methylation and significant age-associated methylation patterns which are CpG island context-dependent. This work provides novel insight into the role of aging and the environment in susceptibility to diseases such as cancer and critically informs the field of epigenomics by providing evidence of epigenetic dysregulation by age-related methylation alterations. Collectively we reveal key issues to consider both in the construction of reference and disease-related epigenomes and in the interpretation of potentially pathologically important alterations.
| The causes and extent of tissue-specific interindividual variation in human epigenomes are underappreciated and, hence, poorly characterized. We surveyed over 200 carefully annotated human tissue samples from ten anatosites at 1,413 CpGs for methylation alterations to appraise the nature of phenotypically, and hence potentially clinically important epigenomic alterations. Within tissue types, across individuals, we found variation in methylation that was significantly related to aging and environmental exposures such as tobacco smoking. Individual variation in age- and exposure-related methylation may significantly contribute to increased susceptibility to several diseases. As the NIH–funded HapMap project is critically contributing to annotating the human reference genome defining normal genetic variability, our work raises key issues to consider in the construction of reference epigenomes. It is well recognized that understanding genetic variation is essential to understanding disease. Our work, and the known interplay of epigenetics and genetics, makes it equally clear that a more complete characterization of epigenetic variation and its sources must be accomplished to reach the goal of a complete understanding of disease. Additional research is absolutely necessary to define the mechanisms controlling epigenomic variation. We have begun to lay the foundations for essential normal tissue controls for comparison to diseased tissue, which will allow the identification of the most crucial disease-related alterations and provide more robust targets for novel treatments.
| While all somatic cells in a given individual are genetically identical (excepting T and B cells), different cell types form highly distinct anatomic structures and carry out a wide range of disparate physiologic functions. The vast repertoire of cellular phenotypes is made possible largely via epigenetic control of gene expression, which is known to play a critical role in cellular differentiation. Epigenetics is the study of mitotically and/or meiotically heritable changes in gene function that cannot be explained by changes in DNA sequence [1], and includes critical normal processes such as X-chromosome inactivation and genomic imprinting. Alterations in epigenetic control have been linked to several human pathologic conditions including cancers, and Rett, ICF, and Beckwith-Wiedemann syndromes [2]–[5]. The most widely studied epigenetic mark is DNA cytosine methylation, most often investigated in the context of CpG dinculeotides in promoter regions which often have concentrations of CpGs known as CpG islands. Normal cells are thought to generally maintain CpG islands in an unmethylated state permissive to transcription [6]. However, emerging work has established the presence of tissue specific methylation patterns in normal tissue at these islands [7]–[10]. Further, just as normal genetic variation is now understood to be associated with a predisposition to a vast array of human diseases [11], it is important that we begin the research needed to define the underlying interindividual differences in tissue specific methylation that lead us to an understanding of the nature of the relationships that govern these crucial tissue specific differences.
We have previously distinguished normal and tumor tissues using methylation profiling [12],[13]. These studies demonstrated variability in the methylation profiles of pleural mesotheliomas and tumors of the head and neck that was, in part, attributable to etiologically important exposures. In a similar manner, there is a basic need for epigenetic profiling of normal tissues to more completely characterize the normal pattern of promoter methylation variation in development, aging, and in response to common environmental exposures such as alcohol and tobacco smoke.
Efforts to describe the methylation profiles of normal tissues are now underway. Recent genome-wide studies of methylation in normal human tissues have shown that DNA methylation profiles are tissue-specific and correlated with sequence elements [8]–[10], [14]–[16]. However, while these studies are groundbreaking in showing that tissues have different patterns of methylation, the underlying causes and extent of tissue-specific and non-specific interindividual variation in DNA methylation patterns remain largely unknown. In fact, in a follow up experiment from a larger effort, Illingworth et al. observed significant variation among individuals when bisulfite sequencing a particular CpG island, and suggested that larger-scale studies are required to determine the extent of interindividual variability in methylation patterns [9]. Epigenetic variation has been hypothesized to cause underlying differences in disease susceptibility among monozygotic twins, and young twin-pairs have been shown to be more epigenetically similar than older monozygotic twins [17]. Therefore, the aging process and differences in environment have been hypothesized to influence clinically significant changes in methylation profiles as individuals accumulate varying exposures with age. In fact, recent work has shown an overall trend of increased methylation associated with older age in normal human prostate and colon tissues in several genes [18],[19]. Although an increase in promoter methylation with aging is generally accepted, recent evidence from Bjornsson et al. suggests a more complex picture. These authors found both increased and decreased intra-individual global methylation levels (enriched for promoter regions) in peripheral blood cell DNA over time [20]. In this background, it is crucial to more extensively characterize the contribution of aging and the environment to tissue-specific interindividual epigenetic variation.
In this study we used Illumina's GoldenGate methylation platform to investigate cytosine methylation in 217 normal human tissue specimens from 10 different anatomic sites in order to begin to understand variation both between and within tissues across individuals. Profiling CpG methylation of normal human tissues allowed us to begin characterizing the role of aging and environmental exposures in interindividual methylation variation, as well as specific gene-loci determinant of normal tissue-specificity. This work highlights the dynamic nature of epigenomes, and begins to disentangle the roles of aging, environmental factors, and innate variability among individual epigenomic profiles, both within, and across tissues.
Array methylation data were first assembled for exploration and visualization with unsupervised hierarchical clustering using Manhattan distance and average linkage for the 500 most variable autosomal CpG loci (Figure 1). Epigenetic profiles among these normal tissues are strikingly different. Applying recursively partitioned mixture modeling (RPMM) [21] to methylation data from all autosomal CpG loci across all 217 normal human tissue samples resulted in 23 methylation classes and their average methylation profiles (Figure 2A). Among the 23 classes in this model, 16 classes (70%) perfectly captured only a single tissue type (Table 1), and methylation profile classes were a highly significant predictor of sample tissue type (permutation P<0.0001). Further, age was significantly associated with methylation classes (P<0.0001). Separating samples into groups as placenta, blood, or other solid tissue, we found a significant association between group and methylation profile classes (P<0.0001).
Random Forests (RF) classification of all samples based on methylation average beta values at all autosomal loci returned a confusion matrix showing: which samples are correctly classified, which are misclassified, and the misclassification error rate for each sample type (Table 2). Overall, 19 samples were confused with different tissue types, giving an overall misclassification error rate of 8.8%, significantly lower than expected under the null hypothesis (P<0.0001). Not unexpectedly, tissue types with larger sample sizes showed a significantly reduced misclassification error rates (P<0.05). The mean and standard deviation of average beta values for all autosomal CpG loci in each tissue type, and values for the decrease in random forest classification accuracy with locus removal are given in Table S1. In a RF analysis that examined whether samples could be correctly classified as placenta, blood, or other solid tissue, no samples were misclassified (misclassification error = 0%, P<0.0001), and the mean and standard deviation of average beta values for all autosomal CpG loci in each of placenta, blood, or solid tissue, and values for the decrease in random forest classification accuracy with locus removal are given in Table S2.
Variation in tissue-specific methylation relative to differences between tissue types was first explored visually. Scatter plots of methylation values for representative samples from two different tissues were less well correlated than similar plots of two representative samples from the same tissue type, though variation in tissue-specific methylation was also evident (Figure S1). Tissue-specific methylation patterns for adult blood, lung, and pleural tissue samples were then modeled with RPMM to investigate potential associations of age and exposures with methylation profiles. An RPMM of adult bloods (n = 30) resulted in two methylation classes (Figure 2B), and age differed significantly by methylation class (P<0.005), though we did not detect significant associations between methylation class and smoking status, packyears, or alcohol consumption. An RPMM of lung tissues (n = 53) resulted in five methylation classes (Figure 2C) where class membership was not associated with age or smoking status. An RPMM of pleural tissues (n = 18) resulted in five methylation classes (Figure 2D), and class membership was not associated with age; yet, an association between methylation class and asbestos exposure approached significance (P<0.07).
While exposures were not strongly associated with array-wide methylation profiles, locus-specific analysis revealed several exposure-related methylation alterations. Among pleural tissues 24 CpG loci had asbestos-related alterations in methylation, all of which were increases in methylation (Q<0.05, Table S3). In adult bloods, increasing packyears of smoking was significantly associated with MLH1 (Q<0.0001), and RIPK3 (Q<0.002) methylation; and over 30 CpG loci had significantly altered methylation in never versus ever drinkers (Q<0.05, Table S4). Among lung tissues, smoking status (never/ever) was associated with altered methylation at 138 CpG loci (Q<0.05, Table S5).
Given our results from RPMM and previous reports of age-related increases in methylation in normal tissues [18],[19],[22] we next focused on age-related methylation at specific CpG loci. We began by examining gene-loci that other investigators have reported to be associated with age and found that ESR1, GSTP1, IGF2, MGMT, MYOD1, RARB, and RASSF1 had significant age-associated methylation alterations, the majority of which were increases (P<0.05, age range >0, n = 139, Table 3). Hypothesizing that alterations in epigenetic regulatory genes or genes involved in aging processes could lead to the observed associations between age and methylation profiles from RPMM, we tested CpG loci in epigenetic regulatory genes, telomere maintenance genes, and a premature aging syndrome gene, again finding significant age related methylation alterations (Table 3). For example, LAMB1 – involved in subchromosome domain positioning [23] – had increased methylation with age. Significant age-related methylation alterations in telomere maintenance gene-loci TERT, ERCC1, RAD50, and the Werner syndrome gene-locus (WRN) were also observed. Additionally, and in contrast to the predominantly increased age associated methylation at other gene-loci, there was a significant age-related decrease in CpG methylation of the de novo methyltransferase DNMT3B; and unlike the vast majority of other CpGs tested, DNMT3B_P352 was not located in a CpG island (Table 3).
To expand the examination of age-associated methylation alterations, we performed array-wide locus-by-locus analysis of CpGs. For all tissues (age range >0, n = 139), after correcting for multiple comparisons, over 300 CpG loci had age-related methylation alterations (Q<0.05, Table 4). Restricting analysis to solid tissues (n = 119) revealed over 250 CpG loci with age-related methylation alterations (Q<0.05, Table 4). Tissue-specific locus-by-locus analysis of age-related methylation was also performed (tissue types with n >10), detailed in Table S6.
There is now a considerable literature that suggests that genome structure affects both the initial placement of DNA methylation marks in development [24] as well as protecting silenced regions from being perturbed later in life [25]. To examine the possibility that genomic structure can affect the changes we observed in normal methylation, we assessed the potential effects of CpG island status on age-related methylation, with the hypothesis that there may be differential susceptibility to changes in DNA methylation in queried regions defined as canonical islands compared to those not in CpG islands. A CpG island is defined according to Takai and Jones [26], as a region of 200 bp with a GC content of >55% with an observed to expected ratio of CpG >0.65. This analysis used Generalized Estimator Equations (GEE), which are robust to within-person correlation and to the influence of aberrant observations [27], and estimated mean associations between age and methylation by CpG island status. Among all solid tissues (n = 119), the direction of correlation between age and methylation differed dependent upon whether the CpG was found in a CpG island. Loci in CpG islands had significantly positive correlations between methylation and aging, while loci not in CpG islands had significant losses of methylation with aging (P = 7.0E-04; Table 5). Similar trends were observed for other solid tissue types; age-related associations with methylation were significantly positive for loci in CpG islands for pleural tissues, and significantly negative for loci not in islands in brain tissues (Table 5). Interestingly, among adult blood samples, significantly negative correlations between age and methylation alterations were observed irrespective of CpG island status (P = 5.2E-05, Table 5).
To investigate CpG-dependent correlations between aging and methylation in more detail we clustered CpGs (rather than samples) with RPMM (aiming to examine classes of CpGs with similar methylation profiles in more detail), grouping CpGs with similar methylation into eight separate classes. The CpG island status of all loci was plotted, and illustrates the well known tendency for CpGs located in islands to be unmethylated, while non-island CpGs tend to be methylated (Figure 3). We again used GEE, here estimating RPMM class-specific mean associations between age and methylation and plotted the estimates with their 95% confidence intervals. In a class-specific model for solid tissue samples, there was a positive correlation between age and methylation in classes whose loci were predominantly located in CpG islands (P = 1.9E-05, Figure 3A). The tissue specific analysis of pleura demonstrated that classes rich in CpG island loci had significant age-associated increases in methylation (P = 2.3E-08, Figure 3B). Interestingly, the pattern of class-specific correlations between age and methylation in adult bloods was similar to those for solid tissue types, though the correlation between age and methylation was shifted towards the negative such that there was a significant decrease in age-related methylation among loci not in CpG islands (P = 6.3E-06, Figure 3C). Lung tissues displayed a similar pattern of class-specific correlations between age and methylation, and the strength of these correlations approached statistical significance (P = 0.13, Figure 3D). Finally, both brain, and head and neck samples demonstrated increased age-associated methylation in classes rich in CpGs island loci, and decreases in age-associated methylation in classes rich in loci not in CpG islands (P = 7.0E-04, P = 5.2E-08, Figure 3E and 3F, respectively).
Bisulfite modified DNA pyrosequencing was performed to validate array results. Array average beta values were significantly correlated with pyrosequencing percent methylation for sequenced array target CpGs; RARA_P176 (P = 0.003), DNMT3B_P352 (P = 0.008), and LIF_P383 (P = 3.0E-06, Figure S2). Consistent with array-based results, increased RARA_P176-local methylation was associated with reported asbestos exposure in pleural samples (n = 16, P = 0.10, Figure 4A). To confirm the observed associations between age or smoking packyears with methylation, specific loci were sequenced in array samples (n = 28) and an independent set of control blood DNAs (total n = 112). Sequencing DNMT3B_P352 both validated the association between decreased methylation and aging, and confirmed it in an independent population (P = 0.03, n = 112, Figure 4B). Similarly, the association between LIF_P383 methylation and packyears smoked from array results (P<0.02) was validated by pyrosequencing (P<0.02, n = 112, Figure 4C). In addition, pyrosequencing FZD9_E458-local CpGs confirmed the association between increased methylation and aging (P<0.001, n = 112, Figure 4D).
Epigenetic patterning and maintenance are of paramount importance for normal cellular functioning and identity. Hence, pursuing the annotation of normal human tissue-specific epigenomes is an important and necessary endeavor. However, such a project is considerably more challenging than sequencing the genome because of the tissue-specific and dynamic nature of epigenomes. Thus, a more complete understanding of what constitutes a normal epigenome, and the degree to which epigenomes vary (in a tissue dependent fashion) based on aging and the environment has the potential to dramatically improve the success of studies of epigenetic alterations in disease. Hence, our work characterized methylation of phenotypically important CpG loci across several human tissue types, elucidating interindividual tissue-specific variation in methylation profiles and the contribution of CpG island context to age associated methylation alterations. This work increases our appreciation for the dynamic nature of the epigenome, and begins to define basic tenets to follow in pursuit of both constructing reference epigenomes and elucidating epigenetic alterations truly indicative of disease states.
Using recursively-partitioned mixture modeling and random forests approaches, we differentiated tissues based on CpG methylation profile, consistent with other recent studies conducting genome-wide DNA methylation profiling [8]–[10],[15],[16]. These studies used high resolution methylation data and together have now shown that tissues have distinct methylation profiles. This novel and consistent body of work has, however, not addressed exposures in relation to interindividual variations in methylation. Not only do our findings confirm that tissue-specific epigenetic patterns can be readily defined with a targeted promoter-based CpG array, but they identify target sets of gene-loci most consistently capable of differentiating tissue types.
Factors known to contribute to methylation alterations include carcinogen exposures, inflammation, and diet. Several carcinogen exposures such as tobacco, alcohol, arsenic, and asbestos have been associated with methylation-induced gene-inactivation in various human cancers including bladder cancer, head and neck squamous cell carcinoma, and mesothelioma [28]–[33]. It is therefore reasonable to suggest that various and potentially accumulating exposures throughout life may directly or indirectly lead to methylation alterations and impact disease susceptibility. Carcinogens are well known to induce genetic abnormalities that can lead to clonal selection and expansion in normal appearing tissues (termed “field effect”). Hence, the association of carcinogen exposures with the occurrence of altered methylation at phenotypically important loci may arise as a consequence of altered (“initiated”) clones. Our data suggest that large epigenetic changes occur in normal appearing tissues, and the relationship of these changes to companion genetic changes is of interest to study in the future.
Cancer is a disease of aging, and initial studies of age-related methylation in normal tissues were motivated in large part by studies of methylation in cancer [34]. An early report from Issa et al. described an association between aging colonic mucosa and estrogen receptor methylation [35]. In general, trends of global (repeat element) hypomethylation and promoter hypermethylation found in cancer also have been observed in normal tissues with aging [36]. In recent reports of age-related methylation in normal human prostate and colon tissues, several CpG-island-containing genes were reported to have age-related increases in methylation [18],[19]. Our results confirm these findings and, in addition, document that age-related alterations in these CpG loci are tissue-dependent. More importantly, our examination of loci with previously reported age-associated methylation alterations, in conjunction with reports from others, suggested that the relationship between aging and promoter CpG methylation is complex. For example, using restriction-landmark genome scanning of over 2000 CpG loci in T lymphocytes comparing newborns, middle age, and elderly people, Tra et al. reported that 29 loci had age-related methylation alterations, with 23 loci displaying increased methylation with age and 6 decreasing with age [37]. In addition, measuring intra-individual global methylation changes over >10 years, Bjornsson et al. found both increased and decreased methylation levels dependent on the individual, with over 50% of participants exhibiting >5% change in methylation [20].
Stratifying our data on CpG-island status of loci, we showed that both the direction and strength of correlation between age and methylation were largely dependent upon CpG island status. More specifically, we found a propensity for CpG-island loci to gain methylation with age, and non-island CpGs to lose methylation with age. Our data are consistent with the literature that has demonstrated age-related increases in methylation at gene-loci found within CpG islands [18],[19],[22], as well as the findings of Tra et al. and Bjornsson et al. who showed bi-modal age-related methylation in normal tissues. A direct comparison, by examination of the data of Bjornsson et al., indicated that a high percentage of their top 50 most age-altered loci (all decreases in methylation) are not located in CpG islands; among 24 of 30 autosomal CpGs in their Table 1 (where CpG island status can be identified by readers), only 5/24 (21%) are located in CpG islands, whereas 70% would be expected. Our results from blood samples corroborate their findings, and extend them to demonstrate similar trends in multiple other tissue types, where the strongest negative correlation between age and methylation occurs at CpGs which are not in CpG islands, and the strongest positive correlation between age and methylation occurs at loci in CpG islands.
The observed pattern of age associated methylation was irrespective of tissue-type, suggesting a common mechanism or dysregulation to explain these alterations. Reduced fidelity of maintenance methyltransferases with aging is one potential explanation for age related decreases in methylation; while age related increases in methylation could potentially reflect the accumulation of stochastic methylation events over time. As the examined tissues do not have a pathologic phenotype, methylated CpGs in these cells may not indicate dramatic functional consequences upon gene expression. However, the (in part selective) accumulation of alterations without readily detectable functional consequences should not be interpreted as biologically insignificant. Age-related drift of normal epigenomes without prominent changes in gene expression may nonetheless confer significantly increased risk of conversion to a pathologic phenotype by enhancing both the likelihood and frequency of methylation events that ultimately result in aberrant expression or altered genomic stability. For example, in the context of acquired “non-functional” CpG methylation in the promoter region of an aged individual, continued stochastic methylation events (e.g. “methylation spreading”) increase the chance of methylation induced silencing at that promoter (or silencing of another locus through action at a distance via silencing of other important regions such as enhancers), and hence, progression to a pathologic phenotype. Certainly, this hypothesis is especially plausible for the many diseases of aging. Alternatively, aberrant CpG methylation that silences a gene on a single allele may not appear to have a functional consequence if the complementary allele can provide compensatory expression. As a result, for example, clusters of cellular clones with mono-allelic gene expression could contribute to an increased risk of progression to a pathologic phenotype (e.g. loss of the 2nd functional allele). Future population-based studies addressing the potential of quantifying age and/or exposure associated methylation alterations indicative of disease risk are necessary.
We have provided clear evidence of interindividual variation in tissue–specific methylation related to aging and environmental exposures at disease-relevant CpGs across 10 normal human tissue types. We have demonstrated both general and tissue specific alterations, uncovered a CpG island context-dependent directionality to age associated methylation alterations, and provided a novel path for examining the mechanistic basis of these alterations. By enumerating the methylation status of a panel of cancer-related genes known to stably control transcription in normal tissues, we have also afforded important controls for comparison to diseased tissues, potentially aiding in identification of the most critical alterations in specific diseases and providing more robust targets for novel treatments. Importantly, we have also begun to disentangle the contributions of aging and environmental factors to methylation alterations in normal tissues. Uncovering age and exposure-related methylation changes and their clear contextual dependence is an important contribution to our basic understanding of epigenetic maintenance as it relates to both aging and the pathologic process, provides a potential avenue to pursue clinically useful biomarkers, as well as to identify novel markers of disease susceptibility.
Normal human tissues were assembled by a collaborative, multi-institutional network of investigators conducting molecular epidemiologic studies of human cancer. Tissues were obtained through Institutional Review Board approved studies already underway at these institutions, or were obtained from the National Disease Research Interchange (NDRI, Philadelphia, PA). Briefly, normal brain tissues (n = 12) were contributed by the Wiencke lab at UCSF through the San Francisco Adult Glioma Study [38]. Normal lung tissues (n = 49) were obtained from adjacent non-tumor portions of lung in patients treated for NSCLC [39] or from the NDRI from non-diseased individuals at autopsy (n = 4). Peripheral blood DNA was obtained from controls enrolled in a study of bladder cancer (n = 15) [40], controls enrolled in a study of head and neck cancer (n = 15) [41], and newborn infants (n = 55) [42]. Non-tumorigenic pleural samples (n = 18) were obtained from grossly disease-uninvolved regions of incident mesothelioma [28]. Head and neck anatomic sites (n = 11), bladder (n = 5), kidney (n = 6), and small intestine (n = 5) were obtained from the NDRI, all from individuals with no gross diseases or tumors of the obtained tissues. Non-pathologic placenta samples (n = 19) were obtained as residual tissues from control infant term births as part of an ongoing hospital-based case-control study of intrauterine growth restriction at Women and Infants Hospital in Providence RI. All tissues obtained from patients with disease (lung, pleura) were histologically confirmed as normal by independent study pathologist review of tissue samples prior to DNA extraction.
Fresh frozen tissue and whole blood DNA was extracted using the QIAamp DNA mini kit according to the manufacturer's protocol (Qiagen, Valencia, CA). DNA was modified by sodium bisulfite to convert unmethylated cytosines to uracil using the EZ DNA Methylation Kit (Zymo Research, Orange, CA) according to the manufacturer's protocol. Illumina GoldenGate methylation bead arrays were used to simultaneously interrogate 1505 CpG loci associated with 803 cancer-related genes. Bead arrays have a similar sensitivity as quantitative methylation-specific PCR and were run at the UCSF Institute for Human Genetics, Genomics Core Facility according to the manufacturer's protocol and as described by Bibikova et al [43].
Quantification of cytosine percent methylation was performed by pyrosequencing bisulfite-converted DNA using the PyroMark MD pyrosequencing system (Biotage). Specific pyrosequencing primers were designed to amplify array CpG sites and as many downstream CpGs as conditions permitted (2 to 5 additional) using Biotage Assay Design Software v1.0.6.
All PCR reactions were carried out in 25 µl, utilized Qiagen Hot Star Taq polymerase, 5× Q solution (except FZD9), and 10× PCR buffer with 15 mM MgCl2 under the following conditions: 95°C 15′, (95°C 30″, 45°C (55°C for RARA) 30″, 72°C 1′) × 45 cycles, and 72°C 5′. Final reaction primer concentration for PCR and sequencing was 0.3 µM, primer details are in Table S7. All PCR reactions included a no template control, unmodified DNA control, and 7 standardized percent methylation controls (0%, 15%, 25%, 45%, 65%, 75%, and 100%) derived from Qiagen EpiTect PCR control DNA set samples. Sequencing reactions used 10 µl of PCR product and were run according to instrument/manufacturer protocols (Biotage).
|
10.1371/journal.pntd.0000827 | In Vivo Imaging of Schistosomes to Assess Disease Burden Using Positron Emission Tomography (PET) | Schistosomes are chronic intravascular helminth parasites of humans causing a heavy burden of disease worldwide. Diagnosis of schistosomiasis currently requires the detection of schistosome eggs in the feces and urine of infected individuals. This method unreliably measures disease burden due to poor sensitivity and wide variances in egg shedding. In vivo imaging of schistosome parasites could potentially better assess disease burden, improve management of schistosomiasis, facilitate vaccine development, and enhance study of the parasite's biology. Schistosoma mansoni (S. mansoni) have a high metabolic demand for glucose. In this work we investigated whether the parasite burden in mice could be assessed by positron emission tomography (PET) imaging with 2-deoxy-2[18F]fluoro-D-glucose (FDG).
Live adult S. mansoni worms FDG uptake in vitro increased with the number of worms. Athymic nude mice infected with S. mansoni 5–6 weeks earlier were used in the imaging studies. Fluorescence molecular tomography (FMT) imaging with Prosense 680 was first performed. Accumulation of the imaging probe in the lower abdomen correlated with the number of worms in mice with low infection burden. The total FDG uptake in the common portal vein and/or regions of elevated FDG uptake in the liver linearly correlated to the number of worms recovered from infected animals (R2 = 0.58, P<0.001, n = 40). FDG uptake showed a stronger correlation with the worm burden in mice with more than 50 worms (R2 = 0.85, P<0.001, n = 17). Cryomicrotome imaging confirmed that most of the worms in a mouse with a high infection burden were in the portal vein, but not in a mouse with a low infection burden. FDG uptake in recovered worms measured by well counting closely correlated with worm number (R2 = 0.85, P<0.001, n = 21). Infected mice showed a 32% average decrease in total FDG uptake after three days of praziquantel treatment (P = 0.12). The total FDG uptake in untreated mice increased on average by 36% over the same period (P = 0.052).
FDG PET may be useful to non-invasively quantify the worm burden in schistosomiasis-infected animals. Future investigations aiming at minimizing non-specific FDG uptake and to improve the recovery of signal from worms located in the lower abdomen will include the development of more specific radiotracers.
| Schistosomiasis is a well studied parasitic disease that is far from eradication despite the development of an effective treatment. The lack of an efficacious vaccine and high re-infection rates after treatment are major factors in its intractable worldwide prevalence. A non-invasive imaging technique like positron emission tomography (PET) could give clinicians and researchers a quantitative and visual tool to characterize the worm burden in infected individuals, determine the efficacy of a candidate vaccine, and provide information about parasite migration patterns and basic biology. We are therefore proposing the novel application of PET imaging to schistosomiasis in order to advance the management and research of this infectious disease. Herein, we demonstrate that schistosome parasites take up 2-deoxy-2[18F]fluoro-D-glucose (FDG). FDG uptake in regions adjacent to or within the liver linearly correlate with the worm number in infected mice, but the correlation was stronger in mice with high infection burdens. We anticipate that this research is a first step in the development of more specific radiotracers optimized for schistosomiasis, and will eventually translate to human studies.
| Schistosomiasis is a disease of chronic morbidity affecting the liver, mesentery, and urogenital tract of infected individuals, caused by the parasitic trematode worms, schistosomes. In certain endemic regions like China, the increasing focus on antischistosomal chemotherapy over the past 30 years has dramatically decreased infection rates [1], [2]. However, global prevalence still exceeds 207 million people because of post-treatment re-infection and inadequate control measures, leaving over 779 million people at risk for future infection [3]. In addition, researchers familiar with the true morbidity of schistosomiasis have indicated a greater disability-adjusted life year (DALY) than was previously thought, estimating up to a 70 million DALY burden [4]. Potential resistance to praziquantel, the current treatment standard also became an emerging concern in 1991 when an initial trial in Senegal reported low cure rates, although these results have not been confirmed [2]. A reduced susceptibility to praziquantel among certain naturally occurring Kenyan isolates of Schistosoma mansoni (S. mansoni) was also reported [5].
Eradicating schistosomiasis will require a multifaceted approach that emphasizes the development of new drugs, vaccines [6] and better diagnostic methods [7], [8]. Currently, such advances are stymied by the inability to accurately assess worm burden in infected humans. Stool examination and urine filtration are still the techniques of choice for diagnosis and testing candidate drug efficacy. These methods are unfortunately time-consuming, labor-intensive, costly, and unreliable because of daily variability and poor sensitivity [7], [9]–[12]. For example, when host egg excretion falls below 100 eggs per gram of stool (equivalent to an infection burden below 40 worms), it is increasingly difficult to accurately assess the disease burden [7], [8], [13]. These shortcomings may overestimate treatment efficacy and delay the detection of praziquantel-resistant strains. Immunodiagnostic techniques are becoming increasingly popular because of their high sensitivity and ease of use, but they rely on circulating antibodies, which have a significant time-delay with respect to infection and treatment [7], [14]–[16]. Furthermore, immunodiagnostic measures are of little use in vaccination trials because subjects have elevated antibody levels. In order to expedite research, vaccine development, and diagnosis of schistosomiasis, these challenges necessitate the development of better tools for determining infection burden.
In the past decade, small animal molecular imaging probes have achieved considerable success for non-invasively assessing disease status. Fluorescence molecular tomography (FMT) using near-infrared imaging probes activated by the abundant cathepsins in the schistosomes' digestive tract can quantify the worm burden in murine models of schistosomiasis [17]–[19]. However, this technology currently cannot be applied to human subjects. Positron emission tomography (PET) imaging is useful in measuring cellular metabolism and has emerged as a key non-invasive imaging tool for the diagnosis, staging and evaluation of treatment response for cancer in human patients [20], [21]. We hypothesized that this in vivo molecular imaging technology may also provide a tool to detect and quantify adult helminth parasites.
Bueding performed the first worm physiology experiments in 1950 and found schistosomes to be demanding consumers of glucose, metabolizing their dry body weight in roughly four hours [22]. We reasoned that 2-deoxy-2[18F]fluoro-D-glucose (FDG), a glucose analogue clinically used as a PET radiotracer for glucose metabolism, could be useful for imaging adult schistosomes. The following work assesses the usefulness of FDG PET to quantify S. mansoni worm burden and treatment efficacy in a mouse model of schistosomiasis.
Unless mentioned otherwise, chemicals and reagents were obtained from Invitrogen (Carlsbad, Ca). In vitro uptake experiments in adult schistosomes were conducted using both female and male, paired and unpaired worms perfused from 5-weeks post-infection CD-1 outbred mice (see below for infection protocol). Perfused worms were washed 4 times in 1.5 ml eppendorf tubes with 1x phosphate-buffered saline (PBS) in which glucose was added at a concentration of 1 g/L. Tubes were arranged in triplicate, each with either 1 or 5 worms per tube. Six hundred microliters of Dulbecco's Modified Eagle Medium (DMEM) low glucose medium (1 g/L) containing 150 µCi of FDG was added to each tube and incubations were carried out at room temperature. Clinical grade FDG was supplied by PETNET solutions/Siemens (Knoxville, TN) with a specific activity larger than 2,000 Ci/mmol. After one hour, each tube was washed four times using cold 1x PBS with 1 g/L of glucose. The tubes were then placed in a Wallac 1282 CompuGamma well counter (Perkin Elmer Life Science; Waltham, Ma) to measure gamma rays emitted from 18F decay in counts per minute (cpm). In some experiments, 20 µl of 10% sodium iodide and 5% iodine was added to 250 µl of PBS containing single adult worms for 5 minutes. The killed worms were washed twice in PBS prior to adding the FDG.
Six-week old female athymic nude mice (nu/nu) were anesthetized using a 500-µl solution of avertin (1∶1 tert-amyl alcohol:tribromoethanol; Sigma Aldrich; St-Louis, MO) diluted to 2.5%. Mice were infected percutaneously with the aim to obtain a wide range of parasite burdens. The animals were kept in isolation facilities for 5–6 weeks until imaging. To minimize background fluorescence during FMT imaging studies, mice were switched to a low-fluorescence diet (Harlan laboratories; Indianapolis, IN) three days prior the experiments. Mice were deprived of food overnight before imaging experiments. All animal studies were conducted in accordance with the regulations and guidelines set forth by the Institutional Animal Care and Use Committee (IACUC) at Case Western Reserve University.
Eight infected and two uninfected (control) mice were injected in the tail vein with 2 nmol of Prosense 680 (Visen Medical Inc.; Bedford, MA) diluted in 100 µl of sterile 1x PBS. Twenty-four hours after injections, mice were anesthetized with isoflurane (Aerrane; Baxter; Deerfield, IL), placed in imaging cassettes and imaged on a FMT 2500 imaging system (Visen Medical Inc.; Bedford, MA) using the 685 nm channel. Anesthesia was maintained with 1.5% isoflurane and oxygen (1 L/min) flowing within the imaging chamber throughout the acquisition. After FMT imaging, wells on the half of the imaging cassette were filled with a solution of 0.05% magnevist (gadopentetate dimeglumine; Bayer Healthcare Pharmaceuticals; Montville, NJ) diluted in distilled water for co-registration purposes. The cassette was transferred to a magnetic resonance imaging (MRI)-compatible bed and imaged on a Biospin 7.0T MRI scanner (Bruker; Billerica, MA). T2-weighted images using a multi-slice multi-echo (MSME) sequence were acquired, producing a set of coronal images (repetition time (TR) = 1250 ms, echo time (TE) = 15 ms, flip angle = 90°, voxel size = 0.0312×0.0312×0.05 cm). Anesthesia was maintained as described above. The animals were allowed to recover from anesthesia in their cages for at least one hour before microPET (μPET) imaging. FMT and MRI images were co-registered using COMKAT, a Matlab-based general-purpose compartment modeling software tool [23].
Thirty-five infected and four uninfected (control) mice were used in the μPET imaging studies (four of the infected mice were scanned twice at an interval of three days and ten other mice had been scanned with FMT and MRI). One hundred minutes before μPET imaging, mice were anesthetized with isoflurane mixed with oxygen via isolated chambers, titrating to a respiratory rate of one breath per second. Periorbital injections of FDG dissolved in 100 µl of sterile saline were administered 90 min before imaging at a target dose of 25 µCi per gram of body weight. Mice were kept under anesthesia with isoflurane until μPET imaging to minimize muscular uptake of the tracer. Twenty minutes before imaging, a 26 gauge fluorinated-ethylene-propylene (FEP) monoject veterinary I.V. catheter coated with an inert lubricant was inserted into the bladder through the urethra. Urine was removed 5 min prior to imaging with the help of gravity to minimize imaging artifacts caused by high FDG concentration in the bladder. μPET imaging was carried out using a Siemens Concord R4 microPET system (Siemens Solutions, Knoxville, TN). The imaging protocol consisted of a 45-min emission scan followed by a 10-min transmission scan using the 57Co point source for attenuation correction. The images were reconstructed using a 2D ordered subset expectation maximization (OSEM) algorithm, yielding a volume of 128×128×63 voxels with a voxel size of 0.85×0.85×1.21 mm and a spatial resolution of approximately 1.84 mm [24]. The μPET scanner was calibrated prior to the imaging studies using a phantom with a known radioactivity concentration measured by a dose calibrator. Verification scans confirmed that the relative error in quantifying the true radioactivity (as measured with a dose calibrator) in a phantom by measuring it using μPET imaging was less than 5%.
Six infected mice were treated with praziquantel using oral gavage. Praziquantel was diluted in 1x PBS, and titrated to 250 mg/kg body mass. Treatments were administered 72, 48, and 24 hours prior to re-imaging. μPET imaging was carried out as described above before and after treatment.
Perfusions were conducted using a 5-ml syringe of citrate solution infused through the right heart ventricle after mice were euthanized with a sodium pentobarbital cocktail. The portal vein was ruptured and the perfusate was collected in a sterile culture plate. Worms and worm pairs were counted and transferred to a 1.5 ml eppendorf tube. Background radioactive signal was washed from worms using an isotonic glucose solution before measuring the radioactivity levels by well counting.
To localize schistosomes in our mouse model of S. mansoni infection, two infected mice (one of which had been imaged with FMT, MRI and μPET) were fixed in OCT, flash frozen in liquid nitrogen and stored at −80°C. The frozen blocs were imaged using a cryomicrotome imaging system developed in house [25]. Bright field color images were collected at 15.6-µm resolution and the thickness of each section was 40 µm.
One infected mouse died approximately 60 min after FDG injection and was therefore not imaged. The liver and digestive tract were rapidly removed and exposed for 30 min to a Molecular Dynamics mounted storage phosphor screen (General Electric Healthcare; Piscataway, NJ) for autoradiography. After exposure, the phosphor screen was imaged with a Typhoon imaging system (General Electric Healthcare; Piscataway, NJ).
The livers of 6 mice were removed after perfusion and kept overnight in paraformaldehyde. The livers were transferred to a 30% sucrose solution overnight. Fixed tissues were embedded in OCT and kept at −80°C. Several 10 µm-thick sections from different areas of the liver were mounted on glass slides using a CM 3050 S cryostat (Leica; Bannockburn, IL) for hematoxylin and eosin (H&E) staining. Perfused worms were also directly mounted on glass slides after FMT and PET imaging. A DM 4000B microscope (Leica; Bannockburn, IL) equipped with a Cy5 filter (663–738 nm band-pass) was used to image all sections. Exposure times for the acquisition of near-infrared images of Prosense 680-labeled worms were set to 50 ms and bright field images were acquired to register fluorescent images with soft tissues using the QCapturePro software (QImaging Corporation; Surrey, BC, Canada).
Three-dimensional regions of interest (ROIs) were drawn around the lower abdomen of the mice imaged with FMT to exclude background signal from the liver. Voxels with values greater than two times the average signal in the lower abdomen of the control mice were used for quantification as previously described [17]. Adjusting the display intensity when needed, ROIs were manually drawn around the portal vein and, if present, around areas of increased FDG uptake within the liver parenchyma using ASIPro VM 6.3.3.0 (Siemens Solutions, Knoxville, TN). In some cases, the internal organs had shifted as a result of animal manipulations prior to imaging and the location of the ROIs was determined using the best judgment from the user. Areas of high FDG uptake within the small intestines or colon as determined by careful examination of multiple adjacent sections were excluded. All radioactivity measurements were decay-corrected to the scan start time for comparison. Statistical data were generated using Origin 8.1 (OriginLab; Northampton, MA). Unpaired and paired t-tests were used to examine the differences between groups in the in vitro uptake and praziquantel treatment experiments, respectively. P-values≤0.05 were considered significant. The square of the Pearson's product-moment correlation (R2) was used to determine correlation between measured signals and worm burden.
In vitro uptake of FDG was examined for adult stage S. mansoni (Figure 1). The uptake in single iodine-killed worms was significantly lower than that in untreated single adult worms (P = 0.04). The total uptake by 5 worms was 3.9 times greater than that by single worms (P = 0.002). Based on the 60-min incubation experiments with 5 worms, the average radioactivity per adult worm was 54 nCi, which would be sufficient for detection by μPET imaging.
Whole animal FMT and MRI scans were performed in two control and eight infected mice (5–6 weeks after infection) and the images were co-registered to provide an anatomical context for the FMT signal. Twenty-four hours after Prosense 680 injection, moderate and high fluorescence levels were observed in the mid-section of the abdomen in control and infected mice (Figure 2A). Anatomical information from MRI confirmed that the most intense fluorescence signals were originating from regions within or adjacent to the liver in infected mice. Liver expresses relatively high levels of the proteases targeted by Prosense probes. Signals of low, but variable intensity were also observed in the lower abdomen of all mice. Worms were collected after imaging and placed on a glass slide. Microscopy demonstrated high fluorescence in the digestive tracts of perfused schistosomes (Figure 2B). Adult S. mansoni parasites are abundant in cathepsins, which supports the basis for imaging schistosomiasis in vivo with a fluorescent probe activated by proteases. Total probe accumulation in the lower abdomen regions was previously found to correlate with worm burden [17]. We did not find a linear correlation between probe levels in the lower abdomen and worm burden (R2 = 0.009, P = 0.80, n = 10; Figure 2C). However, when only considering mice with less than 60 worms (a range comparable to that previously reported [17]) and excluding control mice from the analysis as was done in [17], a linear correlation between probe levels and the lower abdomen and the number of worms was observed (R2 = 0.67, P = 0.09, n = 5; Figure 2C, inset).
We next performed experiments to determine if FDG-μPET imaging would allow noninvasive quantification of adult S. mansoni in infected mice. Four control and 35 infected mice were included in these studies. These included a subset of ten animals (2 control and 8 infected mice) that also underwent FMT imaging as described above. One of the infected mice died approximately 60 min after FDG injection and could not be imaged. In another case, the animal (high infection burden) died before FDG injection so the liver was removed and processed for histology. Four mice were scanned twice at 3-day intervals and one mouse was not perfused, but fixed in OCT after μPET imaging for whole body cryomicrotome imaging, a technique that allows microscopic examination of the entire or regional areas of the interior of the mouse [25]. Forty datasets were thus included in the quantitative studies. Nineteen mice had an infection burden below 40 worms while 17 mice carried between 50 and 242 worms. Intense radioactive signals were observed in the heart, kidneys, bladder and regions of the lower abdomen in control and infected mice (Figure 3A, arrowheads). Autoradiography of the liver and digestive tract in the mouse that died 60 min after FDG injection revealed that the areas of high radioactivity accumulation in the lower abdomen corresponded mainly to the colon and, to a lesser extent, the small intestines (Figure 3B), while the radioactivity levels in the liver were relatively low (not shown). Moderate radioactivity accumulation was observed in small regions inferior to the liver, slightly anterior to the kidneys and superior to the colon (Figure 3A, arrows). In a few mice, abnormal radioactivity accumulation was found in some regions of the liver (Figure 3A, arrows).
Cryomicrotome imaging was carried out after FDG-μPET scans in one mouse with a low (which did not undergo FDG-μPET scanning) and one mouse with a high infection burden. The worms could be readily visualized by their white soft tissue appearance and dark digestive tract containing heme pigment. In the mouse with a low infection burden, only a few worms were seen in the liver parenchyma. We did not find any schistosomes in the mesentery (not shown), but this was only in one animal. In the mouse with a high infection burden, a high number of worms were found in the common portal vein (Figure 3C, arrows). Only a few worms were seen in the liver and scattered in the mesentery (Figure 3D, arrows). H&E sections of the liver of the mouse that died before FDG injection (which had a high target infection burden) confirmed the presence of worms in branches of the portal vein in the liver (Figure 3E). The livers from five other infected mice were also processed for histology after perfusion. Examination of the H&E sections revealed the absence of worms in the liver parenchyma. Ova trapped in the liver parenchyma did not induce an overt inflammatory response (Figure 3F) indicating that regions of abnormal FDG uptake in the liver are unlikely due to granulomatous response to ova. The absence of worms from the mesentery after perfusion was confirmed by careful visual inspection.
Because of high background uptake in the colon and small intestines, three-dimensional ROIs were only drawn around the common portal vein, inferior to the liver and in areas of increased uptake in the liver, if present (Figure S1). A linear correlation between the total radioactivity within ROIs and the number of worms was observed (R2 = 0.58, P<0.001, n = 40; Figure 4A). The background FDG uptake in the common portal vein (y-intercept) was approximately 110 nCi. An even stronger linear relationship between total radioactivity within ROIs and the number of worms was observed in mice with more than 50 worms (R2 = 0.85, P<0.001, n = 17; Figure 4A, inset). The radioactivity levels in the schistosomes from 21 of the infected mice were directly measured by well counting after perfusion. There was a strong linear relationship between the measured radioactivity levels in perfused worms and the number of worms (R2 = 0.85, P<0.001, n = 21; Figure 4B). The total radioactivity within ROIs correlated with the measured radioactivity levels in perfused worms (R2 = 0.74, P<0.001, n = 21; Figure 4C). However, the radioactivity measured by μPET was in general less than that in the perfused worms.
Praziquantel treatment studies were finally performed to investigate the usefulness of FDG-PET for assessing treatment efficacy (Figure 5A). Four infected mice did not receive treatments and were scanned twice, at three-day intervals, serving as positive controls (* and **; Figure S1). Over the three-day period, FDG uptake in the area of the common portal vein increased by 36% on average in untreated mice (P = 0.052; Figure 5B & D). Comparison of uptake in the area of the common portal vein in treated mice showed a 32% average decrease in total radioactivity over the three-day treatment period, although this did not reach significance (P = 0.12; Figure 5C & D). The number of worms reported in Figure 5B and C were determined as described above by perfusing mice after μPET imaging on day 4. The viability of the worms was qualitatively assessed by trypan blue exclusion. Worms recovered from the mouse that exhibited the largest decreased in FDG uptake in the common portal vein (77%) were dead (did not exclude trypan blue; mouse with 20 worms in Figure 5).
Our current understanding of the S. mansoni migration cycle in the host is based on earlier studies in which radiolabeled schistosomulae were injected into mice and tracked with autoradiography [26]. Only in recent years have noninvasive imaging technologies become available for small animal studies of schistosomiasis. FMT was recently applied to schistosomiasis to quantify adult worm burden [17]. While the results showed that the number of worms in infected mice can be quantified using near-infrared probes activated by proteases, only a limited range of worm burdens was assessed and optical imaging techniques currently cannot be applied in humans. In this work, we investigated the possibility of PET imaging of S. mansoni in mice as this technology is readily applicable to humans. Here we show that FDG uptake in adult parasites both in vitro and in vivo correlates with parasite numbers. We also demonstrate that adult worms can be imaged in vivo using μPET to assess worm burden. However, our ability to quantify worm burden by FDG-μPET was limited in mice with low worm burden.
In vitro FDG uptake studies carried out on adult schistosomes first showed that FDG uptake in five adult worms was greater than that in single adult worms. Well counting measurements also showed that the radioactivity levels in worms recovered from infected mice after FDG μPET imaging strongly correlated with the number of worms. These results supported the hypothesis that FDG uptake quantified by μPET imaging may correlate with the worm burden. Furthermore, the lower FDG uptake in iodine-killed single worms compared to untreated single worms suggested that FDG-μPET may be useful in assessing the outcome of treatments such as praziquantel. The residual radioactivity associated with dead worms in Figure 1 likely represents the inability to remove the unmetabolized FDG following repeated washings.
To supplement the μPET data, we quantified worm burden by assaying protease activity using FMT. We first imaged infected mice 24 hours after Prosense 680 injection (Visen Medical Inc.; Bedford, MA). The images revealed high fluorescence signal coming from the liver of all mice, including uninfected animals. This was observed in a previous study unrelated to schistosomiasis comparing small molecular probes with the larger Prosense 680 and 750 probes and may be caused by the hepatic clearance of the released fluorochromes [27]. Furthermore, high levels of the proteases that activate the Prosense probes used are expressed in the liver. Because of this high background in the liver, only probe accumulation in the lower abdomen was quantified. A linear correlation between total probe amount and worm burden was observed, but only in mice with less than 60 worms. Similar results were demonstrated in another study, although only animals with less than 35 worms were included [17]. This demonstrates the presence of adult worms in the mesenteric vasculature, which was confirmed at the time the worms were recovered. The poor correlation in animals with higher worm burdens likely arises from the presence of worms retained in the common portal vein and adjacent liver parenchyma.
In contrast to FMT, FDG imaging showed relatively low FDG uptake in the liver but considerable and variable FDG uptake in the lower abdomen of all mice, regardless of the worm burden (Figure S1). Autoradiography showed that most of the radioactivity in the lower abdomen was in the colon and was not associated with schistosomes. Focal FDG uptake in normal gut can appear as false lesions in PET scans of the abdomen in humans [28]–[30]. These foci of abnormal FDG uptake can be caused by metabolism in the gut flora or by peristalsis for example. In heavily infected animals, FDG uptake in the lower abdomen may also be associated with innate inflammatory response either from products released by worms or colonic flora from compromised integrity of the gut wall. High uptake in the gut of control (uninfected) animals and low uptake in the gut of some infected animals suggests that gut motility and metabolism and a moderate immune response to schistosomiasis infection may all be factors that contribute to FDG uptake not associated with worms in the lower abdomen.
Moderate FDG uptake was observed in regions inferior to or within the liver and total radioactivity levels within these regions appeared to increase with the worm number. A linear correlation between total radioactivity within these regions and the number of worms was demonstrated across a wide range of infection burdens (0–242 worms). This correlation was especially strong in mice with high worm burden, suggesting that in heavily infected animals, most worms remain in the common portal vein and in the liver parenchyma. This was demonstrated by cryomicrotome imaging and by dissection at the time of worm recovery. We speculate that in mice with high infection burden, adults mature more slowly and thus, proportionally fewer worm pairs migrate into the mesenteric vasculature at the time the imaging studies were carried out (5–6 weeks after infection). FMT imaging, cryomicrotome imaging and direct observations at the time of perfusion showed the presence of worms in the mesenteric vasculature in heavily infected animals. However, worms in the mesenteric vasculature are in insufficient numbers to account for the large uptake of FDG observed on the PET images. The low levels of radioactivity associated with worms in the mesenteric vasculature could not be accurately and consistently detected against the high FDG uptake in the lower abdomen of some mice.
To further validate the μPET results, animals were treated with praziquantel, an anti-schistosomal drug that targets adult worms. The FDG signal measured in treated animals decreased after three days of treatment while that in untreated animals increased over the same period. The variability in change of FDG uptake in the region of the common portal vein in response to praziquantel is likely due to the use of athymic nude mice that can result in impaired parasite killing since praziquantel treatment efficacy has been shown to be dependent upon an intact host immune response [31]. Indeed variable efficacy of drug treatment was noted in trypan blue exclusion studies performed on the worms after perfusion. The analysis of the results is further complicated by the fact that the number of worms prior to the start of praziquantel treatment is unknown and thus, treatment efficacy cannot be accurately assessed. Nevertheless, these results suggest that, with some improvements in the technique and recovery of worms in the lower abdomen, PET imaging may be useful to assess treatment outcome using an individual as its own control and ROI-based analysis to assess the number of worms after treatment.
The high non-specific FDG uptake in the colon and, to a lesser extent, small intestines, that limited our ability to detect schistosomes in the mesentery in infected mice may not be such an issue when performing FDG-PET imaging studies in human patients with schistosomiasis. Methods such as colon cleansing by isosmotic solution taken the evening prior to examination, intravenous hydration with 0.45% saline and cleansing of the urinary tract and bladder during scanning were shown to yield artifact-free PET images in humans [30]. Such methods could be applied to decrease non-specific uptake in the gut and allow better detection of worms in the mesentery. Another potential background issue for these studies is that activated immune cells in granulomas formed in response to ova deposition may also uptake FDG and affect quantification of the worm burden in humans. T-cell-deficient mice were used in these experiments to avoid an immune response from the host. Previous studies reported no statistically significant difference in size between worms developing in nude mice and controls [32]. Nude mice also exhibited suppressed granuloma formation and reduced morbidity, further justifying our choice of animal model [33]. We imaged all mice 5–6 weeks post-infection, at a point in development when some worms begin egg release and most worms have matured to full adult stage. Performing these experiments in immunocompetent mice should be achievable since studies conducted in immunocompetent rats showed that treatment with a single dose of methylprednisolone resulted in decreased FDG uptake in granulomas [34]. Such a strategy could be used to distinguish between inflammatory lesions and schistosomes on FDG-PET images.
Finally, the μPET scanner used in our experiments has a spatial resolution of approximately 1.84 mm, but other small animal scanners have reported spatial resolutions as low as 0.7 mm for 10-cm detector ring radius [35], [24]. Adult schistosomes range from 0.6–1.1 cm in the thicker males (0.07 cm in diameter) and 1.2–1.6 cm for the thinner females (0.016 cm in diameter) [36]. Human PET scanners currently have a spatial resolution of approximately 5 mm, but scanners with 2 mm isotropic spatial resolution across the field of view are available [37]. While the spatial resolution of animal or human PET scanners may not be high enough to resolve individual worms, the high sensitivity of this imaging modality may allow non-invasive quantification of the worm burden in infected mice and man through region-based quantitative analysis.
In conclusion, these studies verified the hypothesis that high glucose metabolism in S. mansoni allows for detection with FDG-μPET and quantification of the disease burden in vivo. Because of the possibility to apply PET technology to humans to support the development of new diagnostic tools and for vaccine research, we believe FDG-PET imaging of schistosomiasis should be further investigated. Studies moving forward will include efforts to minimize non-specific radiotracer uptake in the gut (e.g. using antibiotic cocktails, laxatives or combinations of both) and FDG-μPET imaging studies in immunocompetent rodent models infected with schistosomiasis. Finally efforts are already underway to develop radiotracers more specific for parasites, including radiolabeled praziquantel.
|
10.1371/journal.ppat.1006631 | The bacterial virulence factor CagA induces microbial dysbiosis that contributes to excessive epithelial cell proliferation in the Drosophila gut | Gut microbiota facilitate many aspects of human health and development, but dysbiotic microbiota can promote hyperplasia and inflammation and contribute to human diseases such as cancer. Human patients infected with the gastric cancer-causing bacterium Helicobacter pylori have altered microbiota; however, whether dysbiosis contributes to disease in this case is unknown. Many H. pylori human disease phenotypes are associated with a potent virulence protein, CagA, which is translocated into host epithelial cells where it alters cell polarity and manipulates host-signaling pathways to promote disease. We hypothesized that CagA alone could contribute to H. pylori pathogenesis by inducing microbial dysbiosis that promotes disease. Here we use a transgenic Drosophila model of CagA expression to genetically disentangle the effects of the virulence protein CagA from that of H. pylori infection. We found that expression of CagA within Drosophila intestinal stem cells promotes excess cell proliferation and is sufficient to alter host microbiota. Rearing CagA transgenic flies germ-free revealed that the dysbiotic microbiota contributes to cell proliferation phenotypes and also elicits expression of innate immune components, Diptericin and Duox. Further investigations revealed interspecies interactions are required for this dysbiotic CagA-dependent microbiota to promote proliferation in CagA transgenic and healthy control Drosophila. Our model establishes that CagA can alter gut microbiota and exacerbate cell proliferation and immune phenotypes previously attributed to H. pylori infection. This work provides valuable new insights into the mechanisms by which interactions between a specific virulence factor and the resident microbiota can contribute to the development and progression of disease.
| Microbial communities in the gut, termed microbiota are important for human health, and when altered can sometimes promote disease. Infections, such as with the cancer-causing bacterium Helicobacter pylori, can cause altered gut microbiota, but why these alterations occur and whether the altered communities contribute to disease remain unknown. Here, we use Drosophila expressing the H. pylori disease-causing protein CagA, to model this virulence factor’s effect on host pathology and microbiota. We found that expression of CagA in the Drosophila gut causes excessive cell proliferation and immune activation, hallmarks of H. pylori infection. Notably, these traits did not occur when flies were reared in the absence of microbes. Further examination reveals that CagA-expressing flies have an altered gut microbial community that is sufficient to promote cell proliferation even in normal flies. This proliferative activity required the presence of two interacting bacteria, illustrating a new model for disease-promoting microbiota. This work demonstrates how a bacterial protein can cause disease indirectly through altering the microbial ecology of the host, and it suggests future treatments for infections that rely on manipulating the microbiota to mitigate disease pathology.
| Gut microbiota, a complex consortium of bacteria, archaea, viruses and eukaryotes found in the gut, play critical roles in human health. The microbiota is known to facilitate nutrient acquisition, confer resistance to pathogens, and contribute to developmental processes [1]. Pathologic alterations in microbial community composition, termed dysbiosis, result in community dysfunction that is linked to human diseases including inflammatory bowel disease, obesity, pathogen infection and colon cancer [2–5]. Indeed, dysbiotic gut microbiota have the ability to alter immune surveillance mechanisms, and promote proliferation and chronic inflammation within the gastrointestinal tract, processes that play key roles in carcinogenesis [6].
Our current understanding of the mechanisms of microbiota maintenance or induced dysbiosis is limited, but host, environmental, and microbial factors can all influence microbiome structure and interactions [5]. For example, host factors such as chronic inflammation [4] and decreased gut motility [7] modulate microbial community composition to promote disease. Environmental perturbations such as diet and antibiotic usage can allow overgrowth of single pathobionts, such as C. difficle [8]. Conversely, infection with single pathogens can result in altered microbiota composition [3]. For example, Salmonella enterica serovar Typhimurium promotes and thrives on enteric inflammation, creating an environment that also favors other inflammation-adapted Proteobacteria. Alteration in community composition can also enhance virulence of pathogenic microbes. A notable example comes from mice where microbiota transplants from a susceptible population to a previously unsusceptible population conferred susceptibility to the pathogen C. rodentium [9].
Helicobacter pylori is a bacterial resident of the human stomach that in half the world’s population promotes gastritis and an increased risk for gastric cancer development. Infection with H. pylori is associated with shifts in the gastric and colonic microbiota [10], but the extent to which H. pylori-induced dysbiosis contributes to disease is not known. The bacterial-intrinsic mechanisms by which H. pylori exhibits its oncogenic potential are thought to be largely through expression of a potent virulence protein, cytotoxin-associated gene A (CagA), which is injected into the cytoplasm of host gastric epithelial cells [11]. Upon gaining entry to host cells, CagA modulates multiple host pathways such as the Ras/ERK MAPK pathway, resulting in disruption of receptor tyrosine kinase signaling and promotion of cell proliferation [12]. Additionally, CagA activates inflammatory processes through the immune effector NF-κB, which promotes expression of pro-inflammatory cytokines and alters the host response to infection [13].
Profiling of gastric and colonic microbiota in H. pylori-infected humans and rodent models demonstrates that this bacterium has a profound effect on resident microbial communities [10,14]. In a mouse model of H. pylori infection, composition of microbiota prior to H. pylori colonization impacted disease severity and progression, suggesting gastric microbiota modulate H. pylori virulence and contribute to gastric disease [15]. Similarly, H. pylori mono-associated INS-GAS mice showed significantly delayed tumor development and less severe gastritis than when reared with conventional microbiota and H. pylori [16], again suggesting gastric microbiota play a role in disease pathogenesis. Although the direct role of H. pylori mediated gastric disease has been well studied, mechanisms by which the H. pylori-associated gastric microbiota promote disease remain unexplored.
Based on these findings, we hypothesized that the H. pylori virulence protein CagA contributes to pathological proliferation and promotes carcinogenesis via modulation of gastric microbiota, which we tested in a transgenic Drosophila model of CagA expression. Here we took advantage of the relative simplicity of the Drosophila midgut epithelia, microbial community, and genetics to transgenically express CagA within the adult midgut epithelium to investigate the potential effect of CagA expression on resident microbiota. The fruit fly midgut shares many similarities with the mammalian gastrointestinal tract in its tissue organization and programs of homeostasis [17]. Both tissues are continually renewing epithelia with stem cell populations that give rise to transit amplifying cells (called enteroblasts in Drosophila) that then differentiate into absorptive or secretory cell populations. Also like the mammalian digestive tract, the fruit fly gut contains specialized compartments, including an acidified middle midgut that contains specialized H+/K+-ATPase-expressing, acid-secreting Copper cells that function similarly to the parietal cells of the human stomach. In addition to similarities in tissue architecture, many of the same molecular pathways, including Wnt and Notch signaling, regulate these programs of epithelial renewal in flies and mammals. Using the Drosophila model we recapitulated known host cellular phenotypes of CagA expression and explored, for the first time, a role for resident microbiota as a contributor to pathological proliferation mediated by CagA.
We previously developed the Drosophila model of transgenic expression of CagA to elucidate the biochemical and cellular interactions that facilitate pathogenicity of CagA in model epithelial tissues. Specifically our analysis of transgenic expression of CagA in Drosophila photoreceptors revealed CagA’s function as a prokaryotic mimic of the Grb2-associated binder (Gab) adaptor protein that activates SHP-2, a component of receptor tyrosine kinase pathways [18]. Additionally, transgenic expression of CagA in the Drosophila wing and eye imaginal discs revealed that CagA triggers JNK pathway activation and acts to enhance tumor growth and metastasis generated by activated RAS [19]. Here we use transgenic expression of CagA in the adult midgut epithelium to investigate the affects of CagA on the gut epithelium and host-associated microbiota.
CagA has been shown in multiple studies to promote epithelial cell proliferation, and its activity seems to be especially potent in cells with more stem-like characteristics [20]. Recent characterization of H. pylori micro-colonies in human gastric glands found H. pylori closely associated with proliferating progenitor cells [21]. Furthermore, experimental infection of mice revealed H. pylori dependent expansion of Lgr5+ gastric stem cells, with greater expansion occurring with CagA positive H. pylori [21], suggesting that translocation of CagA into stem cells stimulates their proliferation. Using a transgenic model of CagA expression within the adult Drosophila midgut epithelium we examined how expression of CagA within either the intestinal stem cells (ISCs) and their immediate progenitor enteroblasts (EBs) or in the more numerous nutrient absorbing enterocytes (ECs) affects cell proliferation phenotypes. We drove expression of a UAS-CagA transgene [18] along with the UAS-GFP reporter in the ISC and EB population of the adult midgut using the escargot-Gal4 driver. We found expression of CagA in this population resulted in high rates of cell proliferation compared to the control (P<0.0001) (Fig 1A and 1B). The observed proliferation in stem cell populations appeared to be dependent upon phosphorylation of CagA, as a non-phosphorylatable version of CagA (CagAEPISA) [18] showed proliferation rates similar to that of controls (Fig 1A). In contrast to the effect of CagA expression in stem cells, transgenic expression of CagA or CagAEPISA in enterocytes, using the enterocyte specific driver Myo1A-Gal4, did not induce elevated proliferation in the midgut epithelium and instead was significantly lower than controls (P<0.0001) (Fig 1C). These data show expression of CagA within stem cells, but not enterocytes, of the gut epithelium is sufficient to promote excess cell proliferation in the Drosophila midgut.
CagA positive strains of H. pylori have been shown previously to be sufficient to alter host immune signaling pathways in human patients and animal models [12]. To determine whether expression of CagA in the Drosophila midgut is sufficient to alter host immune signaling pathways, we assayed activation of the Drosophila intestinal NF-kB pathway, the immune deficiency (IMD) pathway, by measuring expression of pathway-specific target genes encoding antimicrobial peptides, Diptericin, Attacin and Defensin [22,23]. We found the antimicrobial peptide, Diptericin, is overexpressed an average of 8-fold in CagA transgenic flies compared to the control (P<0.0001) (Fig 1D). This was not the case when we expressed the CagAEPISA transgene suggesting immune activation is dependent upon phosphorylation of CagA (S1 Fig). We found both Attacin and Defensin were similarly expressed in controls and CagA or CagAEPISA transgenic flies (S1 Fig), suggesting the overexpression of Diptericin is a specific response to expression of CagA and not the result of non-specific immune activation.
Reactive oxygen species (ROS) have been shown to play an important role in carcinogenesis by inducing DNA damage and promoting proliferation [23,24]. Additionally, CagA-positive strains of H. pylori have been shown to elicit production of ROS [25]. Therefore, we suspected that similar to its up-regulation of Diptericin, CagA may also activate expression of Drosophila duox, a dual oxidase enzyme responsible for translocation of ROS into the midgut lumen [26]. To test this possibility we assayed expression of duox by qPCR and found nearly 4-fold higher expression of duox in CagA transgenic flies compared to the control (P<0.05) (Fig 1E). Taken together, we conclude that transgenic expression of CagA in the stem cell population of the adult Drosophila midgut is sufficient to specifically activate immune signaling pathways previously shown to be associated with H. pylori infection [27].
Because over-activation of immune signaling pathways is known to be detrimental to long-term survival, we assayed survival of transgenic Drosophila to determine if expression of CagA within ISCs affected life span. We found CagA transgenic flies survive similarly to control flies (S1 Fig), indicating expression of CagA within the midgut stem cells has no net negative effect on whole animal survival. Interestingly, flies expressing the CagAEPISA transgene in two independent lines survived significantly longer than CagA-expressing or control flies. This may indicate a previously unappreciated phosphorylation-independent function of CagA.
Dysbiosis of GI microbiota is well recognized as a contributor to both cell proliferation and inflammatory processes known to facilitate the development and progression of intestinal and colonic cancers [4,5]. The gastric microbiota of individuals with H. pylori infection and gastric cancer is altered from that of control patients [10]. However, whether dysbiosis of gastric microbiota is capable of inducing or contributing to proliferation and disease progression in patients infected with H. pylori remains uncertain. Additionally, it is unclear whether CagA alone is sufficient to alter microbiota or whether this process would also require H. pylori itself. To address this question we took advantage of our transgenic model of CagA expression, where no H. pylori is present, and assayed microbiota of CagA-expressing and control adult Drosophila. We dissected the adult midgut of 7-day-old Drosophila and plated the contents on MRS agar, a modified nutrient agar commonly used for growth of Drosophila gut isolates. Plating revealed that microbiota of the control flies was completely dominated by a single bacterial isolate, whereas microbiota of CagA transgenic flies was made up of this same colony type and another distinct colony morphology. We isolated individual colonies and determined the sequences of their 16S rRNA genes. This analysis identified the single bacterial isolate from control flies as Acetobacter pasteurianus (Ap) of the Acetobacteraceae family (Fig 2A). CagA transgenic flies also contained Ap and a second distinct colony type identified as Lactobacillus brevis (Lb) of the family Lactobacillaceae (Fig 2B). Both of these species of bacteria have been previously described as common inhabitants of Drosophila intestines in both lab-reared and wild populations [28]. Based on the observed differences in microbiota composition in CagA transgenic flies, we conclude that CagA flies harbor an altered gut microbiota from that of control flies. Additionally, flies expressing the non-phosphorylatable CagAEPISA contained communities dominated by Ap (Fig 2C), however, 50% of the flies assayed contained some Lb but at much lower levels than were ever observed in CagA transgenic flies. Occasionally we detected additional taxa identified as L. plantarum and A. tropicalis in both control and CagA flies, however the presence of those microbes was inconsistent between experiments and also found at variable rates even within flies from the same bottle. Due to their low abundance and inconsistency, we focused on the presence and abundance of Ap and Lb for our subsequent analysis. Representative strains A. pasteurianus DORAp21 (Ap21) and L. brevis DORLb22 (Lb22) were collected from CagA transgenic flies reared under standard conditions and used in subsequent experiments.
The Drosophila midgut microbiota is necessary for development and normal homeostatic cell proliferation [29,30], similar to vertebrate gut microbiota [1,31,32]. We therefore asked whether the altered microbiota of CagA transgenic flies could contribute to the excessive cell proliferation observed in these animals by rearing them under germ-free (GF) conditions and assaying cell proliferation. As has been reported previously [33], we found cell proliferation to be slightly higher in conventional (CV) control flies compared to those reared GF (Fig 2D). In the CagA transgenic flies reared GF, we observed a level of cell proliferation intermediate between CV CagA flies and GF control flies (Fig 2D). This suggests that microbiota of CagA transgenic flies contributes to a portion of the total increased cell proliferation observed in CV CagA flies, as compared to CV controls whereas the remaining portion of proliferation is directly due to expression of CagA. We also found that phosphorylation of CagA is required for this microbiota-dependent portion of proliferation, as CagAEPISA flies reared GF showed similar reduction in proliferation to that of GF controls (Fig 2D). These data suggest that the proliferation of midgut cells observed in CagA transgenic flies is multifactorial: expression of CagA induces cell autonomous proliferation and the altered microbiota promotes cell non-autonomous proliferation.
Next we wanted to determine if expression of CagA within the midgut epithelium was responsible for shaping the altered community observed in CagA transgenic flies. Drosophila microbiota can be variable and significantly affected by environmental exposure and access to microbial isolates [34]. We aimed to determine whether, given the same inoculum of bacteria, the CagA expressing intestines would assemble a different community from the control intestines. To test this, we derived flies GF and then immediately exposed them to food inoculated with a 3:1 ratio of Ap21:Lb22. This ratio was determined based on our initial analysis of the community structure observed in CagA transgenic flies. Flies were raised to adulthood and then aged seven days to match the experimental setup used in other assays described above. The midgut was dissected and plated on MRS agar and total Colony Forming Units (CFUs) of each bacterial isolate were determined. We found the associated microbiota from control and CagA transgenic flies contained both Ap and Lb (Fig 2E), confirming that Lb can colonize control flies. However, we found the associated microbiota of CagA transgenic flies contained predominantly Lb compared to CFU counts for Ap (Fig 2D), which deviated significantly from the initial 3:1 (Ap:Lb) inoculum. In addition, the opposite community distribution was seen in control flies, where Ap was predominant and the microbial composition more closely matched the CV condition. Taken together these data demonstrate that Lb can colonize control flies but that expression of CagA enriches for Lb, which could contribute to cell proliferation.
Flies deficient for IMD pathway activation are more sensitive to infection with pathogenic bacteria and experience overgrowth of commensal communities in the midgut [35,36]. Antimicrobial peptide expression in the midgut is normally reserved for combating pathogenic infection and is not typically induced by commensal microbes [37,38]. Therefore, we reasoned that expression of CagA could alter host immune signaling causing the overexpression of Diptericin and Duox, either of which could be sufficient to alter the microbial community. To test this possibility we reared control and CagA transgenic flies GF and used qPCR analysis to measure expression of Diptericin and Duox in the midgut epithelium. Surprisingly, we found the antimicrobial peptide Diptericin was down-regulated (P<0.01) (Fig 3A), and the ROS transporter Duox (Fig 3B) was not over-expressed in control and CagA transgenic flies in the absence of gut microbiota, suggesting that the observed overexpression of these genes in the CV CagA-expressing flies is a consequence of the dysbiotic microbiota. Because overexpression of Diptericin appears to be dependent on a dysbiotic microbiota, we wanted to determine whether overexpressing Diptericin could explain the microbiota-dependent portion of cell proliferation in CagA transgenic flies. To test this hypothesis we expressed a UAS-diptericin transgene using esg-Gal4, UAS-GFP, and then assayed cell proliferation in the midgut epithelium. Using qPCR analysis we first determined that transgenic expression of diptericin resulted in similar transcript levels (11-fold over control) to those observed in CV CagA transgenic flies (8-fold over control) (P<0.05) (Fig 3C). When we assayed cell proliferation, we found that flies overexpressing Diptericin in the midgut epithelium showed rates similar to controls and much lower than those observed in CV CagA transgenic flies (P<0.0001) (Fig 3D). Additionally, gut microbiota was not affected in flies overexpressing Diptericin and remained similar to controls (S2 Fig). Taken together we conclude that overexpression of the antimicrobial peptide Diptericin is a consequence rather than a cause of altered gut microbiota and does not contribute to either the increased cell proliferation or the altered microbiota observed in CagA transgenic flies. Because overexpression of the Duox gene is not sufficient for Duox activation at the membrane, we were unable to use similar experiments to test its role in epithelial cell proliferation or dysbiosis.
Based on our findings that the altered gut microbiota of CagA transgenic flies promotes epithelial cell proliferation in the midgut, we predicted that microbiota of CagA transgenic flies would be sufficient to elicit a similar response in control flies. Previous reports in Drosophila have shown that the composition of microbial communities in the gut and immediate environment are affected by selective processes within the fly gut [39], therefore we expected cohousing flies might result in transmission of the dysbiotic microbiota and could initiate excessive cell proliferation in control flies. To test this possibility, we reared CagA-expressing and control flies to adulthood under CV conditions and then cohoused them for seven days before we assayed cell proliferation and microbiota community structure. Each condition contained 10–12 females and 8–10 males of a given genotype, and each genotype was cohoused with equivalent numbers of control and CagA or CagAEPISA flies to insure population density was sufficient to support bacterial growth and colonization [39]. To determine whether microbiota of CagA transgenic flies was capable of invading a control microbiota, we first assayed the output microbial community from flies in each housing situation. We found microbiota of control flies cohoused with other controls contained primarily Ap and rarely Lb (Fig 4A). These flies also showed low rates of cell proliferation (Fig 4B), similar to those observed previously in controls (Fig 1A). However, when control flies were cohoused with CagA flies, we observed both Ap and Lb in their microbial community (Fig 4A) and we observed a concomitant increase in cell proliferation in the midgut (P<0.0001) (Fig 4B), which was significantly higher than is typically observed in controls. Similarly, CagAEPISA flies cohoused with CagA flies also adopted a CagA-like microbiota of Ap and Lb (Fig 4E) and we observed higher rates of cell proliferation within the midgut epithelium than when they were housed with control or other CagAEPISA flies (P<0.01) (Fig 4F). We did see sporadically the presence of Lb in CagAEPISA flies but the total CFUs were often very low and did not appear to influence epithelial cell proliferation greatly (Fig 4E). Interestingly, cohousing appeared to have no effect on the microbial community of CagA transgenic flies, nor do we see any effect on cell proliferation (Fig 4C and 4D). From these data we conclude that the altered microbiota of CagA transgenic flies is transmissible with the ability to displace a healthy conventional community and promote increased intestinal epithelial cell proliferation even in control flies.
Next we aimed to identify the specific pro-proliferative component of the dysbiotic microbiota in CagA transgenic flies. Lactobacillus was a critical component of the dysbiotic microbiota and is a bacterial genus that has been previously shown to promote intestinal epithelial cell proliferation in Drosophila [40,41]. Based on our finding that Lb was only present in the CagA pro-proliferative microbiota, we predicted that mono-association with Lb might be sufficient for the pathological proliferation observed in CagA transgenic flies. To test this hypothesis we derived control flies GF, raised them to adulthood, and inoculated their food with defined bacterial inocula of Ap21 alone, Lb22 alone, or a 3:1 ratio mixture of Ap21:Lb22, strains isolated from our laboratory CagA transgenic flies. Newly eclosed adult flies were aged for seven days on the defined bacterial inocula before we assayed midgut cell proliferation and the microbial community. First, control flies were mono-associated with Ap21 to recapitulate the CV condition. We found rates of cell proliferation similar to that of CV control flies (Fig 5A), although the final CFUs per midgut were slightly lower (Fig 5B). When control flies were mono-associated with Lb22, rates of proliferation were even lower than those observed upon mono-association with Ap21 or in CV controls (P<0.01) (Fig 5A). We verified Lb22 colonization by plating the dissected midguts and found CFUs were higher than those observed upon Ap21 mono-association (Fig 5B). These data suggest neither Ap21 nor Lb22 alone is sufficient to promote pathological proliferation. In contrast, control flies associated with a 3:1 ratio of Ap21:Lb22 had significantly higher rates of proliferation in the midgut epithelium than control flies mono-associated with either of the individual isolates (P<0.0001) (Fig 5A). We also verified the presence of both microbial isolates from the output community and determined that the total CFUs per midgut recapitulated total CFUs observed in CV CagA flies (Fig 5A). Based on these findings we conclude that the pro-proliferative effects of the dysbiotic CagA microbiota require both Ap and Lb. This was confirmed when we assayed CagA transgenic and CagAEPISA flies, which were either mono-associated with individual isolates or associated with the CagA-like community of Ap21 and Lb22. In each case, mono-association with either Ap21 or Lb22 caused lower rates of cell proliferation (P<0.01 and P<0.0001) than was observed upon community association with Ap21 and Lb22 (Fig 5C and 5E). We also found that the output microbiota of CagA and CagAEPISA flies associated with Ap21 and Lb22 were similar to CV CagA transgenic flies with an Lb-dominate community (Fig 5D and 5F). From these data we conclude that this strain of L. brevis, Lb22, is not sufficient on its own to promote the excess cell proliferation observed in CagA transgenic flies and instead requires interspecies interactions with Ap21 to become pro-proliferative.
Both L. plantarum and L. brevis were previously identified as Drosophila commensals that promote epithelial cell proliferation in the Drosophila midgut. L. plantarum strains promote cell proliferation through the stimulation of cellular ROS [41,42] and both species have been shown to stimulate ROS production through release of uracil [40]. We aimed to determine whether the interspecies interactions of the dysbiotic CagA community use these known mechanisms to promote proliferation in the Drosophila gut. To test this possibility we reared flies GF and mono-associated flies with a Drosophila-derived L. plantarum (Lp) isolate [43] or a 3:1 ratio of Ap21 and Lp. We found flies associated with Lp alone or the Ap21Lp community showed significantly lower average cell proliferation than was observed with the Ap21Lb22 community association; 30 vs 55 cells/midgut in control, 31 vs 94 cells/midgut in CagA transgenic flies and 29 vs 91 cells/midgut in CagAEPISA transgenic flies (S3A Fig). Additionally, we tested the Drosophila-derived L. brevis strain EW previously shown to elicit intestinal epithelial proliferation and activation of Duox through production of uracil [40]. We found neither mono-association with LbEW or association with a 3:1 ratio of Ap21 and LbEW was sufficient to promote excess proliferation; 26 vs 55 cells/midgut in control, 38 vs 94 cells/midgut in CagA transgenic flies and 25 vs 91 cells/midgut in CagAEPISA transgenic flies (S3B Fig). We concluded that the excessive proliferation promoted by the dysbiotic microbiota of CagA transgenic flies is not dependent upon previously known mechanisms and instead requires unknown interspecies interactions to promote proliferation in the Drosophila gut.
Our analysis of the H. pylori protein CagA, expressed transgenically in the Drosophila adult midgut, reveals distinct mechanisms by which this bacterial virulence factor causes excessive epithelial cell proliferation (Fig 5G). First, expression of CagA in intestinal stem cells results in a cell autonomous increase in cell proliferation, independent of the presence of a resident microbiota. This finding is consistent with recent observations of CagA-dependent expansion of gastric stem cells in H. pylori infected murine gastric glands [21,44]. Our gnotobiotic experiments reveal an additional level of excessive cell proliferation that is mediated through the altered microbiota that assembles specifically in the CagA-expressing adult midgut. This non-cell autonomous cell proliferation is recapitulated in wild type animals that receive a dysbiotic CagA-associated microbial community either through cohousing or through gnotobiotic inoculation with a defined bacterial community.
In addition to the excessive cell proliferation in the CagA-expressing guts, we also observed excessive activation of innate immunity as indicated by up-regulation of genes encoding the antimicrobial peptide, Dpt and the dual oxidase, Duox. This activation, however, does not appear to be the cause, but rather a consequence of the CagA-induced dysbiosis, since it does not occur in CagA-expressing animals when they are reared GF. Furthermore, overexpression of Dpt is not sufficient to cause the dysbiosis or excess cell proliferation observed in CagA-expressing guts. We are actively investigating other mechanisms through which CagA expression leads to dysbiosis.
We only observe CagA-induced excessive cell proliferation when we express the protein in intestinal stem cells and their progenitor enteroblasts, but not when we drive expression in nutrient absorbing enterocytes. During H. pylori infection of human and murine stomachs, the bacterium is found in close association with gastric stems cells within gastric glands [20], and this physical proximity is correlated with a CagA-induced increase in stem cell proliferation [21]. It will be interesting to dissect the specific functions of stem cells, as opposed to differentiated epithelial cells, that render them more susceptible to CagA-induced proliferation. One possibility is that CagA expression in the stem cells may impair the polarity and epithelial integrity of the resulting tissue. CagA has been shown to disrupt epithelial polarity in cultured monolayers [45,46], and we noted numerous genetic interactions between CagA and polarity genes in a screen for modifiers of CagA-induced Drosophila retinal epithelial morphology disruption [47].
Gastrointestinal dysbiosis, as analyzed in fecal microbiome samples, is strongly linked to colorectal cancer risk [6], however the effect of dysbiosis within gastric communities is less clear. H. pylori infection itself can be viewed as a gastric dysbiosis characterized by overgrowth of a pathobiont that is a known carcinogen. H. pylori infection is also associated with other alterations in the gastric microbial ecosystem. Patients infected with H. pylori show altered gastric microbiota from that of uninfected individuals [20] and those changes revert back to an uninfected state upon H. pylori eradication [14]. Similarly, several groups have reported altered gastric microbiota in H. pylori infected versus uninfected mice [48,49]. Although the mechanism that could induce gastric dysbiosis in H. pylori infected individuals remains unknown, increased gastric pH, decreased gastric motility and gastric atrophy have all been proposed [50]. Our results are the first to specifically implicate CagA in contributing to microbiota alterations upon H. pylori infection. We plan to use our model of CagA expression within the Drosophila midgut to investigate the role of pH in microbiota maintenance and dysbiosis.
Whether H. pylori-induced microbiota shifts contribute to this bacterium’s pathogenesis in humans is not known. However, experiments in mice have demonstrated that pre-infection microbiota composition modulates the severity of H. pylori-induced pathology [15]. Furthermore, experiments in GF and CV gastrin-overexpressing mice demonstrated that the presence of the microbiota accelerated H. pylori-induced stomach cancer in this model [16], and that specific bacterial taxa from these H. pylori infected mice were sufficient to induce disease acceleration [51].
Dysbiosis is usually associated with either the overgrowth of a single pathobiont, such as adherent-invasive Escherichia coli in a TLR5 deficient model of spontaneous colitis [52], or the loss of a single protective strain, such as the anti-inflammatory Faecalibacterium prausnitzii that is reduced in Crohn’s disease patients [53]. In contrast, the dysbiotic microbiota of CagA transgenic Drosophila requires the presence of two bacteria, Acetobacter pasteurianus (Ap) and Lactobacillus brevis (Lb). These species of Acetobacter and Lactobacillus are well-known members of the Drosophila midgut microbiota [28], and interspecies interactions between closely related species have been shown to determine nutrient allocation in the fly [43].
Lactobacillus species have been previously shown to promote epithelial proliferation in Drosophila via induction of epithelial ROS production [41,42], in some circumstances via bacterial derived uracil [40]. We do not suspect these known mechanisms to be the cause of excess cell proliferation resulting from dysbiotic CagA microbiota because neither association with a ROS-inducing L. plantarum or a uracil-producing L. brevis (LbEW) elicited the high levels of proliferation we observed with the Ap21Lb22 community. Our data instead suggest a novel mechanism, which requires interspecies interactions between the specific Ap and Lb strains we isolated from our Drosophila colony. Metabolic cross feeding between Acetobacter and Lactobacillus species has been well described (e.g. [54]) and it is plausible to imagine that the microbial metabolites produced by these strains during mono-association in the Drosophila midgut differ significantly from those produced when both strains are present. We are currently exploring the genomic and metabolic properties of these bacterial strains. To our knowledge, this is the first reported demonstration of an interspecies interaction being the etiological agent of dysbiosis-associated disease.
Based on our results, we propose a new paradigm for bacterial pathogenesis by which bacterial virulence factors induce dysbiosis that contributes to disease pathology. This view expands the traditional view of virulence factors as modifiers of host cell biology and considers their capacity to modify host microbial ecology. Salmonella enterica serovar Typhimurium provides an exemplar of this mechanism as a pathogen that promotes intestinal inflammation as an adaptive metabolic strategy [3]. A consequence of this host inflammation, which requires the invA and spiB virulence factors, is the expansion of other pro-inflammatory Proteobacteria that promote disease pathology [55]. An implication of this view of bacterial virulence factors is that treating the pathogen-associated dysbiosis could mitigate the pathology of infectious disease. With a dwindling arsenal of effective antibiotics, future treatments for infectious disease may rely more heavily on therapies such as fecal microbiota transplants and probiotics. Simple model systems for dissecting the mechanisms of dysbiosis will provide useful tools for advancing these therapeutic approaches.
Drosophila melanogaster (Wolbachia-free) were reared at 25°C, 12h:12h light:dark cycle in a humidified chamber on standard cornmeal agar medium. All assays were performed on mated adult females. The following Drosophila lines were used: Sp/CyO; UAS-CagA, UAS-CagA; Dr/TM3 Sb and UAS-CagAEPISA [18] w; UAS-Dipt,imd,DiptD/CyO; spz/TM3 Sb, [22], esg-Gal4, UAS-GFP; Dr/TM6B Tb [56] and Myo1A-Gal4 [57]. Drosophila gut microbiota members were isolated on MRS agar from aseptically dissected 7–10 day old adult female guts.
Fresh laid eggs (<18 hr old) were collected from apple juice agar plates and dechorionated in 50% bleach for 3 minutes then rinsed in 2 consecutive washes of 70% ethanol followed by 2 washes in sterile H2O. Sterile embryos were then aseptically transferred to sterile fly food and maintained at 25°C. Inocula for gnotobiotic flies were prepared as follows and added to the food after aseptic transfer of eggs: 100 μl of cell suspension were added to each gnotobiotic vial to give 5 x 106 cells per vial. For inculcation with the two species community microbes were added in a 3:1 ratio of Ap:Lb to make up the total inoculum. Strains used: A. pasteurianus DORAp21, L. brevis DORLb22, L. brevis EW [40] and L. plantarum [43]. Axenia was confirmed by homogenizing a single representative larva from each bottle in 200 μl sterile 1X PBS. Homogenates were plated on MRS agar and incubated at 30°C for 2–4 days to evaluate bacterial growth. Axenic flies were transferred to sterile fly food 0–2 days after eclosing and allowed to age for 7 days at 25°C. As has been previously reported [58,59] axenic flies took longer to eclose than CV flies. We noted a similar delay in eclosure with Lb mono-associated flies. We also noted slight delays in ecolsure of CV CagA transgenic flies, which may reflect the significant Lb fraction of their microbiota. All data shown represent data collected from 7–10 day old mated adult females.
Microbial density was determined to assess the presence and/or abundance of each bacterial species associated with the host. In all experiments 7–10 day old adult female flies were dissected in sterile 1X PBS. The dissected midgut was immediately placed in 200μl 1X PBS and homogenized with a handheld pestle grinder for 20–30 seconds/gut. The resulting homogenate underwent serial dilutions and was then plated on MRS agar plates using sterile glass beads and incubated at 30°C for 2–4 days under aerobic conditions. CFU counts were determined after manually counting each plate. The limit of quantification (LOQ) was defined as 200 CFU per plate [60].
Aseptically-dissected 7–10 day old female guts were dissect in sterile 1X PBS then fixed for 30 minutes to 1 hour in fresh 4% Paraformaldehyde/1X PBS. Guts were washed 3 times for 15 minutes with 1X PBS containing 0.1% Triton X-100 (PBST) then blocked with the same solution plus 0.02% BSA (PBSTB) for 30 minutes at room temperature. Primary antibodies were applied either for 2 hours at room temperature or overnight at 4°C and include Rabbit anti-phospho histone H3 (1:500; Millipore) and Chicken anti-GFP (1:500; AVES labs). Guts were then washed 3 times for 15 minutes with PBSTB and incubated with AlexaFluor 594 Goat anti-Rabbit and AlexaFluor 488 Goat anti-Chicken, for 2 hours at room temperature. Guts were then washed 3 times for 15 minutes with PBSTB and mounted on glass slides with ProLong Diamond with DAPI anti-fade mounting media (Life Technologies). pH3+ cells were counted manually on a Nikon compound microscope. Total cell count includes pH3+ cells from the base of the proventriculus to midgut/hindgut junction at the posterior end of the midgut.
Bacteria were grown statically (Lactobacillus) or shaking (Acetobacter) at 30°C to late-log phase and genomic DNA was isolated using the Qiagen DNeasy Blood and Tissue Kit. Lysozyme digestion was used as a pre-treatment procedure. The University of Oregon Genomics Core Facility performed amplification of the V4 variable region of the 16S Ribosomal gene and sequenced products. The resulting sequences were submitted to a standard nucleotide BLAST that identified isolates as A. pasteurianus, which we gave the strain name DORAp21 and L. brevis, which we gave strain name DORLb22.
Total RNA was extracted from 5 pooled adult female midguts/per sample, using TRIzol reagent and the Qiagen RNeasy Mini Kit according to manufacturer’s protocol. cDNA was prepared using the Thermo Fisher Scientific SuperScript III Reverse Transcriptase Kit. cDNA was analyzed using gene-specific primers in triplicate, for at least three independent experiments. Data were analyzed by relative quantification by normalization to the gene rp49. Primers sequences were previously published [37] and are listed below: rp49: Forward 5’AGA TCG TGA AGA AGC GCA CCA AG 3’ Reverse 5’ CAC CAG GAA CTT CTT GAA TCC GG 3’; Diptericin: Forward 5’ GGC TTA TCC GAT GCC CGA CG 3’ Reverse 5’ TCT GTA GGT GTA GGT GCT TCC C 3’; Duox: Forward 5’ GCT GCA CGC CAA CCA CAA GAG ACT 3’ Reverse 5’ CAC GCG CAG CAG GAT GTA AGG TTT-3’; Attacin: Forward 5’ ACG CCC GGA GTG AAG GAT GTT 3’ Reverse 5’ GGG CGA TGA CCA GAG ATT AGC AC 3’; Defensin: Forward 5’ TGC AGC ATA GCC GCC AGA A 3’ Reverse 5’ TTG CAG TAG CCG CCT TTG AAC C3’.
<24 hour-old adult flies were collected and placed on fresh food vials and left to mate for 24 hours. Mated adults were then sorted into 3-vials of 20 females and 20 males each, with 3 replicates of each genotype. Vials were scored 3 times per week and the number of dead flies was recorded each day until all the flies were recorded dead or the vial was empty.
|
10.1371/journal.pcbi.1004072 | Input-Dependent Frequency Modulation of Cortical Gamma Oscillations Shapes Spatial Synchronization and Enables Phase Coding | Fine-scale temporal organization of cortical activity in the gamma range (∼25–80Hz) may play a significant role in information processing, for example by neural grouping (‘binding’) and phase coding. Recent experimental studies have shown that the precise frequency of gamma oscillations varies with input drive (e.g. visual contrast) and that it can differ among nearby cortical locations. This has challenged theories assuming widespread gamma synchronization at a fixed common frequency. In the present study, we investigated which principles govern gamma synchronization in the presence of input-dependent frequency modulations and whether they are detrimental for meaningful input-dependent gamma-mediated temporal organization. To this aim, we constructed a biophysically realistic excitatory-inhibitory network able to express different oscillation frequencies at nearby spatial locations. Similarly to cortical networks, the model was topographically organized with spatially local connectivity and spatially-varying input drive. We analyzed gamma synchronization with respect to phase-locking, phase-relations and frequency differences, and quantified the stimulus-related information represented by gamma phase and frequency. By stepwise simplification of our models, we found that the gamma-mediated temporal organization could be reduced to basic synchronization principles of weakly coupled oscillators, where input drive determines the intrinsic (natural) frequency of oscillators. The gamma phase-locking, the precise phase relation and the emergent (measurable) frequencies were determined by two principal factors: the detuning (intrinsic frequency difference, i.e. local input difference) and the coupling strength. In addition to frequency coding, gamma phase contained complementary stimulus information. Crucially, the phase code reflected input differences, but not the absolute input level. This property of relative input-to-phase conversion, contrasting with latency codes or slower oscillation phase codes, may resolve conflicting experimental observations on gamma phase coding. Our modeling results offer clear testable experimental predictions. We conclude that input-dependency of gamma frequencies could be essential rather than detrimental for meaningful gamma-mediated temporal organization of cortical activity.
| Almost 350 years ago the physicist and polymath Christiaan Huygens first observed the synchronization between two pendulum clocks attached to a common support. Since then synchronization has been recognized as a universal phenomenon from astronomy to biology. The phase-locking (synchrony) and the phase-relation between the two pendulums are determined by two principal forces: the synchronization force exerted over the connection and the tendency to desynchronize due to frequency (speed) differences. We propose that gamma synchronization (25–80Hz) among oscillating cortical neurons in the brain can be understood according to the same principles—like a field of many connected pendula—with the critical addition that input changes the frequency of gamma oscillations, as shown by recent experimental studies. It has been assumed that input-dependent changes in oscillation frequency are detrimental for a meaningful role of gamma synchronization in neural processing. To the contrary, our theoretical analysis demonstrates that because input can change the frequency of the oscillation, phase-locking and phase-relations among neurons relate systematically to input. By analogy, it is because a local push to a pendulum will change its frequency, that resulting changes in phase-locking and phase-relation among the pendula can be used to derive the external force applied.
| How the millions of neurons in the brain are coordinated to permit meaningful computations is one of the fundamental questions of neuroscience. Spike synchrony and relative spike timing play important roles in dynamically coordinating neural activity [1–7] with substantial impact on neuronal function [8–12]. Synchronization often goes hand in hand with neural oscillations, of which gamma-band oscillations (∼25–80Hz) have received broad attention [13–15]. Gamma oscillations occur in various brain regions and species [13–16]. Gamma oscillations arise locally from mainly direct interactions between inhibitory and excitatory neurons [14,15,17,18]. Modulations of gamma oscillation properties (power, frequency) have been found for various cognitive functions including perception [19–21], attention [22–24], working memory [23] as well as in psychiatric disorders like psychosis [25,26] and ADHD [27,28]. At the neuronal level, different roles (that are not mutually exclusive) have been suggested; they include neural grouping by phase-locking within [21,29–31] and between cortical areas [13,32,33], phase coding [15,18,34–37], neuronal plasticity [38,39], gain control [18] and normalization [40].
However, the role of gamma oscillations in neural computation is controversial, with judgments ranging from fundamental [13,14,21] to epiphenomenal [41–43]. Experimental studies have given conflicting evidence on the role of gamma phase coding of input drive. For example, Vinck et al. [34] have shown that visual cortical neurons receiving different input drive (through varying stimulus orientation) can exhibit reliable spike timing differences in the gamma oscillation range. However, Montemurro et al. [44] using natural stimuli could not find any contribution of visual cortical gamma phase to the encoding of the input. Similarly, McLelland and Paulsen [45] did not find a rate-to-phase transform for gamma oscillations, which would assign a specific level of input to a specific phase of gamma. Moreover, although various experimental studies [29,31,43] have shown input(stimulus)-dependent changes in gamma synchronization, theoretical models [21,46,47] have fallen short in convincingly including the local and variable nature of gamma oscillations. For example, the dependence of gamma oscillation frequency on stimulus attributes (e.g. visual contrast [33,42,43]) as well as the limited spread of gamma phase-locking over cortical distance [48,49] are seen as conflicting with a functional role of gamma oscillations in neural processing [40–42,50].
Here, we used computational modeling techniques to develop a deeper understanding of input-dependent cortical gamma synchronization. We focused on the underlying organization principles of phase and frequency coding of input drive and its relation to spatial synchronization and network connectivity. Mathematically, the synchronization principles of interacting limit-cycle oscillators (and other types, [51]) is well understood [52–54]. In particular, the theory of weakly coupled oscillators (TWCO) (see [55] for review) has proven to be useful and has been applied in many scientific domains, including neuroscience [52–54,56–60]. In TWCO the phase of an oscillator (neuron, group of neurons) is defined by an intrinsic (natural) frequency. The interaction with other oscillators is characterized by the phase response curve (PRC, [61]) which defines how the phase is modified by the interaction. Crucially, the phase-locking between oscillators depends on the intrinsic frequency difference (described as the detuning level) as well as interaction strength (or coupling strength), defining the so called Arnold tongues (region of synchronization defined by the interplay of detuning and coupling) [18,55,62]. Note that in TWCO, the coupling strength is considered to be ‘weak’, meaning that the interactions among oscillators mainly change the phases but not the oscillation amplitudes.
A few prior studies have concretely considered TWCO for explaining input-dependent cortical gamma synchronization [18,53,54,63–65]. Of most relevance here, Tiesinga and Sejnowski [18] first used TWCO in a biophysically realistic gamma network for explaining gamma phase coding in visual cortex [34]. Several interconnected pyramidal-interneuron-gamma networks (PING) synchronized on a common frequency, despite receiving different levels of input currents, and converted input differences into phase-differences.
Despite these important advances, the organization principles of gamma oscillations in cortical networks, characterized by local synchrony and input-dependent oscillation frequencies over cortical space, have so far not been systematically investigated. In particular, so far, the theoretical principles that determine the phase-locking and phase-relationship among interconnected gamma-oscillating neurons receiving different input levels are as yet not well understood. In the present study, we study whether TWCO may offer a framework to describe these organization principles. Moreover, it is currently poorly understood how much information about the stimulus input is encoded in the phase-relation and frequency differences among neurons. To answer these questions, we investigated spatially-defined excitatory-inhibitory (PING) networks that were, similarly to cortical networks, topographically organized with spatially local connectivity and spatially-varying input drive. The network exhibited local spatial synchrony and could express different gamma frequencies at different locations at the same time. The gamma frequency ranges were set to match our own observations in awake monkey V1. We used several networks varying in size and complexity. In all of them, we observed that phase-locking, phase-relations and frequency differences among neurons resulted from an interplay between detuning (Δintrinsic frequency) and coupling strength, in accord with TWCO and the Arnold tongue. Critical for the behavior was the property of gamma oscillations to shift their preferred frequency with input drive. Phase and frequency coding of input was largely complementary in accordance to the Arnold tongue concept, whereby conditions inside the Arnold tongue lead to phase coding, and conditions outside lead to frequency coding. A combined frequency and phase coding could best reconstruct the stimulus input. Importantly, the Arnold-tongue based phase coding implied a relative Δrate-to-phase transform and therefore gamma phase told little about absolute input levels. Our work has clear theoretical implications leading to experimentally testable predictions that are elaborated in the Discussion.
Experimental observations in Fig. 1 and associated methods of data collection shown have been described in a previous publication [33]. We show here only data from monkey S V1 for illustration purposes only. We re-analyzed the LFP spectra obtained during stimulation (Stim) with static square-wave grating (2 cycles per degree), using a multi-taper method with discrete prolate spheroid sequences for frequencies 20 to 60Hz (smoothing ± 3Hz) in non-overlapping 500ms windows starting 350ms after stimulus onset. LFP power in the pre-stimulus baseline (Base) was calculated from the 500ms period before stimulus onset. Relative power was calculated as (Stim-Base)/Base, where Stim and Base were calculated separately after averaging over trials. In Fig. 1B the quantifications of maximum of peak gamma power as well as frequency of peak power is shown for the Michelson contrast conditions 6.1%, 9.7%,16.3%,35.9%,50.3% and 72%. The stimulus contrast conditions 2.5% and 3.7% had very low induced gamma power and no clear peak in the power spectrum and therefore not included.
Intel(R) Xeon(R) CPU E5-1620 0 @3.6GHz with 16GB RAM.
Minimal single-compartment Hodgkin-Huxley models [66] were used to construct E-cells (regular-spiking excitatory neurons, RS) and I-cells (fast-spiking inhibitory interneurons, FS). For the network simulations shown in Fig. 1 and Fig. 2 the networks consisted of 80 E-cells and 20 I-cells. For Fig. 3 the network consisted of 160 E-cells and 40 I-cells neurons. The E-cells in Fig. 3 had particularly high firing rate matching the network gamma frequencies. This was done to increase oscillatory stability of the small network which was limited in size due to computational constraints. We use Izhikevich-type neurons [67] for replicating our findings in larger E-and I-cells networks (see S1 Text). In Fig. 3 neurons were ordered along a ring to avoid network border effects (continuous connectivity). Numerical simulations were computed using a variable step size Runge-Kutta method of order 8 according to the Dormand and Prince algorithm [68]. The simulation code was written in FORTRAN95. Analysis of the simulation output was performed with Matlab (MathWorks, R2012b).
Regular-spiking (RS) E-cell:
C
m
dV
dt
=−
g
Leak
(
V−
E
Leak
)−
I
Na
−
I
K
−
I
M
Fast-spiking (FS) I-cell:
C
m
dV
dt
=−
g
Leak
(
V−
E
Leak
)−
I
Na
−
I
K
The leakage reversal potential and conductance were ELeak = −70mV and gLeak = 0.0205mS / cm2 for E-cells and gLeak = 0.015mS / cm2 for I-cells. The membrane capacitance was Cm = 1 mF/cm2. All kinetic parameters were according to a temperature of 36ºC using standard conductance equations.
Sodium current INa:
RS:
E
N
a
Ê
=
50
Ê
m
V
,
g
¯
Na
=50mS/c
m
2
V
T
=−61.5mV
FS:
E
Na
=50
mV
,
g
¯
Na
=46mS/c
m
2
V
T
=−61.84mV
I
Na
=
g
¯
Na
m
3
h(
V−
E
Na
)
dm
dt
=
α
m
(
V
)(
1−m
)−
β
m
(
V
)m
dh
dt
=
α
h
(
V
)(
1−h
)−
β
h
(
V
)h
α
m
=
−0.32(
V−
V
T
−13
)
exp[
−
V−
V
T
−13
4
]−1
β
m
=
0.28(
V−
V
T
−40
)
exp[
−
V−
V
T
−40
5
]−1
α
h
=0.128exp[
−
V−
V
T
−17
18
]
β
h
=
4
1+exp[
−
V−
V
T
−40
5
]
Delayed-rectifier potassium current IKd:
RS:
E
K
d
=
−
90
m
V
,
g
‒
K
d
=
4.8
m
S
/
c
m
2
V
T
=
−
61.5
m
V
FS:
E
Kd
=−
90
mV
,
g
¯
Kd
=5.1
mS/c
m
2
V
T
=−61.84mV
I
Kd
=
g
¯
Kd
n
4
(
V−
E
k
)
dn
dt
=
α
n
(
V
)(
1−n
)−
β
n
(
V
)n
α
n
=
−0.032(
V−
V
T
−15
)
exp[
−
V−
V
T
−15
5
]−1
β
n
=0.5exp[
−
V−
V
T
−10
40
]
Slow non-inactivating potassium current IM:
RS:
E
Km
=−90
mV
,
g
¯
Km
=0.15
mS/c
m
2
V
T
=−61.5mV
,
τ
max
=
1123.5
ms
I
M
=
g
¯
M
p
(
V−
E
k
)
dp
dt
=(
p
∞
(
V
)−p
)/
τ
p(
V
)
p
∞
(
V
)=
1
1+exp[
−
V+35
10
]
τ
p
(
V
)=
τ
max
3.3exp[
V+35
20
]+exp[
−
V+35
20
]
Synaptic excitatory AMPA and inhibitory GABA-A potentials were modeled based on [69]. The synaptic current into neuron α was:
I
syn
,
α
=
G
in
,
α
(
V
α
−
V
in
)
+
G
ex
,
α
(
V
α
−
V
ex
)
Here, the total synaptic conductance from inhibitory presynaptic neurons was:
G
in,α
=
∑
β
in
g
β→α
s
β
The expression for the excitatory synaptic conductance was of the same form. It was assumed that the dynamics of all synapses of a given (presynaptic) neuron were perfectly synchronized. Hence a synaptic gate, though physically located on the post-synaptic neuron α, followed the potential Vβ of the pre-synaptic neuron β with parameters shown in Table 1. For maximum conductance values gβ→α see above.
For Fig. 1–3 the network connectivity parameters (in mS/cm2) are listed in Table 2. The connectivity matrix network was based on the number of neighbor connections m. For example, m = 8 meant that a neuron connected to the closest 8 other neurons with unit connection strength. For E-I connections, m meant that I-cell received input from m E-cells (afferent). For I-E, m meant that I-cell sent input to m E-cells (efferent). The connectivity parameters were normalized (divided) by the number of connections m. The chosen parameters are listed in Table 3. We describe effects of changing coupling parameters in S1 Fig.
The input to each neuron consisted of an external excitatory input plus internal excitatory and inhibitory input via network connections. The external input consisted of a train of AMPA synaptic conductance spikes (double exponentials: rising constant = 1ms, decaying constant = 5.2ms) with Poisson statistics at a rate of 800Hz (±SD = 100) and spike amplitudes of default 0.02mS/cm2 (±SD = 0.002). The default mean AMPA input level to each neuron was 0.01mS/cm2 for FS neurons. For Fig. 1 the mean amplitude of the AMPA synaptic potentials were modulated from 0.02 to 0.08mS/cm2 for RS by modulating the spike amplitude. In Fig. 2 the mean AMPA conductance input level was 0.06 mS/cm2. In Fig. 3 each neuron received a spatially specific input level depending on its position in the ring architecture. The amplitude of the sinusoidally modulated AMPA conductance was of 0.006mS/cm2 and the mean conductance AMPA input 0.06mS/cm2 and 0.055 mS/cm2 for S1 Fig.
The voltage used as the spike detection level was −17.5mV for both E- and I-cells. The local field potential (LFP) was estimated in units of microVolt (μV) for Fig. 1 & 2 as an overall network signal. The LFP was the extracellular electrical field potential LFP = LFP (r0,t) at an electrode position r0. We treated neuron i at position ri as a point-current source I = I (ri,t) (total transmembrane current into the neuron) in a homogeneous extracellular medium with conductivity σ (1/σ = 0.3 kΩcm), taken from [70]: 0.2–0.4 kΩcm). We summed the individual neuron contributions according to the quasistatic Maxwell equations:
LFP=
1
4πσ
∑
i
N
I(
r
i
,t
)
R
with
R=|
r
i
−
r
0
|
the distance of the point source to the electrode (R = ∼1mm). The extracellular voltage signal was smoothed with a pseudo-Gaussian function (width = 4ms).
For computing local synchronous rhythmic activity of local population of neurons in S1 Fig. to show effects on noise on rhythmic population activity and rhythmic single neuron activity, we derived a local population average signal (LPA) based on the spike trains of the neurons. For each position in the network, we aggregated the spike activity of the whole network weighted by a spatially exponentially decaying function.
With ri being the binary spiking variable, Di,j being the spatial distance between neurons (defined circularly on the ring in radians). S corresponds to the spatial decay constant which was chosen to be 0.4. The rationale of the value (similar results were observed for a large range of values) was that it was large enough to allow for sufficient aggregation to quantify oscillatory activity and spatially specific enough to reveal the spatial change of phase-locking and phase-relation. The LPA was further smoothed with a pseudo-Gaussian function (width = 4ms).
We obtained natural images from the Berkeley segmentation dataset (BSDS500, [71]). We took the first 100 gray-scale natural images (comprising the Berkeley training dataset). The natural images were first resampled and squared to fit it to the 100×100 lattice. Then the local root-mean squared contrast C = C (xc, yc) at image position (xc, yc) (RMS, [72]) was computed:
C=
∑
i=1
N
w
i
(
L
i
−L
)
2
L
2
/
∑
i=1
N
w
i
w
i
=0.5(
cos(
π
p
(
x
i
−
x
c
)
2
+
(
y
i
−
y
c
)
2
)+1
)
L is luminance and i is the pixel index. The summation was over pixels within a patch radius p of 3 pixels. The local contrast values C were then transformed into intrinsic frequencies ν by approximating the experimentally observed relationship ν = 25 + 0.25C between gamma frequency and contrast. We defined a minimum (25 Hz) and slope of 0.25Hz per contrast value (estimated over both monkeys,[33]). Gamma power was not taken into account in the phase-oscillator model. 20 of 100 images were excluded because not enough segmentation borders (see criteria below) suitable for phase-locking analysis (minimum of > 10 per image) could be obtained.
We used a modified version of the Kuramoto model [56] as a basic model of the dynamics of a limit-cycle oscillators that has been used to investigated synchronization between coupled oscillators. The network input was set by natural images transformed into local contrast. The intrinsic frequency of each oscillator was set by the local contrast at the oscillator’s corresponding pixel. For each image, the simulation run was 10s with a time step of 2ms. Each oscillator started with a random phase. During the simulation run the phase of each oscillator was determined by an intrinsic (natural) frequency (ω), a noise term (ζ) and an interaction term describing the impact (phase response curve, PRC) by other coupled oscillators depending on the coupling constant (K).
The interaction term (infinitesimal PRC, [56]) was a sinusoidal function such that the coupled oscillators tended to engage in zero-phase synchrony. The coupling constant was an exponential function of distance (D) (in contrast to the all-to-all connectivity in the Kuramoto model), with a scaling constant (s = 0.4 for ring-network (radians) and s = 0.5 for 2D lattice network (pixel)) and strength (C = 0.003).
The noise term was pink noise with a power scaling exponent of 1. The strength C was scaled at a sufficient level for the model to reach near-zero coherence when oscillators were uncoupled. The noise was spatially correlated (smoothed with spatial kernel of 3 pixels), to reduce spurious phase locking over the lattice. This step eased the computation of ‘true’ phase-locking between distant clusters having very close frequencies (without the use of phase-perturbation techniques), because synchronous clusters cannot easily average out noise (correlated between members of a cluster). We also included a time-delay term as function of distance as conduction delay of cortical horizontal connections can be significant for longer cortical distances [73,74].
Where the time-delay Ti,j was a linear function of distance (pixel units). The slope v was 0.4 and v0 was 2ms. Ti,j was then made then discrete to change in steps of 2ms (simulation time step). The inclusion of the time-delay factor was not critical for the results of the paper. Natural images and segmentations by human observers were taken from the Berkeley segmentation dataset [71]. Images were downsampled from 350×450 pixels to 100×100 pixels using the Matlab in-built ‘imresize’ function, to fit the size of the lattice model.
In Fig. 1 as well as S2 Fig. we used the Matlab in-built power spectral density function (psd) with multitaper estimation for estimating the power spectrum. For the time-frequency representation (TFR) in Fig. 2c we used the Matlab in-built spectrogram function (Short-time Fourier transform).
The instantaneous phase (IP) was derived for the LFP (Fig. 1–2) or LPA (S2 Fig.) signals by taking the Hilbert-transform (HT, [75]) of the signal. The HT gives the analytical signal (complex numbers) from which the IP can be obtained by taking the argument of the complex number. The HT is well defined for signals characterized by a single oscillation (mono-component) which was the case in our simulations. The IP was the output variable of the phase-oscillator model. The instantaneous frequency (IF) was obtained by taking the derivative of the IP. IF was estimated by unwrapping the IP, then first applying smoothing (half-cycle rectangular points smoothing) followed by computing the first derivative. For the phase-oscillator model (Fig. 5–8), we could directly use the output phase-traces to compute IF. We averaged the IF estimation of each time point over the whole simulation (excluding the first 200ms) period to obtain a mean frequency. For single neurons spike trains we used the spike rate, computed as n spikes per second, as our frequency estimation.
A. Based on Instantaneous phase (LFP,LPA and phase-oscillators)
The phase relation was defined as the mean circular phase difference between two signals (averaged in the complex domain).
θ
ij
=arg(
1
N
∑
t=1
N
exp(
i
(
φ
i
−
φ
j)
))
with a range of [−π, π]. Arg is the argument function and φ is the IP. For estimating phase-locking we computed the phase-locking value (PLV, [76]). The PLV was computed by averaging the complex values with unit amplitude
The PLV ranges from 1, corresponding to full phase consistency, to 0, corresponding to fully random.
B. Based on spike trains
For computing the phase-relation and locking between two neurons we applied cross-correlation.
C
C
ij
(
l
)=
1
N
(
∑
n=0
N−1
r
i
*
[ n ]
r
j
[
n+l ]
)
with
r
i
*
being the complex conjugate (*) of spike train (r) of neuron i. The cross-correlation CCij between neuron I and j was computed with lags not exceeding +/− half mean rate (time window is assumed to be the period of the oscillation the neurons are locked to). The spike timing difference (in ms) was defined as
s
t
ij
=arg(
max(
C
C
ij
)
)
and the locking as
ψ
ij
=max(
C
C
ij
)
We converted the spike timing differences into phase-values by dividing twice timing difference by the mean spike rate of the respective neurons and then multiplied by π.
The matrices represent the phase-locking or the phase-relation between all possible pairs of neurons or oscillators. The diagonal is always 1 (phase-locking with itself) in the phase-locking matrix and 0 in the phase-relation matrix. For the phase-locking matrix the color were from 0 (black) to 1 (yellow-white), if not otherwise stated. For the phase-relation matrix the color were – pi/2 (blue) to pi/2 (red). A negative phase relation (blue) means that the neuron/oscillator X from the x-axis has an earlier/leading phase compared to the neuron/oscillator Y from the y-axis. The phase-relation matrix was threshold for illustration purposes, because phase-relations from non-synchronized neurons/oscillators are randomly distributed over –pi to pi making the plot difficult to interpret visually. The threshold was defined as being equal to 3 times the mean phase-locking value between uncoupled neurons/oscillators.
We used the image segmentations performed by several human observers (n = 30) from the Berkeley segmentation dataset (BSDS500,[71]). All subjects did not segment all the images, instead, segmentations from a subset of the observers was available (n∼ = 5) for each image. For each image the segmentation-border analysis based on different observers was averaged. We selected 1-dimensional spatial windows of ±15 pixels centered on segmentation lines that fulfilled the following criteria: (1). A vertical or horizontal segmentation line should consist of three consecutive pixels. (2) Within the spatial window no other line should be present.
For the analysis, the horizontal and vertical line segments were concatenated. We then computed the averaged phase-locking matrix (S4 Fig.) between all oscillators as well as the averaged absolute spatial derivative of contrast values. For computing significance thresholds (permutation testing, [77]) we constructed a null distribution by choosing random positions for the same number of spatial windows.
The stimulus to be reconstructed Sorig(i) for each network position i was the excitatory (AMPA) input drive to E-cells for the PING networks and the intrinsic frequency for the phase-oscillator model. For ring-networks the spatial variation of Sorig was defined by a sinusoidal function and for the 2D phase-oscillator lattice network by natural image local contrast, where each network position corresponded to one pixel. A seemingly easy way of estimating the stimulus Sorig is by using frequency coding Sest (ω). If it is defined at single neuron level, it is often termed spike rate. At neuronal population level, the code might be based on the oscillation frequency. In the Discussion section we discuss these different type of frequency coding and their relation (see also S1 Fig.). The spike rate is defined as the number of spikes per second for a given time window (spike count code,[78]). This was used for Fig. 3. The oscillation frequency was determined as the mean instantaneous frequency [79] over the simulation period. This was used for Fig. 2 (LFP), Fig. 6 and Fig. 8 (phase-oscillator) and S1 Fig. (LPA). The Sest (ω) was simply defined as the frequency of the neuron i
The stimulus Sorig, estimated by phase differences between neurons, was defined as follows:
θij is the phase-relation, ψij the phase-locking and Kij the connectivity strength between the reference neuron i and the neighbor neuron j. θij was determined using the whole simulation period. For analysis based on spike trains, the phase-relation and strength was determined by cross-correlation analysis (see above), whereas for LPA or phase-oscillator analysis it was based on the instantaneous phase variable (for LPA determined by Hilbert transform). The connectivity strength was defined prior to simulation (see above) and was an exponential function decaying over distance (phase-oscillator model) or was of nearest neighbor type with unit strengths (PING network). For all network types, the interaction strength (determined by direct and indirect connections) decayed over distance approximated as an exponential decay function over space with the same parameter used for all network types. Including the connectivity term improved the Sest, in particular for the 2D phase-oscillator lattice model (MI = 0.60 to 0.67). This is because the detuning-to-phase conversion is coupling dependent. The phase value of each neuron i, was computed by averaging phase-relations to all other neurons in the network weighted by phase-locking strength and coupling strength. Sest(θ) is an assembly code using the spike relation between neurons to obtain more information about the stimulus.
For the combined frequency and phase code Sest(ω,θ) the stimulus level Sest for a given neuron was then given as
S
est
(
ω,θ
)=
S
est
(
ω
)+
S
est
(
θ
)*F
where F is a scaling factor determining the contribution of the phase code. The scaling factor F which maximized the stimulus reconstruction performance was chosen. Intuitively, the optimal scaling factor is the slope of the function between intrinsic frequency and the phase variable (see red line in Fig. 2J & 8IV) for a given coupling value. One common scaling factor was chosen for all 80 natural image simulations.
We estimated the reconstruction performance as the Shannon mutual information I(X;Y) [80] between the intrinsic frequency image X and the reconstructed intrinsic frequency image Y. The direct method approach gives:
I(
X;Y
)=
∑
y∈Y
∑
x∈X
p(
x,y
)log(
p(
x,y
)
p(
x
)p(
y
)
)
where p(x,y) is the joint probability, p(x) and p(y) are the marginal distributions. We normalized I(X;Y) by dividing by I(X;X). We compare the phase-code, frequency-code and the combined frequency/phase code. For the natural images, we computed a baseline reconstruction performance, for each image we computed the normalized MI between the simulation output from that image with the intrinsic frequency maps from all other images. We averaged the 79 MI values (79 MI values per 80 images, each image compared with all other images) to get an estimate of baseline reconstruction for each image. We used a repeated measures ANOVA to test for significant effect of coding types. For post-hoc pairwise comparisons [81] between different coding types we used the Tukey’s HSD (honest significant difference) test (which corrects for multiple comparisons). The Tukey’s HSD was computed as follows:
HSD=
(
μ
1
−
μ
2
)
MSE/N
With μ the mean value of a condition, MSE is the mean sum squared error and N is the number of values within a condition.
During active information processing, a cortical network will receive variable afferent input drive reflecting sensory variables. By input drive we mean the net excitatory drive to a population of neurons resulting from the sum of afferent excitatory and inhibitory connections. The dependence of gamma oscillations on input drive is central for the understanding of its role in neural processing. Theoretical [17,33,42] and experimental observations [33,42,43,82,83] have shown that excitatory drive increases the frequency of gamma oscillations. For example, recent experimental studies on gamma oscillations in primate visual cortex have shown a striking relationship between visual contrast, which is considered a proxy for excitatory drive [33,42,43], and the frequency and power of gamma oscillations. Results from our own experimental work [33] demonstrated the effect of contrast on gamma oscillations in primary visual cortex V1 and in V2 of rhesus monkeys (Fig. 1A). We found a monotonic increase in the frequency at which the gamma frequency spectrum peaks (Fig. 1B, top) with increased contrast, and a non-monotonic modulation of gamma power (Fig. 1B, bottom).
These findings fit with theoretical studies of the two most common gamma oscillation generating mechanisms [17], the interneuron-gamma network (ING, e.g. [84,85]) and the pyramidal-interneuron gamma network (PING, e.g. [33,86]), which are characterized by increasing oscillation frequency with increasing excitatory drive. We replicated this relationship in a model network consisting of 20 I-cells (fast-spiking type) and 80 E-cells (regular spiking type) using model neurons based on the Hodgkin-Huxley formalism (Fig. 1C, see Methods,[66]). Model neurons interacted through model synapses [23] that included AMPA and GABA-A connections. Each neuronal class received independent external excitatory input, yet the main excitatory input for I-cells was internally generated by E-cells (Fig. 1C). The network exhibited pyramidal-interneuron gamma oscillations (PING) characterized by I-cell spikes lagging E-cell spikes (Fig. 1D). We then (Fig. 1E) systematically modulated external excitatory input to the network (modeled as a train of AMPA-spikes), with the mean level of input ranging from 0.02 to 0.08 milliSiemens per area (mS/cm2). GABA-A decay time constant (20ms) was defined such that frequencies were in the range as observed in our own experimental V1 LFP recordings. However, the exact frequency range is not critical for the conclusions of the paper. We observed input-dependent effects on the model power spectra based on the estimated LFP (from transmembrane currents, see Methods). Gamma oscillation frequency increased monotonically with input-drive over a range of ∼20–25Hz [33], as did the spike rates (Fig. 1F, top). Oscillation power (Fig. 1F bottom) showed a nonlinear relationship with oscillation frequency, with peak power at intermediate levels of input [33]. In line with previous work (for review [17]), the main time constant of the PING network oscillation was set by the inhibitory GABA-A decay time constant and by the time needed for the E-cells to escape the inhibition, the latter being reduced by higher excitatory drive. We assume here that synaptic time constants and connectivity strengths did not change within the time-scale considered here for the stimulation (several 100ms to a few seconds).
As described above, there is substantial evidence that gamma oscillations adapt their frequency as a function of input-drive. But what happens if input-drive varies over cortical space? An experimental study in macaque V1 [43] with contrast-varying stimuli has shown that the frequency of gamma oscillation can vary over a short cortical distance, with higher contrast producing higher frequencies. Hence, nearby cortical location can show different oscillation frequencies. This supports older studies in V1 [31,87] that showed that gamma phase-locking decayed rapidly over cortical distance at the spatial scale of horizontal connectivity. In the light of those findings, a theoretical model of cortical gamma oscillation should be able to express different oscillation frequencies at nearby spatial locations. Such results cannot be simulated in gamma network models that are characterized by global synchronization and express one dominant frequency at a time [88]. Gamma oscillation networks with predominantly local spatial connectivity with locally varying input drive could thus be a promising framework for cortical gamma oscillations. We therefore aimed to gain an understanding of the underlying principles that cause these networks to organize themselves depending on spatially varying input drive.
A theoretical framework for understanding the self-organization principles of such a network with spatially local emerging oscillations is offered by the theory of weakly coupled oscillators (TWCO). The TWCO describes under which conditions interacting (coupled) oscillators synchronize. The ability of a population of coupled oscillators [56] to synchronize is controlled by two opponent forces [55]: their detuning (Δintrinsic frequency) and their interaction strength (here through synaptic interactions), to which we refer as coupling strength. The region in the two-dimensional parameter space of coupling strength and detuning within which synchronization occurs is called the ‘Arnold tongue’ [55]. For conditions within the Arnold tongue (Fig. 2A), the oscillators converge on a common emergent frequency. Within the Arnold tongue, the initial (intrinsic) frequency difference between the pair is replaced by a consistent phase difference, where the oscillator with the higher intrinsic frequency leads in phase. Outside the Arnold tongue, intrinsic frequency differences are maintained, precluding a consistent phase relationship (i.e., phase precession instead of synchronization).
We first illustrate these ideas in simulations from a gamma model consisting of two interconnected PING networks (Fig. 2B, see also [18]). The two PING networks were both identical to the network introduced in Fig. 1C-F, with inhibitory neurons only projecting locally within their own network to excitatory (I→E) and inhibitory cells (I→I). The inter-network connectivity was comprised of excitatory-to-inhibitory connections (E→I, I receiving input from 8 E) and excitatory-to-excitatory connections (E→E, 8 per neuron). In different model simulations, the two inter-network connection types (coupling) were modulated jointly from 0 to 0.07 mS/cm2 (note that these values are an order of magnitude lower than in intra-network coupling, see Methods). Fig. 2C shows an example of simulation output with estimated LFP traces (top) and corresponding time-frequency representations (TFRs, bottom). In this example, the drive to the two coupled networks (here coupled with 0.004mS/cm2) was very similar (network 1/2 = 0.069/0.0635 mS/cm2). This resulted in closely matching oscillation behavior. We then used this model to study the effects of varying input drive differences and of varying coupling strength.
Fig. 2D-F shows the detailed effects of three combinations of coupling strength and detuning (intrinsic frequency difference) on the ability of coupled oscillating networks to synchronize. Phase locking was estimated here based on the population response of each PING network (here LFP, see Methods). In Fig. 2D, network 2 (neurons 101–200 in simulated spike histograms) received (network 1/2 = 0.0598/0.0635 mS/cm2) more excitatory input than network 1 (neurons 1–100). Because of a sufficiently small difference in input and intrinsic frequency at the chosen coupling strength (a parameter constellation falling within the Arnold tongue, 0.02 mS/cm2), the networks synchronized at a common emergent frequency (∼35.5Hz). This is visible (from left to right) in the overlapping power spectra, in the consistent time difference between spikes of network 1 and 2 in the population spike raster, and in the narrow phase difference distribution (see Methods). The spike raster and the phase difference distribution also show that spikes in network 2 were leading spikes in network 1. In Fig. 2E, the excitation level difference in the networks were approximately reversed (network 1/2 = 0.069/0.0635 mS/cm2) while keeping coupling strength constant. Again, the networks synchronized at a common frequency (∼37Hz), but spikes of network 2 now lagged network 1. In Fig. 2F the coupling (cross E-E, E-I) between networks 1 and 2 was decreased from 0.02 mS/cm2 to 0.004 mS/cm2, while keeping the detuning constant (network 1/2 = 0.069/0.0635 mS/cm2), creating a condition falling outside the Arnold tongue region. The two networks therefore did not synchronize but oscillated at different frequencies (network 1/2 = ∼36Hz / ∼38Hz). By systematically modulating the coupling strength and the detuning between the two networks, the Arnold tongue could be fully reconstructed: it emerged as a region of high phase-locking (Fig. 2G) characterized by a common emergent frequency (Fig. 2H) and systematic phase differences (Fig. 2I). Fig. 2J shows a comparison of the intrinsic frequency of one network observed in the absence of coupling (dashed line), and its emergent frequency when coupling was set to 0.04mS/cm2 (black solid line). As in Fig. 1F, the intrinsic frequency depended linearly on the input level. However, the emergent frequency during coupling displayed a non-linear function, whereby frequency was constant within the range of the Arnold tongue. Within that range synchronization was observed; meaning that a consistent phase relationship emerged (Δ phase, red line). The phase relationship was linearly related to the input level difference.
We described above how the behavior of two interacting PING-networks can be understood in the framework of TWCO. However, to understand the self-organization principles of gamma oscillation activity in a cortical area, one needs to take into account interactions among large numbers of interconnected neurons that constitute multiple potential local PING networks. The local networks may be more easily comparable to anatomically distinct ‘columns’ (which may or may not underlie functionally defined columns) in some sensory systems (e.g., barrel cortex) than in others (e.g., visual cortex), but this correspondence is not critical to our argument. In this study, we used continuous local connectivity and spatially specific input drive as the more general case. We chose a model architecture in which neurons were organized along a ring (Fig. 3A, see Methods), to avoid border effects and thus facilitate analysis. For the generation of the PING mechanism, E-cells and I-cells were designed to have relatively strong local interactions (E-I, I-E) between neighbors (E-I = 0.23 mS/cm2, 10 per neuron, I-E = 0.13 mS/cm2, 10 per neuron). Inhibitory-to inhibitory connections (I-I) further supported the PING mechanism (I-I = 0.1 mS/cm2, 4 per neuron), yet they were not critical [17,18]. In addition, weak but numerous RS to RS excitatory connections (E-E) were added (E-E = 0.03 mS/cm2, 25 per neuron) [16]. The topographic input to RS neurons was modulated sinusoidally around the ring (shading in Fig. 3A). I-cells received most of their input drive from nearby located E-cells (for details, see Methods). In the ring-network simulations E-cells had similar spike rates to I-cells and both had spike rates close to the gamma oscillation frequency. This allowed us to use smaller but stably synchronized networks to increase computational efficiency. However, we will describe below that our results can be extended to large sparse-firing gamma oscillation networks in which E-cells (RS) fire much less than the I-cells (FS) and below the gamma frequency.
In Fig. 3B, an example simulation output is shown, displaying the spike raster for the entire network (red, E-cells; blue, I-cells). Neuronal spiking was synchronized in the gamma range (∼25–35Hz), but individual neurons displayed spike timing differences relative to each other that were related to the input-drive differences. Fig. 3C-F describes the detailed relationship between synchronization and input. This relationship will be described both in terms of the strength of phase locking among neurons and of the phase differences among synchronized neurons. Phase locking was estimated by the peak of the cross-correlation histogram computed over the respective simulated spike trains, and the phase difference by the lag of the cross-correlation peak (divided by cycle length, see Methods). We focus here on E-cells but all observations for E-cells have been replicated for I-cells (see similar behavior of I-and E-cells in Fig. 3B).
We found that the spatial extent of phase-locking in the network differed along the sinusoidal input function (Fig. 3C) in close relation to the level of detuning approximated here by the squared input derivative (red line). Neurons receiving their input around the trough and the peak of the spatial sinusoidal input engaged in spatially larger ensembles of synchronized rhythmic activity compared to neurons along the slope of the sinusoidal input (Fig. 3D). To illustrate this, phase locking strength distributions are shown for three reference neurons. Each distribution refers to locking between a reference neuron and all other neurons in the network. The three reference neurons were labeled as neuron a (red), neuron b (green), and c (blue), located respectively near the trough, slope, and near the peak of the sinusoidal input function. The spatial extent of phase-locking (Fig. 3D) decreased with increases in the slope of the sinusoidal input, corresponding to increases in detuning, yielding much larger distributions for neurons a and c than for b. Specific features of the input also influenced the distribution of phase locking strength. The spatial distribution of phase-locking for neuron b, situated where the slope of the sinusoid was steepest, was not only small but also symmetric. By contrast, the larger distributions for neurons a and c, were asymmetric, with a skew towards neurons receiving more similar input drive. Hence, despite the symmetric synaptic spatial coupling for each reference neuron with its neighbors, their spatial phase-locking distributions with neighboring neurons differed. This reflected the spatial variation of input drive to neurons in the vicinity of reference neurons a, b and c. Moreover, the relation between input drive and synchronization also led to characteristic phase differences among synchronized neurons (Fig. 3E). This is illustrated for each of the same three reference neurons (only phase-relations shown if >.25 phase-locking, see Methods). The reference neurons had systematically leading phase relationships with neurons receiving a lower input drive, and a lagging phase relationship with respect to neurons receiving a higher input drive.
We now consider the combined results of all E-cells in the network. Fig. 3F shows a phase-locking matrix in which the phase-locking values between all possible pairs of E-cells in the network (160 × 160 E-cell pairs) are shown. Neurons around the peak or trough of the sinusoidal input function formed large assemblies of synchronized units. Neurons along the steepest slope of the input function only synchronized with their immediate neighbors (narrow regions of bright color at the centre and extreme ends of the diagonal). Note that neurons close to, but not exactly on, the peak/trough had asymmetric distributions of phase-locking values, in spite of their symmetric connectivity.
In Fig. 3G, phase differences are also shown for all E-cell pair combinations. Within regions of high synchronization, neurons with higher input drive (negative lag shown in blue) led neurons with lower input drive (positive lag shown in red). Both the behavior of phase-locking and phase-relation as a function of detuning are in agreement with TWCO. The detuning magnitude (large at the sinusoidal slope and small around the peak/trough) strongly determined whether neurons could synchronize. If synchronized (within the Arnold tongue), the sign and magnitude of detuning defined the phase-relation.
The synchronization properties of the ring-PING network in Fig. 3 depended on the connectivity patterns. First, sufficiently strong E→I as well as I→E connections were required to allow for PING type synchronization [17]. Further, we observed that the exact synchronization properties depended on the number of excitatory connections in relation to the number of inhibitory connections (S1 Fig.). In the case of more numerous (or stronger) E-E connectivity, the spatial extent of synchronization was larger around the sinusoidal peak compared to the trough. In contrast, in case of strong I-I or I-E connectivity, the spatial extent of synchrony was larger around the trough compared to the peak. This seemingly odd result can be understood if one considers the influence of excitatory vs. inhibitory input in terms of the phase response curve (PRC). Whereas excitatory connectivity will tend to advance the phase of a next spike, inhibitory connectivity will delay the occurrence of a next spike. Neurons synchronizing to other neurons with excitatory connections alone will most optimally entrain neurons with intrinsically lower frequencies (as excitatory connections speed them up) and hence synchronization extends further around the peak. In comparison, inhibitory connections entrain best neurons with intrinsically higher frequencies and will therefore lead to stronger synchronization around the trough of the sinusoidal input. Therefore the balance of inhibitory and excitatory interaction is critical for understanding how PING networks will self-organize depending on input-drive. We stress that future studies should further investigate how connectivity patterns affect the gamma-mediated temporal organization.
In the simulation analysis presented in Fig. 3, network performance was analyzed in terms of simulated spike output, where phase-locking strength and phase differences were derived from spike cross correlations. In experimental studies, gamma oscillation properties are often investigated in terms of the Local Field Potentials (LFP), which is a population aggregate signal (mainly reflecting synaptic potentials, [89]). We therefore conducted a similar analysis on E-cell population activity (representing a LFP-like signal) to test whether the same phase-locking and phase-relation behavior could be observed. Further, we were interested whether input noise affected single neuron spike rates differently than the local population oscillation frequency. To estimate the local LFP-like measure, we aggregated the network spiking activity at each E-cell reference position with an exponentially decaying spatial function (see Methods). We termed this the local population average (LPA), to make clear that it is not the LFP, yet sharing the property of being a population signal. Results are shown in S2 Fig. We observed the same behavior of phase-locking and phase-relation patterns for the LPA estimates. The properties of the network were relatively robust against input noise (S2 Fig.). Generally, the higher the input noise, the smaller the extent of synchrony [18,55]. Further, at higher input noise levels, we observed that the spike rates were no longer highly locked to the local gamma frequency (estimated based on LPA) and, if rates were estimated over a long time window (here 5 sec), the single spike rates could reflect input differences between neurons, despite being locked to the same gamma oscillation frequency. The relationship between population gamma frequency and single neuron spike rates as well as important issues related to noise and the encoding time window will be elaborated further in the Discussion section.
One limitation of the above presented ring-PING network was that the E-cells and I-cells had similar spike rates, both in the range of the local population gamma frequency. However, experimental studies suggest that neurons, in particular pyramidal neurons (RS-type, [66,90,91]), have spikes rates lower (sparser) than the gamma oscillation frequency (they do not spike each gamma cycle). It has been shown in theoretical studies that sparsely firing PING network regimes exist as long as the number of neurons is sufficiently large [92]. We therefore replicated the findings shown in S3 Fig. in a larger network with Izhikevich neuron models [67], which have higher computational efficiency than the Hodgkin-Huxley neuronal model, but still generate realistic RS and FS spiking patterns. The ring-PING network consisted of 4000 E-cells and 1000 I-cells. Whereas the I-cells still spiked close to the gamma range being around ∼40Hz, the E-cells had spike rates around ∼12Hz. Although the E-cells showed spike rates much lower than the gamma oscillation frequency, we still observed the phase-locking and phase-relation among E-cells as described in Fig. 3.
The results from the two interacting PING-networks (Fig. 2) suggested that reliable phase differences corresponded to small local differences in input (small detuning), whereas frequency differences reflected larger input differences. The same could be observed for the ring-PING network driven by spatially varying input (Fig. 3). Neurons interacting with small detuning (at the peak or trough of the sinusoidal input) exhibited reliable phase differences, whereas neurons interacting at larger detuning values (at the steepest slope of the input function) showed reduced synchrony and large (emergent) frequency differences. This indicates that information about input drive differences might be present both in frequency and phase in a complementary manner.
We therefore extended our analysis of the ring-PING network to investigate neural coding by quantifying explicitly the relationship between the input patterns and the neuronal responses in terms of their frequency and phase-relation. We will first describe the coding types and their derivations. The stimulus (Sorig) was the spatially-defined sinusoidal excitatory drive to the E-cells. The first coding type was the (emergent) ‘frequency code’ (Sest (ω)). We explicitly mean the frequency that would be (experimentally) measurable in a network. In our ring-PING network described above, single neuron spike frequencies (rates) were close to the (LPA) gamma frequency and neurons were strongly locked to the rhythm. Therefore LPA gamma frequencies or single spike rates gave here similar estimates (see Discussion below). The second coding type is the ‘phase code’ (Sest (θ)). We described above the phase-relations between neurons in the ring-PING network as function of the spatial sinusoidal input. When neurons were synchronized, hence sharing a common frequency, the neuron with higher drive occupied a leading (earlier) phase. In our network, multiple oscillatory frequencies were present and we therefore had to define the common oscillation frequency by the group of neurons to which it had substantial phase-locking. This was implemented by weighting each phase-relation between neurons by their phase-locking strength. The Arnold tongue relationship states also that a phase-difference between two oscillators depends on their coupling strength. Hence, to achieve more exact estimates of the input differences from the phase differences, we needed to make them independent of coupling. This was implemented by multiplying a given phase-difference by the coupling strength between neurons (see Methods and Discussion). This operation was necessary as a phase-difference between strongly coupled neurons corresponds to a higher input difference than the same phase-difference between more weakly coupled neurons. To summarize, the phase-code was calculated for each neuron as the average phase-relation to all other neurons weighted by their phase-locking strength and coupling strength. In the combined ‘frequency and phase code’ (Sest (θ,ω)) both the phase code and the frequency code were summated.
In Fig. 4B, we reconstructed the spatial sinusoidal input (Sori) of the ring-PING network based on the E-cell spike count (frequency code Sest (ω)), the phase-relation between E-cell spike trains (phase code Sest (θ)) or by combining both sources of information (combined code Sest (θ,ω)). The frequency code exhibited plateaus around the peak and trough of the sinusoid, where synchronization was strongest. The phase code followed the variation around the peak and trough of the sinusoid, but could not follow the larger input differences (e.g. overall difference between peak and trough). By combing both coding types the variation of the original sinusoidal input was well reconstructed. This was quantified by computing the mutual information (MI, [80], see Methods) between Sorig and Sest. The lowest reconstruction performance was achieved by the phase-code (MI = 0.18), followed by the frequency code (MI = 0.65), with the highest MI for the combined code (MI = 0.92). The contribution of each coding type will depend on the exact input characteristics. For example, the higher the synchrony within a network (e.g. by lower amplitude of the sinusoidal input function) the more information the phase code will add to the frequency code (see discussion). These results indicate 1) that phase coding is most suited to resolve fine (small input differences) and local (high coupling) input variation; 2) that phase coding represents a relative Δrate-phase transform [45]); 3) that phase coding depends on both input difference and coupling strength (Arnold tongue); and 4) that phase coding can add complementary information to frequency coding.
The Hodgkin-Huxley PING network simulations in Fig. 1–4 have shown that the input-dependent gamma synchronization can be well understood within the TWCO and the Arnold tongue [55]. In the following result section, we show that the PING spiking neural network can be successfully reduced to a basic model [52,53] of weakly-coupled oscillator networks, the phase-oscillator model (Kuramoto model, [56]). The reduction to the phase-oscillator model allowed us to investigate the oscillatory properties of much larger topographic networks exposed to natural complex input patterns due to the computational efficiency. However, first we will describe the rationale of reducing local PING networks to abstract phase-oscillators. We will then show that the exact same behavior of the ring-PING network can be reproduced by a ring-phase-oscillator network.
A single phase-oscillator is characterized by an intrinsic (natural) frequency that determines how fast the phase-variable (the central variable of the model) evolves over time. The intrinsic frequency is the frequency that characterizes the oscillator in the absence of interactions with nearby phase-oscillators. A local PING network consisting of a few E- and I- cells is considered here as equivalent to one phase-oscillator (Fig. 5A). The population frequency of the local PING network (LPA, Fig. 5B) would be similar to the frequency of the phase-oscillator. The instantaneous phase [75] of the local population rhythm (Fig. 5C) would be equal to the output-variable of the phase-oscillator model (Fig. 5D). In a network of connected (coupled) oscillators, there is not only a reduction in terms of units (from a number of E- and I- cells to a single phase-oscillator), but also a reduction in the complexity of connectivity. Connectivity (coupling) is defined here in terms of one oscillator advancing or delaying the phase of the other in a manner that is defined by the phase relation between them. Coupling strength refers to the magnitude of the modulation of the phase-variable in two oscillators which is a function of the ongoing phase-relation between the phase-oscillators. This function is referred to as the phase-response curve (PRC). It is a sinusoidal function as defined in the Kuramoto model of weakly coupled oscillators, with a clear attractor at phase 0. For example, if oscillator 1 at a given moment is trailing oscillator 2, oscillator 2 will push the phase of oscillator 1 forward while oscillator 1 will delay the phase of oscillator 2 (assuming sufficiently similar frequencies). This means that given equal intrinsic frequencies, a network initialized with random phases in each oscillator will tend to converge towards the attractor phase. In the case of a local-PING network, the interactions with nearby neurons are exerted through excitatory connections (E-E, E-I) and inhibitory connections (I-E, I-I) which together determine the effective coupling strength and the phase response curve (PRC) of our network. A further difference between oscillator networks and PING networks is the manner in which input modulates ongoing interactions in the network. In PING networks, the oscillation frequency emerged from the interaction between network properties and excitatory input drive. In phase oscillator networks, the intrinsic frequency of a phase-oscillator was set before simulation based on a function (as established experimentally) linking input strength to oscillation frequency. Further, whereas in the PING network the gamma rhythm might be not sustainable in some conditions, e.g. due to low input drive, a phase-oscillator will always oscillate at any arbitrary frequency. Overall, it must be emphasized that even though not all complexities of the PING network can be captured by a phase-oscillator model; we argue that it captures the most characteristic properties of PING network behavior.
To illustrate that phase-oscillator networks capture the behavior of the PING model, Fig. 6 describes a ring-phase oscillator neural network similar to the ring PING network (Fig. 6A), in which 160 phase-oscillators were locally coupled along a ring (see Methods). The intrinsic frequency of each phase-oscillator was set by a sinusoidal input function (defined over the ring). The simulation output can be seen in Fig. 6B, which shows the output phase-traces of the 160 phase-oscillators. In Fig. 6C-D, we computed the phase-locking and the phase-relation matrix between all phase-oscillators. The results shown resembled those obtained for the ring-PING network (Fig. 3F-G). The spatial extent of phase-locking was larger around the peak and trough of sinusoidal function. The phase-relation patterns within the region of phase locking were the same as in the ring-PING-network, so that phase-oscillators with higher intrinsic frequency had a leading phase compared to other phase-oscillators. In Fig.6 E-F, we tested the contribution of the different coding types to input reconstruction (see Fig. 4B-C for comparison). Notice that Sorig (stimulus input) corresponds here to the intrinsic frequency set by an input function and not to excitatory drive. The same results (here with sinusoidal intrinsic frequency fluctuation of +/−3Hz) were obtained as in the ring-PING network. The best input reconstruction was given by the combined frequency and phase code (MI = 0.95), followed by the frequency code (MI = 0.73), and lastly the phase-code (MI = 0.17). We also reproduced the asymmetries in spatial synchronization (around peak and trough of sinusoid) which in the ring-PING network were induced by changing the amount of excitatory connections in relation to the amount of inhibitory connections (see S1 Fig.). The asymmetries in spatial extent of synchronization were obtained in the phase oscillator network by modifying the phase-response curve (PRC). To model a dominance of excitatory connections, we set all values below zero (phase delay) to zero, and to model a dominance of inhibitory connections we did the opposite. This resulted in the same asymmetries in synchronization around the peak or trough of the sinusoidal function (S2D Fig.).
We took advantage of the computational efficiency of the phase-oscillator model and extended our analysis to a 2D 100×100 lattice networks. The network consisted of 10.000 phase-oscillators with a total of 108 possible connections. Connections among phase-oscillators decreased exponentially in strength as a function of distance on the lattice as an approximation for cortical horizontal connectivity (e.g. V1, [73]). Note that the lattice network was an abstract model of a cortical area aimed to reflect only the essential characteristics of a sensory cortical area (topographical map, spatially local connectivity, feature map). The model was aimed to capture the essential input-dependent self-organization principles of cortical oscillations, in particular gamma oscillations. We tested the network behavior using visual images representing natural and complex intrinsic frequency variation better (Fig. 7A). Inspired by experimental observations of a close link between visual contrast and gamma oscillation frequency in macaque visual cortex V1 and V2 [33], we used the local contrast of natural visual stimuli to define the intrinsic frequencies of the phase-oscillators. To that aim, we used 80 (grayscale) natural images from an online database (Berkeley segmentation dataset, see Methods). The images were down-sampled such that each pixel of the image corresponded to one phase-oscillator. Local contrast was estimated using a root-mean square measure [25] with a spatial kernel of 3 pixels. From the online database, information on the location of boarders between objects, or segments in the image (segmentation borders), defined by 30 human observers [71], was also available.
In Fig. 7B, we show in the left column the synchronization (black = no phase-locking, yellow/white = max phase-locking of 1) and in the right column the phase-relations (blue = earlier phase, red = later phase, white = below threshold, see Methods) of two reference oscillators (black dots) as compared to all other oscillators in response to an example image. We will use these examples to show how the synchronization fields and phase-relations adapt to local changes of contrast/intrinsic frequency. The ‘synchronization field’ refers to the spatial extent over which a reference oscillator synchronizes with neighboring oscillators [46]. The synchronization fields were asymmetric, despite symmetric coupling in the network. Within the synchronization fields, off-zero phase-relations could be observed. This was due to the positions of the two example reference phase-oscillators, which were chosen to be close to an object border (here a bear) with one reference oscillator located just outside the main object (top row Fig. 7B) and the other just within (bottom row Fig. 7B). The object border was associated with high local contrast variation (hence large detuning). The asymmetry of the synchronization fields followed from the fact that synchronization drops with the rapid increase of local contrast when moving from the reference oscillator towards the border (increasing detuning) while the converse was true when moving from the reference oscillator away from the border (decreasing detuning). Here, synchronization extended far from the reference oscillator towards the interior or ulterior surfaces (because the small detuning within surfaces permits synchronization over larger spatial extents). Furthermore, within the synchronization fields, the oscillators closer to the border led in phase compared to the reference oscillator, because the border had higher contrast/intrinsic frequency, while the converse was true when considering oscillators away from the border. These example synchrony fields indicate that phase-oscillators in the 2D lattice model behaved similarly to those in the ring-phase oscillator and ring-PING network models. Moreover, the data suggest that the synchronization fields might capture specific aspects of the statistics of contrast distributions in natural images. To further explore this point, we tested systematically in Fig. 7C how synchronization was affected around segmentation borders in 80 natural images (available from the Berkeley segmentation dataset [71], see Methods for more details). We first tested whether contrast variation was significantly modulated around segmentation borders (see Methods). To quantify these effects, we defined 1-dimensional spatial windows of ± 15 pixels centered on segmentation borders. We then aligned the different windows and averaged them for each image separately (see Methods for more details). Population statistics are based on these average windows per image. We then calculated the averaged absolute spatial derivative of contrast values (equivalent to detuning) along each window. Fig. 7C (top) shows a steep change in contrast as a function of distance to the border. Mean contrast variability at the center of the window, at the border, was significantly different from the extremities of the window (paired t-test: t = 7.35, df = 79, p <0.001). We then quantified the change in synchronization as a function of distance to a border. For initial population analysis (N = 80), we selected three reference oscillators. Reference oscillators a and c (Fig. 7C bottom) were located 3 pixels away from each side of the border, and reference oscillator b was on the border. Relative to reference oscillators a and c, synchronization fields showed a much more rapid decline of phase-locking strength towards the border than away from the border (see S4 Fig. for statistics and more details). Thus overall, the asymmetry of synchronization fields around reference oscillators a and c in Fig. 7E matched with the asymmetry of synchronization fields around neurons a and c in Fig. 3C-E of the ring-PING network model. This input-dependency of spatial synchronization suggests, in line with previous studies [21,29,46,47,93,94], that oscillatory synchronization might be a useful tool for clustering operations, for example for visual segmentation.
So far, we have described the behavior of the phase-oscillator model using particular image examples. We now analyze the network behavior in terms of the principles that underlie its self-organization behavior independent of the specific image providing the input. If the input-dependent self-organization of the phase-oscillators is mainly governed by principles of the TWCO, then one should be able to reconstruct the Arnold tongue from the simulation output. To test this, we determined for each phase-oscillator pair its coupling strength (direct connections) and detuning (as derived from the 80 input images). Fig. 8A shows that in the two-dimensional parameter space of detuning and coupling strength an Arnold tongue could be observed in terms of phase-locking (Fig. 8A I), frequency difference (Fig. 8A II) and phase-difference (Fig. 8A III). Fig. 8A IV represents a horizontal cross-section of the Arnold tongue (coupling strength = 0.62) where frequency differences (black) and reliable phase differences (red) are plotted. The dashed line denotes the intrinsic frequencies. As described in Fig. 2 for PING networks, phase differences better resolved smaller intrinsic frequency variation, whereas frequency differences reflected best the larger differences. The Arnold tongue reconstruction was reliable and reconstructions from individual images looked very similar to the averaged one shown in Fig. 8A. The Arnold tongue properties were similar to the one described from the two interacting PING networks (Fig. 2G-J). This analysis confirmed that the phase-oscillator lattice model with complex natural detuning (intrinsic frequency variation) behaved very similar to the ring-PING network model driven by simple sinusoidal excitatory drive. Hence, information about the natural image stimulus should be available in a complementary manner at the level of frequency variation as well as phase variation. We therefore quantified the amount of information present in the above defined coding types, frequency and phase coding, as well as a combined coding type (Fig. 8). We used the same approach as used for the ring-PING network and ring- phase oscillator network. The stimulus Sorig to be reconstructed was the intrinsic frequency image defined by the local natural image contrast. The frequency code Sest(ω) was the mean (emergent) frequency of a phase-oscillator. The phase code Sest(θ) for a given phase-oscillator was the phase differences with all other phase-oscillators weighted by phase-locking and coupling strength (see Methods). The combined code Sest(ω,θ) was the summation of both former coding types. In Fig. 8B, the stimulus and the reconstruction estimates are shown for an example image. The stimulus reconstruction based on frequencies Sest(ω) appeared smoothed compared to the original stimulus Sorig reflecting the loss of fine spatial details. The reconstruction based on phase-relations Sest(θ) resembled a second derivative of the original stimulus Sorig. The phase code reflected well local and fine details but it did not reflect the absolute contrast/intrinsic frequency level. A fair reconstruction of the original stimulus was achieved by the combined frequency and phase code Sest(ω,θ), which indicates that information from the frequency code and phase code were complementary. In Fig. 8C, we quantified information content, by estimating the (normalized) mutual information (MI) between the intrinsic frequency image Sorig and the reconstructed image estimates Sest ([80], see Methods) For the example image in Fig. 8B, the MI was 0.44 for Sest(ω), the MI for Sest(θ) was 0.31, and the MI for Sest(ω,θ) was 0.69. Over the population of 80 natural images (Fig. 8c), the MI was 0.46 (SEM = ±9*10−3) for the frequency code Sest(ω), 0.28 (SEM = ±5 *10−3) for the phase code Sest(θ) and 0.67 (SEM = ±7*10−3) for the combined frequency and phase code Sest(ω,θ). A repeated measures ANOVA showed that all three codes were significantly different from one another (F(2,158) = 881, p<0.001), and all pair-wise comparisons were highly significant according to the Tukey’s HSD tests).
In the following, we will first discuss the underlying assumptions and limitations of our cortical gamma network model and relate them to previous modeling approaches. Then, we will turn to the implications of our findings for gamma phase coding and its relation to frequency/rate coding, stressing the distinction between the coding of larger versus smaller input variations, and the effect of noise, encoding time window and connectivity. Further, we will discuss the experimental and theoretical implications of the relative rate-to-phase transform, as proposed here for gamma synchronization, compared to an absolute rate-to-phase transform. This is followed by considerations on input-dependent spatial synchronization, in particular in the context of the phase-oscillator network response to natural images, and comparisons to related modeling approaches more specifically designed for image segmentation. We end with testable experimental predictions that follow directly from the present study.
‘Single oscillator model’ vs. ‘multiple oscillator model’. The exact underlying mechanism of gamma oscillations is still under debate, even though significant advances have been achieved over the last decades. A primary distinction [17] has been made between interneuron-network gamma (ING) versus pyramidal-interneuron network gamma (PING). However, we will first focus on another key model distinction that has received much less attention so far. In one class of models, inhibition acts as reference clock to which excitatory neurons at different cortical locations are entrained at different phases [15,84,95,96], whereas in another class of models I- and E-cells at different cortical locations represent different ‘clocks’ (oscillators) that synchronize at different phases [18,52,54]. We call the former the ‘single-oscillator’ model and the latter the ‘multiple-oscillator model’. The critical distinction is whether I-cells receive spatially local excitatory drive or whether the I-cell network acts as a (indiscriminative) single unit. In the latter case, the network can express only one dominant gamma frequency at a time. In principle, both types of models can be implemented in either PING or ING mode. Reliable phase-coding in a single oscillator model regime is not easy to achieve [18] and requires strong inhibition and relatively high spike timing precision. Phase-coding is determined mainly by how fast E-cells recover from inhibition (similar to latency coding, [97]). In the ‘multiple-oscillator model’ a precise entrainment of E-cells by their nearby I-cells is not essential as long as a gamma rhythm is produced. Phase-relations are established between nearby gamma rhythmic E- and I-cells through synchronization [55], during which phase-relations are determined by detuning and coupling strength. In our network models, we observed that the behavior of the PING network was largely consistent with the multiple-oscillator model. However, the two models do not necessarily exclude each other.
ING versus PING. It has been shown that an ING [83,84] network of mutually interacting inhibitory neurons can produce robust gamma oscillations, if the network is driven with sufficient excitatory input. In this model, local pyramidal neurons receive rhythmic inhibition from the I-cells. An extension of the model is the PING model, where the excitatory drive to the I-cells originates from the local pyramidal E-cells themselves [15,86]. In the PING model, the excitatory state of the pyramidal neurons influences the network rhythm, in particular the frequency. In comparison, in the ING model the frequency is determined solely by the excitability of the inhibitory neurons.
Both ING and PING mechanism likely coexist in cortical networks [92] and the dominance of one to the other may switch depending on network state [90]. Experimental studies suggest that cortical gamma oscillation in a stimulus-driven state show properties consistent with the PING model [15,90]. However, the essential TWCO properties we described in our networks are not restricted to PING networks. An ING network with locally defined connectivity and excitatory drive will exhibit similar behavior. Hence, our modeling results are expected to be independent of specific PING and ING network configurations. However, in the case of E-cells receiving rhythmic inhibition from an ING network that either does not receive a spatially-defined drive or is coupled in a manner such that the ING network acts as a single unit, then the regime is not expected to be in agreement with TWCO (‘single-oscillator model’, see above). The same is expected for a PING network where the E→I connections are all-to-all, such that all I-cells receive the same mean input from E-cells.
Assumptions of the theory of weakly coupled oscillators. Hoppenstaedt and Izhikevich [52] formulated two basic assumptions in which oscillatory interactions in the cortex would be expected to occur in the regime of weakly coupled oscillators. First, they assumed that oscillations are internally (autonomously) generated and second, they assumed that the coupling is weak between oscillating neural populations.
The first assumption was fulfilled for our study by the PING mechanism. The mean network input had no rhythmic components. Oscillations were generated by the interaction between E-and I-cells. The second assumption of weak coupling can be understood as oscillatory interactions between units leading to phase shifts (as defined by the PRC), but not to substantial amplitude changes or perturbation of the rhythm-generating mechanism (or quenching, [55]). In the case of the two PING networks (Fig. 2), the cross-network connections were an order of magnitude smaller than the within-network connections. We observed small phase-dependent amplitude fluctuations (corresponding to a partially synchronized state), but they did not substantially affect the phase-trajectory of the network oscillations. In the ring-PING network (Fig. 3), connectivity was spatially continuous and no columnar structures were assumed, thus there was no distinction between within and cross network connectivity. The synaptic connectivity strengths were in the range as normally used for PING networks [18] and the behavior was stable for a large range of different connection strengths. As described in Fig. 5, for the reduction of the ring-PING network to the ring-phase-oscillator model we assumed that single phase-oscillators could be equated to pairs of E-cells and I-cells of the PING network. Further, we assumed that the complex interactions (E-E, I-E, E-I, I-I) could be approximated by a single PRC-defined connection type. Weak coupling in this context means that the synaptic coupling between neurons did shift the spike timing and but neither increased the firing rate substantially nor interfered with the spike generation mechanism. The comparison of the ring-PING network with the ring-phase oscillator network indeed revealed striking similarities (compare Fig. 3 with 6). In conclusion, we showed that our weakly coupled oscillator network conformed to the assumptions of the TWCO. In addition, our modeling data indicates that discrete columnar network structure (coupling of many individual PING networks) is not a necessary condition to investigate PING-type oscillations in the weakly coupled oscillator regime.
Note that principally, the TWCO can be applied to any frequency band. However, to apply the TWCO in a valid manner to rhythms recorded in the brain, the mechanism generating the rhythm should be characterized by a link between excitation and intrinsic frequency, and by the possibility for different frequencies to exist in neighboring neural populations. Arguably, the generative mechanisms are best understood for gamma [17], and the emergence of localized differences in frequency has to the best of our knowledge only been demonstrated in the gamma range [43]. Therefore, the implications of our work are meant to be restricted to gamma oscillations.
Networks with sparsely firing neurons. In Fig. 3, the E-cells had firing rates close to the gamma oscillation frequency (∼30–40Hz), which might be considered as unusual for regular spiking pyramidal neurons (RS, [66]). E-cells often spike at much lower rates than the gamma frequency, and in the hippocampus spike rates can be as low as a few Hertz despite gamma oscillations in the 20 to 80Hz range [98]. It has also been described in the neocortex that Layer 2/3 networks display more sparse-firing properties compared to Layer 4 [99]. In contrast, I-cells of the fast-spiking type (FS) have firing rates that can be close to the network gamma rhythm [15]. To demonstrate that the network behavior described above was not restricted to networks with fast firing E-cells, we constructed a network in which the E-cells had much lower spike rates. It has been described that such ‘sparse’ networks need a sufficient number of neurons to reach stability [92], as each cell contributes a spike only every few cycles. For computational efficiency we used Izhikevich-type neurons [67] instead of Hodgkin-Huxley (HH) type to increase the network size by an order of magnitude. We were able to replicate the findings from the smaller HH-network in a large Izhikevich-type network where E-cells had 3–4 times lower firing rates than the gamma oscillation frequency. Hence, the essential behavior described generalizes to more sparsely firing networks.
Topographic phase-oscillator lattice model. The topographic phase-oscillator neural network was aimed to represent a simplified sensory cortical area (inspired by V1) [73] with predominantly local connectivity where input features are smoothly represented over cortical space (feature map). We emphasize however that our topographic network model is not a strict model of a particular sensory area (such as V1 with columns and hyper-columns, layers). Moreover, we used natural visual contrast stimuli as an example for natural input pattern to a visual topographic network, but the described principles of self organization may well apply across different sensory systems. However, we suggest that the main principles underlying the network self-organization are perhaps most easily experimentally testable in a cortical area like V1.
Conduction delay and asymmetric connectivity. In the PING networks the conduction delay in the interaction was mainly determined by the synaptic postsynaptic potentials [1–2ms]. The conduction delay did not change with distance between neurons. However, conductional delays of horizontal interactions over cortical space can be significant [73,74]. Conduction delays may affect the phase-relation as well as phase-locking. For the phase-oscillator lattice model we included a time-delay term as a linear function of spatial distance (slope of 0.4 ms/pixel, offset 1ms). By including this term we observed that the spread of spatial synchronization became more limited, helping to restrict the ‘synchronization fields’. We did not include the time-delay term explicitly in the reconstruction coding formula, but because it affected the phase-locking strength, it was implicitly included in the weighting of phase-relations by phase-locking.
An additional factor that is of importance in our models was the assumption of symmetric coupling, and how it interacts with time-delays. Whereas under conditions of symmetric coupling we expect time-delays to affect phase-locking but not phase-relations, under asymmetric coupling we expect pronounced effects also on phase-relations [59]. More research will be needed to investigate the effect of time delays and asymmetries in connectivity on input-dependent gamma synchronization.
Background. The problem of neural coding of sensory signals is a central topic in neuroscience with a long history (for review [1,100–102]). Despite substantial advances over the last decades, many fundamental issues remain unresolved. We will give a short review of different perspectives on this issue. A first important conceptual opposition in the literature is that between rate and time coding. The ‘rate-coding hypothesis’ has a long tradition founded on early discoveries of a close relationship between spike rate (frequency) and sensory variables. Rate-codes however are constrained by the length of the encoding time window (integration time constant), which defines its resolution. Another limitation is represented by the saturation properties of many spiking neurons in the low and high input range, that is, the limited dynamic range of neuronal spike rate. In this framework, variability in spike times over time is considered as noise to be averaged out either over time, or over many neurons. As an alternative to rate coding, the ‘time-coding hypothesis’ was introduced later, stating that precise spike timing contains a significant amount of information about the stimulus [103,104]. Time-codes require short integration time windows (coincidence detectors, [105]). It is well established that single-neuron spike timing contains additional information about time-varying stimuli [102]. However, the idea that spike-timing relation between different neurons is meaningful has received less acceptance in the neuroscience community [2]. A further distinction has been made between ‘independent single-neuron coding’ and ‘assembly/ensemble coding’. In an ‘independent’ scenario, each neuron codes its input by its rate and position within a feature map (position coding) independent of other neurons. All the information is represented in the rate of the single neuron (input-rate transform) and the position it has within the network. However, experimental observations have shown in many studies [106] that correlations exists between spike trains at various spatial and temporal scales. We consider here precise correlations of spiking timing between neurons (synchrony) in contrast to slow time varying (trial-by-trial) fluctuations in spike rate (known as noise correlation, [107]). There have been various experimental studies showing precise spike synchrony between neurons on a fast-time scale [21,108,109]. The observations of spike synchrony indicates that additional information about the stimulus might be present in the relative spike-timing between neurons, which has led to the formulation of the ‘assembly/ensemble coding’ hypothesis [3,6,47,110]. Here, information is represented in the exact spiking pattern of several neurons. The relationship between oscillation phase coding (assembly code) and single spike rates or population frequency has not been well studied.
In the following sections, we will discuss in more depth the implications of our theoretical results for the understanding of gamma phase coding and its relationship to rate/frequency coding.
Complementary coding between frequency and phase. Our theoretical analysis has shown that oscillating neural networks with local connectivity represent input patterns in frequency and phase according to the TWCO. Of particular importance is the Arnold tongue that describes, as a function of coupling strength and input difference, the transition of frequency coding to phase coding. Reliable phase coding can only exist if neurons are synchronized (phase-locked, [18]), or in other words converge on a common frequency. That is the reason why frequency and phase coding are in principle complementary, because the process of synchronization ‘transfers’ the information (detuning magnitude) represented at the level of frequency differences into phase differences (outside vs. inside the Arnold tongue).
We have shown in our models that phase coding can add significant information about the stimulus. The level of contribution of phase coding will depend both on the input as well as on the coupling characteristics. A network will rely more on frequency coding if the coupling (interaction) between neurons is weak/sparse and input variability is high, whereas it will rely more on phase coding if coupling is strong and input variability is low. We argue that in many cases the network will be situated between these two extremes and would profit from combining both coding types.
Previous theoretical studies have already suggested the TWCO as the underlying model of phase-coding. Hoppenstaedt and Izhikevich [52] discussed the theory of weakly coupled oscillators in the context sensory cortical columnar processing. Moreover, Tiesinga and Sejnowski [18] discussed the behavior of multiple interconnected PING networks using TWCO. They were able to reproduce the experimentally observed gamma phase coding of stimulus orientation in primary visual cortex [34]. In their modeling study, each PING network had a different orientation tuning. During ‘presentation’ of a particular stimulus orientation, the PING networks synchronized and the PING network with strongest input (optimal orientation) led in phase. In other words, the PING networks operated within the Arnold tongue regime (in which detuning was translated into phase differences). They also described that PING networks with weak coupling showed stronger phase shifts than PING networks with stronger coupling, in line with experimental data [34]. We replicated the main observations of Tiesinga and Sejnowski [18], but also extended their results. First, we confirmed their observations also in networks with continuous spatial connectivity (no assumption of columns). Second, we described in a systematic way networks which ‘translated’ input/intrinsic frequencies into phase-relations as well as (emergent) frequencies. Moreover, we quantified explicitly the contribution of frequency and phase information to the encoding of simple as well as complex natural stimuli.
We described above the fundamental complementary nature of frequency and phase coding, as predicted by the Arnold tongue. However, in assessing the complementary aspects of frequency and phase coding, several factors need to be taken into account. First, the transition between frequency (asynchrony) and phase coding (synchrony) is not sharp, but characterized by a state of partial synchrony in which both frequency and phase coding can be expressed. Second, noise has an important effect on the level of synchrony/partial synchrony. Third, the encoding time window is critical for the assessment of the contribution of phase and frequency coding. These points are also critical to evaluate the relation of oscillation frequency to single spike rates as described below.
Noise and partial synchrony. Our study indicates that oscillators with different input strengths/intrinsic frequencies are not necessarily precluded from synchronization, however, in order to synchronize they must first arrive at an emergent ‘compromise’ frequency. This is apparently at odds with experimental studies showing that different neurons can be phase-locked to a rhythm, yet still express different spike rates [34]. Further, this seems at odds with a study that reported significant gamma synchronization between separate neuronal populations despite different (mean) gamma frequencies [43]. Importantly, their findings can be accounted for within the TWCO by ‘partial synchrony’ [55], corresponding to an attraction towards a synchronous state during brief time periods, including the adoption of a common frequency, interspersed with periods of asynchrony and separate frequencies.
Noise, largely present in biological systems, plays an important role. As reported also by Tiesinga & Sejnowski [18], noise shrinks the border of the Arnold tongue, and increases the amount of partial synchrony [55]. Arnold tongue reconstruction from the PING-networks (Fig. 2) and the phase-oscillator lattice model (Fig. 8) exhibited large regions of partial synchrony with only a very small region where phase-locking was perfect (close to 1). Therefore, because partial synchrony inter-mixes states of synchrony and desynchrony (inside/outside of Arnold tongue) information about stimulus input in a noisy network will be represented (on average, yet not at the same time) in both the frequency as well as in the phase differences.
Single spike rates and encoding time window. Single neuron spike rates will be close or equal the common population frequency if neurons are highly synchronized to the population rhythm (as in Fig. 3). In noisy and sparsely firing networks however, the synchronization can be low and a single neuron spike rate can be relatively independent of the population frequency (yet still exhibit phase coding, [18]). This applies particularly to neurons with spike rates much lower than the gamma oscillation frequency, because higher-order Arnold tongues are narrow and hence much more sensitive to noise. This fits with experimental data where a rather low locking of single neurons to an oscillation rhythm is observed, especially for pyramidal neurons [90]. In this regime, single neuron spike rates contain additional information compared to the local population rhythm (see S2 Fig.) and both phase-relation and single neuron spike rate might contain overlapping (‘redundant’) information.
Yet, an important aspect to consider, is how much time is needed to reliably retrieve the relevant information [78,103,104]. Compared to a rate code, the phase-code can retrieve the information in much shorter time windows (in one or a few oscillation cycles). For example, considering a rate code, a 1Hz spike rate difference can be retrieved with a minimal 1000ms encoding (integration) time window, whereas with the phase-code a time window of 25–100ms might be sufficient for a 40Hz oscillation (1–4cycles) to get reliable estimates. Hence, although information about small input differences might be present in both phase-relation and spike rates of single neurons over longer time windows, a phase-code could be particularly beneficial for stimulus reconstruction on a fast-time scale (e.g. within typical saccade intervals of ∼300ms).
Δrate-phase transform. Analogous to the common input-rate transform of single neurons, it has been assumed that oscillations implement a rate-phase transform [15,44,45]. That means that the higher the input strength to a neuron, the earlier in phase the neuron will spike within the oscillation cycle. Therefore the phase gives a direct estimate of the absolute input levels. There is experimental evidence for this type of transform for slower frequency oscillations, in particular the delta/theta rhythm [44,45]. Strong experimental evidence has been obtained from hippocampal theta oscillations, where neurons spike at different phases depending on their input [111]. An explicit experimental test for a rate-phase transform was carried out by McLelland&Paulsen [45] for the hippocampal theta oscillations. They found that whereas theta oscillations implemented a rate-phase transform, they did not observe the same for gamma oscillations. This seemed to be in line with other observations that did not find a systematic relation between the input level and the gamma phase under natural stimulation conditions [44]. In addition, stimulus contrast modulations did not yield phase shifts in macaque V1 [43]. However, other experimental studies in macaque V1 have shown that gamma phase can code for orientation tuning [34], suggesting a role of gamma phase in stimulus encoding. Hence there are conflicting results whether gamma phase coding can represent input in the same way as has been reported for slower oscillations. Our computational analysis might help resolve these seemingly contradictory results. We have shown, in line with previous studies [18,52,54,59], that the gamma phase coding can be understood in terms of the Arnold tongue, where phase-relations depends on the input difference/detuning (for a given coupling value). Hence, the essential parameter is not the absolute input level, but the input difference between interacting neurons. We term this coding the ‘Δrate-phase transform’. This transform represents a relative encoding of relative input differences (detuning) between nearby neurons, irrespective of mean input levels. Changes in absolute input levels are in turn represented in the frequency of gamma oscillations. This is in line with the lack of rate-phase transform findings in the gamma range [45], with the finding of no phase-shifting with contrast [43] and the finding of phase-coding with orientation [34]. Orientation tuning is locally defined in visual cortex where nearby neurons are driven slightly differently by a given stimulus orientation. In this case, as shown in Tiesinga and Sejnowski [18], phase-coding should reflect the input differences between synchronized columns, independently however of overall input drive. We predict therefore that the gamma phase-coding of orientation should be insensitive to overall input strength, e.g. stimulus contrast.
Link between coupling and phase coding. According to the Arnold tongue, phase-relations between oscillating neurons are determined by input (intrinsic frequency) differences as well as coupling strength. This implies that for an exact interpretation of input differences, knowledge about the coupling values is required. In our reconstruction formula we multiplied the phase-differences with coupling values, such that phase-differences of more strongly coupled oscillators were weighted more strongly than from more weakly coupled oscillators. We conceptualize the included coupling term as a ‘prior’ (representing the general connectivity structure of a network) that might be used by the brain to optimize the input reconstruction performance based on phase. In general, we argue that the dependence of gamma phase coding on connectivity is of high interest and should be investigated in future studies, because it makes the code also sensitive to information (e.g. memory) imprinted in the network connectivity structure [60].
We observed synchronization between nearby locations and the formation of gamma synchronization fields which were shaped by anatomical connectivity constraints and by the spatial pattern of input. Note that the input variable (excitatory drive) in the present paper has been mapped on visual contrast, but we expect to see the same network self-organization for stimulus features other than contrast, such as orientation and motion, which also modulate gamma frequency [42]. The use of visual contrast as an example parameter is related to the fact that contrast changes induce especially robust gamma frequency modulations [33,42,43], and that spatial changes in contrast have been demonstrated to lead localized differences in gamma frequency in nearby neural populations in visual cortex [43]. To our knowledge, this latter finding, although expected, has not been empirically demonstrated yet for sensory features other than visual contrast. In sum, we suggest that the input to the model can be mapped on many sensory variables.
In our model simulations, synchronization fields emerged in regions of high local input similarity (low detuning) where nearby neurons shared similar input properties. The shapes of those fields were input-specific; being small in regions of high local input variance and large in regions of low local variance. Their shapes were asymmetric around reference oscillators close to large discontinuities in input such as borders in a visual image. An analysis of the topographic network’s response to natural images showed that synchronization fields extended away from segmentation borders (as indicated by human observers), and did not cross them, in agreement with the network behavior described above. This network response matches with the statistics of natural images, in which segmentation borders are often associated with large local contrast changes [112], whereas the interior of surfaces often shows more modest variations in local contrast. This indicates that input-dependent spatial synchronization may be meaningful way to cluster/integrate nearby neurons based on input similarity.
The potential of oscillating neural networks for meaningful segmentation of input patterns has been well established in computational neuroscience studies [113–115] (in particular visual segmentation), which were inspired by experimental studies on stimulus-specific gamma synchronization over the last decades [2,21,30]. However, the proposed segmentation mechanisms differ between studies. In some studies, the clustering is based on a phase-code only [49,88,94,113], whereas in others it is mainly based on de-/synchronization [47]. Clustering capacity of a neuronal network model similar to the one proposed in our study has been demonstrated [116]. Our model architecture differs strongly from model networks characterized by global synchrony, like the LEGION model (local excitatory global inhibitory oscillator network, [88]) or the PCNN (pulse-coupled neural network, [113]), where clustering is based on phase alone and the network has a single main frequency at any given moment. LEGION and PCNN are powerful for image segmentation tasks, yet, they are not accurate models of cortical gamma oscillations, which are characterized by local synchrony and variable oscillation frequencies. However, they might be more appropriate model for slower oscillation phase [95] or latency coding [117]. In our simulations we did not explicitly investigate the clustering/segmentation performance of the neural network per se. However, our results give a new perspective into that matter that might guide future research in field of image segmentation. In particular, the TWCO offers a more precise understanding on how synchrony and phase in a self-organizing network relate to stimulus input characteristics and network connectivity. According to the Arnold tongue, input variations are transformed into both frequency (synchrony) and phase variations, and therefore both might be useful for clustering. We further suggest that clustering based on phase and synchrony/frequency would be complementary and represent fine and coarse spatial scales respectively. Finally, the Arnold tongue is an appropriate framework in linking clustering based on connectivity with clustering based on synchronization.
From our theoretical analysis clear predictions can be derived that can be tested experimentally. We offer three key predictions:
For measuring the information content of gamma phase coding, e.g. during natural image stimulation in visual cortex, we predict that it is critical to measure the excitation differences between nearby locations, because we predict that gamma phase coding does not capture absolute input. We also expect that the frequency of gamma oscillations will contain a significant amount of information about the stimulus.
By manipulating both excitatory drive independently in two nearby cortical locations as well as the distance (coupling) between the locations, one should be able to reconstruct an Arnold tongue, assuming that cortical distance is correlated with a declining probability of connectivity. We expect that large excitation differences are expressed in oscillations frequency differences, whereas lower excitation differences will permit synchronization and a translation of the excitation differences into phase differences. This idea could be tested with optogenetic technology [118] by driving pyramidal cells with different intensities at different locations or by manipulating sensory stimulus features known to manipulate excitatory drive (e.g. contrast in V1).
Our ‘multiple oscillator model’ predicts that I-cells (e.g. FS) will also exhibit phase-differences, not just E-cells.
|
10.1371/journal.pgen.1008339 | A Sir2-regulated locus control region in the recombination enhancer of Saccharomyces cerevisiae specifies chromosome III structure | The NAD+-dependent histone deacetylase Sir2 was originally identified in Saccharomyces cerevisiae as a silencing factor for HML and HMR, the heterochromatic cassettes utilized as donor templates during mating-type switching. MATa cells preferentially switch to MATα using HML as the donor, which is driven by an adjacent cis-acting element called the recombination enhancer (RE). In this study we demonstrate that Sir2 and the condensin complex are recruited to the RE exclusively in MATa cells, specifically to the promoter of a small gene within the right half of the RE known as RDT1. We also provide evidence that the RDT1 promoter functions as a locus control region (LCR) that regulates both transcription and long-range chromatin interactions. Sir2 represses RDT1 transcription until it is removed from the promoter in response to a dsDNA break at the MAT locus induced by HO endonuclease during mating-type switching. Condensin is also recruited to the RDT1 promoter and is displaced upon HO induction, but does not significantly repress RDT1 transcription. Instead condensin appears to promote mating-type donor preference by maintaining proper chromosome III architecture, which is defined by the interaction of HML with the right arm of chromosome III, including MATa and HMR. Remarkably, eliminating Sir2 and condensin recruitment to the RDT1 promoter disrupts this structure and reveals an aberrant interaction between MATa and HMR, consistent with the partially defective donor preference for this mutant. Global condensin subunit depletion also impairs mating-type switching efficiency and donor preference, suggesting that modulation of chromosome architecture plays a significant role in controlling mating-type switching, thus providing a novel model for dissecting condensin function in vivo.
| Sir2 is a highly conserved NAD+-dependent protein deacetylase and defining member of the sirtuin protein family. It was identified 40 years ago in the budding yeast, Saccharomyces cerevisiae, as a factor that silences the cryptic mating-type loci, HML and HMR. These heterochromatic cassettes are utilized as templates for mating-type switching, whereby a programmed DNA double-strand break at the MATa or MATα locus is repaired by gene conversion to the opposite mating-type. This directional switching is called donor preference, which in MATa cells, is driven by a cis-acting DNA element called the recombination enhancer (RE). It was believed the only role for Sir2 in mating-type switching was silencing HML and HMR. However, we now show that Sir2 also regulates expression of a small gene (RDT1) in the RE that is activated during mating-type switching. The promoter of this gene is also bound by the condensin complex, and deleting this region of the RE drastically changes chromosome III structure and alters donor preference. The RE therefore appears to function as a complex locus control region (LCR) that links transcriptional control to chromatin architecture, thus providing a new model to investigate the underlying mechanistic principles of programmed chromosome architectural dynamics.
| Since the first descriptions of mating-type switching in budding yeast over 40 years ago, characterization of this process has led to numerous advances in understanding mechanisms of gene silencing (heterochromatin), cell-fate determination (mating-type), and homologous recombination (reviewed in [1]). For example, the NAD+-dependent histone deacetylase, Sir2, and other Silent Information Regulator (SIR) proteins, were genetically identified due to their roles in silencing the heterochromatic HML and HMR loci, which are maintained as silenced copies of the active MATα and MATa loci, respectively [2–4]. The SIR silencing complex (Sir2-Sir3-Sir4) is recruited to cis-acting E and I silencer elements flanking HML and HMR through physical interactions with silencer binding factors Rap1, ORC, and Abf1, as well as histones H3 and H4 (reviewed in [5]).
HML and HMR play a critical role in mating-type switching. Haploid cells of the same mating-type cannot mate to form diploids, the preferred cell type in the wild. Therefore, in order to facilitate mating and diploid formation, haploid mother cells switch their mating-type by expressing HO endonuclease, which introduces a programmed DNA double-strand break (DSB) at the MAT locus [6]. The break is then repaired by homologous recombination using either HML or HMR as a donor template for gene conversion [6, 7]. This change in mating-type enables immediate diploid formation between mother and daughter. HO is deleted from most standard lab strains in order to maintain them as haploids, so expression of HO from an inducible promoter such as PGAL1 is commonly used to switch mating-types during strain construction [8].
There is a “donor preference” directionality to mating-type switching such that ~90% of the time, the HO-induced DSB is repaired to the opposite mating-type [9]. For example, MATα cells preferentially switch to MATa using HMR as the donor. However, while both silent mating loci can be utilized as a donor template, usage of HML by MATa cells requires a 2.5 kb intergenic region located ~17 kb from HML called the recombination enhancer (RE) [10]. Donor preference activity within the RE has been further narrowed down to a 700 bp segment containing an Mcm1/α2 binding site (DPS1) and multiple Fkh1 binding sites [10]. The RE is active in MATa cells, requiring Mcm1 and Fkh1 activity at their respective binding sites [10–12]. The RE is inactivated in MATα cells due to expression of transcription factor α2 from MATα [13], which forms a repressive heterodimer with Mcm1 (Mcm1/α2) to repress MATa-specific genes [1]. Current models for donor preference posit that Fkh1 at the RE helps position HML in close proximity with MATa by interacting with threonine-phosphorylated H2A (γ-H2AX) and Mph1 DNA helicase at the HO-induced DNA DSB [14, 15].
Sir2-dependent silencing of HML and HMR has two known functions related to mating-type switching. First, HML and HMR must be silenced in haploids to prevent formation of the a1/α2 heterodimer, which would otherwise inactivate haploid-specific genes such as HO [16]. Second, heterochromatin structure at HML and HMR blocks cleavage by HO, thus restricting its activity to the fully accessible MAT locus [17, 18]. Here we describe new roles for Sir2 and the condensin complex within the RE during mating-type switching. ChIP-seq analysis revealed strong overlapping binding sites for Sir2 and condensin at the promoter of a small gene within the RE known as RDT1. Here, Sir2 was found to repress the MATa-specific transcription of RDT1, which is also translated into a small 28 amino acid peptide. RDT1 expression is also dramatically upregulated during mating-type switching when Sir2 is lost from the RDT1 promoter and instead associates with the HO-induced DNA DSB at MATa. Furthermore, eliminating Sir2/condensin recruitment to the RDT1 promoter disrupts chromosome III architecture such that donor preference is partially impaired. The RDT1 promoter region therefore functions like a classic locus control region (LCR) in MATa yeast cells, regulating localized transcription as well as long-range chromosome interactions.
We previously characterized global sirtuin distribution using ChIP-Seq to identify novel loci regulated by Sir2 and its homologs [19]. Significant overlap was observed between binding sites for Sir2, Hst1, or Sum1 with previously described condensin binding sites [19, 20], suggesting a possible functional connection. ChIP-Seq was therefore performed on WT and sir2Δ strains in which the condensin subunit Smc4 was C-terminally tagged (13xMyc) (Fig 1A). To avoid “hyper-ChIPable” loci that can appear in yeast ChIP-seq experiments, we also ran nuclear localized GFP controls [21]. Genes closest to Sir2-dependent condensin peaks after subtraction of GFP are listed in S1 Table, and are distributed throughout the genome. One of the strongest peaks overlapped with a Sir2-myc binding site on chromosome III between KAR4 and SPB1 that was not enriched for GFP (Fig 1A). The specificity of Sir2 enrichment at this peak, as opposed to the adjacent SPB1 gene, was independently confirmed by quantitative ChIP using an α-Sir2 antibody (Fig 1B), with enrichment comparable to levels observed at the HML-I silencer (Fig 1A and 1B). Sir2-dependent condensin binding was also confirmed for Myc-tagged Smc4 and Brn1 subunits (Fig 1C). The ~2.5 kb intergenic region between KAR4 and SPB1 was previously defined as a cis-acting recombination enhancer (RE) that specifies donor preference of mating-type switching in MATa cells [10, 13]. Quantitative ChIP assays revealed that Sir2 and Brn1-myc enrichment at the RE was also MATa-specific (Fig 1D and 1E), which was notable because the ChIP-seq datasets in Fig 1A happened to be generated from MATa strains. We next considered whether the condensin binding defect in the MATa sir2Δ mutant was due to HMLALPHA2 expression caused by defective HML silencing. To test this idea, we re-examined Brn1-myc ChIP signal at the RE in strains lacking HML, and found that deleting SIR2 no longer affected condensin recruitment (Fig 1F). Similarly, a MATa condensin mutant (ycs4-1) known to have an HML silencing defect [22] reduced Sir2 recruitment to the RE, but had no effect when HML was also deleted (Fig 1G). Sir2 and condensin are therefore independently recruited to the RE only in MATa cells.
Donor preference activity ascribed to the RE was previously narrowed down to a KAR4 (YCL055W)-proximal 700 bp domain defined by an Mcm1/α2 binding site (Fig 2A, DPS1) [10, 11, 13]. The Sir2 and condensin ChIP-seq peaks we identified were located outside this region, between a second Mcm1/α2 binding site (DPS2) and a small gene of unknown function called RDT1 [23] (Figs 1A and 2A). We noticed the location of RDT1 coincided with the smallest of several putative non-coding RNAs (ncRNA) previously reported as being transcribed from the RE, but not annotated in SGD (Fig 2A, [13]). Quantitative RT-PCR and analysis of publicly available RNA-seq data from BY4741 (MATa) and BY4742 (MATα) revealed that RDT1 expression was indeed MATa specific (Fig 2B and S1A Fig).
We next asked whether Sir2 and/or condensin regulate histone acetylation and RDT1 expression when recruited to the RE. Sir2 normally represses transcription at HML, HMR, and telomeres as a catalytic subunit of the SIR complex where it preferentially deacetylates H4K16 (reviewed in [5]). Accordingly, deleting SIR2, SIR3, or SIR4 from MATa cells increased H4K16 acetylation at the RDT1 promoter (Fig 2C), consistent with the observed enrichment of Sir3-myc and Sir4-myc at this site (S1B Fig). Furthermore, re-introducing active SIR2 into the sir2Δ mutant restored H4K16 to the hypoacetylated state, whereas catalytically inactive sir2-H364Y did not (Fig 2D). While H4K16ac is a preferred Sir2 substrate for silencing at HML and HMR, the recruited SIR complex also maintains lysine deacetylation of the other N-terminal histone tails [24]. Since RDT1 transcription is MATa specific, and Mcm1/α2 represses MATa-specific genes by recruiting the Ssn6/Tup1 corepressor complex and Class I/II HDACs such as the H3/H2B-specific histone deacetylase HDA1 [25, 26], we also tested the effect of deleting SIR2 on H3K9/14 acetylation at the RDT1 promoter, predicting it may remain hypoacetylated due to the SIR complex being replaced by Ssn6/Tup1/HDA1. Indeed, H3K9/14 acetylation was reduced in the sir2Δ mutant relative to WT, but was significantly elevated in the hmlΔ sir2Δ double mutant (Fig 2E). RDT1 expression was similarly reduced in the sir2Δ mutant and strongly upregulated when HML and SIR2 were both deleted (Fig 2F). An even stronger expression effect was observed in an hmlΔ sir2Δ hst1Δ triple mutant that eliminates any possibility of redundancy between the Sir2 and Hst1 paralogs (S1C Fig). On the other hand, RDT1 was not upregulated in an hmlΔ ycs4-1 condensin mutant (Fig 2G). Taken together, these results provide strong evidence that the SIR complex represses RDT1 in MATa cells by establishing a generally hypoacetylated chromatin environment at the promoter, while condensin has a functional role independent of transcriptional regulation.
We next attempted to prevent Sir2 and condensin recruitment to the RDT1 promoter by precisely deleting a 100bp DNA sequence underlying the shared enrichment region (coordinates 30701–30800), while not disturbing the adjacent Mcm1/α2 site (Fig 3A). Sir2 and Brn1-myc binding to the RE as measured by ChIP was greatly diminished in this mutant (Fig 3B and 3C), despite unaltered Sir2, Brn1-myc, or Smc4-myc expression levels (S2A–S2C Fig). Furthermore, RDT1 RNA expression level was significantly increased by the 100bp deletion exclusively in MATa cells (Fig 3D), consistent with the loss of Sir2-mediated repression.
Because Sir2 and condensin were not present at the RDT1 promoter in MATα cells, we reasoned that their binding should require a MATa specific transcription factor. This made the 2nd Mcm1/α2 binding site (DPS2) upstream of the Sir2/condensin ChIP-seq peaks an ideal candidate, because it had not previously been ascribed a function other than redundancy with DPS1 for donor preference [13]. Deleting MCM1 is lethal, so alternatively, we precisely deleted the 2nd Mcm1/α2 binding site (ChrIII coordinates 30595 to 30626, S3A Fig) and then retested for Sir2 and Brn1-myc enrichment. As shown in S3B and S3C Fig, respectively, Sir2 and Brn1-myc enrichment at the Mcm1/α2 binding site (DPS2) and the RDT1 promoter (defined as the Sir2/condensin peaks) was significantly reduced in the binding site mutant. These results suggest that Mcm1 nucleates a complex that recruits the SIR and condensin complexes to the RDT1 promoter in MATa cells, and also provides a possible mechanism of blocking the recruitment in MATα cells due to the interaction of Mcm1 with α2.
Ribosome Detected Transcript-1 (RDT1) was originally annotated as a newly evolved gene whose transcript was associated with ribosomes and predicted to have a small open reading frame of 28 amino acids [23]. Our work suggested that RDT1 and the putative non-coding R2 transcript were the same (Fig 2A). To determine if RDT1/R2 codes for a small protein, the ORF was C-terminally fused with a 13x-Myc epitope in MATa and MATα cells. As shown in Fig 3E, a fusion protein was exclusively detected in exponentially growing MATa WT cells and also highly expressed in the 100bpΔ strain, correlating with the increased RNA level observed for that mutant in Fig 3D.
Since additional MATa-specific RNAs are derived from the minimal 700bp RE domain (Fig 2A; R1L and R1S) [13, 27], we next tested whether Sir2/condensin recruitment to the RDT1 promoter had any effect on expression of these upstream ncRNAs from a distance. Quantitative RT-PCR for the R1L/S transcripts indicated their expression level in the WT strain was comparable to RDT1, and was also reduced in a sir2Δ mutant because of the pseudodiploid phenotype (Fig 3F). However, while RDT1 was strongly upregulated by the 100bp deletion of the Sir2/condensin binding site, R1L/S expression was unaffected (Fig 3F). As a control, we also measured expression of the SPB1 gene located immediately downstream of RDT1 (see Fig 2A schematic). SPB1 expression is not mating-type specific, it encodes a rRNA methyltransferase required for maturation of the large 60S ribosomal subunit [28], and interestingly, also functions in silencing at HML and HMR [29]. It was therefore intriguing that SBP1 expression was increased 2- to 3-fold in the sir2Δ and 100bpΔ mutants (Fig 3G), suggesting that Sir2 at the RDT1 promoter has a modest downstream repressive effect on SPB1, but not on the upstream R1L/S ncRNAs. Notably, steady state RDT1 and R1L/S RNA levels were relatively low compared to SPB1 and the ACT1 loading control, even in the 100bpΔ mutant (Fig 3G).
We next asked if Sir2 played any role in regulating RDT1 during mating-type switching. Sir2 was previously shown to associate with a HO-induced DSB at the MAT locus during mating-type switching, presumably to effect repair through histone deacetylation [30]. SIR complex association with DSBs occurs at the expense of telomere binding [31], so we hypothesized that the HO-induced DSB at MAT could also trigger loss of Sir2 from the RDT1 promoter, thus facilitating increased RDT1 transcription. To test this idea, HO was induced at time 0 with galactose and then turned off 2 hours later by glucose addition to allow for repair/switching to occur (Fig 4A and 4B). By the 3 hr time point (1 hr after glucose addition), ChIP analysis indicated Sir2 was maximally enriched at the MAT locus (Fig 4C), corresponding to the time of peak mating-type switching ([30] and Fig 4B). Interestingly, Sir2 was significantly depleted from the RDT1 promoter within 1 hr after HO induction, and by 3 hr there was actually stronger enrichment of Sir2 at MAT than RDT1 (Fig 4C). Critically, this shift in Sir2 distribution correlated with maximal induction of RDT1 mRNA and the Myc-tagged Rdt1protein (Fig 4D and 4E, 3 hr). Once switching was completed by 4 hr (2 hr after glucose addition), RDT1 transcription was permanently inactivated and Sir2 binding never returned because most cells were now MATα. The Myc-tagged Rdt1 protein, however, remained elevated for the rest of the time course (Fig 4E), suggesting that it is relatively stable, at least when epitope tagged. A parallel ChIP time course experiment was performed with condensin (Brn1-myc), indicating significant depletion from the RDT1 promoter within 1 hr (Fig 4F), similar to the timing of Sir2 loss. Unlike Sir2, Brn1-myc enrichment at the HO-cleaved MAT site did not increase, suggesting that condensin normally associates with MATa in non-switching cells and is then displaced in response to the HO-induced DSB, perhaps to facilitate structural reorganization associated with switching.
Since RDT1 was highly expressed during mating-type switching, we next asked whether the small protein encoded by this gene had a direct function during the switching process when using the galactose-inducible system employed in this study. The 28 amino acid ORF was precisely deleted using the delitto perfetto method [32], and the efficiency of switching from MATa to MATα was then tested across a time course by PCR (S4A and S4B Fig). No significant differences were observed between the WT and rdt1Δ strains. Next, donor preference was tested using a strain previously developed by the Haber lab [15], whereby HMRa was replaced with HMRα containing a unique BamHI site (HMRα-B) (S4C Fig). Following completion of Gal-HO-induced switching from MATa to MATα, the proportion of donor utilization was determined by BamHI digest of a MATα PCR product. As shown in S4D and S4E Fig, deleting RDT1 also had no effect on donor preference, indicated by low utilization (~9%) of HMRα-B. Therefore, although RDT1 gene expression strongly correlates with switching, a specific function for its gene product remains elusive. Therefore, we shifted our attention to a possible function for the RDT1 promoter.
The coupling of Sir2 and condensin distribution with RDT1 transcriptional regulation during mating-type switching was reminiscent of classic locus control regions (LCR) that modulate long-range chromatin interactions [33, 34]. We therefore hypothesized that the RDT1 promoter region functions as an LCR to modulate long-range chromatin interactions of chromosome III. To test this hypothesis, we performed Hi-C analysis with WT, sir2Δ and the 100bpΔ strains. Genomic contact differences between the mutants and WT were quantified using the HOMER Hi-C software suite [35], and the frequency of statistically significant differences for each chromosome calculated (Fig 5A). Chromosome III had the most significant differences in both mutants, so we focused on this chromosome and used HOMER to plot the observed/expected interaction frequency in 10kb bins for each strain as a heat map (Fig 5B). In a WT strain (ML1) there was strong interaction between the left and right ends of chromosome III, mostly centered around the HML (bin 2) and HMR (bin 29) loci. Interestingly, HML (bin 2) also appeared to sample the entire right arm of chromosome III, with the interaction frequency increasing as a gradient from CEN3 to a maximal observed interaction at HMR, thus also encompassing the MATa locus at bin 20. This distinct interaction pattern was completely disrupted in the sir2Δ mutant, whereas some telomere-subtelomere contacts were retained in the 100bpΔ mutant (Fig 5B), suggesting there was still limited interaction between the left and right ends of the chromosome. We confirmed the changes in HML-HMR interaction for these strains using a quantitative 3C-PCR assay to rule out sequencing artifacts (Fig 5C), and to confirm an earlier sir2Δ 3C result from the Dekker lab [36]. Importantly, despite the loss of HML-HMR interaction in the 100bpΔ mutant, heterochromatin at these domains was unaffected based on normal quantitative mating assays (S5A Fig), and unaltered Sir2 association with HML (S5B Fig).
We next analyzed the Hi-C data using an iterative correction method that reduces background to reveal interacting loci that potentially drive the overall chromosomal architecture, rather than passenger locus effects [37]. HML (bin 2) and HMR (bin 29) again formed the dominant interaction pair off the diagonal in WT, which was lost in the sir2Δ or 100bpΔ mutants (Fig 5D, red arrows). Importantly, a prominent new interaction between HMR (bin 29) and MATa (bin 20) appeared in both mutants (Fig 5D, black arrows), as would be predicted if normal donor preference of MATa cells was altered. We conclude that the RDT1 promoter does function like an LCR in MATa yeast cells, regulating localized transcription and establishing a long-range chromatin interaction between HML and HMR that appears to prevent HMR from strongly associating with MATa (Fig 5E).
Sir2/condensin binding was observed in the right half of the RE (Fig 1A), but this region was previously reported as dispensable for donor preference activity [10]. Considering that HMR was aberrantly associated with the MATa locus in sir2Δ and 100bpΔ mutants (Fig 5B–5E), we proceeded to test whether these mutants had any alterations in donor preference. As was done for the rdt1Δ mutant in S4C and S4D Fig, the 100bpΔ mutation was introduced into a reporter strain with HMRα-B on the right arm of chromosome III (Fig 6A; [15]). After inducing switching to MATα with galactose, the proportion of HMLα or HMRα-B used for the switching was determined by BamHI digestion of a MATα-specific PCR product (Fig 6B; [15]). As expected for normal donor preference, HMRα-B on the right arm was only utilized ~9% of the time in the WT strain (Fig 6C and 6D). Strikingly, donor preference was lost in the sir2Δ mutant, similar to a control strain with the RE deleted (Fig 6C and 6D), and consistent with the clear interaction between HMR and MATa bins observed for the sir2Δ mutant in Fig 5D and 5E. This interaction was less prominent in the 100bpΔ mutant (Fig 5D), and the corresponding donor preference defect was also less severe (~25% HMRα-B), though still significantly different from WT (Fig 6C and 6D).
Since the donor preference assay is an endpoint experiment, we next tested whether the sir2Δ or 100bpΔ mutations temporally impacted the efficiency of switching from MATa to MATα in the same ML1 background strains used for Hi-C analysis. As shown in Fig 6E and 6F, switching efficiency was dramatically impaired in the sir2Δ strain, but unaffected in the 100bpΔ mutant. These results suggest a model whereby condensin and Sir2 recruitment to the RDT1 promoter supports donor preference by organizing chromosome III into a structure that limits HMR association with the MATa locus, but is not required for the mechanics of mating-type switching. We suspect at least part of the strong sir2Δ effect on chromosome III organization and donor preference is caused by HMLALPHA2 derepression, which inactivates the RE due to formation of the Mcm1/α2 heterodimer [13]. We also considered a possibility that the sir2Δ heterochromatin defect at HML and HMR could make them highly accessible to HO cleavage [38], which would prevent their usage as donor templates. As a measure of HO cleavage at MATa, HML, or HMR, we assayed for reduced PCR amplification across the recognition site following Gal-HO induction (S6 Fig). While MATa was equally cut by HO in WT and sir2Δ strains (S6A and S6B Fig), HML was only cut in the sir2Δ strain (S6A and S6C Fig), consistent with the idea of reduced HML availability for switching. Unexpectedly, HMR remained available as a template in the sir2Δ strain (S6A and S6D Fig). We confirmed the difference in HO accessibility between HML and HMR using real-time qPCR (S6E and S6F Fig), which could contribute to the extreme sir2Δ donor preference defect (Fig 6C and 6D). In the 100bpΔ mutant, which locally eliminates condensin recruitment at RDT1, the continued maintenance of heterochromatin at HML/HMR and telomeres (S5 Fig), as well as residual telomere clustering (Fig 5D), may partially buffer the resulting donor preference defect by still limiting contact between the subtelomeric HMR locus and MATa.
Condensin recruitment to the RDT1 promoter does not appear critical for the mechanics of mating-type switching, as indicated by normal timing of switching in the 100bpΔ mutant (Fig 6E and 6F). However, condensin could still potentially impact the switching process independent of the RE. In order to test this hypothesis, we C-terminally tagged the Brn1 condensin subunit with an auxin-inducible degron (AID) fused with a V5 epitope, which allows for rapid depletion of tagged proteins upon the addition of auxin [39]. Indeed, Brn1-AID was effectively degraded within 15 min of adding auxin to cells, as measured by western blotting (S7A Fig), or ChIP at the RDT1 promoter (S7B Fig). Importantly, even after 1 hr of auxin treatment, there were no significant changes in RDT1 or HMLALPHA2 gene expression as measured by qRT-PCR (S7C and S7D Fig), thus indicating that silencing of HML was unaffected, unlike the ycs4-1 condensin mutant used in Fig 1G [22]. The efficiency of ML1 switching from MATa to MATα was then tested across a time course with or without auxin treatment (Fig 7A). As shown in Fig 7B and 7C, auxin significantly slowed the efficiency of switching to MATα, indicating that the Brn1 subunit of condensin is important for normal mating-type switching.
Since the 100bpΔ mutant caused a modest donor preference defect that we partially attributed to a loss of condensin (Fig 6C and 6D), it was important to also test for a donor preference defect when condensin was depleted. Indeed, Brn1-AID depletion produced a significant defect in donor preference using the HMRα-B reporter strain (Fig 7D) that was similar in magnitude to that observed for the 100bpΔ strain (Fig 6D). Taken together, these results support a working model whereby condensin recruitment to the RDT1 promoter in MATa cells organizes chromosome III into a conformation that limits the association of HMR with MATa, thus partially contributing to donor preference regulation. We hypothesize that upon HO cleavage of MATa, the increased expression of RDT1 caused by loss of Sir2, displaces condensin from the promoter, which then allows the left half of the RE to physically direct HML to MATa for use as a donor (Fig 8).
SIR2 was identified ~40 years ago as a recessive mutation unlinked from HML and HMR that caused their derepression [3, 4], and has been extensively studied ever since as encoding a heterochromatin factor that functions not only at the HM loci, but also telomeres and the rDNA locus (reviewed in [5]). In this study we describe a previously unidentified Sir2 binding site that overlaps with a major non-rDNA condensin binding site within the RE on chromosome III in MATa cells. Here, Sir2 regulates a small gene of unknown function called RDT1, which is transcriptionally activated during mating-type switching due to loss of repressive Sir2 from the RDT1 promoter that correlates with binding to the HO-induced DSB at MATa. The RDT1 RNA transcript is also translated into a small protein, but we have not yet been able to assign a function to the gene or protein because deleting the 28 amino acid ORF does not measurably alter mating-type switching when using GAL-HO based assays. It remains possible that deleting RDT1 would have a significant effect on switching in the context of native HO expression, which is expressed only in mother cells during late G1, whereas the GAL1-HO is overexpressed in all cells throughout the cell cycle. Furthermore, RDT1 and the R1L/S ncRNAs are themselves cell cycle regulated, similarly showing peak expression around late G1 [27]. Therefore, even with the Gal-HO induced system, the strong RDT1 expression observed during switching (Fig 4D) suggests significant enrichment of G1 cells in the population. It is also possible that RDT1 functions as a non-coding RNA that happens to be translated into a small non-functional peptide. Alternatively, transcription of RDT1 could directly function in chromosome III conformation by altering local chromatin accessibility at the promoter. Such a model was proposed for regulation of donor preference by transcription of the R1S/R1L non-coding RNAs [13, 27]. Dissecting the function(s) of RDT1 through the cell cycle therefore remains an area of active investigation for the lab, and perhaps the key to fully understanding how its promoter functions as an LCR.
While we do not yet know the molecular function of RDT1 in mating-type regulation or other cellular processes, the promoter region of this gene clearly controls the structure of chromosome III. Three-dimensional chromatin structure has long been proposed to influence donor preference [40, 41]. However, deleting the minimal 700bp (left half) of the RE alters donor preference without a large change in chromosome III conformation. Furthermore, deleting the right half of the RE, which includes RDT1, changes chromosome III conformation without a dramatic change in donor preference [10, 13, 42]. Based on these findings it was proposed that the RE is a bipartite regulatory element [42], with the left half primarily responsible for donor preference activity and the right half for chromosome III structure. Our results support this view and narrow down the structural regulatory domain of the RE to a small (100bp) region of the RDT1 promoter bound by the SIR and condensin complexes. Importantly, deleting this small region not only altered chromosome III structure, but also had a significant effect on donor preference, though not as strong as the sir2Δ mutation.
The coordination of RDT1 expression with loss of Sir2/condensin binding at its promoter during mating-type switching, together with the loss of HML-HMR interaction in the 100bpΔ mutant, makes this site intriguingly similar to classic locus control regions (LCRs) in metazoans, which are cis-acting domains that contain a mixture of enhancers, insulators, chromatin opening elements, and tissue-specificity elements [33]. The minimal RE was previously described as an LCR in the context of donor preference [10], and transcription of the R1S/R1L long non-coding RNAs via activation by the 1st Mcm1/α2 binding site (DPS1) appears to be important for this activity in MATa cells [27]. We find that Sir2 indirectly supports donor preference from the left half of the RE in MATa cells by silencing HMLALPHA2 expression, which prevents transcriptional repression by an Mcm1/α2 heterodimer, and by protecting the HML template from HO cleavage. Similarly, the loss of Sir2 also represses RDT1 expression and condensin recruitment in the right half of the RE due to HMLALPHA2 expression. It remains possible that SIR-dependent heterochromatin formation also directly contributes to the HML-HMR interaction through clustering. More clearly, however, Sir2 directly represses RDT1 through localized histone deacetylation. How the loss of RDT1 regulation and condensin recruitment changes chromosome III structure in the sir2Δ mutant remains unknown, but we propose that the HMR-MATa interaction is a default state, while the HML-HMR association has to be actively maintained by condensin and likely additional factors co-localized to this element, as well as SIR-dependent heterochromatin.
Interestingly, there also appears to be a functional relationship between the RE and silencing at the HML locus, such that deleting the left half of the RE specifically stabilizes HML silencing in MATa cells [43]. The mechanism involved remains unknown, but we hypothesize that eliminating this part of the RE could potentially allow the SIR and condensin complexes bound at the RDT1 promoter to encroach and somehow enhance the heterochromatic structure at HML. Under this scenario, the left half of the RE could be insulating HML from the chromosomal organizing activity that occurs at the RDT1 promoter.
The RDT1 promoter is a major condensin binding site identified by ChIP-seq (Fig 1), and given the strong Hi-C interaction between HML and HMR, we initially hypothesized that condensin at the RDT1 promoter would crosslink with another condensin site bound on the right arm of chromosome III. However, this turned out to be unlikely because ChIP-seq of Smc4-myc did not reveal any strong peaks near HMR. The S. cerevisiae condensin complex was recently shown to catalyze ATP-dependent unidirectional loop extrusion using an in vitro single molecule assay [44]. The mechanism involves direct binding of condensin to DNA, followed by one end of the bound DNA being pulled inward as an extruded loop. Applying this model to the strong binding site at the RDT1 promoter, this region could act as an anchor bound by condensin, with DNA to the right being rapidly extruded as a loop until pausing at CEN3. Extrusion would then continue at a slower rate toward HMR, allowing HML time to sample the entire right arm of chromosome III, until clustering with HMR (Fig 8). HOMER analysis of the Hi-C data in Fig 5B provides evidence for such a model because there is an ascending gradient of HML interaction frequency with sequences extending from the centromere region (bin 12) toward HMR, suggesting that HML “samples” the right arm of chromosome III. Once brought in contact, HML and HMR would then remain associated due to their heterochromatic states and shared retention at the nuclear envelope [45]. Formation of this loop appears to limit HMR association with MATa, but since the 100bpΔ mutation has no effect on the timing of switching, we do not believe that condensin at the RDT1 promoter functions directly in the homologous recombination process. Rather, general Brn1 subunit depletion could slow the switching process by affecting chromosome III flexibility or conformational dynamics.
Condensin, and Sir2 each strongly associated with the RDT1 promoter exclusively in MATa cells, though it is not clear if they bind at the same time, or are differentially bound throughout the cell cycle. Since DPS2 was required for Sir2 and condensin recruitment, and derepression of HMLALPHA2 from HML also eliminated binding, we hypothesized and then demonstrated (S3 Fig) that Mcm1 was a key DNA binding factor involved. Mcm1 is a prototypical MADS box combinatorial transcription factor that derives its regulatory specificity through interactions with other factors, such as Ste12 in the case of MATa haploid-specific gene activation, or α2 when repressing the same target genes in MATα cells [46]. This raises the question of whether Mcm1 directly recruits the SIR and condensin complexes, or perhaps additional factors that work with Mcm1 are involved. The latter is likely true because condensin and Sir2 are not recruited to the leftmost Mcm1/α2 binding site in the left half of the RE, as indicated by the ChIP-seq data in Fig 1A. At the RDT1 promoter, specificity for Sir2/condensin recruitment could originate from sequences underlying the condensin/Sir2 peaks. There are no traditional silencer-like sequences for SIR recruitment within the deleted 100bp (coordinates 30702 to 30801), and yeast condensin does not appear to have a consensus DNA binding sequence [47]. Closer inspection of the RDT1 promoter indicates an A/T rich region with consensus binding sites for the transcription factors Fkh1/2 and Ash1, each of which regulates mating-type switching [11, 48, 49]. Fkh1 and Fkh2 also physically associate with Sir2 at the mitotic cyclin CLB2 promoter during stress [50]. Ash1 is intriguing because it represses HO transcription in daughter cells [49, 51], raising the possibility of RDT1 repression in daughter cells. Mcm1 activity in MATa cells could also indirectly establish a chromatin environment that is competent for Sir2/condensin recruitment, rather than direct recruitment through protein-protein interactions. In MATa cells, Mcm1 appears to prevent the strong nucleosome positioning across the RE that occurs in MATα cells [27], and indicative of an actively remodeled chromatin environment. Perhaps condensin is attracted to such regions, which is consistent with the association of condensin with promoters of active genes in mitotic cells, where enrichment was greatest at unwound regions of DNA [52]. Furthermore, nucleosome eviction by transcriptional coactivators was found to assist condensin loading in yeast [53], though the mechanism of loading remains poorly understood. Recruitment of condensin to the RDT1 promoter LCR therefore provides an outstanding opportunity for dissecting mechanisms of condensin loading and function.
Yeast strains were grown at 30°C in YPD or synthetic complete (SC) medium where indicated. The SIR2, or HST1 open reading frames (ORFs) were deleted with kanMX4 using one-step PCR-mediated gene replacement. HML was deleted and replaced with LEU2. Precise deletions of the 100bp condensin/Sir2 binding site within the RDT1 promoter (chrIII coordinates 30701–30800), DPS2 (chrIII coordinates 30557–30626), or the RDT1 ORF (chrIII coordinates 30910–30996) were generated using the delitto perfetto method [32]. Endogenous SIR2, BRN1, or SMC4 genes were C-terminally tagged with the 13xMyc epitope (13-EQKLISEEDL). Deletion and tagged gene combinations were generated through genetic crosses and tetrad dissection, including Brn1 tagged with a V5-AID tag (template plasmids kindly provided by Vincent Guacci). All genetic manipulations were confirmed by PCR, and expression of tagged proteins confirmed by western blotting. The pGAL-HO-URA3 expression plasmid was kindly provided by Jessica Tyler [30]. Strain genotypes are provided in S2 Table and oligonucleotides listed in S3 Table.
Sir2 ChIP-seq was previously described [19]. For other ChIP-seq datasets, log-phase YPD cultures were cross-linked with 1% formaldehyde for 20 min, pelleted, washed with Tris-buffered saline (TBS), and then lysed in 600 μl FA140 lysis buffer with glass beads using a mini-beadbeater (BioSpec Products). Lysates were removed from the beads and sonicated for 60 cycles (30s “on” and 30s “off” per cycle) in a Diagenode Bioruptor. Sonicated lysates were pelleted for 5 min at 14000 rpm in a microcentrifuge and the entire supernatant was transferred to a new microfuge tube and incubated overnight at 4°C with 5 μg of anti-Myc antibody (9E10) and 20 μl of protein G magnetic beads (Millipore). Following IP, the beads were washed once with FA140 buffer, twice with FA500 buffer, and twice with LiCl wash buffer. DNA was eluted from the beads in 1% SDS/TE buffer and cross-links were reversed overnight at 65°C. The chromatin was then purified using a Qiagen PCR purification kit. Libraries were constructed using the Illumina Trueseq ChIP Sample Prep kit and TrueSeq standard protocol with 10ng of initial ChIP or Input DNA. Libraries that passed QC on a Bioanalyzer High Sensitivity DNA Chip (Agilent Technologies) were sequenced on an Illumina Miseq (UVA Genomic Analysis and Technology Core).
Biological duplicate fastq files were concatenated together and reads mapped to the sacCer3 genome using Bowtie with the following options:—best,—stratum,—nomaqround, and—m10 [54]. The resulting bam files were then converted into bigwig files using BEDTools [55]. As part of the pipeline, chromosome names were changed from the sacCer3 NCBI values to values readable by genomics viewers e.g. "ref|NC_001133|" to "chrI". The raw and processed datasets used in this study have been deposited in NCBI’s GEO and are accessible through the GEO series accession number GSE92717. Downstream GO analysis was performed as follows. MACS2 was used to call peaks with the following options:—broad,—keep-dup, -tz 150, and -m 3, 1000 [56]. GFP peaks in the WT or sir2Δ backgrounds were subtracted from the WT SMC4-13xMyc and sir2Δ SMC4-13xMyc peaks, respectively, using BEDTools “intersect” with the–v option. The resulting normalized peaks were annotated using BEDTools “closest” with the -t all option specified, and in combination with a yeast gene list produced from USCS genome tables. The annotated peaks were then analyzed for GO terms using YeastMine (yeastmine.yeastgenome.org).
Log-phase cultures were cross-linked with 3% formaldehyde for 20 min and quenched with a 2x volume of 2.5M Glycine. Cell pellets were washed with dH2O and stored at -80°C. Thawed cells were resuspended in 5 ml of 1X NEB2 restriction enzyme buffer (New England Biolabs) and poured into a pre-chilled mortar containing liquid N2. Nitrogen grinding was performed twice as previously described [57], and the lysates were then diluted to an OD600 of 12 in 1x NEB2 buffer. 500 μl of cell lysate was used for each Hi-C library as follows. Lysates were solubilized by the addition of 50 μl 1% SDS and incubation at 65°C for 10 min. 55 μl of 10% TritonX-100 was added to quench the SDS, followed by 10 μl of 10X NEB2 buffer and 15 μl of HindIII (New England Biolabs, 20 U/μl) to digest at 37°C for 2 hr. An additional 10 μl of HindIII was added for digestion overnight. The remainder of the protocol was based on previously published work with minor exceptions [58]. In short, digested chromatin ends were filled-in with Klenow fragment (New England Biolabs) and biotinylated dCTP at 37°C for 1 hr, then heat inactivated at 70°C for 10 min. Ligation reactions with T4 DNA ligase were performed at 16°C for a minimum of 6 hr using the entire Hi-C sample diluted into a total volume of 4 ml. Proteinase K was added and cross-links were reversed overnight at 70°C. The ligated chromatin was phenol:chloroform extracted, ethanol precipitated, then resuspended in 500μl dH2O and treated with RNAse A for 45 min. Following treatment with T4 DNA polymerase to remove biotinylated DNA ends that were unligated, the samples were concentrated with a Clean and Concentrator spin column (Zymogen, D4013) and sheared to approximately 300bp with a Diagenode Bioruptor. Biotinylated fragments were captured with 30 μl pre-washed Streptavidin Dynabeads (Invitrogen), then used for library preparation. Hi-C sequencing libraries were prepared with reagents from an Illumina Nextera Mate Pair Kit (FC-132-1001) using the standard Illumina protocol of End Repair, A-tailing, Adapter Ligation, and 12 cycles of PCR. PCR products were size selected and purified with AMPure XP beads before sequencing with an Illumina Miseq (UVA Genomic Analysis and Technology Core) or Hiseq (HudsonAlpha Institute for Biotechnology, Birmingham, AL). Raw and mapped reads are deposited at GEO (GSE92717).
Iteratively corrected heatmaps were produced using python scripts from the Mirny lab hiclib library, http://mirnylab.bitbucket.org/hiclib/index.html. Briefly, reads were mapped using the iterative mapping program, which uses Bowtie2 to map reads and iteratively trims unmapped reads to increase the total number of mapped reads. Mapped reads were then parsed into an hdf5 python data dictionary for storage and further analysis. Mapped reads of the same strains were concatenated using the hiclib library’s “Merge" function. Both individual and concatenated mapped reads have been deposited in GEO. Mapped reads were then run through the fragment filtering program using the default parameters as follows: filterRsiteStart(offset = 5), filterDuplicates, filterLarge, filterExtreme (cutH = 0.005, cutL = 0). Raw heat maps were further filtered to remove diagonal reads and iteratively corrected using the 03 heat map processing program. Finally, the iteratively corrected heatmaps were normalized for read count differences by dividing the sum of each row by the sum of the max row for a given plot, which drives all values towards 1 to make individual heatmaps comparable.
Observed/Expected heatmaps were created using HOMER Hi-C analysis software on the BAM file outputs from the iterativemapping program of the hiclib library python package [35]. Tag directories were created using all experimental replicates of a given biological sample and the tbp -1 and illuminaPE options. Homer was also used to score differential chromosome interactions between the WT and mutant Hi-C heatmaps. The resulting list of differential interactions was uploaded into R where the given p-value was adjusted to a qvalue with p.adjust. An FDR cutoff of 0.05 was used to create a histogram of significantly different chromosome interactions in the mutants compared to WT. The histogram was further normalized by dividing the total number of significant differential interactions for a chromosome by total number of interactions called in the WT sample for that chromosome to account for size differences in the chromosomes. Thus, frequency represents the number of interactions that changed out of all possible interactions that could have changed.
RNA-Seq data was acquired from GEO accessions GSE73274 [59] and GSE58319 [60] for the BY4742 (MATα) and BY4741 (MATa) backgrounds, respectively. Reads were then mapped to the sacCer3 genome using Bowtie2 with no further processing of the resulting BAM files visualized in this paper.
Chromosome Conformation Capture (3C) was performed in a similar manner to Hi-C with a few exceptions due to assay-specific quantification via quantitative real-time PCR rather than sequencing. Most notably, digested DNA ends were not filled in with dCTP-biotin before ligation and an un-crosslinked control library was created for each 3C library. Furthermore, all PCR reactions were normalized for starting DNA concentration using a PDC1 intergenic region that is not recognized by HindIII, in addition to PCR of the un-crosslinked control for all tested looping interaction primer pairs.
Total RNA (1 μg) was used for cDNA synthesis with oligo(dT) and Superscript II reverse transcriptase as previously described [61]. Expression levels are indicated in figures relative to the level of ACT1 mRNA, with this ratio then normalized to 1.0 for a specific strain or condition indicated for each experiment.
Proteins were blotted using standard TCA extraction followed by SDS-PAGE as previously described [19]. Myc-tagged proteins were incubated with an anti-Myc primary antibody 9E10 (Millipore) at a 1:2000 dilution while tubulin was incubated with anti-Tubulin antibody B-5-1-2 (Sigma-Aldrich) at a 1:1500 dilution. The V5-AID tagged Brn1 was detected with anti-V5 antibody (Invitrogen, R96025) at a 1:4000 dilution. Primary antibodies were detected with an anti-mouse secondary antibody conjugated to HRP (Promega) at 1:5000 dilution in 5% fat-free milk. Bands were then visualized with HyGlo (Denville Scientific) capture on autoradiography film (Denville Scientific).
For tracking the efficiency of switching, strains were transformed with pGAL-HO-URA3, pre-cultured in SC-ura + raffinose (2%) medium overnight, re-inoculated into the same medium (OD600 = 0.05) and then grown into log phase. Galactose (2%) was added to induce HO expression for 45 min. Glucose (2%) was then added and aliquots of the cultures were harvested at indicated time points. Genomic DNA was isolated and 10 ng used for PCR amplification. MATα was detected using primers JS301 and JS302. The SCR1 gene on chromosome V was used as a loading control (primers JS2665 and JS2666). PCR products were separated on a 1% agarose gel stained with ethidium bromide and then quantified using ImageJ. Donor preference with strains containing HMRα-B was performed as previously described [15]. Briefly, MATa was amplified with primers Yalpha105F and MATdist-4R from genomic DNA 90 after switching was completed (90 min), and then digested with BamHI. Ethidium stained bands were quantified using ImageJ. For the conditional V5-AID degron strains, degradation of V5-AID-fused Brn1 protein was induced by addition of 0.5 mM indole-3-acetic acid (Auxin, Sigma # 13750).
For assaying HO cleavage across MATa, HML, and HMR, the WT and sir2Δ strains containing the pGAL-HO-URA3 plasmid were induced with galactose for 0 to 2 hrs following growth in raffinose. Genomic DNA was then isolated and PCR across the HO-cleavage site performed with primers specific to each locus, and SCR1 used as a loading control. MATa was detected with JS301 and JS854, HML with JS3101 and JS3103, and HMR with JS3097 and JS3100. PCR was performed in the linear range and bands on ethidium stained agarose gels quantified with ImageJ. Real-time qPCR of HML and HMR with the same genomic DNA was performed with primers JS3101-JS3103, and JS3097-JS3098, respectively.
|
10.1371/journal.pntd.0002158 | Intensified Surveillance and Insecticide-based Control of the Chagas Disease Vector Triatoma infestans in the Argentinean Chaco | The elimination of Triatoma infestans, the main Chagas disease vector in the Gran Chaco region, remains elusive. We implemented an intensified control strategy based on full-coverage pyrethroid spraying, followed by frequent vector surveillance and immediate selective insecticide treatment of detected foci in a well-defined rural area in northeastern Argentina with moderate pyrethroid resistance. We assessed long-term impacts, and identified factors and procedures affecting spray effectiveness.
After initial control interventions, timed-manual searches were performed by skilled personnel in 4,053 sites of 353–411 houses inspected every 4–7 months over a 35-month period. Residual insecticide spraying was less effective than expected throughout the three-year period, mainly because of the occurrence of moderate pyrethroid resistance and the limited effectiveness of selective treatment of infested sites only. After initial interventions, peridomestic infestation prevalence always exceeded domestic infestation, and timed-manual searches consistently outperformed householders' bug detection, except in domiciles. Most of the infestations occurred in houses infested at baseline, and were restricted to four main ecotopes. Houses with an early persistent infestation were spatially aggregated up to a distance of 2.5 km. An Akaike-based multi-model inference approach showed that new site-level infestations increased substantially with the local availability of appropriate refugia for triatomine bugs, and with proximity to the nearest site found infested at one or two preceding surveys.
Current vector control procedures have limited effectiveness in the Gran Chaco. Selective insecticide sprays must include all sites within the infested house compound. The suppression of T. infestans in rural areas with moderate pyrethroid resistance requires increased efforts and appropriate management actions. In addition to careful, systematic insecticide applications, housing improvement and development policies that improve material conditions of rural villagers and reduce habitat suitability for bugs will contribute substantially to sustainable vector and disease control in the Gran Chaco.
| Vector-borne transmission of Chagas disease has not been effectively controlled in large parts of Latin America, particularly in the Gran Chaco ecoregion. To better understand the challenges in this region, we assessed the effectiveness of an intensified insecticide-based spraying strategy in suppressing the major vector Triatoma infestans from a well-defined rural area in northern Argentina. After an initial community-wide spraying, we intensively monitored infestation every 4–7 months for 35 months and applied insecticides selectively to the detected foci. In addition to the moderate levels of pyrethroid resistance reported in parallel, we found that selectively spraying only infested sites performed poorly. Prespraying bug abundance and the characteristics of infested sites before the initial interventions were reliable predictors of postspraying site infestation, including the type and use of the site and availability of refuges for the vector. We conclude that professional vector control based on residual insecticide spraying in an area with moderate pyrethroid resistance requires intense monitoring of house infestation, systematic insecticide applications and appropriate management actions. Operational, economic and political constraints to sustainable vector and disease elimination require complementary tools and approaches that favor changes in material conditions which reduce habitat suitability for the vector.
| Field trials carried out in Brazil and Argentina in 1948 demonstrated the effectiveness of hexachlorocyclohexane for suppressing domestic infestations with Triatoma infestans, one of the major vectors of Trypanosoma cruzi [1], [2]. More than 60 years later, residual insecticide spraying continues to be virtually the only tactic applied to triatomine control. Chagas disease vector control programs typically have an initial ‘attack phase’ (in which full-coverage applications of insecticide are made) followed by a ‘surveillance phase’, in which vector detection surveys and selective insecticide sprays are implemented [3]. Decades of vector control actions and screening of blood donors dramatically reduced the numbers of infected people and population at risk; Brazil, Chile and Uruguay were declared free of blood-borne and vector-borne transmission of T. cruzi mediated by T. infestans, and the extent and intensity of infestations were substantially reduced in some sections of Argentina, Bolivia and Paraguay [4], [5].
The major obstacle to attain effective control of the major vectors of Chagas disease using residual insecticide spraying has been the reappearance of triatomine bugs and the difficulties in addressing this recurrent process. Reinfestation of human habitations and peridomestic structures after insecticide application has been documented for most of the main triatomine species [4], [6]–[8]. Sources of reinfestation for T. infestans have usually been associated with passive bug transport by people, active dispersal of bugs from residual or untreated foci, and more rarely and with less supporting evidence, from sylvatic foci [9]–[14]. For several species such as Triatoma dimidiata, Triatoma brasiliensis and Rhodnius ecuadoriensis, sylvatic foci represent the major source of bugs [15]–[18]. These species pose particular problems to vector control programs because they inhabit nearby vegetation where chemical control is hampered or infeasible.
Despite progress in vector control status, T. infestans and Chagas disease persist as a major public health problem in many rural and some periurban communities in the Southern Cone countries [11]. The initial goal of eliminating T. infestans set by the Southern Cone Initiative in 1991 has not been reached yet in the Gran Chaco region –a 1.1 million km2 semiarid plain covering large parts of northern and central Argentina, southeast Bolivia, and central and western Paraguay [19]. Several key factors converge in the Gran Chaco to maintain house infestation with T. infestans: suitable environmental conditions, hosts and habitats for bug development; poor living conditions; irregular vector control activities coupled with intrinsic operational difficulties (e.g., access through dirt roads, limited transportation); relatively few resources assigned to vulnerable populations with low political visibility and high disease burden; diminished effectiveness of pyrethroid insecticides because of environmental conditions [11], [20].
In the Argentinean dry Chaco, house reinfestation after insecticide spraying was mainly associated with the occurrence of residual foci in peridomestic structures [21]–[23]. Randomized field trials demonstrated that a double-dose application of pyrethroid insecticides produced a greater initial impact on peridomestic populations of T. infestans than standard doses and reduced house reinfestation rates in the dry Chaco [24], [25]. However, the understanding of house reinfestation dynamics still is very limited because only a few field trials assessed insecticide effectiveness in more than 100–200 houses during one year or more and monitored infestations once or twice per year [11], [26], [27]. These methodological details are relevant because the generation time of T. infestans may range from 4 to 6 months depending on temperature and resource availability [28], [29]. In addition, resistance to pyrethroid insecticides in T. infestans has been detected in northern Argentina and Bolivia [30]–[33]. The reasons for the lack of success of the regional elimination of T. infestans may be multiple and remain unclear.
As part of a multi-site research program on the eco-epidemiology and control of T. infestans in the Gran Chaco, we assessed the long-term impacts on house infestation and bug abundance of an intensified control strategy based on full-coverage pyrethroid spraying followed by frequent vector surveillance and immediate selective treatment of the detected foci in a well-defined rural area in northeastern Argentina. Before initial control interventions, a multi-model inference analysis showed that availability of appropriate refuges for T. infestans, use of cardboard as a building material, and household numbers of domestic hosts were strongly and positively associated with site-specific bug infestation and abundance, whereas reported insecticide use by householders was negatively related to infestation [33]. No sylvatic foci of T. infestans were detected [34]. Monitoring of house infestation during the first 12 months postspraying (MPS) revealed unexpected vector control failures associated with moderate levels of pyrethroid resistance [35]. By extending these observations with unprecedented levels of spatio-temporal detail and extent up to 35 MPS, we here focus on persistence of infestation at site or house level over time and space; assess the effects of selective treatments with a standard or double dose of pyrethroids, and conduct a multi-model inference analysis of factors putatively related to new infestations at site level detected at 12 MPS or subsequently. This investigation identifies several constraints operating on surveillance and subsequent insecticidal treatments that challenge Chagas disease vector suppression attempts in general.
Fieldwork was conducted in a well-defined rural section (450 km2) of the municipality of Pampa del Indio (25°55′S 56°58′W), Province of Chaco, Argentina (see map and photos in [33], doi:10.1371/journal.pntd.0001349.g001). The study initially encompassed all existing 353 houses and 37 public buildings in 13 neighboring rural villages. Newly-built houses during the three-year follow-up led to a final count of 411 different houses. The two main ethnic groups are Creole and Toba. Vector control activities in the area had historically been very sparse. The last community-wide insecticide spraying campaign conducted locally by vector control personnel was carried out in 1995; a few houses were treated by villagers or local hospital staff in 2006. Before community-wide residual spraying with pyrethroid insecticides in December 2007, the prevalence of infection with T. cruzi in bugs (27.4%), dogs (26%) and cats (29%) was indicative of active domestic and peridomestic transmission (M.V. Cardinal et al., unpublished results).
A prospective cohort study was conducted between late 2007 and 2010. Surveys aiming at complete house coverage (i.e., a community-level census) were conducted at baseline and every 4–7 months during 35 months. A community-wide spraying with pyrethroid insecticides of all sites within each house compound was conducted immediately after the baseline survey [35]. Further interventions involved selective insecticide sprays of sites or house compounds infested during the follow-up period (Table 1). This study was approved by Institutional Review Board N° 00001678 (NIH approved) in Buenos Aires, Argentina.
Demographic and entomological surveys were conducted at baseline, during insecticide spraying, and at 4, 8, 12, 17, 22, 28 and 35 MPS. All existing houses were visited and its status recorded (inhabited, closed, abandoned, re-occupied, demolished, new). A sketch map of the spatial setting of all sites within each house compound was drawn, and each site was georeferenced and given an individual code in September or November–December 2007 (baseline, 0 MPS). The sketch map was updated during each visit. A house compound encompassed a domiciliary area with human habitations (sometimes in two separate buildings that counted as two domestic sites) and all sites within the peridomestic area (i.e., peridomicile) –usually a storeroom, a kitchen, an oven, one or more sites for chickens and other poultry (trees, coops, nests), one or more corrals, and a latrine. Each of these habitats characterized by some typical physical structure and use was considered an “ecotope”. A site (i.e., a patch) was any individual structure built and/or given a defined use by householders which might provide refuge for bugs.
All sites within each house were searched for triatomine bugs by timed manual collections (TMC) conducted by two skilled bug collectors using 0.2% tetramethrin (Espacial, Argentina) as a dislodging agent. Human habitations were inspected by one person for 20 min and each peridomestic site was searched by a second person for 15 min. In practice, each house compound averaged three peridomestic sites inspected and therefore the total search effort averaged one person-hour per house. In addition, most sites were inspected thoroughly before the stipulated time, and therefore search efforts were roughly similar across sites of different size. In several houses, bugs were also collected after the stipulated search time (after-manual collections), or by insecticide knock-down during insecticide applications (by the spray team) or a few days later (by householders) [35]. Local villagers were encouraged to capture bugs and hand them on to the research team during the subsequent visit. The collected triatomine bugs were transported to the field laboratory in plastic bags labeled with unique codes for house and bug collection site, identified taxonomically and counted according to species, stage and sex as described elsewhere [33].
The treatment criteria, insecticides and doses applied at different times are described in Table 1. The initial community-wide intervention sprayed all sites from 348 houses (including 325 inhabited and 23 vacant houses) with suspension concentrate (SC) deltamethrin (K-Othrin, Bayer, Argentina) at standard dose (25 mg/m2) applied by vector control personnel using backpack manual compression sprayers (Guarany, Brazil, and Hudson, Illinois) as described elsewhere [35]. Only four households refused insecticide spraying (not bug inspections) because they frequently sprayed themselves and their houses apparently were not infested, and another vacant house could not be accessed for treatment (Table S1). Selective sprays of all individual sites found infested with T. infestans at 4 or 8 MPS (including adjacent sites) were performed with deltamethrin upon completion of the 8 MPS survey. Likewise, sites found infested with T. infestans at 12 MPS (including adjacent sites within the same house compound and other sites that had not been sprayed at 8 MPS) were sprayed with SC β-cypermethrin (the only insecticide available to the vector control program at that time). To assess the impact of double-dose insecticide application on persistent infestations, standard (50 mg/m2) and double-dose (100 mg/m2) treatments with SC β-cypermethrin were assigned at random to infested peridomestic sites while a standard pyrethroid insecticide dose was applied in domiciles.
In view of the infestation levels recorded, from 17 MPS and thereafter the spray criterion was modified to full-spray coverage of infested house compounds (i.e., all sites within a house with 1 or more sites infested with T. infestans were treated with insecticide). Double-dose β-cypermethrin was used at 17 MPS. Field and laboratory-based evidence of local pyrethroid resistance [35] supported the application of a standard dose of malathion (1 g/m2) –the only effective alternative to pyrethroids available that was authorized by the corresponding federal agency at that time– to the few house compounds still infested with T. infestans at 22 and 28 MPS. At 22 MPS, one house was left unsprayed by mistake, and the owner of another house refused spraying; both houses were sprayed at 28 MPS.
All data reported correspond only to inhabited houses unless otherwise noted; no public building and only two uninhabited houses were ever found infested with T. infestans in the study area. The prevalence of infestation and colonization by T. infestans was computed either for sites or house compounds. Infestation was defined by the catch of at least one live T. infestans nymph or adult, and colonization by the catch of at least one T. infestans nymph. Persistent infestation of a site (or house) at time t was defined as the occurrence of infestation in a given site (or house) both at time t−1 and t. Estimates of infestation prevalence were based on the combined results of TMC and knock-down bug collections. Householders' bug collections were only considered when provided with precise information on date and site of capture; these and other data were used to distinguish between occasional invasions and established infestations. Bug abundance was computed as the number of live T. infestans collected in a specific site per 20 (domiciles) or 15 (peridomestic sites) person-minutes of search effort by TMC. As a measure of insecticide spraying effectiveness at site- or house-level, the percentage of sites (or houses) infested and sprayed at time t that were again found infested at time t+1 (i.e., apparently were persistently infested) was calculated for each selective spray round.
Data on reported insecticide use, ecotope, building materials, refuge availability, household numbers of people and domestic animals, and host resting places were collected in every survey starting on 4 MPS. The corresponding data for 0 MPS were extrapolated from the 4 MPS survey as explained elsewhere [35]. Demographic data for the 35 MPS survey were taken from the preceding survey at 28 MPS; although this procedure may introduce some inaccuracies, these should be trivial because only three sites were found infested at 35 MPS.
The association between new site-level infestations detected at 12 MPS or subsequently and refuge availability, reported insecticide use by householders, and distance to the closest infested site at t−1 and t−2 surveys was evaluated by means of multiple logistic regression analysis. Apparently new infestations were only considered from 12 MPS onwards; houses found infested at 4 or 8 MPS were excluded from this analysis because they had high chances of being locally persistent foci –at site or house level– after initial interventions. Thus, for this particular analysis, an infestation occurring at time t was considered new (i.e., not persistent) if it was found at 12 MPS (or subsequently) in a house considered uninfested at 4 and 8 MPS (or at t – 1). A multi-model inference approach based on Akaike's Information Criterion (AIC) was used to assess the relative importance (RI) of each variable [36] as detailed elsewhere [33]. The maximum value RI can take is 1, representing maximum relative importance, whereas RI = 0 represents no importance at all relative to the set of variables considered. Parameter estimates for each predictor variable were based on averaging the parameter value in each model including the predictor weighted by the Akaike weight of the respective model. Analyses and calculations were performed in R 2.7.0 [37].
The spatial distribution of houses with persistent infestations at 4 MPS was evaluated with respect to house infestation at 0 MPS (i.e., most bug colonies with late stages found at 4 MPS were very unlikely to have established after the initial insecticide spray because of the long generation time of T. infestans ranging from 4 to 6 months). The null model was built maintaining the pattern of infestation at 0 MPS (pattern 1) fixed, and randomizing the status of infestation of houses at 4 MPS (pattern 2) among all existing houses. The O-ring statistic O12(r) [38] was used to evaluate if the number of points of the randomized pattern 2 within a ring of radius r and a given width, centered at each point of the fixed pattern 1, corresponded on average to a random process (i.e., a homogeneous Poisson process; O12(r) = 1); aggregation of 2 relative to 1 (O12(r)>1), or regularity of 2 with respect to 1 (O12(r)<1). This procedure was implemented in Programita [39]. The grid size for analysis was 100 m; ring width, 400 m; maximum radius, 5 km; 999 simulations were performed, and the upper and lower 25th simulations were used as a 95% confidence envelope. A goodness-of-fit test was used to evaluate the overall fit of the observed pattern to the expected distribution [39].
Using the prospective data available, we assessed the hypothetical effects of adopting alternative insecticide spray criteria to the ones actually adopted (Table 1). The alternative criteria were either to spray all the sites within a house compound with one or more infested sites (i.e., the criterion applied originally at 17 MPS and thereafter), or to spray all sites within a given radius from the sites found infested. For each of the first four selective sprays rounds, we identified the sites that would have been sprayed at survey t if an alternative criterion had been applied; e.g., for the first criterion, by identifying all the sites within a house compound with an infested site at a given survey. For these identified sites, we searched for sites that were infested at survey t+1 and recorded the outcome at survey t+2, had these sites been sprayed at survey t+1. The outcomes at t+2 were then taken as the hypothetical outcomes for t+1 under the alternative criterion for the identified sites; for other sites, the observed infestation for t+1 was considered. Taken together, these outcomes represented the hypothetical infestation status under the alternative spray criteria for each survey. Distances between 0.1 and 5 km at 0.05 km increases were considered as hypothetical spray radii. As these criteria imply spraying more sites than those that were actually sprayed with insecticides, infestation would decrease solely because of the fact of treating more sites and not because nearby sites were treated. This procedure was used to calculate confidence envelopes for the hypothetical infestations. For each distance considered, as many sites as those that would be sprayed were randomly selected and the same calculations as with the actual sample were performed. This procedure was repeated 1,000 times for each distance and survey, and the upper and lower 25th values were taken as the 95% confidence envelope. Calculations were performed in Matlab 7.3.0 [40].
All of the 411 houses enumerated during the period September 2007–October 2010 were included in this study, although not all of them occurred at the same time point. Few (4–6%) houses were vacated between consecutive surveys, whereas newly-built or re-occupied houses represented 4–5% of the total number of inhabited houses at each survey. Very few households refused searches for bugs through the follow-up (Table S1). The main reason for lost-to-house inspection was that residents were repeatedly absent and access to closed premises through neighbors could not be arranged. A total of 4,053 sites was inspected for infestation at least once.
The initial community-wide insecticide spraying reduced the overall prevalence of house-level infestation with T. infestans from 49.5% before interventions to 12.3% and 8.9% at 4 and 8 MPS, respectively (Figure 1). After each of the first two selective treatment rounds conducted at 8 or 12 MPS, overall house infestation remained at 6.5–7% at 12 or 17 MPS. After the third selective treatment at 17 MPS, when all sites in any infested house were sprayed with a double-dose of pyrethroids (Table 1), overall house infestation fell to 3% at 22 or 28 MPS, and below 1% after selective treatments with malathion.
We assessed the effectiveness of selective pyrethroid sprays in suppressing site-specific infestations. TMC searches conducted 4–6 months after each of the first four pyrethroid spray rounds revealed that 5–13% of the treated sites were persistently infested between successive surveys (Figure 2A). No significant differences in effectiveness were detected between selective spray rounds regardless of the time elapsed after initial intervention (χ2 = 3.74, df = 3, P>0.25), treatment coverage (i.e., community-wide versus selective, χ2 = 3.20, df = 2, P>0.20), and insecticide dose (standard versus double dose, χ2 = 1.73, df = 1, P>0.15). Persistent infestations were detected 4–5 months after selective applications of pyrethroids with standard dose in 2 (4%) of 55 infested sites, and with a double dose in 3 (10%) of 30 infested sites. In sites negative before selective applications, infestations were subsequently detected in 1 of 21 sites sprayed with a standard dose, and in none of 9 sites sprayed with a double dose.
Regarding the effectiveness of selective treatments at house-compound level, TMC searches found persistent infestations (in at least one site per house) in 13–37% of the treated houses within 4–6 months after each of the four selective spray rounds (Figure 2B). No significant differences in effectiveness between spray rounds were detected despite variations in treatment criteria (χ2 = 3.70, df = 3, P>0.30). Infestation persisted to the subsequent survey conducted 4–5 months later in 10% of the 31 infested houses that were fully sprayed with a double dose of pyrethroids at 17 MPS, whereas none of the 292 negative houses not sprayed with pyrethroids at that time had a subsequent infestation.
The impacts of the initial community-wide pyrethroid spray on house infestation were not homogeneous across the study area (Figure 3). Houses with a persistent infestation at 4 MPS were spatially aggregated up to a distance of 2.5 km from houses infested at 0 MPS (Figure 4). Although the western and eastern sections had similar house infestation prevalence at 0 MPS (50.9% and 46.9%, respectively; Fisher's exact test, P = 0.49), infestation at 4 MPS in the western section (15.8%) was three times higher than in the eastern section (4.9%; Fisher's exact test, P = 0.005) (Figure 3). Considering the entire follow-up period, most (76%) of the infestations detected after initial intervention occurred in houses that had been infested before community-wide spraying with insecticides. When houses with a putative persistent infestation were excluded from consideration, postspraying infestation was still significantly more frequent among houses infested before initial spraying (17 of 122, 13.9%) than among those that had not been infested at baseline (10 of 190, 5.3%) (Fisher's exact test, P<0.001).
Of the eight sites (each in a different house) infested with T. infestans at 22 MPS that were immediately sprayed with malathion, two were infested at 28 MPS (see Text S1 for further details on apparent rainstorm effects). At 28 MPS, the 10 sites found infested (at 10 houses) were sprayed with malathion and none of them were found infested at 35 MPS.
Peridomestic infestation prevalence exceeded that in domestic sites during the entire follow-up after initial interventions, even though prespraying infestation was slightly higher in domiciles (Figure 1). The most frequently infested ecotopes before the initial community-wide spraying with pyrethroids were also the ones most frequently infested after selective treatments, including domiciles, kitchens or storerooms, fowl coops and ‘nideros’ –an elevated shelf made of wood or sometimes bricks where chickens, and occasionally turkeys or ducks, nested (Figure 5). Corrals and other types of ecotope (latrines, ovens, trees with chickens, and others) were rarely infested. The relative frequency of infested sites at 4 MPS increased with increasing bug abundance determined by TMC before initial interventions (Figure 6).
The reported use of insecticides by householders varied significantly from 41.4% to 69.0% during the follow-up (Table S2) (χ2 = 63.14, df = 4, P<0.0001). These variations were mainly caused by an increase in the number of households using domestic insecticide aerosols at least every two months. Householders reported applying insecticides in domiciles, kitchens and storerooms.
The results of the multi-model inference analysis of factors putatively related to new infestations at the first year of interventions or subsequently are presented in Table 2. Refuge availability showed the highest RI (0.98) followed by distance to the nearest site found infested at the preceding survey (RI = 0.81) and, much less important, by distance to the nearest site found infested two surveys earlier (RI = 0.60). Reported insecticide use had a low RI (0.39). The chance of new infestations increased with more refuges for bugs and with more proximity to the nearest infested site at the preceding surveys.
TMC significantly outperformed householders' bug detection before and after interventions except in domiciles after initial insecticide spraying (Table 3). However, householders contributed to enhanced detection of T. infestans by capturing and handing bugs on to the research team in 108 occasions, 66% of which occurred in sites negative by TMC at the survey that immediately followed householders' collections. Conversely, TMC detected T. infestans bugs in 328 occasions, 89% of which occurred in the absence of householders' collections. Householders collected only one adult T. infestans in 48 (44%) of the occasions; >1 adult in 50 (46%) occasions, and nymphs in 40 (37%) of their bug collections. Householders captured bugs most frequently in domiciles (72 of 108), followed by kitchens, storerooms and ‘nideros’.
For each of the first four selective spray rounds (8–22 MPS) we assessed the hypothetical effects of spraying all sites within a house compound that at least had an infested site, and spraying all sites within a given radius around the detected focus. The number of hypothetically infested sites declined gradually up to a radius of 2–3 km while the total number of hypothetically sprayed sites increased considerably (Figure 7). Such decline in infested sites exceeded that expected by chance up to 0.5 km (∼1,500 sites sprayed) for the first selective spray round (Figure 7A) and up to 2–3 km (∼2,300 sites sprayed within 2 km) for the second round (Figure 7B). Decreases in the number of hypothetically infested sites in the two subsequent rounds –when the entire infested house compounds were actually sprayed– were within the confidence envelope expected by chance (Figure 7C,D).
Overall figures of the impacts of insecticide spraying on infestation provide a broad picture while hiding some relevant observations. Below we report on several specific cases (sites or houses), and relate their infestation status to defined events or processes:
Our results show that although house infestation was dramatically reduced after initial interventions, the local elimination of T. infestans was not achieved even after three years of intensified vector surveillance and frequent selective insecticide sprays conducted by professional personnel under close supervision. The local occurrence of bug populations with moderate resistance to pyrethroid insecticides [35] and the limited effectiveness of selective treatment of infested sites contributed to the challenging scenario emerging in the Gran Chaco.
The full-coverage pyrethroid spraying did not reach the expected target of vector control programs (<5% of infested houses within 6–12 MPS, to account for technical errors including suboptimal spray coverage) despite the occurrence of lower baseline infestation levels than in other similar rural areas with no recent vector control actions [11], [27], [28], [41], [42]. Subsequent selective sprays with pyrethroids did not perform better, and its effectiveness remained approximately stable across several spray rounds (Figure 2) despite bug colonies had been decimated. Unlike in previous trials [21], [22], both simple- and double-dose applications had similarly limited effects consistent with the occurrence of diminished susceptibility to pyrethroid insecticides [35].
We recorded a large degree of spatial and temporal heterogeneity in infestation, the apparent effectiveness of insecticide spraying, and people's practices related to infestation (Text S1). The spatial distribution of pyrethroid resistance was apparently aggregated, as determined from the occurrence of early persistent infestations. None of these facts would have been noticed without frequent monitoring of infestations at site level and upscale. Few vector control failures associated with pyrethroid resistance have been reported [30], [31], [35], whereas the number of resistant T. infestans populations scattered through most of its current distribution range has gradually increased [32]. Alerted by our findings, in 2010 the Chagas vector control program detected a new focus of high pyrethroid resistance associated with repeated control failures at 100 km from our study area [43], in a district which also had a history of sporadic house spraying with insecticides. Adequate monitoring of treatment effects in routine operations of vector control programs may reveal hitherto unknown resistant foci and the actual effectiveness of control interventions.
Selective pyrethroid sprays of sites infested only performed poorly. Frequently the treated bug colonies were apparently suppressed, yet in a few months other infestations became detectable in adjacent sites within the same house compound (Figure S1), either because they had not previously been detected by TMC or because bugs dispersed actively from the detected foci around the time of selective treatment. In our study, the multi-model based association between new site-level infestations and the proximity of other foci detected at one or two preceding surveys is consistent with: (i) spatial aggregation of T. infestans foci occurring at various scales [21], [44]–[46]; ii) a six-month time lag between detection of foci and dispersal events of T. infestans inferred from the spatio-temporal dynamics of reinfestation patterns [47], [48], and iii) frequent flight or walking dispersal of T. infestans during spring-summer [e.g., 49]. These findings support the extension of spray coverage to all sites within the house compound and the consideration of the compound as the minimal vector control unit, rather than the individual infested sites.
The simulation results further suggest that enhanced bug control would be achieved if all suitable sites within 500–1,000 m of the detected foci were sprayed. Similarly, a longitudinal study conducted in the Argentinean dry Chaco reported significant spatial aggregation of reinfested sites at 25–500 m around residual foci, and recommended extending selective insecticide sprays up to a distance of 450–500 m around the detected foci [21], [44]. This tactic (justifiable if the goal were vector elimination) implies a substantial increase in the frequency of sites sprayed and the resources needed. Therefore, its relative merits must be framed within the stringent operational and economic constraints of vector control programs in the study region; variations in the spatial layout of villages (i.e., connectivity), and eventual landscape effects on vector dispersal (e.g., barriers).
Refuge availability, the main bug habitats, and prespraying bug abundance were closely related to postspraying site-level infestation after initial interventions, in agreement with existing evidence [35]. The multi-model inference approach showed that refuge availability was highly important for explaining variations in site infestation before and after interventions. Similarly, nearly all detected infestations occurred in the same four key ecotopes before and after interventions. Moreover, peridomestic ecotopes were more frequently infested than domiciles after initial interventions –a recurrent pattern related to the higher exposure of peridomestic structures (especially chicken coops and ‘nideros’) to sunlight, rain and dust, all of which undermine the activity of pyrethroids [22], [24], [25]. The relative occurrence of a persistent site-level infestation at 4 MPS increased substantially with increasing prespraying bug abundance, as recorded elsewhere [22], [24], [25], [50]. Conversely, reported insecticide use by householders had a low RI for explaining new infestations despite it was closely associated with prespraying infestation in domiciles, kitchens and storerooms [33]. Such differences are probably related to the few domiciles, kitchens and storerooms found infested at 12 MPS or subsequently. Taken together, these results are highly relevant for improved vector control and imply that: i) factors with substantial effects on infestation remain approximately stable before and after insecticide spraying (e.g., refuge and host availability), and ii) prespraying data on the main types of infested ecotopes and bug abundance provide valuable information on the future effectiveness of insecticide spraying and may be used for identifying sites, houses or village sections most likely to be problematic for vector control.
House infestation prevalence fell below 1% only after multiple inspection and selective spray rounds with pyrethroids and finally, selective treatment of highly persistent foci with malathion. During the follow-up, several other local events potentially affecting infestation may have confounded the specific effects of insecticide spraying. Some of the cases analyzed (Text S1) illustrate modifications introduced by householders (e.g., physical structure of sites, removal of infested sites, host management, and non-professional insecticide use) combined with adverse effects related to occasional rainstorms, operational problems during insecticide application (e.g., sites difficult to spray adequately, as in a storeroom full of corn, or errors in procedures or planning), and imperfect detection methods [41], [51]–[53]. These factors may enhance or diminish substantially the effects of insecticide spraying.
Timed manual searches conducted by skilled personnel using a dislodging agent is the standard method used to assess infestations in intervention trials despite its limited sensitivity and precision, especially at low bug densities [8], [35], [51], [53]. In our study, its shortcomings were partially compensated by recurrent, very frequent searches of bugs in identified sites (averaging approximately one person-hour per house compound) and promotion of householders' bug collections. The overall frequency of infestations detected by TMC (and bug catches) was much larger than those achieved by householders, unlike in other settings with lower bug densities [51], [54]. Local villagers were aware that insecticidal treatments would continue regardless of their compliance with capturing and keeping the bugs, and therefore may have been less motivated to do so. Community-based vector surveillance has played an increasing role over recent decades [8], [24], [42], [55], [56], yet the ability of householders to detect bugs is widely variable depending on various factors [8], [51], [57] and may be more difficult to standardize. Householders detected proportionally more infestations in domestic rather than peridomestic ecotopes [51], perhaps because they were more motivated to suppress bugs from sleeping quarters or were there when bugs emerged from refuges. They also detected several infestations missed by subsequent TMC searches, several of which may have been recent invasions (not established bug colonies). More attention needs to be given to vector surveillance in peridomestic sites either through appropriate training of rural villagers or using baited traps [14].
One limitation of our study is that we have not assessed the impact of interventions on parasite transmission, as vector-borne transmission of T. cruzi to humans and dogs may occur at very low infected-bug densities in high-risk areas [58]. The use of a house infestation prevalence of 5% as a threshold for parasite transmission mediated by T. infestans “is not supported by rigorous evidences but rather derived from data on Triatoma infestans in Brazil without scientific justification” [59]. The validity of the refuge availability index and other predictors was discussed before [35]. Although most of the bugs collected after interventions most likely survived treatment at site level (i.e., residual foci), immigrant bugs from other sources may explain in part new infestations. Use of microsatellite markers and wing geometric morphometry [e.g.], [ 12,18] may provide concluding evidence on their relative contribution. A major strength of our intervention trial was the detailed information collected systematically at site level in a sizable number of house compounds every 4–7 months over a three-year period.
Evidence of the obstacles to suppress T. infestans in the Gran Chaco ecoregion [e.g.], [ 11], [21,28] led the Southern Cone Initiative to turn from the initial goal of vector elimination into the less ambitious one of controlling house infestations and interrupting vector- and blood-borne transmission in recent years [60]. Our current results document substantial geographic variations in the characteristics of persistent foci in the region [21], [28], [29], [33], and provide guidance on the effort levels needed to suppress T. infestans in an area with moderate pyrethroid resistance. The chronic limitations in personnel and resources in the Latin American health sector (more so in rather remote rural areas) pose serious obstacles to vector and disease suppression efforts. Research on the real functioning of disease control programs and their capacity to operate and modify what they do and how they do it is needed to push knowledge closer to its effective application [61], [62]. The ultimate limitations of insecticide-based control strategies are that they do not change the material conditions that favor the occurrence and spread of domestic vectors [63] such as T. infestans, and the eventual emergence of insecticide resistance. In addition to careful, systematic residual insecticide applications, our findings confirm that housing modifications and development policies that improve material conditions of rural villagers and reduce habitat suitability for T. infestans [35] may contribute substantially to sustainable vector and disease control in the Gran Chaco.
|
10.1371/journal.pgen.1003774 | An Alteration in ELMOD3, an Arl2 GTPase-Activating Protein, Is Associated with Hearing Impairment in Humans | Exome sequencing coupled with homozygosity mapping was used to identify a transition mutation (c.794T>C; p.Leu265Ser) in ELMOD3 at the DFNB88 locus that is associated with nonsyndromic deafness in a large Pakistani family, PKDF468. The affected individuals of this family exhibited pre-lingual, severe-to-profound degrees of mixed hearing loss. ELMOD3 belongs to the engulfment and cell motility (ELMO) family, which consists of six paralogs in mammals. Several members of the ELMO family have been shown to regulate a subset of GTPases within the Ras superfamily. However, ELMOD3 is a largely uncharacterized protein that has no previously known biochemical activities. We found that in rodents, within the sensory epithelia of the inner ear, ELMOD3 appears most pronounced in the stereocilia of cochlear hair cells. Fluorescently tagged ELMOD3 co-localized with the actin cytoskeleton in MDCK cells and actin-based microvilli of LLC-PK1-CL4 epithelial cells. The p.Leu265Ser mutation in the ELMO domain impaired each of these activities. Super-resolution imaging revealed instances of close association of ELMOD3 with actin at the plasma membrane of MDCK cells. Furthermore, recombinant human GST-ELMOD3 exhibited GTPase activating protein (GAP) activity against the Arl2 GTPase, which was completely abolished by the p.Leu265Ser mutation. Collectively, our data provide the first insights into the expression and biochemical properties of ELMOD3 and highlight its functional links to sound perception and actin cytoskeleton.
| Autosomal recessive nonsyndromic hearing loss is a genetically heterogeneous disorder. Here, we report a severe-to-profound mixed hearing loss locus, DFNB88 on chromosome 2p12-p11.2. Exome enrichment followed by massive parallel sequencing revealed a c.794T>C transition mutation in ELMOD3 that segregated with DFNB88-associated hearing loss in a large Pakistani family. This transition mutation is predicted to substitute a highly invariant leucine residue with serine (p.Leu265Ser) in the engulfment and cell motility (ELMO) domain of the protein. No biological activity has been described previously for the ELMOD3 protein. We investigated the biochemical properties and ELMOD3 expression to gain mechanistic insights into the function of ELMOD3 in the inner ear. In rodent inner ears, ELMOD3 immunoreactivity was observed in the cochlear and vestibular hair cells and supporting cells. However, ELMOD3 appears most pronounced in the stereocilia of cochlear hair cells. Ex vivo, ELMOD3 is associated with actin-based structures, and this link is impaired by the DFNB88 mutation. ELMOD3 exhibited GAP activity against Arl2, a small GTPase, providing a potential functional link between Arf family signaling and stereocilia actin-based cytoskeletal architecture. Our study provides new insights into the molecules that are necessary for the development and/or function of inner ear sensory cells.
| Many molecular components that are necessary for the development and maintenance of hearing have been discovered by identifying the genes that underlie hearing impairment in humans and mice [1]–[4]. Hearing requires the precise and efficient functioning of intricately structured mechanosensory hair cells and supporting cells in the inner ear [3]. One of the key structures in the mechanotransduction process is the hair cell stereocilium. Protruding from the apical surface of the hair cells, stereocilia are organized in three rows of decreasing height in a staircase pattern. Each stereocilium is composed of an actin core that contains cross-linked and bundled γ- and β-actin microfilaments that are uniformly polarized, with the barbed (positive) ends localized at the tip. At the tapered end of the stereocilium, the actin filaments form a rootlet that has been proposed to anchor the structure in the actin-rich meshwork of the cuticular plate [5]. Interestingly, among the identified hearing loss-associated genes, nineteen encode proteins that interact with actin [6], [7]. Numerous studies have demonstrated that actin cytoskeleton-associated proteins are involved in the development, maintenance and stabilization of the stereocilia (for review, see [6]).
Continuous depolymerization of actin filaments at the base and polymerization at the barbed end, termed treadmilling, is thought to be critical to the maintenance of the length of stereocilia [8], [9]. However, a recent study demonstrated a rapid turnover of the actin filaments only at the tip of the stereocilia, without a treadmilling process [10], emphasizing the specific role of proteins at the stereocilia tip in the regulation of actin filaments. Regardless of the precise site, it is quite clear that the proper regulation of actin dynamics is critical to the generation and maintenance of stereocilia as sensory structures.
The Rho/Rac/Cdc42 family of GTPases is well known as a regulator of actin. Rho and Rac in the inner ear are involved in the morphogenesis and growth of the otocyst [11], [12]. The depletion of Rac1 or both Rac1 and Rac3 in the murine inner ear leads to a shorter cochlear duct with an abnormal sensory epithelium. Rac may participate in cell adhesion, proliferation, and movements during otic development [11], [12]. Several studies have suggested that the activation/inhibition of Rho pathways control the actin depolymerization rate in the outer hair cells [13], [14]. Although best known for their roles in the regulation of membrane traffic, there is growing evidence that GTPases in the Arf family can also act via changes in actin. [15].
Here, we report the identification of a new deafness gene, which encodes an ELMO/CED 12 domain containing protein, ELMOD3. Our biochemical studies demonstrated that ELMOD3 possesses GAP activity against a small GTPase in the Arf family, Arl2, providing a functional link between Arf family signaling pathways and stereocilia actin-based cytoskeletal architecture. GAPs are regulators and effectors of the Ras superfamily of GTPases, which are increasingly recognized as providing specificity as well as temporal and spatial regulation to GTPase signaling [16]. Thus, we believe that the identification of ELMOD3 role in the inner ear provides new insights into signaling processes that are important to hearing in humans.
Family PKDF468 (Figure 1A) was recruited after obtaining Institutional Review Board approval and written informed consent. The family history revealed that the onset of hearing loss was pre-lingual, with no clear vestibular impairment among the deaf individuals. Pure-tone bone and air-conduction audiometry revealed severe-to-profound mixed (conductive and sensorineural) hearing loss in the affected individuals of family PKDF468 (Figure 1B). Individual V:2 exhibited severe-to-profound mixed hearing loss, with bone conduction thresholds for the right ear displaying a mild downward slope to the severe hearing loss range. The left ear displayed slightly better bone conduction thresholds with normal values for lower frequencies and a downward slope to severe hearing loss at higher frequencies (Figure 1B). The audiograms of individual V:5 revealed bilateral severe-to-profound mixed hearing loss, with a large conductive component in both ears. The bone conduction thresholds exhibited a mild downward slope to moderately severe hearing loss for the right ear and were slightly better on the left, for which the thresholds ranged from borderline normal to moderate hearing loss ranges (Figure 1B). The clinical evaluation revealed no clear signs of skin, renal, or retinal abnormalities.
To determine the temporal bone malformation, we performed computed tomography (CT) scans of two affected (V:2 and V:11) along with a normal hearing sibling (V:7). CT scan of individual V:2 revealed all three semicircular and internal auditory canals were intact on both sides. The middle ear and mastoid appeared well-aerated bilaterally. Imaging of individual V:11 demonstrated a slightly narrow appearing internal auditory canal on the right side only. The mastoid air cells and middle ear cleft were well-aerated bilaterally. The external auditory canal appeared normal as well for both affected individuals.
We initially observed that deafness in family PKDF468 did not co-segregate with short tandem repeat (STR) markers for 74 of the reported recessive nonsyndromic deafness loci (data not shown). We therefore performed a genome-wide linkage analysis and observed that the deafness phenotype of family PKDF468 exhibited significant evidence of linkage to STR markers on chromosome 2p12-p11.2 (Figure 1A). Additional STRs on 2p were genotyped, and haplotype analysis revealed a 0.91 Mb linkage interval that was delimited by the markers D2S1387 and D2S2232 (Figure 1A). Under a recessive model of inheritance, with a disease allele frequency of 0.001 and full penetrance, a maximum two-point LOD [17] score of 4.74 (θ = 0) was obtained for the marker D2S2333. These results define and delimit DFNB88 [Human Genome Nomenclature Committee (HGNC) approved locus symbol], a novel recessive deafness locus on chromosome 2p11.2.
The DFNB88 locus partially overlaps with the dominant deafness locus DFNA43 (Figure 1C) [18]. Four known candidate genes were identified within the DFNB88/DFNA43 overlapping linkage region (Figure 1C). However, Sanger sequencing of these genes did not reveal any pathogenic variants. Approximately 85% of the disease-causing mutations in Mendelian disorders reside in coding regions or in exon-intron canonical splice junctions [19]. We therefore performed exome sequencing of an affected individual from family PKDF468. The sample was enriched using the NimbleGen SeqCap EZ Exome Library v2.0 (Roche Diagnostics; San Francisco, CA), and 100 bp, paired-end sequencing was performed on the Illumina HiSeq 2000 platform (Illumina). An average of 78.94% of bases were sequenced with 20× coverage within the targeted regions. This yielded a total of 64,863 single-nucleotide variants, of which 1,928 were not found in the dbSNP133 database (Table S1). Based on the recessive mode of inheritance evident in the pedigree, we analyzed genes with homozygous changes and potential compound heterozygous changes. Additionally, we removed all of the variants that were present in six ethnically matched control samples (Table S1). No mutation segregating with hearing loss in family PKDF468 was identified in any of the known deafness-causing genes (Table S1).
We identified one homozygous transition mutation, c.794T>C (p.Leu265Ser), in ELMOD3 (Figure S1) on chromosome 2p11.3 (Figure 1C) that segregated with DFNB88-linked deafness (Tables S1 and S2). The c.794T>C change was not present near the canonical splice junctions and was not predicted to create any aberrant splice site. However, to confirm that c.794T>C did not affect splicing of ELMOD3 transcripts, we generated cDNA libraries using the total RNA extracted from the white blood cells of two affected and one normal hearing individual. Sanger sequencing of sub-cloned PCR products, amplified using primers in either exons 9 and 11 or in exons 9 and 12 (Figure S2), did not reveal any aberrant splicing product in affected individuals. Thus, the likely pathogenic affect of the c.794T>C change is substitution of a highly conserved leucine residue at amino acid position 265 of the human ELMOD3 protein with serine (Figure 2C). No carrier of c.794T>C was identified among 524 ethnically matched control chromosomes, in the 1000 Genome database or in the 6500 individuals who are listed in the NHLBI-ESP variant database (http://evs.gs.washington.edu/EVS/). Moreover, Polyphen-2 [20], SNPs3D [21], MutationTaster [22], PMut [23], and SIFT [24] predicted that the ELMOD3 mutation would be deleterious (Table S3).
To further confirm that the p.Leu265Ser allele of ELMOD3 is the only mutation that was associated with hearing loss at the DFNB88 locus, we sequenced the coding, non-coding, and approximately 75 bp flanking sequences of the exon-intron boundaries of all the known candidate genes present within the linkage region in two affected individuals of family PKDF468 (Figure 1C). No other potentially pathogenic mutation was identified in the affected individuals of family PKDF468. Although, ELMOD3 is located outside the reported linkage interval of DFNA43 (Figure 1C) [18], nevertheless we sequenced DNA samples of two affected individuals from the original DFNA43 family and no mutation was found.
We next examined the gene structure and expression of ELMOD3. Seven alternatively spliced isoforms of human ELMOD3 were identified (Figure 2A). Isoform A (reference sequence NM_ 032213.4) has a translation initiation codon (AUG) in exon 2, ten coding exons that encode a polypeptide of 391 residues (Figure S1B). Exons 7 to 11 encode the engulfment and cell motility (ELMO or CED12) domain, which consists of 164 amino acid residues (Figure S1B; blue box). ELMOD3 isoforms B to D include alternatively spliced exons in the 5′ untranslated region (UTR) but harbor the same coding exons and encode identical 381 residue polypeptides that differ from isoform A only at their carboxy termini (Figures 2A and S1B). The human ELMOD3 isoforms, A and B, share 87% identity, with all the differences clustered near the C-terminus. Isoforms E, F, and G do not encode the full-length ELMO domain due to alternate splicing of exons in the carboxy terminus (Figure 2A). The c.794T>C transition mutation is predicted to result in the substitution of serine for a highly conserved leucine in all of the ELMO domain-containing isoforms of ELMOD3 (Figures 2A and 2C). In comparison to the human sequence, mouse Elmod3 includes only three known alternatively spliced transcripts (Figure 2A). RT-PCR and real-time quantitative PCR analysis of multiple human and mouse tissue cDNAs (Tables S4 and S5) revealed the ubiquitous expression of isoforms A/a and B–D/b–c (Figures S3 and 2B). We also assayed the relative mRNA expression of murine Elmod3 isoforms a and b–c with real-time quantitative RT-PCR of RNA that was extracted from cochlear and vestibular inner ear tissues from postnatal day 0 (P0), P10, and P30 C57BL/6J mice (Figure 2D). The expression of Elmod3 isoform b–c was several-fold higher than isoform a, in both cochlear and vestibular tissues at all of the time points examined (Figure 2D). Therefore, we focused on the ELMOD3 isoform B for the subsequent biochemical and cellular studies. The mouse ELMOD3 protein is 70% and 80% identical to the A and B isoforms of human ELMOD3, respectively, and again the differences are greatest at the C-terminus, although single amino acid changes are scattered throughout the alignments.
To characterize the cellular localization of ELMOD3, we produced a rabbit polyclonal antiserum against synthetic peptide immunogens from mouse ELMOD3 isoform b. The sensitivity and specificity of the ELMOD3 antibody was validated in immunoblot and immunofluorescence analyses, in transfected cells and mouse tissues (Figures S4 and S7). Our antibodies specifically recognized ELMOD3 isoform b but not murine ELMOD1, ELMOD2, or ELMOD3 isoform a (Figure S4). We next performed immunolocalization of ELMOD3 in the rat and mouse organ of Corti (Figures 3 and 4). In rat cochlea, ELMOD3 immunoreactivity was observed in the stereocilia, kinocilia and cuticular plate of developing hair cells (Figures 3 and S5). Before P07, ELMOD3 staining was very weak in the inner hair cells stereocilia. By P07, in auditory hair cells, patchy labeling of ELMOD3 immunostaining was detected along the length of stereocilia (Figure 3). In contrast to actin staining, ELMOD3 immunoreactivity was not uniform along the length of each stereocilium and the protein seemed to be excluded from a region near the tip (Figure 3E). ELMOD3 immunoreactivity was also found in the supporting cells, including pillar and Dieters' cells (Figure 3). Similar to that seen in the rat (Figure 3D), the stereocilia of inner hair cells in the mouse organ of Corti were more intensely labeled than those of outer hair cells (Figure 4A). In contrast to the cochlear hair cells, ELMOD3 antibody labeling was observed within the hair cell bodies in the vestibular end organs of both rat and mouse inner ear, but no prominent immunoreactivity was observed in the hair bundles (Figures 4B–4C and S6). These observations suggest a unique role for ELMOD3 in cochlear sensory cells and may reflect the functional or structural differences between cochlear and vestibular hair bundles. No specific immunoreactivity was observed when the primary antibody was omitted (data not shown) or when the antibody was pre-incubated with the ELMOD3 peptide antigen (Figure S7).
We examined LLC-PK1-CL4 epithelial (CL4) cells to understand the mechanism and effect of the hearing loss-associated allele of ELMOD3. CL4 cells contain actin-rich microvilli and have been used as in vitro models of stereocilia to examine F-actin and protein dynamics [25]. We transiently co-transfected GFP-ELMOD3 constructs with Espn constructs, where the latter was used to over-elongate the microvilli at the CL4 cell surface [26] (Figure 5A–5B). We observed a significant expression of GFP-ELMOD3 in the apical (microvillar) plasma membrane twenty-four hours post-transfection (Figure 5A). We also observed expression of GFP-ELMOD3 in the cytosol of transfected cells (Figure 5A). In contrast to the wild type protein, the p.Leu265Ser mutation in the ELMO domain yielded a protein that displayed either weak or no labeling in the microvilli of the transfected CL4 cells. Additionally, the protein appeared to be diffusely located throughout the cytoplasm, with a nuclear concentration in approximately half of the transfected cells (Figure 5B). Identical results were observed with tdTomato-tagged wild-type and mutant ELMOD3 constructs (data not shown).
To determine the effect of p.Leu265Ser mutation on the localization of ELMOD3 in the mouse inner ear, we performed gene gun-mediated transfection of wild-type and p.Leu265Ser mutant GFP-tagged ELMOD3 cDNA constructs in organotypic cultures of inner ear sensory epithelia of P2 C57BL/6J mice (Figure 5C–5D). Over-expressed wild-type GFP-ELMOD3 localized along the length of the stereocilia of cochlear hair cells (Figure 5C). We also observed homogeneous distribution throughout the hair cell bodies (Figure 5C). Similar to the results that were observed in CL4 cells, GFP-ELMOD3 harboring the p.Leu265Ser mutation failed to target to the stereocilia, and the protein was apparently distributed throughout the cochlear hair cell bodies (Figure 5D). Taken together, these results support our conclusion that ELMOD3 localizes to actin-based microvilli and stereocilia (Figure 5) but that a point mutation in the ELMO domain can prevent its normal localization and potentially affect its function in the stereocilia.
To further investigate the ELMOD3-actin association, we transfected GFP-tagged ELMOD3 into MDCK cells, which is a highly polarized cell model system (Figure S8). Forty-eight hours post-transfection, GFP-ELMOD3 accumulation was apparent at the periphery of the transfected cells near the plasma membrane (Figure S8A). The expression of GFP-ELMOD3 harboring the p.Leu265Ser mutation in MDCK cells resulted in a protein that failed to target or accumulate at the plasma membrane and instead, appeared to concentrate in the nuclei (Figure S8B). To determine whether GFP-ELMOD3 associates with the actin cytoskeleton at the plasma membrane (Figure 6A), we treated the cells with cytochalasin D (cyto-D), which is a potent inhibitor of actin polymerization, to disrupt the actin cytoskeleton [27], [28]. We hypothesized that if GFP-ELMOD3 associated with the actin cytoskeleton at the cell membrane, then treatment of the cells with cyto-D would also affect ELMOD3 localization. Indeed, we observed a significant decrease in the GFP-ELMOD3 signal at the cell membrane following disruption of the actin cytoskeleton (Figure 6B, 6D). Four hours following cyto-D treatment (i.e., the recovery period for actin re-polymerization) [27], we observed that ELMOD3 re-accumulated at the cell membrane (Figure 6C, 6D). These results suggest that the localization of ELMOD3 is dependent on the actin cytoskeleton and/or may contribute to a mechanism that supports its maintenance.
To decipher the link between F-actin and ELMOD3, we performed a two-color stochastic optical reconstruction microscopy (STORM) imaging of EGFP-ELMOD3 transfected MDCK cells. While conventional confocal acquisitions revealed co-localization of ELMOD3 and the actin-cytoskeleton, this high resolution imaging technique allowed us to determine more precisely the relative positions of ELMOD3 and actin (Figure 7). ELMOD3 and actin were each found in close apposition to the plasma membrane and in irregularly shaped puncta (Figure 7C). Many regions of extensive overlap in staining between ELMOD3 and actin, suggest the possibility that a subset of the actin-based structures may contain ELMOD3 but that each protein is also found localized independently of the other at the plasma membrane (Figure 7C).
To test the possibility that ELMOD3 binds to actin-based structures, we performed a high-speed co-sedimentation assay that pellets actin along with its associated proteins. To obtain a source of purified ELMOD3, His6-Trigger factor-ELMOD3 (TF-ELMOD3) fusion protein was expressed in bacteria, and the recombinant protein was purified by Ni-NTA chromatography. The TF-ELMOD3 or control proteins were incubated with polymerized F-actin and subjected to high-speed centrifugation at 150,000× g for 1.5 hrs (Figure S9). TF-ELMOD3 co-sedimented, albeit weakly or incompletely, with F-actin in this assay (Figure S9). Under these conditions, the p.Leu265Ser mutation did not significantly impact the level of TF-ELMOD3 that co-sedimented with F-actin (Figure S9).
Human ELMOD1 and ELMOD2 each possess Arl2 GAP activity [29]. We therefore investigated whether ELMOD3 also possesses GAP activity against Arl2. Previous tests of bacterially expressed human ELMOD3 as either maltose binding proteins or trigger factor fusion proteins were negative, but the homologous preparations of ELMOD1 and ELMOD2 were found to possess very low specific activities as Arl2 GAPs, compared to the preparation purified from bovine tissues. To obtain a potentially more active preparation of human ELMOD3, we expressed ELMOD3 and the mutant p.Leu265Ser in HEK293T cells as N-terminal GST-fusion proteins to facilitate protein purification. Protein expression and purification from ∼108 HEK293T cells that expressed GST-ELMOD3 or the mutant each yielded ∼0.6 mg protein. These preparations were stable at 4°C and against freeze-thaw cycles, as judged by either GAP activity or lack of precipitation. We expressed and purified GST alone and used it as a negative control in all of our assays. We also evaluated the effect of cleavage of the GST fusion tag by TEV protease on ELMOD3 activity; no changes in activity in the Arl2 GAP assay were observed compared to un-cleaved proteins (data not shown). Thus, we believe that the presence of the GST moiety at the N-terminus does not interfere with access to the substrate or with enzymatic activity in our assay.
When we varied the amount of GST-ELMOD3 protein in the Arl2 GAP assay, we observed a dose-dependent response and evidence of saturation at higher protein concentrations (data not shown). Using the lower concentrations of GST-ELMOD3 to estimate the initial rates of GAP-dependent activity and estimating the purity of the preparation at 50% (based on visual inspection of Coomassie blue-stained gels), we obtained a specific activity of 24 pmol of GTP hydrolyzed/min/mg (Figure 8). This specific activity of GST-ELMOD3 as an Arl2 GAP is approximately 32-fold lower than that determined for GST-ELMOD1 and nearly 1000-fold lower than that of GST-ELMOD2 or bovine testes ELMOD2, which is the most active reported preparation of any Arl2 GAP [29]. Thus, in contrast to our earlier report that it is inactive, GST-ELMOD3 does exhibit Arl2 GAP activity, and we believe that its lower specific activity when expressed in bacteria likely contributed to the earlier negative findings [29]. The differences in specific activity among the three human GST-ELMOD preparations from HEK293T cells are predicted to result from differences in substrate specificity, sensitivities to co-activators (as known for Arf GAPs), or both. Thus, more studies are required to determine whether the biologically relevant substrate of ELMOD3's GAP activity in the inner ear is Arl2 or a related GTPase.
We next assessed the effects of the Leu265Ser point mutation on the Arl2 GAP activity of GST-ELMOD3. The mutant protein was expressed at the same levels in HEK293T cells and was purified in the same way, resulting in equivalent amounts of protein, indicating that the protein is equally stable in mammalian cells and in solution. However, when assayed for Arl2 GAP activity, the mutant was inactive (Figure 8). Although we observed small amounts of activity over our no protein control, this level of activity seen for the mutant was not different from that observed with GST alone (Figure 8). These activities are so low as to be at or near the lower limits of our assay. Thus, we can safely conclude that the point mutant has at least a 10-fold lower specific activity than the wild-type protein, but it might be completely inactive as an Arl2 GAP.
Our study revealed that ELMOD3 is important for hearing in humans as a missense mutation in the gene leads to profound hearing impairment. ELMOD3 belongs to the engulfment and cell motility (ELMO) protein family, which includes six known members in mammals (ELMO1-3 and ELMOD1-3). Our ex vivo studies reveal that fluorescently tagged ELMOD3 localized with the actin-based microvilli of LLC-PK1-CL4 epithelial cells, in the stereocilia of sensory hair cells of mouse organ of Corti explants, and to a lesser extent to the actin cytoskeleton of MDCK cells, whereas the deafness-associated allele (p.Leu265Ser) was deficient in each case. Similarly, we show that human ELMOD3 possesses Arl2 GAP activity but the mutant has at least a 10-fold loss in activity.
While ELMOD3 antibody reactivity was detected in outer hair cells stereocilia at P02, more pronounced accumulation of ELMOD3 immunoreactivity was detected in rat cochlear inner hair cell stereocilia only by P12, which is when hair bundles are in the late phase of maturation. During this period, the inner hair cell stereocilia undergo a rapid elongation [30]. The observed staining suggests that ELMOD3 might be necessary for the initial development of the outer hair cell stereocilia or the organization of the bundle in a staircase pattern but may play a different role in the stereocilia of inner hair cells. Nevertheless, it is tempting to speculate that ELMOD3 may play a role in the maturation or maintenance of the cochlear stereociliary bundle. Recently, two spontaneous mutations (rda and rda2J) in mouse Elmod1 were shown to result in profound deafness and vestibular dysfunction [7], demonstrating that the function of ELMOD1 is essential for regulating the shape and maintenance of inner ear hair cell stereocilia in mice [7]. Elmod1 has been shown to be part of a large cluster of genes expressed in the developing inner ear, while Elmod3 level was below the detectable range [31]. These observations are consistent with findings from other studies, like the SHIELD database, which demonstrated that the level of Elmod3 mRNA is ∼100-fold lower than that of Elmod1 in the developing inner ear (P0–P1).
Besides stereocilia bundles, immunoreactivity for ELMOD3 was also detected along the kinocilium in developing cochlear hair cells. Recently, it has been shown that ELMO1 can act at the interface between the actin-cytoskeleton and microtubule network by interacting with ACF7 (Actin crosslinking family 7) [32]. Moreover, microtubule polymerization depends on Arl2 activity [33], and we have shown that ELMOD3 exhibits a GAP activity against Arl2. Therefore, ELMOD3 expression in the kinocilium might have a role in the assembly of the kinocilium architecture and in pathways regulating planar cell polarity.
The ELMO family proteins are functionally poorly characterized, and more information is currently available for the ELMOs than for the ELMODs, with no structural information available for any of them. So far, only one activity has been ascribed to ELMOD proteins: we previously reported that recombinant human ELMOD1 and ELMOD2 display in vitro Arl2 GAP activity, whereas ELMOD3 and bacterially expressed ELMO1-3 did not [29]. More recently, we performed additional phylogenetic and functional analyses of the ELMO domain that led us to re-examine whether ELMOD3 shares the Arl2 GAP activity of ELMOD1 and ELMOD2. Our data contrast with the earlier-published claim that ELMOD3 lacks Arl2 GAP activity; we determined that it does indeed possess this activity, albeit at a substantially lower specific activity than that of its two closest human paralogs. The large differences in specific activities observed in the Arl2 GAP assay may be due to differences in the specificities of ELMODs as GAPs for different GTPases, including the lack of one or more binding partners (e.g., one that is perhaps analogous to Dock180 binding to ELMO1 or a co-activator for activity such as has been proposed for COP-I and ArfGAP1 [34]), and/or the lack of post-translational modification.
Our in vitro experiment revealed that ELMOD3 harboring the p.Leu265Ser mutation, unlike ELMOD3, has no or few GAP activity against Arl2. The lack of Arl2 GAP activity of the mutant may suggest a reduced affinity for the Arl2 GTPase, which may play an important role in ELMOD3 localization. Elmod3 and Arl2 are expressed in developing mouse cochlear tissues and weakly in vestibular tissues (Figure S10). Even though ELMOD3 is active and has been defined as “Arl2 GAPs”, we expect it to be active against other GTPases in the Arf family as well. We therefore speculate that ELMOD3 functions as a GAP for Arl2 and perhaps other GTPases that participate in actin organization, polymerization or depolymerization in the cochlear hair bundles. If ELMOD3 is an active GAP for other GTPases, these GTPases are likely to be part of the Arf family given that GAPs are not known to cross family boundaries within the Ras superfamily. However, it is plausible that ELMOD3 functions in a signaling pathway that includes Arl2 (or an Arf family GTPase), Rac, Rho, and, ultimately, affects the actin cytoskeleton.
Future studies will address the specific roles of ELMOD3 in the development of the inner ear sensory epithelium, cytoskeletal organization, and ELMOD3-mediated signaling pathways. Revealing the interacting partners, substrate specificities for its GAP activities, as well as the means of regulation of ELMOD3 and other ELMO family proteins, will shed light on the overlapping functions of the Ras superfamily in the inner ear. These fundamental functions of this unique protein family are likely to be important in all eukaryotic cells.
Family PKDF468 was enrolled in the present study from the Punjab province of Pakistan, and written informed consent was obtained from all participating family members. The Institutional Review Boards at the Center for Excellence in Molecular Biology (Pakistan), at the National Institute on Deafness and Other Communication Disorders, and at Cincinnati Children's Hospital (USA) approved the present study. Hearing loss in the affected family members was evaluated using pure-tone audiometry, which tested frequencies that ranged from 125 Hz to 8 kHz. The family medical history stated that the onset of hearing loss was pre-lingual, and we observed no evidence of vestibular dysfunction or other balance issues using the Romberg and tandem gait tests. There were no other significant findings from the clinical exam, and the affected members had basic metabolic panel results within the normal range, indicating that they had nonsyndromic hearing loss.
We conducted a genome-wide scan on family PKDF468 using 388 STR markers and performed linkage analysis using GeneMapper software (Applied Biosystems; Carlsbad, CA). The LOD score was calculated using a recessive model of inheritance assuming a fully penetrant disorder and a disease allele frequency of 0.001. The primers were designed with Primer3 to sequence all of the coding exons and 75 bp of the exon-intron boundaries of all of the known genes within the DFNB88 locus (Table S2). The products were amplified using either Taq polymerase (Genscript; Piscataway, NJ) or Amplitaq Gold 360 (Applied Biosystems) for the GC-rich regions. The chromatograms were read using SeqMan software (DNAStar; Madison, WI).
Exome sequencing was conducted on one affected individual from family PKDF468 and was enriched using the Nimblegen SeqCap EZ Exome v2.0 Library (Roche Diagnostics; San Francisco, CA). One hundred base pair paired-end sequencing was performed on an Illumina Hi-Seq 2000 system. The sequencing data were analyzed following the guidelines that are outlined in the Broad Institute's Genome Analysis Toolkit [35], [36]. The row data were mapped using the Burrows Wheeler Aligner [36], the variants were called using the Unified Genotyper, and the data underwent further processing and quality control [35], [36]. Low-quality reads (less than 10× coverage) were removed, and the remaining variants were filtered against the dbSNP133 database and all of the known variants in the NHLBI 6500 Exome Variant database that had a minor allele frequency (MAF) of greater than 0.05%. We also filtered out additional variants that were observed in six ethnically matched control exomes. Primers were designed, using Primer3, to screen the remaining candidate gene variants, and we performed segregation analysis by performing Sanger sequencing of the variants of all of the participating family members.
Human and mouse ELMOD3/Elmod3 isoform-specific primers and TaqMan probes were designed, using Primer3 web-based program, and the transcripts were amplified from human and mouse cDNA libraries (Clontech Laboratories; Mountain View, CA).
Mouse inner ear tissues were harvested from 3 or more C57BL/6J mice at P0, P10, and P30. The cochlea and vestibular system were separated from the inner ear, and the total RNA was extracted from each tissue using TRIreagent (Life Technologies, Grand Island, NY). The RNA was reverse-transcribed into cDNA using the SMARTscribe Kit (Clontech). Real-time PCR was performed in triplicate on a StepOne Plus instrument (Applied Biosystems). The data were analyzed using the comparative Ct method, with Gapdh as the endogenously expressed reference gene. RT-PCR was performed by using LA Taq (Clontech). The products were run on a 2% agarose gel that was stained with ethidium bromide and each isoform was verified by sequencing.
To determine the in vivo effect of c.794T>C allele, if any, on the splicing of ELMOD3 transcripts, total RNA was isolated from fresh blood samples of two affected individuals (V:2 and V:11) and one normal hearing individual (V:3) by use of TRIzol reagent (Life Technologies). Oligo dT and randomly primed first strand cDNA libraries were generated using SMART 1st strand cDNA synthesis kit (Clontech). Touchdown PCR was performed with GenScript Taq (GenScript) and 1.5 mM MgCl2 at an annealing temperature of 63°C for 30 cycles using ELMOD3-specific primer pairs with a common forward primer in exon 9 and reverse primer either in exons 11 (hELMOD3_ex9-11; Table S5) or in exon 12 (hELMOD3_ex9-12; Table S5). GAPDH was used as a control and amplified under the same conditions. PCR fragments were subcloned into pCR-TOPO cloning vector (Life Technologies), and the sequences were verified.
Human ELMOD3 B isoform, murine Elmod1, Elmod2 and Elmod3 isoforms a and b open reading frame have been amplified from commercially available human and mouse cDNA libraries (Clontech) and inserted in pEGFP-C2 vector (Clontech) to generate proteins with GFP fused to their N-termini. The construct encoding p.Leu265Ser ELMOD3 was prepared through site-directed mutagenesis (Agilent Technologies, Santa Clara, CA) using the wild-type ELMOD3 isoform B as a template. The full-length open reading frame of human ELMOD3 (isoform B) was cloned into the pCOLD-TF (Takara Bio, Inc.; Otsu Shiga, Japan) vector at the BamHI and SalI sites by PCR amplification of the cDNA using primers that inserted the appropriate restriction sites. This generated a fusion protein with a His6 tag at the N-terminus. This tag was followed by trigger factor (∼48 kDa), a thrombin cleavage site, and the ELMOD3 open reading frame. To insert the Leu265Ser mutation, we performed site-directed mutagenesis on the construct using the QuikChange Lightening Kit (Agilent Technologies). Full-length open reading frames of human ELMOD3 (Isoform B) and the Leu265Ser mutant were cloned into the pLEXm-GST vector [37] using KpnI and SphI sites that were inserted into the PCR primers, with subsequent confirmation of the correct DNA sequence. The parent vector was used to express GST alone.
Inner ear explants were harvested from C57BL/6J mice at P2. The explants were cultured in a glass-bottom Petri dish (MatTek, Ashland, MA) that was coated with Matrigel (BD Biosciences, San Jose, CA) and were maintained in DMEM that was supplemented with 7% fetal bovine serum (FBS) (Life Technologies) for 24 hrs at 37°C with 5% CO2. The cultures were transfected using a Helios gene gun (Bio-Rad, Hercules, CA), as described elsewhere [38].
HEK293T, CL4 and MDCK cells were grown in DMEM that was supplemented with 10% FBS, 2 mM L-glutamine, and penicillin/streptomycin (50 U/ml) (Life Technologies) and were maintained at 37°C in 5% CO2. The cells were transfected using Fugene HD Transfection Reagent (Promega; Sunnyvale, CA), according to the manufacturer's instructions. The cells were then cultured for an additional 48 hrs prior to immunostaining. Forty-eight hrs following transfection, we added 2.5 µM Cytochalasin D (EMD Millipore, Billerica, MA) in fresh DMSO medium to the cells for two hrs to disrupt the actin cytoskeleton. The Cytochalasin D was then washed out, and the cells were grown for four additional hrs in complete medium.
ELMOD3 antiserum was raised in rabbits against two synthetic mouse ELMOD3-specific peptides (corresponding to residues 143–156 and 346–361 of the mouse sequence [GenBank accession number GI:358679299]). The immunizations and sera collections were performed by Covance (Princeton, New Jersey). The antiserum was affinity purified (AminoLink Plus Immobilization Kit; Thermo Scientific, Rockford, IL) either using both synthetic peptides in combination or individually.
Antibody specificity was assessed by transfections (Fugene HD Transfection Reagent; Promega) of GFP-tagged mouse ELMOD1, ELMOD2 and ELMOD3 into HEK293T cells followed by western blot analysis, as described elsewhere [39]. Inner ear and olfactory bulbs tissues were harvested from C57BL/6J mice at P30 and followed by western blot analysis, as described elsewhere [39]. Antigen competition was performed by incubating the primary antibody for 30 min at room temperature with the two immunizing peptides prior to use in Western Blot or immunofluorescence analyses.
C57BL/6J mice were obtained from Jackson Laboratories (Bar Harbor, ME) and bred in house. Sprague–Dawley rats were purchased from Charles River Breeding Laboratories (Raleigh, North Carolina) and bred in house. All experiments and procedures were approved by the Institutional Animal Care and Use Committee of the Cincinnati Children's Hospital Medical Center.
The inner ears from rats and mice were fixed with 4% PFA at 4°C overnight. P12 and P14 rat cochlea were incubated for one day, at 4°C, in 0.25M EDTA. The sensory epithelia were dissected in PBS. Following permeabilization with 0.25% Triton X-100 for 45 min, the samples were incubated in blocking solution (5% normal goat serum in PBS). The samples were then incubated overnight at 4°C with primary antibody (anti-ELMOD3; anti acetylated Tubulin (Sigma-Aldrich, St Louis, MO)) in 3% NGS/PBS. This step was followed by three washes in PBS and consecutive incubation with Alexa-488 conjugated secondary antibody and Alexa-647 conjugated secondary antibody (Life Technologies) at 1∶500 dilution and with rhodamine phalloidin (Life Technologies) at 1∶200 dilution in 3% NGS/PBS for one hour. After three washes with PBS, the samples were mounted using ProLong Gold Antifade Reagent (Life Technologies).
Transfected cells and inner ear explants were washed with PBS and fixed for 20 min in 4% paraformaldehyde. Filamentous actin was detected with rhodamine phalloidin (Life Technologies) in PBS/0.1%Triton-X100 for one hour. Following subsequent washes, the coverslips were mounted using FluoroGel medium (Electron Microscopy Sciences, Hatfield, PA) for the cell monolayers or with ProLong Gold Antifade Reagent (Life Technologies) for the explant cultures.
MDCK cells were washed with PBS, fixed for 20 min in 4% paraformaldehyde and blocked with 10%NGS/PBS/0.1% Triton-X100. The cells were incubated overnight at 4°C with a primary antibody (ZO1, Life Technologies; anti-ELMOD3) in 3% NGS/PBS/0.1%Triton-X100. The cells were then washed in PBS/0.1%Triton-X100 and incubated with the Alexa-fluor 546 conjugated secondary antibody (Life Technologies) in 3% NGS/PBS/0.1% Triton-X100 for one hour at room temperature. Filamentous actin was detected with Alexa647-phalloidin (Life Technologies) in PBS/0.1%Triton-X100 for one hr. Following subsequent washes, the coverslips were placed using FluoroGel mounting medium (Electron Microscopy Sciences).
All images were acquired using a Zeiss LSM 700 scanning confocal microscope that was equipped with 63× and 100× objectives, and the analyses were performed using ImageJ software. The pixel intensity analyses were performed using ImageJ software on images that were acquired with the same microscope settings. The statistical analyses were performed using GraphPad Prism software and the ANOVA test function.
Transfected MDCK cells were washed with PBS and fixed for 15 min in 4% paraformaldehyde. Following a one hour incubation in blocking solution (10% normal goat serum in PBS with 0.1% Triton X-100), the cells were then incubated one hour with primary antibody: polyclonal chicken anti-GFP (Aves Labs Inc., Tigard, OR) custom conjugated with APEX Alexa Fluor 568 Antibody (Life Technelogies), in PBS/0.1%Triton-X100/3%NGS. Filamentous actin was labeled concomitantly with Alexa647-phalloidin (Life Technologies). After extensive washes in PBS/0.1%TritonX-100/3%NGS, the cells were fixed 15 min in 4% paraformaldehyde and stored at 4°C in PBS.
For imaging, PBS was replaced by the following imaging medium: 2-mercaptoethanol, buffer B (10% glucose, 50 mM Tris-HCl pH = 8, 10 mM NaCl) and the GLOX system (14 mg Glucose Oxidase, 50 ul Catalase, 200 ul Tris buffer) in a 1∶100∶1 volume ratio. N-STORM imaging was performed with a Nikon N-STORM super-resolution microscope system (Nikon Instruments Ltd, Melville, NY) based on an inverted microscope Nikon A1Rsi equipped with a perfect-focusing system and a 100× TIRF APO NA 1.49 oil objective. A 561 nm wavelength laser was applied for bleaching and excitation of Alexa Fluor 568 while a 647 nm wavelength laser was applied for Alexa Fluor 647. The images were acquired with an Andor Xion 897 EMCCD camera using 16 ms exposition with one frame of imaging for 5000 cycles. The analysis was performed with Nikon Elements Storm software.
BL21 (DE3) Gold E. coli were transfected with the pCOLD-TF-ELMOD3 and pCOLD-TF-ELMOD3 Leu265Ser plasmids, and single colonies were used to inoculate cultures in LB medium with 100 µg/ml ampicillin, which were grown at 37°C with shaking. When OD600 = 0.5, the cells were moved to 15°C for 30 min without shaking. IPTG was then added to 1 mM, and the culture was grown overnight at 15°C with shaking. The cells were collected and lysed by passage through a French press in 20 mL of 20 mM HEPES (pH 7.5), 150 mM NaCl, and 5 mM imidazole with a protease inhibitor cocktail (Sigma-Aldrich). The lysates were clarified by centrifugation at 100,000× g for one hour at 4°C, and the supernatants were loaded onto a Ni-NTA column (GE Healthcare) that was pre-equilibrated in the same buffer. The column was washed with 20 mL of 20 mM HEPES (pH 7.5), 150 mM NaCl, and 55 mM imidazole prior to elution in 15 mL buffer that contained 250 mM imidazole. The protein was further purified and buffer-exchanged by gel filtration chromatography using a Superdex S200 column (24 mL) that was run in 20 mM HEPES (pH 7.5), 150 mM NaCl, and 1 mM dithiothreitol. Typical yields from this protocol were ∼10 mg/L TF-ELMOD3 that was >90% pure, as estimated by visual inspection of a Coomassie blue-stained gel. The actin binding experiment was performed with purified TF-ELMOD3 and TF-ELMOD3 Leu265Ser using an Actin Binding Protein Biochem Kit (Cytoskeleton Inc, Denver, CO), following the manufacturer's protocol and later repeated with the GST-ELMOD3 proteins purified from HEK cells.
GST-ELMOD3, the point mutant, or GST alone was expressed in HEK293T cells using the pLEXm-GST vector (a kind gift from Dr. James Hurley (NIDDK)) and a modification of the method that was described in Aricescu et al [37]. Briefly, HEK293T cells (10×10 cm plates) were transfected at 90% confluency with 1 µg/mL DNA after mixing with polyethyleneimine (PEI-MAX; Polysciences, Inc.; Warrington, PA) at a 1∶3 ratio of DNA∶PEI in Opti-MEM medium (Life Technologies). The mixture was then added to cells in DMEM medium that contained 2% FBS and grown for two days. The cells were collected by centrifugation and lysed by resuspension in 1.5 ml 25 mM HEPES (pH 7.4), 100 mM NaCl, and 1% CHAPS. The solution was clarified by centrifugation for 30 min in a microfuge at a maximum speed (∼14,000× g) at 4°C. Glutathione Sepharose 4B (GE Healthcare) beads were added and incubated at 4°C for 3 hrs with mixing. The beads were then pelleted, washed twice in 25 mM HEPES (pH 7.4) and 100 mM NaCl and eluted (2×0.5 mL) in the same buffer containing 20 mM glutathione. The eluted protein was concentrated to 0.25 mL in a spin concentrator (Amicon Ultra-4; EMD Millipore). The protein concentration was determined using a Bradford assay. Typical yields of preparations from 10×10 cm plates were ∼600 µg of GST-ELMOD3 or mutant and ∼6 mg of GST. The purified proteins were quick frozen and stored at −80°C.
The Arl2 GAP assay was performed as described previously by Bowzard et al [29]. Briefly, 2 µM purified recombinant Arl2, prepared as described by Clark et al [40], was pre-loaded with [γ-32P]GTP in 25 mM HEPES (pH 7.4), 2.5 mM MgCl2, 100 mM NaCl, 1 mM EDTA, 25 mM KCl, and 0.5 mM ATP in a total volume of 100 µL. The incubation was performed at 30°C for 30 min. The GAP reaction was performed in a buffer that contained 25 mM HEPES (pH 7.4), 2.5 mM MgCl2, 100 mM NaCl, 1 mM dithiothreitol, 2 mM ATP, 1 mM GTP and the pre-loaded Arl2 [γ-32P]GTP. The total 50 µL reaction was initiated by the addition of the sample that contained the Arl2 GAP and stopped after 4 min at 30°C by the addition of 750 µL of ice-cold activated charcoal (5% activated charcoal (Sigma-Aldrich, St. Louis, MO) in 50 mM Na2HPO4 (pH 7.4). The samples were clarified by centrifugation, and 400 µL was taken for counting in a liquid scintillation counter. Each GAP sample was also assayed in parallel in a tube that contained all of the above reagents except Arl2 (i.e., the same amount of [γ-32P]GTP). The resulting “blank” values were subtracted from the results that were obtained in the presence of Arl2 to determine the amount of hydrolyzed 32Pi that was dependent on Arl2 GAP activity. More detail and descriptions regarding how the specific activities were calculated from this assay can be found in the report by Bowzard et al [29].
|
10.1371/journal.pgen.1002811 | A Key Role for Chd1 in Histone H3 Dynamics at the 3′ Ends of Long Genes in Yeast | Chd proteins are ATP–dependent chromatin remodeling enzymes implicated in biological functions from transcriptional elongation to control of pluripotency. Previous studies of the Chd1 subclass of these proteins have implicated them in diverse roles in gene expression including functions during initiation, elongation, and termination. Furthermore, some evidence has suggested a role for Chd1 in replication-independent histone exchange or assembly. Here, we examine roles of Chd1 in replication-independent dynamics of histone H3 in both Drosophila and yeast. We find evidence of a role for Chd1 in H3 dynamics in both organisms. Using genome-wide ChIP-on-chip analysis, we find that Chd1 influences histone turnover at the 5′ and 3′ ends of genes, accelerating H3 replacement at the 5′ ends of genes while protecting the 3′ ends of genes from excessive H3 turnover. Although consistent with a direct role for Chd1 in exchange, these results may indicate that Chd1 stabilizes nucleosomes perturbed by transcription. Curiously, we observe a strong effect of gene length on Chd1's effects on H3 turnover. Finally, we show that Chd1 also affects histone modification patterns over genes, likely as a consequence of its effects on histone replacement. Taken together, our results emphasize a role for Chd1 in histone replacement in both budding yeast and Drosophila melanogaster, and surprisingly they show that the major effects of Chd1 on turnover occur at the 3′ ends of genes.
| Nucleosomes prevent transcription by interfering with transcription factor binding at the beginning of genes and blocking elongating RNA polymerase II across the bodies of genes. To overcome this repression, regulatory proteins move, remove, or structurally alter nucleosomes, allowing the transcription machinery access to gene sequences. Over the body of a gene, it is important that nucleosome structure be restored after a polymerase has passed by; failure to do so may lead to activation of transcription from internal gene sequences. Interestingly, although nucleosomes constantly move on and off of promoters, they are relatively stable over the bodies of genes. Thus, the same nucleosomes that are removed to allow a polymerase to pass by must be reassembled in its wake. Here, we examine the role of an ATP–dependent chromatin remodeling protein, Chd1, in regulating nucleosome dynamics. We find that Chd1 is important for exchange of the histone H3 in both yeast and Drosophila and that, surprisingly, while it promotes exchange of histones at the beginning of genes, it prevents exchange at the ends of genes. Finally, we show that Chd1 helps determine the characteristic pattern of chemical modifications of histone H3 found over actively transcribed gene sequences.
| Eukaryotic genomes are packaged as chromatin, whose fundamental repeating subunit, the nucleosome, is composed of 147 bp of DNA wrapped 1.7 times around an octameric histone core. Nucleosomes may interact with each other to form higher-order levels of chromatin packaging necessary to compact an entire genome within a nucleus. This genome packaging strategy leads to a dominant theme in eukaryotic gene regulation: nucleosomes tend to repress gene expression, and a large array of gene regulatory mechanisms in eukaryotes operate by strengthening or weakening the repressive effects of nucleosomes on gene expression [1].
Genome-wide nucleosome mapping studies indicate that although the majority of a eukaryotic genome is typically covered with regularly spaced nucleosomes, nucleosome depleted or nucleosome free regions are frequently found over promoters and at the 3′ ends of genes (reviewed in [2]). Although these studies give a fixed snapshot of chromatin organization, other analyses indicate that chromatin is dynamic. Studies in which histones were pulse-labeled with radioisotopes or tagged with GFP demonstrated that histones can be actively exchanged on chromatin, even in the absence of DNA replication [3], [4]. More recent work has utilized induction of epitope-tagged alleles of histones in G1-arrested yeast cells followed by chromatin immunoprecipitation to examine histone H3 dynamics genome-wide [5], [6]. These studies show that histone H3 exchanges at a high rate on promoters and in other intergenic regions such as downstream of the 3′ ends of genes. With the exception of highly-transcribed genes, the bodies of genes, even those that are transcribed at moderate rates, exhibit much lower H3 exchange rates.
Although nucleosomes over transcribed genes appear to be relatively stable in vivo, nucleosomes form a strong barrier to elongating RNA polymerase II (RNA Pol2) in vitro [7]. Thus, it is likely that accessory factors assist in transcription elongation to alleviate this barrier. These factors may promote the temporary disassembly or displacement of nucleosomes permitting the passage of elongating RNA Pol2, and furthermore, they may assist in nucleosome (re)assembly after polymerases have passed. A wide variety of factors have been implicated in the dynamics and maintenance of chromatin structure over transcribed sequences. These include ATP-dependent chromatin remodeling enzymes, enzymes that post-translationally modify histones, histone chaperones and transcription elongation factors [8]. Interestingly, mutations affecting a number of these factors cause a cryptic transcription initiation phenotype, in which disruption of chromatin in the body of genes leads to activation of internal, normally quiescent promoters [9].
One factor implicated in the regulation of transcribed chromatin is the ATP-dependent chromatin remodeling enzyme Chd1. Chd1 is the founding member of a family of highly conserved chromatin remodeling enzymes found throughout eukaryotes [10]. Although budding yeast only express a single Chd1 protein, at least 9 CHD family proteins are expressed in humans. Mammalian CHD family members have been implicated in diverse roles including promotion of normal organismal development, and the maintenance of pluripotency and prevention of heterochromatin formation in mouse embryonic stem cells [10]. In addition, mutations in CHD protein genes are implicated in several human cancers and CHARGE syndrome, which is characterized by a phenotypically heterogeneous set of developmental defects [10], [11].
CHD proteins typically have a pair of N-terminal chromodomains, a central Snf2/Swi2 type helicase domain and a C-terminal domain that mediates DNA or nucleosome binding [10]. The chromodomains of human Chd1 bind histone H3 tails methylated at lysine 4 (H3K4me) suggesting a mechanism for recruitment [12], [13]. However, yeast Chd1 does not bind H3K4-methylated tails [13], and in Drosophila melanogaster, the chromodomains do not play an important role in its localization to chromatin [14]. Recent structural and biochemical studies suggest that rather than mediating chromatin localization, the chromodomains may regulate enzyme activity [15]. In vitro assays show that Chd1 has the ability to assemble, remodel, slide and promote regular spacing of nucleosomes [16]–[18]. Chromatin immunoprecipitation in budding and fission yeast, and immunostaining of Drosophila polytene chromosomes show that Chd1 associates with both promoters and transcribed regions of active genes [19]–[23]. Consistent with its localization on genes, genetic studies in yeast have implicated Chd1 in the regulation of transcription initiation, elongation and termination [22], [24]–[28]. Although Chd1 can be purified as a monomer, its association with several complexes that regulate initiation and elongation, which include mediator, FACT, the Paf1 complex, SAGA and SLIK, provides further support to these conclusions [22], [29]–[33]. Chd1 also associates with histone chaperones Nap1 in fission yeast, and HirA, a histone chaperone for histone H3.3, in fruit flies [19], [34].
Several studies suggest mechanisms for how Chd1's biochemical activity may relate to these biological functions. Chd1 can promote transcription and catalyze activator dependent, promoter specific nucleosome remodeling in vitro [35], [36]. Furthermore, in Schizosaccharomyces pombe, Chd1 (Hrp1) acts at a subset of promoters to disassemble nucleosomes close to the transcription initiation site [19]. In Drosophila, following fertilization of an egg, sperm chromatin is decondensed, protamines are removed and replaced with nucleosomes whose only form of histone H3 is the replication-independent variant H3.3 [37]. Interestingly, in chd1 mutants, H3.3 levels in decondensing sperm chromatin are greatly reduced and unevenly distributed, suggesting a role for Chd1 in the replication-independent assembly or distribution of H3.3 nucleosomes [34], [38].
A recent high-resolution genome-wide nucleosome mapping study in budding yeast points to an in vivo role for Chd1's nucleosome remodeling activity. Nucleosomes are typically regularly positioned over genes in wild type yeast cells [39]. However, in a chd1Δ mutant, this positioning is largely lost over gene bodies [40]. Specifically, nucleosome free regions at the 5′ and 3′ ends of genes and the first (+1) nucleosome over the transcribed region were minimally affected by loss of Chd1, but downstream nucleosomes (particularly those starting at the +3 position) were dramatically delocalized in chd1Δ yeast cells. Curiously, micrococcal nuclease digestion patterns of bulk chromatin are not affected in a chd1 mutant, suggesting that Chd1 affects the positioning of nucleosome arrays primarily over the transcribed body of genes, rather that the precise spacing between any given pair of nucleosomes [40], [41]. Although chd1 mutations have modest effects on gene expression in yeast, and are virtually indistinguishable from wild type strains in phenotypic assays, they do cause a cryptic initiation phenotype, consistent with the loss of nucleosome organization over the body of genes [9], [28], [42], [43].
Although these data clearly demonstrate a role for Chd1 in nucleosome positioning in vivo, the mechanism underlying its in vivo function and its relationship to transcription remains unclear. In this study, we examine the role(s) of Chd1 in governing the replication-independent exchange of newly-expressed histone H3 onto chromatin in budding yeast and Drosophila using genome-wide methodologies. Chd1 mutants have dramatic defects in the localization of the replication-independent histone variant H3.3 in flies, while in Saccharomyces cerevisiae, chd1Δ mutants exhibit dramatic defects in H3 turnover in coding regions. Surprisingly, Chd1 predominantly affects histone H3 exchange at the 3′ ends of coding regions, and this effect on turnover depends on gene length – H3 turnover at 3′ ends is fairly concordant between wild type and chd1Δ strains for genes 1 kb and shorter, whereas Chd1 appears to specifically stabilize nucleosomes over the 3′ ends of longer genes. Finally, we show that loss of Chd1 globally alters histone modification patterns related to active transcription, with H3K36me3 in particular shifting in concert with the changed patterns of H3 replacement. Together, our results show that Chd1 plays a key role in histone H3 dynamics, and surprisingly, that yeast Chd1's influence on H3 dynamics is most apparent at the 3′ ends of genes.
Previously, Fyodorov and colleagues examined the distribution of epitope-tagged, full length H3.3 in the Drosophila syncytial blastoderm and only observed a modest defect in H3.3 distribution in chd1 null mutants [34]. Because the H3.3 N-terminal tail, which is required for replication-dependent assembly of H3.3 [44], was intact in this experiment, we reasoned that any defect in replication-independent assembly of the tagged H3.3 might have been obscured.
To reassess Chd1's role in replication-independent deposition of H3.3, we imaged GFP-tagged histone H3.3core in live salivary glands from chd1 mutant larvae. We utilized a transgenic fly expressing an AB1-GAL4 driver and a Gal inducible histone H3.3core-GFP [44]. Because the H3.3core protein encoded by the transgene lacks the N-terminal tail, it is only incorporated into chromatin via the replication-independent pathway [44]. In an otherwise wild type background, H3.3core-GFP was deposited into the polytene chromosome arms of salivary glands (Figure 1A). In some cases, we also observed a nucleoplasmic GFP signal in which the entire nucleus, including non-chromosomal territories, exhibited a strong GFP signal, although a chromosomal banding pattern was still evident. In flies that were heterozygous or homozygous for chd15, a null allele of Chd1 [21], we observed salivary gland nuclei with GFP signals similar to those of wild type, i.e. chromosomal or broad nucleoplasmic GFP fluorescence. However, we also observed nuclei with a novel, “non-chromosomal” phenotype where the polytene arms appear almost devoid of GFP signal and a substantial nuclear, non-chromosomal H3.3core-GFP signal was still apparent (Figure 1B, 1C). We determined the relative frequencies of these phenotypes in wild type and mutant flies by blind scoring, and observed that the predominant chromosomal fluorescence pattern observed in wild type cells declined dramatically in chd1 mutants, whereas the nucleoplasmic and non-chromosomal patterns increased in frequency (Figure 1D). We observed similar phenotypes when we repeated these experiments with independently derived chd15 flies using a different balancer chromosome (data not shown). These results do not appear to be due to any peculiarity of the AB1-GAL4 driver as we observed similar fluorescence patterns when we used sgsGAL4 and eyelessGAL4 drivers (data not shown). Furthermore, we did not observe obvious differences in the strength of H3.3core-GFP signals between flies with the three analyzed genotypes (wild type, +/chd15 heterozygous and chd15/chd15 homozygous), nor between nuclei with the three observed staining patterns (chromosomal, non-chromosomal and nucleoplasmic) (Figure S1), suggesting that the observed localization patterns were not due to differences in H3.3core-GFP expression. Rather, we favor the idea that the variability observed here reflects perdurance of maternally contributed Chd1, which has been observed previously [21]. Immunostaining of fixed polytene chromosomes similarly revealed a reduction of H3.3core-GFP on chromosomes derived from chd15 mutant larvae, while levels of full length H3.3-GFP were not affected by loss of Chd1 (Figure S2), consistent with the ability of full length H3.3 to incorporate through both replication-dependent and –independent pathways. Consistent with our observations in the chd15 mutants, we observed decreased association of H3.3core-GFP with polytene chromosomes when we knocked down Chd1 levels with either of two RNAi constructs (Figure S2 and data not shown). Overall, these data are consistent with the possibility that Chd1 may contribute to replication-independent assembly of H3.3 containing nucleosomes.
To further examine roles of Chd1 in nucleosome dynamics in vivo, we turned to budding yeast. To test the idea that Chd1 may modulate replication-independent nucleosome assembly or dynamics, we took advantage of the observation that the yeast H3 N-terminal tail is important for normal chromatin structure [45]. Reasoning that the H3 N-terminal tail deletion mutation likely interferes with replication-dependent assembly of H3, as is the case in Drosophila and Physarum polycephalum, [44], [46], we predicted that loss of this function would sensitize cells to defects in other chromatin assembly or maintenance pathways, we used a plasmid shuffle strategy to create CHD1+ and chd1Δ yeast strains expressing either wild type histone H3 (H3WT) or a histone H3 N-terminal deletion mutation, H3Δ4-30. Consistent with prior observations, the chd1Δ H3WT strain grew indistinguishably from wild type cells, and the CHD1 H3Δ4-30 strain exhibited a moderate growth defect (Figure 2). Interestingly, the chd1Δ H3Δ4-30 double mutant grew much more poorly than the CHD1 H3Δ4-30 single mutant, indicating that Chd1 and the N-terminal tail of H3 share a redundant function.
In contrast to other model organisms, the budding yeast genome expresses only a single non-centromeric form of histone H3. However, the major histone H3/H4 chaperones, including the H3.3 chaperone HirA, are conserved, suggesting that yeast retain distinctive replication dependent and independent chromatin assembly pathways [47]. We have obtained data consistent with this idea in a screen for genetic suppressors of a cold-sensitive allele of transcription elongation factor SPT5. Among these suppressors were mutations in CHD1, mutations in the H3K4 and H3K36 histone methyltransferases SET1 and SET2, histone H3K4 and H3K36 substitutions, and mutations in members of the RPD3S histone deacetylase complex. Further characterization of these suppressors led us to propose that they act by lowering the chromatin barrier to efficient transcription elongation [28].
Given the observations described above, we recently screened a randomly mutagenized plasmid library for histone H3 mutations that suppress spt5Cs- (to be described in detail elsewhere). Among the suppressor mutations obtained in that screen, we isolated a mutation, H3-S87P/G90S, which simultaneously alters two of the four residues that distinguish histone H3.1 from H3.3 in other eukaryotes. Yeast expressing the S87P/G90S form of histone H3 from the normal HHT2 locus are viable, indicating that this mutation is unlikely to strongly perturb replication coupled chromatin assembly. As with several other of the mutations that suppress spt5Cs- (e.g., H3K36R, set2, mutations affecting Rpd3s), the H3-S87P/G90S mutant caused cryptic initiation of transcription (Figure S3).
We therefore examined genetic interactions between the H3-S87P/G90P, chd1Δ and the H3Δ4-30 mutations using the plasmid shuffle assay described above (Figure 2). Interestingly, the chd1Δ H3-S87P/G90S double mutant exhibited no new mutant phenotypes, whereas combining H3-S87P/G90S with the H3Δ4-30 deletion resulted in a very poor growth phenotype and the chd1Δ H3-S87P/G90S H3Δ4-30 triple mutation showed an even more severe growth defect. Thus, like Chd1, residues 87 and 90 of histone H3 function redundantly with the H3 N-terminal tail. It is tempting to argue that these data indicate that Chd1 interacts with histone H3 via a surface defined by residues 87 and 90. However, the fact that the phenotype of the chd1Δ H3-S87P/G90S H3Δ4-30 triple mutant is more severe than that of the chd1Δ H3Δ4-30 double mutant suggests that H3 residues S87 and G90 may retain functions that are redundant with the H3 tail, even when Chd1 is absent.
The data presented above suggest that Chd1 affects replication independent dynamics of histone H3. To test this idea directly in budding yeast, we used a yeast strain carrying galactose-inducible Flag-tagged H3, coupled with chromatin immunoprecipitation and tiling microarray (ChIP on chip) analysis, to follow the incorporation of newly-synthesized H3 genome-wide in cells arrested in the cell cycle [6]. Briefly, wild type or chd1Δ yeast strains are arrested in G1 phase using alpha factor, then Flag-H3 is induced with galactose, and after 60 minutes Flag-H3 and total H3-associated DNA are subject to ChIP enrichment and competitively hybridized on ∼250 bp resolution tiling microarrays. Resulting Flag/total H3 ratios provide locus-specific estimates of H3 turnover rates.
Figure 3A shows a “metagene” analysis of H3 turnover in 3 biological replicate samples for wild type (blue) and chd1Δ (red) strains. The wild type profile recapitulates previous results from multiple labs [5], [6], [48] – H3 replacement is highest over promoters and at the 5′ ends of genes, with coding regions being remarkably protected from H3 replacement, and modest levels of turnover being seen at the 3′ ends of genes.
Conversely, chd1Δ mutants exhibit H3 turnover patterns in which genes appear to effectively reverse polarity. Turnover is still lowest over coding regions, but the trough of minimal turnover has shifted 5′ along coding regions. Promoter and 5′ turnover are slower in chd1Δ cells, whereas maximal H3 replacement is instead observed at the 3′ ends of genes. This behavior is highly unusual, as several published [5], [6], [49], [50] and a large number of unpublished (OJR, unpublished data) mutants exhibit quite distinct turnover defects. We confirmed the increased 3′ H3 replacement at two model genes (Figure S4) using an entirely independent assay for histone replacement based on Cre-mediated recombination of C-terminal H3 epitope tags [51]–[53].
As a separate visualization, Figure 3B shows the average H3 turnover for various classes of genomic elements [6], [54]. Even though a previous microarray analysis showed only a very modest effect of chd1Δ on transcription [18], we considered the possibility that the altered H3 turnover in chd1 cells could be due to a large shift in cellular transcription. However, we observed strong concordance of ChIP on chip of RNA Pol2 signals for wild type and chd1Δ cells (Figure S5). Moreover, as noted below, Chd1's effects on H3 replacement are strongly gene length-dependent, but we find no correlation between mRNA abundance changes and gene length or transcription frequency (Figure S6). Thus, Chd1's effects on turnover are not secondary effects of altered transcription.
We sought to understand what factors might contribute to Chd1 recruitment or function at gene ends. To this end, we first examined the genes with the greatest changes in H3 replacement at their 3′ ends in chd1Δ mutants. Notably, we observed that the genes with the greatest changes in 3′ end H3 turnover were among the longest (>3 kb) genes in budding yeast. We therefore systematically analyzed the effects of gene length on Chd1's role in H3 replacement.
Figure 4 shows H3 turnover levels for wild type and chd1Δ yeast cells at gene ends (the first and last 500 bp of coding regions) as a function of gene length. At both gene ends there is strong length dependence for H3 turnover in wild type yeast cells, with turnover decreasing as a function of gene length. Notably, for both 5′ end and 3′ end H3 turnover, Chd1's effect on H3 turnover was greatest at unusually long genes. In addition, we found that Chd1's effect on 3′ turnover was greater at highly transcribed genes (Figure S7).
The length dependence for 3′ end H3 replacement (Figure 4B) is particularly remarkable – H3 turnover is nearly identical in wild type and chd1Δ strains for genes of up to roughly 1 kb in length, at which point 3′ end turnover continues to decrease with gene length in wild type cells but stays essentially constant in chd1Δ cells. In other words, the role of Chd1 in wild type cells seems to be to help stabilize nucleosomes at the 3′ ends of genes over 1 kb in length.
Chd1's effects on H3 turnover are greatest at genomic loci that are enriched in H3K36me3 or H3K4me3 modified nucleosomes [55], and chd1 mutants exhibit synthetic genetic interactions with the H3K4 and H3K36 methyltransferases Set1 and Set2 [56], [57]. We therefore determined if chd1Δ mutants affect histone modification patterns by genome-wide mapping of H3K4me3 and H3K36me3 in wild type and chd1Δ yeast cells. Crosslinked chromatin from these two strains was digested with micrococcal nuclease, immunoprecipitated with H3K4me3 or H3K36me3 antisera and competitively hybridized to microarrays with micrococcal nuclease digested input DNA.
Figure 5 shows average H3K4me3 and H3K36me3 patterns in chd1Δ cells. On average, H3K4me3 patterns were minimally affected by loss of Chd1, although we noticed a subtle increase in H3K4me3 at the 3′ ends of many genes. This may be a consequence of the fact that chd1Δ mutants show increased transcription from “cryptic” internal promoters [9],[28]. Interestingly, the gain in H3K4me3 at the 3′ ends of genes was greatest at longer genes (Figure S8), which also exhibited the greatest defects in H3 turnover.
More dramatically, H3K36me3 patterns were extensively altered in chd1Δ cells, with loss of H3K36me3 at the 3′ ends of genes and a shift in the H3K36me3 peak towards the 5′ ends of genes. Consistent with the loss of H3K36me3 at the 3′ ends of genes, we previously observed increased H3K9/K14 acetylation at the 3′ ends of several genes in a chd1Δ mutant [28], as would be expected since reduced H3K36me3 results in reduced recruitment or activity of the Rpd3S deacetylase complex [27], [58], [59].
In our prior study, we did not observe any significant change in total levels of H3K4me3 or H3K36me3 in a chd1 mutant [28]. As H3K36me3 typically anticorrelates with H3 turnover [6], we hypothesize that the altered H3K36me3 profile observed here is a consequence of Chd1's effects on H3 turnover – increased H3 turnover at 3′ ends of genes likely results in loss of H3K36me3 at these regions. Consistent with this hypothesis, we found that loss of 3′ H3K36me3 was greatest at longer genes (Figure S8).
We present evidence that Chd1 modulates replication-independent turnover of histone H3 in both Drosophila and budding yeast. Chd1's effects on H3 turnover are greatest at genomic loci that normally coincide with peaks of H3K4me3 and H3K36me3 modified nucleosomes. This observation is consistent with prior reports that Chd1 reduces nucleosome density at promoters, can catalyze activator-dependent nucleosome removal and promote transcription in vitro, and that it modulates the efficiency of transcription termination [19], [24], [36].
Chd1's effects on H3 turnover may reflect a direct role in histone eviction or deposition during replication-independent histone exchange, consistent with its ability to catalyze ATP-dependent assembly of nucleosomes in vitro, or it could reflect a role for Chd1 in stabilization of pre-existing nucleosomes. Here, we observed that the predominant effect of Chd1 on H3 turnover in budding yeast was to repress turnover over the 3′ ends of genes. While we do not yet understand the mechanism underlying this observation, we favor the idea that Chd1 acts upon nucleosomes that have been perturbed by elongating RNA polymerase II, restoring them to their normal structures or positions and thereby stabilizing them. Importantly, we do not favor the alternative model, that Chd1's effects on chromatin are secondary to perturbation of transcription; we and others do not observe significant alterations of gene expression in chd1 mutants in yeast (Figure S6 and [21]) and Pol II phospho-Ser2 staining of Drosophila polytene chromosomes is normal in chd1 mutants [21].
Chd1's effects on H3 turnover at the 3′ end of genes depended strongly upon gene length (Figure 4), and was also correlated with transcription rate (Figure S7). Given the model above, it is possible that in the absence of Chd1, perturbation of nucleosome positioning by transcription complexes increases with gene length and nucleosome number. Alternatively, Chd1's function may relate to supercoiling changes driven by transcription. To test this we have preliminarily investigated whether additional deletion of the major topoisomerase Top1 affects the chromatin changes observed in chd1Δ yeast mutants. However, we have not observed any suppression of the chd1Δ turnover phenotype in chd1Δtop1Δ double mutants (not shown). Thus, at present we have no additional evidence that supercoiling per se mediates the length dependence of Chd1 on H3 turnover, although given the ability of other topoisomerases to compensate for loss of Top1 we still consider this an appealing hypothesis.
Previous results show that Chd1 has dramatic effects on nucleosome positioning over coding regions [40]. Our results extend this characterization by showing that Chd1 also has dramatic effects on H3 turnover over coding regions, raising the question of whether these two roles for Chd1 in chromatin structure are related. In other words, does Chd1's effect on H3 replacement follow from its role in establishing wild type nucleosome positions, or vice versa? We have no evidence for either possibility, but note that our prior genetic analyses suggest that chd1Δ mutations lower the nucleosomal barrier to RNA Pol2 elongation [22], [28]. Thus, we speculate that disorganized nucleosomes in chd1Δ mutants could be unusually susceptible to eviction by RNA Pol2. This model is consistent with a recent suggestion that elongating polymerases could cause collisions and eviction of adjacent nucleosomes if they are spaced inappropriately [60]. However, arguing against this are observations that nucleosome ladders are little affected in chd1 mutants [40], [41]. Future studies will be needed to address these mechanistic questions.
Taken together, our results identify an evolutionarily conserved role for Chd1 in histone turnover in yeast and flies. Most surprising is our finding that the major site of Chd1 function appears to be at the 3′ ends of genes, suggesting that this enzyme may be recruited or regulated by 3′ histone marks such as H3K36me3. Finally, we find that Chd1 largely affects H3 turnover over longer coding regions, raising the question of whether resolving superhelical tension could be a key role for Chd1 in maintaining wild type chromatin architecture.
Flies were raised on cornmeal, agar, yeast, and molasses medium, supplemented with methyl paraben and propionic acid. To drive the P[UHS-H3.3core-GFP] transgene [44], [61] in the salivary gland, flies were crossed to P{GawB} AB1-Gal4 flies (Bloomington Stock Center). Mutant chd15 flies were described previously [21]. All crosses were carried out at 18°C.
Live analysis of polytene chromosome phenotypes was performed as described previously [62]. To analyze the effect of chd15 on H3.3core-GFP incorporation, chd15 b c sp/BcGla; P[UHS-H3.3core-GFP]/TM6B Tb Hu flies were crossed to chd15 b c sp/BcGla; P{GawB}AB1/TM6B Tb Hu flies at 18°C. Flies with chd15 balanced by CyO Kr-GFP instead of BcGla were also analyzed and yielded similar results. Salivary glands were dissected and imaged from heterozygous and homozygous chd15 third instar larvae. For control nuclei, P{GawB} AB1-Gal4 flies were crossed to P[UHS-H3.3core-GFP]/TM6B Tb Hu flies. H3.3core-GFP expression was quantitated by calculating sum pixel intensity in polytene nuclei using the Volocity software package as described previously [62].
Polytene chromosomes were prepared and fixed as described [63] and immunostained using primary antibodies directed against CHD1 ([21], 1∶300 dilution), H5 anti-RNA polymerase II (specific for the Ser 2-phosphorylated form of Pol II CTD, Covance; 1∶50 dilution), and the JL-8 anti-GFP (Clontech, 1∶300 dilution). Secondary antibodies donkey anti-rabbit IgG-Cy3, donkey anti-mouse IgM-Cy2, and donkey anti-mouse IgG Fc2a-DyLight 649 (Jackson ImmunoResearch Laboratories, 1∶200 dilutions) were tested with each individual primary antibody to ensure specificity. Images were examined on an Olympus 1X81 inverted fluorescence microscope and acquired using Image-Pro6.3. Control and mutant chromosomes were photographed using identical exposure times, and images were processed identically in Adobe Photoshop CS3.
All S. cerevisiae strains used in this study (see Table S1) were constructed by standard procedures, are isogenic to S288c and are GAL2+ [64]. Yeast media was made as described previously [65].
Plasmids used in this study are described in Table S2. Plasmid pJH18-A06 was obtained by random PCR mutagenesis (GAH, TKQ and Araceli Ortiz unpublished). pJH18-Δ4-30, S87P/G90S was created by site-directed mutagenesis of pJH18-A06. PGAL-H4-FlagH3 contains a KpnI-NotI fragment carrying pGAL-driven Flag-H3 from plasmid MDB61 [50], in pRS416.
Strains transformed with pGAL-H4-FlagH3 were grown to ∼1.2×107 cells/ml in SC-Ura media with raffinose as the carbon source. Cells were G1 arrested with alpha factor and Flag-H3/H4WT expression was induced by addition of galactose (2% final concentration). ChIP assays were preformed as described previously [66]. 60 minutes after addition of galactose, cells were crosslinked with 1% formaldehyde for 15 min, disrupted by bead beating and chromatin was sonicated using a Diagenode Bioruptor to obtain an average size of 500 bp. Chromatin was immunoprecipitated using 40 µl (1∶2 slurry) Anti-Flag M2 Affinity gel (A2220; Sigma) or 1 µg of a rabbit polyclonal antibody against the C-terminus of H3 (ab1791; Abcam). Chelex 100 resin (BioRad) was added to the immunoprecipitated material and Input-DNA samples, and the suspensions were placed at 100°C for 10 min to reverse crosslinks. Samples were treated with proteinase K and DNA was recovered.
Initial characterization and confirmatory analyses of ChIP samples were performed by qPCR in a Corbett Life Science Rotor Gene 6000 machine using SYBR Green as the detection dye (qPCR MasterMix Plus for SYBR Green, Eurogentec). The fold difference between immunoprecipitated material (IP) and total Input sample for each qPCR amplified region was calculated as described in [67], following the formula IP/Input = (2InputCt - IPCt). H3 turnover rates were measured as the final ratio between Flag-tagged H3 and total H3 (Flag-H3/Input vs total H3/Input). The sequences of oligonucleotides used in these PCR reactions are listed in Table S3.
The immunoprecipitated DNA was initially PCR amplified using random hexamer primers as described in [68]. The number of cycles used to amplify the samples was adjusted to between 28 and 37 so that there was equal amplification of DNA in the IP vs. Flag-tagged H3 and the IP vs. total H3 samples. Amplified DNA was visualized on a 1% agarose gel and checked for a visible smear of DNA between 500 and 1.2 kB. Amplified DNA from Flag-tagged H3 and total H3 ChIPs samples were labeled and competitively hybridized to tiling microarrays as described below.
Wt and chd1Δ cells were grown to log phase and fixed with 1% formaldehyde. Cell pellets (from 100 mL cells) were resuspended in 8.8 ml Buffer Z (1 M sorbitol, 50 mM Tris-Cl pH 7.4), with addition 6.5 µl of ß-ME (14.3 M, final conc. 10 mM) and 350 µL of zymolyase solution (10 mg/ml in Buffer Z; Seikagaku America), and the cells were incubated at 30°C shaking at 220 rpm. After spinning at 4000× g, 10 min, 4°C, spheroplast pellets were resuspended in 600 µl NP-S buffer (0.5 mM spermidine, 1 mM ß-ME, 0.075% NP-40, 50 mM NaCl, 10 mM Tris pH 7.4, 5 mM MgCl2, 1 mM CaCl2) per 100 ml cell culture equivalent. 25–40 units (depending on yeast strain and cell density) of micrococcal nuclease (Worthington Biochemical) were added and spheroplasts were incubated at 37°C for 20 minutes. The digestion was halted by shifting the reactions to 4°C and adding 0.5 M EDTA to a final concentration of 10 mM.
All steps were done at 4°C unless otherwise indicated. For each aliquot, Buffer L (50 mM Hepes-KOH pH 7.5, 140 mM NaCl, 1 mM EDTA, 1% Triton X-100, 0.1% sodium deoxycholate) components were added from concentrated stocks (10–20×) for a total volume of 0.8 ml per aliquot. Each aliquot was rotated for 1 hour with 100 µl 50% Sepharose Protein A Fast-Flow bead slurry (Sigma) previously equilibrated in Buffer L. The beads were pelleted at 3000× g for 30 sec, and approximately 100 µl of the supernatant was set aside for the input sample. With the remainder, antibodies were added to each aliquot (equivalent to 100 ml of cell culture) in the following volumes: 10 µl anti-H3K36me3 (Abcam polyclonal), or 7 µl anti-H3K4me3 (Millipore monoclonal). Immunoprecipitation, washing, protein degradation, and DNA isolation were performed as previously described [69]. The samples were amplified, with a starting amount of up to 75 ng for ChIP samples, using the DNA linear amplification method described previously [54].
2.5 µg of aRNA produced from the linear amplification were labeled via the amino-allyl method as described on www.microarrays.org. Labeled probes (a mixture of Cy5 labeled input and Cy3 labeled ChIP'ed material) were hybridized onto an Agilent yeast 4×44 whole genome array. The arrays were scanned at 5 micron resolution with the Agilent array scanner. Image analysis and data normalization were performed as previously described [54].
Microarray data have been deposited in GEO (Accession #GSE38540).
|
10.1371/journal.ppat.1003017 | Toll-Like Receptor 8 Ligands Activate a Vitamin D Mediated Autophagic Response that Inhibits Human Immunodeficiency Virus Type 1 | Toll-like receptors (TLR) are important in recognizing microbial pathogens and triggering host innate immune responses, including autophagy, and in the mediation of immune activation during human immunodeficiency virus type-1 (HIV) infection. We report here that TLR8 activation in human macrophages induces the expression of the human cathelicidin microbial peptide (CAMP), the vitamin D receptor (VDR) and cytochrome P450, family 27, subfamily B, polypeptide 1 (CYP27B1), which 1α-hydroxylates the inactive form of vitamin D, 25-hydroxycholecalciferol, into its biologically active metabolite. Moreover, we demonstrate using RNA interference, chemical inhibitors and vitamin D deficient media that TLR8 agonists inhibit HIV through a vitamin D and CAMP dependent autophagic mechanism. These data support an important role for vitamin D in the control of HIV infection, and provide a biological explanation for the benefits of vitamin D. These findings also provide new insights into potential novel targets to prevent and treat HIV infection.
| Cells use macroautophagy (autophagy - ‘self-eating’, lysosome-dependent degradation and recycling of intracellular components in response to stress) as a mechanism to detect intracellular pathogens through pattern-recognition receptors such as Toll-like receptors (TLRs) that recognize signature molecules of pathogens that are essential for their survival. One such Toll-like receptor, TLR8, which is located in human macrophage endosomes, recognizes both imidazoquinoline compounds and uridine-rich single-stranded RNA such as human immunodeficiency virus type-1 (HIV) single-stranded RNA. In the present study we report that TLR8 activation in human macrophages induces the expression of the human cathelicidin microbial peptide (CAMP), the vitamin D receptor (VDR), and cytochrome P450, family 27, subfamily B, polypeptide 1 (CYP27B1), which 1α-hydroxylates the inactive form of vitamin D, 25-hydroxycholecalciferol, into its biologically active metabolite. Moreover, we demonstrate that TLR8 activation induces autophagy in human macrophages through a vitamin D and CAMP dependent mechanism, and that the induction of autophagy by TLR8 agonists inhibits HIV. These data support an important role for vitamin D in the control of HIV infection, and provide a biological explanation for the benefits of vitamin D. These findings also provide new insights into potential novel targets to prevent and treat HIV infection.
| As an obligatory intracellular parasite, human immunodeficiency virus type-1 (HIV) survival is dependent upon its ability to exploit host cell machinery for replication and dissemination, and to circumvent cellular processes that prevent its growth. One such intracellular process is macroautophagy (hereafter referred to as autophagy). Autophagy is a degradation pathway whereby cytosolic double membrane-bound compartments termed autophagosomes engulf cytoplasmic constituents such as sub-cellular organelles and microbial pathogens. These autophagosomes then fuse with lysosomes, resulting in the degradation of the engulfed components. HIV relies on several components of autophagy for its replication with silencing of autophagy proteins inhibiting HIV replication [1]–[6]. In macrophages, HIV group-specific antigen (Gag)-derived proteins colocalize and interact with microtubule-associated protein 1 light chain 3B (LC3B), and are present at LC3B-II enriched membranes suggesting that autophagy may be involved in Gag processing and the production of nascent virions [5]. This is consistent with the hypothesis that HIV is assembled on endocytic membranes that intersect with recycling endosomes [7], [8]. Despite the requirement for autophagy, HIV actively downregulates autophagy regulatory factors, reducing both basal autophagy and the numbers of autophagosomes per cell [9]–[11]. The HIV negative elongation factor (Nef) protein has been shown to protect HIV from degradation by inhibiting autophagosome maturation [5], and enhances HIV replication through interactions with immunity-associated GTPase family M (IRGM) protein. IRGM interacts with the autophagy-associated proteins autophagy related 5 homologue (ATG5), ATG10, LC3B and SH3-domain growth factor receptor-bound protein 2-like endophilin B1, inducing autophagosome formation [6]. However, inducers of autophagy including amino acid starvation, rapamycin, and 1α,25-dihydroxycholecalciferol (1,25D3), the active form of vitamin D, overcome the imposed phagosome maturation block leading to inhibition of viral replication [2], [3], [11]. Interestingly, the HIV envelope glycoprotein expressed on the surface of infected cells has been reported to induce cell death in uninfected bystander CD4+ T cells through autophagy [12], [13].
Recent research has focused on the role of autophagy in the innate and adaptive immune systems. Cells use autophagy as a mechanism to detect intracellular pathogens through pattern-recognition receptors (PRRs) which recognize signature molecules of pathogens termed pathogen-associated molecular patterns (PAMPs) that are essential for their survival. There are several classes of PRRs: Toll like receptors (TLRs), retinoic acid-inducible gene-I-like receptors and nucleotide-binding oligomerization domain-like receptors. These PRRs recognize PAMPs in various cell compartments and trigger the release of inflammatory cytokines and type I interferons for host defense [14], [15]. Human TLR2/1 recognizes Mycobacterium tuberculosis lipoproteins, and upon activation induces the expression of cytochrome P450, family 27, subfamily B, polypeptide 1 (CYP27B1) which 1α-hydroxylates the inactive form of vitamin D3, 25-hydroxycholecalciferol (25D3), into its biologically active metabolite, the steroid hormone 1,25D3. TLR2/1 agonists also induce the activation and upregulation of the vitamin D (1,25D3) receptor (VDR) leading to the induction of the human cathelicidin microbial peptide (CAMP), autophagic flux and the killing of intracellular M. tuberculosis [16], [17]. We have recently demonstrated that 1,25D3 inhibits mycobacterial growth and the replication of HIV through the CAMP-dependent induction of autophagy [3].
TLR8 is phylogenetically and structurally related to TLR7 [18], [19] and is expressed in endosomes of myeloid cells such as monocytes, macrophages and myeloid dendritic cells, and in regulatory T cells [20]–[22]. TLR8 recognizes both uridine-rich single-stranded RNA (ssRNA) and imidazoquinoline compounds [23], [24]. Upon stimulation, TLR8 agonists activate nuclear factor kappa-light-chain-enhancer of activated B cells via the myeloid differentiation primary response gene (88) adaptor protein that leads to the induction of a cascade of antiviral effector functions including the induction of autophagy in murine cells [25] and proinflammatory cytokines in human cells [21]. HIV ssRNA encodes for multiple PAMPs that can be recognized by TLR8 expressed in macrophage endosomes [23], [26] and suppresses HIV replication in acute ex vivo human lymphoid tissue of tonsillar origin and renders peripheral blood mononuclear cells (PBMC) barely permissive to HIV infection [20]. Interestingly, HIV downregulates interleukin-1 receptor-associated kinase 4, which is essential for virtually all TLR signaling [27].
Despite the immune defense mechanisms that the host deploys against HIV and improved antiretroviral therapies, the virus persists in long-lived cells including macrophages and dendritic cells. Major questions remain as to the mechanism by which TLR8 agonists inhibit HIV and whether HIV antigens can activate autophagy in human cells through TLR8. In the present study, we demonstrate that TLR8 ligands, in the presence of 25D3, inhibit HIV replication in macrophages through a vitamin D and CAMP-dependent mechanism involving autophagy.
Both ssRNA40 and the imiquimod R837 promote autophagic responses in murine RAW 264.7 cells [25] through a beclin-1 (BECN1) dependent mechanism. However, the ability of TLR8 ligands to induce an autophagic response in primary human macrophages has not been investigated. Therefore, the ability of TLR8 agonists to induce autophagy in human macrophages was determined in monocyte-derived macrophages cultured in RPMI 1640 supplemented with 10% (v/v) charcoal/dextran treated, heat-inactivated fetal bovine serum (FBS), 10 ng/mL macrophage colony stimulating factor and 100 nmol/L 25D3 as described in the Materials and Methods. The effect of ssRNA40 and the imidazoquinoline CL097 on the formation of the class 3 phosphoinositide-3-kinase (PIK3C3) kinase complex was initially assessed. The PIK3C3 kinase complex is essential for the induction of autophagosome formation at the vesicle elongation step and is formed when BECN1 physically interacts with PIK3C3. Co-immunoprecipitation followed by immunoblotting demonstrated enhanced binding of BECN1 to PIK3C3 forming the PIK3C3 kinase complex following ssRNA40 or CL097 treatment (Figure 1A).
During autophagy, cytosolic LC3B-I is converted to LC3B-II by a ubiquitin-like system that involves ATG7, ATG3 and the ATG12–ATG5 complex. The ATG12–ATG5 complex ligates LC3B-II to the nascent autophagosome membrane through phosphatidylethanolamine with the LC3B-II associated with the inner membrane degraded after fusion of the autophagosome with lysosomes. Therefore, the conversion of LC3B-I to LC3B-II and its turnover is an indicator of autophagy induction and flux [28]. Activation of macrophages with ssRNA40 and CL097 for 24 h led to an increase in LC3B-II similar to that observed with rapamycin, an inducer of autophagy through inhibition of the mammalian target of rapamycin (MTOR) complex 1 (MTORC1), and was increased in the presence of the lysosomal protease inhibitor pepstatin A indicative of autophagic flux (Figure 1B).
During the formation of autophagosomes, LC3B redistributes from a soluble diffuse cytosolic pattern to an insoluble autophagosome-associated vacuolar pattern that can be quantified using fluorescence microscopy [29]. Both ssRNA40 and CL097 induced a significant increase in both the quantity of LC3B per cell and the number of cells with increased LC3B puncta formation in the absence of pyknosis, karyorrhexis, or plasma membrane blebbing and was similar to that observed after rapamycin treatment (Figure 1C). To verify that the increase in the number of autophagosomes in TLR8 agonist treated cells versus control cells represents increased autophagic flux rather than an accumulation of LC3B-positive autophagosomes, the degradation of the polyubiquitin-binding protein sequestosome 1 (SQSTM1) was quantified. Inhibition of autophagy leads to an increase in SQSTM1 protein levels while autolysosomes degrade SQSTM1- and LC3-positive bodies during autophagic flux [30]. Both TLR8 activation and rapamycin treatment of macrophages for 24 h led to a decrease in SQSTM1 protein levels corresponding to the stimulation of autophagic flux (Figure 1D). Moreover, the TLR8 ligands also decreased SQSTM1 protein levels in a dose-dependent manner (Figure 1D).
TLR8 and TLR7 both contribute to the recognition of viral ssRNA and are both found in human macrophages [23], [31]. Therefore, the role of TLR8 and TLR7 in the induction of autophagy in macrophages post-ssRNA40 and CL097 was examined. RNA interference (RNAi) of TLR8 (Figure 2A) significantly inhibited both the ssRNA40- and CL097-mediated LC3B lipidation (Figure 2B) and the increase in the number of LC3B-positive autophagic vesicles (Figure 2C). Although TLR7 is involved in sequence-specific sensing of ssRNAs in human macrophages [31], TLR7 protein expression was undetectable in primary macrophages; therefore, the role of TLR7 in this system is unknown. However, the complete abrogation of TLR8 agonist-induced LC3B lipidation in the presence of RNAi for TLR8 suggests that TLR8 is the mediator of the effect of ssRNA40 and CL097 on autophagy induction.
Recent studies have demonstrated that TLR2/1 activation of human monocytes/macrophages upregulates the expression of vitamin D related genes including CYP27B1 [17] and that 1,25D3 induces autophagy [2], [32]. Given this background, the effect of ssRNA40 and CL097 stimulation on the expression of CYP27B1 and the VDR in macrophages and the role of the vitamin D pathway in TLR8-mediated autophagic flux was investigated. The TLR8 agonists induced a dose-dependent increase in both CYP27B1 (Figure 3A) and VDR (Figure 3B) mRNA and protein expression. Macrophages were then transduced with short-hairpin RNA (shRNA) specific to CYP27B1, VDR or a scrambled non-specific control followed by TLR8 stimulation. Figure 3C shows that CYP27B1 silencing abrogates the lipidation of LC3B in response to TLR8 activation but not in response to rapamycin. Similar results were observed post-VDR silencing. Furthermore, LC3 puncta in CYP27B1 or VDR silenced cells post-TLR8 activation was significantly reduced (Figure 3D). To determine whether differences in vitamin D concentration affect the ability of TLR8 agonists to stimulate autophagy induction, the concentration of 25D3 in the media was reduced to 45 nmol/L reflecting the lower levels observed in vitamin D deficient individuals [33]. At this concentration, the TLR8-mediated induction of autophagy was significantly impaired with little to no LC3B lipidation observed (Figure 3E).
CCAAT/enhancer binding protein β (CEBPB) activation is thought to be a required transcription factor controlling immune-mediated transcription of CYP27B1 [34]. Therefore, to assess the role of CEBPB in CYP27B1 expression, macrophages were transduced with shRNA specific to CEBPB, followed by TLR8 stimulation. Figure 4 shows that CEBPB silencing significantly reduced the expression of CYP27B1 in macrophages post-TLR8 activation.
Previous studies have shown that TLR8 agonists inhibit HIV replication in ex vivo infected lymphoid tissue while inducing virion release from transformed cell lines [20], [35]. We therefore determined whether the TLR8 agonists influence HIV infection and replication in primary macrophages by comparing the extent to which CL097 and ssRNA40 pre-treatment influenced p24 antigen accumulation in the supernatants of macrophages that were subsequently infected with HIV. Both ssRNA40 and CL097 induced a dose-dependent inhibition of HIV replication. This inhibition became significant across all concentrations tested by day 3 post-infection (p<0.01) with the magnitude of the inhibition increasing until cultures were discontinued on day 10 post-infection (Figure 5).
To confirm that the inhibition of HIV observed in macrophages post-CL097 stimulation is predominantly through TLR8, we employed RNAi for TLR8. In the scrambled control RNAi treated cells, CL097 inhibited HIV p24 levels by 90% and 74% at 5 and 1 µg/mL, respectively by day 10 post-infection (p<0.029; Figure 6A). Conversely, TLR8 silencing reduced the inhibitory effect of CL097 to <6% at both concentrations tested, which was not significantly different to the vehicle control treated cells (p>0.48; Figure 6A). Thus, although human macrophages may express low levels of TLR7 [31], TLR8 is the predominant signaling pathway through which CL097 inhibits HIV.
Based on our observations that: i) 1,25D3 inhibits HIV replication through the induction of autophagy [2], [3], ii) TLR8 activation significantly increases the expression of both CYP27B1 and the VDR, and iii) silencing either CYP27B1 or the VDR inhibits TLR8-mediated autophagy, we sought to determine whether the autophagic response induced by TLR8 through the vitamin D pathway was responsible for the observed inhibition of HIV. Silencing of CYP27B1 resulted in the markedly decreased inhibition of HIV by CL097 to levels that were not significantly different to the vehicle control treated cells (p>0.05; Figure 6A). Similarly, silencing the VDR significantly reduced the inhibition to control levels (p>0.1; Figure 6A) suggesting that the vitamin D pathway is important during the inhibition of HIV by TLR8 agonists. To confirm this, and to determine whether differences in the availability of vitamin D affects the ability of TLR8 agonists to inhibit HIV replication, we reduced the concentration of 25D3 in the media to 45 nmol/L, reflecting the lower levels observed in vitamin D deficient individuals. Under these conditions, we observed a significantly diminished capacity of both ssRNA40 and CL097 to inhibit HIV replication (Figure 6B).
To determine whether TLR8-induced autophagy contributes to the CL097-mediated inhibition of HIV by CL097, we assessed the effect of BECN1 and ATG5 silencing on HIV infection post-TLR8 activation. BECN1 silencing reduced the 5 µg/mL CL097 mediated inhibition of HIV at day 10 from 90% to 50% (p = 0.028; Figure 7A). We next assessed the effect of ATG5 silencing. During autophagy, cytosolic LC3B-I is converted to LC3B-II by an ubiquitin-like system that involves ATG7, ATG3 and the ATG12–ATG5 complex. The ATG12–ATG5 complex ligates LC3B-II to the nascent autophagosome membrane through phosphatidylethanolamine. Therefore, RNAi of ATG5 inhibits autophagosome formation. ATG5 RNAi abrogated the CL097 mediated inhibition of HIV by day 10 (90% versus 22% inhibition; p<0.028; Figure 7B).
We next investigated whether autophagosome acidification, a late stage event during autophagy, is required for the TLR8-mediated autophagic inhibition of HIV. During autophagy, lysosomes fuse with autophagosomes to form autolysosomes. Macrophages were treated with bafilomycin A1, an inhibitor of the vacuolar H+ ATPase and autophagosome-lysosome fusion, and subsequently infected with HIV. Bafilomycin A1 reversed the TLR8-mediated inhibition of HIV (Figure 7C) suggesting that the acidic pH of autolysosomes is required for the autophagy-mediated control of HIV.
After lysosomes fuse with autophagosomes to form autolysosomes, the sequestered components are degraded by lysosomal hydrolases and released into the cytosol by lysosomal efflux permeases. We investigated whether lysosomal hydrolases are important for TLR8-mediated inhibition of HIV through autophagy using SID 26681509, a novel thiocarbazate specific inhibitor of the lysosome hydrolase cathepsin L. In the absence of TLR8 ligands, SID 26681509 induced no net inhibition of HIV (Figure 7C). Moreover, in the presence of TLR8 ligands, SID 26681509 abrogated the HIV inhibition (Figure 7C).
Previous studies have demonstrated that CAMP expression is upregulated by 1,25D3, that it is required for 1,25D3 mediated autophagy [3], [32], and that it is involved in the autophagic inhibition of HIV in human macrophages [3]. Moreover, monocytes express CAMP in response to TLR2/1 agonists [16]. Therefore, to determine the role of CAMP in the TLR8-mediated autophagic response, we first investigated whether TLR8 agonists induce the expression of CAMP. Both ssRNA40 and CL097 induced the expression of CAMP mRNA by 13- and 10-fold, respectively over the vehicle control (p = 0.029; Figure 8A). These data indicate that TLR8 activation triggers CAMP expression in human macrophages.
A functioning vitamin D signaling pathway is required for the expression of CAMP in response to TLR2/1 agonists [36]. To assess whether TLR8 activation of CAMP expression was dependent on the presence of 25D3, CAMP expression post-TLR8 activation was investigated in macrophages in 25D3 sufficient and deficient media. TLR8-induced CAMP expression was observed in cultures containing 100 nmol/L 25D3, but not in vitamin D deficient culture medium (Figure 8B).
To address the role of CAMP in TLR8-induced autophagy and antimicrobial activity, RNAi for CAMP was employed. Transduction of shCAMP into macrophages significantly blocked endogenous LC3B lipidation post-TLR8 activation, whereas macrophages transduced with a scrambled control (shNS) showed increased LC3B-II conversion consistent with autophagosome formation and the induction of autophagy (Figure 8C). Consistent with the findings that autophagy is required for the restriction of HIV replication, CAMP silencing reduced TLR8-mediated inhibition of HIV to insignificant levels (p>0.09; Figure 8D). Collectively, these data suggest that 25D3 is required for the TLR8 induced expression of CAMP and that CAMP expression is required for TLR8-mediated antimicrobial activity in human macrophages.
The antimicrobial effects of vitamin D have been well documented and association studies have linked low levels of 25D3 and/or 1,25D3 with increased risk of, or severity of infection with HIV [37], [38]. The present study identifies how vitamin D deficiency may influence innate immunity against HIV infection. Stimulation of human macrophages with TLR8 agonists upregulates the expression of CYP27B1 and the VDR leading to the induction of CAMP and autophagic flux. Moreover, when serum was 25D3 deficient, or when the vitamin D signaling pathway was silenced, TLR8 agonists were unable to induce autophagy. Thus, the presence of 25D3 and a functional vitamin D signaling pathway are required for TLR8-induced autophagy.
Previous studies have demonstrated that 1,25D3 induces autophagy in primary macrophages through a CAMP dependent mechanism [3], [32] and inhibits HIV replication in macrophages [2], [3]. Consistent with published data, CL097 and the guanosine- and uracil-rich oligonucleotide ssRNA40, but not RNA41 in which all uracils were replaced with adenosines, inhibited HIV replication in primary macrophages [20]. The current study expands on these findings and demonstrates that TLR8 agonists inhibit HIV replication in macrophages through a vitamin D3- and CAMP-dependent mechanism involving autophagy. Indeed, the TLR8-mediated inhibition of HIV replication occurred only in the presence of vitamin D-sufficient media or in cells with an intact vitamin D signaling pathway.
The present data demonstrate that in CAMP silenced cells, TLR8 activation failed to induce LC3B-II lipidation and inhibit HIV. Endogenous CAMP has been implicated in a number of cellular functions including the regulation of inflammatory responses [39] and the formation and maturation of autophagosomes [32]. CAMP has also been shown to play an important role in the activation of mitogen activated protein kinases and CEBPB which contribute to the transcriptional activation of BECN1 and ATG5 in response to 1,25D3 [32]. Moreover, during autophagy, autophagosomes recruit CAMP through an AMP kinase, Ca2+ and calcium/calmodulin-dependent protein kinase kinase 2 beta dependent mechanism where it is involved in microbial killing [32]. Further work is necessary to determine the precise role of CAMP in TLR-activated autophagy and antiretroviral activity.
Vitamin D deficiency is conservatively defined by most experts as <50 nmol/L 25D3 [33]; 52–72 nmol/L 25D3 is considered to indicate insufficiency and >73 nmol/L considered sufficient [33]. In contrast to this, the estimated mean concentration of 25D3 present in people worldwide is just 54 nmol/L [40]. The major source of vitamin D is through the endogenous photochemical conversion of 7-dehydrocholesterol in the skin to pre-vitamin D3 by ultra-violet B light exposure which then undergoes a 1,7-sigmatropic hydrogen transfer forming cholecalciferol. This is then transferred from the skin by the vitamin D binding protein and is subsequently 25-hydroxylated by cytochrome P450, family 2, subfamily R, polypeptide 1 (CYP2R1) in hepatocytes to form 25D3 in a poorly regulated manner. Lesser amounts of vitamin D3 metabolites are also consumed through fortified dairy products and oily fish. Vitamin D status, therefore, is largely dependent upon the availability of cholecalciferol. Why HIV-infected individuals tend to have lower levels of 1,25D3 and/or 25D3 is largely unknown but it is possible that inadequate renal 1α-hydroxylation mediated by pro-inflammatory cytokines and/or a direct effect of antiretroviral drugs play a role [37]. Four genes contribute to the variability of serum 25D3 concentrations: 7-dehydrocholesterol reductase (involved in cholesterol synthesis and the availability of 7-dehydrocholesterol in the skin), 25-hydroxylase CYP2R1, and CYP24A1 (cytochrome P450, family 24, subfamily A, polypeptide 1) (degrades and recycles 1,25D3), and GC (group-specific component [vitamin D binding protein]) which encodes for the vitamin D binding protein. Genetic variations at these loci were recently identified to be significantly associated with an increased risk of 25D3 insufficiency [41].
The characterization of the TLR8/vitamin D mediated antimicrobial mechanism in macrophages provides further evidence of the link between vitamin D and the immune system. In a recent study, 25D3 levels were negatively correlated with the expression of TLR8 in human monocytes. In the same study, it was observed that in healthy individuals circulating 25D3 levels and TLR8 expression decreased with age and that this decrease coincided with a decrease in CAMP expression [42]. Unlike the parathyroid-hormone responsiveness of renal CYP27B1, extra-renal CYP27B1 is not subject to the same feedback control so that the local synthesis of 1,25D3 in macrophages probably reflects the availability of 25D3. Therefore, the intracrine nature of this mechanism suggests that the ability of TLR8 to promote HIV killing could be affected by the availability of 25D3 and the efficiency of the synthesis of 1,25D3 by macrophages.
TLR7 and TLR8 expression in peripheral blood monocytes decreases with disease progression and monocytes from HIV-infected individuals produce less tumor necrosis factor following TLR8 activation than those from uninfected individuals while successfully inhibiting HIV infection [43]. Moreover, these monocyte responses are negatively correlated with CD4+ T cell count and positively associated with HIV viral load [44]. The ability of cells to respond strongly to a TLR8 agonist in the presence of high HIV viremia means that ongoing chronic immune activation can be continuously driven by HIV-encoded PAMPs. Despite this, there is no tolerance induction towards TLR8 agonists [35], [44]. Persistent immune activation during HIV infection contributes to the pathogenesis of disease by disturbing the functional organization of the immune system with induction of high levels of cytokines and chemokines. Therefore, chronic stimulation of the innate immune system by TLR ligands may result in the chronic production of proinflammatory cytokines which drive disease progression through generalized immune activation [45]. Supporting this model is the association of a single-nucleotide polymorphism in TLR8 (TLR8 A1G; rs3764880) which confers a significant protective effect against HIV disease progression [46]; however, this same polymorphism increases male susceptibility to pulmonary tuberculosis [47], [48]. Despite these apparent limitations, TLR8 agonists given as vaccine adjuvants with HIV proteins in non-human primate models enhance the magnitude and quality of the anti-HIV Th1 and CD8+ T cell responses [49]. Finally, as TLR8 activation of the latently infected cell lines U1 and OM10 results in a marked increase in HIV replication [20], TLR8 triggering of latently infected macrophages may result in the increased release of HIV in vivo. Therefore, it may be possible to use TLR8 agonists to purge latently infected cells while inhibiting new infections. Thus, further research on the effect of TLR8 agonists on latently infected macrophages from HIV-infected individuals is warranted.
Collectively, this study demonstrates that TLR8 agonists inhibit HIV replication in macrophages through the induction of autophagy that is dependent upon both available 25D3 and a functioning vitamin D signaling pathway as well as the induction of CAMP. Moreover, this study also expands the known PAMP that induce vitamin D-dependent autophagy to include TLR8. Well-controlled clinical trials are needed to determine if vitamin D supplementation is of value as adjunctive treatment in HIV-infected persons. Dissecting the molecular mechanisms by which HIV utilizes autophagy has the potential to lead to the identification of novel drug candidates to prevent and treat HIV infection and related opportunistic infections including tuberculosis.
Venous blood was drawn from HIV seronegative subjects using a protocol that was reviewed and approved by the Human Research Protections Program of the University of California, San Diego (Project 08-1613) in accordance with the requirements of the Code of Federal Regulations on the Protection of Human Subjects (45 CFR 46 and 21 CFR 50 and 56). Written informed consent was obtained from all blood donors prior to their participation.
Peripheral blood mononuclear cells (PBMC) were isolated from whole blood of HIV seronegative donors by density gradient centrifugation over Ficoll-Paque Plus (GE Healthcare). PBMC were then incubated overnight at 37°C, 5% CO2 in RPMI 1640 (Gibco) supplemented with 10% (v/v) charcoal/dextran treated, heat-inactivated FBS (Gemini Bio-Products) and 10 ng/mL macrophage colony stimulating factor (R&D Systems), after which non-adherent cells were removed by aspiration. Monocyte derived macrophages were obtained by further incubating the adherent population in RPMI 1640 (Gibco) supplemented with 10% (v/v) heat-inactivated FBS and 10 ng/mL macrophage colony stimulating factor (R&D Systems) for 10 d at 37°C, 5% CO2. All experiments were performed in RPMI 1640 supplemented with 10% (v/v) charcoal/dextran treated, heat-inactivated FBS, 10 ng/mL macrophage colony stimulating factor and 100 nmol/L 25D3 (Sigma) unless otherwise stated.
CL097, ssRNA40 and ssRNA41 were obtained from Invivogen and were described previously [23]. Pepstatin A, bafilomycin A1, SID 26681509 and rapamycin were purchased from Sigma. Bafilomycin A1 was used at 100 nmol/L, SID 26681509 at 50 nmol/L, and pepstatin A at 10 µg/mL with pretreatment for 1 h before addition of TLR8 ligands or rapamycin.
HIVBa-L was obtained through the AIDS Research and Reference Reagent Program, from Dr. Suzanne Gartner and Dr. Robert Gallo [50], [51]. Virus stocks and titers were prepared as previously described using the Alliance HIV p24 antigen ELISA (Perkin Elmer) [52]. Cells were infected with 105 TCID50/mL HIVBa-L per 5×105 cells for 3 h after 24 h pretreatment with TLR8 ligands or rapamycin unless otherwise stated.
LC3B (D11), PIK3C3 (D9A5), SQSTM1 (D5E2), BECN1 (D40C5), and ATG5 antibodies were obtained from Cell Signaling; VDR (N-terminal), TLR8 (4C6), and β-actin (AC-74) antibodies were from Sigma; CYP27B1 (H-90) antibody was from Santa Cruz Biotechnology. Cell lysates were prepared using CelLytic M (Sigma) supplemented with protease inhibitors (Thermo Scientific). For co-immunoprecipitation, 50 µg anti-BECN1 was immobilized in a coupling gel then 50 µg of the cell lysates were incubated with the antibody-immobilized coupling gel using the ProFound-Co-Immunoprecipitation kit (Thermo Scientific). For immunoblot analyses, cell lysates were resolved using 2-[bis(2-hydroxyethyl)amino]-2-(hydroxymethyl)propane-1,3-diol buffered 12% polyacrylamide gel (Novex) and transferred to polyvinylidene difluoride membranes (Thermo Scientific), followed by detection with the WesternBreeze chemiluminescence kit (Novex) as described previously [52]. Relative densities of the target bands compared to the reference β-actin bands were analyzed using ImageJ (NIH).
Cells were fixed and permeabilized in Dulbecco's phosphate buffered saline supplemented with 4.5% (w/v) paraformaldehyde and 0.1% (v/v) saponin for 30 min, washed, then probed with rabbit anti-LC3B (D11) for 30 mins followed by goat anti-rabbit Alexa Fluor 488 conjugated antibodies (Molecular Probes) for 30 mins and counterstained with Hoechst 33342. Cells and LC3B puncta were imaged and counted using an Olympus IX71 inverted fluorescence microscope as described previously [2].
Lentiviral transduction of macrophages with MISSION lentiviral particles containing shRNAs targeting ATG5 (SHCLNV-NM_004849/TRCN0000150940), BECN1 (SHCLNV-NM_003766/TRCN0000033551), CAMP (SHCLNV- NM_004345/TRCN0000118645), CYP27B1 (SHCLNV-NM_000785/TRCN0000064365), TLR8 (SHCLNV-NM_138636/TRCN0000359246), CEBPB (SHCLNV-NM_005194/TRCN0000007440), VDR (SHCLNV-NM_000376/TRCN0000277001), or scrambled non-target negative control (Scr, SHC002V) was performed according to the manufacturer's protocol (Sigma). Macrophages were transduced with non-specific scrambled shRNA (shNS) or target shRNA and selected using puromycin (Gibco). Five days later, cells were analyzed for target gene silencing and used in experiments.
mRNA quantification was measured by real time PCR using the LightCycler 1.5 Instrument and the FastStart RNA Master SYBR Green I kit (all Roche Applied Science). PCR reactions were carried out in a 20 µL mixture composed of 3.25 mM Mn(CH3COO)2, 0.5 µM of each primer, 1 µL sample and 1-fold LightCycler RNA Master SYBR Green I. Primers were synthesized by Integrated DNA Technologies and were CYP27B1 sense 5′-GTTTGTGTCCACGCTG-3′, antisense 5′-CCCGCCAATAGCAACT-3′; VDR sense 5′-GTTGCTAAACGAGTCAATCC-3′, antisense 5′-AGTAACGGCACGATCT-3′; CAMP sense 5′-CTCGGATGCTAACCTCT-3′, antisense 5′-CATACACCGCTTCACC-3′; polymerase (RNA) II (DNA directed) polypeptide A (POLR2A) sense 5′-GCACCACGTCCAATGACAT-3′, antisense 5′-GTGCGGCTGCTTCCATAA-3′. Reaction mixtures were initially incubated at 61°C for 20 min to reverse transcribe the RNA. Samples were then heated to 95°C for 30 sec to denature the cDNA followed by 45 cycles consisting of following parameters: CYP27B1 5 s at 95°C, 15 s at 55°C and 16 s at 72°C; VDR 5 s at 95°C, 20 s at 58°C and 25 s at 72°C; CAMP 10 s at 95°C, 10 s at 61°C and 7 s at 72°C each with a single fluorescent reading at the end of each cycle followed by a melting curve analysis. To exclude contamination with DNA, Alu-PCR and minus reverse transcriptase controls were performed. Results were calculated using the Pfaffl method [53] and are expressed as the ratio between the target gene and the reference gene POLR2A and normalized so that CYP27B1, VDR and CAMP mRNA expression in unconditioned cells equals 1.00.
Intracellular staining of endogenous CAMP was performed as previously described [2] using goat anti-LL-37 antibodies (Santa Cruz Biotechnology) and Alexa Fluor 647 conjugated donkey anti-goat antibodies (Invitrogen).
Comparisons between groups were performed using the nonparametric two-sided Mann-Whitney U test. Differences were considered to be statistically significant when p<0.05.
|
10.1371/journal.pgen.1006679 | COLEC10 is mutated in 3MC patients and regulates early craniofacial development | 3MC syndrome is an autosomal recessive heterogeneous disorder with features linked to developmental abnormalities. The main features include facial dysmorphism, craniosynostosis and cleft lip/palate; skeletal structures derived from cranial neural crest cells (cNCC). We previously reported that lectin complement pathway genes COLEC11 and MASP1/3 are mutated in 3MC syndrome patients. Here we define a new gene, COLEC10, also mutated in 3MC families and present novel mutations in COLEC11 and MASP1/3 genes in a further five families. The protein products of COLEC11 and COLEC10, CL-K1 and CL-L1 respectively, form heteromeric complexes. We show COLEC10 is expressed in the base membrane of the palate during murine embryo development. We demonstrate how mutations in COLEC10 (c.25C>T; p.Arg9Ter, c.226delA; p.Gly77Glufs*66 and c.528C>G p.Cys176Trp) impair the expression and/or secretion of CL-L1 highlighting their pathogenicity. Together, these findings provide further evidence linking the lectin complement pathway and complement factors COLEC11 and COLEC10 to morphogenesis of craniofacial structures and 3MC etiology.
| The 3MC syndrome is a unifying term amalgamating four rare recessive genetic disorders with overlapping features namely; Mingarelli, Malpuech, Michels and Carnevale syndromes. It is characterised by facial malformations including, high-arched eyebrows, cleft lip/palate, hypertelorism, developmental delay and hearing loss. We previously reported that lectin complement pathway genes COLEC11 and MASP1/3 were mutated in 3MC syndrome patients. Here we describe a new gene from the same pathway, COLEC10, mutated in 3MC patients. Our results show that COLEC10 is expressed in craniofacial tissues during development. We demonstrate how CL-L1, the protein expressed by COLEC10, can act as a cellular chemoattractant in vitro, controlling cell movement and migration. We overexpressed constructs carrying COLEC10 non-sense mutations found in our patients, CL-L1 failed to be expressed and secreted. Moreover, when we expressed a missense COLEC10 construct, CL-L1 was expressed but failed to be secreted. In sum, we discovered a new gene, COLEC10, mutated in 3MC syndrome and we propose a pathogenic mechanism for 3MC relating to the failure of CL-L1 function and its craniofacial developmental consequences.
| 3MC syndrome (MIM 257920;265050;248340) is a unifying term amalgamating four rare autosomal recessive disorders with overlapping features namely; Mingarelli, Malpuech, Michels and Carnevale syndromes. 3MC syndrome is characterized by facial features including hypertelorism, cleft lip/palate, high-arched eyebrows, craniosynostosis, developmental delay and hearing loss [1–3]. We previously reported that mutations in COLEC11 and MASP1/3 genes were responsible for several cases of 3MC syndrome [4]. Since then, further novel mutations in MASP1/3 and COLEC11 have been reported in 3MC patients [5–7]. COLEC11 and COLEC10 encode CL-K1 (also known as CL-11) and CL-L1 (also known as CL-10) respectively, members of the collectin family with an N-terminal collagen-like domain linked to C-terminal carbohydrate-recognition domains (CRDs). CL-K1 and CL-L1 are able to bind to microorganisms including bacteria, fungi and viruses, through their CRDs. This binding capacity to antigens, followed by their interaction with MASP proteins, is their main role in lectin complement pathway activation. [8–12]. However the pathogenic mechanism of lectin complement related proteins in 3MC syndrome is not yet understood [4].
CL-K1 and CL-L1 can also work in partnership in complement activation [13]. Usually CL-K1 and CL-L1 form homodimers, as is generally the case with CDR-domain containing proteins but CL-K1 and CL-L1 can form CL-K1/CL-L1 (also known as CL-LK) heterodimers in plasma and in vitro. These CL-LK heterodimers can also interact and form complexes with MASP-1, MASP-2 and MASP3 [10].
MASP1/3 encodes for 3 alternative products MASP-1, MASP-3 and MAp44 [14]. MASP-1 collaborates with MASP-2 to activate C4. MAp44 has the MASP1 H domain truncated and inhibits MASP1 and MASP2 complement activation. MASP3 shares H chain domain with MASP1 and have a unique protease domain. The precise role of MASP-3 in complement signalling is still unclear, but it has been proposed to form a complex with CL-LK and MASP-2 [10]. It remains to be determined whether these interactions play a role in embryological development, perturbation of which gives rise to the diverse morphological features of 3MC syndrome.
Recently it has been shown that 3MC mutations in COLEC11 inhibit secretion of CL-K1 in mammalian cells, reducing the normal serum levels of CL-K1 and probably disrupting interaction with MASPs or CL-L1 [15].Another report describes how three exonic polymorphisms in COLEC11 and COLEC10 also have an effect in reducing levels of circulating CL-K1 and CL-L1 in serum [16]. Those findings hint how mutations and polymorphisms in both COLEC11 and COLEC10, can directly affect CL-K1 and CL-L1 secretion.
The skeletal phenotype of 3MC patients is the result of complex embryological processes, including neural crest cell (NCC) induction, migration, morphogenesis and differentiation [17]. Correct migration of cNCC is essential for the formation of many tissues in the head from cartilage and bones to muscle and ganglia [18–21]. The regulation and control of NCC migration is complex involving multiple genetic pathways including Wnt, Shh and transcription factors such as Hox and Dlx genes [18,22,23]. Complement factors, such as C3a, have been recently established to play a role in NCC cohesion during migration. Mayor and collaborators have established how complex collective cell migration of NCC requires complement proteins. For example, C3a and its receptor C3aR work together to co-attract each other in order to maintain the coordinated migration of NCC [24–26].
In the present study we describe mutations in a novel lectin alternative pathway gene, COLEC10, in 3MC patients, adding to the body of evidence implicating the complement pathway in human development. We also present new COLEC11 and MASP1/3 mutations found in our cohort of 3MC patients. To validate COLEC10 mutations as causative of 3MC syndrome we determine its expression pattern in the developing mouse embryo and we further demonstrate the in vitro functional consequences of COLEC10 mutations, and present evidence that CL-L1 act as a cellular chemoattractant. Finally we propose a pathogenic mechanism for 3MC relating to the failure of CL-L1 function and its developmental consequences in 3MC.
We collected a bank of patient DNA samples comprising diagnoses of Carnevale, Mingarelli, Michels and Malpuech syndromes. Our cohort currently consists of 45 3MC families of Asian, Middle Eastern and European origin. We previously demonstrated that mutations in COLEC11 and MASP1/3 lectin complement pathway related genes are causative of 3MC syndrome in 11 families and 16 patients. Therefore, we screened for COLEC11 and MASP1/3 mutations by Sanger sequencing in the remaining 34 families and 36 patients in this heterogenous group of patients.
We found three novel homozygous mutations in COLEC11 (NM_024027.4) in three patients and a single homozygous mutation in MASP1/3 (NM_139125.3) in one patient (see Table 1 and Fig 1A and 1B). Of these, two patients were from consanguineous families; MC35.1 (Pakistani) and MC37.1 (Somalian). Both harbored non-synonymous homozygous mutations in COLEC11 leading to a predicted premature termination codon, c.309delT (p.Gly104Valfs*29, exon 4) and a predicted damaging missense, c.G496A (p.Ala166Thr, exon 6) respectively. For patient M35.1 we sequenced the parents, demonstrating that the mutations segregated with the disorder. Parental samples were not available for patient MC37.1.
We found in patient MC29.1 a deletion of 10 nucleotides in COLEC11 (c.89_98delATGACGCCTG, exon 2) which predicts a frameshift change and the introduction of a premature stop codon (p.Asp30Alafs*68). None of the COLEC11 mutations was present in the Exome Aggregation Consortium Database (ExAC), (Cambridge, MA URL http://exac.broadinstitute.org). Overall, two of the new COLEC11 mutations lead to premature terminations (p.Gly104Valfs*29 and p.Asp30Alafs*68), or the missense mutation p.Ala166Thr. This last missense change lies, within the CRD, as shown in Fig 1A, and probably disrupts its recognition function.
In our 3MC cohort we also found a new mutation affecting the second previously described gene mutated in 3MC, MASP1/3 (NM_139125.3). Patient MC27.1, with a consaguinous family, presents a homozygous nonsense mutation (c.9G>A) leading to premature truncation of the protein recently been reported by [6].
These results corroborate our previous finding that genes involved in the lectin complement pathways cause 3MC. However, mutations in COLEC11 and MASP1/3 were excluded in the remaining 30 families and 32 patients. Therefore, we performed whole exome sequencing (WES) in six 3MC patients from consanguineous families, without mutations in COLEC11 or MASP1/3, in order to identify new causative gene associations. We found one patient diagnosed with Michels syndrome harbouring deletions in COLEC10 (NM_006438.4), another member of the collectin family. Despite parental consanguinity in this family, we discovered that the proband, MC19.1 harboured compound heterozygous mutations, c.25C>T; p.Arg9Ter in exon 1 and c.226delA; p.Gly77Glufs*66 in exon 3. We confirmed these mutations segregated with disease by Sanger sequencing (Table 2 and Fig 1D and 1E). The affected sibling, MC19.2, also harboured the same compound heterozygous mutations in COLEC10.
Next we Sanger sequenced COLEC10 in the remainder of our patient cohort. These patients were previously screened for COLEC11 and MASP1/3 mutations, with none identified. We identified another patient (25.1) with the p.Arg9Ter COLEC10 mutation accompanied by a new missense mutation c.528C>G, p.Cys176Trp (exon6) in the other allele (Table 2 and Fig 1D and 1E). The unaffected sibling or parents were not available for testing, therefore we cannot conclusively state that both mutations in patient 25.1 could be in -cis.
The p.Gly77Glufs*66 mutation is not present in the ExAC database and p.Cys176Trp (position Chr8:120118124 C / G, not found in dbSNP) has a frequency of 1 in 120850 chromosomes in the same database. The p.Arg9Ter mutation (rs149010496) is present in only 4 alleles out of 121220 (ExAC). Collectively, these data strongly support the notion that pathogenic mutations in COLEC10 cause a subset of 3MC diagnoses.
COLEC10 mutations c.25C>T; p.Arg9Ter and c.226delA; p.Gly77Glufs*66 both lead to early termination and are likely to produce either truncated proteins or undergo non-sense mediated decay. However, the missense mutation p.Cys176Trp lies in the CRD domain of CL-L1 (Fig 1C), affecting a cysteine residue Cys176 that forms a disulphide bond with C270 [9] and is predicted by PolyPhen-2 to be damaging (http://genetics.bwh.harvard.edu). We next used the SWISS-MODEL Workspace application (http://swissmodel.expasy.org) to predict how the p.Cys176Trp mutation might affect the secondary structure of the CL-L1 protein. Residue 176 on the second helix-loop-helix domain of the protein is predicted to change the tridimensional structure of the protein (Fig 1F), probably affecting the C-type lectin domain function. Table 3 shows detailed clinical features for all of described patients.
To further characterise the function of COLEC10 we assessed intracellular localisation of CL-L1 in ATDC5 cells, a murine chondrocyte cell line. Consistent with previous results for COLEC11 [4], we observed expression of CL-L1 in the Golgi apparatus consistent with a secreted peptide, colocalising with the TGN marker 58K, and with cytosolic expression (Fig 2A). We also found CL-L1 colocalised with laminin, a major component of the basal lamina (Fig 2B). This expression is similar to the cellular colocalisation we found between CL-K1 and laminin (Fig 2C).
Next, we analysed the expression of CL-L1 during murine craniofacial development. We detected CL-L1 expression in the epithelium and mesenchyme of the palate shelf and jaw in E18.5 embryos (Fig 2D). Moreover, we found by immunofluorescence that this particular mandibular epithelial expression is present as early as E13.5, revealing coexpression between CL-L1 and laminin, where CL-L1 is clearly visible in the basement membrane in the palate area (Fig 2E).
We investigated the ability of CL-L1 to act as a chemoattractant in the context of human cells. We spotted 1% (w/v) low melting point agarose discs mixed with PBS, BSA or recombinant human CL-L1. As reported previously, when the same experiment was performed for CL-K1 [4], cells were observed to invade the protein-containing agarose disc. To quantify this effect, we calculated the cell invasion index as shown in Fig 3A. We found that PBS and BSA containing discs failed to attract any cells (Fig 3B and 3C and S1 and S2 Movies respectively) which was in stark contrast to CL-L1 containing discs that exhibited extensive migration/invasion into the discs with an invasion index score of 140.0±22.9 (Fig 3C and S3 Movie).
Having demonstrated a role for CL-L1 in normal craniofacial development we sought to confirm that the mutations found in our 3MC patients were pathogenic. We predicted that COLEC10 mutations c.25C>T; p.Arg9Ter and c.226delA; p.Gly77Glufs*66 would lead to either truncated or absent protein. However, we expected that the missense mutation c.528C>G, p.Cys176Trp, affecting a crucial cysteine residue, would likely lead to abnormal protein folding and possibly affects secretion, as seen with three disease-associated mutations in COLEC11 [15]. To test this hypothesis, we transfected COLEC10WT, COLEC10Arg9Ter and COLEC10Gly77Glufs*66 constructs into HeLa and HEK293 cell lines and detected CL-L1 expression.
Immunoblotting demonstrated that CL-L1 protein was present in both cell extracts and supernatants when COLEC10WT plasmid was transfected into HEK293 cells. By contrast, no protein was detected when the mutant plasmids COLEC10Arg9Ter and COLEC10Gly77Glufs*66 were transfected, suggesting that both transcripts underwent nonsense-mediated decay. Transfection of COLEC10Cys176Trp plasmid allowed CL-L1 expression but not secretion as demonstrated by detection of CL-L1 in the cell lysates but not in the supernatant (S1 Fig).
Western blot data were further supported by quantitative ELISA (Fig 3D). The results showed highest levels of CL-L1 protein in pellets of cells transfected with COLEC10Cys176Trp plasmid than cells transfected with COLEC10WT (HeLa COLEC10Cys176Trp 2518.3±21.3ng/mL vs HeLa COLEC10WT 1823.3±7.2ng/mL, p<0.001; HEK293 COLEC10Cys176Trp 1302.7±3.7ng/mL vs HEK293 COLEC10WT 632.0±3.6ng/mL, p<0.001. Fig 3D and S1 Table). However, secretion of CL-L1 was severely reduced in the COLEC10Cys176Trp transfected cells compared with COLEC10WT supernatant transfections (HeLa COLEC10Cys176Trp 12.5±0.2ng/mL vs HeLa COLEC10WT 200.3±1.5ng/mL, p<0.001; HEK293 COLEC10Cys176Trp 5.7±0.1ng/mL vs HEK293 COLEC10WT 390.2±4.1ng/mL, p<0.001. Fig 3D and S1 Table). These results suggest that accumulation of CL-L1 in cell pellets in COLEC10Cys176Trp is the result of severely reduced levels of CL-L1 secretion. Besides, no CL-L1 expression was observed for COLEC10Arg9Ter and COLEC10Gly77Glufs*66 transfected cells, which served as a negative control.
We previously showed COLEC11 and MASP1/3 lectin alternative pathway genes were mutated in 3MC patients. Since our initial discovery, several groups reported mutations in COLEC11 and MASP1/3 in their 3MC cohorts [5–7]. Here we report four new mutations for COLEC11 affecting four further 3MC patients from consanguineous families. None of these mutations has been found in the ExAc database, supporting pathogenicity and indicating their private nature in these pedigrees. We also identified another MASP1/3 mutation in the homozygous state, c.9G>A, in our cohort confirming a prior report of this mutation by Urquhart et al. [6].
These results increase the percentage of patients with known mutations in our 3MC cohort; 23% carry a COLEC11 mutation and 12% now carry a MASP1/3 mutation. In the remaining patients we identified a second member of the collectin family, COLEC10, found to be mutated in 3MC. The addition of these 2 families in COLEC10 (5%) increase the coverage to 40% of known genes of our patients. Therefore, over 60% of our 3MC cohort is still without molecular confirmation of disease and that at least one further gene remains to be identified.
In contrast with COLEC11 patient mutations, all three COLEC10 patients have compound heterozygous COLEC10 mutations, which is slightly surprising as they come from consanguineous families. They all share the terminating mutation c.25C>T;Arg9Ter, found in ExAc in the general population at a low frequency (0.00003300) (Table 2), whereas the mutations c.226delA and c.528C>G were not present in the ExAc database.
In recent years a very well documented evidence implicating cNCC migration in craniofacial cartilage and bone morphogenesis has accumulated (reviewed in [19]). Our data suggests the failure of NCCs to migrate correctly is the principal factor leading to craniofacial abnormalities in 3MC patients. We confirmed that CL-L1 has chemotactic properties, most likely through recognition of carbohydrates on the cell surface, providing a potential explanation on how its absence can lead to abnormal NCC migration in 3MC. This is not surprising as other complement pathway proteins have previously been shown to play important roles in cell migration. For example in the first steps of the regulation of NCCs, crest cells are co-attracted by the complement fragment C3a and its receptor C3aR. When the C3aR function is inhibited enteric neural crest cell adhesion and migration is affected, and there is an increase in NCC dispersion [24,26]. It is worth noting that the lectin complement pathway can also induce cleavage of C3 to C3a [25] which in turn can regulate NCC migration.
Furthermore, other complement factors also regulate cell migration and morphology. C3 regulates epithelial-mesenchymal transition via TWIST1 activation [27]. C3a also controls radial intercalation during early gastrulation and tissue spreading [28]. An important common functionality of C3a is its capacity to act as a chemoattractant to pull cells together and force them to migrate collectively.
In the lectin complement pathway CL-L1 can form a complex with CL-K1, called CL-LK, and bind to MASP1/3 and MASP2 [10] to activate the lectin complement pathway.
We propose here that the role of CL-L1 and CL-K1 lies in regulating cell migration via cell attraction in 3MC syndrome. We know that CL-L1 and CL-K1 can act by themselves to attract cells but both can also form the heteromeric complex CL-LK that can also bind to MASP1/3 and MASP2 with higher affinity than CL-K1 homodimers [10]. Therefore, it is possible that the NCC migration in vivo requires cooperation of heteromeric interactions between CL-L1 and CL-K1. That is supported by the observations that COLEC11 and COLEC10 genetic variants strongly influence the circulating serum levels of CL-K1 and CL-L1 and that a major proportion of these proteins are circulating in the form of heterocomplexes [16]. As such, whilst we have demonstrated CL-L1 can in itself induce cell migration and invasion, the exact molecular pathway leading to NCC migration regulation requires further investigation.
We did not observe any COLEC10 expression in cells pellets and supernatant when overexpressing the mutations 9G>A; ArgXTer and c.226delA; p.Gly77Glufs*66 (Fig 3D). However, the missense c.528C>G, p.Cys176Trp mutation did not affect COLEC10 expression, although it did prevent cellular secretion of the protein into the supernatant. Furthermore, 3MC patient mutations in COLEC11 also show a similar secretory phenotype disruption [15]. These data suggest that the mechanism of disease could be linked to abnormal CL-L1 secretion. The fact that we observe continuous expression of CL-L1 in E13.5 embryos and P0 pups in the mandibular epithelium could indicate there is an additional role for maintaining cellular adhesion even after NCC migration is complete; further data are required to prove this hypothesis.
In summary, we have described here a new gene, COLEC10, that when mutated causes 3MC syndrome. Further mutations identified in COLEC11 and MASP1/3 further confirm clinical suspicions of disease in several 3MC patients but leaves a sizeable proportion (60%) without molecular confirmation and implicate one or more further genes. We propose that the lectin complement pathway acts as a chemottractant to guide and possibly to maintain cNCC adhesion. We believe that in future more genes linked to the lectin complement pathway and with roles in cellular adhesion and guidance will be found to be mutated in 3MC syndrome patients and other craniofacial conditions.
Patients and families samples were screened by whole-exome sequencing, including the proband and both parents when available. In each case, genomic DNA was enriched for exonic regions using the SureSelect All Exon 50Mb Targeted Enrichment kit (targeting 202,124 exons from 20,718 genes) from Agilent Technologies, according to the manufacturer's protocol. Captured libraries were sequenced on an Illumina HiSeq 2000 instrument using Illumina sBot clustering and HiSeq chemistries v1.0, under a paired-end 100-bp read-length protocol, with four samples per flow cell lane to achieve minimum median coverage of 60×. All exomes for COLEC11, COLEC10 and MASP1/3 have a coverage of at least x15. For specific exonic coverage of 3MC family 19 see S1 Methods Table. The variant annotation and interpretation analyses were generated through the use of Ingenuity Variant Analysis software version 3.1.20140902 from Ingenuity Systems. For the recessive model, homozygous/compound heterozygous variants in the affected individual were retained. Intronic and exonic synonymous variants were filtered out; exonic and splice variants (up to 2 base pairs into intron or predicted pathogenic on MaxEntScan) with a public databases (ExAC, 1000 Genomes and ESP Exomes) frequency <0.01% (3MC phenotype) were retained. All disease causing variants (COLEC10) were validated by Sanger sequencing. Filtering pipelines for variants, ingenuity and a final list of all variants identified are presented in S2 Methods Table, S3 Methods Table and S4 Methods Table.
HEK293 and HeLa cells were cultured in DMEM (Invitrogen) supplemented with 10% (v/v) foetal bovine serum and incubated in humidified 5% CO2 at 37oC.
An agarose spot assay was used to assess chemotactic invasion potential of CL-L1. Briefly, a 2% (w/v) solution of low-melting point agarose (Invitrogen) in phosphate-buffered saline was boiled and when the solution cooled to around 50oC it was mixed 1:1 with solutions of PBS, bovine serum albumin (BSA), recombinant CL-K1 (Abnova, H00078989-P01) and/or recombinant CL-L1 (Abnova, H00010584-P01). 10μL of the agarose-protein mix was then spotted onto the wells of plastic tissue culture plates, allowed to polymerise at room temperature for around 10 minutes and cells added. Cell migration and invasion was monitored at 37°C with 5% CO2 for around 48 hours using an Axiovert 135 microscope (Zeiss) equipped with a motorized stage that captured 1 image per 15 minutes (Volocity software v6.3, PerkinElmer). Migration and invasion was quantified using ImageJ software by measuring the area within the agarose-protein discs that had been occupied by cells (Fig 3A).
Patient mutations c.25C>T,p.Arg9Ter; c.226delA,p.Gly77Glufs*66 and c.528C>G;p.Cys176Trp were introduced into a plasmid encoding wild-type human CL-L1 (pCMV6-XL5-COLEC10; OriGene, SC303774) using QuickChange II Site-Directed Mutagenesis kit (Agilent) with hCOLEC10Arg9Ter, hCOLEC10Gly77Glufs*66 and hCOLEC10Cys176Trp primers (S5 Methods Table). hCOLEC10WT, hCOLEC10Arg9Ter, hCOLEC10Gly77Glufs*66 and hCOLEC10Cys176Trp plasmids were complexed with 25kDa branched polyethylenimine (Sigma) and transfected into HEK293 and HeLa cells. A negative control with untransfected HEK293 and HeLa cells was used to show CL-L1 expression was not innate cell endogenous expression.
Western blot was performed using standard protocols. Briefly, 48 hours post-transfection cell-culture supernatant was collected and clarified by centrifugation at 13,000 rpm for 10 minutes and pellet discarded. To obtain cell extract, cells were lysed by incubating on ice with chilled cell extraction buffer (Invitrogen) supplemented with cOmplete, mini protease inhibitor cocktail (Roche) and 1mM phenylmethylsulfonyl fluoride (PMSF; Sigma) for 30 minutes with vortexing every 10 minutes. Cell extract was then clarified by centrifugation at 13,000 rpm for 10 minutes and pellet discarded. Proteins in supernatant and cell lysate were separated by SDS-PAGE (Tris-Acetate 4–15% gels, Invitrogen), blotted onto nitrocellulose membranes (Bio-Rad) and detected using primary antibodies against CL-L1 (Generon; CSB-PA896556LA01HU, 2μg/mL) and GAPDH (Generon; CSB-PA00025A0Rb, 2μg/mL) with HRP-conjugated secondary antibodies (Dako). Blots were developed with enhanced chemiluminescence (Pierce).
To obtain cell extract for ELISA, cells were lysed by incubating on ice with chilled ELISA cell extraction buffer (100mM Tris; pH7.4, 150mM NaCl, 1mM EGTA, 1mM EDTA, 1% Triton X-100 and 0.5% sodium deoxycholate) supplemented with cOmplete, mini protease inhibitor cocktail (Roche) and 1mM PMSF (Sigma) for 30 minutes with vortexing every 10 minutes. Cell extract was then clarified by centrifugation at 13,000 rpm for 10 minutes and pellet discarded.
For cell immunofluorescence ATDC5 cells were fixed with cold methanol -20°C, washed with PBS and blocked for 1 hour with 1% BSA. Cells were incubated overnight with the following antibodies and concentrations: CL-L1/100 (Novus Biologicals H00010584-M01), CL-K1 (Novus Biologicals H00010584-M01), Laminin (Abcam, ab11575). Cells were washed with PBS and incubated for 1 hour with Mouse or Rabbit Alexa Fluor 488 and 568 secondary antibodies (1/1000) (ThermoFisher). E18.5 mouse embryos were harvested and fixed in 4% paraformaldehyde overnight at 4°C, dehydrated and embedded in paraffin. 10μm sections were cut. Slides were rehydrated and blocked with 5% BSA with 10% of sheep serum. The samples were incubated with a rabbit in house made CL-L1 primary antibody (1/100) overnight at 4°C, washed in PBS and developed with a Horseradish peroxidase conjugated secondary antibody and diaminobenzidine staining.
1000 Genomes, http://www.1000genomes.org
Ensembl Genome Browser, http://www.ensembl.org/index.html
ExAC Browser, http://exac.broadinstitute.org/
OMIM, http://www.omim.org/
PolyPhen-2, http://genetics.bwh.harvard.edu/pph2/
SIFT, http://sift.bii.a-star.edu.sg/
All work involving human subject research was approved by the UCL-ICH/Great Ormond Street Hospital Research Ethics Committee (08/H0713/82) (REC reference 08/H0713/82, Protocol number HBD2008v1).All patients and families included in this work have given written consent for the use of their biological samples for research purposes under the HT act 2004 ethics committee. All animal work has been conducted under the UK Home Office regulation, Animals (Scientific Procedures) Act 1986 and was approved by the Home Office with Procedure Project License PPL number 70/7892.
|
10.1371/journal.ppat.1005119 | Plasmodium Infection Is Associated with Impaired Hepatic Dimethylarginine Dimethylaminohydrolase Activity and Disruption of Nitric Oxide Synthase Inhibitor/Substrate Homeostasis | Inhibition of nitric oxide (NO) signaling may contribute to pathological activation of the vascular endothelium during severe malaria infection. Dimethylarginine dimethylaminohydrolase (DDAH) regulates endothelial NO synthesis by maintaining homeostasis between asymmetric dimethylarginine (ADMA), an endogenous NO synthase (NOS) inhibitor, and arginine, the NOS substrate. We carried out a community-based case-control study of Gambian children to determine whether ADMA and arginine homeostasis is disrupted during severe or uncomplicated malaria infections. Circulating plasma levels of ADMA and arginine were determined at initial presentation and 28 days later. Plasma ADMA/arginine ratios were elevated in children with acute severe malaria compared to 28-day follow-up values and compared to children with uncomplicated malaria or healthy children (p<0.0001 for each comparison). To test the hypothesis that DDAH1 is inactivated during Plasmodium infection, we examined DDAH1 in a mouse model of severe malaria. Plasmodium berghei ANKA infection inactivated hepatic DDAH1 via a post-transcriptional mechanism as evidenced by stable mRNA transcript number, decreased DDAH1 protein concentration, decreased enzyme activity, elevated tissue ADMA, elevated ADMA/arginine ratio in plasma, and decreased whole blood nitrite concentration. Loss of hepatic DDAH1 activity and disruption of ADMA/arginine homeostasis may contribute to severe malaria pathogenesis by inhibiting NO synthesis.
| During a malaria infection, the vascular endothelium becomes more adhesive, permeable, and prone to trigger blood clotting. These changes help the parasite adhere to blood vessels, but endanger the host by obstructing blood flow through small vessels. Endothelial nitric oxide (NO) would normally counteract these pathological changes, but NO signalling is diminished malaria. NO synthesis is inhibited by asymmetric dimethylarginine (ADMA), a methylated derivative of arginine that is released during normal protein turnover. We found the ratio of ADMA to arginine to be elevated in Gambian children with severe malaria, a metabolic disturbance known to inhibit NO synthesis. ADMA was associated with markers of endothelial activation and impaired tissue perfusion. In parallel experiments using mice, the enzyme responsible for metabolizing ADMA, dimethylarginine dimethylaminohydrolase (DDAH), was inactivated after infection with a rodent malaria. Based on these studies, we propose that decreased metabolism of ADMA by DDAH might contribute to the elevated ADMA/arginine ratio observed during an acute episode of malaria. Strategies to preserve or increase DDAH activity might improve NO synthesis and help to prevent the vascular manifestations of severe malaria.
| Current estimates of world-wide mortality due to malaria range from 367,000 to 755,000 deaths per year, mostly in African children [1,2]. Prompt treatment with parenteral artesunate improves survival in children with severe malaria but mortality remains high in those presenting with complications such as coma or acidosis [3]. Development of effective therapies for these patients will require improved understanding of the pathophysiology of severe malaria.
Patients with severe malaria exhibit impaired endothelium-dependent vasodilation [4] and reduced nitrite and nitrate concentrations in plasma and urine [5], indicating decreased nitric oxide synthesis. Impaired NO signalling has been implicated in microcirculatory dysfunction [6], loss of blood-brain barrier integrity [7,8] and cytoadherence of infected erythrocytes to the vascular endothelium in mice [9]. Similar pathology has been directly observed in human malaria, but the importance of NO signalling in these processes is less certain [10–13]. NO production by nitric oxide synthase (NOS) is dependent in part on the relative bioavailability of arginine, the NOS substrate, and asymmetric dimethylarginine (ADMA), an endogenous NOS inhibitor released during hydrolysis of proteins that have been methylated by protein arginine methyltransferase [14–16] (Fig 1). By inhibiting NOS, ADMA not only causes vasoconstriction, increased blood pressure, increased systemic vascular resistance and decreased forearm blood flow in vivo [17–19], but also affects adhesion, inflammation, thrombosis, barrier integrity, motility, growth and repair in vitro [20–31]–endothelial functions that are relevant to the pathophysiology of malaria.
Dimethylarginine dimethylaminohydrolase 1 (DDAH1) metabolizes ADMA at a rate inversely proportional to arginine concentration [32] and thus stabilizes the ratio of ADMA to arginine when arginine levels vary [33,34]. In Gambian children, an intronic DDAH1 polymorphism is associated with susceptibility to severe malaria [35], raising the possibility that DDAH1 might be functionally linked to disrupted ADMA/arginine homeostasis and impaired NO synthesis in severe malaria. In this study, we identify dysregulation of ADMA/arginine homeostasis in Gambian children with severe malaria and hypothesize that ADMA clearance is impaired by hepatic DDAH1 inactivation. To test this hypothesis, we infected mice with P. berghei ANKA and assessed changes in DDAH1 expression, protein levels and activity in hepatic tissue, a major site of ADMA metabolism [29,36–38].
We determined ADMA and arginine concentrations in blood plasma obtained from Gambian children with severe or uncomplicated malaria at initial presentation and 28 days later. Healthy afebrile aparasitemic Gambian children served as an additional control group. Baseline characteristics of the study populations are presented in Table 1.
Plasma ADMA was lower in children with severe malaria (median [IQR]: 0.40 [0.30–0.51] μmol/L) or uncomplicated malaria (0.40 [0.33–0.47] μmol/L) compared to healthy children (0.61 [0.56–0.69] μmol/L, p < 0.0001 vs. uncomplicated; p < 0.0001 vs. severe; Fig 2 and Table 2). ADMA remained low at the 28-day follow-up visit for patients recovered from malaria. Plasma arginine was profoundly depleted in children with severe malaria compared to children with uncomplicated malaria or healthy children (severe malaria: 31.7 [23.0–40.6] μmol/L; uncomplicated malaria: 45.0 [35.4–55.7] μmol/L, p < 0.0001 vs severe; healthy: 88.7 [79.3–102.5] μmol/L, p < 0.0001 vs severe; Fig 2 and Table 2). By the 28-day follow up visit, plasma arginine concentration increased to 56.7 [42.1–78.9] μmol/L among children who recovered from severe malaria and to 70.8 [58.6–85.1] μmol/L among children who recovered from uncomplicated malaria (p < 0.0001 vs acute, Fig 2), but remained lower than the arginine concentration observed in healthy Gambian children (88.7 μmol/L; p < 0.0001 for either comparison).
ADMA is a competitive inhibitor of NOS, and the ratio of ADMA to arginine determines NOS activity [39]. The ratio of ADMA to arginine was elevated among children with severe malaria (13.5 [11.2–17.1] ×10−3) compared to children with uncomplicated malaria (8.8 [7.2–11.6] ×10−3, p < 0.0001) or healthy children (6.7 [5.8–8.2] ×10−3, p < 0.0001, Fig 2). After recovery from severe malaria, the ADMA/arginine ratio returned to the level observed in healthy Gambian children (recovered from severe malaria: 7.4 [5.9–10.1] ×10−3, p < 0.0001 vs. acute; p = 0.25 vs. healthy children; Fig 2 and Table 2). Thus elevation of the ADMA/arginine ratio appears to be an acute metabolic disturbance associated with a symptomatic episode of malaria (modeled in S2 Fig).
In healthy Gambian children, plasma ADMA concentration was correlated with plasma arginine concentration (r = 0.43, p < 0.05); children with lower arginine tended to have lower ADMA (S1 Fig). In children with uncomplicated malaria, the correlation was stronger (r = 0.59, p < 0.0001) and the slope of the linear regression was steeper (p < 0.0001 vs healthy; S1 Fig). In children with severe malaria, the correlation was stronger (r = 0.77, p < 0.0001) and the slope of the linear regression was steeper still (p < 0.0001 vs uncomplicated; S1 Fig). At the day 28 follow up visit, the relationship between ADMA and arginine (ie, the slopes of the linear regressions) had returned to normal (S1 Fig), though the absolute levels of ADMA and arginine remained lower than in health children.
Lactate is a biomarker of impaired tissue perfusion that is associated with mortality from severe malaria in children [40–43]. In our study, lactate (median [IQR]) was elevated in children with severe malaria (5.0 [3.2–7.0] mmol/L) compared to children with uncomplicated malaria (2.8 [2.2–4.1] mmol/L, p < 0.0001, Table 1). Lactate correlated positively with ADMA among children with severe malaria (r = 0.34, p = 0.004, S3 Fig), implying that tissue perfusion was impaired among those with higher ADMA levels. Lactate did not correlate with arginine (r = 0.16, p = 0.20, S3 Fig) but did correlate with the ADMA/arginine ratio (r = 0.28, p = 0.02, Table 3). In multiple linear regression analysis using ADMA and arginine as explanatory variables, ADMA was positively related to lactate (β = 0.758, p = 0.002) while arginine was negatively and non-significantly related to lactate (β = -0.393, p = 0.09; Table 4).
We measured soluble vascular cell adhesion molecule (sVCAM) as a biomarker of endothelial activation. Plasma sVCAM was elevated in children with severe malaria at the time of admission compared to children with uncomplicated malaria or healthy children (severe malaria admission: 1266 [828–1798] ng/mL; acute uncomplicated: 905 [726–1313] ng/mL, p < 0.05 vs severe, healthy children: 905 [773–1078] ng/mL, p < 0.05 vs severe, Table 1). sVCAM returned to normal at day 28 among children who had severe malaria (841 [655–1147] ng/mL, p < 0.001 vs admission). In contrast, the sVCAM level of children with uncomplicated malaria was similar to the level measured in healthy children (p = 0.84). Endothelial activation appears to be a distinctive feature of acute severe malaria.
Soluble VCAM was positively correlated with plasma ADMA in severe malaria patients (r = 0.60, p < 0.0001, Table 3 and S3 Fig). This observation suggests that ADMA, a NOS inhibitor, is associated with endothelial activation and release of sVCAM into circulation. sVCAM was also positively correlated with arginine (r = 0.59, p < 0.0001, Table 3 and S3 Fig). In multiple linear regression analysis, ADMA (β = +0.413, p = 0.02) was more significantly related to sVCAM levels than was arginine (β = +0.332, p = 0.06; Table 4).
Haptoglobin becomes depleted from plasma during acute intravascular hemolysis [44]. Haptoglobin was low at the time of admission in children with severe malaria or uncomplicated malaria (severe: 0 [0–4.1] mg/dL; uncomplicated: 1.3 [0–49.3] mg/dL), and increased by the 28-day follow up visit (severe day 28: 13.8 [1.2–44.3] md/dL, p < 0.0001 vs admission; uncomplicated day 28: 18.7 [0.2–59.4] mg/dL, p < 0.07 vs admission). Admission haptoglobin values, but not day 28 values, were significantly lower than in healthy children (44.5 [15.6–79.9] mg/dL).
Because haptoglobin was undetectable in many children with severe malaria, we analyzed the correlation with ADMA and arginine using Spearman’s method. The correlation between haptoglobin and ADMA was weak (r = -0.29, p = 0.02, Table 3 and S3 Fig), and weaker still with arginine (r = -0.24, p = 0.06, Table 3 and S3 Fig). There were however, moderate negative correlations with hemoglobin, a measure of anemia that may be partially reflective of hemolysis (correlation with Hb and ADMA: r = -0.44, p <0.0001; Hb and Arg r = -0.32, p = 0.004; Table 3 and S3 Fig). Multiple linear regression analysis again revealed that this correlation was primarily due to the association of hemoglobin with ADMA (β = -3.039, p = 0.003) and not with arginine (β = +0.404, p = 0.66; Table 4).
P. falciparum histidine-rich protein 2 (PfHRP2), a circulating marker of parasite biomass, was higher in children with severe malaria compared to uncomplicated malaria (severe: 249 [133–605] ng/mL vs uncomplicated: 118[54–226] ng/mL, p = 0.0001, Table 1). However, PfHRP2 was not correlated with ADMA or arginine (Table 3; S3 Fig).
To determine whether P. berghei infection was associated with systemic changes in ADMA and arginine, we analyzed plasma from P. berghei ANKA-infected mice. Similar to Gambian children with malaria, plasma ADMA concentrations were lower in infected animals compared to uninfected controls on day 6 post-inoculation (0.44 [0.40–0.49] vs. 0.59 [0.54–0.63] μmol/L, p < 0.0001, Fig 3A). Arginine concentrations were also lower in infected mice compared to uninfected controls (46.7 [39.2–53.3] vs. 78.1 [66.5–101.9] μmol/L, p<0.0001, Fig 3B). Arginine decreased to a greater extent than ADMA, resulting in an increased ratio of ADMA to arginine among infected mice (9.83 [8.65–12.49] ×10−3 vs. 7.10 [5.94–8.81] ×10−3, p<0.0001, Fig 3C). The murine findings recapitulated our observations in Gambian children with malaria.
Whole blood nitrite is reflective of NOS activity [45]. We determined nitrite concentrations in whole blood samples from P. berghei ANKA-infected mice and uninfected controls using a gas-phase chemiluminescent assay. Whole blood nitrite was decreased in infected mice (0.36 [0.28–0.45] μmol/L) compared with uninfected controls (0.49 [0.41–0.71] μmol/L, p = 0.0001, Fig 3D), suggesting that P berghei ANKA infection causes a decrease in systemic NO production in mice.
DDAH1 is highly active in the liver and plays a key role in regulating circulating levels of ADMA [29,33,36–38]. To determine whether severe malaria affects hepatic DDAH1 function, we assessed hepatic Ddah1 gene expression and protein levels in liver tissue from C57BL/6 mice 6 days after inoculation with P. berghei ANKA. Using quantitative RT-PCR to assess hepatic Ddah1 mRNA transcript number relative to Gapdh, we found that Ddah1 gene expression was not changed by P. berghei ANKA infection (median [IQR] fold change: 1.1 [1.0–1.1], p = 0.07, Fig 3E). In contrast, Western blot analysis revealed a decrease in hepatic DDAH1 protein from P. berghei ANKA-infected mice compared to uninfected control mice (median [IQR] fold change: 0.33 [0.28–0.46], p < 0.0001, Fig 3F). These data demonstrate that P. berghei ANKA infection decreases DDAH1 protein abundance by a post-transcriptional mechanism.
To determine the functional impact of hepatic DDAH1 inactivation by Plasmodium infection, we quantified ADMA clearance in liver homogenates by measuring the rate of de novo citrulline production in the presence of saturating concentrations of ADMA substrate (assay validation presented in S5 Fig). Hepatic ADMA clearance was lower in mice infected with P. berghei ANKA compared with uninfected controls (infected: 3.83 [3.22–4.19] nmol citrulline × mg protein-1 × hr-1 vs uninfected control: 6.48 [5.23–7.49] nmol citrulline × mg protein-1 × hr-1, p < 0.0001, Fig 3G). To assess the impact of P. berghei infection on hepatic ADMA metabolism, we determined intracellular ADMA concentrations in PBS-perfused liver samples from infected and control mice. ADMA was increased in liver tissue from infected mice compared with uninfected controls (infected: 126.5 [88.9–198.2] pmol/mg protein vs uninfected control: 49.7 [35.3–77.8] pmol/mg protein, p < 0.0001, Fig 3H).
We calculated the correlation between hepatic DDAH activity and hepatic ADMA concentration in healthy mice and found a positive correlation (r = 0.46, p = 0.01), i.e., hepatic DDAH activities were greater in mice that had higher tissue levels of ADMA. This may reflect induction of DDAH activity in response to tissue levels of ADMA. Among P. berghei-infected mice, the correlation was similar (r = 0.42, p = 0.04) though the tissue levels of ADMA were higher, and the DDAH activities were lower than in uninfected mice (S4 Fig and S1 Table). The partial correlation coefficient between hepatic DDAH activity and hepatic ADMA concentration, accounting for infection status, remained positive (rpart = 0.34, p = 0.01). We also calculated the correlation between hepatic DDAH activity and plasma ADMA/Arginine ratio in mice, and found a negative trend, i.e., mice with lower DDAH activity tended to have higher ADMA/Arginine ratios in plasma (rpart = -0.23, p = 0.09; S4 Fig and S1 Table).
We have analyzed ADMA and arginine concentrations in plasma from children with severe malaria to determine whether malaria is associated with disruption of ADMA/arginine homeostasis. We found that children with acute severe malaria have uncompensated hypoargininemia, i.e., low arginine with an elevated ADMA/arginine ratio. The hypoargininemia persisted over the 28 days of follow-up, while the ratio of ADMA to arginine returned to normal. We interpret this as a transient inability to metabolize ADMA at a sufficient rate to compensate for low arginine during acute infection. Although plasma ADMA levels were below normal in patients with severe malaria, ADMA was positively correlated with lactate, a biomarker of severity, and sVCAM, a biomarker of endothelial activation, suggesting that higher ADMA levels are associated with adverse pathophysiologic changes. DDAH1 metabolizes ADMA, so we examined changes in DDAH1 activity in mice infected with a Plasmodium berghei ANKA, a model of severe malaria. P berghei infection caused inactivation of hepatic DDAH1, accumulation of intracellular ADMA in liver tissue, elevation of the ADMA to arginine ratio in plasma, and decreased levels of nitrite in blood. Although these findings in the mouse model cannot be directly extrapolated to human malaria, it raises the possibility that the elevated ADMA/arginine ratio observed in children with severe malaria could be due in part to inactivation of DDAH1.
Our results extend upon a previous report of ADMA and arginine levels in Tanzanian children. Weinberg et al. observed ADMA/arginine ratios of 13.2 [11.1–16.4] ×10−3 in cerebral malaria, 12.3 [10.0–15.1] ×10−3 in non-cerebral severe malaria, 12.6 [10.7–15.1] ×10−3 in moderately severe malaria and 7.1 [5.8–9.0] ×10−3 in healthy children [46]. These values are consistent with the ADMA/arginine ratios observed in our study. We found the ADMA/arginine ratio to be significantly greater in children with severe malaria compared to children with uncomplicated malaria, while Weinberg et al found no differences among the ADMA/arginine ratios of cerebral malaria, non-cerebral severe malaria and moderately severe malaria groups. This discrepancy may be explained by the increased severity of the moderately severe malaria group in Weinberg, et al. that differed from our uncomplicated group by the inclusion of patients who could not tolerate oral medication [46]. As a result, these children may have had greater dietary insufficiency of arginine and arginine precursors than our group of uncomplicated malaria patients. Although the plasma arginine concentration was lower in Gambian children with severe malaria compared to the Tanzanian children with cerebral malaria, the rise from admission to day 28 in the Gambian children (31.7 to 56.7 umol/L, an increase of 25 umol/L) was similar to the rise from admission to day 7 in the Tanzanian children (45 umol/l to 70 umol/L, an increase of 25 umol/L).
In both Gambian and Tanzanian children with severe malaria, ADMA and the ADMA/Arg ratio were each correlated with lactate. This could be mediated through the vasoconstrictive or pro-adhesive effects of ADMA on vascular endothelium especially in the setting of hypoarginemia, with subsequent impairment of tissue perfusion leading to anaerobic glycolysis and lactate generation. Inter-individual differences in hepatic blood flow could also be responsible for the strong correlation between ADMA and lactate, since the clearance of each is dependent on hepatic perfusion. Impaired perfusion of liver tissue has been observed in adults with severe malaria [47] and could limit hepatic clearance of plasma ADMA [33,36–38].
In both Gambian and Tanzanian children with severe malaria, ADMA was correlated with biomarkers of endothelial activation (sVCAM and Angiopoietin-2, respectively). This could be through direct effects of ADMA on endothelial cells [48] or via the pro-adhesive effects of ADMA on circulating immune cells that interact with endothelium [49,50].
P. falciparum histidine-rich protein 2 (PfHRP2) has been previously assessed as a quantitative marker of parasite biomass [51,52]. PfHRP2 did not correlate significantly with ADMA, arginine or the ADMA/arginine ratio among children with severe or uncomplicated malaria (Table 3). Our findings are in agreement with the prior study [46] and together suggest that in children host ADMA metabolism is not determined by parasite biomass.
Arginine depletion and disruption of ADMA/arginine homeostasis have also been observed in adults with moderately severe and severe malaria [53]. In contrast to African children, Indonesian adults with severe malaria demonstrated elevated ADMA [53]. The apparent discrepancy in plasma ADMA in children and adults with severe malaria might be explained by differences in severe malaria pathophysiology observed in older versus younger patients [54]. Changes in plasma ADMA also differ between children and adults with acute sepsis; compared to age-matched healthy controls, ADMA was decreased in pediatric sepsis, but studies of adult sepsis found ADMA to be either unchanged or increased [55–58].
Disruption of ADMA/arginine homeostasis in children with severe malaria could be due to increased protein methylation, accelerated proteolysis of methylated proteins or impaired clearance of free ADMA. One might expect plasma ADMA to be directly elevated during a severe malaria infection, due to the combination of increased release of ADMA from erythrocytes undergoing hemolysis [59] and the impaired activity of hepatic DDAH that we present here. Instead, we observed lower plasma ADMA concentrations in both human and mouse malaria, consistent with prior measurements in children with malaria [46]. This could be due to increased uptake of ADMA from plasma into cellular compartments as has been observed in vitro after LPS, TNF or IL-1 stimulation [60]. In addition, plasma ADMA was strongly correlated with plasma arginine in our study, suggesting that arginine deficiency might lead to lower plasma ADMA. Mechanisms that could potentially link plasma ADMA to plasma arginine are inadequately understood, but might include the requirement for protein-incorporated arginine as the substrate for PRMTs that generate ADMA [16], upregulation of the cationic transporters that allow both ADMA and arginine to cross cell membranes [61,62], and negative feedback of arginine on DDAH activity [32].
DDAH1 is known to regulate ADMA/arginine homeostasis: heterozygous knock-out of DDAH1 in mice increased plasma ADMA, decreased NO-dependent vasodilation, and elevated blood pressure [63]. Conversely, transgenic over-expression of DDAH1 decreased plasma ADMA concentrations, increased urinary nitrites/nitrates and decreased blood pressure [64]. A second DDAH isoform (DDAH2) has been identified [65], but in contrast to DDAH1, suppression of DDAH2 expression did not result in altered plasma ADMA concentrations [34]. The liver expresses DDAH1 [65] and metabolic tracer studies identified the liver as a major site for clearance of circulating ADMA [37]. Induction of DDAH1 expression in liver significantly lowered plasma ADMA [33]. Endothelial cell-specific knock-out of DDAH1 revealed hepatic DDAH1 expression not only in hepatic endothelial cells but also in hepatocytes [29,66], which appear to be primarily responsible for systemic ADMA metabolism. Moreover, DDAH1 may regulate the release of ADMA from non-endothelial cell sources that affect local ADMA levels and vascular function. Patients with hepatic failure had elevated plasma levels of ADMA [38,67,68] that decreased after liver transplantation [67]. Conversely, patients with acute rejection of their liver graft had elevated ADMA compared to patients without episodes of rejection [67]. Taken together, results from human patients and animal studies implicate hepatic DDAH1 as a key regulator of circulating ADMA. In severe malaria, renal insufficiency [17], in addition to the hepatic DDAH1 dysfunction we present here, could contribute to dysregulation of ADMA/arginine homeostasis.
Plasmodium infection appears to accelerate the degradation of DDAH1 protein in hepatic tissue. Oxidative stress is a potential trigger of DDAH degradation that is present during malaria infection [69,70]. Overexpression of the p22phox subunit of NADPH oxidase in smooth muscle cells increased oxidative stress, decreased DDAH protein levels, decreased DDAH activity and caused accumulation of both intracellular and extracellular ADMA [71]. Treatment of p22phox-transfected smooth muscle cells with the proteasome inhibitor epoxomicin raised DDAH protein concentrations and reduced intracellular ADMA, demonstrating that oxidative modification of DDAH protein may target it for degradation by the proteasome. Plasmodium infection causes oxidative stress in liver tissue of mice [72], which could be sufficient to accelerate the degradation of DDAH1 in liver endothelium.
The liver may be exposed to reactive oxygen species generated by the increased populations of neutrophils and pigment-laden monocytes found in the hepatic vasculature during malaria infection [72,73]. Increased cell-free heme due to hemolysis promotes neutrophil infiltration and resulting liver damage [72], raising the possibility that hemolysis may contribute to hepatic DDAH dysfunction. Hemolysis may also result in direct release of ADMA into circulation. Human erythrocytes contain total (free plus protein-incorporated) ADMA concentrations in the range of 47.85 ± 1.68 μmol/L, extrapolated from a concentration of 15.95 ± 0.56 μmol/L reported for hydrolysates of erythrocyte samples diluted 1:3 in water [59]. Rat erythrocytes contain a similar concentration of total ADMA, estimated to be 40.6 ± 7.2 μmol/L [74]. Following hemolysis, methylated erythrocyte proteins are exposed to proteases that disproportionately release ADMA relative to arginine [74,75]. Thus hemolysis could both increase ADMA release and inhibit ADMA clearance by promoting DDAH degradation.
Elevation of ADMA relative to arginine favors NOS inhibition because ADMA is a competitive inhibitor of NOS [14]. ADMA also competes with arginine for cellular uptake, which could limit arginine availability for NO synthesis [61]. The intracellular concentration of ADMA in endothelial cells is approximately 10-fold higher than extracellular levels, reaching a concentration of 3–5 μmol/L which is near the Ki of eNOS [76,77]. Even small elevations in extracellular ADMA concentration to 2 umol/L had profound effects on brain NOS activity and gene expression profiles of endothelial cells in culture or in mice [78,79]. Impaired NO synthesis has been implicated in impaired vasoregulation, loss of blood-brain barrier integrity and cytoadherence of parasitized erythrocytes to the vascular endothelium during severe malaria. In a mouse model, treatment with an NO-donor improved cerebral microcirculation, reduced cerebral hemorrhages and prevented blood-brain barrier break-down [6,7]. NO synthase inhibition by ADMA downregulates tight junction protein expression [24], which may explain the beneficial effect of NO on endothelial barrier integrity. Impaired NO synthesis is associated with increased adhesion molecule expression [80] and L-NAME (a synthetic NOS inhibitor) increased cytoadherence of parasitized red blood cells to vascular endothelial cells in vitro [9]. Thus, impaired NO signaling may contribute to microhemorrhage, vascular leak and sequestration of parasitized red blood cells observed in children with fatal cerebral malaria [13]. Taken together, these findings suggest that disruption of ADMA/arginine homeostasis could contribute to severe malaria pathogenesis by inhibiting NO synthesis.
Therapeutic strategies that preserve or enhance DDAH activity during Plasmodium infection are needed to establish a causal relationship between DDAH degradation and disruption of ADMA/arginine homeostasis. While restoring ADMA/Arginine homeostasis might be necessary to improve endothelial NO synthesis, it might not be sufficient: impaired endothelial NO synthesis is likely to be limited by arginine deficiency and oxidation of tetrahydrobiopterin. NO that is produced will have a limited half-life due to reactions with cell free hemoglobin, superoxide, and other radicals that have increased abundance during malaria infection.
In summary, through clinical observational studies of Gambian children and controlled experiments mice, we have identified hepatic DDAH dysfunction as a potential mechanism disturbing ADMA/Arginine homeostasis and limiting nitric oxide synthesis in severe malaria.
Patient enrollment and sample collection were conducted following ethical review and approval by the Gambian Government/MRC Joint Ethics Committee and the Ethics Committee of the London School of Hygiene & Tropical Medicine (SCC 670, SCC 1002, SCC 1003, SCC 1077 & SCC 1113 [healthy children]). Analysis of plasma samples and de-identified clinical data at the National Institutes of Health was exempted from further ethical review by the NIH Office of Subjects Research (Exemption #5161). Written informed consent was provided by a parent or guardian on behalf of children enrolled in the study. Families who declined to participate were provided standard medical care.
Animal studies (described in detail in the supplementary information, S1 Methods) were specifically approved by the National Institutes of Allergy and Infectious Diseases (NIAID) Animal Care and Use Committee (ACUC) under the protocol identification LMVR 18E. The NIAID ACUC complies with the U.S. Government Principles for the Utilization and Care of Vertebrate Animals, the Public Health Service (PHS) Policy on Humane Care and Use of Laboratory Animals, and the Animal Welfare Act.
Children with severe malaria or uncomplicated malaria were enrolled at health centers in a peri-urban area around Fajara, The Gambia as previously described [81]. Enrollment sites included the Royal Victoria Teaching Hospital, the Brikama Health Centre, the MRC Fajara Gate Clinic and the Jammeh Foundation for Peace Hospital in Serekunda.
Acute uncomplicated malaria was defined as asexual P. falciparum parasitemia of >5000 parasites/μl detected by slide microscopy with an episode of fever (temperature >37.5°C) within the previous 48 hrs and the absence of severe criteria. Acute severe malaria was defined as parasitemia of >5000 parasites/μl, a history of fever and one or more of the following: severe anemia (Hb < 6g/dl), severe acidosis (serum lactate >7 mmol/L), cerebral malaria (Blantyre coma score 2 or less in the absence of hypoglycemia or hypovolemia with the coma lasting for at least 2 hrs), and severe prostration (inability to sit unsupported in children >6 months or inability to suck in children <6 months). Patients with severe malaria were admitted and treated with quinine, and patients with uncomplicated malaria were treated with chloroquine plus sulfadoxine-pyrimethamine according to Gambian Government Treatment Guidelines [82].
Ninety-six children with severe malaria and 102 children with uncomplicated malaria were enrolled during the 2005–2008 malaria seasons. Four milliliters of blood were collected in heparinized vacutainers (BD) at the time of initial presentation and at a follow-up visit 28 days later. Blood samples were immediately refrigerated, placed on ice for transport, and processed within 2 hours of collection. Plasma was frozen at -80°C. Three patients with severe malaria died and 65 completed follow-up visits; plasma sample were available from 47 of them. Eighty-five patients with uncomplicated malaria completed follow-up visits; plasma samples were available from 65 of them. Thirty-one healthy, afebrile, aparasitemic Gambian children of similar ages were also studied at a single visit.
P. falciparum parasitemia was determined by bright-field microscopy of giemsa-stained blood smears. 50 fields were counted at high power. Full blood counts were obtained with an automated instrument (Clinical Diagnostics solutions, Inc., Fort Lauderdale, FL, USA). Lactate was measured with a handheld Lactate Pro device (Arkray, Edina, MN, USA). Soluble VCAM (sVCAM) and plasma haptoglobin concentrations were determined by ELISA (sVCAM: R&D Systems, Minneapolis, MN, USA; haptoglobin: Alpco, Salem, NH, USA) according to the manufacturer’s instructions.
P. falciparum HRP2 was measured in duplicate in plasma by ELISA (Cellabs) according to the manufacturer’s instructions. A standard curve was constructed using serial dilutions of the PfHRP2 standard and run with every plate. Laboratory staff were unaware of clinical status of the subjects.
Each plasma sample was diluted in PBS containing NG-monoethyl-L-arginine (MEA) as an internal standard before undergoing solid-phase extraction (Oasis MCX 96-well μElution Plate, Waters Corporation, Milford, MA, USA). The eluted cationic amino acids were dried, resuspended in water, and derivatized with ortho-phthalaldehyde (OPA) in 3-mercaptopropionic acid. Derivatized samples were separated by reverse-phase liquid chromatography over a 1×100 mm C18(2) column (Phenomenex, Torrance, CA, USA) and fluorescence detected at excitation and emission wavelengths of 340nm and 455nm, respectively. Concentrations were determined by integrating peak area with reference to the internal standard (MEA) and daily external standards (Arg, ADMA, and MEA).
Data are expressed as median and inter-quartile range (IQR). Statistical analyses were performed with GraphPad Prism 6.02 software and the R computing environment. P-values of less than 0.05 were considered significant. Mann-Whitney test was used to compare median values between healthy children and children with uncomplicated or severe malaria. Wilcoxon matched-pairs signed rank test was used to compare acute and recovery values in children with uncomplicated or severe malaria. Correlations were calculated on transformed data using Pearson’s correlation test, except in the case of correlation with haptoglobin in which Spearman’s method was used. When two separate groups were combined, partial correlations were calculated. Multiple linear regression analysis was used to assess the independent contributions of ADMA or arginine to correlations with hemoglobin, HRP2, sVCAM, or lactate. In these linear models, arginine and ADMA were the explanatory variables, and the direction, strength and significance of the association was assessed by the beta value and p-value. Mann-Whitney test was used to compare values between infected mice and uninfected controls. Data are available in supplemental files S1 Data and S2 Data.
|
10.1371/journal.ppat.1006237 | Nematode neuropeptides as transgenic nematicides | Plant parasitic nematodes (PPNs) seriously threaten global food security. Conventionally an integrated approach to PPN management has relied heavily on carbamate, organophosphate and fumigant nematicides which are now being withdrawn over environmental health and safety concerns. This progressive withdrawal has left a significant shortcoming in our ability to manage these economically important parasites, and highlights the need for novel and robust control methods. Nematodes can assimilate exogenous peptides through retrograde transport along the chemosensory amphid neurons. Peptides can accumulate within cells of the central nerve ring and can elicit physiological effects when released to interact with receptors on adjoining cells. We have profiled bioactive neuropeptides from the neuropeptide-like protein (NLP) family of PPNs as novel nematicides, and have identified numerous discrete NLPs that negatively impact chemosensation, host invasion and stylet thrusting of the root knot nematode Meloidogyne incognita and the potato cyst nematode Globodera pallida. Transgenic secretion of these peptides from the rhizobacterium, Bacillus subtilis, and the terrestrial microalgae Chlamydomonas reinhardtii reduce tomato infection levels by up to 90% when compared with controls. These data pave the way for the exploitation of nematode neuropeptides as a novel class of plant protective nematicide, using novel non-food transgenic delivery systems which could be deployed on farmer-preferred cultivars.
| Plant parasitic nematodes (PPN) reduce crop plant yield globally, undermining food security. Many of the chemicals used to kill these parasites are non-specific and highly toxic, and are being phased out of general use through governmental and EU regulation. The withdrawal of these chemicals is beneficial to the environment, but limits our ability to protect crops from infection. Efforts must now focus on developing environmentally safe PPN controls. PPNs can absorb various molecules directly from the environment into their nervous system, including peptides and proteins. Here we profiled the feasibility of using PPN neuropeptides, small signalling molecules, to interfere with normal PPN behaviour. We exposed PPNs to a variety of neuropeptides, and found that they could interfere with behaviours that are important to host-finding and invasion. We then developed soil-dwelling microbes that could generate and secrete these neuropeptides into the soil where the PPN infective juveniles are found. These transgenic microbes can protect host plants from infection, and represent a completely new approach to controlling PPNs in crop plants. Importantly, these neuropeptides appear to have no impact on other beneficial nematodes found in the soil.
| Plant parasitic nematodes (PPNs) are responsible for an estimated 12.3% reduction in crop yield each year, which equates to losses of around $US80 billion worldwide [1, 2]. Traditionally PPNs have been controlled through the use of fumigant, carbamate and organophosphate nematicides which are being withdrawn over environmental health and safety concerns, through global and EU level directives [3]. The fumigant methyl bromide was used extensively to control PPN infestations for more than 60 years, however the identification of ozone-depleting characteristics was recognised within the Montreal Protocol which aimed to eliminate methyl bromide use by 2010 [4]. Likewise, dibromochloropropane (DBCP), a highly lipophilic brominated organochlorine was first used as a nematicide in the mid 1950’s before animal safety tests in the 1960’s demonstrated endocrine disrupting, and carcinogenic properties, alongside an increased incidence of developmental defects following exposure. Later studies further demonstrated strong mutagenic properties, and workers at the Occidental Chemical plant in California, which produced DBCP, displayed significantly higher rates of spermatogenic abnormalities relative to the rest of the population [5]. The carbamate nematicide aldicarb also triggers toxicity in non-target organisms through disruption of cholinergic neurons. Initial withdrawal of use across the USA in 1990 was followed by re-introductions to counteract a serious shortfall in alternative control options in 1995; similar dispensations have been afforded to EC states. The extensive withdrawal of frontline nematicides has left a significant shortfall in our ability to control PPNs.
Transgenic approaches could provide a cost-effective means of PPN control. Much effort has focused on the development of in planta RNA interference (RNAi) to silence PPN genes necessary for successful parasitism [6–9]. Whilst many such studies have shown promise, concerns surround the persistence of RNAi trigger-expressing traits. It remains to be established if DNA methylation and transcriptional silencing of double stranded (ds)RNA-expressing transgenes is an issue in plants other than Arabidopsis thaliana [10]. Efforts to inhibit PPN nutrient acquisition through transgenic expression of cystatins that inhibit intestinal protease activity have also proven an effective strategy [6]. The utility of peptide resistance traits has also been demonstrated [7], resulting in field level resistance and high target specificity [8]. Indeed, stacking peptide and cystatin resistance traits has proven extremely effective in plantain, triggering a 99% reduction in PPN infection levels at harvest, with a corresponding 86% increase in plantain yield [9].
Peptides have traditionally been viewed as poor drug candidates due to issues surrounding cellular uptake and half-life. However it has long been known that nematodes display an unusual neuronal uptake mechanism which is exploited by amphid dye-filling methods [11]. The amphid neurons assimilate exogenous peptides which subsequently accumulate in cells of the central nerve ring [11], where they can interact with available receptors.
Neuropeptides are highly enriched and conserved amongst nematodes, coordinating crucial aspects of physiology and behaviour [12–21]. The model nematode Caenorhabditis elegans encodes at least 113 neuropeptide genes, producing over 250 mature neuropeptides [16]. It is thought that this neurochemical diversity underpins the wide array of complex behaviours which are found within such neuroanatomically simple animals [16, 22]. Many neuropeptides are known to be expressed within the anterior neurons of nematodes [16, 22–24], and it is likely that their cognate receptors are expressed in these or adjacent cells. The retrograde transport of exogenous peptides suggest that these receptors could be amenable to activation through signalling molecules following their uptake from the external environment. Conceptually, the mining of native neuropeptide complements for novel nematicides is an attractive prospect, based on the a priori assumption of bioactivity. An additional positive quality of neuropeptides is their characteristically high potency when acting on cognate receptors [13, 25–30]. Furthermore, the high degree of phylogenetic sequence conservation suggests that neuropeptides could represent broad-spectrum nematicides as they share significant sequence similarity within and between parasite species [17, 22, 31, 32]. Disrupting PPN behaviour through the dysregulation of native neuropeptide signalling could hinder the development of resistance traits anchored on target receptor mutation. Selective pressure drives the propagation of drug target variants which escape agonism / antagonism, or the development of enhanced efflux mechanisms [33, 34]. Conceptually, the development of resistance to neuropeptides which coordinate crucial aspects of PPN biology would seem less likely.
Nematode neuropeptide complements are organised into three broad groupings: i) the FMRF-amide Like Peptides (FLPs); the INSulin like peptides (INSs); and iii) the Neuropeptide-Like Proteins (NLPs). FLPs represent the most widely studied and best understood family, characterised by a C-terminal RFamide motif, and are known to coordinate motor and sensory function [14, 16, 22]. In particular, C-terminal amidation is necessary for biological function, and so precludes FLPs from most transgenic delivery methods. INSs coordinate and integrate sensory signals with developmental circuits [35] and they share characteristic domain organisation and tertiary structure with vertebrate insulin peptides [16, 36–40]. Specific proteolytic processing requirements suggest that INSs do not represent ideal candidates for transgenic delivery methods. The NLPs represent the least studied grouping of neuropeptides, comprising every neuropeptide that does not conform to the biosynthetic and structural characteristics of FLPs or INSs and encompassing multiple peptide families. Little is known about their function in nematodes, however many NLPs are expressed in anterior neurons and do not appear to require post-translational modifications [20, 24, 40–45], making them more amenable to generation and delivery by transgenic systems than FLPs or INSs. A key gap in assessing the potential of unamidated NLPs as nematicides is the lack of data on their bioactivity in PPNs.
Here we aimed to characterise the NLP complements in silico for two economically important PPNs that display different modes of infection and parasitism, M. incognita and G. pallida. Subsequently we aimed to screen NLPs for their ability to dysregulate the normal behaviour of infective stage juveniles (J2s) when applied exogenously and, simultaneously, to develop and assess novel transgenic delivery methods as next generation plant protection platforms.
Pro-peptide sequences of C. elegans NLPs predicted to be unamidated (no C-terminal glycine; uNLPs) were used as queries to conduct a BLASTp analysis of the predicted protein complements of both M. incognita and G. pallida [46, 47]. A total of four nlp genes encoding 25 predicted uNLPs were found within the G. pallida genome, and seven nlp genes encoding 28 predicted uNLPs within the M. incognita genome (Table 1).
Predicted uNLPs were synthesised and screened against M. incognita and G. pallida J2s for plant protective qualities. Chemotaxis, host-invasion, and stylet thrusting behaviours were assayed following J2 exposure to 100 μM of each uNLP for 24 h. Eleven of 27 tested uNLPs were found to disrupt normal chemotaxis towards root exudate collected from tomato cv. Moneymaker (Fig 1A). Of particular interest is our observation that six of eight predicted Mi-nlp-15 peptides inhibit chemosensation (Mi-NLP-15a/d, b, c, e, f). Analysis of the sequence similarity between these peptides suggest that the amino terminal variation of Mi-NLP-15g and h are responsible for observed functional differences. Two predicted Mi-nlp-9 peptides also inhibit chemosensation (Mi-NLP-9b, f), however no clear amino acid differences correlate with functionality across predicted nlp-9 peptides.
Likewise, 13 uNLPs were found to disrupt M. incognita host invasion (tomato cv. Moneymaker) compared to controls (Fig 1B). Multiple active uNLPs originated from single nlp genes, however no obvious amino acid conservation could predict bioactivity across Mi-nlp-8, -9, -18 or -15 peptides. In contrast, Mi-NLP14a and c share a common ALDMxEGDDFIGG motif. Three predicted uNLPs (Mi-NLP-15b, 9f, and 18a) inhibited both chemosensation and host invasion.
Eleven uNLPs were also found to disrupt the rate of serotonergic-induced M. incognita stylet thrusting (positively or negatively) compared with controls (Fig 1C). Multiple such uNLPs originated from Mi-nlp-15 and -18, however no common amino acid motif could be found relative to the inactive predicted peptides from either gene. Mi-NLP-14a and c were observed to differentially stimulate the inhibition and excitation of stylet thrusting respectively. The amino acid sequences of both peptides suggest that this difference must be mediated by differences in the 5th amino acid position, and/or differences at the carboxyl terminus. Mi-NLP-15b was the only predicted peptide to differentially regulate chemosensation, host invasion and stylet thrusting behaviours of M. incognita J2s (Refer to supplemental S1 Data).
12 of 25 tested uNLPs were found to disrupt chemotaxis of G. pallida J2s towards root exudate (tomato cv. Moneymaker), originating from Gp-nlp14, -15 and -21 (Fig 2A). Bioactivity of predicted peptides from Gp-nlp-15 and -21 does not correlate with an obvious amino acid sequence or motif, however Gp-NLP-14a and e peptides share an amino terminal ALDIL motif.
Five predicted Gp-NLP-21 neuropeptides were found to disrupt G. pallida host invasion (tomato cv. Moneymaker) relative to controls (Fig 2B). No obvious amino acid sequence or motif was predictive for bioactivity relative to the other inactive Gp-nlp-21 peptides. Both Gp-NLP-21b and g were found to inhibit both chemosensation and host invasion of G. pallida J2s.
Three uNLPs were also found to modulate serotonergic-induced stylet thrusting of G. pallida J2s relative to controls groups (Fig 2C). Gp-NLP-21h and -21i do not share any obvious amino acid similarity that is predictive for bioactivity relative to other inactive Gp-nlp-21 peptides. Gp-NLP-21h, -21i and Gp-NLP-15c were found to inhibit chemosensation alongside modulating stylet thrusting rates (Refer to supplemental S1 Data).
The potency of Mi-NLP-15b-induced disruption of chemotaxis and host invasion was assessed by exposing M. incognita J2s to various concentrations of synthetic Mi-NLP-15b for 24 h. Normal chemotaxis of M. incognita towards root exudate was inhibited across a range of dilutions, indicating high potency (Fig 3A). We found that M. incongita J2 invasion was also inhibited across a range of Mi-NLP-15b concentrations (Fig 3B; refer to supplemental S1 Data).
Innoculation of C. reinhardtii cultures secreting selected uNLPs into the tomato invasion assay arena inhibited M. incognita invasion relative to untransformed C. rehinhardtii: Mi- NLP-9f (10.32% +/-10.32, p<0.0001), Mi-NLP-15b (10.82% +/-6.574, p<0.0001) (Fig 4A). Likewise, innoculation of B. subtilis cultures secreting selected uNLPs, significantly inhibited M. incognita invasion: Mi-NLP-15b (26.63% +/-8.12, p = 0.0003), Mi-NLP-40 (23.72% +/-5.448, p = 0.0002) (Fig 4B). C. reinhardtii expressing Gp-NLP-15b also inhibited G. pallida invasion relative to controls (30.95% +/-9.021, p = 0.0042) (Fig 4C). Similarly, innoculation with B. subtilis secreting Gp-NLP-15b inhibited G. pallida invasion relative to control groups (51.98% +/-13.29), p = 0.0203 (Fig 4D). Secretion of a His-tagged NLP-15b peptide from B. subtilis was confirmed by ELISA, indicating active secretion of 193.8 ±81.3 ng/ml in LB broth culture (refer to supplemental S2 Data).
BLAST was used to identify NLP-15b homologues across available expressed sequence tags (ESTs) or genomes of PPNs and non-target nematode species. PPNs with diverse life history traits share high levels of NLP-15b sequence similarity, however sequence similarity is reduced in non-target nematode species (Table 2).
Incubation of mixed-stage C. elegans in selected PPN uNLPs (100 μM, 24 h) had no statistically significant impact on chemotaxis towards the attractants: sodium acetate, pyrazine, benzaldehyde or diacetyl, relative to controls (Fig 5D). Exposure of S. carpocapsae infective juveniles (IJs) to selected PPN uNLPs also had no statistically significant impact on insect host-finding (Fig 5E).
We have identified seven nlp genes that putatively encode 27 mature unamidated peptides in the root knot nematode, M. incognita (Mi-nlp-2, -8, -9, -14, -15, -18, -40). Likewise, four nlp genes predicted to encode 24 mature unamidated peptides were identified in the potato cyst nematode, G. pallida (Gp-nlp-8, -14, -15, -21) (Table 1). Several predicted unamidated NLPs share high levels of amino acid sequence similarity between M. incognita and G. pallida, with one predicted peptide, designated NLP-15b, perfectly conserved between the two. Indeed, NLP-15b is highly conserved at the sequence level across PPN species with diverse life history traits; less sequence similarity is observed between NLP-15b from PPNs and non-target species such as S. carpocapsae, C. elegans or P. pacificus for example (see Table 2).
Selected M. incognita and G. pallida peptides had a negative impact on PPN chemosensation and host-finding behaviours, but not on chemosensory or host-finding behaviours of mixed stage C. elegans or S. carpocapsae infective juveniles (Figs 1, 2 and 5). This may be due to NLP sequence dissimilarity, or to different peptide uptake efficiencies between species. The attractants used to assay C. elegans chemotaxis operate via distinct neuroanatomical and biochemical pathways; sodium acetate is detected by the ASE neurons, benzaldehyde by the AWC neurons and prazine and diacetyl are both detected by the AWA neuron. The ASE, AWC and AWA neurons mediate aspects of water soluble and volatile chemotaxis in C. elegans [48, 49]. Off-target NLP impacts were also assessed as a factor of host-finding ability in S. carpocapsae which will involve numerous neuroanatomical and biochemical pathways. Whilst these data on C. elegans and S. carpocapsae are far from exhaustive, they suggest that neuropeptide treatments that produce strong disruptive effects on the behaviours of M. incognita and G. pallida may be specific to PPNs.
PPNs use a hollow protrusible stylet in order to pierce plant cells on entry to the plant root, and to secrete various parasitism effectors related initially to cell wall degradation, and subsequently to the re-programming of plant cells into giant cell (RKN) or syncytial (PCN) feeding sites. Our data reveal that both agonistic and antagonistic disruption of stylet thrusting can reduce host invasion rates, however modulation of stylet thrusting does not always correlate with modified invasion behaviour under the conditions tested here. Mi-NLP-14c, Mi-NLP-18b, and -18d agonise serotonergic stylet thrusting, have no negative impact on J2 chemosensory ability, and yet also reduce host invasion rates. Mi-NLP-15f reduces stylet thrusting rate, but does not impact on host invasion, whereas Mi-NLP-14a reduces stylet thrusting rate and does inhibit host invasion. None of the three uNLPs that dysregulate stylet thrusting in G. pallida have an impact on host invasion rate. We hypothesise that enhanced stylet thrusting rate may be beneficial for initial invasion events, however it seems likely that coordinated stylet thrusting behaviour is more beneficial during feeding site development for example. Our data do not point to an obvious outcome in this regard, however we do find that dysregulation of behaviour tends to lower plant invasion levels of both M. incognita and G. pallida J2s.
Whilst it is tempting to extrapolate something on native NLP functionality from these data, we do not know if the aberrant phenotypes observed are due to interactions between tested NLPs and their cognate receptors. However, we do observe that exogenous NLPs can interact with endogenous neurophysiological circuits, interfering with host-finding, invasion and serotonergic stylet-thrusting behaviours of both M. incognita and G. pallida juveniles (Figs 1 and 2). This supports our initial hypothesis that nematode neuropeptides represent a valuable repository of nematicide candidates, which may elicit broad-spectrum activities against PPN species, but not off-target nematode species. Serial dilution of Mi-NLP-15b inhibited M. incognita chemosensation at concentrations as low as 10 pM, demonstrating high uNLP potency, which is a known characteristic of interactions between nematode neuropeptides and their cognate receptors [13, 25–30, 50, 51] (Fig 3). While the potency of this peptide would support the specificty of the associated phenotypic impact, we advise some caution when interpreting these data as indicative of NLP function within either M. incognita or G. pallida J2s due to the potential for peptide interaction with other, non-cognate receptors.
In order to further assess the efficacy of exogenous NLPs as nematicides, we developed two transgenic synthesis and delivery systems which could be deployed in the field, potentially through seed treatments or soil amendments. Gram positive Bacillus spp. are a major component of rhizosphere microbial communities [52, 53], and are frequently categorised as Plant Growth Promoting Rhizobacteria (PRPR) [54, 55]; B. subtilis has also been shown effective in controlling Meloidogyne species [56]. More generally, B. subtilis represents an important organism for many biotechnology applications, and is classified as GRAS (generally regarded as safe) by the FDA [57, 58]. It is increasingly well served by the development of synthetic biology tools [59], and can persist in soil for long periods through the production of spores [60]. We modified B. subtilis to secrete a number of PPN NLPs, and found that transformed B. subtilis cultures confer significant levels of protection on tomato cv. Moneymaker against both M. incognita and G. pallida infective juveniles (Fig 4). This proof of concept demonstration employed a commercial B. subtilis strain and signal peptide sequence. It has however been reported that signal peptide identity can have a significant influence on the level of protein / peptide secreted by B. subtilis [61, 62]. Unfortunately, we were unable to raise a suitable antisera to NLP-15b over several commercial synthesis rounds, due to the lack of NLP-15b immunogenicity. This restricted our ability to confirm amphidial uptake of the uNLPs, and to quantify microbial secretion of the uNLPs. Whilst we aimed to deliver proof of principle for this approach using commercially available and independently validated microbial synthesis and secretion systems, we confirmed secretion of a His-tagged NLP-15b from B. subtilis by ELISA (S2 Data). We anticipate that signal peptide optimisation efforts could increase secretion and correspondingly enhance plant protection levels. Likewise, assessing other rhizobacteria strains may enhance efficacy. The secretion of uNLP nematicides could also be more targeted if driven by a plant root exudate-responsive promoter [63, 64, 65, 66].We also utilised the soil-dwelling microalgae, C. reinhardtii as a novel synthesis and delivery platform. Like B. subtilis, C. reinhardtii benefits from an improving suite of synthetic biology tools [67]. C. reinhardtii cultures secreting selected PPN NLPs also provided significant levels of protection to tomato cv. Moneymaker when challenged by either M. incognita or G. pallida infective juveniles (Fig 4).
The NLP screening approach employed here may underestimate the efficacy achievable through a continuous transgenic delivery (Figs 1 and 2). For example, exogenous NLP-15b exposure inhibits G. pallida chemotaxis, but does not inhibit host invasion (Fig 2). However, when NLP-15b is delivered continuously to G. pallida infective juveniles via microbial secretion, we observe a significant inhibition of tomato invasion relative to J2s exposed to unmodified B. subtilis (Fig 4). This discrepency may be due to the recovery of G. pallida infective juveniles over the 24 hour timecourse of the tomato invasion assay. We expect that this may result in some false negative determinations in our NLP pre-screening approach.
Our data demonstrate that unamidated NLPs represent a new class of potent and specific plant protective nematicide that could be deployed as a transgenic trait in crop plants, or through soil microorganisms such as the B. subtilis and C. reinhardtii systems developed here. In particular, these non-crop delivery approaches could facilitate rapid deployment to many different crop plant species and cultivars. A key consideration in the development of PPN resistance traits must be the maintenance of genetic diversity across crop cultivars and isolates. This reduces the chance of widespread pathology from other pests as a result of genetic bottlenecks introduced by a single preferred transgenic cultivar.
The predicted NLP complement of C. elegans [16] was used in a simple BLASTp and tBLASTn analysis of available genomic / transcriptomic sequence data of G. pallida and M. incognita [46, 47]. All returned hits were curated by eye, and NLPs identified as per McVeigh et al. [17].
M. incognita were maintained in tomato plants (cv. Moneymaker) under greenhouse conditions. 8 weeks post infection M. incognita eggs were harvested from the roots by washing away excess soil and by briefly treating cleaned roots in 5% sodium hypochlorite to soften the root tissue and release the eggs. Eggs were cleaned from debris by passage through nested sieves (180 micron, 150 micron and 38 micron) and washed thoroughly with water. Eggs were separated from remaining soil / silt by centrifugation (2000 rcf for 2 minutes) in 100% sucrose solution and collected in a thin layer of spring water (autoclaved and adjusted to pH 7). Eggs were treated in antibiotic / antimycotic solution (Sigma) overnight, placed in a nylon net with a 38 micron pore size, immersed in spring water and maintained in darkness at 23°C, until infective juveniles emerged. Freshly hatched juveniles were used for each assay.
G. pallida were maintained in potato (cv. Cara) at the Agri-Food and Biosciences Institute (AFBI), Belfast. Soil was collected surrounding potato roots, dried for one week and washed through sieves to collect cysts. Cysts were incubated in potato root diffusate in the dark at 17°C until infective juveniles emerged. Freshly hatched juveniles were used for each assay.
Predicted uNLPs from both M. incognita and G. pallida were synthesised by EZBiolab and dissolved into pH adjusted ddH2O to make a 5 mM stock which was aliquoted and stored at -20°C. J2s of both M. incognita and G. pallida were incubated for 24 hours in 200 μl of each peptide in a 24 well plate (SPL Lifesciences, South Korea) at a defined concentration.
A 60 mm Petri dish was divided into two segments, a positive and a negative side, with a 0.5 cm 'dead zone' either side of the centre point. The petri dish was filled with 15 ml of 0.25% w/v agar which was allowed to solidify. 3 ml of 0.25% w/v agar slurry in spring water (pH 7, agitated with a magnetic stirrer for several hours to give a smooth consistency) was added to the petri dish and spread evenly over the surface. Root diffusate (attractant) and water only (control) 0.25% agar plugs were embedded in the agar slurry, either side of the assay arena. Root diffusate was collected from 10 tomato plants, aged 3–6 weeks in 1 litre pots, by pouring 500 ml of ddH20 through the soil three times. Diffusate from each plant was combined, filter sterilised and stored at 4°C for a maximum of 1 month. Root diffusate agar plugs were made by melting 1.25% agar in ddH20, cooling to 50°C before mixing with 4 parts of root diffusate. The agar was then allowed to solidify at room temperature. 100 uNLP pre-treated M. incognita or G. pallida J2s were added by pipette to the centre of the plate. J2s which moved out of the 'dead zone' after 3 hours were counted and their location (+/-) scored. The distribution of J2s were used to create a chemotaxis index [68] for each plate, which formed one replicate, a total of 10 replicates where completed for each uNLP treatment.
Tomato seeds were sterilised with 2.5% NaOCl for 15 minutes, washed 5 times in ddH20 and germinated on 0.5% Murashige and Skoog plates at 23°C. An agar slurry was prepared by autoclaving 0.55% (w/v) agar (using autoclaved spring water adjusted to pH 7) which was mechanically agitated overnight until it had a smooth consistency. Invasion assays were performed by mixing 500 pre-treated M. incognita or G. pallida J2s with agar slurry and a single tomato seedling (2 days post germination) in a 6 well plate. Assays were left at 23°C for 24 hours in the case of M. incognita and at 18°C for 24 hours in the case of G. pallida under a 16 hour light and 8 hour darkness cycle. Seedlings were stained using acid fuschin [69] and the number of nematodes within the roots counted. At least five seedlings were used for M. incognita infections assays, with at least 15 seedlings used for G. pallida infections assays (due to increased variation).
Stylet thrusting assays where performed by incubating 100 M. incognita or G. pallida J2s for 15 minute in 5 mM or 2 mM serotonin (Sigma Aldrich, USA), respectively. J2s were placed on a glass slide and stylet thrusts were counted for randomly selected J2s, for 1 minute each. Counts for a given cohort of J2s were taken in a maximum interval of 15 minutes. Longer counting intervals, making for longer serotonin incubations, yielded inconsistent results. At least 30 J2s were counted for each neuropeptide treatment.
B. subtilis were grown overnight in LB media containing ampicillin (100 μg/ml) at 37°C with shaking, and harvested in the log phase of growth determined by measuring OD600nm. Five ml of culture at 0.5 OD was spun down and the pellet mixed with 3 ml of agar slurry and 500 J2s from either G. pallida or M. incognita. C. reinhardtii clones were grown at 23°C with shaking, cultures in the log phase were measured at OD750 and 5 ml of culture at 0.5 OD was pelleted by centrifugation. C. reinhardtii pellets were mixed with 3 ml of agar slurry and 500 J2s from either G. pallida or M. incognita. Plant invasion assays were performed as described above.
C. elegans wild-type N2 Bristol strain were obtained from the C. elegans Genomics Center and maintained on a Escherichia coli (strain OP50) lawn on nematode growth medium (NGM) agar plates (3 g/l NaCl, 17 g/l agar, 2.5 g/l peptone, 5 mg/l cholesterol, 25 mM KH2PO4 (pH 6.0), 1 mM CaCl2, 1 mM MgSO4) at 20°C [70]. Chemotaxis assays were performed in a 9 cm diameter Petri dish on NGM agar which was split into a positive and negative side with a central ‘dead zone’ of 1.5 cm diameter. 100 mixed-staged C. elegans were washed three times in M9 buffer and soaked in 100 μM PPN uNLP, or M9 vehicle control for 24 hours. 2 μl of 50 mM sodium acetate, 0.5% pyrazine, 0.5% benzaldehyde or 0.5% diacetyl was spotted onto the positive side, 2 μl of ddH20 was spotted onto the negative side. Pyrazine, benzaldehyde and diacetyl volatile attractants were assayed immediately whereas the water soluble sodium acetate was assayed 18 hours following addition to the plate. Assays were maintained in the dark at 20°C, and counted after 1 hour. Eight replicates were conducted for each C. elegans attraction assay.
S. carpocapsae were cultured in Galleria mellonella at 23°C. Infective juveniles (IJs) were collected using a White trap [71] in PBS. Freshly emerged IJs were used for each assay. 100 IJs were incubated for 24 hours in 100 μM of selected uNLPs, and host-finding assays (n = 5) performed as in Morris et al. [45].
Codon optimised DNA sequences coding for the desired neuropeptide flanked by restriction sites necessary to clone into the C. reinhardtii expression vector pChlamy_3 (Life Technologies, USA) or the B. subtilis expression vector pBE-S (Clontech, USA) were synthesised by GeneArt Gene Synthesis (Life Technologies, USA).
uNLP secretion inserts, and vector pChlamy_3 were digested using KpnI/XbaI (New England Biolabs, USA), ligated using T4 ligase (New England Biolabs, USA), and cloned into Escherichia coli One Shot TOP10 chemically competent cells (Life Technologies, USA) following manufacturer’s instructions. Ampicillin (Sigma Aldrich, USA) was used to select E. coli containing the pChlamy_3 plasmid, which was subsequently extracted using the High Pure Plasmid Isolation Kit (Roche) and sequenced (Eurofins Genomics, UK) to identify correct clones. C. reinhardtii was transformed by electroporation following manufacturer’s instructions (GeneArt Chlamydomonas Engineering Kit, Life Technologies) and individual colonies grown on TAP-Agar-Hygromycin plates (10 μg/mL) (Sigma Aldrich, USA) at 23°C. Colonies were picked and grown at 23°C in 100 ml TAP growth media (Invitrogen, USA) with constant orbital agitation. qRT-PCR was performed to identify clones with the highest level of uNLP expression, which were then selected for downstream assays (pChlamy universal FWD: CACTTTCAGCGACAAACGAG, nlp-15b REV: CTACTAGTCGAGGCCGGTA; Mi-nlp-9f REV: GAACGGGCGGATGAAGTAG).
uNLP secretion inserts, and vector pBE-S were digested using XbaI/MluI (New England Biolabs, USA), ligated using T4 ligase (New England Biolabs, USA), and cloned into E. coli One Shot TOP10 chemically competent cells (Life Technologies, USA) following manufacturer’s instructions. Ampicillin (Sigma Aldrich, USA) was used to select E. coli containing the pBE-S plasmids, which were subsequently extracted using the High Pure Plasmid Isolation Kit (Roche) and sequenced (Eurofins Genomics, UK) to identify correct clones. B. subtilis RIK1285 competent cells (Takara, USA) were transformed according to manufacturer’s instructions and grown overnight at 37°C on kanamycin selective plates (10 μg/mL) (Sigma Aldrich, USA). Individual colonies were picked and grown in LB broth overnight at 37°C. qRT-PCR (pBE-S universal FWD: GGATCAGCTTGTTGTTTGCGT, nlp-15b REV: CCTGGCCCAGTGAAAGAGTC, Mi-nlp-40 REV: TACCGGCTGCCAAGATACCA) was performed to confirm the expression of uNLP secretion cassettes.
Codon optimised NLP-15b, tagged with six histidine residues and an upstream aprE signal peptide, were cloned into the pBE-S vector (GeneArt, Life Technologies) and transformed into B. subtilis following manufacturer’s instructions (Takara Bio, Inc.). NLP-15b transformed B. subtilis were grown at 37°C in 50 ml of LB (kanamycin 10 μg/ml) and wild type B. subtilis in 50 ml of LB without selection. Once growth passed the exponential phase (OD 660) one tablet of cOmplete Protease Inhibitor Cocktail (Roche) was added to 10 ml of bacterial suspension and allowed to dissolve. Bacteria were removed from the LB by centrifugation at 10,000g for 10 minutes. Supernatant was collected and peptides were isolated by a MWCO 3 kDa filter (Amicon, Sigma). Histidine tagged peptide concentration assessed using the His Tag protein ELISA Kit (Cell Biolabs, Inc.) following manufacturer’s instructions. The ELISA results were measured (OD 450) using the FLUOstar Omega microplate reader (BMG Labtech). A line of best fit was plotted and the slope used to calculate the concentration of peptide across individual samples (n = >11).
Data pertaining to behavioural and invasion assays were assessed by Brown-Forsythe and Bartlett’s tests to examine homogeneity of variance between groups. One-way ANOVA was followed by Fisher’s Least Significant Difference (LSD) test. All statistical tests were performed using GraphPad Prism 6.
|
10.1371/journal.pntd.0006702 | Community-directed vector control to supplement mass drug distribution for onchocerciasis elimination in the Madi mid-North focus of Northern Uganda | Onchocerciasis a neglected tropical disease that historically has been a major cause of morbidity and an obstacle to economic development in the developing world. It is caused by infection with Onchocerca volvulus, which is transmitted by black flies of the genus Simulium. The discovery of the potent effect of Mectizan (ivermectin) on O. volvulus microfilariae and the decision by its manufacturer to donate the drug for onchocerciasis spurred the implementation of international programs to control and, more recently, eliminate this scourge. These programs rely primarily on mass distribution of ivermectin (MDA) to the afflicted populations. However, MDA alone will not be sufficient to eliminate onchocerciasis where transmission is intense and where ivermectin MDA is precluded by co-endemicity with Loa loa. Vector control will likely be required as a supplemental intervention in these situations.
Because biting by the black fly vectors is often a major nuisance in onchocerciasis afflicted communities, we hypothesized that community members might be mobilized to clear the breeding sites of the vegetation that represents the primary black fly larvae attachment point. We evaluated the effect of such a community based "slash and clear" intervention in multiple communities in Northern Uganda. Slash and Clear resulted in 89–99% declines in vector biting rates. The effect lasted up to 120 days post intervention.
Slash and clear might represent an effective, inexpensive, community- based tool to supplement ivermectin distribution as a contributory method to eliminate onchocerciasis and prevent recrudescence.
| River blindness is one of the most important causes of morbidity in the developing world. The discovery of ivermectin and the decision by its manufacturer to donate the drug for river blindness spawned the development of programs to eliminate river blindness through mass treatment of afflicted populations. But ivermectin alone will not eliminate river blindness in much of Africa; additional interventions are necessary. We show that a simple community-based approach to controlling the black fly vector results in dramatic reductions in the vector population. Such a community-based approach to vector control will be compatible with the community-driven mass drug administration programs distributing ivermectin for onchocerciasis elimination in Africa. This should help reduce the time needed to obtain elimination and help prevent recrudescence once elimination is attained.
| Onchocerciasis (river blindness) is a neglected tropical disease that is historically one of the most important causes of blindness worldwide [1]. The disease is caused by the human filarial parasite, Onchocerca volvulus. The major pathogenic manifestation of the infection, ocular damage leading to blindness, commonly affects individuals beginning in the second decade of life, disabling them as they enter their most productive period. Other manifestations include pruritis and severe skin disease. In hyper-endemic settings, the care of blinded individuals places a severe stress on the community, often leading to their dissolution [2]. The parasite is transmitted by black flies (primarily Simulium damnosum sensu lato in Africa) that develop as larvae in fast running rivers. Thus, transmission of the parasite is most intense in large river basins, rendering many such areas uninhabitable [2]. Unfortunately, these areas contain much of the fertile land in the African savanna. By preventing their use for agriculture, onchocerciasis has retarded economic growth in many of the poorest countries of Africa.
In the 1980s, Mectizan (ivermectin) was shown to be a potent microfilaricide against O. volvulus [3] and that mass treatment of an afflicted population could reduce parasite transmission [4, 5]. Merck, the manufacturer of ivermectin, announced that it would provide the drug free of charge for the treatment of onchocerciasis, “as much as needed for as long as needed” [6]. Because of this, several programs were begun to control or eliminate onchocerciasis, employing a strategy of ivermectin mass drug administration (MDA) to the afflicted communities. These included the African Programme for Onchocerciasis Control (APOC) in Africa and the Onchocerciasis Elimination Program of the Americas (OEPA). OEPA, employing a strategy of semi-annual distribution of ivermectin MDA has succeeded in eliminating onchocerciasis in four of the six formerly endemic countries in Latin America and has interrupted transmission in all but one binational focus in the region [7]. In Africa, studies conducted in Mali and Senegal [8] and Nigeria [9] demonstrated that APOC- administered annual ivermectin MDA eliminated onchocerciasis from some foci.
Despite these successes, ivermectin MDA is not a panacea in the struggle to eliminate onchocerciasis. Where vector densities are high, models suggest that ivermectin MDA alone will not be sufficient to interrupt transmission [10, 11]. These predictions have recently been supported by field data from Cameroon and Uganda, where transmission of O. volvulus continues despite 15 and 18 years of annual ivermectin MDA, respectively [12, 13]. Second, large portions of Central Africa afflicted with onchocerciasis are co-endemic for Loa loa. Individuals with high L. loa parasitemias are susceptible to developing severe side effects when given ivermectin [14]. This has complicated the use of ivermectin MDA in areas where O. volvulus and L. loa are co-endemic and has prevented the implementation of ivermectin MDA altogether in some areas.
Vector control as a tactic to combat onchocerciasis in Africa has a long history. The use of larvicides to eliminate adult black flies and block transmission of Onchocerca volvulus was first implemented in Kenya in1946 [15]. Elimination of the vector (Simulium neavei) was successful and follow-up studies conducted in 1964 confirmed that the parasite had been eliminated from that country [16]. This successful program was used as a model for the first international onchocerciasis control program in Africa, the Onchocerciasis Control Programme in West Africa, or OCP. The OCP was a large-scale, vertically integrated control program whose aim was to eliminate blinding onchocerciasis as a public health problem throughout eleven countries in West Africa through vector control. A great deal of public health value was accomplished by this landmark effort. Skin disease was significantly reduced, more than 200,000 cases of blindness were prevented and the size of the O. volvulus population was substantially decreased [17]. More recently, Uganda has demonstrated the power of utilizing a combination of vector control and ivermectin MDA. Uganda has used a strategy that combines vector control (local larviciding of breeding sites) with semi-annual MDA. This has resulted in the apparent interruption of transmission in 15 of the 17 foci in Uganda, a finding that has been confirmed in the Wadelai [18], Itwara [19] and Mt. Elgon [20] foci. These outcomes are similar to those reported from the island of Bioko where a combination of vector control followed by ivermectin treatment has resulted in elimination of the parasite [21].
These data suggest that vector control, used in combination with ivermectin MDA, is a powerful, synergistic strategy to eliminate onchocerciasis. However, traditional vector control of S. damnosum s.l. with larvicides has several drawbacks. Larvicides are expensive to apply, have potential detrimental environmental consequences and require technically trained individuals to calculate proper dosage. We hypothesized that community members might be mobilized to remove the trailing vegetation at the breeding sites that represent a primary attachment point for the black fly larvae as an alternative method to larviciding. Here we report a study whose overall objective was to determine if simple community-based removal of larval attachment sites could reduce the biting rate of the black fly vectors. We report the results of a number of different trials to evaluate the effect that such a "slash and clear" approach had upon the biting rates of the vector in onchocerciasis endemic communities in Northern Uganda. The data indicate that such a community-driven vector control initiative may be an effective and inexpensive tool to supplement ivermectin MDA and accelerate the effort to eliminate onchocerciasis from Africa.
This study was conducted in communities located in Northern Uganda within the Madi-mid North focus of onchocerciasis (Fig 1). All communities were located in the districts of Amuru and Nwoya in the Madi-Mid North focus of Uganda. The vector of O. volvulus in this region is the savannah dwelling species Simulium damnosum sensu stricto [23]. While Uganda committed to a program of onchocerciasis elimination in 2007 [22] and began a program of country wide twice per year ivermectin distribution in onchocerciasis foci at that time, regular treatments were delayed in the Madi mid-North focus due to political unrest. Initial evaluations conducted prior to the start of the regular treatment program suggested that the region was hypo-endemic for onchocerciasis. The prevalence of skin snip positive individuals in the Amuru district was 1.5%, while no skin snip positive individuals were encountered in a survey of four villages in the Nwoya district. Regular twice per year treatments began in this region in 2010 and have been maintained since then. Mean therapeutic coverages in both districts have exceeded 90% in all treatment rounds since 2010.
The communities were each visited by the field team to validate their choice for inclusion in the study. Team members informally questioned the residents of the village about their knowledge of biting black flies, helping to determine if the flies represented a significant nuisance. Predicted breeding sites located within 1km of the village were validated by ground prospection to confirm the presence of S. damnosum larvae. As described below, all chosen communities were found to have similar biting rates prior to the start of the interventions.
Selected villages were divided into pairs, with one village of each pair randomly assigned to the control group and one to the intervention group. Baseline collections using standard human landing techniques were carried out to establish the biting rate at each community. Following the baseline collections, young men (16–22 years of age) were recruited to carry out slash and clear in the intervention villages. The recruits were brought to the breeding sites located 1km upstream and downstream of the village and were instructed in the process of cutting the trailing vegetation from the water and throwing it on the river banks to dry, thereby killing the adherent black fly larvae. In the initial trials, two slash and clear cycles of intervention were conducted, with the first cycle conducted on days 8 and 9 of the study (i.e. after the 7-day baseline period) and the second conducted on days 19 and 20 of the study, thereby eliminating nearby larval substrates. Landing collections of adult flies were carried out daily throughout the study period (31 days total). In the long-term studies, collections were carried out daily throughout the first 20 days of the study, and twice per week thereafter. All community members, including the individuals participating in the study, were given ivermectin twice per year as part of the Uganda Onchocerciasis Elimination Program of the Uganda Ministry of Health.
The number of flies collected in the intervention and control communities were compared at day 8 (at the start of the intervention), at day 18 (at the start of the second intervention) and at the end of each trial. Fly counts in the control communities at the start and end of the first two trials were also compared. The data were analyzed using a basic linear model that treated the river as a blocking effect and treatment type as the variable of interest. The basic factorial design model had the form
count=rivertreatmentriver*treatment
Because the data were counts, a negative binomial distribution model was used with SAS PROC GENMOD. A complete description of the statistical analysis and the results may be found in the S1 Supplemental Material.
The initial trial of the slash and clear intervention was carried out on breeding sites near the villages of Gonycogo and Adibuk, located along the Ayago river, a small river in North Central Uganda (Fig 1). The villages of Laminlatoo and Ayago/Nile, also located along the Ayago, served as control communities. In all cases, the village inhabitants were found to be acutely aware of the extreme nuisance posed by blood-feeding black flies. The initial trials commenced on May 9, 2015, in the beginning of the rainy season in Northern Uganda. Trailing vegetation was removed from all breeding sites located within 1km of the intervention villages on days 8 and 9 and 19 and 20 (Fig 2, Panel A). No interventions were conducted in the control villages. Daily biting rates in the control and intervention communities were not significantly different from one another prior to commencement of the interventions (p> 0.1). Fly biting rates were seen to decline beginning at day 16, 6 days following completion of the first intervention, and were significantly lower than those in the control villages by day 18 (Fig 2, Panels B and C; p <0.0001). Fly numbers continued to decline through the end of the study, at which point the mean biting rate in the intervention villages was 11% of the mean biting rate in the control villages (mean daily biting rate of 32 in the intervention villages versus 296.5 in the control villages; Fig 2, Panels B and C; p < 0.0001). No significant change in the biting rate was seen in the control villages throughout the trial period (Fig 2; p = 0.9).
The Ayago river is a small stream, averaging 2m in width. Thus, it was of interest to determine if a similar approach could be used along larger rivers. A second study was carried out along the Aswa river, one of the largest rivers in the district, averaging 11m in width (Fig 3, Panel A). Again, four villages were identified using the methods described above. Two villages were assigned to the intervention group and two to the control group. Interventions were carried out on days 8 and 9 of the study and on days 19 and 20 (Fig 3, Panels B and C). This study was conducted in late August and early September 2015, which is during the peak fly biting season in this area. The results were similar to those obtained in the initial trial, with mean fly biting rates in the intervention communities declining to just 1% of the mean biting rates found in the control communities at the end of the study (mean daily biting rate in the intervention villages of 3.5 versus a mean daily biting rate in the control villages of 412; Fig 3, Panels B and C; p < 0.0001).
These initial studies were limited to approximately one month during the rainy season. Thus, it was of interest to determine how long the effect of the slash and clear intervention lasted. To begin to answer this question, a third trial was conducted on the Aswa river, which began on July 15, 2016, in the middle of the rainy season when fly numbers are high. Fly collections continued until November 30, 2016, near the end of the rainy season. Similar to the previous trials, the maximum reduction in the biting rate in the intervention villages was reached on day 29, where the mean biting rate in the intervention villages was 2.4% of that seen in the control villages (mean daily biting rate of 4.5 in the intervention villages versus a mean daily biting rate of 183.5 in the control villages; Fig 4, Panels A and B). This degree of reduction was maintained through day 67 of the study (September 19, 2016), at which point the fly numbers in the intervention villages began to slowly recover. At day 139 (the end of the trial), the mean biting rate in the intervention villages had reached 32% of the mean biting rate in the control villages (mean daily biting rate of 55.5 in the intervention villages versus 173.5 in the control villages; Fig 4).
As the fly populations did not recover completely by the end of the first long-term trial, a second long-term trial was carried out beginning in May, 2017. The goal of this trial was two-fold; first to determine how long it took for the fly populations to recover to levels indistinguishable from the control sites, and second to obtain baseline data on the fly biting rates throughout the year. In this trial, a total of six communities were enrolled. Three were randomly assigned to the control group and three to the intervention group as before. A single slash and clear intervention was carried out towards the end of May, 2017 in the three intervention villages. The slash intervention was completed on May 24, 2017. Flies were collected twice per week from all villages for a total of one year (May, 2017-April, 2018). The daily biting rate in the intervention villages again fell dramatically, reaching 3.3% of the mean daily biting rate in the control villages on June 15, 2017 (a mean daily biting rate of 2.67 in the intervention villages versus a mean daily biting rate of 80.3 in the control villages; Fig 5, Panel A). This effect lasted through the end of July, when the mean daily biting rate in the intervention villages was 9.9% of that in the control villages (a mean daily biting rate of 4.67 in the intervention villages, versus a mean daily biting rate of 47.3 in the control villages; Fig 5, Panel A). In August, fly numbers fell precipitously in the control villages as a result of a flood that occurred in early August that removed much of the trailing vegetation from the control breeding sites. Fly numbers then began to recover in both the control and intervention sites in November, but then the fly numbers declined dramatically in all sites in December, marking the start of the dry season (Fig 5, Panel B). Fly numbers began to recover in all sites as the rains began to return in March, 2018 (Fig 5).
The data presented here suggest that removal of trailing vegetation by community members resulted in a dramatic reduction in daily vector biting rates in the Madi mid-North focus of onchocerciasis in Uganda. The maximum reductions we observed in the daily biting rate ranged from 89% to 99%, occurring roughly 20 days following the initial intervention. These results are similar to the only other study reporting an evaluation of a similar vegetation removal strategy conducted in Sudan in 1984 [26]. In that study, biting rate reductions of roughly 80% were reported. However, in that trial, vegetation removal was conducted in conjunction with larviciding, making it difficult to separate the effects of the two methods. This study demonstrates that vegetation removal alone resulted in a highly significant reduction in biting rates.
The effect of vegetation removal was quite long lasting. This is perhaps not surprising, as the vegetation clearance removed most or all of the trailing vegetation at the breeding sites, and we observed that the trimmed vegetation took months to recover. This is in contrast to larviciding, which, though it removes developing S. damnosum s.l. larvae in the breeding site does not alter the breeding site habitat. Thus, the breeding site can be rapidly re-populated by gravid adult flies that are not affected by larviciding [27]. In contrast, slash and clear removes vegetation substrates from the breeding habitat, limiting re-population until the substrates recover. Importantly, the duration of the effect of the slash and clear treatments means that if the goal is achieving a sustained reduction in biting rates, the removal process may need to be repeated only once every two months during the breeding season. Finally, slash and clear avoids the prospect of insecticide resistance, a fairly common phenomenon associated with long-term use of insecticides that significantly contribute to the cost and complexity of control programs [28]. As such, this approach fits well with a recent proposal that alternative vector control strategies should be considered to mitigate the impact of resistance [29].
We did not attempt to accurately assess the effect that the reduction in biting rate combined with MDA had on parasite transmission in these studies. This is because, as is the case for most previously endemic foci in Africa, the population in this area had received multiple treatments with ivermectin and transmission in the focus was dramatically reduced as a result. Furthermore, due to the success of the slash and clear process, insignificant numbers of flies were collected in the intervention sites once the interventions were performed. However, based on traditional studies where vector control drastically reduces annual biting rates and transmission, we believe that the impact would be significant on transmission of O. volvulus and would eventually threaten the existence of the parasite population.
We believe that there are three features inherent in the slash and clear method that may increase its potential for sustainability in the affected communities. First, in all the communities enrolled in the study, residents could readily identify the vector flies and all reported that biting from the flies represented a significant nuisance. The community members were motivated to become involved in any program that promised to reduce the number of biting flies plaguing them. Second, the expenses involved in carrying out the slash and clear interventions are minimal and involve materials already available and used routinely in the community (e.g. machetes and rubber boots). Thus, the investment required by the community to undertake and maintain a slash and clear intervention is minimal. Finally, we predict it is likely that as the fly densities return to a nuisance level, the community members will be motivated to independently conduct slash and clear interventions, thereby keeping biting rates low. Studies to test this hypothesis are currently underway.
It is notable that we observed dramatic reductions in the biting rate in all the intervention communities, despite the fact that only breeding sites located within a 1km radius of each village were targeted. This finding indicates that most black flies biting the community members were derived from nearby breeding sites. This suggests that targeting only the nearby breeding sites will be sufficient to dramatically reduce the biting rate in a given community, thereby likely reducing the annual transmission potential, a key epidemiological statistic. The data also suggest that as most of the flies take bloodmeals locally, the large majority of parasite transmission may also be driven by locally produced flies. This has implications for defining so called “transmission zones” when planning and implementing elimination programs. Previous studies have suggested that the savanna dwelling species of Simulium damnosum can travel long distances in West Africa, migrating on seasonal winds [30]. This suggests that transmission zones may be quite large, on the order of hundreds of kilometers in diameter. However, if most of the biting is the result of flies that are breeding in nearby breeding sites, it is likely that most of the transmission will be carried out by locally breeding flies, which would tend to shrink the effective size of the transmission zone and the contribution of migrating flies to overall transmission. Additional work will be necessary to quantify the contribution of migrating flies to the overall level of transmission.
Models of the dynamics of transmission of O. volvulus suggest that transmission intensity is strongly affected by the rate of host-vector contact [10, 11]. Thus, reducing vector densities can be an effective method of reducing or suppressing transmission [31]. In fact, the first large scale international program that eliminated blinding onchocerciasis as a public health problem, the Onchocerciasis Control Programme in West Africa (OCP), relied almost exclusively on vector control to reach their goal [32]. We hypothesize that the slash and clear approach to vector control may be applicable in at least three situations in the effort to eliminate onchocerciasis: 1. As a supplement to ivermectin MDA to reduce the time necessary to achieve elimination; 2. As an adjunct intervention to the selective use of ivermectin (employing the "test and not treat" strategy in areas co-endemic for L. loa and O. volvulus), and; 3. As a way to prevent recrudescence from occurring once onchocerciasis has been eliminated from a given focus. The effectiveness of slash and clear in these situations could be estimated by using the data generated in this study in models that are being developed to assist in the onchocerciasis control and elimination efforts in Africa [31]. Such studies are currently underway.
While the data presented here are very encouraging, it is likely that slash and clear will not be a panacea for every effort to eliminate onchocerciasis from Africa. First, these studies targeted the savanna dwelling species of S. damnosum s.l. and involved relatively small rivers. It is unlikely that this strategy will be as successful when applied in communities located near breeding sites in some of the large powerful rivers in Africa (such as the Nile). Furthermore, although the savanna dwelling species represent the major vectors of O. volvulus throughout most of sub-Saharan Africa, applying this method to foci in which some of the other sibling species of S. damnosum s.l. are vectors may prove to be more difficult. For example, breeding in forested onchocerciasis foci occurs in small streams, making the identification of all the breeding sites located near a community difficult. Additional studies will be necessary to evaluate the effect of slash and clear in such environments.
|
10.1371/journal.pntd.0003475 | Surveillance of Aedes aegypti: Comparison of House Index with Four Alternative Traps | The mosquito Aedes aegypti, vector of dengue, chikungunya and yellow fever viruses, is an important target of vector control programs in tropical countries. Most mosquito surveillance programs are still based on the traditional household larval surveys, despite the availability of new trapping devices. We report the results of a multicentric entomological survey using four types of traps, besides the larval survey, to compare the entomological indices generated by these different surveillance tools in terms of their sensitivity to detect mosquito density variation.
The study was conducted in five mid-sized cities, representing variations of tropical climate regimens. Surveillance schemes using traps for adults (BG-Sentinel, Adultrap and MosquiTRAP) or eggs (ovitraps) were applied monthly to three 1 km2 areas per city. Simultaneously, larval surveys were performed. Trap positivity and density indices in each area were calculated and regressed against meteorological variables to characterize the seasonal pattern of mosquito infestation in all cities, as measured by each of the four traps.
The House Index was consistently low in most cities, with median always 0. Traps rarely produced null indices, pointing to their greater sensitivity in detecting the presence of Ae. aegypti in comparison to the larval survey. Trap positivity indices tend to plateau at high mosquito densities. Despite this, both indices, positivity and density, agreed on the seasonality of mosquito abundance in all cities. Mosquito seasonality associated preferentially with temperature than with precipitation even in areas where temperature variation is small.
All investigated traps performed better than the House Index in measuring the seasonal variation in mosquito abundance and should be considered as complements or alternatives to larval surveys. Choice between traps should further consider differences of cost and ease-of-use.
| Dengue vector surveillance programs regularly use household surveys for searching breeding sites that are positive for Ae. aegypti larvae. Infestation indices are calculated as the percentage of positive houses, or percentage of positive containers, and are used to guide control actions and to issue alerts. However, these indices are costly and prone to error due to variation in searching effort and the sometimes cryptic nature of the mosquito egg laying behavior. In the recent years, many devices for trapping mosquitoes were developed. One problem when deciding which trap to choose for surveillance is the lack of a gold standard to compare trap-based indices with. In this scenario, choice relies on the behavior of the trap indices, that is, how well they inform about mosquito population growth, decline, or spread. Our study compared infestation indices produced by four different trap schemes as well the immature survey. We observed that any trap is more sensitive in detecting the mosquito presence than the immature survey, and they are better in capturing the temporal variation in mosquito abundance as well. However, traps are very different in terms of cost, ease-of-use and sensitivity and trap choice should further consider these factors.
| The mosquito Aedes aegypti, vector of dengue, chikungunya and yellow fever viruses, is an important target of vector control programs in tropical countries. Traditional Ae. aegypti surveillance is based on periodic household inspections for the presence of larvae-bearing containers, which inform health agents on the most productive breeding sites and trigger control strategies in the form of container removal or chemical treatment. Household surveys also provide measures of infestation in the form of House (HI) and Breteau indices (BI). Based on the former, risk of disease transmission is empirically defined as low if HI<1.0%, or high, if HI> = 4.0%, and these thresholds guide control initiatives [1].
House indices face many criticisms: household surveys are costly to be performed with the frequency required for surveillance; indices are highly dependent on both agent’s effort and householder availability; these surveys only provide qualitative measures of abundance as the number of immatures per container or premise is not computed, only their presence/absence; moreover, larval density is not a precise measurement of mosquito adult density, the stage involved in dengue virus transmission [2].
Traps are promising alternatives to larval surveys: they transfer the searching effort from the health agents to the mosquitoes themselves (this time saved allows more frequent surveys); and traps provide qualitative (% of positive traps) and quantitative (number of captures per trap) indices [2, 3].
There are currently several traps for Ae. aegypti, varying in their attractiveness, specificity to different mosquito life stages, ease-of-use, and cost. The classical trap is the patent-free ovitrap, a tool designed to attract and collect eggs of female mosquitoes searching places to oviposit. This trap, in use since 1965, is employed together with routine larval surveys in many countries, due to its high sensitivity and low cost [4–6]. When mosquito infestation is low, ovitraps are more sensitive for detecting Ae. aegypti presence than larval surveys [7]. Ovitraps have disadvantages as well: variation in the local availability of breeding sites may interfere with their attractiveness by affecting the probability of egg deposition in these traps, potentially impairing comparison among areas; the skipping oviposition behavior of Aedes females may also affect the number of eggs deposited in individual traps and the reliability of adult abundance estimation derived from egg counts.
Traps against adults are often presented as alternatives to the ovitrap and to larval surveys (see [2] for a review). In principle, the advantage of adult catching traps is to obtain estimates of the population that is directly involved in transmission. Currently available adult capturing traps are designed to attract only a subset of the adult population: the ovipositing females are attracted to Adultrap [8] and to sticky traps, like MosquiTRAP; host seeking females are attracted to the BioGent-Sentinel [9]. The reliability of the infestation indices will depend on a series of parameters, as efficiency, sensibility and specificity.
Attractiveness is an important issue as any trap has to compete for the“attention” of the target individuals [10]. Ovipositing traps compete with other breeding sites, and traps emulating a host compete with hosts themselves. The quality of a trap depends on how their attractiveness remains unchanged as the environment changes.
Traps against adults (and ovitraps) do not provide absolute measurements of the adult population (which should be expressed in“mosquitoes/area” or“mosquitoes/person” units) [11, 12]. Trap indices are relative measures of abundance, having“mosquitoes per trap” unit. Conversion from relative to absolute measurements requires assumptions regarding the area effectively covered by a trap. The development of statistical methods for estimating absolute mosquito abundance (mosquitoes/person or mosquitoes/premise) from adult trap data is a recent subject [13].
One problem when deciding which trap to choose for surveillance is the lack of a gold standard to compare trap-based indices with. In this scenario, choice relies on the behavior of the trap indices, that is, how well they inform about mosquito population growth, decline, or spread. Ideally, traps’ statistical properties (sensitivity, specificity) should be invariant (that is, not influenced by anything but fluctuations in mosquito abundance), and consistently behave under both low and high abundances.
Here, we present the results of a two-year study designed in the context of a routine Ae. aegypti surveillance program to compare the standard larval survey and four traps: ovitrap, MosquiTRAP, Adultrap and BioGent-Sentinel. Five cities representing four climatically distinct dengue endemic regions in Brazil were simultaneously monitored by the different traps. We investigated the temporal consistence of the entomological indices produced by the different traps, comparing positivity and density indices, and trap indices versus larval indices.
Brazil extends from the Equator to sub-tropical latitudes (05°15′05″N to 33°45′09″S). Except for the southernmost region, dengue is endemic throughout the country. Within this dengue endemic zone, five mid sized cities representing different climate regimes were chosen for this study (Figs. 1 and 2). In each city, three non-adjacent 1 km2 areas with roughly equivalent demography and geographical characteristics were chosen as study sites, totalizing 15 study areas:
Santarém (STR), North Region. 2°26′35″S, 54°42′29″W, 20m elevation. Located in the Amazon region, this city has a tropical climate with no dry season(Köppen-Geiger type Af), wit warm temperatures year around (23–33°C) and a monsoon season from January to May when precipitation may reach 500 mm in a single month. The dengue fever season tends to coincide with the monsoon period.
Parnamirim (PNM), Northeast Region. 5°54′56″ S, 35°15′46″ W, 26m elevation. It is a tropical city (Köppen-Geiger type As), with warm climate year around (summer: 24–29°C, winter: 22–27°C). Compared to Santarém, Parnamirim is considerably drier, with monthly precipitation reaching its maximum in April-May, with less than 200 mm / month. Dengue transmission intensifies during the wet-warmer season.
Duque de Caxias (DQC) and Nova Iguaçu (NIG), Southeast Region. 22°47′08″ S, 43°18′42″, 7m elevation. Adjacent cities located in a lowland region, with tropical climate (Köppen-Geiger type Aw/Am). Heavy precipitation occurs during summer (December to February), with occasional floods. Temperature differences between summer and winter are more pronounced than in the North and Northeast Regions. Dengue incidence often peaks in March, at the end of the rainy season. Due to logistic reasons, survey was conducted in DQC in the first year and shifted to NIG during the second year. Both cities belong to the same metropolitan area, are similar in demography, land use and geographical features (S1 Table in S1 Text).
Campo Grande (CGR), Central-West Region. 20°26′34″ S, 54°38′47″ W, 600m elevation. With a highland tropical climate (Köppen-Geiger type Aw), this city presents the highest temperature and precipitation amplitudes among the localities under study. The winters are particularly dry and cool (Fig. 2). Dengue incidence historically peaks at late summer, at the end of the rainy season.
Four traps were employed: Adultrap (ADT), BG-Sentinel (BGS), MosquiTRAP (MQT) and Ovitrap (OVT). Simultaneously, larval surveys were performed according to the routine policies of the Brazilian Dengue Control Program [1]. In all cases, trap installation or larval survey depended on householders’ oral consent. To ensure good spatial coverage, each 1 km2 area was subdivided into 4 sub-areas of 250×250 m and 1/4 of the traps was installed in each sub-area. To select houses, a random set of geographical coordinates was taken and once in the field, the closest house was chosen. The number of traps installed per km2 varied between trap types and seek to obey their manufacturer’s recommendations or, in its absence, previous recommendations from the scientific literature [14].
Except for DQC and NIG, in each locality, fieldwork lasted 24 months (details in S1 Table in S1 Text) with monthly samplings. In DQC and NIG, the survey lasted 12 months each. A brief description of the main characteristics of each trap, as well as the amount and schedule of traps installation in each 1 km2 area are detailed below.
Adultrap (ADT). Adultrap is designed to capture gravid Ae. aegypti females during oviposition, using water as its principal attractant. A large hole on the top is the main entrance for attracted females that become trapped in the interior chamber. Water remains confined in a compartment at the bottom of the trap that cannot be reached by trapped mosquitoes, avoiding egg laying [15].
Adultrap was tested using two different approaches: exposure of 240 trap units /km2 during 24 hours (in the first year) or 100 units/km2 during 4 days (in the second year). Although the first approach is the one recommended by the manufacturer, the percentage of positive traps never reached 5% in the first year. Modifications of the protocol aimed at improving the surveillance ability of this trap.
BG-Sentinel (BGS). It is a collapsible bucket with a white gauze covering its opening. In the middle of the gauze cover, there is a black tube through which a down flow is created by an electric power exhauster fan that captures mosquitoes flying in the vicinity of the opening into a catch bag. An attractant (BG-Lure) releases sinthetic human skin odors that attracts preferentially host-seeking females [10].
A total of 24 BG-Sentinels were installed during 24 hours per 1 km2 area. Differently from the other traps, BGS were installed indoors due to their energy requirement.
Ovitrap (OVT). This trap consists of a black plastic container filled with 300ml of hay infusion. A wooden paddle held vertically on the wall serves as substrate for mosquito oviposition [4]. After a few days paddles are removed and laid eggs counted. In this study, in each 1 km2 area, 120 ovitraps were installed and exposed for five days in a place located in a shaded peridomestic environment. The number of ovitraps was based on [14].
MosquiTRAP (MQT). This trap is designed to collect gravid Ae. aegypti females [9]. It is made of a matte-black container filled with 300 ml of water, and requires a synthetic attractant (AtrAedes), and an adhesive card. Attracted ovipositing females stick to the adhesive card.
Monthly, 32 MosquiTRAP units/km2 were installed, and removed after seven days. The sample size was twice the size recommended by the manufacturer. This choice was based on results from [13] and sought to improve its sensitivity. Identification and counting was carried out in the field, with the help of a magnifying glass, as recommended by the manufacturer. However, identification of these same samples in the laboratory revealed significant differences both in the total numbers of mosquitoes recorded and in the amount of specimens identified as Ae. Aegypti [15]. MosquiTRAPs were introduced in the study only in the second year.
Larval surveys. This survey was carried out concomitantly with the Adultrap installation, always in the same dwellings. Briefly, after inspection of all potential breeding sites, samples of mosquito immatures were collected, and brought to the laboratory to be identified up to the species level. The House Index (HI) waa calculated for each 1 km2 area.
For each trap, the positivity index was defined as the proportion of traps with at least one capture (one mosquito or one egg), relative to the total units installed and successfully retrieved. In the same way, density indexes were calculated dividing the total number of mosquitoes (or eggs) captured in a given area by the total number of inspected traps.
Daily meteorological data were obtained from the Brazilian Environmental Information System (SISAM http://sisam.cptec.inpe.br/sisam/). Weekly temperature and precipitation statistics were calculated as mean, minimum and maximum values. For each of these variables, lagged values (l = 0, 1, 2 and 3 weeks), were also calculated, considering the 7 days before trap collection as lag 0.
Comparison between positivity and density indices. Within each city, scatter plots of density versus positivity indices evidenced a nonlinear relationship. This pattern was expected, as positivity indices are bounded to [0, 100] while density indices are unbounded. In other words, as population increases, the probability of ovipositing in an empty trap decreases. To formalize this observation and compare areas, a linear model of the form Positivity = b * Density and a nonlinear asymptotic model of the form Positivity = a*Density/(K+density) were fitted to each trap data from each locality. A likelihood ratio test was used to check if the nonlinear model provided a significantly better fit than the linear model. The likelihood ratio test compares the difference of the likelihoods of the two models (multiplied by two) to a chi-square distribution. As the fitted models for the same trap tended to be similar among cities, further modeling was carried out combining data from all localities into a single model. Model fitting used the nls function in R 2.12.1 [16] and the likelihood ratio test used the function lrtest in library lmtest [17]
Regression models. To infer associations between meteorological variables and trap indices, linear regression model with the identity link function was used. The outcome variable was either the density or the positivity index. Separate models were fit to each trap and each city. To attend the assumption of normality, density and positivity indices were square-root transformed before modeling.
Modeling was carried out step-wisely. First, the best lag for each meteorological variable (Tmin_l, Tmed_l, Tmax_l, Rain_l, l = 0, 1, 2 or 3 weeks) was chosen based on the Akaike Information Criterion (AIC). Secondly, the variable Neighborhood (Neig, meaning each of the 1km2 areas) was included in all models. The variable Year (referring to the study year, first or second year) was also included in the studies that lasted two years. This variable accounted for differences in surveillance efficiency between years, an approach similar to [16]. Only in the Southeast cities, NIG and DQC, models did not include Year, as their surveillance lasted one year only. Since residual analysis of the final models suggested significant autocorrelation of residuals at lag 1, a first order autocorrelation term was also added to all models.
Following this procedure, a model for each city was developed containing Neig, Year and the set of meteorological variables that were significant for at least one trap within that city. With this procedure, for each city, a common model was available for all concomitantly used traps, allowing comparison between trap systems. Interactions between Neig and climate and between Year and climate were also assessed and found to be significant in some instances. In these cases, the interaction was included in all trap models for that city. With this approach, we were able to check if a given meteorological variable similarly affected all traps within a city, even if non-significantly.
All statistical analysis were carried out using the software R 2.12.1 [17], library mgcv [18]. Data is available in S1 Dataset in Supplementary material.
Fig. 3 shows the range of House Index (HI) and trap density indices measured during the survey. Equivalent panels with trap positivity indices are found in S1 Fig. in S1 Text. Overall, the Southeast cities, NIG and DQC, presented the highest infestation levels among the evaluated sites. Although these two cities have been surveyed in different years, both presented very similar infestation magnitudes according to OVT and BGS measurements—the two traps employed in both cities with the same protocol (see Methods section: in the second year MQT was introduced and ADT protocol was changed). The larval survey detected high premise indices in DQC but failed to detect infestation in NIG. In contrast, according to ADT, infestation was higher in NIG. However, this may have resulted from the increased ADT exposure applied in the second year in NIG (4 days), compared to the 1-day exposure of this trap in the first year, in DQC.
The House Index was consistently low in all cities, except in DQC. The median was always 0, that is, in 50% of months, not a single household was found positive. CGR, which experienced a large dengue outbreak during the survey, exhibited the lowest HI. The traps, on the other hand, rarely produced null indices, pointing to their greater sensitivity in detecting the presence of Ae. aegypti in comparison to the larval survey. STR, CGR and PNM tended to present lower infestation indices when compared to DQC and NIG, according to all traps. Ovitraps in CGR and PNM captured the least number of eggs; BGS captured the least number of Ae. aegypti in PNM and STR; ADT exposed during 4 days tended to capture more mosquitoes than after only 1 day exposure.
Fig. 4 shows the scatterplot of trap positivity versus density indices. The nonlinear model was the best model for all traps but the ADT in Campo Grande (p-value = 0.02) (S2 Table in S1 Text). Remarkably, for each specific trap, the nonlinear relationship between positivity and density was very consistent among cities, despite the differences in climate and mosquito abundance. The parameters of the nonlinear model fitted to the combined data from all cities are shown in Table 1. It is evident from both Fig. 4 and Table 1 that for each trap, positivity indices plateau at different mosquito densities. For example, the ovitrap positivity index tends to plateau at 80%, when egg density exceeds ca. 50 eggs/trap. BGS positivity index shows signs of nonlinear behavior as mosquito density exceeded ca. 1.0 mosquito/trap. MQT positivity never reached values as high as those attained by OVT and BGS. Positive MQT rarely exceeded 40%. Nonlinear relationship between MQT positivity and density indices is stronger when mosquito density is above 0.5 mosquitoes/MQT.
ADT showed less evidence of nonlinear association between positivity and density indices. It was also the trap with the lowest positivity indices, never exceeding 20%. This weak nonlinearity is explained by the fact that each ADT rarely captures more than 1 mosquito, thus, density and positivity indices tend to equate.
Figs. 5–9 show the time series of Ae. aegypti positivity and density indices using ADT, BGS, MQT and OVT traps for each studied city. House indices and meteorological time series are also shown. Overall, positivity and density indices produced very similar temporal patterns. This agreement is confirmed by the qualitative similarity of the regression models fitted to both indices (S3–S7 Tables in S1 Text). Below, we describe the association between infestation and climate referring only to the density indices, but the conclusions are directly applicable to the positivity indices as well.
Ae. aegypti dynamics in Central Brazil (CGR). The highland tropical city of Campo Grande has the most marked seasonality among the study sites. The surveyed period consisted of two typical warm-wet summers and two dry-cool winters (Fig. 5). Temperature and precipitation are weakly positively correlated (Pearson’s r = 0.2, p = 0.09). Mosquito density, as measured by ADT, BGS and OVT (used during the 2 years) presented very consistent patterns among study areas and traps, with mosquito population peaking during the summer months, and drastically dropping during the dry winters. Ae. aegypti density, as measured by these three traps, is strongly and positively correlated with minimum or mean air temperature at lags 1 week (OVT) or 2 weeks (ADT and BGS) (S3 Table in S1 Text). MQT, included in the study during the second year, presented a similar pattern, but not in all areas. Precipitation also showed correlation with mosquito index, but lost significance in a model with temperature. Among the adult traps, BGS was the most efficient trap, generating indices up to 3 mosquitoes/trap, while ADT and MQT never exceeded 1 mosquito/trap.
The House Index in CGR presented very low values throughout the study, with no seasonal signal. HI was null during the 2010 dengue epidemic, which attacked 3.8% of the population, with a peak incidence of 1105 cases: 10.000 inhabitants in February. Considering an alert threshold of 1 mosquito/trap [19], BGS would launch an alert for this epidemic.
Ae. aegypti dynamics in Northeastern Brazil (PNM). Differently from CGR, climate in Parnamirim is warm year around, with a temperature regimen that is always favorable to the development of Ae. aegypti (Fig. 6). Precipitation usually spreads through the first six months of the year, increasing in intensity during the winter months.
In the course of the study, PNM presented contrasting precipitation regimens, the first year being very dry, with 6 months without precipitation, while the second year was very wet. The house index showed strong spatial heterogeneity, with one site presenting HI > 1% in most of the study, and 2 sites with HI < 1%. No trap confirmed this difference among neighborhoods. This result may indicate variation in the types of breeding sites affecting the immatures’ search success in different areas.
No adult trap ever exceeded 1 mosquito/trap in PNM, as was observed in CGR, suggesting lower infestation. Ovitrap indices, on the other hand, reached higher values in PNM than in CGR.
ADT and OVT detected higher mosquito abundance in the second year than in the first year, but only for ADT this difference between the two years was significant (S4 Table in S1 Text). This is certainly due to the protocol modification implemented for ADT in the second year (although the same is not observed in the other cities). Regression models show that minimum or maximum temperatures at lags 1 or 3 weeks are strong positive predictors of mosquito abundance in Parnamirim, according to BGS, ADT, and OVT. On the other hand, precipitation did not show a consistent association with mosquito density. While OVT and BGS detected positive effects at lag 1 week, ADT and MQT detected weak negative effects at lag 0. ADT, OVT and MQT are traps that attract female mosquitoes searching a place for oviposition. The ADT and MQT’s negative association with precipitation at lag 0 may be explained by the fact that strong precipitation affects flight behavior, does reducing the searching effort. Association with precipitation at lag 1, detected by the BGS, may reflect direct effects of precipitation on egg hatching and subsequent adult recruitment.
Ae. aegypti dynamics in Northern Brazil (STR). Santarém contrasts with the previous cities by its constant high temperatures and its monsoon season (Fig. 7). The two years of study were very similar in climate but very different in mosquito abundance. The second year presented significantly less mosquitoes than the first year, and this was attributed to the implementation of a new vector control program during the second year. This difference between years, perceived by all methods, was treated in the regression model by fitting the climate effects within each year (S5 Table in S1 Text).
Although small, a temperature variation does occur in Santarém, in response to clouding. In months with high precipitation, less solar incidence reduces the temperature a few degrees. This effect causes a negative correlation between precipitation and temperature (Pearson’s r = -0.19, p = 0.1). The regression models identified a negative effect of minimum temperature at lag 0 on mosquito abundance in Santarém, according to ADT, BGS and OVT. Positive association with precipitation was found in the second year (ADT and BGS), while a negative association was found in the first year (BGS). The only trap giving a very different seasonal signal was MQT, which detected greater mosquito abundance during the dry season instead of the wet season. A study conducted in Manaus (similar climate) comparing BGS and MQT found the same results: while BGS tended to find more mosquitoes during the wet season, MQT found more in the dry season [20]. One possible explanation for MQT increased capture during the dry season could be the effect of reducing the number of competing breeding sites; however, as the same effect was not observed in the OVT and ADT time series, this effect might be related to other causes, for example, the efficiency of the adhesive card or other technical features of the MQT.
Ae. aegypti dynamics in Southeastern Brazil (DQC and NIG). Duque de Caxias (DQC), where field work was conducted in the first year, suffered an extreme flooding event during the summer which caused severe disruption of the city’s services (Fig. 8). In DQC, all traps (ADT, OVT and BGS) as well as the HI, detected higher mosquito abundance during the summer months with a positive association with minimum temperature at lags 2 or 3 (S6 Table in S1 Text). Precipitation at lag 0 was negatively associated with infestation, likely due to the flooding event. This seasonal pattern is consistent throughout the whole study area as measured by ADT and OVT, but appears only in some neighborhoods, according to BGS. MQT was not used in Duque de Caxias.
Nova Iguaçu (NIG), the adjacent city, entered the study in the second year, replacing DQC. Climate is similar in both cities, but during the second year, no flooding events were recorded (Fig. 9). Trap indices in NIG tended to increase during the summer months. BGS, MQT and OVT detected strong association with temperature at lags 1 to 3 (S7 Table in S1 Text). On the other hand, ADT and HI behaved differently, with a negative association between trap positivity and temperature. These two measurements were taken exactly in the same dwellings. An interaction between temperature and area was detected by the ovitrap and the MosquiTRAP, but the latter was not significant. Association between precipitation and mosquito abundance was inconsistent, while BGS detected a significant negative association, all the other traps detected positive associations.
Dengue vector surveillance is a time and resource consuming activity in many tropical countries. In Brazil, it is estimated that more than 300 million dollars are spent in this activity every year. In this country and many other dengue endemic countries, surveillance protocols are based on larval inspections [21]. Larval surveys are good for identifying key containers, but often fail in providing fast and localized measurements of mosquito abundance. Traps are presented as a complementary approach for dengue vector surveillance and several studies have tested their sensitivity and efficacy. This study expands this discussion by presenting results from a large scale project including five dengue endemic cities simultaneously monitored by four different trap schemes besides the standard immature mosquito survey. Our goal was to reproduce, as well as possible, the real conditions to be faced by a surveillance program using the human and infrastructure resources present in each city. It is important to note, however, that we chose Brazilian municipalities with a prominent record of dengue vector control initiatives and a better than average infrastructure.
Overall, during the study, the house indices rarely reached values above the 4% alert threshold, the only exception being DQC. Still, all cities recorded dengue transmission during the study period and CGR reported a large epidemic. For many reasons, the standard larval survey was not capable to issue proper alerts. For example, failure to detect larval breeding sites despite the presence of high trap indices may indicate that female mosquitoes are choosing more cryptic places to deposit their eggs, such as clogged rain gutters and pipes [7].
In comparison, all traps detected increased mosquito infestation during the dengue transmission seasons, indicating their ability to detect mosquito density variation. Previous studies have compared larval surveys to ovitraps finding the latter more sensitive [7, 22–24], and cost-effective at low mosquito densities [25]. This greater sensitivity of traps to mosquito density variation is probably due to their ability to cover more than one premise while immature surveys only encompass those houses included in the sample. In CGR, BGS was the only adult trap to capture more than 1 mosquito/trap during the dengue epidemic. The other traps exhibited less efficient mosquito collection, but note that MQT was not in use during this period.
This is not the first study to compare traps. Ovitraps were more sensitive than MosquiTRAPs in the low infestation season in Belo Horizonte, Brazil [26] and Pedro Leopoldo, Brazil [19]. In mark-release-recapture experiments, MosquiTRAPs captured more marked mosquitoes than Adultraps [27]. In Rio de Janeiro, Brazil, an experiment with the concomitant distribution of ovitraps and MosquiTRAPs, resulted in greater ovitrap positivity indices [14, 28]. Despite these variations, in general traps perform better than the larval surveys.
One practical question that arises in a trap based surveillance program is the possibility of using positivity as a proxy for the more time consuming density indices. If both measures were linearly correlated, this approximation would be easily defended. However, our results suggest that under field conditions, this does not occur for any of the traps. This is attributed to the aggregated spatial distribution of Ae. aegypti, typical of many insects [29, 30]. Mogi et al [25] and Ho et al [24] found that ovitrap data fit reasonably well to the empirical model developed by Gerrard and Chaing [31] for aggregated mosquito distributions. Differently from these authors, here we used the Michaelis-Menten model to represent this association. One advantage of the Michaelis-Menten model is the interpretation of its parameters. The parameter“a” stands for the maximum positivity, that is, the positivity index as density tends to infinite; the parameter“K” stands for the mosquito density when positivity index is“a“/2. In other words, above positivity = a/2, the association between positivity and density tend to vanish as positivity saturates.
Ae. aegypti positivity-density relationship varied among traps but was consistent among cities, despite the differences in climate (Fig. 4). These intrinsic sensitivity differences among traps have surveillance implications, that is, premises within the study area can be considered negative or positive depending on the method. Reasons for this difference can be attributed to trap attractiveness (more attractive traps covering larger geographical areas), and to different target populations.
A trap based surveillance program should be able to launch alerts if the trap index crosses a predefined threshold. Ideally, such threshold should be estimated based on the minimum mosquito density required to sustain dengue transmission. However, this absolute number is not known for any of the studied traps. A practical alternative would be to define an empirical value, based on the range of densities observed in the past during dengue transmission seasons. For instance, in Australia, control measures are strengthened when sticky ovitrap density index are above 1 mosquito/trap [32]; in Brazil, the MI-Dengue surveillance service uses a threshold of 2 mosquitoes/MQT to launch an alert signal [33].
Now, let’s suppose this threshold is a certain quantity of mosquitoes/trap (mc). A trap positivity index will be a good proxy for this density index only if mc < K. Under this condition, the positivity-density conversion is done in a situation where positivity and density indices show high covariation. If mc > K, on the other hand, decision is done in a parameter region where positivity index has low precision and small variations in positivity may represent large variations in density. Applying this rule to the traps under study, we conclude that for ADT, the condition mc < K is likely to be met in most conditions, that is, ADT positivity index is a good proxy for ADT density index. This is a straightforward conclusion anyway, since ADT often captures no more than one mosquito per trap. In the other extreme, OVT’s estimated K was 28–36 eggs/trap which is relatively low compared to the observed range of density observed. Thus, the ovitrap density index will only be reasonably approximated by the positivity index if mc is equal to low to moderate mosquito densities. The same seems to apply to BGS and MQT (see S2 Table in S1 Text). In the absence of a well established threshold, density measurements are preferable, for more sensitive traps, as BGT and OVT.
Aedes aegypti geographical range is roughly limited to the intertropical area (35°N–35°S), where temperature is mostly above 10°C year around. Our study covered several latitudes within this range and showed that different tropical climates have distinct seasonal patterns of mosquito abundance. Aedes aegypti population oscillations were clearly related to seasonality in the tropical climates, even at the equator. Several studies have analyzed the effect of temperature, precipitation and even relative humidity on mosquito abundance in different parts of the globe, with sometimes inconsistent results [7, 14, 22, 24, 26, 29, 32, 34–41]. Compared to other tropical climates, the Af type has the weaker association between mosquito infestation and temperature, although association with precipitation remains significant.
Precipitation sometimes exerts negative effects on mosquito abundance. We found situations in which heavy precipitation events may have disturbed the traps and altered mosquito flight behavior (DQC, PNM and STR). This negative association has also been observed in Selangor, Malaysia [34]. As precipitation contributes to the creation of new breeding sites, it may also reduce the attractiveness of oviposition traps, causing low catching rates. In contrast, positive association of precipitation and mosquito abundance was found in CGR and NIG. Both are cities with typical tropical wet summers and dry winters. It is also relevant that the high correlation between temperature and precipitation makes it difficult to isolate the effects of each factor by regression analysis. The negative association between temperature and mosquito abundance in the equatorial STR was not expected. Since in this city lower temperature is associated with higher precipitation, this result was interpreted as an indirect measurement of precipitation.
Sensitivity to mosquito density variation is an important feature of a trap based surveillance scheme. Among the studied cities, CGR represents the best scenario to test such sensitivity, as it is the only one with strong seasonal variation. We found that Ovitraps, Adultraps and BG Sentinels were all capable of detecting large mosquito variations throughout the year, in a consistent way. This feature is less evident for MosquiTRAPs, which showed strong variation between the three sites. One disclaimer, though, is that MosquiTRAPs were used for a single year and further studies should be done to confirm these initial results.
Many authors have supported the use of adult traps in the place of ovitraps for surveillance. Our results show that ovitraps, although not measuring directly the adult population, do capture its variation very well. Actually, of all traps, ovitrap was the one with best sensitivity (never presented null indices), and strongest association with climate, and consistently followed the adult mosquito patterns detected by the adult traps. These results confirm the usefulness of ovitraps for Ae. aegypti surveillance, even if it does not produce direct indices of adult mosquito abundance.
At last, independently of the trap chosen, it should be considered that, although to a lesser extent than larval surveys, trap based indices are still dependent on the work quality of the field personnel. This is especially relevant in the scope of a broad and regular surveillance program and relates to aspects such as traps installation and collection, as well as specimens counting or identification.
This study was carried out to support the development of trap based surveillance programs in dengue endemic countries. Our main conclusions are that all investigated traps are valuable tools and could be considered in combination with vector control strategies to improve our response to dengue and other diseases transmitted by Ae. aegypti. Household larval surveys and trap based surveillance systems are not interchangeable approaches though. Household surveys are required for the identification of the major mosquito breeding sites in a given locality. This allows the design of adequate control or elimination strategies. Traps are useful for monitoring adult infestation levels and the impact of control strategies. Used together, they synergistically optimize both surveillance, prevention and control. Future studies should assess the cost-benefit of such integrated strategies. Other features should be also evaluated before choosing a trap for surveillance: specificity, low cost, ease of distribution, a consistent sampling profile [3]. This will be subject of future studies.
|
10.1371/journal.pgen.1003773 | Specific Tandem Repeats Are Sufficient for Paramutation-Induced Trans-Generational Silencing | Paramutation is a well-studied epigenetic phenomenon in which trans communication between two different alleles leads to meiotically heritable transcriptional silencing of one of the alleles. Paramutation at the b1 locus involves RNA-mediated transcriptional silencing and requires specific tandem repeats that generate siRNAs. This study addressed three important questions: 1) are the tandem repeats sufficient for paramutation, 2) do they need to be in an allelic position to mediate paramutation, and 3) is there an association between the ability to mediate paramutation and repeat DNA methylation levels? Paramutation was achieved using multiple transgenes containing the b1 tandem repeats, including events with tandem repeats of only one half of the repeat unit (413 bp), demonstrating that these sequences are sufficient for paramutation and an allelic position is not required for the repeats to communicate. Furthermore, the transgenic tandem repeats increased the expression of a reporter gene in maize, demonstrating the repeats contain transcriptional regulatory sequences. Transgene-mediated paramutation required the mediator of paramutation1 gene, which is necessary for endogenous paramutation, suggesting endogenous and transgene-mediated paramutation both require an RNA-mediated transcriptional silencing pathway. While all tested repeat transgenes produced small interfering RNAs (siRNAs), not all transgenes induced paramutation suggesting that, as with endogenous alleles, siRNA production is not sufficient for paramutation. The repeat transgene-induced silencing was less efficiently transmitted than silencing induced by the repeats of endogenous b1 alleles, which is always 100% efficient. The variability in the strength of the repeat transgene-induced silencing enabled testing whether the extent of DNA methylation within the repeats correlated with differences in efficiency of paramutation. Transgene-induced paramutation does not require extensive DNA methylation within the transgene. However, increased DNA methylation within the endogenous b1 repeats after transgene-induced paramutation was associated with stronger silencing of the endogenous allele.
| Paramutation is a fascinating process in which genes communicate to efficiently establish changes in their expression that are stably transmitted to future generations without any changes in DNA sequences. While paramutation was first described in the 1950s and extensively studied through the 1960s, its underlying mechanism remained mysterious for many years. Over the past ten years paramutation at the b1 locus in maize was shown to require transcribed, non-coding tandem repeats located 100 kb upstream of b1. These repeats generate small RNAs, and mutations in multiple genes mediating small RNA silencing at the transcriptional level prevent paramutation. While underlying mechanisms are shared, current models for RNA-mediated transcriptional silencing that are based on experiments with S. pombe and Arabidopsis do not explain many aspects of paramutation. In this manuscript we used a transgenic approach to demonstrate that the b1 non-coding tandem repeats are sufficient to send and respond to the paramutation signals and that this occurs even when the repeats are not at their normal chromosomal location.
| Paramutation is a trans-interaction between specific alleles or transgenes that leads to a meiotically heritable change in the expression of one of the participating alleles or transgenes. Originally described for the maize (Zea mays L.) r1 (red1) [1] and b1 (booster1) [2] genes, paramutation has since been reported for several other genes in plants (see e.g. [3]–[9]). Paramutation-like interactions have also been described in other species, including Drosophila [10], mammals and humans (for review see [11]).
Paramutation at the b1 locus provides a powerful system for dissecting the underlying mechanism of paramutation (reviewed in [12]). The b1 gene encodes a transcription factor that activates the purple anthocyanin biosynthesis pathway. Alterations of b1 expression lead to a visual change in plant pigmentation, and the amount of pigment is a read-out of the b1 transcription level [3]. The two b1 alleles that participate in paramutation are B-I (B-Intense) and B'; B-I is highly expressed and specifies dark purple pigmentation of the husk, sheath and tassel of the maize plant, while B' is expressed at a much lower level and specifies light streaky pigmentation in the same plant tissues as B-I [3], [13]. The high expressing B-I allele is unstable and can spontaneously change to B' at variable frequencies (can be up to 10%; [13]). In contrast, B' is very stable and does not change to B-I in wild-type genetic backgrounds [13], [14]. Paramutation occurs when B' and B-I alleles are combined in one nucleus by crossing. The “paramutagenic” B' allele turns the “paramutable” B-I allele into B' at a 100% frequency. The new B' allele (B-I in the previous generation) is as heritable and paramutagenic as the original B' allele [13]. Alleles that do not participate in paramutation are referred to as neutral [14].
Genetic screens in maize have uncovered a number of genes required for paramutation (reviewed in [12], [15], [16]). All but one gene [17] identified to date share homology with genes involved in the RNA-directed transcriptional silencing pathway in Arabidopsis [18], strongly indicating a requirement of this pathway for paramutation.
A necessary step towards further dissecting the mechanism of paramutation is knowledge of the key sequences mediating paramutation, the subject of this work. Previous fine structure recombination studies between B' or B-I and neutral b1 alleles revealed that paramutation requires a region spanning ∼6 kb located ∼100 kb upstream of the b1 transcription start site [19], [20]. This region was also required for high b1 expression. In B' and B-I, this region contains seven tandem repeats of an 853-bp sequence that is unique to this location within the maize genome. Notably, an allelic series in which alleles differed only by the number of repeats revealed that multiple repeats are required for paramutation. Alleles with seven and five repeats were fully paramutagenic, alleles with three repeats had reduced paramutagenicity, and alleles with a single repeat were neutral to paramutation [19].
The B' and B-I alleles are epialleles as they have identical DNA sequences [20]. Consistent with epigenetic regulatory mechanisms defining the B' and B-I states, the hepta-repeats have distinct chromatin structures in B' and B-I [19], [21]. The epigenetic mark that correlates best with paramutation ability is DNA methylation. The B' repeats have extensive DNA methylation, while the B-I repeats have low levels of DNA methylation [21]. There are differences between the alleles in histone modifications and the extent of chromosomal looping between the repeats and the b1 promoter, but these differences correlate mainly with tissue-specific expression, not the heritable silencing associated with paramutation [21], [22].
The b1 tandem repeats are transcribed [23] and generate siRNAs [24], yet repeat siRNAs are produced even from alleles that do not participate in paramutation, suggesting b1 siRNAs are not sufficient for paramutation in the tissues analyzed [24]. However, when the repeat sequence is expressed as a hairpin RNA from a transgene, which generates much higher levels of siRNAs than the endogenous alleles, heritable silencing and paramutation can be reconstructed [24]. This contrasts with two other examples of siRNAs generated from hairpin RNA producing transgenes in maize. These siRNAs effectively silenced homologous promoters, yet that silencing was not heritable [24]. Similar studies using hairpin RNAs to silence promoters in Arabidopsis did not report on heritability (e.g. [25]).
We hypothesize that the tandem repeats of the B-I and B' epialleles have special properties, which confer the ability to establish and heritably transmit the silenced paramutagenic state of B'. In this study, we test this hypothesis by asking whether the tandem repeats themselves are sufficient to send and respond to trans-acting paramutation signals, using a series of transgenes containing b1 tandem repeats. Our results are consistent with the above hypothesis. While paramutation was effectively reconstituted, the repeat transgene-induced silencing of B-I was less frequent and showed reduced stability in the next generation relative to endogenous paramutation, which occurs 100% of the time and is always stably transmitted.
To test whether the b1 sequences upstream of the transcription start site (TSS) could induce silencing of the B-I allele from a non allelic position, two constructs carrying the b1 repeats and surrounding sequences were used to generate transgenic maize lines: pB, containing the 5′ part of the b1 transcription unit and 106.2 kb of sequences upstream of the ATG (Figure 1A, [19], [20]) and pBΔ, which had 91.6 kb deleted between the tandem repeats and the proximal promoter of the b1 transcription unit relative to pB (Figure 1A). These constructs allowed us to also address if, in addition to the tandem repeats, other sequences upstream of the TSS were required for paramutation. For example, the observed transcription of the repeats [23], [24] is likely to be required for paramutation and the promoter sequences driving this transcription might be located outside of the repeats.
The Hi-II maize stock used for transformation carried recessive neutral b1 alleles (designated as b-N) that do not participate in paramutation and do not confer anthocyanin plant pigment (V. Chandler, unpublished data), enabling the monitoring of silencing activity of the transgenes after crossing the regenerated transgenic plants to B-I. To test whether the sequences within either construct could mediate B-I silencing, the primary transgenic plants were crossed with plants carrying the paramutable B-I allele and a neutral b-N allele (Figure 2A). The presence of the neutral allele provided a means to propagate the transgenes in the absence of B-I (Figure 2A), which was done for multiple generations by crossing with b-N testers (Figure S1). To test the ability of transgenes to induce silencing, transgenic plants at different generations of propagation were crossed with B-I (Figure S1, Table S1). Scoring of plant pigment of the B-I/b-N progeny carrying transgene loci (TG/-) revealed that four out of ten pB, and five out of nine pBΔ transgene loci induced silencing of B-I (Figure 2B). In the transgenic events with silencing, the frequencies of silencing varied across multiple generations, ranging from 17 to 100% (Figure 2B, Table S2). The phenotypes of plants showing transgene-induced silencing of B-I were very similar to those showing B'-induced paramutation of B-I (Figure 2A and data not shown). In this paper, the transgene-induced silenced state of B-I is noted as B'# to signify the transgenic origin of this state, in contrast to paramutation induced by the endogenous B' allele. Non-transgenic sibling plants (B-I/b-N) served as controls for spontaneous paramutation of B-I to B', which can happen frequently [14]. Data from families showing spontaneous paramutation in non-transgenic siblings were not included in this paper.
Our results indicate that silencing of B-I can be mediated by sequences in ectopic, i.e. non-allelic, locations, paving the way for using a transgenic approach to further dissect the minimal sequences required for paramutation. Furthermore, these results demonstrate that a sub-fragment of the b1 locus, containing primarily the tandem repeats and the 5′ part of the b1 transcription unit, is sufficient to establish B-I silencing.
The most prominent feature within the 16.3 kb sequence contained in the pBΔ construct are the seven 853 bp tandem repeats, and as paramutation strength correlates with the number of repeats [19], they were strong candidates for the minimal sequences mediating paramutation. To determine which part of the repeat sequence is needed to induce silencing, the 853 bp tandem repeat unit was dissected into two halves based on their different GC content; one half (hereafter referred to as FA) is 48% AT-rich, while the other half (hereafter referred to as FB) is 68% AT-rich [19] (Figure 1B). PCR-amplified sub-fragments (FA or FB halves) were ligated in head-to-tail orientation to form seven tandem repeats (Figure 1B). Constructs carrying the FA and FB hepta-repeats, pFA and pFB, were then transformed into maize and the resulting twelve transgenic events were tested for their ability to induce B-I silencing, similar to the approach used for pB and pBΔ transgenic loci (Figure 2A). Results revealed that all four pFA transgenic events induced B-I silencing at 100% frequency, indicating the pFA transgene contains all sequences sufficient for trans-silencing (Figure 2B, Table S2). None of the eight pFB transgenic events induced B-I silencing (Figure 2B, Table S2), suggesting the FB sequences were not sufficient for trans-silencing. Because pFB transgenic events do not induce silencing they serve as controls demonstrating that specific repeated sequences mediate silencing of B-I.
One of the defining features of b1 paramutation is that B' is fully paramutagenic to B-I and the silencing is heritable [13]. To assay whether the transgene-induced B'# silenced state was heritable and paramutagenic, plants carrying B'# alleles, induced by three independent pBΔ and two independent pFA transgenic loci, were crossed with plants heterozygous for the paramutable B-I and a neutral b-N allele (Figure S2A). Assaying the phenotype of the resulting non-transgenic B'#/b-N progeny revealed that the silenced B'# phenotype was heritable in the majority (78–100%) of the non-transgenic plants (Table 1b). Assaying the B'#/B-I non-transgenic progeny revealed that the B'# states were often paramutagenic (41–100%; Table 1d). To distinguish the various epigenetic states, we use B'∧ to signify a B-I allele silenced by B'#. Together, our results demonstrate that pBΔ and pFA-induced silencing of B-I to B'# can recapitulate the two key characteristics of paramutation; the silenced B'# state can be transmitted to progeny and it can be paramutagenic, inducing the B'∧ silenced state in the absence of the inducing transgene.
Unlike the state induced by the B' allele, the heritability and paramutagenicity of the B'# state was not fully penetrant and the frequency varied between the different pBΔ transgenic events. To test whether prolonged exposure to the pBΔ transgenes would increase the heritability and paramutagenicity of B'#, B'#/b-N; TG/- plants carrying B'# alleles that had been exposed to the transgenes for two subsequent generations were crossed with either b-N or B-I (Figure S2B). For all three pBΔ transgenic events tested, subsequent generation in the presence of the transgene increased the heritability and paramutagenicity of the B'# state to 100% (Table 1ce). This could be because of prolonged in trans interactions between the transgene and B'#. Spontaneous paramutation of B-I can, however, not be completely ruled out.
Roughly half of the pB and pBΔ transgenic events, and all of the pFB transgenic events were not paramutagenic (Figure 2B, Table S2). As the endogenous B' and B-I alleles have identical DNA sequences but differ in chromatin structure, expression levels and paramutation properties [19], [21], one possibility was that the transgenic events that were not paramutagenic might have assumed a B-I-like epigenetic state upon integration. If that was true, such transgenes should become paramutagenic upon exposure to B'. To test this hypothesis, b-N/b-N; TG/- F1 plants (as indicated in Figure 2A and Figure S3), which had never been crossed to B-I, but whose siblings crossed to B-I demonstrated they carried non-paramutagenic or weakly paramutagenic transgenic events, were crossed to B'. The resulting transgenic progeny plants were then crossed to B-I to determine if the paramutagenicity of the transgenes had increased (crosses described in Figure S3). Results shown in Table 2 demonstrate that four out of seven pB, and three out of four pBΔ transgenic events became highly paramutagenic.
One potential explanation for the increased paramutagenicity could be spontaneous paramutation of the transgenic loci instead of an interaction with B'. The frequency of spontaneous paramutation of the transgenes can be estimated by carrying the transgenes for multiple generations with only neutral b1 alleles and then testing their ability to induce paramutation of B-I (shown in Figure S1 and Table S1). While there was some variability from generation to generation among the weakly paramutagenic events, none of the weakly paramutagenic transgenes became fully paramutagenic unless crossed to B'. For example, with event number 3-46, its paramutation frequency ranged from 36 to 85% over six generations with neutral alleles. In contrast, after one generation with B', its paramutation frequency was 100%. Similarly, several transgenes only became paramutagenic upon crossing with B'. For example, event 4-06 was not paramutagenic when carried for four generations with neutral b-N alleles (0% paramutagenicity, Table S1), but became highly paramutagenic (97%) after only one generation with B' (Table 2). We refer to these transgenic events as paramutable to distinguish them from the paramutagenic transgenes, which did not require crosses with B' to become paramutagenic. The ability of certain transgenes to become paramutagenic only after exposure to B' suggested that upon integration these transgenes initially assumed a B-I-like state. The transgenic events that did not become paramutagenic, even after crossing with B', are referred to as neutral. In contrast to the majority of the pB and pBΔ transgenic events, none of the seven pFB transgenic events tested showed any paramutagenicity after exposure to B' (Table 2), suggesting that the repeat sequences in the pFB transgenes were not sufficient to receive and/or heritably transmit the paramutation signal.
Failure of some transgenic events to participate in paramutation could be attributed to several factors. Transgenes may be truncated and not carry tandem repeats, which are absolutely required for endogenous paramutation [19], or they may have integrated in genomic locations that prevent establishment of silencing. To determine how many events had the intact hepta-repeat fragment and to estimate the number of repeat units present in each event, DNA blot analyses (Materials and Methods) were performed on paramutagenic, paramutable and neutral events (see previous section for definitions). As is typical for biolistic transformation, the DNA blot analysis revealed that the pB and pBΔ transgenic plants contained multiple copies of the transgenes, including complete and truncated fragments (Figure 3A), which segregated as a single locus in each independent event. Six of the paramutagenic transgene loci carried an intact hepta-repeat fragment (Figure 3A, black arrow, 7 kb) and three paramutagenic events did not. None of the paramutable or neutral events carried an intact hepta-repeat. Thus, an intact hepta-repeat fragment was associated with paramutagenicity but was not absolutely necessary for an event to be paramutagenic or paramutable. As all of the insertions are complex we cannot rule out that one or more of the transgenic lines also contain repeats in an inverted orientation, a sequence arrangement known to mediate silencing [25], [26]. We favour our hypothesis that it is the tandem repeats mediating paramutation because it is unambiguous from the fine structure mapping that tandem repeats mediate endogenous paramutation [19] and all the transgenic events with an intact tandem hepta-repeat were paramutagenic.
The number of repeat units present within each event was estimated by normalizing to an endogenous fragment containing a single repeat unit (Materials and Methods). In each functional category, paramutagenic, paramutable or neutral, there are examples of transgenic events that have relatively high or low numbers of repeat units (Figure 3A). All transgenic events, except one neutral event (4–12), carried more than one copy of the 853 bp repeat unit. There was not an absolute correlation between the number of the repeats and paramutation activity in the transgenic events (Figure 3A), although all intact hepta-repeat events were paramutagenic. A similar lack of correlation was observed with the pFA and pFB transgenic events (Figure 3B). The pFA transgenic events, which were all highly paramutagenic, had lower copy numbers (6–9 repeat units) than most of the pFB events (seven out of eight events had 16 or more repeat copies), which showed no paramutation ability. In addition, most of the pFB events had an intact fragment containing seven repeats, while none of the pFA events did (Figure 3B). These results confirm that the pFA sequences are sufficient for paramutation, while the pFB sequences are not.
Relative to B-I, the paramutagenic B' allele has high levels of cytosine methylation within the tandem repeats [21]. To determine if there was a correlation between the frequency of paramutation and DNA methylation levels at the transgenic repeats, two pBΔ transgenic events, 3-39 and 3-46, were selected for DNA blot analysis. These two events have relatively simple transgene integrations; one intact hepta-repeat fragment and only a few other, truncated repeat-containing fragments (Figure 3A), enabling the interpretation of the DNA blot results. Representative examples of the 3-39 and 3-46 transgenic loci that were in the presence of neutral b-N alleles (in the immediate progeny of regenerated transgenic plants) and had not been exposed to B-I or B', are shown in Figure 4A. The transgenic repeats were mostly unmethylated within the assayed restriction sites (Figure 4A, open and grey arrows; a total of four 3-39 and seven 3-46 plants were examined). The repeat DNA methylation levels were not only lower than those previously observed for B' and for plants undergoing spontaneous paramutation of B-I to B', but were also lower than those observed for B-I (Figure 4B and 4D; Figure S4) [19], [21]. These results indicate that paramutation can be mediated by transgenic repeats that do not have the DNA methylation levels typical of B'.
To test if the methylation levels of the transgene increased after it had mediated paramutation, we examined the 3-39 transgene after it had segregated from the F1 between the primary 3-39 transgenic plant and B-I [In this F1, paramutation occurred at a frequency of 90%, (Table S1)]. The segregating 3-39 repeat transgene was extensively methylated, equivalent to B' (Figure 4B; three plants examined). Thus, after paramutation and segregation the transgene was extensively methylated. This could be due to spontaneous increases in DNA methylation or due to interactions between the transgene and the endogenous allele (resulting in paramutation of B-I to B'#), or both. To test for spontaneous DNA methylation within the repeats, we examined the 3-39 transgene maintained in the presence of neutral b1 alleles for four generations (never exposed to B-I or B'). We observed a spontaneous increase in the DNA methylation levels in the transgenic repeats (Figure 4C, black arrows, four plants tested) up to the levels observed for the endogenous B' repeats (Figure 4B and 4D). Thus, the increased methylation observed within the 3-39 transgenic repeats after encountering B-I could be due to spontaneous events.
The 3-46 transgenic event had very low levels of DNA methylation (Figure 4A) in the immediate progeny of the primary transgenic event, and when crossed with B-I plants, paramutation occurred at a frequency of 66% (Table S1). After crossing the 3-46 transgene with B' and then outcrossing to B-I, 100% paramutation was observed. With this one event, we saw that after crossing with B', both the transgene and B'# had acquired extensive DNA methylation (summarized in Figure 4D and data not shown; a total of six B' TG/- plants, and 11 B'# TG/- plants were tested). This result indicates that transgenic repeats with low levels of DNA methylation can acquire higher DNA methylation, but more events and individuals need to be examined to determine if increased paramutagenicity correlates with DNA methylation.
A key difference between transgene-mediated and endogenous allele-mediated paramutation is that the resulting silencing of B-I to B'# is less stable when induced by the transgenes than by B' (Table 1). To determine whether this difference in silencing, as measured by plant phenotypes, might correlate with the extent of repeat DNA methylation in the endogenous allele, non-transgenic progeny plants segregating B'# and displaying a range of pigment phenotypes were examined. These individuals derived from outcrossing the B'#/b-N; TG/- F1 to b-N (Figure S2A). Notably, DNA methylation levels within the B'# repeats, induced by the 3-39 transgene, varied and this variation correlated with the extent of silencing; the more DNA methylation, the lower the plant pigment levels, which are a read-out of the level of B'# silencing (Figure 4B and data not shown). The same correlation between the extent of silencing and DNA methylation was observed for the 3-46 transgene (data not shown). While the number of individuals examined is small (four 3-39 and six 3-46 plants looked at in total), these data are consistent with a correlation between the level of B'# silencing and extent of DNA methylation within the endogenous repeats.
Paramutation by the endogenous B' allele requires the Mop1 gene [23], which encodes a protein with high similarity to RDR2, a putative RNA-dependent RNA polymerase required for RNA-directed transcriptional silencing in Arabidopsis [27]. To test whether MOP1 is required for the transgene-induced paramutation, the appropriate crosses were done to assay the ability of three pBΔ transgenes to paramutate B-I in the presence of the mop1-1 mutation (Figure S5). If paramutation was prevented, the segregating progeny should have the B-I phenotype, whereas if paramutation occurred, most progeny should have the B' phenotype. Analysis of the segregating non-transgenic progeny revealed that the majority of the plants had a B-I phenotype (Table 3), indicating that the mop1-1 mutation prevented the pBΔ transgenes from paramutating B-I to B'#. A few light B' plants were observed in three out of twelve testcross families. These could be the result of spontaneous paramutation of B-I to B', or because paramutation was not fully prevented in all plants. The observation that MOP1 is required for the transgenes to silence B-I demonstrates RNA-mediated mechanisms are involved in transgene-induced paramutation of B-I.
In addition to mediating silencing, multiple b1 tandem repeats are required for high B-I expression [19]. It is, however, not known if the repeats are sufficient to mediate high expression or whether additional sequences are needed. To test if the repeats can mediate high expression a construct was produced in which the seven tandem repeats of B', B-I (b1TR) were fused to the minimal −90 bp Cauliflower Mosaic Virus 35S promoter (35S) [28] and the GUS (beta-glucuronidase, [29]) reporter gene to generate the pb1TR::GUS transgene (Figure 5A, Materials and Methods). As a negative control, a construct was made that carried only the minimal −90 bp 35S promoter fused to GUS (p35S::GUS). Both constructs were used to generate transgenic maize lines; only lines carrying intact GUS reporter genes were examined for GUS activity (Materials and Methods). Sheath and husk tissues were stained for GUS activity and scored using a graded scale shown in Figure 5B. High GUS activity was observed in pb1TR::GUS events 36-7, 36-21 and 36-31, but not in the event 36-11 (Figure 5C). Southern blot analysis (not shown) revealed that the GUS transgenes in events showing high GUS activity carried about ∼7 repeats (36-7, 1 transgene copy), 6 and 1.5 repeats (36-21, 2 copies), and 4 and 3 repeats (36-31, 2 copies), while the transgenes in the event showing weak GUS activity carried about 3.5 and 2.5 repeats (35-11, 2 copies). Three p35S::GUS control events that contained no repeats showed low GUS activity, while one had high GUS activity (34-10, Figure 5D).
The high GUS activity in the p35S::GUS event 34-10 was unexpected and was hypothesized to be caused by integration of the transgene near an endogenous transcriptional regulatory element. If this hypothesis was correct, the expectation was that the GUS activity should not be silenced by B'. In contrast, if the high expression in the pb1TR::GUS events 36-7, 36-21 and 36-31 was mediated by the repeats, B' should silence that expression. To test these hypotheses, the p35S::GUS event (34-10) and the three pb1TR::GUS transgenic events strongly expressing GUS (36-7, 36-21 and 36-31) were crossed with the paramutagenic B' allele. The three pb1TR::GUS transgenic events were also crossed with two highly paramutagenic pBΔ transgenic events. Results shown in Figure 5E revealed that the expression of p35S::GUS event 34-10 was not affected by B', consistent with the hypothesis that its high expression is caused by integration near an endogenous regulatory element that is insensitive to B'. In contrast, all three pb1TR::GUS transgenic events exhibited a significant reduction in GUS activity after exposure to B' (Figure 5F) or the paramutagenic pBΔ transgenes (Figure 5G), providing additional support that the high expression was not simply due to insertion next to an endogenous enhancer. The silencing of the pb1TR::GUS transgenic loci in the presence of the paramutagenic pBΔ transgenes was not due to spontaneous paramutation, because for all three pb1TR::GUS loci control transgenic siblings segregating only the pb1TR::GUS transgenes showed higher GUS activity (Figure 5G). Together, these data suggest that the b1 tandem repeats are sufficient to trigger expression of a heterologous gene and that this expression is sensitive to paramutation.
To determine if the transcriptional regulatory activity within the repeats could be further delineated, transgenic lines containing seven FA or seven FB tandem repeats fused to the minimal p35S::GUS reporter gene were generated (Figure S6). GUS expression was observed in all the four intact pFA::GUS events and the one intact pFB::GUS event. However, because there was only one intact pFB::GUS event available, more experiments will be required to delineate where the transcriptional regulatory activity maps.
Previous studies have shown that the tandem repeats in B-I and B' produce siRNAs [24]. Therefore various repeat transgenes were tested for the production of b1 repeat siRNAs from their ectopic locations. As b1 alleles that have a single copy of the repeat unit, and do not participate in paramutation, also produce b1 repeat siRNAs [24], non-transgenic siblings with the same b1 genotype as their transgenic counterparts were tested alongside (Figure 6). Transgenic pBΔ 3-39 plants with the full length repeats showed slightly increased levels (∼2–3 fold) of b1 repeat siRNAs relative to their non-transgenic siblings, suggesting that either the transgenic locus was producing b1 repeat siRNAs and/or it triggered an increase in the production of b1 repeat siRNAs from the endogenous alleles. Similar increases in b1 repeat siRNAs were seen with pFA::GUS and pFB::GUS transgenes (Figure 6 and Figure S6A). Notably, the b1 oligoprobe used in this experiment hybridizes to the FA part of the repeats, indicating that, at least in the pFB::GUS event, the b1 siRNAs detected are derived from the endogenous b1 repeat sequences. In spite of similar siRNA levels, the pBΔ 3-39 and pFA::GUS transgenic events were paramutagenic, while the pFB::GUS transgenic event was not (Figure 2 and data not shown), suggesting that the increased production of siRNAs was not sufficient to establish paramutation. A similar lack of correlation with paramutagenic ability and production of siRNAs was previously reported for endogenous b1 alleles [24].
The observation that the b1 tandem repeats are sufficient to recapitulate paramutation with a heterologous reporter gene in maize suggested that it might be possible to transfer the maize b1 paramutation system to Arabidopsis thaliana. For Arabidopsis, a large set of well-characterized mutations affecting epigenetic regulation exist that could be tested for their involvement in paramutation. The first step was to generate transgenic loci in Arabidopsis that would be dependent on the b1 tandem repeats for their expression. Constructs were generated with three to seven b1 tandem repeats fused to the minimal −90 35S promoter and the luciferase reporter gene (Figure 7A). As a control, sequences upstream of the repeats (Figure 7A) or b1 proximal promoter sequences (not shown) were used. Extensive analysis of the transgenic Arabidopsis plants containing intact transgenes revealed that all transgenic events carrying the b1 repeats exhibited a low level of luciferase activity similar to that displayed by control events with no b1 sequences (Figure 7A and Table S3). One possibility was that the transgenes integrated into a B'-like epigenetic state, which is associated with DNA methylation [19], [21]. Analyses of methylation using DNA blot analyses (Figure 7B and 7C, Figure S7A) revealed low levels of DNA methylation within the repeats and no detectable methylation in sequences upstream or downstream of the tandem repeats. All 7-repeat-containing transgenic events analyzed (pEN-MS1 and pEN-MS2) showed similar DNA methylation patterns compared to each other and also to that of the maize transgenes with seven repeats (Figure 4A). Such uniformity among transgenic events is unusual as methylation patterns between independent transgenic events are typically more variable [30]–[33]. The transgenic events carrying four and three b1 repeats (pEN-MS3 and pEN-MS4) also displayed low methylation levels within the repeats, but there was more variation between the different independent transgenic events (Figure 7B and 7C, and data not shown), similar to that seen for the endogenous maize three-repeat allele [19]. Together, these results demonstrate that, in the primary transgenic plants, the transgenic repeat sequence acquired similar sparse DNA methylation in maize and Arabidopsis.
During the Arabidopsis transformation process de novo DNA methylation occurs [34], [35]. We hypothesized that preventing any DNA methylation from occurring may enable the detection of the transcriptional regulatory function of the b1 repeats. To investigate this hypothesis, an Arabidopsis line in which the de novo DNA methyltransferases drm1 and drm2 (DOMAIN REARRANGED DNA METHYLASE 1 and 2; [34]) were mutated, was transformed with pb1::GFP constructs carrying b1 repeat- or b1 proximal promoter sequences fused to the minimal 35S promoter and GFP (Green Fluorescent Protein) coding region (Figure S7B). As a positive control, the 35S enhancer was fused to the GFP reporter gene (p35S::GFP). None of the pb1::GFP transgenic events showed GFP expression, while all of the p35S::GFP events did (Figure S7B and S7C, and Table S4). DNA blot analyses revealed that the drm1 drm2 double mutant background did prevent DNA methylation within the b1-repeats (Figure S7D), indicating that the lack of GFP expression was not due to DNA methylation.
Two other possible explanations for a lack of GFP expression, RNA-directed transcriptional or post-transcriptional silencing, were tested using the appropriate Arabidopsis mutants. Constructs with either seven or three b1 repeats (Figure S7B) were introduced into the rdr2-1 [35] and sgs2-1/rdr6 [36] mutants. RDR2 mediates RNA-directed transcriptional gene silencing, and RDR6 post-transcriptional gene silencing. None of the transgenic plants showed GFP expression (Figure S7B, Table S4), suggesting that neither RNA-directed transcriptional or post-transcriptional silencing is responsible for the lack of GFP expression. Taken together these data suggest that the maize b1 repeats do not have transcriptional regulatory activity in Arabidopsis. As one needs transcription to study transcriptional silencing this approach is not viable to study paramutation in Arabidopsis.
Results of the transgenic analysis presented in this paper demonstrate that specific tandem repeats are sufficient to both send and respond to the paramutation signal and that the repeat sequences need not be in an allelic position to communicate. The Mop1 gene, necessary for endogenous paramutation, is also required for transgene-induced paramutation, suggesting common mechanisms. The sequences required and sufficient for paramutation are localized in the first half of the b1 repeat unit. The tandem repeats are furthermore sufficient to enhance the expression of a heterologous reporter gene in maize, but not Arabidopsis. While transgenes are capable of inducing paramutation, several key differences exist between endogenous- and transgene-induced paramutation. Endogenous b1 paramutation is stable, fully penetrant and associated with dense DNA methylation within the b1 repeats, while transgene-induced paramutation displays variation in stability, penetrance and DNA methylation levels within the transgenic and endogenous b1 repeats.
Repeats have been implicated in multiple examples of paramutation [5], [19], [37], [38] and other silencing phenomena (e.g. [39]–[41]), but detailed mechanisms for why multiple copies are quantitatively required is not known in any system. Multiple models postulating which properties of the repeats are being counted have been discussed (reviewed in [42]). Models include a quantitative increase in a repeat product such as siRNAs [43], the quantitative binding of regulatory proteins to the repeats [44], the extent of DNA methylation within the repeats [21], or the creation and amplification of a unique junction fragment [21]. The transgenes were able to slightly elevate the production of siRNAs in immature ears but as we previously observed [24] there was no correlation between levels of siRNAs and the ability to participate in paramutation. These results do not exclude the possibility of a correlation between repeat siRNA levels and paramutation in other tissue types and/or developmental timepoints. Our results that tandem repeats of either the full repeat unit or the FA half are both strongly paramutagenic, yet they have distinct junctions, argues against a critical role for the junction regions. Furthermore, our observation that multiple repeats of FB have no paramutation activity strongly suggests tandem repeats of a specific sequence within FA are being counted during paramutation.
The FA and FB fragments differ in several properties that could be contributing to their ability to mediate paramutation. The FA half is much more GC rich relative to FB and as such, it contains most of the differentially DNA methylated region, including “the seed region” which becomes methylated very early in development in plants undergoing endogenous paramutation [21]. One possibility is that the AT richness of FB (68%) and the resulting lower capacity for cytosine methylation may prevent it from receiving and/or transmitting silencing signals. Intriguingly, the FA transgenes tended to be more strongly paramutagenic than those with the full repeat, suggesting that removal of the FB sequence increases the strength of paramutation. A full repeat is likely to have a lower overall density of DNA methylation than an FA repeat, which could be the signal being counted. It is also possible that FA, but not FB contains the regulatory sequences necessary to generate RNA silencing signals. The endogenous FB sequence is transcribed at a lower level and produces lower amounts of siRNAs relative to FA [24], [45]. Even though FB is neither required nor sufficient for paramutation in the transgenic assay, it may contribute to endogenous paramutation. Support for this hypothesis is that overexpression of a protein that binds to FB can induce a heritable and paramutagenic silenced state at the endogenous B-I allele [44]. Future experiments such as further dissecting the minimal sequences required for paramutation, mapping the key sequences mediating transcription of the b1 repeats and characterization of additional DNA binding proteins, should help to distinguish between hypotheses.
Two broad classes of models have been proposed for the allelic interaction that mediates endogenous paramutation, diffusible trans-acting signals or pairing between the repeats - these models are not mutually exclusive. Our observation that many different transgenic loci, located at distinct genomic sites, efficiently induce paramutation is most consistent with a diffusible trans-acting signal mediating the initial communication establishing paramutation. Consistent with this hypothesis, mutations in multiple genes involved in the RNA-directed transcriptional silencing pathway prevent the establishment of paramutation (reviewed in [42]), suggesting RNA may be the signal. However, our transgene experiments do not eliminate repeat pairing, as there are examples of pairing between homologous sequences in non-allelic positions in other systems [46]–[48]. Future experiments employing cytological methods may be able to shed light on whether there is a role for DNA pairing in paramutation.
Fine structure recombination mapping and chromosome conformation capture studies demonstrated that the b1 tandem repeats are also required for transcriptional activation of b1 [19], [22], but those studies could not distinguish between a direct role, i.e. the repeats carry transcriptional regulatory sequences, versus an indirect role, i.e. they mediate the ability of regulatory sequences located elsewhere to activate b1. Our maize transgenic results demonstrate that the b1 repeats do carry sequences that can mediate transcriptional activation of heterologous reporter genes, most consistent with a direct role of the repeats in transcriptional activation. Previous chromatin immunoprecipitation experiments demonstrated that upon transcriptional activation of B-I, the repeats are relatively depleted for nucleosomes and those that remain are enriched for H3ac histone marks [21]. These two properties, which strongly correlate with active transcriptional regulatory sequences [49], [50], are observed in both the FA and FB halves [21].
There is only one other paramutation system (p1, pericarp color) in which the sequence mediating paramutation has been defined [5], and that sequence also contains transcriptional regulatory activity [51], [52]. However, simply having a transcriptional regulatory element is not sufficient for paramutation as there are two transcriptional regulatory elements at p1 and only one of them can induce paramutation [5]. In contrast to the observations in maize, the b1 tandem repeats did not function as a transcriptional activator in Arabidopsis, suggesting that the transcription factors recognizing this sequence are not conserved between maize and Arabidopsis.
When B-I is paramutated by the repeat transgenes, the resulting transgene-induced B'# state, while heritable, often induced paramutation at a lower frequency and was less stable relative to the endogenous B' allele-induced B' state, in spite of the sequences being identical. The fact that after the transgenes are crossed to B', they induced a much more stable B'# state, indicates that their non-allelic positions or the structure of the transgenic loci cannot be responsible for the original reduced penetrance and heritability. Furthermore, the observation that a generation together with B' increased the transgenes' paramutagenicity, relative to carrying the transgenes over neutral alleles, suggests some type of heritable epigenetic mark is accumulating. Precedence for a role for DNA methylation has been reported in Arabidopsis where the RNA-directed transcriptional silencing machinery requires the presence of pre-existing DNA methylation on the endogenous FWA locus for effective silencing of an incoming FWA transgene [40]. This may not be the case with paramutation in maize, as two transgenes with very low DNA methylation levels could induce paramutation of the endogenous allele. Our results do indicate that specific sequences within the FA region of the repeat are a critical component and given that most of the DNA methylation marks are within this region, it remains possible that DNA methylation marks contribute to the strength of paramutation. Further studies of multiple transgenic events will be required to test this hypothesis.
The b1 stocks were initially acquired from a variety of sources and have been maintained in the Chandler laboratory for a number of years. The B', B-I and neutral b1 alleles were obtained from E.H. Coe, Jr. (University of Missouri, Columbia) and B-P was obtained from M.G. Neuffer (University of Missouri). All maize plant stocks used in this study carry functional alleles for all biosynthetic genes and the other regulatory genes required for anthocyanin biosynthesis, unless otherwise indicated. All genetic tests were conducted in the irrigated field conditions in Tucson, Arizona.
The seed stocks used were wild type Arabidopsis thaliana (ecotype WS) and the previously described mutants drm1 drm2 (ecotype Ws-2; [34]), rdr2 (ecotype Col-0, SAIL_1277H08; [35]) and rdr6 (sgs2, Col-0 [36]). All Arabidopsis plants were grown under standard greenhouse conditions.
The pB clone (Figure 1A) contains 106.6 kb of sequences upstream of the b1 transcription start site plus exon one, two and part of exon three (also named pBACB'1 in [20]; accession AY078063). The pBΔ clone was produced by digesting pB with the SwaI restriction enzyme, removing 91.6 kb of internal sequences and religation of the remaining sequences [20]. To produce the pFA and pFB transgenes, the two halves of the repeat were PCR amplified and inserted one by one in the BamHI/BglII digested P1.0b::GUS plasmid [51]. The p35S::GUS construct (Figure 5A) was the same as −90 35S::GUS described in [53] and contained the minimal −90 bp Cauliflower Mosaic Virus 35S promoter (35S), the maize adh1gene intron1, the omega leader, the beta-glucuronidase (GUS) coding region, and the potato PinII terminator. To produce the pb1TR::GUS construct, the seven 853 bp repeat array was inserted in the p35S::GUS construct upstream of the 35S promoter. To produce the pFA::GUS and pFB::GUS constructs, the FA and FB tandem repeats were ligated upstream of the 35S promoter of the p35S::GUS construct, respectively. Primer information and detailed information on cloning and vectors used for plasmid construction is presented in the Methods S1.
Transgenic maize plants were generated at the Iowa State University Plant Transformation Facility using biolistic particle bombardment of Hi-II immature embryos, which carry a neutral b1 allele (b-N) [54], [55]. The plasmid pBAR184 carrying the BAR gene, which confers resistance to the herbicide bialaphos, was co-bombarded with each construct [55]. Herbicide resistant calli were screened for DNA of interest using DNA blot analysis. Transformation events carrying transgene copies of the b1 repeat DNA were regenerated from calli.
The first set of plasmids used for Arabidopsis transformation carried the luciferase reporter gene (Figure 7A). These plasmids were made by inserting fragments of the maize b1 gene in front of a −90 35S promoter fused to the omega leader, luciferase coding region and nopaline synthase (nos) polyadenylation signal. The second set of the plasmids contained a GFP reporter gene (Figure S7). These plasmids were produced either by replacing the luciferase reporter gene by a GFP reporter gene from the pFLUAR100 plasmid [56] or by transferring the b1 sequences to an intermediate plasmid containing the 90 bp-35S promoter-GFP-nos gene cassette. A detailed description of the cloning steps and vectors used for plasmid construction is provided in the Methods S1.
Arabidopsis plants were transformed as described by [57] using 5% sucrose, 0.05% Silwet L-77, 0.5× Murashige & Skoog basal salts (micro and macro elements; Duchefa). The dipped plants were covered with Saran wrap, placed in the dark the first night and then grown in the greenhouse to maturity. To screen for transgenic plants, depending on the binary vector used, fluorescent seeds were either selected using the Leica MZ FLIII stereo fluorescence microscope with a dsRed filter or seedlings were sprayed with 0.5% BASTA (Glufosinate) twice, two and three weeks after sowing in soil, and surviving plants were transferred to individual pots. Transgenic plants were examined for reporter gene expression. Luciferase activity was evaluated using the Luciferase Assay System (Promega) and GFP activity was examined using the Leica MZ FLIII stereo fluorescence microscope with a GFP2 and GFP3 filter.
Transgenic maize calli were ground in liquid nitrogen and incubated with extraction buffer (200 mM Tris-HCl pH 7.5; 250 mM NaCl; 25 mM EDTA pH 8.0; 0.5% SDS) for 10 minutes, followed by phenol∶chloroform (1∶1) and chloroform extraction. DNA was precipitated with 1/10 of the volume of 3 M NaOAc and an equal volume of isopropanol. Pelleted DNA pellet was washed with 70% ethanol and resuspended in TE (10 mM Tris-HCl pH 8.0; 1 mM EDTA). DNA extraction from maize leaves and Arabidopsis flower heads was performed according to [58], [59], respectively. For DNA blot analysis 4–5 µg of maize and 0.5–2.5 µg of Arabidopsis genomic DNA was digested with the appropriate restriction enzyme(s) following the manufacturer's specifications, size-fractionated by electrophoresis in 0.5× TBE 0.8–1.5% agarose gels, transferred to positively charged nylon membranes, fixed by UV fixation and hybridized with 32P labeled DNA probes as described [26]. All blots that contained samples digested with DNA methylation sensitive enzymes were probed with a fragment (Probe A [19]) that recognizes sequences that are not methylated in maize to confirm all restriction enzymes cut the DNA to completion [19] followed by hybridization to the b1 repeat probe. Details describing probe fragments and restriction enzymes used for DNA blot analysis of maize and Arabidopsis transgenes are in Methods S1. Copy number of b1 repeat units in maize transgenic plants was estimated using the software packages Quantity One (Biorad) for pB and pBΔ, and ImageJ [60] for pFA and pFB. Copy number was calculated and normalized to the intensity of a single copy band of one the endogenous b1 allele present in each lane. Description of PCR-based genotyping of the endogenous b1 alleles and the mop1-1 mutation is presented in Methods S1.
Small RNA fractions were extracted from young, immature (∼5 cm) maize ears as described by [24]. RNA was separated on denaturing polyacrylamide gels, hybridized with 32P end-labeled DNA/LNA b1 repeat (VC1657F, [24]) and U6 (5′-CGTGTCATCCTTGCGCAGGGGCCATGCTAATCTTCTCTGTATCGT-3′) oligos. Results were analysed similarly to described previously [24].
Tissues from transgenic plants (Figure 5 and Figure S6) were collected between ∼50–60 days after germination and incubated with 1 ml of 0.1% X-GLUC solution (5-bromo-3-chloro-2-indolyl-b-D-glucuronic acid, Sigma) in the dark at 37°C for 24 hours [52]. Chlorophyll pigment was removed by repeated incubations in 70% ethanol. Stained tissues were analyzed under a binocular microscope and categorized according to the staining levels shown in Figure 5B and Figure S6B.
|
10.1371/journal.pgen.1002001 | RNF12 Activates Xist and Is Essential for X Chromosome Inactivation | In somatic cells of female placental mammals, one of the two X chromosomes is transcriptionally silenced to accomplish an equal dose of X-encoded gene products in males and females. Initiation of random X chromosome inactivation (XCI) is thought to be regulated by X-encoded activators and autosomally encoded suppressors controlling Xist. Spreading of Xist RNA leads to silencing of the X chromosome in cis. Here, we demonstrate that the dose dependent X-encoded XCI activator RNF12/RLIM acts in trans and activates Xist. We did not find evidence for RNF12-mediated regulation of XCI through Tsix or the Xist intron 1 region, which are both known to be involved in inhibition of Xist. In addition, we found that Xist intron 1, which contains a pluripotency factor binding site, is not required for suppression of Xist in undifferentiated ES cells. Analysis of female Rnf12−/− knockout ES cells showed that RNF12 is essential for initiation of XCI and is mainly involved in the regulation of Xist. We conclude that RNF12 is an indispensable factor in up-regulation of Xist transcription, thereby leading to initiation of random XCI.
| In all placental mammals, the males have only one X chromosome per diploid genome, as compared to the females who have two copies of this relatively large chromosome, carrying more than 1,000 genes. Hence, the evolution of the heterologous XY sex chromosome pair has resulted in an inevitable need for gene dosage compensation between males and females. This is achieved at the whole-chromosome level, by transcriptional silencing of one of the two X chromosomes in female somatic cells. Initiation of X chromosome inactivation (XCI) is regulated by X-encoded activators and autosomally encoded suppressors controlling Xist gene transcription. Spreading of Xist RNA in cis leads to silencing of one of the X chromosomes. Previously, we obtained evidence that the X-encoded E3 ubiquitin ligase RNF12 (RLIM) is a dose-dependent XCI activator. Here, we demonstrate that RNF12 exerts its action in trans and find that RNF12 regulates XCI through activation of transcription from the Xist promoter. Furthermore, analysis of female Rnf12−/− knockout ES cells shows that RNF12 is essential for initiation of XCI and that loss of RNF12 resulted in pronounced and exclusive down-regulation of Xist. It is concluded that RNF12 is an indispensable factor in Xist transcription and activation of XCI.
| X chromosome inactivation (XCI) in placental mammals is random with respect to the parental origin of the X chromosome that undergoes inactivation, during early embryonic development [1]. In contrast, in marsupials and mouse extra-embryonic tissues XCI is imprinted. Imprinted XCI always targets the paternally inherited X chromosome (Xp), and is initiated during the early cleavage divisions [2], [3], [4]. In the inner cell mass (ICM) of the mouse blastocyst, the inactive X chromosome is reactivated, after which random XCI is initiated around 5.5 days of embryonic development.
In mouse, two non-coding X-linked genes, Xist and Tsix, play a central role in the random XCI mechanism. Upon initiation of XCI, Xist is up-regulated on the future inactive X chromosome (Xi), and the transcribed RNA spreads along the X in cis, directly and indirectly recruiting chromatin modifying enzymes acting to establish the Xi [5], [6], [7]. Tsix is a negative regulator of Xist; the Tsix gene overlaps with Xist but is transcribed in the anti-sense direction [8], [9].
Random XCI is a stochastic process in which each X chromosome has an independent probability to become inactivated [10], [11]. Initiation of XCI is thought to be regulated by X-encoded activators and autosomally encoded inhibitors [11], [12]. With two active X chromosomes, female cells will have a concentration of XCI activators two-fold higher than male cells, sufficiently different to drive XCI in female cells only. Rapid down-regulation of XCI activator genes in cis, after initiation of XCI on either one of the X chromosomes, prevents initiation of XCI on the second X chromosome.
XCI inhibitors are involved in maintaining a threshold for XCI to occur. So far, several XCI inhibitors have been identified, acting through different mechanisms, in mouse. YY1 and CTCF act as positive regulators of Tsix, by binding the DXpas34 Tsix regulatory element [13]. The pluripotency factors OCT4, SOX2 and NANOG were proposed to regulate XCI by binding to intron 1 of Xist and suppressing Xist expression directly [14]. OCT4 and SOX2 have also been implicated in the positive regulation of Tsix and Xite, the latter being an enhancer of Tsix [15]. These findings indicate that several proteins and pathways act in concert to suppress Xist transcription and to block Xist RNA spreading in cis.
XCI activators could act by activation of Xist, but also by suppression of negative regulators of Xist such as Tsix and the Xist intron 1 region. Recently, we identified RNF12 (RLIM) as the first X-linked activator of XCI [16]. This E3 ubiquitin ligase is involved in regulation of LIM-homeodomain transcription factors and telomere length homeostasis, through degradation of LDB1 and TRF1, respectively [17], [18]. Previously, we found that additional transgenic copies of the Rnf12 gene encoding this protein resulted in induction of XCI on the single X in transgenic male cells, and on both X chromosomes in a high percentage of female cells. XCI was also affected in Rnf12+/− ES cells supporting a dose-dependent role for RNF12 in activation of XCI. In the present study, we aimed to dissect the role of RNF12 in XCI, and we obtained evidence that RNF12 regulates XCI in trans, by activation of the Xist promoter. In addition, the generation and analysis of Rnf12−/− ES cells indicated that RNF12 is required for the XCI process and appears to be involved in XCI mainly by activation of Xist. The results reinforce that RNF12 is a key player in regulation of the XCI process.
XCI is regulated by several cis elements, and Rnf12 is located in close proximity to Xist (∼500 kb). Therefore, we aimed to test whether all the activity of RNF12 is mediated in trans. Our previous studies showed that Rnf12+/− female ES cells induce XCI in a reduced number of ES cells. Here, we rescued 129/Sv/Cast/Ei (129/Cas) polymorphic Rnf12+/− female ES cells by introducing a 129 BAC (RP24-240J16) construct covering Rnf12. RT-PCR analysis followed by RFLP detection confirmed expression of the transgenic copies of Rnf12 (Figure 1A). Xist RNA-FISH analysis, to detect the Xist coated inactive X chromosome (Xi) in day 3 differentiated transgenic ES cell lines with one additional copy of Rnf12, shows that XCI was restored to wild type level (Figure 1B). In line 20, with 5 transgenic copies of Rnf12 the percentage of cells with one or two Xi's is even more pronounced, supporting a dose dependent role of RNF12 in XCI (Figure 1B, 1C). XCI is skewed in wild type 129/Cas female ES cells towards inactivation of the 129 X. This is due to the presence of different X-linked cis elements (Xce) that affect random choice [19]. RT-PCR detecting a length polymorphism was used to distinguish Xist emanating from either the 129 or the Cas alleles. We observed that skewed XCI is more pronounced in the Rnf12+/− cells, as compared to XCI in wild type cells at day 3 of differentiation (Figure 1D). This could be caused by selection against cells inactivating the wild type X chromosome, which would result in complete loss of RNF12 from these cells. However, RNF12 possibly is not essential for cell survival, also of differentiated cells, so that selection against cells inactivating the wild type X chromosome might point to a role for RNF12 in maintaining Xist expression. In the rescued cell lines, Xist was up-regulated from both alleles at day 3 of differentiation (Figure 1D). This result demonstrates that RNF12 activates XCI in trans.
One possible mechanism for regulation of XCI by RNF12, might be a direct interaction with Xist RNA to target chromatin components. However, examination of day 3 differentiated female cells by immunocytochemistry detecting RNF12, together with the Polycomb protein SUZ12 which accumulates on the Xi [20], [21], excludes this possibility (Figure 2A). Interestingly, we noticed that the RNF12 staining intensity was much more dynamic in female compared to male cells (Figure 2B, Figure S1). Also, in female cells, a SUZ12 coated Xi appeared mainly in cells with low RNF12 staining (Figure 2A, Figure S2, and data not shown). Immunostaining of differentiating female ES cells indicated a negative correlation between expression of RNF12 and NANOG, although expression was not completely mutually exclusive (Figure 2C). To analyze this in more detail, we targeted an Rnf12 promoter-mCherry construct into ES cells, also harboring a knock-in GFP transgene in the Nanog and Oct4 loci. We analyzed expanded individual clones and pooled clones and obtained similar results. FACS analysis, prior to differentiation and at different time points after differentiation of these double transgenic ES cell lines, showed a negative correlation between RNF12-mCherry and NANOG-GFP expression, but not for RNF12-mCherry and OCT4-GFP (Figure 2D, 2E, Figure S3). Our findings therefore suggest specific counteracting regulatory roles for RNF12 and NANOG in XCI, which might include an inhibitory effect of NANOG on Rnf12 transcription. Interestingly, NANOG has been implicated in the regulation XCI by direct suppression of Xist in ES cells, and Xist suppression in the ICM of the developing blastocyst corresponds with up-regulation of NANOG expression [22]. Therefore, mutual exclusive expression of RNF12 and NANOG may be required for initiation of XCI.
Recently, the first intron of Xist has been identified as a region involved in recruitment of three pluripotency factors, OCT4, NANOG and SOX2 [14]. It was shown that down-regulation of Nanog and Oct4, through gene ablation, resulted in an increase in Xist expression, and initiation of XCI in male cells. Interestingly, the intron 1-mediated suppression of XCI was suggested to directly act on Xist, without involvement of Tsix. To study if RNF12 might regulate XCI by interfering with binding of pluripotency factors to the intron 1 region of mouse Xist, we removed 1.2 kb of Xist intron 1 including all reported NANOG, OCT4 and SOX2 binding sites by homologous recombination with a BAC targeting construct, without disturbing the integrity of the Xist transcript. Targeted clones were screened by PCR amplification of a targeted RFLP (BsrgI) in female F1 2-1, 129/Cas polymorphic ES cells, which was confirmed by Southern blotting, followed by Cre mediated loop-out of the kanamycin/neomycin resistance cassette (Figure 3A, Figure S4). Xist RNA FISH at different time points of differentiation of several Xistintron1+/− ES cell lines indicated that XCI is initiated with the same kinetics as in wild type cells, and showed that the intron 1 region is not required for repression of Xist in undifferentiated ES cells or early during initiation of XCI (Figure 3B, 3C, and Figure S4G). Nevertheless, Xist specific RT-PCR, detecting a length polymorphism distinguishing 129 and Cas Xist, showed enhanced skewing at day 3 of differentiation towards 129 Xist expression, suggesting a role for the intron 1 region in suppressing Xist at later stages of differentiation, when NANOG, OCT4 and SOX2 are expressed at a lower level (Figure 3D). To test an involvement of the intron 1 region in RNF12-mediated activation of XCI, we introduced an Rnf12 BAC transgene into the Xistintron1+/− ES cell lines. Additional copies of Rnf12 resulted in induction of Xist, even in undifferentiated ES cells (Figure 3E, 3F, 3I), confirming our previous findings [16]. However, allele specific RT-PCR did not point to an increased preference for expression of the mutated or wild type allele, in undifferentiated ES cells (Figure 3G, 3H), indicating that RNF12-mediated action on XCI does not require the Xist intron 1 region (Figure 3J). At day 3 of differentiation, in several cell lines, we found higher expression of Cas Xist in Rnf12 transgenic Xistintron1+/− cells compared to Xistintron1+/− only cells. We attribute this finding to an increase in the percentage of cells with two Xist clouds. We conclude that the Xist intron 1 region is not essential for suppression of XCI in undifferentiated ES cells, but may play a role later during differentiation. Furthermore, RNF12-mediated activation of XCI is independent from the Xist intron 1 region.
RNF12 could regulate XCI through activation of Xist or suppression of Tsix, or both. Previously, we analyzed Xist transgenic male ES cell lines with a BAC RP24-180B23 integration covering Xist only [16], or a BAC RP23-338B22 sequence containing both Xist and Tsix (Figure 4A). These male transgenic ES cell lines also contained 16 copies of an ms2 bacteriophage repeat sequence located in exon 7 of the endogenous Xist gene, allowing separate detection by RNA-FISH of autosomal versus endogenous Xist spreading [23]. Differentiation of transgenic male ES lines containing the Xist-Tsix transgene resulted in expression of Xist from the autosomal integration site in cell lines containing multicopy integrations. Autosomal spreading of Xist in these cell lines is most likely due to accumulation of enough Xist RNA to silence at least one copy of Tsix, allowing spreading of Xist in cis. Integration of truncated transgenes that lack Tsix would facilitate this process [16]. This also explained autosomal Xist spreading in BAC RP-24-180B23 single copy male transgenic ES cell lines upon differentiation, because Tsix is not covered by this BAC [16]. We used two of these, Xist only, BAC RP-24-180B23 ES cell lines to introduce 129 BAC RP24-240J16 transgenes covering Rnf12, and found Xist spreading on the single endogenous X (Figure 4B and 4C), confirming previous results. We also found a significant increase in the number of cells with autosomal Xist spreading, indicating that RNF12 activates XCI through Xist. Next, we introduced an Rnf12 transgene (BAC RP24-240J16) in a single copy Tsix male transgenic ES cell line that lacks transgenic Xist (BAC RP23-447O10). These double transgenic ES cell lines contain a Cas X chromosome which allowed RFLP mediated discrimination of endogenous (Cas) and transgenic (129) Tsix. Analysis of these cell lines indicated that transgenic over-expression of RNF12 does not lead to down-regulation of Tsix, as measured by qPCR and by RNA-FISH examining the relative number of Tsix pinpoint signals (Figure 4D, 4E, 4G). Interestingly, allele specific RT-PCR indicated that endogenous Tsix (Cas) is even down-regulated in samples with higher Xist expression, indicating Xist-mediated silencing of Tsix in cis (Figure 4F). Taken together, these results indicate that Xist and not Tsix is the functionally most important downstream target of RNF12.
We previously found that the rate of initiation of XCI is reduced in differentiating female Rnf12+/− ES cells, compared to wild type ES cells [16]. The RNF12 protein level in these Rnf12+/− female cells is equal to that in male cells [16], but XCI is still occurring at a higher rate than in male cells. This indicated the presence of additional X-encoded XCI activators, but did not exclude the possibility that RNF12 is essential for XCI. To address this point, we generated Rnf12−/− female ES cells by targeting the wild type Cas Rnf12 allele in Rnf12+/− ES cells (Figure 5A). Correct targeting was confirmed by RT-PCR, showing loss of a targeted RFLP located in exon 5 of Rnf12 (Figure 5B). The presence of two X chromosomes in these Rnf12−/− female ES cells was ascertained by X chromosome DNA FISH analysis and amplification of an RFLP in the Xist gene (Figure 5C, and data not shown). Western blotting analysis confirmed the absence of RNF12 protein in the knockout cells (Figure 5D). RT-PCR and qRT-PCR of pluripotency associated genes and differentiation markers gave information that differentiation of the Rnf12−/− ES cells was not different from that of wild type ES cells (Figure 5E, 5F and Figure S5). However, Xist RNA FISH analysis showed that differentiating Rnf12−/− ES cells only sporadically initiate XCI (Figure 5G, 5H and 5I). QPCR analysis confirmed that Xist is not detectably up-regulated when measured for a population of Rnf12−/− cells upon differentiation. Moreover, DNA-FISH detecting a whole chromosome X paint probe at day 7 and 10 of differentiation excluded X chromosome loss (Figure S5). The few Rnf12−/− cells that initiated XCI appeared in clusters, suggesting clonal expansion of a few cells that initiated XCI (Figure S5). We therefore conclude that RNF12 is an essential factor in XCI.
Evidently, the Rnf12−/− knockout cells present the possibility to study control of gene expression by RNF12. Therefore, we next performed micro-array expression analysis comparing day 3 differentiated Rnf12−/− and wild type cells. We found that Xist was the only gene that was subject to differential regulation, showing pronounced down-regulation (Figure 5J). Interestingly, none of the known downstream targets of RNF12 appeared affected in our analysis. This may be due to our ES cell differentiation system resulting in a mixed population of cells at different stages of differentiation. In addition, the 3-day-time span allowed in our studies for cell differentiation may have prevented detection of effects on downstream targets which are expressed at later stages of differentiation. Nevertheless, our results indicate that the main function of RNF12 at this early stage of differentiation concerns the regulation of XCI. The observed dependency of Xist transcription on RNF12 might be effectuated by RNF12 acting through the Xist promoter. To test this, we expressed Xist promoter luciferase reporter constructs, both transiently and stably, in wild type female and Rnf12−/− ES cell lines and differentiated these cells for 3 days. The results revealed an unequivocal correlation between RNF12 expression and luciferase expression (Figure 5K). Our results therefore demonstrate that RNF12 activates the Xist promoter, although this does not exclude a role for other cis regulatory sequences, further away from the Xist promoter, in RNF12-mediated activation of XCI.
Here, we present evidence that RNF12 is an essential activator of random XCI. RNF12 acts in trans on the Xist promoter, in differentiating mouse ES cells, to activate Xist transcription, leading to Xist RNA cloud formation and spreading of the silencing complex over the future inactive X chromosome in cis. Although our results show that RNF12 acts in trans, it is to be expected that the close proximity of the Rnf12 gene to the Xist locus, taken together with the dose-dependent action of RNF12, is quite crucial for well-tuned regulation of XCI. Such proximity most likely facilitates rapid down-regulation of Rnf12 in cis upon initiation of XCI, leading to a lower nuclear RNF12 content, thereby preventing inactivation of the second X chromosome.
Whole genome expression analysis suggests that the major function of RNF12 in ES cells is its regulation of Xist RNA expression, hence XCI. This is a very surprising finding, as RNF12 has been implicated in many other biological pathways. Apparently, in the present cell differentiation system, loss of expression of RNF12 does not cause a deviation from the wild type differentiation process to such an extent that it affects gene expression other than that of Xist. However, also based on our studies we do not exclude a function for RNF12 at later stages of cell differentiation, or in mouse development. In addition, redundant pathways or proteins such as RNF6, a close homologue of RNF12, may prevent full phenotypic expression of loss of RNF12. However, RNF12 exerts a predominant role in targeting Xist, as evidenced by our observation that Xist is largely silenced in the RNF12 deficient cells.
While our manuscript was under review, Shin et al. (2010) published a paper suggesting that RNF12 might be required in particular for imprinted XCI in mice [24]. Remarkably, that study included the observation that RNF12 depletion did not prevent initiation of random XCI in a significant percentage of Rnf12−/− ES cells derived from mouse blastocysts. This discrepancy with our findings might be explained by experimental differences, such as differences concerning the design of the knockout, the genetic background of the ES cells, or the cell derivation and culture procedures. Differences in cell differentiation protocols have been shown to have a pronounced impact on the XCI process [25]. Also, ES cells derived from embryos with a different genetic background could express XCI activators and XCI inhibitors at different levels, allowing XCI in either a lower or a higher percentage of Rnf12−/− cells. Future studies comparing the two independently generated Rnf12−/− ES cell lines will yield useful information about these points.
Although our observations provide evidence that RNF12 is an essential factor for the XCI process to occur in differentiating ES cells, we anticipate that other XCI activators act in parallel, and might independently regulate Xist or Tsix, or both. Dosage compensation mechanisms in species such as D. melanogaster and C. elegans also involve multiple factors and pathways, possibly leading to increased fidelity of these mechanisms [26]. In such a mechanism involving multiple factors, RNF12 would be the dose-dependent factor that is required to exceed the cumulative threshold limit to proceed towards initiation of XCI. It is feasible that female Rnf12−/− cells sometimes do initiate XCI (Figure 6A), as a consequence of the stochasticity of the process. This would be compatible with a mechanism, in which the combined total activity of all putative XCI activators exclusive of RNF12 is just below or around the threshold to initiate XCI. Interestingly, Xist cloud formation is also sporadically found in male cells, but in contrast to female Rnf12−/− cells, this represents a lethal condition and will be selected against.
Our studies indicate that RNF12 participates in Xist promoter activation, through an action which requires the presence of the minimal promoter. Although the direct protein target(s) of RNF12 remain elusive, its reported E3 ubiquitin ligase activity [17] would be compatible with RNF12 targeting an inhibitor of Xist transcription through proteasome-mediated degradation. This does not exclude that RNF12 might be involved, in addition or alternatively, in activation of a transcription factor driving Xist expression through positive regulation of transcription. Furthermore, RNF12 could be involved in regulation of cis-regulatory sequences other than the Xist promoter, yet to be identified and further away from the Xist locus.
Selection against cells inactivating the X chromosome containing the wild type allele of Rnf12 in the heterozygous Rnf12+/− ES cells could point to a continued requirement for Rnf12 in maintaining Xist expression, following the early stages of differentiation. From the fact that male Rnf12−/Y knockout male mice are viable [24], it can be concluded that RNF12 deficiency is compatible with survival of differentiated cells in which XCI does not play any role. Hence, it would be difficult to explain the observed selection against cells inactivating the wild type X chromosome in the heterozygous Rnf12+/− ES cells by loss of any possible function of RNF12 independent of XCI. If RNF12 would be required for maintaining Xist expression and XCI, the cells inactivating the wild type allele and becoming deficient in RNF12 can be expected to lose Xist expression and to reactivate the Xi. In contrast, cells inactivating the X chromosome containing the mutated allele, keeping one functional allele of Rnf12, will be able to maintain Xist expression and XCI. In a population of cells this will lead in skewed XCI of the mutated allele. In fact, such a mechanism might also be relevant to explain the reported defect in imprinted XCI resulting from an Rnf12 mutation [24].
Imprinted XCI involves activation of Xist on the Xp, and the observed phenotype concerns lack of this imprinted XCI of the Xp when the mutant Rnf12 allele is inherited from the mother. It was observed that no female embryos were born, inheriting a mutated Rnf12 allele from either a Rnf12−/− or a Rnf12+/− mother in crosses with wild type males. In contrast, the mutated allele was transmitted to male offspring. Maternal storage of RNF12 in the oocyte was proposed to play a crucial role in imprinted silencing of the Xp in the early embryo [24]. Rnf12 is at a 46 cM distance of the centromere, so that it can be expected that many haploid oocytes generated by the first meiotic division (the reduction division) of Rnf12+/− oocytes, which occurs at the time of ovulation, will contain both wild type and Rnf12 mutated alleles, as a consequence of meiotic recombination. Hence, we anticipate that there will be ongoing expression of Rnf12 in a high percentage of oocytes transmitting the mutated Rnf12 allele, until fertilization triggers meiotic division II. The recombined wild type and mutant alleles which are present within one haploid oocyte, will be exposed to the same maternal storage of RNF12. Taken together with the observation that Rnf12+/− oocytes did not give rise to female offspring carrying the mutant allele, whereas female offspring carrying the wild type allele were obtained at the expected mendelian ratio from these oocytes [24], this argues against a predominant role for maternal storage in imprinted XCI. Rather, we favor the hypothesis that continued transcription of Rnf12 throughout ovulation and after fertilization is required for sustained expression of RNF12, activation of Xist from the Xp, and maintenance of the inactive Xp. Future research will be required to address this hypothesis.
Our results indicate a negative correlation between NANOG and RNF12 expression. NANOG and the other pluripotency factors OCT4 and SOX2 have been shown to be recruited to the Xist intron 1 region in undifferentiated ES cells, and were proposed to play a role in Tsix independent suppression of Xist [14]. In this regulatory mechanism, ablation of Tsix did not result in up-regulation of Xist in undifferentiated ES cells, and Tsix was not required for repression of Xist located on the inactivated paternal X chromosome in the inner cell mass. This pointed to an important role for recruitment of NANOG, OCT4 and SOX2 to Xist intron 1 in suppression of Xist in ES and ICM cells [14]. However, the present findings show that the intron 1 region is dispensable, in silencing the XCI process in undifferentiated ES cells. Deletion of Xist intron 1 caused an effect, but only in the form of skewing of XCI, which was notable at later stages of differentiation. Interestingly, a previous study analyzing an Xist mutant allele that lacks the intron 1 region but leaves the Xist promoter intact, also did not show up-regulation of the mutated allele in undifferentiated ES cells [27]. Although these latter results support our findings, they should be interpreted with caution because the selection cassette was still present in the cells analyzed by Marahrens et al. [27].
Like for the role of RNF12, this points to the presence of additional mechanisms, involved in suppression of XCI. Tsix and Xite are the most likely candidate genes taking part, and the combined action of these repressive mechanisms may be sufficient to suppress Xist. However, even with all the repressive elements in place RNF12 can induce Xist expression and XCI in undifferentiated ES cells [16]. This points towards another mechanism involved in Xist suppression, in which the nuclear concentration of the XCI activator may be too low in undifferentiated ES cells and ICM cells to allow Xist expression and initiation of XCI, even in the absence of repressive elements such as the intron 1 region. Future research should clarify whether these mechanisms indeed act synergistically in silencing the XCI process.
The negative correlation of RNF12 and NANOG expression that we report could reflect the differentiation state of the ES cells, and does not necessarily entail a cross-regulatory role for these proteins. Nevertheless, NANOG and other pluripotency factors are also recruited to the Rnf12 promoter in ES cells, where it might be involved in down-regulation of Rnf12 (Figure 6B) [28], which opens the intriguing possibility that NANOG might also be implicated in regulation of the initiation of XCI through suppression of Rnf12. This highlights the complexity of the overall mechanism and the interconnection of the different players involved in XCI, but also reinforces the predominant role of RNF12 in this process.
ES cells were grown in standard ES medium containing DMEM, 15% foetal calf serum, 100 U ml−1 penicillin, 100 mg ml−1 streptomycin, non-essential amino acids, 0.1 mM β-mercaptoethanol, and 1000 U ml−1 LIF. To induce differentiation, ES cells were split, and pre-plated on non-gelatinised cell culture dishes for 60 minutes. ES cells were then seeded in non-gelatinised bacterial culture dishes containing differentiation medium to induce embryoid body (EB) formation. EB-medium consisted of IMDM-glutamax, 15% foetal calf serum, 100 U ml−1 penicillin, 100 mg ml−1 streptomycin, non-essential amino acids, 37.8 µl l−1 monothioglycerol and 50 µg/ml ascorbic acid. EBs were plated on coverslips 1 day prior to harvesting, and allowed to grow out.
For the Rnf12 rescue experiments, an Ampicilin-Puromycin resistance cassette was inserted in the backbone of BAC RP24-240J16 by homologous recombination in bacteria. The modified BAC was electroporated in to female heterozygous Rnf12+/− cells [16], and colonies were picked after 8–10 days of Puromycin selection, expanded and differentiated. BAC copynumber was determined by qPCR, and transgene specific expression was determined by allele specific RT-PCR, as described previously [16].
To generate the female homozygous Rnf12 −/− ES cell line, the previously generated Rnf12+/− ES cell line was targeted with an Rnf12 BAC targeting construct containing an Ampicilin-Puromycin cassette disrupting the open reading frame of Rnf12. To generate this targeting construct, targeting arms were PCR amplified using primers GCCTTCGAACATCTCTGAGC, GAGCCGGACTAATCCAAACA, cloned into pCR-BluntII-TOPO (Invitrogen), and linearized with NheI to introduce an Ampicilin-Puromycin cassette from pBluescript. The targeting cassette was inserted in a Cast/Ei Rnf12 BAC RP26-81P4 by homologous recombination in bacteria, and the resulting construct was used to target specifically the Cast/Ei X chromosome of the Rnf12 +/− ES cell line. Colonies were selected under Neomycin and Puromycin selection, and the absence of Rnf12 expression was confirmed by Western analysis.
To generate the Xist intron 1 deletion, a BAC targeting construct was generated by homologous recombination, replacing intron 1 by a floxed Neomycin cassette. Targetting arms were PCR amplified using primers 5′Forw:CATCAGGCTTGGCAGCAAGT, 5′R: CCTTGTTGGTCCAGACGACTATT and 3′Forw: CCAGACCAGGTCTTTGTATGCA, 3′Rev: GTGCTCCTGCCTCAAGAAGAA. Correctly targeted clones were identified by allele specific RFLP analysis using primers CAGTGGTAGCTCGAGCCTTT and CCAGAAGAGGGAGTCAGACG, followed by BsrGI digestion. The Neomycin cassette was removed by transient transfection with a CrePAC vector and selection with puromycin. The final cell lines were verified by Southern blotting.
To generate the Rnf12 promoter cherry reporter cell lines, the Rnf12 promoter was PCR amplified using previous described primers [29], and cloned into pCR-BluntII-TOPO and sequence verified. The Rnf12 promoter was then released from pCR-BluntII-TOPO by digestion with SacI and KpnI, and blunt cloned into an AseI-BamHI fragment from pmCherry-N1 (Clonetech), thereby replacing the pCMV promoter of pmCherry-N1 with the Rnf12 promoter. The resulting construct was used to electroporate in Oct-GFP and Nanog-GFP ES cell lines. Both pooled cell lines and single colonies were expanded, and cherry expression was analysed by FACS analysis using a BD FACSAria apparatus.
The Xist promoter was amplified using primers: TCCCAAGGTATGGAGTCACC, and GGAGAGAAACCACGGAAGAA, and cloned into pGL3-basic vector. As a control, the promoter less pGL3-basic vector was transfected.
Stable pooled cell lines of wild type or Rnf12 −/− ES cells were generated by co-transfection with a puromycin or hygromycin selection vector. Expression of Luciferase was determined using the Bright-Glo luciferase assay system (Promega) and measured using a Promega luminometer. Results were normalized to the amount of protein present in the cell lysate measured by nanodrop, and copynumber of Xist promoter integration determined by qPCR. qRT-PCR using primers detecting luciferase (TCTAAGGAAGTCGGGGAAGC and CCCTCGGGTGTAATCAGAAT) confirmed the results obtained. For transient luciferase experiments, cells were co-transfected using the Xist reporter constructs and a control Renilla construct, using Lipofectamine 2000. Luciferase activity was measured using the Dual Glo luciferase system (Promega).
Xist RNA-FISH was performed as described [11], [16]. Immunofluorescence was performed using standard procedures. RNF12 and NANOG were detected using a mouse anti- RNF12 antibody (1∶250, Abnova), and a rabbit anti-NANOG antibody (1∶100, SC1000, Calbiochem). ImageJ software was used to measure staining intensities; at least 100 cells were measured for each indicated time point, and background correction was performed. Western blotting was performed as previously described [16].
RNA was isolated using Trizol reagent (Invitrogen) using manufacturers instructions. DNAse treatment was performed, and cDNA was prepared using SuperScriptII (Invitrogen), using random hexamers. qRT-PCR was performed using a Biorad thermocycler, using primers described in Table S1. Results were normalized to Actin, using the ΔCT method.
Whole genome wide expression analysis of female wild type and Rnf12−/− ES cells differentiated for 3 days was performed with Affymetrix Mouse Genome 430 2.0 Arrays. Differentially expressed genes were identified using Limma (Bioconductor package) in R software.
|
10.1371/journal.pbio.1000218 | Bayesian Modeling of the Yeast SH3 Domain Interactome Predicts Spatiotemporal Dynamics of Endocytosis Proteins | SH3 domains are peptide recognition modules that mediate the assembly of diverse biological complexes. We scanned billions of phage-displayed peptides to map the binding specificities of the SH3 domain family in the budding yeast, Saccharomyces cerevisiae. Although most of the SH3 domains fall into the canonical classes I and II, each domain utilizes distinct features of its cognate ligands to achieve binding selectivity. Furthermore, we uncovered several SH3 domains with specificity profiles that clearly deviate from the two canonical classes. In conjunction with phage display, we used yeast two-hybrid and peptide array screening to independently identify SH3 domain binding partners. The results from the three complementary techniques were integrated using a Bayesian algorithm to generate a high-confidence yeast SH3 domain interaction map. The interaction map was enriched for proteins involved in endocytosis, revealing a set of SH3-mediated interactions that underlie formation of protein complexes essential to this biological pathway. We used the SH3 domain interaction network to predict the dynamic localization of several previously uncharacterized endocytic proteins, and our analysis suggests a novel role for the SH3 domains of Lsb3p and Lsb4p as hubs that recruit and assemble several endocytic complexes.
| Significant diversity exists in protein structure and function, yet certain structural domains are used repeatedly across species to execute similar functions. The SH3 domain is one such common structural domain. It is found in signaling proteins and mediates protein–protein interactions by binding to short peptide sequences generally composed of proline. To investigate both the generality and selectivity of peptide binding by SH3 domains, we examined peptide specificity for almost all SH3 domains encoded within the proteome of the budding yeast, Saccharomyces cerevisiae, using a range of experimental methods. We found that although most of the intrinsic binding specificity for SH3 domains can be summarized by the two previously described canonical binding modes, each individual SH3 domain that we studied utilizes unique features of its cognate ligand to achieve binding selectivity. Moreover, some domains exhibit binding specificities that are distinct from the two canonical classes. We integrated peptide-SH3 domain binding data from three complementary screening techniques using a Bayesian statistical model to generate a protein–protein interaction network for the budding yeast SH3 domain family. This network was highly enriched in endocytosis proteins and their interactions. By examining these interactions in detail, we show that our SH3 domain network can be used to predict the temporal localization of several previously uncharacterized proteins to dynamic complexes that orchestrate the process of endocytosis.
| Families of peptide recognition modules (PRMs), such as PDZ (PSD-95/Discs-large/ZO-1), SH2 (Src homology 2), and SH3 (Src homology 3) domains bind peptide motifs within proteins to mediate protein–protein interactions required for the assembly of stable or transient biological complexes [1]. Thus, PRMs function to dynamically orchestrate biological pathways [1]. PRM family members can be identified directly from whole-genome sequences; therefore, it is possible to explore the recognition specificity of entire families using a variety of different experimental approaches [2],[3]. Here, we explore the potential for mapping SH3 domain protein interaction networks by a Bayesian integration of results from three complementary experimental screening approaches: phage display, peptide array, and yeast two-hybrid analysis.
In general, PRMs engage in protein–protein interactions by recognizing a core motif common to a domain family as well as additional ligand features that are more specific for each family member as is the case for PDZ domains [3]. Initial studies determined that SH3 domains bind to proline-rich sequences containing a core PXXP motif (where X is any amino acid) flanked by a positively charged residue [4],[5]. Class I domains bind to ligands conforming to the consensus +XXPXXP (where + is either arginine or lysine), and do so in an orientation opposite to that of class II domains, which recognize PXXPX+ motifs [6],[7]. More recently, a number of alternative SH3 domain binding motifs have been identified, highlighting a wider breadth of SH3 specificities [8]–[11].
A general genome-wide analysis of PRMs would involve defining all the domains from their primary sequence, profiling their ligand-binding specificities in detail, predicting natural ligands for each domain, and mapping large-scale protein–protein interaction networks for each domain family. Here, we present the first high-resolution analysis of the yeast SH3 domain family. First, we used large-scale phage-displayed peptide libraries and extensive sequencing to generate high-resolution binding profiles, which we show accurately represent binding specificity across multiple SH3 domain ligand positions. Second, we used the resulting specificity profiles to identify putative interactions in the yeast proteome, which were subsequently confirmed using oriented synthetic peptide arrays. Third, we conducted large-scale yeast two-hybrid screens to generate a physical interaction network mediated by the set of yeast SH3 domains. Finally, the datasets were integrated using Bayesian networks to generate a global SH3 domain interaction map in yeast.
Applying the integrated probabilistic network revealed an intricate array of SH3-mediated interactions amongst proteins that make up the endocytic machinery. Investigation and comparison of the dynamics of protein localization within this network showed that the modular network predictions of the spatiotemporal dynamics of several novel endocytotic components were correct. In particular, our analysis predicts that the SH3 domains from Lsb3p and Lsb4p interact with multiple endocytic proteins and therefore might act as hubs to cluster these proteins at sites of endocytosis.
We used peptide phage display to conduct a large-scale analysis of yeast SH3 domain specificity. We cloned DNA fragments encoding all 27 unique yeast SH3 domains using boundaries taken as the union of the domain lengths identified by three domain detection tools, BLAST [12], PFAM [13], and SMART [14], with an additional ten amino acids included on either side of the overlapping domain region to facilitate cloning (Table S1). We expressed the domains in bacteria as proteins fused to the C-terminus of glutathione S-transferase (GST) and purified 24 out of 27 fusion proteins in a stable, soluble form. For two of the three recalcitrant domains, the C-terminal domain of Bem1p (Bem1-2) and the Bud14p domain, we extended the sequence boundaries by examining the conservation of the domain regions across diverse fungal species. Based on this analysis, the domain boundaries for these two SH3 domains were extended, enabling the isolation of stable GST fusion proteins (Table S2 and Text S1). The third recalcitrant domain, the N-terminal domain of Sla1p (Sla1-1), could only be purified in tandem with Sla1-2, and we denoted the dual domain protein as Sla1-1/2.
The GST-SH3 domain fusion proteins were used as targets in binding selections with a combination of random and biased peptide–phage libraries. We were successful in obtaining ligands for all SH3 domains except Bud14 and Cdc25, and we isolated a total of 1,871 unique peptides. These results extend results from our previous study [2] and represent nearly an 8-fold increase in binding data. The set of aligned ligands for each domain was used to compile a position weight matrix (PWM), which captures the frequency of amino acid preferences at each ligand position. Some ligand sets contained two distinct groups of ligands, and for these, two separate PWMs were compiled (see below). From each PWM, a sequence logo [15] was generated to graphically represent the specificity at each amino acid position in each ligand set.
To compare the binding specificities for the yeast SH3 domain family on a global scale, we clustered all domains in an unrooted tree based on their specificities (Figure 1 and Figure S1). We generated a set of 10,000 random peptides from the yeast proteome and used these to score each phage-derived PWM. The match of each PWM with each peptide was calculated using an information score yielding a 10,000-dimensional profile vector for each PWM. This profile vector describes the binding specificity in a cellular context by sampling the sequences that the domains would be exposed to in the cell. The similarity between PWMs was computed as the Pearson correlation between these vectors. PWMs were then clustered according to this similarity measure using a complete linkage algorithm. Hence, the tree represents natural yeast SH3 domain specificity as it clusters binding profiles based on endogenous protein ligands. Overall, our results are consistent with previous findings [2]; in addition, this higher resolution analysis reveals that each domain exhibits specificity across multiple ligand positions, including the core motif and flanking positions.
Our specificity map reveals that the majority of yeast SH3 domains have specificities that can be defined as class I or II, with eight and 12 representatives, respectively. Notably, the SH3 domains from Cyk3p and Rvs167p, and a protein fragment containing the two N-terminal domains from Sla1p (Sla1-1/2) exhibit dual specificity for both ligand classes (Figure 1). Furthermore, the specificity map uncovered many specificity profiles that do not cluster with either of the canonical classes. For instance, the SH3 domains of Bem1p, Hof1p, Myo3p, and Myo5p comprise a distinct cluster, which we denote as class III, and are characterized by their preference for poly-proline ligands, without the requirement for flanking charged residues. The SH3 domains of the paralogs Pin3p and Lsb1p exhibit dual specificity, recognizing class II ligands and a ligand set (+XXXPXP) that resembles class I ligands, but with different residue spacings, thus was left unclassified. The SH3 domain of Pex13p also exhibits dual specificity for class II ligands and for a second motif characterized by a positively charged residue located between proline residues, which does not fit any defined class. The specificity profiles for the paralogs Boi1p and Boi2p (PXXXPX+) resemble class II, but with proline spacings that differ from the canonical binding motif, and have also been left unclassified. Finally, as observed previously [2], the SH3 domain of Fus1p exhibits a unique specificity profile that does not include prolines.
To compare the intrinsic specificities of yeast SH3 domains, we quantified the specificities using a specificity potential (SP) score, which was applied previously to the PDZ domain family [3]. The SP value summarizes the specificity observed in each column of a PWM as a numerical value ranging from zero (least specific) to one (most specific; Table S3). We had sufficient peptide data (n>10) to calculate reliable SP scores for 26 distinct specificity profiles. By summing the SP score across all PWM columns, we calculated a total SP (SPt) score for each SH3 domain specificity profile. Most yeast SH3 domains exhibit similar intrinsic specificities with SPt values ranging from four to six (Figure 2A). Furthermore, domains that recognize more than one class of ligands do so with approximately the same level of specificity for each class. This analysis reveals that the Cyk3p SH3 domain [16]–[18] has an unusually high SPt value for class II ligands, which stems from its strong preference for an Asp-Tyr motif downstream of the Arg residue of the canonical class II motif (Figure 1).
To assess the specificity contribution from different elements in the binding profiles, we quantified separately the SP scores for the positions within or outside the core motif for the various specificity profiles (Figure 2B). The core positions for classes I and II only contribute roughly half of the SPt value, with the other half being contributed by other positions that define distinct specificity niches. Analogously, residues outside the core positions contribute approximately the same level of specificity for the unique sets of ligands recognized by Lsb1/Pin3 and Boi1/Boi2 (Figure 2B). For class III domains, we found that recognition of proline accounts for approximately 60% of the SPt. Taken together, these results highlight the importance of residues outside the core positions for mediating specificity in SH3 domain–ligand interactions.
Phage display generally selects high-affinity ligands through an iterative panning process, and high-resolution PWMs have been used to predict preferences in selectivity that reflect binding affinities for PDZ domain–ligand interactions [3],[19]. To assess the accuracy of our phage-derived data for SH3 domains, we examined the SH3 domain of Sho1p and determined the correlation of PWM scores derived from phage display to differences in Gibbs free energy (ΔΔG) derived from previous in vitro binding assays with synthetic peptides [20] (Table S4). We observed an excellent correlation between the two datasets (r2 = 0.97; p = 7.8×10−5; Figure 3A), and a significant correlation was also observed for similar datasets for the SH3 domain of Abp1p [21] (r2 = 0.73; p = 2.1×10−4; Figure S2 and Table S5). For the SH3 domain of Sho1p, the correlation between binding affinity and PWM score match is proportional to the number of peptides used to generate the PWM, and good correlation is observed for datasets containing >30 peptides (r2>0.8). Notably, 22 of our SH3 domain specificity profiles are derived from 30 or more ligands, suggesting that the majority of our phage-derived data can predict accurately the relative in vitro affinities of peptide ligands for SH3 domains.
The use of synthesized peptide arrays offers an alternative and independent approach to query PRM–ligand interactions. In an ideal scenario, unique peptides representing the entire proteome of the organism would be spotted onto an array and assayed individually for interactions with a PRM of interest. However, in practice, a filtering step is required to generate an array of manageable size. In a strategy dubbed WISE (whole interactome scanning experiment), natural ligands for PRMs are identified by computationally scanning the proteome for sequence patterns similar to known ligands, and these are tested for interactions using synthetic peptide array (SPOT) technology. Proteome scanning can use regular expressions (REs), which describe discrete text patterns, or PWMs, which describe probabilistic positional frequency-based patterns. Although both methods rely heavily on the quality of the information they are based upon, REs run a higher risk of missing candidate ligands (higher false-negative rate), whereas PWMs often fail to catch strict position-specific rules. Following identification of putative natural ligands by either filtering method, the peptides are tested for interactions by SPOT.
We used the WISE approach to generate a yeast SH3 domain interaction network independently by creating a set of 15 REs based on SH3 domain specificity profiles identified in this study and previously [2], and searching for matches in the yeast proteome (Text S1). The stringency of the REs was set very low in order to maximize the number of putative ligands tested on the array. Although this approach potentially identifies a number of false positives, the goal is to capture as many interactions as possible, thus minimizing the number of false negatives. This analysis identified 2,953 peptides within 1,693 proteins (almost one-third of all yeast ORFs; Table S6). This defined set of peptides was synthesized on cellulose membranes according to a modified SPOT synthesis approach [22]. Subsequently, peptide arrays were screened for binding individually with 26 SH3 domains. In total, we identified 295 peptides that showed a positive signal with at least one SH3 domain (Table S7).
Peptides identified by either PWMs or REs address the ability of a domain to bind to a ligand outside of its protein and cellular context, but the peptides are identified by independent computational approaches with different strengths and weaknesses. To address this, we used the PWMs to define a set of peptides of similar size to the one defined by the REs. Interestingly, this analysis revealed only an approximately 30% overlap between the peptide sets defined by REs and PWMs. To examine the PWM-defined peptides experimentally, we tested in the SPOT assay the ten peptides with the highest PWM score for each SH3 domain. Of the 230 PWM high-scoring peptides, 113 were not included in the original WISE interactome, and approximately 55 of these gave a significant SPOT signal with at least one SH3 domain (Table S8). The 55 peptides predicted by PWM but missed by RE that yielded a significant SPOT signal can be regarded as false-negative interactions for the RE approach; therefore, the false-negative hit rate for the RE-defined peptides appears to be approximately 20%. Notably, of the 230 PWM high-scoring peptides, 69 did not generate a SPOT signal, which suggests that the PWM false-positive rate is on the order of approximately 30%.
The SPOT approach is semiquantitative, so we also examined the correlation between interaction signals and dissociation constants, but we found that, as reported previously [22], it was much poorer than that observed with the phage-display score (unpublished data). Thus, although SPOT assays can be used to validate PWM-predicted interactions, further development of the method is required to obtain highly accurate quantitative signals. Taken together, our SPOT analysis of the yeast SH3 interactome yields a weighted graph of more than 5,000 edges, which served as an additional source of semiquantitative information to be integrated into a map of yeast SH3 domain interactions.
To complement the phage display and SPOT experiments, we performed large-scale yeast two-hybrid screens. We screened 22 yeast SH3 domain baits against a novel yeast activation domain ORFeome library [23], which tests for interactions with full-length proteins, using an array-based approach as described previously [24] and repeating each screen twice (Table S9). In addition, 26 SH3 domain baits were screened against a randomly fragmented genomic library (gDNA), which tests for interactions with protein fragments [25] (Table S9). In total, we identified 801 unique interactions, consisting of 241 and 587 interactions from the ORFeome or gDNA library screens, respectively (Table S10). Only 26 interactions were identified in both screens (10.7% or 4.4% of the interactions identified by the ORFeome or gDNA screens, respectively). Using the ORFeome screen, we identified an average of 11.0 interactions per SH3 domain, whereas we detected an average of 22.6 interactions per SH3 domain in the gDNA screen. One major reason for the difference in these numbers is that we sequenced approximately 200 positive single colonies from each gDNA library screen in an attempt to saturate the system. Furthermore, although we repeated the ORFeome screening twice, this is not expected to achieve complete saturation according to a recent analysis [23]. In total, we sequenced 3,965 yeast two-hybrid–positive colonies, and some interactions were captured multiple times (593 interactions were captured at least twice) by each screening technique (Table S10).
To assess the potential of identifying biologically relevant interactions, we examined the number of literature-validated interactions that were identified by each approach. To do so, we curated a comprehensive “gold-standard” set of 42 SH3 domain interactions from the literature (see below). Within this gold-standard set, only five were identified by the ORFeome screen, whereas 28 were identified by the gDNA screen. Thus, with yeast SH3 domains, gDNA two-hybrid screening has a 2.5-fold lower false-negative rate than ORFeome analysis (Figure S3), which may reflect both that our screening of the gDNA library was more extensive and that it contains gene fragments corresponding to protein domains, which often behave better in the two-hybrid system [26]. Taken together, these results highlight the complementary nature of ORFeome and gDNA screening methods to experimentally identify protein interactions for PRMs.
As yeast two-hybrid and phage display potentially query different regions in interaction space, we sought to determine the overlap between the two methods. The phage-derived PWMs were used to search the yeast proteome for matching peptide ligands based on a PWM-scoring algorithm. For each SH3 domain, the yeast proteins were ranked according to their associated PWM score. Subsequently, the fraction of yeast two-hybrid hits containing predicted ligands with a rank higher than a defined threshold (x = 1, 2,…N, where N is the size of the yeast proteome) was determined. We find that approximately 10% of two-hybrid positives rank among the top ten hits predicted by the PWM of the associated SH3 domain (Figure 4, dashed line). The fraction of yeast two-hybrid hits with peptide sequences ranked among the top ten PWM-predicted ligands is increased to more than 25% when considering interactions that are captured at least six times, suggesting that these interactions have a higher likelihood of representing bona fide SH3 domain ligands (Figure 4, solid line). The high fraction of yeast two-hybrid positives with high-scoring PWM matches, compared to those predicted for random interactions, suggests that the detailed binding specificity uncovered by phage-derived PWMs was recapitulated using the yeast two-hybrid system.
Each experimental method has different strengths and biases, and the integration of data from independent techniques increases the accuracy of the resulting dataset substantially [27]. We generated a yeast SH3 domain protein–protein interaction network and used a statistical approach based on Bayesian networks [27] to assign each interaction a probability score. This score is based on the confidence level of the experimental data that defined the interaction benchmarked by the gold-standard set (see Materials and Methods and Table S11). A Bayesian networks formalism was chosen for the machine learning because it has been shown previously to perform well at integrating heterogeneous biological data [27],[28].
The gold-standard set represents a list of manually curated interactions known to be mediated by a specific SH3 domain, compiled through an exhaustive literature search. Each interaction in the gold-standard set is supported by multiple experiments reported in one or more focused studies, which show the direct binding of the SH3 domain to its target, and its functional relevance.
Each technique utilized in our analysis encompasses a quantitative measure: first, the phage-derived PWMs accurately represent relative binding affinities; second, interactions identified by SPOT peptide arrays can be binned and ranked based on intensity (see Materials and Methods); and third, interactions captured multiple times by yeast two-hybrid can be assigned a higher score than those captured only once. Furthermore, the different methods have complementary features. Whereas the phage display and SPOT peptide array signals correlate with and predict binding affinity, the yeast two-hybrid system identifies putative in vivo interactors of SH3 domains. We therefore integrated these datasets into a Bayesian model to identify highly likely SH3 domain–ligand interactions.
All interactions in the gold-standard set were mapped specifically to an SH3 domain and, where applicable, to the peptide sequence within the interacting partner (see Materials and Methods and Table S12). We generated a negative set using random protein pairs under the constraint of never sharing or being in “adjacent” cellular compartments (see Materials and Methods). To determine the sensitivity of each technique individually, we plotted their respective receiver-operating characteristic (ROC) curve, a standard assessment of accuracy, and examined the area under the curve (AUC; Figure 5). The phage-derived PWMs were found to exhibit the highest AUC (0.91; Figure 5A and Figure S4), with the SPOT peptide array and yeast two-hybrid exhibiting a lesser value (0.85 in both cases; Figure 5B and 5C, respectively). Remarkably, the Bayesian network, which integrates the data from all three techniques, results in an AUC of 0.94 (Figure 5D; p = 1.2×10−10; Figure S5), suggesting that our probabilistic interaction network captures the vast majority of literature-validated interactions. The entire set of yeast SH3 domain–ligand interactions predicted by our Bayesian model is represented as a network diagram and summarized in table format (Figure S6 and Table S12).
To assess how profile specificity translates into specificity at the level of the network, we computed for each SH3 domain, the fraction of its interactors in the Bayesian network that are targeted by at least one other domain (Figure S7). Our results show that many proteins (61%) are targeted by only one SH3 domain. The other proteins (39%) are predicted to bind to more than one SH3 domain. Furthermore, important differences between SH3 domains can be observed, some of them having very unique specificity (e.g., Fus1p SH3 domain), whereas others share most of their interactors with other domains. The latter is especially true for SH3 domains from paralogous proteins such as Boi1p/Boi2p, Lsb1p/Pin3p, Lsb3p/Lsb4p, and Myo3p/Myo5p.
To study specificity further, we also distinguished the different predicted binding sites on each protein (binding sites are predicted by the best PWM hit in the protein sequence), since a protein can be targeted by multiple SH3 domains but at different binding sites. Interestingly, the fraction of binding sites targeted by more than one SH3 domain is lower than the fraction of proteins targeted by more than one SH3 domain (29% against 39%), revealing that some proteins have multiple unique binding sites recognized by individual SH3 domains (Figure S7, grey bars). However, cases of possible competition are not completely removed by distinguishing the different binding sites.
To assess the contribution of SH3 domains from the same protein and highly similar SH3 domains, we merged Bzz1-1 and Bzz1-2, Sla1-1/2 and Sla1-3, and the four pairs of close paralogs (Boi1p/Boi2p, Lsb1p/Pin3p, Lsb3p/Lsb4p, and Myo3p/Myo5p), treating each of them as a single protein since they have highly similar specificity profiles. In this case, we found that 33% of all interactors are targeted by more than one SH3-containing protein in our network (Figure S8). As previously, we distinguished the different binding sites in each protein target and found that 23% of binding sites are targeted by more than one SH3-containing protein (Figure S8). Thus, the majority of interactions are expected to be insulated from competition effects, due to sequence differences among binding sites alone, though some competition among domains is likely.
As one approach to assessing the biological relevance of interactions identified by the Bayesian model, we examined biological process annotation associated with the putative SH3 domain ligands, defined by Gene Ontology (GO). We found a significant number of overrepresented biological processes known to be associated with yeast SH3 domain biology such as establishment of cell polarity and endocytosis (p = 3×10−7 and p = 9×10−8, respectively). Moreover, from a recently published set of approximately 60 known and putative endocytosis proteins, 29 were found to be connected with at least one SH3 domain in our interaction network [29] (Figure S6). In addition, by searching for highly interconnected nodes in the Bayesian interaction network, we identified a core of 31 proteins that engage in at least six interactions with each other (k-core = 6; Figure S9). Consistent with the GO term enrichment analysis described above, 14 of the proteins that emerge from the k-core analysis (e.g., Las17p, Myo3p, and Vrp1p) have well-defined roles in endocytosis with a GO enrichment p-value of 5×10−8 [29],[30]. Hence, we decided to focus on the SH3-mediated interactions underlying endocytosis in more detail.
Endocytosis is a complex cellular process in which a dynamic array of protein interactions are sequentially coordinated to drive endocytic site initiation, membrane invagination, and vesicle scission [31]. Live-cell imaging analyses uncovered a detailed spatiotemporal map for the dynamic recruitment of numerous proteins to endocytic sites in budding yeast [31],[32]. These studies proposed the existence of four dynamic protein modules that cooperate to drive vesicle formation: (1) the endocytic coat module, (2) the Wiskott-Aldrich syndrome protein (WASP)-myosin (WASP/Myo) module, (3) the scission (or amphiphysin) module, and (4) the actin module.
Proteins in the endocytic modules arrive sequentially at sites of endocytosis with precisely defined dynamics and their assembly drives the steps of endocytic internalization. The first step in the endocytic pathway is the recruitment to the plasma membrane of the coat module proteins, which include clathrin, Sla1p, Pan1p, End3p, and Sla2p. The assembly of the coat module occurs prior to and independent of actin assembly. However, the subsequent movement of the coat proteins into the cell, and subsequent coat disassembly, are dependent upon actin polymerization. One to two minutes following coat module assembly, Las17p (the yeast ortholog of WASP) is recruited, which activates the Arp2/3 complex to promote actin assembly. The SH3 protein Sla1p is thought to inhibit the actin polymerizing function of Las17p. This inhibition appears to be relieved by the recruitment of members of the WASP/Myo complex, including Vrp1p and the SH3 proteins Bbc1p, Myo3p/Myo5p, and Bzz1p. Actin polymerization triggered by the WASP-myosin complex leads to recruitment of the actin module proteins, which include actin, Cap1p, Cap2p, Sac6p, Abp1p (SH3 protein), and the Arp2/3 complex, leading to further actin polymerization. As the vesicle begins its movement into the cell, the scission module, consisting of Rvs161p and the SH3 protein Rvs167p, is recruited. Although the exact scission mechanism is unclear, the scission module promotes the release of the nascent endocytic vesicle [29],[31]. In contrast to components of the coat module, proteins of the WASP/Myo module remain immotile at the plasma membrane as actin is being polymerized, and disassemble as the nascent vesicle is internalized [29],[31].
Spatiotemporal characterization of protein dynamics by live-cell imaging has provided a detailed view of endocytosis, but our understanding of this pathway is far from complete. It has been established that numerous proteins arrive at sites of endocytosis in a choreographed manner, but it is not known how the sequential recruitment, assembly, and functions of endocytic proteins are achieved. Our Bayesian interaction network contains 29 of the 60 or so known yeast endocytosis proteins, including ten that contain SH3 domains. To gain insights into the roles of SH3-mediated interactions in endocytosis, we screened for putative ligands for these ten SH3 domains using our Bayesian scoring algorithm (Table S13). The interacting proteins were grouped with the respective protein modules described above, and the putative SH3-mediated interactions at each stage of endocytosis were determined (Figure 6 and Table S13). This analysis uncovered a vast array of putative SH3 domain–mediated interactions, with 53 connections among 19 known or putative endocytic proteins, and suggested that interactions are likely to become more prevalent as additional proteins are recruited to the endocytic site (Figure 6). Furthermore, the interaction network suggests that the majority of SH3 domain–mediated interactions are established 35 to 15 s prior to vesicle internalization (Figure 6C to 6E). This timing suggests that SH3 domains play a particularly important role at the stages encompassing assembly of the WASP/Myo module, actin polymerization, membrane invagination, and vesicle scission.
The network allows us to map potential interactions onto the temporal order of protein recruitment at the site of endocytosis, and these interactions likely mediate assembly of protein modules and coordinate activities between the modules (Figure 6). We therefore examined in greater detail the relationships between SH3 domain–mediated interactions and protein dynamics during endocytosis. For each protein, we summed Bayesian probability scores (or interaction scores) based on interactions with proteins from within its corresponding module compared to interactions with proteins from external modules (Table S13). This analysis revealed that proteins have the highest total interaction score for interactions occurring within the same module. This was the case for 11 of 13 endocytic proteins for which a suitable Bayesian probability score and dynamic data were available (Table S13). For instance, the network identified a large number of interactions between members of the WASP/Myo module (Bbc1p, Bzz1p, Las17p, Vrp1p, Myo3p, and Myo5p), which arrive following the coat module, 35 to 25 s prior to vesicle internalization (Figure 6C and 6D). Summing their Bayesian probability scores across all modules revealed that each of these proteins has the highest combined interaction score for interactions within the WASP/Myo module. This finding provides support for the conclusion that the SH3 domain–mediated interactions are required for the assembly of this module, and that interactions between these proteins are established upon their temporal recruitment to the endocytic site.
Subsequent to the formation of an SH3 domain–mediated network within the WASP/Myo module, the network analysis points to the formation of an SH3 domain–mediated network within the actin module (e.g., Abp1p, Ark1p, Prk1p, and Sjl2p [33]), at 25 to 10 s prior to vesicle internalization (Figure 6D and 6E). Interestingly, proteins from the actin module also appear likely to engage in many interactions with members of both the WASP/Myo and actin modules, suggesting extensive cross-talk between these two modules (Figure 6). However, the interaction scores for proteins within the same module were higher than those for proteins in different modules, underscoring the predictive potential of using interaction scores to place endocytic components into their respective modules (Table S13).
Our network analysis, which incorporates both SH3 domain–mediated interactions and dynamics of endocytic proteins, suggests that members from the same endocytic module engage in tighter SH3 domain–mediated interactions and have similar spatiotemporal dynamics. This raises the possibility of predicting the dynamics of putative endocytic proteins based on their SH3 domain interaction profile. Thus, an uncharacterized endocytic protein is predicted to be part of the module within which it registered the highest interaction scores. For example, if an uncharacterized protein is implicated as a member of the WASP/Myo module because it has high scores with SH3 domains within the WASP/Myo module, then we predict that its dynamics will follow a similar pattern to those of other proteins in that module. Analogously, an uncharacterized SH3 domain protein would be predicted to be part of the module containing its best-predicted binding partners.
To test our hypothesis, we quantitatively examined the protein dynamics of five uncharacterized endocytosis proteins (Scd5p, Aim21p, Scp1p, Bsp1p, and Lsb3p), each of which had a high SH3 interaction score with at least one of the established endocytic modules (Table S13). We predicted that Scd5p, a protein first identified as a suppressor of defects in cells depleted of clathrin heavy chain (Chc1p) [34], arrives with and is part of the late coat module (with Sla1p, and Sla2p, etc., but not with the early coat protein, clathrin) and/or WASP/Myo module. We also predicted that Aim21p, a fungal-specific protein, is a component of the WASP/Myo module, and that Scp1p, a conserved member of the Calponin/transgelin family of actin-associated proteins [35], is part of the actin module.
Two closely related SH3 domain proteins, Lsb3p and its paralog Lsb4p, had high interaction scores across several modules, most significantly with early (e.g., coat and WASP/Myo) and late (actin) modules (Figure 6 and Table S13). Notably, we observed that the score for a particular module did not exceed the median interaction score across all other modules by more than 2-fold. This unique pattern of interactions suggests that Lsb3p and Lsb4p may play a role to cluster and to coordinate the activities of several module components at the site of endocytosis. In addition, Bsp1p, an adapter that links the yeast synaptojanins, Inp52p and Inp53p, to the cortical actin cytoskeleton and participates in actin contractile ring function [36], showed a similar interaction profile, and therefore, we speculated that it too might be a cross-module protein together with Lsb3p and Lsb4p.
To test our predictions in vivo, each protein was C-terminally tagged with GFP and expressed from its endogenous locus in yeast cells. The dynamics of each protein were analyzed individually and in tandem with Abp1p-RFP. The dynamic localization analysis validated our approach for assigning proteins to endocytic modules based on their interaction scores (Figure 7 and Table S13).
In agreement with earlier observations [37], we found that Scd5p-GFP patches had a lifetime of 22±6 s (Figure 7A). Simultaneous, two-color analysis with Abp1p-RFP, a marker for actin polymerization, revealed that Scd5p-GFP arrives prior to actin polymerization (Figure 7B). However, Scd5p-GFP patches were immotile throughout their lifetime, like proteins of the WASP/Myo module (Figure 7C). These dynamics establish Scd5p as a component of the WASP/Myo module with similar dynamics to Bzz1p, suggesting that it might participate in late coat formation and/or coordinate this module with the WASP/Myo module. Moreover, Scd5p was recently reported to have a role in phospho-regulation of the endocytic coat complexes and its spatial dynamics may have a role in this essential function [37].
For Aim21p-GFP, we observed that it is located in immotile patches with a lifetime of 10±1 s, similar to the patch dynamics reported for Bbc1p [29] (Figure 7A and 7B). Furthermore, Aim21p-GFP arrives when actin begins to polymerize, as revealed by two-color analysis with Abp1p-RFP (Figure 7C). Thus, as predicted, Aim21p localizes as a component of the WASP/Myo module (Figure 6D and Table S13).
Scp1p is expected to be part of the actin module as it is predicted to bind the SH3 domain of Abp1p. Indeed, Scp1p-GFP formed patches with a lifetime of 15±2 s (Figure 7A) and colocalized with Abp1p (Figure 7B and 7C) [30]. These patch dynamics are indicative of proteins in the actin module. Scp1p patches had shorter lifetimes than Abp1p. However, Scp1p-GFP exhibited weak fluorescence intensity, which likely accounted for this lifetime decrease. Two-color analysis revealed strong colocalization between Scp1p and Abp1p with the fluorescence intensity of the patches peaking together (Figure 7C; unpublished data).
As mentioned above, Lsb3p and Lsb4p scored highly across all modules, predicting a long lifetime at the patch. As expected, Lsb3p-GFP patches were long-lived with a lifetime of 36±9 s (Figure 7A). Lsb3p-GFP patches arrived at the cell cortex as an immotile patch, but showed an initial slow movement into the cell to a depth of about 200 nm. The initial slow movement was then followed by a fast, more randomly directed movement further into the cell (Figure 7B). Two-color simultaneous imaging with Abp1p-RFP revealed that, like Sla1p, Lsb3p-GFP arrived early at endocytic sites but persisted late with the actin module proteins (Figure 7C) [30]. These dynamics are consistent with the prediction that the Lsb3p and Lsb4p SH3 domains interact with Sla1p and several actin module proteins (Figure 6). Thus, Lsb3p and Lsb4p appear to provide continuity in the context of a continuously evolving endocytic protein composition.
Finally, we analyzed the dynamics of Bsp1p, which our model suggested interacts with proteins in all modules, similar to Lsb3p and Lsb4p. However, in contrast to the Lsb proteins, Bsp1p-GFP patches were short-lived with a lifetime of 13±2 s (Figure 7A). Bsp1p-GFP colocalized well with Abp1p-RFP and displayed an Abp1p-like motility behavior (Figure 7B and 7C) [30], suggesting that Bsp1p functions within the actin module. Two-color analysis revealed that Bsp1p consistently arrived approximately 1 to 2 s after Abp1p-RFP, in a manner similar to descriptions for Cof1p, Ark1p, or Prk1p [38],[39]. Moreover, unlike other patch proteins, Bsp1p-GFP had an additional stable localization to the bud neck as described previously [40]. Bsp1p is not well characterized, and further studies are necessary to understand the nature of the discrepancy between its predicted interactions with proteins of multiple modules and its appearance only late in the pathway during the burst of actin assembly.
Our SH3 domain network contains only approximately half of the 60 proteins implicated in endocytosis and, as underscored by the Bsp1p example, a number of SH3-independent interactions must control endocytosis protein localization. To emphasis this point, we also analyzed the dynamics of yeast twinfilin (Twf1p), a highly conserved actin monomer-sequestering protein involved in regulation of the cortical actin cytoskeleton [41], which was not predicted to bind to any SH3 domain. Similar to Scp1p and Bsp1p, Twf1p localized to the patch with a lifetime of 15±2 s (Figure 7A). The patches were initially immotile at the cell surface but subsequently underwent a highly motile phase, after which the patch moved long range into the center of the cell (Figure 7B), in a manner characteristic of proteins comprising the actin module.
In summary, SH3 domain interactions are powerful predictors of spatiotemporal localization of yeast SH3 domain proteins. The putative SH3 domain–mediated interaction networks allowed us to accurately predict the dynamics of several previously uncharacterized proteins in the endocytic pathway and provided a detailed spatiotemporal map of the endocytic pathway (Figure 8).
We generated a specificity map for the SH3 domain family of budding yeast based on 1,871 unique peptide ligands isolated against 25 of the 27 domains. This map reveals that SH3 domains have a high level of intrinsic specificity and different domains recognize distinct sets of ligands. Notably, specificity was observed for ligand positions outside of the core positions, suggesting that SH3 domains utilize multiple features of their peptide ligands to achieve binding selectivity.
A major challenge in functional proteomics is the development of accurate protein interaction networks. We have integrated the data from three independent screening techniques (phage display, peptide arrays, and yeast two-hybrid) into a Bayesian model to generate a yeast SH3 domain interaction map. Each technique has a semiquantitative measure that was captured by the probabilistic model. Our interaction map captures a significant proportion of literature-validated interactions and therefore serves as an accurate reference for additional in-depth studies of yeast SH3 domain biology. Proper interpretation and use of our interaction map requires consideration of additional factors such as cellular concentration, localization, and competition from other SH3 domains to identify physiologically relevant interactions.
Applying our model to proteins involved in endocytosis revealed that there is a significant connection between the time at which a protein arrives at the endocytic patch and its predicted SH3 domain interactions. This correlation was used to accurately predict the spatiodynamics of several uncharacterized endocytic proteins. We found that Scd5p and Aim21p are both components of the WASP/Myo module, which drives vesicle internalization by nucleating actin filament assembly and generating myosin motor-based forces on the actin filaments [29]. Future studies will reveal how these proteins contribute to the function of the WASP/Myo module, but the presence of Scd5p in the WASP/Myo module may be important for its role in phospho-regulation of the endocytic machinery [37]. We also found that Scp1p, Bsp1p, and Twf1p are components of the actin module. Both Scp1p and Twf1p are known actin-binding proteins and may play a role in modulating actin dynamics at endocytic sites [35],[41]. The novel dynamics observed for Lsb3p indicate that it is present across all modules. Based on conserved SH3 predictions and sequence homology, we propose that Lsb4p has similar dynamics. Their numerous predicted interactions, and their dynamics and association with multiple endocytic modules, suggest that Lsb3p and Lsb4p may play an important role in coordinating the activities of the various endocytic modules.
The SH3 interaction predictions did not agree with the dynamics of Sla1p and Bsp1p. Sla1p appears in the coat module, whereas its predicted interactions are more consistent with it being a component of the WASP/Myo module. This may be explained by its established essential role in regulating the WASP/Myo module [29]. Perhaps Sla1p integrates its cargo adaptor role [42] with its roles in actin assembly to prevent premature actin nucleation, and perhaps its departure from the cell surface with the coat proteins separates it from the WASP/Myo proteins, further relieving its inhibition of actin polymerization. Unlike Sla1p, Bsp1p is less well studied and lacks any obvious homology with other proteins. Bsp1p has been linked to the actin module protein Sjl2p, which regulates phosphatidylinositol 4,5-bisphosphate levels [36]. Furthermore, Bsp1p plays a role in actomyosin ring function [40], but it is unclear how this relates to its role at endocytic sites [40]. The delayed recruitment of Bsp1p to the actin module also suggests a role in endocytic site disassembly alongside Sjl2p, Ark1p, Prk1p, and Cof1p [38],[39].
Given the conserved nature of endocytosis from yeast to human [31], it will be of great interest to examine SH3 domain interaction networks in more complex organisms. We emphasize the feasibility of the approach presented here, given the recent discovery that orthologous protein modules generally have very similar specificity profiles [3]. Recent studies of PDZ domains have shown that PRMs are more specific than previously appreciated [3],[43], and we show that the same holds true for SH3 domains. The intrinsic specificity observed at the level of the protein domain itself suggests that there is significant selective pressure driving the domain into a specificity niche not utilized by other domains. As postulated previously [20], an interplay between positive specificity selection acting on the protein interaction module and negative selection acting upon its cognate ligands would ensure high specificity without the requirement of a high-affinity interaction. It appears that such specific interactions have evolved and are necessary to create the dynamic and intricate signaling pathways required for cellular functions.
For cloning, the SH3 domain boundaries were defined as the union of the domain regions identified by BLAST [12], PFAM [13], and SMART [14], plus an additional ten amino acids (where applicable) on either side as described previously [3]. DNA fragments encoding the identified domains were amplified from S. cerevisiae genomic DNA by the polymerase chain reaction (PCR) and cloned into a vector designed for the expression and purification of SH3 domains fused to the C-terminus of glutathione S-transferase, as described [44]. All plasmid constructs were verified by DNA sequencing.
Phage-displayed peptide libraries (>1010 unique members) fused to the N-terminus of the gene-8 major coat protein of M13 filamentous phage were used to select peptide ligands for the collection of purified GST-SH3 fusion proteins. All domains were first screened using a random decapeptide library (X10, where X is any amino acid). Domains that failed to select peptides with the decapeptide library were subsequently screened using a biased peptide library (X6-PXXP-X6, where P is proline). Three SH3 domain proteins (Cyk3, Lsb4, and Sla1-1/2) were also tested using a biased library containing a fixed positive charge (X7-R/K-X7, where R and K are arginine or lysine, respectively). Phage display selections were carried out as described [44]. Individual binding clones were tested for positive interactions with cognate yeast SH3 domains by phage ELISA as described [44]. The sequencing of approximately 3,000 clones resulted in the isolation of 1,871 unique peptide sequences, which were manually aligned by an expert (RT). The phage library used to select peptides for each domain is indicated in Tables S1 and S2. As some peptide files contain peptides selected from different libraries (Cyk3-class II, Lsb4, and Sla1-1/2-class II), the library from which each peptide was isolated is also indicated in each sequence file.
For each SH3 domain, the set of peptide ligands was used to create a binding profile statistical model as a PWM. The specificity potential (SP) for a given column (position) of a PWM was calculated as is done for the letter height in a sequence logo [15], except normalized to range from 0 to 1 instead of 0 to 4.32 (log 20). A SP value of one means the given PDZ domain is completely specific for a single amino acid at that position, and a value of zero means that there is no preferred amino acid at that position. We have also included a p-value to assess the statistical significance of these scores. The p-values were computed by statistical sampling: for each PWM, we generated 107 sequences of N randomly chosen amino acids, with N equal to the number of different peptides used to build the PWM. For each sequence, we computed the SP score, and from the distribution of SP scores, we computed the p-value of the SP scores for each column in the initial PWM (Table S3).
Peptide arrays were semi-automatically prepared on cellulose-(3-amino-2-hydroxy-propyl)-ether membranes [22] (CAPE membranes) using a SPOT robot (Intavis) and the standard SPOT synthesis protocol [45]. Array design was generated using the in-house software LISA. To exclude false-positive spots in the incubation experiment, all cysteine residues were replaced by serine. The CAPE membranes were used because of the better signal to noise ratio in the incubation experiments.
The peptide arrays were incubated with the GST-SH3 fusion proteins, as described [22]. Analysis and quantification of membrane-bound GST-SH3 fusion proteins was carried out using a chemiluminescence substrate and a Lumi-Imager (Roche Diagnostics). Analysis and quantification of SPOT signal intensities (SI) were executed with the software Genespotter (MicroDiscovery) following previously described rules [46].
DNA fragments encoding SH3 domains were amplified by PCR from a S. cerevisiae genomic DNA library, using sequence specific primers fused to common sequences used for homologous recombination cloning. Specifically, the forward primer was composed of the bait-specific primer and a 23-nucleotide common sequence (CGACCCCGGGAATTCAGATCTAC), which is homologous to the upstream sequence of the SpeI site on pPC97 [23]. The reverse primer was composed of the bait-specific primer and a 23-nucleotide common sequence (CGGGGACAAGGCAAGCTAAACTA), which is homologous to the 5′ of the KanMX6 cassette. The KanMX6 cassette was amplified by PCR with the forward primer (TTTAGCTTGCCTTGTCCC) and the reverse primer (ATAGATCTCTGCAGGTCGACGGATCCCCGGGAATTGCCATTTTTCGACACTGGATGGC), using a KanMX6 cassette carrying plasmid, p2076, as template. Along with the PCR-amplified KanMX6 cassette and the SpeI-cut pPC97, bait coding sequence PCR product was transformed into Y8930 (MATα trp1-901 leu2-3,112 ura3-52 his3-200 gal4Δ gal80Δ LYS2::GAL1-HIS3 GAL2-ADE2 met2::GAL7-lacZ cyhR), which was generated from PJ69-4α [47]. G418 positive yeast transformants were selected on SD-Leu+G418 medium, and yeast DNA was purified and transformed into Escherichia coli DH5-α. Constructed plasmid was purified from kanamycin-positive clone, verified by DNA sequencing, and transformed into Y8930 for Y2H screening.
The whole library was assembled in an 1,536-spot array format on agar plates with each clone represented twice. ORFeome Y2H screening was performed as described [24] with some modifications. The optimal concentration of 3-amino-1,2,4-triazole (3-AT) was tested for each bait before performing the screen. SD-Leu-Trp-His+3-AT selective medium was used for screening. Plates were incubated at 30°C for 5 to 10 d before scoring positive colonies.
All pOBD plasmids were taken from Tong et al. [2], and the pBDC plasmids were cloned by homologous recombination as described above (Table S9) and verified by sequencing. All bait plasmids were transformed into Y8930. A genomic DNA library [25] was transformed into Y8800 (same genotype as Y8930 except opposite mating type). Screening was performed by mating methods as described previously [48] on SD-Leu-Trp-His+3AT plates. Up to 192 positive single colonies were picked from each screen. The identity of each positive colony was determined by colony-PCR and sequencing.
To compile a comprehensive list of yeast SH3 domain–ligand interactions supported by one or more experiments (referred to as the gold-standard set), we used a combination of automatic text mining and database searches to retrieve abstracts from the literature. The DOMINO database, specialized in domain–peptide interactions, already contained 22 entries for yeast SH3 domains, curated according to the MIMix standards from 14 papers [49],[50]. A text-mining approach looking for co-occurrence in the abstract of names of yeast proteins together with “SH3” and a list of nouns and verbs indicating interactions yielded only two papers containing relevant information after manual inspection. An additional 19 papers were captured by manual searching PubMed (http://www.ncbi.nlm.nih.gov/sites/entrez) and the Saccharomyces Genome Database (SGD; http://www.yeastgenome.org/) [51],[52]. These 21 new papers, which were not already present in DOMINO, were read, and the information supporting interactions mediated by SH3 domains was captured in a MIMix format. A total of 56 new interactions were added to the DOMINO database by this curation effort. Since some interactions are supported by more than one report, this amounts to a total of 41 nonredundant interactions mediated by SH3 domains and supported by at least one experiment. One paper reported a yeast two-hybrid interaction between Las17p and a protein fragment encoding both SH3 domains of Bzz1p. As the SH3 domains were not tested individually for interaction with Las17p, we counted the interaction twice to account for the two SH3 domains, thus resulting in 42 total SH3-mediated interactions [53]. The Bzz1 domains were tested individually by GST pull-down and Western blot analysis, and both domains interact with Las17p (A. Soulard and B. Winsor, personal communication). The curated gold-standard list is contained in Table S11.
The peptides from phage display were converted into position weight matrices (PWMs) by calculating the probability of occurrence for each amino acid at each position. Despite the large number of peptide sequences, we still substantially undersampled sequence space, and hence added pseudocounts. We scaled the number of pseudocounts added by the entropy of each position [54]. Each matrix was used to scan the yeast proteome to identify the best matches. We used the MOTIPS analysis pipeline to identify possible binders for each domain. Only the proteome-scanning module of the pipeline was utilized, which performs a highly optimized search in the yeast proteome for optimal matches to a given PWM. It works in an analogous fashion to earlier proteome scanners (e.g., the Scansite server) [55].
We employed the Bayesian Network algorithm as implemented in the WEKA 3.4.13 Java libraries [56]. All pre- and post-processing of the data was carried out using custom code written in Perl and Java. Bayesian networks can efficiently integrate different types of data and accurately estimate the probability of interactions based on different experiments [27]. The different data sources were first preprocessed as follows: the Y2H hits were put in one of two bins, depending on whether the associated clone was found once or more than once and given scores of one or two, respectively. The resulting discrete data were then fed directly into the learning algorithm. In the preprocessing step, the SPOT peptide binding data was discretized into four bins. The discretized data were then used as one feature of the learning algorithm.
To ensure a reliable set of gold-standard true-positive interactions for efficient machine learning, we used the curated list of 42 bona fide domain–peptide interactions for the yeast SH3 domains deposited in the DOMINO database, as described above [49]. We generated the true-negative set by using the “random with constraints” logic. Specifically, we included only pairs of proteins where protein A is annotated to localize to the cell membrane and where protein B is annotated to localize to the nucleus. Proteins with overlapping annotations were excluded as well. Although the first member of each gold-standard negative set was chosen to be one of the proteins containing SH3 domains, its interacting partner was under no such constraint. Since the proportion of real interactions is very low in the space of possible interactions, one can use random domain–ligand pairs to get a set likely to contain only negatives. However, we improved upon this set by filtering out only those pairs that do not occur in known interaction databases and are annotated to occur in nonadjacent cellular compartments. Specifically, we included only pairs of proteins where protein A is annotated to localize to the cell membrane and where protein B is annotated to localize to the nucleus. Proteins with overlapping annotations were excluded as well.
Performance of each data source was evaluated using the AUC (area under the curve) in the ROC curve. This corresponds to an evaluation of how well each data source corresponds to the gold-standard data. Finally, the performance of the Bayesian data integration was assessed using the AUC in a ROC curve analysis with 10-fold cross-validation. Ten-fold cross-validation corresponds to splitting the gold standard into a training (9/10) and a testing (1/10) set ten times in a rotating fashion and evaluating its accuracy for each split. Using the gold-standard set, we classified the discretized input data into the “True” (interacting) and “False” (noninteracting) labels as well as a probability score of the interaction. We report all interactions assigned a probability score of >0.6 (Table S12). The networks were created using Cytoscape 2.6 [57]. On the basis of affinity data for the Sho1p and Abp1p SH3 domains, we estimate that this cutoff represents a dissociation constant of approximately Kd = 1.5 µM.
Yeast strains were grown at 25°C in standard rich medium (YPD) or synthetic medium (SD) supplemented with appropriate amino acids. GFP tags were integrated chromosomally to generate C-terminal fusions of each protein, as described [58]. All strains expressing fluorescent fusion proteins had growth properties similar to the corresponding untagged strains.
For microscopy, cells were grown in SD medium without tryptophan (to minimize autofluorescence) at 25°C until early log phase. Cells were attached to coverslips coated with concanavalin A, which were sealed to slides with vacuum grease (Dow Corning). Imaging was done at room temperature using an Olympus IX81 or IX71 microscope equipped with 100× NA 1.4 objectives, and Orca II cameras (Hamamatsu). Simultaneous two-color imaging was done using an image splitter (Optical Insight) to separate the red and green emission signals to two sides of the camera sensor using a 565-nm dichroic mirror, and 530/30-nm and 630/50-nm emission filters. To excite GFP or RFP, we used a 488-nm Argon ion laser (Melles Griot) or a mercury lamp filtered through a 575/20-nm filter, respectively. The excitation beams from these two light sources were combined using a beam splitter. After each experiment, images of immobilized microbeads that fluoresce at both green and red wavelengths were captured. These images were used to align the cell images.
Image analysis was done as described [30]. Tracking of patches was done from single-color GFP movies to achieve the best signal-to-noise ratio. ImageJ (http://rsbweb.nih.gov/ij/) was used for general manipulation of images and movies.
|
10.1371/journal.pntd.0004996 | An Extended Multilocus Sequence Typing (MLST) Scheme for Rapid Direct Typing of Leptospira from Clinical Samples | Rapid typing of Leptospira is currently impaired by requiring time consuming culture of leptospires. The objective of this study was to develop an assay that provides multilocus sequence typing (MLST) data direct from patient specimens while minimising costs for subsequent sequencing.
An existing PCR based MLST scheme was modified by designing nested primers including anchors for facilitated subsequent sequencing. The assay was applied to various specimen types from patients diagnosed with leptospirosis between 2014 and 2015 in the United Kingdom (UK) and the Lao Peoples Democratic Republic (Lao PDR). Of 44 clinical samples (23 serum, 6 whole blood, 3 buffy coat, 12 urine) PCR positive for pathogenic Leptospira spp. at least one allele was amplified in 22 samples (50%) and used for phylogenetic inference. Full allelic profiles were obtained from ten specimens, representing all sample types (23%). No nonspecific amplicons were observed in any of the samples. Of twelve PCR positive urine specimens three gave full allelic profiles (25%) and two a partial profile. Phylogenetic analysis allowed for species assignment. The predominant species detected was L. interrogans (10/14 and 7/8 from UK and Lao PDR, respectively). All other species were detected in samples from only one country (Lao PDR: L. borgpetersenii [1/8]; UK: L. kirschneri [1/14], L. santarosai [1/14], L. weilii [2/14]).
Typing information of pathogenic Leptospira spp. was obtained directly from a variety of clinical samples using a modified MLST assay. This assay negates the need for time-consuming culture of Leptospira prior to typing and will be of use both in surveillance, as single alleles enable species determination, and outbreaks for the rapid identification of clusters.
| Leptospirosis is a zoonotic disease with more than 1 million cases per year globally and epidemics are increasingly reported. In this setting rapid typing is essential to identify potential clusters and transmission pathways. Typing of bacteria commonly requires bacterial isolates but culturing Leptospira is difficult and time consuming and requires invasive samples, such as blood or cerebrospinal fluid. We modified an existing typing scheme to lower the limit of detection and were able to amplify and sequence alleles directly from clinical specimens. Samples included blood (whole blood, serum, or buffy coat) and urine from patients diagnosed by PCR with leptospirosis between 2014 and 2015 in the United Kingdom and the Lao Peoples Democratic Republic. Using the sequences in phylogenetic analysis we identified the predominant Leptospira species in both countries as L. interrogans. With its increased sensitivity the modified assay allows for typing and species determination of Leptospira directly from blood or urine. It will be of use during epidemics and outbreaks for rapid identification of clusters and can support surveillance without the need to culture fastidious isolates.
| Leptospirosis is a zoonotic disease caused by pathogenic species of Leptospira that can be carried naturally by most mammalian species [1–3]. Transmission to humans most commonly occurs via direct animal contact or via water contaminated with animal urine [2, 4]. Symptoms range from a mild febrile illness to severe disease with pulmonary haemorrhage or central nervous system involvement [3, 5]. In its early stages leptospirosis resembles many other febrile illnesses, hampering clinical diagnosis. The highest disease burden is in tropical low and middle income countries, driven by high humidity, close human-animal contact, and inadequate sewage disposal and water treatment [3]. Annual worldwide case number was estimated at around 1 million with the majority of cases and death occurring in tropical regions [6]. Despite these relatively high numbers the epidemiology of leptospirosis is not well understood. Epidemics in humans and animals are increasingly reported and are often related to natural events like floods [3, 7]. In these settings rapid typing is essential to identify potential clusters and transmission pathways.
The gold standards for laboratory diagnosis of leptospirosis are culture or a four-fold rise in antibody titre between admission and convalescent samples by the microscopic agglutination test (MAT). Culture of Leptospira spp. is time consuming and diagnosis by MAT is retrospective by nature, hence both methods have disadvantages as diagnostic tools. To enable early detection several quantitative real-time PCR assays have been developed, some of which allow for species distinction [8–20].
Three MLST schemes are currently hosted by the public MLST database [21–23], two of which have been tested directly on clinical samples from humans [24–26]. Only two studies tried to amplify all seven loci and showed that MLST is possible directly from serum and whole blood. However the bacterial load required was high (~5x104 leptospira/mL) with only 21% and 5% or 10% success rates for partial and full profiles, respectively [24, 26]. The objective of this study was to develop an assay based on a published MLST scheme that lowers the limit of detection (LoD) to enable rapid provision of typing data directly from patient specimens whilst minimising costs for subsequent sequencing [22].
Specimens included in the study were not collated specifically for this study. Specimens included those within a collection of specimens submitted to the Public Health England Leptospira Reference Laboratory received routinely for Leptospira testing, identification of infecting species, confirmation of infection and for epidemiological investigation. Specimens were anonymised prior to testing. IRB board approval was not required as this involved routine specimens submitted for Leptospira testing by MLST as a secondary test for confirmation of infection and species identification and for the provision of epidemiological information.
The protocol was validated on 25 isolates from the WHO recommended Serovar panel (data in S1 Table) which is currently used for serological diagnostic and serovar identification. The assay was tested using 104 clinical specimens (45 serum, 6 whole blood, 13 buffy coat, 40 urine) from the UK (n = 35) and the Lao PDR (n = 69), (Mahosot Hospital Microbiology Laboratory, Vientiane). For initial laboratory diagnosis samples were tested with a triplex qPCR assay targeting the 16S rRNA gene (rrs) containing three different probes to distinguish between pathogenic, intermediate and environmental strains [27]. Using this assay, 44 samples (23 serum, 6 whole blood, 3 buffy coat, 12 urine) tested positive for pathogenic Leptospira spp. and 15 were negative. In addition, 16 environmental and 29 intermediate Leptospira spp. positive samples were included in the panel as negative controls as they should not be detected by the MLST scheme. Testing was performed blinded. A detailed list of pathogenic Leptospira spp. positive samples and origin can be found in the table in S2 Table.
For each sample, 200 μl sample material was used for extraction. For urine samples from Lao PDR 1.5 mL was spun down at 14000 rpm for 15 minutes before it was used for extraction. DNA from bacterial isolates and Lao PDR samples was extracted using the QIAmp DNA Mini Kit (Qiagen, Germany) according to manufacturer‘s instructions. DNA from UK samples (C1-C10) was extracted on the MagNA Pure Compact (Roche, Germany) using the DNA_Bacteria Protocol. These samples and bacterial isolates were eluted once in 50 μL nuclease-free water. Samples from Lao PDR were eluted twice in 50 μl nuclease-free water to reach a final volume of 100 μL. UK samples P1-P25 were extracted on the EZ1 investigator platform (Qiagen, Germany) and eluted in 120 μL.
MLST was performed based on a published scheme targeting seven loci (glmU, pntA, sucA, tpiA, pfkB, mreA, caiB) of seven pathogenic Leptospira species (L. alexanderi, L. borgpetersenii, L. interrogans, L. kirschneri, L. noguchii, L. santarosai, L. weilii) [22]. The protocol was adapted by using the HotStar Taq Master Mix (Qiagen, Germany) in a 20 μl reaction including additional 100 nmol MgCl2 for locus 4 (tpiA) only, 5 pmol of each primer, and 40–60 ng DNA. For clinical samples, 5 μl DNA extract was used. Cycling conditions remained unchanged, except for additional initial 15 minutes incubation at 95°C to activate the enzyme. Further to the published protocol, nested primers were designed for all loci in the original MLST scheme (Table 1) to improve the LoD. Primer sequences were based on multi-sequence alignments of all serovars available in this study. To facilitate downstream sequencing primers were extended with M13 anchor primers.
The nested PCR was performed in 20 μl reaction using 5 pmol of each primer and 2 μl of the first-round PCR product. Cycling conditions were as follows: 10 min at 95°C, 5 cycles of 30 sec at 95°C, 30 sec at 46°C, 30 sec at 72°C. This was followed by 10 cycles with the annealing temperature increasing by 1°C per cycle and 20 cycles with an annealing temperature of 56°C. The final extension period was 7 min at 72°C. To avoid contamination different processes were performed in physically separated rooms. For detection of possible cross-contamination between samples that could occur during transfer of the amplicon from first to second round PCR non-template controls were included in all experiments and handled last. Further, only one sample was opened at a time and stringent cleaning measures were applied after each experiment.
To compare the detection limits serial dilutions of six DNA extracts from Leptospira isolates (Serovars Canicola, Grippotyphosa, Copenhageni, Hardjo, Mini, Pyrogenes) were tested using the original typing scheme and the second round PCR of the modified assay. Initial DNA concentration was 4 ng/μl, corresponding to 800,000 copies of genomic DNA (gDNA) or 8 x 105 organisms (calculations based on the size of the genome of L. interrogans strain Fiocruz L1130 (4.6 Mb); 1 genome is ~5 fg). Serial dilutions were tested from 10−2 to 10−5 and PCR products were visualised on 2% agarose E-gels (Thermo Fisher Scientific, USA). In addition, 15 patient specimens (P1-P15) were tested with the modified assay first and second round PCRs.
PCR products were purified on an automated liquid handling robot (Biomek NXP) using Ampure XP paramagnetic beads (Beckman Coulter, USA). Sanger sequencing was carried out on the Applied Biosystems 3730XL Genetic Analyser (Thermo Fisher Scientific, USA). Sequences were assembled, edited, and trimmed using BioNumerics version 6.1 (Applied Maths NV). Sequence types (ST) were assigned by BioNumerics using allelic profiles in the order glmU-pntA-sucA-tpiA-pfkB-mreA-caiB. The same order was used to concatenate sequences for phylogenetic analysis. All new sequences have been submitted to the leptospira MLST database (http://pubmlst.org/leptospira/).
For species assignment sequences from all patient samples were included in phylogenetic analyses along with isolates from the WHO panel for which the species are known. Sequences were aligned in seaview4 [28] and used to construct maximum likelihood trees in MEGA version 6 [29] using the best suitable and available model for each alignment as determined by jModeltest [30].
The modified scheme allowed for amplification of all pathogenic Leptospira species covered by the scheme and represented in the WHO serovar panel (25/28). Two had new ST assigned (allelic profiles serovar Saxkoebing strain Mus 24: 24-69-30-35-37-26-51 [ST 219]; serovar Shermani strain 1342 K: 57-53-47-49-79-61-43 [ST 220]; data in S1 Table).
Fifteen clinical specimens (P1-P15) were tested using the first-round MLST assay and none gave a positive result. Applying the improved nested MLST assay five of these yielded at least one amplified locus; two samples gave full allelic profiles (P1 and P12).
In total, using the improved nested assay on 44 clinical samples PCR positive for pathogenic Leptospira species, 22 yielded a result in at least one allele detected that could be sequenced (50%). Full allelic profiles were obtained from 10 (23%) specimens, and partial allelic profiles from 12 specimens (27%, Table 2). No nonspecific amplicons were observed in any of the clinical samples. All negative control samples (including those positive for environmental and intermediate Leptospira species) were negative by MLST.
Out of the twelve positive urine specimens, three gave full allelic profiles (25%), and two a partial profile (4 and 5 loci). In total, eleven new alleles were detected and five of the specimens revealed allelic profiles representing new ST. Despite several attempts three samples resulted in ambiguous nucleotides in sequences of two (L29, sucA and caiB) and one (C4 and P8, pfkB) loci. No numbers could be assigned to those alleles. The locus that was amplified most often from clinical samples was caiB (19/44, 43.2%), followed by glmU (18/44, 40.9%) (data in S3 Table).
Using the nested approach it was possible to lower the LoD of the assay. The minimum DNA concentration for simultaneous detection of all loci (42 PCRs) using the nested MLST scheme was 4x10-4 ng, corresponding to 80 copies of genomic DNA (gDNA; S1 Fig). In contrast, after the first round of amplification weak bands were visible for only eight loci (8/42, 19%). When using eight gDNA copies per reaction in the nested assay only two PCRs did not yield a detectable product (strain Hardjoprajitno /pntA and Salinem /pfkB) while no product was detectable using the first round PCR only.
For species assignment sequences from all patient samples were included in the phylogenetic tree along with isolates from the WHO panel for which species are known (S1 Table). A maximum likelihood tree showing all samples for which a full allelic profile could be obtained is shown in Fig 1. Trees based on separate alleles, are in concordance with the full-profile tree (S2 Fig). L. interrogans was the most frequently detected species in 17 samples (17/22, 77%). Table 3 shows the different species detected in each country.
Using the developed nested amplification approach presented in this study it was possible to increase the MLST assay’s analytical sensitivity and obtain typing information of pathogenic Leptospira species directly from a variety of clinical samples. The developed assay is based on an established MLST scheme supported by a public website (http://leptospira.mlst.net/) and it will therefore not negatively impact comparability of already typed leptospires. The simplified PCR setup along with the anchor primers incorporated in the nested assay enables sequencing using two primers for all loci which will reduce costs. No nonspecific amplification was observed in any of the clinical samples. Consequently, in resource-limited settings where quantitative real-time PCR facilities are not available, the assay (or defined loci only) may be a useful diagnostic tool when applied with all necessary precautions to avoid cross-contamination between samples.
Sample numbers in the presented study are too low to make any inferences as to which specimen type is most promising for molecular typing. Success rates between different samples varied between 40–100%. The highest proportion of full allelic profiles was obtained from buffy coat (2/3) and whole blood (3/6), followed by urine (3/12). Due to the dynamics of the disease Leptospira may be found in blood or urine at different time points [8, 31]. Consequently, choice of specimen type and sampling time post symptom onset may prove critical for molecular MLST determination direct from specimens. In addition, as for any PCR based assay, detection is influenced by the genomic sequence of the strain present. Most primers used in the modified typing scheme were degenerated to account for sequence differences between the different strains, leading to variable specificity. Samples used for the present study were extracted using different platforms and elution volumes. However, all extracts were tested using the same diagnostic qPCR method and there does not appear to be a correlation between the original CT values and whether full or partial profiles were obtained (data in S3 Table). Similarly, there was no correlation between sample type or Leptospira species and successfully amplified locus. Interestingly, the locus that performed best in the nested assay (caiB) was the least reliable in a study from Argentina using the unmodified MLST scheme [26]. Overall, using the nested approach the success rate of detecting full or partial profiles could be improved by more than two fold when compared to previous studies applying the original MLST scheme directly on clinical specimens [24, 26].
Typing results of samples from the WHO serovar panel are 100% concordant with previously published results. Of note, the panel does not include an isolate of L. alexanderi and none of the clinical samples turned out as such. Boonsilp et al. (2013) characterized 325 isolates that resolved into 190 different ST and showed that L. alexanderi is detected by the original MLST scheme [22]. All loci represent conserved genes and the nested primers fit a representative sequence of L. alexanderi. It hence can be assumed that the nested assay would detect L. alexanderi, enabling it to detect all pathogenic Leptospira species, as well as ST that could not be tested for in the present study.
Single alleles amplified from clinical specimens allow for species determination when used in phylogeny, opening up the possibility for the assay to support surveillance. Currently, most human leptospirosis cases are not identified to species level, so it is difficult at this point to draw any further conclusions from the presented results. A recent survey conducted in Southeast Asia identified four pathogenic species in native rodents: L. weilii, L. kirschneri, L. interrogans and L borgpetersenii, the latter being the most prevalent [32]. This is consistent with the findings of our study. Similarly, in the UK and Europe, L. interrogans was identified in indigenous rodents [33, 34]. The variety of species found in the UK patients might be attributable to the fact that many cases in the UK are diagnosed in returning travellers. Of the 34 cases diagnosed in the UK, 15 reported a travel history (44%). Of these, 9 (26%) had travelled to South East Asia (Malaysia, Thailand and Indonesia). One case found to be infected with L. weilii had travelled to Thailand and one case infected with L. santarosai reported travel to Central America. The ability to obtain typing data directly from clinical specimens is ideal for pathogens that are difficult and slow to isolate in culture. The use of direct typing on urine specimens allows for non-invasive sampling and in some cases the provision of typing information in the absence of data from blood samples. One patient was positive for pathogenic Leptospira spp. in both serum and buffy coat by qPCR. MLST in this patient yielded a full profile from buffy coat, but only a partial profile (5 loci) from serum. While this is consistent with our finding that success rates for amplifying MLST loci were higher in buffy coat than in serum it has to be interpreted with caution due to low sample numbers.
Despite several attempts one sample resulted in ambiguous nucleotides in two loci (L29) and two samples in one locus (C4 and P8). This could indicate active infection with more than one strain. Another possibility is that more than one copy of the gene is present in the genome, as has been shown for the mompS gene of several Legionella strains [35].
In summary, the reported improved MLST assay represents a fast and specific tool for typing of Leptospira direct from clinical specimens, including non-invasive samples such as urine. It may be of use during epidemics and outbreaks by enabling rapid identification of Leptospira species and MLST types without the inherent delay involved in Leptospira culture.
|
10.1371/journal.pbio.2005840 | In vitro–transcribed guide RNAs trigger an innate immune response via the RIG-I pathway | Clustered, regularly interspaced, short palindromic repeat (CRISPR)–CRISPR-associated 9 (Cas9) genome editing is revolutionizing fundamental research and has great potential for the treatment of many diseases. While editing of immortalized cell lines has become relatively easy, editing of therapeutically relevant primary cells and tissues can remain challenging. One recent advancement is the delivery of a Cas9 protein and an in vitro–transcribed (IVT) guide RNA (gRNA) as a precomplexed ribonucleoprotein (RNP). This approach allows editing of primary cells such as T cells and hematopoietic stem cells, but the consequences beyond genome editing of introducing foreign Cas9 RNPs into mammalian cells are not fully understood. Here, we show that the IVT gRNAs commonly used by many laboratories for RNP editing trigger a potent innate immune response that is similar to canonical immune-stimulating ligands. IVT gRNAs are recognized in the cytosol through the retinoic acid–inducible gene I (RIG-I) pathway but not the melanoma differentiation–associated gene 5 (MDA5) pathway, thereby triggering a type I interferon response. Removal of the 5’-triphosphate from gRNAs ameliorates inflammatory signaling and prevents the loss of viability associated with genome editing in hematopoietic stem cells. The potential for Cas9 RNP editing to induce a potent antiviral response indicates that care must be taken when designing therapeutic strategies to edit primary cells.
| Clustered, regularly interspaced, short palindromic repeat (CRISPR)–CRISPR-associated 9 (Cas9) genome editing is transforming fundamental research, as it allows researchers to make targeted changes to the genome of cells. For efficient editing, the Cas9 protein (a DNA nuclease) and a guide RNA (gRNA), which leads the nuclease to the correct location in the genome, have to be introduced into cells. One recent advancement is the delivery of a Cas9 protein and an in vitro–transcribed (IVT) gRNA as a precomplexed ribonucleoprotein (RNP). This approach allows editing of more sensitive cell types such as immune cells and hematopoietic stem cells. However, the consequences of introducing foreign Cas9 nuclease and gRNA into mammalian cells are not fully understood. Here, we show that in many cell types, the IVT gRNAs trigger a potent innate immune response—a natural defense mechanism against RNA viruses. We show that the innate immune response causes cell death in primary hematopoietic stem cells but that removal of the 5’-triphosphate from gRNAs by phosphatase treatment can ameliorate the immune response and prevent the loss of viability. Hence, CRISPR-Cas9 RNP editing has the potential to induce a potent antiviral response, and we suggest that care must be taken when designing therapeutic strategies to edit primary cells.
| Clustered, regularly interspaced, short palindromic repeat (CRISPR)–CRISPR-associated (Cas) genome editing has rapidly become a widely used tool in molecular biology laboratories. Its ease of use and high flexibility allows researchers to modify and edit genomes in cell lines [1], stem cells [2], animals and plants [3,4], and even human embryos [5]. The Cas protein complexes with a target-specific CRISPR RNA (crRNA) and a trans-activating crRNA (tracrRNA), which keeps the Cas protein catalytically active [6]. In experimental procedures, the two RNAs are often combined to generate a single guide RNA (gRNA), which means that at least two components must be successfully delivered into cells during genome editing: the Cas protein, such as Cas9, and gRNA to direct the Cas9 protein to its target site. For in vitro–cultured cells, this can be done by transfecting plasmids encoding gRNA and Cas9 protein. However, transfection of plasmid DNA into sensitive cell types such as primary and stem cells is challenging and inefficient. The introduction of plasmids can also lead to undesired integration of DNA at the cut site [7], increased off-target activity through prolonged expression of the CRISPR-Cas9 components [8], and a delay in editing while the cell expresses gRNA and Cas protein [9].
The delivery of gRNA and Cas9 protein as a precomplexed ribonucleoprotein (RNP) sidesteps issues related to plasmid expression and has proved to be a successful strategy to edit human primary cells, including T cells [10,11], hematopoietic stem cells [12–15], and neurons [16]. This makes RNP editing a particularly attractive approach for therapeutic applications, but relatively little is known about the nonediting consequences of introducing a foreign gRNA and Cas9 protein. Human cells have evolved multiple defense mechanisms to guard against foreign components, and genome editing reagents have the potential to activate these systems. For example, recent data suggest that humans may have a preexisting adaptive immune response to the Cas9 protein [17]. But cellular responses to the gRNAs used to program Cas9 editing have so far not been well explored.
Cells respond to infection by RNA viruses with an innate immune response that protects the host cell from invading foreign genetic material [18]. Foreign RNAs are recognized by pathogen-associated molecular pattern (PAMP) binding receptors in the cytosol that include retinoic acid–inducible gene I (RIG-I) and melanoma differentiation–associated gene 5 (MDA5) [19]. This triggers a cascade of events mediated by the mitochondrial antiviral signaling (MAVS) protein, resulting in the transcriptional activation of type I interferons and interferon-stimulated genes (ISGs) [20–22]. RNA PAMPs usually contain exposed 5’-triphosphate ends [19], which may also be present in gRNAs made via T7 in vitro transcription [23,24]. Given that Cas9 has a picomolar affinity for targeting gRNA [25], it is not clear that the 5’-triphosphate would be available to stimulate an innate immune response.
We asked whether in vitro-transcribed (IVT) gRNAs complexed with Cas9 cause an innate immune response and here show that introduction of RNPs into cells induces up-regulation of interferon beta (IFNβ) and interferon-stimulated gene 15 (ISG15) in a variety of human cell types. This activity depends upon RIG-I and MAVS but is independent of MDA5. The extent of the immune response depends upon the protospacer sequence, but removal of the 5’-triphosphate from gRNAs avoids stimulation of innate immune signaling. The potential for Cas9 RNP editing to induce an antiviral response indicates that care must be taken when designing therapeutic strategies to edit primary cells.
To investigate if mammalian cells react to IVT gRNA/Cas9 with an innate immune response, we first performed genome editing in human embryonic kidney 293 (HEK293) cells using Cas9 RNPs. To separate innate immune response from genome editing, we performed these experiments with a nontargeting gRNA that recognizes a sequence within blue fluorescent protein (BFP) and has no known targets within the human genome [26]. Constant amounts of recombinant Cas9 protein were complexed with different amounts of nontargeting IVT gRNA, and RNPs were transfected into HEK293 cells using CRISPRMAX lipofection reagent [27]. We harvested cells 30 h after transfection and measured transcript levels of interferon beta 1 (IFNB1) and ISG15 by quantitative real-time PCR (qRT-PCR; Fig 1A). Introduction of gRNAs caused a dramatic increase in both IFNB1 and ISG15 levels, and the presence of Cas9 protein did not have an effect on the outcome. Cas9 on its own did not induce IFNB1 or ISG15 expression. To our surprise, as little as 1 nM of gRNA was sufficient to trigger a 30–50-fold increase in the transcription of innate immune genes. We further found that a commonly administered amount of 50 nM gRNA can induce IFNB1 by 1,000-fold, which is equal to induction by canonical IFNβ inducers such as viral mRNA from Sendai virus [28] or a hepatitis C virus (HCV) PAMP [21,29] (Fig 1B).
RNPs can be delivered into cells via different transfection methods, and while lipofection is cost-effective and easy to use, many researchers prefer electroporation for harder-to-transfect cells. We wondered if the transfection method would affect the IFNβ response and compared gRNA transfection via lipofection (Lipofectamine 2000 and RNAiMAX) to nucleofection (Lonza) (Fig 1C). Lipofection led to a strong increase in IFNB1 and ISG15 transcript levels after as little as 6 h posttransfection, and the response was sustained for up to 48 h. Nucleofection also caused an increase in innate immune signaling at early time points, but the response was milder than in lipofected samples and was greatly diminished by 48 h.
Next, we asked if the innate immune response to gRNAs is a common phenomenon across different cell types and compared IFNβ activation in seven commonly used human cell lines of various lineages: human embryonic kidney cells 293 SV40 large T antigen (HEK293T), HEK293, Henrietta Lacks cells (HeLa), Jurkat, HCT116, HepG2, and K562 (Fig 2A). While the magnitude of induction varied between cell lines, all tested cell lines responded to IVT gRNA transfection with activation of IFNB1 expression. The sole exception was K562 cells, which have a homozygous deletion of the IFNA and IFNB1 genes [30]. We also measured transcript levels of two major cytosolic pathogen recognition receptors, RIG-I (DExD-H-box helicase 58 [DDX58]) and MDA5 (interferon induced with helicase C domain 1 [IFIH1]), and noticed that all cell lines except K562 up-regulated these transcripts in response to introduction of gRNAs. We also confirmed these results on the protein level in HEK293 cells (Fig 2B).
The RIG-I and MDA5 receptors complement each other by recognizing different structures in foreign cytosolic RNAs, but the exact nature of their ligands is not yet fully understood [31,32]. To investigate if IVT gRNAs are recognized via RIG-I or MDA5, we generated clonal knockout (KO) cell lines for RIG-I, MDA5, and their downstream interaction partner MAVS in HEK293 cells using CRISPR-Cas9. As the expressions of both RIG-I and MDA5 are themselves stimulated by IFNβ, we confirmed successful KO after transfection with gRNAs by genomic PCR, Sanger sequencing, and western blot (S1A–S1C Fig). MAVS KO cells were confirmed by western blot (S1D Fig). Strikingly, activation of IFNB1 expression after introduction of gRNAs was absent in RIG-I and MAVS KO cells, while MDA5 KO cells did not show a significant decrease in IFNB1 transcript levels (Fig 2C). This indicates that IVT gRNAs are exclusively recognized through RIG-I to trigger a type I interferon response.
As the structural requirements of RIG-I ligands are still not completely understood, we wondered if different 20-nucleotide protospacers in gRNAs vary in their potency to trigger an innate immune response via RIG-I. We designed 10 additional nontargeting gRNAs that we in vitro transcribed and transfected into HEK293 cells. Surprisingly, we found that the cells responded to different protospacers with a wide range of IFNB1 expression. Several gRNAs produced very little innate immune response, and one gRNA (gRNA11) yielded no IFNB1 activation at all (Fig 3A). We speculated that the differential response may be correlated with the purity of the RNA product after in vitro transcription or the stability of the secondary structure of the RNA [33,34]. However, we found that there was no obvious correlation between the immune response to certain gRNAs and their purity; predicted protospacer secondary structure; full secondary structure, including the constant region; or predicted disruption of the constant region by mispairing with the protospacer (S2 Fig). When we separately nucleofected five of these gRNAs into primary CD34+ human hematopoietic stem and progenitor cells (HSPCs), we found that all gRNAs induced a strong immune response. Only gRNA11, which showed no immune stimulation in HEK293 cells, resulted in half the amount of ISG15 transcript (Fig 3B). These results indicate that RIG-I recognition patterns of IVT gRNAs are complex and difficult to anticipate a priori based on predicted properties of the variable protospacer and cell type.
One well-established structural requirement of RIG-I ligands is the presence of a 5’-triphosphate group [35]. We asked if preparations that remove the 5’ triphosphate might avoid or reduce the innate immune response to IVT gRNAs. We first used a synthetic gRNA that lacks a 5’-triphosphate and verified that this gRNA does not induce IFNB1 expression when transfected into HEK293 cells (Fig 3C). Synthetic gRNAs are becoming more commonplace but are still an order of magnitude more expensive than in vitro transcription of gRNAs. This limits their application for high-throughput interrogation of gene function in primary cells. We therefore asked if treatment of IVT gRNA with phosphatases that remove the 5’-triphosphate would reduce IFNB1 induction. We tested calf intestinal alkaline phosphatase (CIP), shrimp alkaline phosphatase (SAP), 5’-RNA polyphosphatase (PP), and thermosensitive alkaline phosphatase (AP) and found that phosphatase treatment with CIP and AP abolished the IFNB1 response, while SAP and PP treatment only resulted in a reduction of the response (Fig 3C). We also compared purification of IVT gRNAs by solid-phase reversible immobilization (SPRI) beads to column purification and established that SPRI bead cleanup is not sufficient to completely avoid an immune response, even when more phosphatase is used (S3A–S3B Fig). Taken together, these results indicate that 5’-triphosphate is a necessary requirement for gRNA-induced IFNB1 activation through RIG-I but that additional structural properties of the gRNAs also influence the magnitude of the immune response.
Next, we asked if phosphatase treatment alters the genome editing potential of gRNAs. As gRNA1 targets the BFP gene, we used a HEK293T cell line with a stably integrated BFP reporter [26], nucleofected cells with phosphatase-treated gRNA-Cas9 RNPs, and monitored editing outcomes by T7 endonuclease I assay (S3C Fig). We did not observe any significant difference in editing outcomes between synthetic, IVT, and phosphatase-treated gRNAs, suggesting that phosphatase treatment does not affect the function of the gRNA.
When a cell initiates an antiviral immune response, it also undergoes cellular stress that can affect cell viability [36,37]. Hence, we asked if there is a correlation between the IFNβ response and cell viability after transfection with synthetic, IVT, or CIP-treated IVT gRNA. Not surprisingly, the viability of the very robust HEK293 cell line was not affected by the antiviral immune response (S3D Fig). We then turned to HSPCs, which are a much more sensitive cell type. We first nucleofected HSPCs with RNPs targeting the hemoglobin subunit beta (HBB) gene [12] and compared synthetic and IVT gRNA interferon stimulation and cell viability posttransfection. Double-strand breaks (DSBs) have been reported to cause innate immune stimulation and can themselves cause decreases in cell fitness [38,39]. Therefore, we performed controls using nuclease-dead Cas9 (dCas9) to form RNPs and confirmed by Sanger sequencing and TIDE analysis that dCas9-RNPs did not induce DSBs [40] (S3E Fig).
We found a significant decrease in HSPC viability using both of the IVT gRNA RNPs that had an increase in IFN-stimulated genes ISG15 and RIG-I (Fig D-E). We did not see a substantial difference in viability or ISG expression between Cas9 and dCas9 RNPs, suggesting that nuclease activity leading to DNA damage did not cause the immune response. Next, we asked if CIP treatment of gRNAs could reverse the decrease in viability in HSPCs. We nucleofected HSPCs with dCas9 RNPs targeting a noncoding intron of Janus kinase 2 (JAK2) or Cas9 RNPs targeting the HBB gene and compared synthetic, IVT, and CIP-treated IVT gRNAs. Strikingly, CIP treatment restored viability in HSPCs (Fig 3F). We were also interested in editing outcomes in these samples and performed amplicon next-generation sequencing (NGS) for the HBB locus. While the phosphatase-treated gRNA performed similarly to the synthetic gRNA, the IVT gRNA resulted in slightly fewer insertions and deletions (indels) (Fig 3G).
We have found that IVT gRNAs used with Cas9 RNPs for many genome-editing experiments can trigger a strong innate immune response in many mammalian cell types (Fig 4). Lipofection results in a stronger and longer-lasting response than nucleofection, possibly because lipofection delivers gRNAs to the cytosol, while nucleofection delivers mainly to the nucleus. Using isogenic KO clones, we found the gRNA-induced response is mediated via the antiviral RIG-I pathway and results in expression of genes that initiate an antiviral immune response. While introduction of IFN-stimulating gRNAs does not affect viability in HEK293 cells, we found that viability of primary HSPCs is negatively affected by the antiviral immune response. While DSBs have on their own been reported to induce an innate immune response [38], we found triphosphate-containing gRNAs complexed with dCas9 induce an immune response and cell death in HSPCs. Only removal of the triphosphate is sufficient to reduce gRNA-induced innate immune signaling.
These results have several implications. We suggest that the gene signature associated with type I interferon stimulation should be considered when studying the transcriptome of recently edited bulk populations of cells. Furthermore, all mammalian cells can both produce type I interferons and also respond to them through the ubiquitously expressed receptor interferon alpha and beta receptor subunit 1 (IFNAR1) [41]. Even cells that have not been successfully transfected with RNPs could sense the IFNβ produced by neighboring cells and activate downstream antiviral defense mechanisms. This could be an important consideration during in vivo genome editing applications, as RNP delivery into one set of cells could provoke a widespread innate immune response in the surrounding tissues.
We found that synthetic gRNAs completely circumvent the RIG-I mediated response, offering a valuable path to avoid innate immune signaling during therapeutic editing. However, synthetic gRNAs can become expensive when performing experiments that require testing or using many gRNAs. We found that a cost-effective phosphatase treatment to remove the 5’-triphosphate before transfection reduces the immune response and increases posttransfection viability in HSPCs. Furthermore, editing outcomes in cell lines with phosphatase-treated gRNA were comparable to those of IVT gRNAs, suggesting that removal of 5’-phosphate groups does not abolish gRNA function. In fact, in sensitive HSPCs, phosphatase-treated gRNA slightly outperformed IVT gRNA, which is possibly due to reduced viability in samples transfected with IVT-RNPs. Thus, consideration of a potential innate immune stimulation prior to choice of genome editing reagents, study design, and implementation of controls is critical when performing genome editing using RNPs in mammalian cells.
While we were preparing this manuscript for submission, the Kim group reported similar results in HeLa cells and primary human CD4+ T cells [42]. They confirmed that the type I interferon response is dependent on the presence of a 5’-triphopsphate group and that CIP treatment can increase viability by avoiding the antiviral response. These results are very much in alignment with our findings and extend the potential problem of innate immune signaling to additional cell types.
Our study adds extra depth by further outlining the mechanisms by which gRNAs are sensed. We show that gRNA sensing depends upon RIG-I and MAVS, but MDA5 KO cells are fully capable of inducing IFNβ after IVT gRNA transfection. Hence, gRNA sensing is independent of the MDA5 PAMP receptor, consistent with RIG-I’s preference for short double-stranded RNA (dsRNA) structures and MDA5’s preference for long dsRNA fragments [43]. Furthermore, we show that in addition to a 5’-triphosphate, the protospacer sequence is also critical to determine the intensity of the IFNβ response. Not only do different gRNAs induce different innate immune responses, but some gRNAs induce no response at all. However, this seems to be cell-type specific, as we found that sensitive cells such as primary HSPCs react to the same gRNAs with a strong immune response independently of the protospacer. It has been proposed that 5’-base-paired RNA structures are required to activate antiviral signaling via RIG-I, but we found no correlation between signaling and a variety of predicted RNA properties, including secondary structure [33]. Our results therefore suggest that the mechanism of gRNA sensing by the RIG-I pathway is relatively complex in that it requires 5’-triphosphates but that this moiety is not sufficient to induce the response. Additionally, we have not ruled out the possibility that gRNAs could be recognized by Toll-like receptors (TLRs), though we and others [42] have found that KO of RIG-I is sufficient to completely abrogate gRNA-induced signaling in multiple cell contexts. The role of TLR recognition could be addressed in future work to delineate the full set of molecular features responsible for gRNA activation of innate immunity, which might yield accurate predictors of innate immune signaling in general.
gRNA was synthesized by assembly PCR and in vitro transcription as previously described [12]. Briefly, a T7 RNA polymerase substrate template was assembled by PCR from a variable 58–59 nt primer containing T7 promoter, variable gRNA guide sequence, the first 15 nt of the nonvariable region of the gRNA (T7FwdVar primers, 10 nM, S1 and S2 Tables for gRNA sequences), and an 83 nt primer containing the reverse complement of the invariant region of the gRNA (T7RevLong, 10 nM), along with amplification primers (T7FwdAmp, T7RevAmp, 200 nM each). The two long primers anneal in the first cycle of PCR and are then amplified in subsequent cycles. Phusion high-fidelity DNA polymerase was used for assembly (New England Biolabs). Assembled template was used without purification as a substrate for in vitro transcription by T7 RNA polymerase, using the HiScribe T7 High Yield RNA Synthesis kit (New England Biolabs) following the manufacturer’s instructions. Resulting transcription reactions were treated with DNAse I (New England Biolabs), and RNA was purified by treatment with a 5X volume of homemade SPRI beads (comparable to Beckman-Coulter AMPure beads) and elution in RNAse-free water.
gRNAs were treated with phosphatases as follows: CIP (New England Biolabs, 30 U), SAP (New England Biolabs 10 U), PP (Lucigen, 20 U), and FastAP AP (Thermo Fisher Scientific, 10 U) were added per 20 μl in vitro transcription reaction, and samples were incubated at 37°C for 3 h before proceeding to purification and DNAseI treatment. gRNA was purified using a Qiagen RNeasy Mini Kit (Qiagen) or by 5X volume of homemade SPRI beads (comparable to Beckman-Coulter AMPure beads). The detailed protocol and additional notes are available online (dx.doi.org/10.17504/protocols.io.nghdbt6).
HCV PAMP in vitro transcription template [21] was generated by annealing HCV fwd and rev (5 μM each) oligos (S1 Table). In the subsequent in vitro transcription reaction, 2 μl of the annealed product was used as DNA template, using HiScribe T7 High Yield RNA Synthesis kit (New England Biolabs).
The plasmid containing the SeV DI RNA[28] was a gift from Prof. Peter Palese, Icahn School of Medicine at Mount Sinai, New York. Plasmid was digested with HindII/EcoRI before in vitro transcription with HiScribe T7 High Yield RNA Synthesis kit (New England Biolabs). The sequence of the IVT DI, including the T7 promoter, hepatitis delta virus ribozyme, and the T7 terminator, is TAATACGACTCACTATAACCAGACAAGAGTTTAAGAGATATGTATCCTTTTAAATTTTCTTGTCTTCTTGTAAGTTTTTCTTACTATTGTCATATGGATAAGTCCAAGACTTCCAGGTACCGCGGAGCTTCGATCGTTCTGCACGATAGGGACTAATTATTACGAGCTGTCATATGGCTCGATATCACCCAGTGATCCATCATCAATCACGGTCGTGTATTCATTTTGCCTGGCCCCGAACATCTTGACTGCCCCTAAAATCTTCATCAAAATCTTTATTTCTTTGGTGAGGAATCTATACGTTATACTATGTATAATATCCTCAAACCTGTCTAATAAAGTTTTTGTGATAACCCTCAGGTTCCTGATTTCACGGGATGATAATGAAACTATTCCCAATTGAAGTCTTGCTTCAAACTTCTGGTCAGGGAATGACCCAGTTACCAATCTTGTGGACATAGATAAAGATAGTCTTGGACTTATCCATATGACAATAGTAAGAAAAACTTACAAGAAGACAAGAAAATTTAAAAGGATACATATCTCTTAAACTCTTGTCTGGTGGCCGGCATGGTCCCAGCCTCCTCGCTGGCGCCGGCTGGGCAACATTCCGAGGGGACCGTCCCCTCGGTAATGGCGAATAGCATAACCCCTTGGGGCCTCTAAACGGGTCTTGAGGGGTTTTTTG.
The sequence of the SeV DI is highlighted in boldface.
Both HCV PAMP and SeV DI RNA were purified by treatment with a 5X volume of homemade SPRI beads (comparable to Beckman-Coulter AMPure beads) and elution in RNAse-free water.
Chemically synthesized gRNAs, which were purified using high-performance liquid chromatography (HPLC), were purchased from Synthego.
IVT gRNAs were analyzed using a Bioanalyzer. This was performed by the UC Berkeley Functional Genomics Laboratory (FGL) core facility. gRNAs were denatured for 5 min at 70°C before analysis on bioanalyzer.
The Cas9 construct (pMJ915) contained an N-terminal hexahistidine-maltose binding protein (His6-MBP) tag, followed by a peptide sequence containing a tobacco etch virus (TEV) protease cleavage site. The protein was expressed in Escherichia coli strain BL21 Rosetta 2 (DE3; EMD Biosciences) grown in TB medium at 16°C for 16 h following induction with 0.5 mM IPTG. The Cas9 protein was purified by a combination of affinity, ion exchange, and size exclusion chromatographic steps. Briefly, cells were lysed in 20 mM HEPES pH 7.5, 1 M KCl, 10 mM imidazole, 1 mM TCEP, 10% glycerol (supplemented with protease inhibitor cocktail [Roche]) in a homogenizer (Avestin). Clarified lysate was bound to Ni-NTA agarose (Qiagen). The resin was washed extensively with lysis buffer, and the bound protein was eluted in 20 mM HEPES pH 7.5, 100 mM KCl, 300 mM imidazole, 1 mM TCEP, 10% glycerol. The His6-MBP affinity tag was removed by cleavage with TEV protease, while the protein was dialyzed overnight against 20 mM HEPES pH 7.5, 300 mM KCl, 1 mM TCEP, 10% glycerol. The cleaved Cas9 protein was separated from the fusion tag by purification on a 5 ml SP Sepharose HiTrap column (GE Life Sciences), eluting with a linear gradient of 100 mM–1 M KCl. The protein was further purified by size exclusion chromatography on a Superdex 200 16/60 column in 20 mM HEPES pH 7.5, 150 mM KCl, and 1 mM TCEP. Eluted protein was concentrated to 40 uM, flash-frozen in liquid nitrogen, and stored at −80°C.
Cells were obtained from ATCC and verified mycoplasma-free (Mycoalert LT-07, Lonza). HEK293, HEK293T, HCT116, HepG2, and HeLa cells were maintained in DMEM supplemented with 10% FBS and 100 μg/mL penicillin-streptomycin (all Gibco). K562 and Jurkat cells were maintained in RPMI supplemented with 10% FBS and 100 μg/mL penicillin-streptomycin.
All transfections in cell lines were performed in 12-well cell culture dishes using 2 × 105 cells per transfection. For lipofection, we used Lipofectamine CRISPRMAX-Cas9, Lipofectamine RNAiMAX, or Lipofectamine 2000 Transfection Reagent (all Invitrogen) in reverse transfections according to the manufacturer’s protocols. Unless stated otherwise, 2 × 105 cells were transfected with 50 pmol of RNA to a final concentration of 50 nM and harvested 24–30 h posttransfection for RNA extraction.
HSPCs from mobilized peripheral blood (Allcells) were thawed and cultured in StemSpan SFEM medium (StemCell Technologies) supplemented with StemSpan CC110 cocktail (StemCell Technologies) for 48 h before nucleofection with dCas9 or Cas9 RNP (75 pmol of dCas9, 75 pmol of gRNA). Then, 1.5 × 105 HSPCs were pelleted (100 × g, 10 min) and resuspended in 20 μl Lonza P3 solution, mixed with 10 μl dCas9 or Cas9 RNP, and nucleofected using ER100 protocol in Lonza 4D nucleofector. Viability of the cells was measured 24 h postnucleofection using trypan blue exclusion test. RNA was harvested 16 h postnucleofection.
Cell cultures were washed with PBS prior to RNA extraction. Total RNA was extracted using RNeasy Miniprep columns (Qiagen) according to the manufacturer’s instructions, including the on-column DNAseI treatment (Qiagen). One μg of total RNA was used for subsequent cDNA synthesis using Reverse Transcription Supermix (Biorad). For qRT-PCR reactions, a total of 20 ng of cDNA was used as a template and combined with primers (see S3 Table), and EvaGreen Supermix (Biorad) and amplicons were generated using standard PCR amplification protocols for 40 cycles on a StepOnePlus Real-Time PCR system (Applied Biosystems). Ct values for each target gene were normalized against Ct values obtained for GAPDH to account for differences in loading (ΔCt). To determine “fold activation” of genes, ΔCt values for target genes were then normalized against ΔCt values for the same target gene for mock-treated cells (ΔΔCt).
For CRISPR/Cas9 genome editing, we used a plasmid encoding both the Cas9 protein and the gRNA. pSpCas9(BB)-2A-GFP (px458) was a gift from Feng Zhang (Addgene plasmid #48138). We designed gRNA sequences using the free CRISPR KO design online tool from Synthego. Two different gRNA sequences were designed for RIG-I and MDA5, respectively (see S3 Table).
Using a Lonza 4D nucleofector (Lonza) with the manufacturer’s recommended settings, 2 × 105 HEK293 cells were nucleofected with 2 μg of px458 plasmids containing both targeting gRNAs in a 1:1 ratio. After 48 h, cells were harvested and subjected to fluorescence-activated cell sorting (FACS). Cells expressing high levels of GFP were single-cell sorted into 96-well plates to establish clonal populations.
For the screening process, genomic DNA (gDNA) from clonal populations was extracted using QuickExtract solution (Lucigen). For KO of RIG-I and MDA5, we screened clones by genomic PCR, looking for a PCR product that is significantly smaller in size than that of WT HEK293 cells (see S4 Table for primers). PCR products were then Sanger sequenced by the UC Berkeley DNA Sequencing facility using the forward primers of the PCR reaction as sequencing primers.
Cells were harvested and washed with PBS. Cells were lysed in 1x RIPA buffer (EMD Millipore) for 10 min on ice. Samples were spun down at 14,000 × g for 15 min, and protein lysates were transferred to a new tube. Fifty μg of total protein was separated for size by SDS-PAGE and transferred to a nitrocellulose membrane. Blots were blocked in 4% skim milk in 50 mm Tris-
HCl (pH 7.4), 150 mm NaCl, and 0.05% Tween 20 (TBST) and then probed for RIG-I, MDA5, MAVS, or GAPDH protein using antibodies against RIG-I (D14G6), MDA5 (D74E4), MAVS (D5A9E), or GAPDH (14C10), respectively (all Cell Signaling Technologies). This was followed by incubation with secondary antibody IRDye 800CW Donkey anti-Rabbit IgG (Li-Cor). Protein standards (GE Healthcare) were loaded in each gel for size estimation. Blots were visualized using a Li-Cor Odyssey Clx (Li-Cor).
Cells were harvested 24 h after transfection and washed with PBS. gDNA was extracted using QuickExtract solution (Lucigen) following the manufacturer’s protocol. PCR across the target site in the BFP gene was run using the BFP amplicon primer set (S4 Table). Two hundred ng of PCR product was heated to 100°C and slowly cooled down to let DNA reanneal. Annealed DNA was digested with T7 endonuclease I (NEB) for 20 min at 37°C. DNA was then analyzed by agarose gel electrophoresis.
PCR products were generated with target-specific HBB primer set 1, sequenced, and Sanger traces were then analyzed with the TIDE webtool (http://tide.nki.nl).
Using primer set 1, 50–100 ng of gDNA from edited CD34+ cells was amplified at HBB sites (S4 Table). The PCR products were SPRI cleaned, followed by amplification of 20–50 ng of the first PCR product in a second 12-cycle PCR using primer set 2 (S4 Table). Then, the second PCR products were SPRI cleaned, followed by amplification of 20–50 ng of the second PCR product in a third 9 cycle PCR using illumina-compatible primers (primers designed and purchased through the Vincent J. Coates Genomics Sequencing Laboratory [GSL] at University of California, Berkeley), generating indexed amplicons of an appropriate length for NGS. Libraries from 100–500 pools of edited cells were pooled and submitted to the GSL for paired-end 300 cycle processing using a version 3 Illumina MiSeq sequencing kit (Illumina, San Diego, CA) after quantitative PCR measurement to determine molarity.
Samples were deep sequenced on an Illumina MiSeq at 300 bp paired-end reads to a depth of at least 10,000 reads. A modified version of CRISPResso [44] was used to analyze editing outcomes. Briefly, reads were adapter trimmed and then joined before performing a global alignment between reads and the reference sequence using NEEDLE [45]. Indel rates were calculated as any reads in which an insertion or deletion overlaps the cut site or occurs within 3 base pairs of either side of the cut site, divided by the total number of reads.
|
10.1371/journal.pbio.1001083 | Sonic Hedgehog Dependent Phosphorylation by CK1α and GRK2 Is Required for Ciliary Accumulation and Activation of Smoothened | Hedgehog (Hh) signaling regulates embryonic development and adult tissue homeostasis through the GPCR-like protein Smoothened (Smo), but how vertebrate Smo is activated remains poorly understood. In Drosophila, Hh dependent phosphorylation activates Smo. Whether this is also the case in vertebrates is unclear, owing to the marked sequence divergence between vertebrate and Drosophila Smo (dSmo) and the involvement of primary cilia in vertebrate Hh signaling. Here we demonstrate that mammalian Smo (mSmo) is activated through multi-site phosphorylation of its carboxyl-terminal tail by CK1α and GRK2. Phosphorylation of mSmo induces its active conformation and simultaneously promotes its ciliary accumulation. We demonstrate that graded Hh signals induce increasing levels of mSmo phosphorylation that fine-tune its ciliary localization, conformation, and activity. We show that mSmo phosphorylation is induced by its agonists and oncogenic mutations but is blocked by its antagonist cyclopamine, and efficient mSmo phosphorylation depends on the kinesin-II ciliary motor. Furthermore, we provide evidence that Hh signaling recruits CK1α to initiate mSmo phosphorylation, and phosphorylation further increases the binding of CK1α and GRK2 to mSmo, forming a positive feedback loop that amplifies and/or sustains mSmo phosphorylation. Hence, despite divergence in their primary sequences and their subcellular trafficking, mSmo and dSmo employ analogous mechanisms for their activation.
| Hedgehog (Hh) signaling governs embryonic development and adult homeostasis in species ranging from Drosophila to human, and its malfunction has been implicated in a wide range of human disorders. Hh signal is received by the twelve-transmembrane receptor Patched and transmitted intracellularly by the seven-transmembrane protein Smoothened (Smo). How vertebrate Smo is activated in order to transmit the Hh signal remains poorly understood. Here we investigate the molecular mechanism of mammalian Smo (mSmo) activation and find it is similar to that described for Drosophila Smo despite the marked sequence divergence between them. We show that mSmo is activated via phosphorylation at multiple sites by the serine/threonine kinases CK1α and GRK2. We provide evidence that Sonic hedgehog (Shh; the best studied of the three mammalian pathway ligands) can regulate the accessibility of mSmo to these kinases and that phosphorylation promotes the ciliary accumulation of this transmembrane protein in its active conformation. Moreover, increasing concentrations of Shh induce a progressive increase in mSmo phosphorylation that fine-tunes mSmo activity. Thus, our results provide novel insights into the biochemical mechanism of vertebrate Hh signal transduction and reveal a conserved mode of Smo activation.
| The Hh family of secreted proteins plays pivotal roles during embryonic development and adult tissue homeostasis [1]–[3]. Aberrant Hh signaling contributes to numerous human disorders including congenital diseases and cancers [4],[5]. In a number of developmental contexts, Hh functions as a morphogen that specifies distinct cell fates in a concentration-dependent manner [1],[2]. For example, in vertebrate neural tube patterning, Shh secreted by the notochord and floor pate forms a ventral to dorsal concentration gradient that specifies distinct pools of neural progenitor cells [6].
Hh exerts its biological function through a signaling cascade that ultimately controls a balance between activator and repressor forms of the Gli family of transcription factors [2]. In the absence of Hh, Gli2 and Gli3 are processed into truncated repressor forms (GliR). Hh signaling blocks Gli processing and converts full-length Gli2/3 into activator forms (GliA). The reception system for the Hh signal consists of a twelve-transmembrane protein Patched (Ptc) as the Hh receptor and a seven-transmembrane protein Smoothened (Smo) as the obligatory Hh signal transducer [2],[3]. Ptc inhibits Smo substoichiometrically through a poorly defined mechanism in the absence of Hh [7]. Binding of Hh to Ptc and the Ihog/Cdo family of proteins alleviates Ptc inhibition of Smo [8]–[14], leading to Smo activation and signal transduction. How Smo is activated and how it transduces the Hh signal to regulate GliR and GliA are still poorly understood.
In mammals, Hh signaling depends on the primary cilium, a microtubule-based membrane protrusion found in almost all mammalian cells [15]. Key components in the Hh pathway are found in cilia and exhibit dynamic patterns depending on the Hh signaling state. For example, in the absence of Hh, Ptc localizes to cilia and prevents Smo from accumulating in the cilia; binding of Hh to Ptc triggers reciprocal trafficking of Ptc and Smo, with Ptc moving out of and Smo accumulating in the cilia [16],[17]. Ciliary accumulation of Smo correlates but is not sufficient for Hh pathway activation [16]–[19]. Additional mechanisms, including conformational change in Smo, are likely to be responsible for Smo activation [20]–[22]. Indeed, fluorescence resonance energy transfer (FRET) analysis indicates that both Drosophila and mammalian Smo proteins exist as constitutive dimers/oligomers, but in the absence of Hh, Smo C-tails adopt a closed conformation that prevents their association. Hh induces a conformational switch in Smo, leading to dimerization/oligomerization of the C-tails [22]. The mechanisms underlying mammalian Smo ciliary accumulation, conformational change, and activation are largely unknown.
In Drosophila, Hh and Ptc reciprocally control Smo cell surface accumulation and conformation through regulating Smo phosphorylation [22]–[25]. In response to Hh, Smo is phosphorylated by protein kinase A (PKA) and casein kinase 1 (CK1) at multiple sites in its C-tail, and these phosphorylation events activate Smo by promoting its cell surface accumulation and active conformation [22],[25]–[27]. However, vertebrate Smo C-tails diverge significantly from that of Drosophila Smo and do not contain the PKA/CK1 phosphorylation clusters found in Drosophila Smo C-tail [22]. In addition, a systematic mutagenesis study did not reveal any Ser/Thr residues as essential for mammalian Smo activation [28]. These and other observations led to a proposal that mammalian Smo and Drosophila Smo are regulated by fundamentally distinct mechanisms [28],[29].
Several studies suggested that G protein coupled receptor kinase 2 (GRK2) positively regulates Shh signaling [30]–[32]. Metabolic labeling experiments revealed that GRK2 is required for the basal phosphorylation of an exogenously expressed Smo [30]. However, it is not clear whether GRK2 directly phosphorylates Smo and how GRK2 activates Shh signaling. In addition, direct evidence for Hh-induced mammalian Smo phosphorylation is lacking. A recent kinome siRNA screen identified CK1α as a positive regulator for Shh signaling, but its mechanism of action remains unknown [33].
In this study, we investigate the activation mechanism of mammalian Smo (henceforth referred to simply as Smo). We demonstrate that Smo is activated via multiple phosphorylation events mediated by CK1α and GRK2 that induce its ciliary accumulation and active conformation. We provide evidence that graded Shh signals induce increasing levels of Smo phosphorylation that fine-tune Smo ciliary localization, conformation, and activity. In addition, we provide evidence that Shh promotes Smo phosphorylation by regulating the accessibility of Smo to its kinases.
A previous study revealed that CK1α siRNA blocked Shh pathway activation in C3H10T1/2 cells [33]. To determine how CK1α positively regulates Shh signaling, we tested whether CK1α activates Smo. Coexpression of CK1α with Smo in NIH3T3 cells activated a Gli-luciferase (Gli-luc) reporter gene, although the fold of activation was less dramatic compared with Shh stimulation (Figure 1A). In line with a previous finding [31], coexpression of GRK2 with Smo also activated Gli-luc in NIH3T3 cells (Figure 1A). Coexpression of CK1α and GRK2 with Smo had a slightly stronger effect on Gli-luc expression than overexpression of each kinase alone (Figure 1A). Overexpression of another GRK family member (GRK5) with Smo activated the Gli-luc reporter gene similarly to GRK2 (Figure 1A), indicating that overexpressed GRK5 and GRK2 have a similar activity in the Shh pathway.
We also examined the effect of CK1α/GRK2 overexpression on Gli-luc expression in the absence of exogenously expressed Smo. Consistent with previous findings [31],[33], overexpression of CK1α, GRK2, or both only slightly increased the expression of Gli-luc reporter gene (Figure S1G). Thus, CK1α/GRK2 overexpression synergized with Smo overexpression to drive Shh pathway activation.
Our previous FRET analysis indicated that Shh induces a conformational change in Smo from a closed to an open conformation [22]. In the closed conformation, Smo exists as a dimer/oligomer through an N-terminal interaction(s), which results in high basal FRET between CFP and YFP fused to the N-termini of two Smo molecules (FRETN); however, Smo C-tail folds back and is in close proximity to the intracellular loops, resulting in high intramolecular FRET between CFP inserted in the second intracellular loop (L2) and YFP fused to the C-terminus (FRETL2C) and low intermolecular FRET between CFP and YFP fused to the C-termini of two Smo molecules (FRETC) (Figure 1B–D) [22]. Shh induced a marked decrease in FRETL2C and a concomitant increase in FRETC without affecting FRETN (Figure 1B–D) [22], suggesting that Smo C-tails move away from the intracellular loops and form dimers/oligomers. To determine whether CK1α and GRK regulate Smo conformation, we carried out FRET analysis using the Smo biosensors indicated in Figure 1B–D. We found that overexpression of CK1α, GRK2, or GRK5 resulted in a significant increase in FRETC (Figure 1B) and a marked decrease in FRETL2C (Figure 1C). In contrast, overexpression of these kinases did not cause a significant change in FRETN (Figure 1D). These results suggest that excessive CK1α and GRK2/5 kinase activities induce a conformational change in Smo similar to that induced by Shh stimulation.
Having established that CK1α and GRK2 act upstream of Smo, we then determined whether CK1α and GRK2 could promote Smo phosphorylation using a Phos-tag gel that specifically retards phosphorylated proteins [34]. We found that coexpression of CK1α, GRK2, or both with a Myc-tagged Smo (Smo-Myc) resulted in a clear mobility shift of Smo-Myc on Phos-tag PAGE that was abolished by phosphatase treatment (Figure 1E, lanes 5–10), suggesting that CK1α and GRK2 can promote Smo phosphorylation.
We next determined whether Shh normally induces Smo phosphorylation and whether it does so through CK1α and GRK2. Treating Smo-Myc transfected cells with a Shh-conditioned medium but not a control medium induced a marked mobility shift of Smo-Myc that was abolished by phosphatase treatment (Figure 1E, lanes 3–4). Importantly, Shh-induced Smo-Myc mobility shift was greatly reduced by treating cells with a CK1 inhibitor CKI-7 [35] and/or a GRK inhibitor heparin [36] (Figure 1F), suggesting that Shh induces Smo phosphorylation through CK1 and GRK kinase activities.
To establish that CK1α and GRK2 are required for Shh-induced Smo phosphorylation, we generated cell lines stably expressing shRNA targeting CK1α, GRK2, or GRK5. Two independent shRNA constructs that effectively and selectively knocked down the targeted kinase were employed in our assay (Figure S1A–B). In line with previous findings [31]–[33], CK1α or GRK2 shRNA inhibited Shh pathway activity in the Gli-luc reporter assay (Figure 1G, Figure S1C–D, H). In contrast, GRK5 shRNA did not alter Shh-induced Gli-luc expression (Figure 1G, Figure S1E). Importantly, CK1α and/or GRK2 shRNA but not GRK5 shRNA reduced Shh-induced mobility shift of Smo-Myc (Figure 1H, Figure S1F), suggesting that CK1α and GRK2 are required for Shh-induced Smo phosphorylation. We note that Shh-induced Smo mobility shift was not completely abolished by silencing CK1α and GRK2, likely due to an incomplete elimination of these kinase activities by the RNAi approach (Figure S1B). However, it is also possible that the residual Smo-Myc phosphorylation in the presence of CK1α and GRK2 shRNA could be due to the involvement of another kinase(s).
To determine whether CK1 and GRK directly phosphorylate Smo, we developed an in vitro kinase assay in which purified GST-fusion proteins containing different regions of Smo C-tail were incubated with a recombinant CK1 (CK1δ from New England Biolabs) or GRK (GRK5 from Cell Signaling Technology) in the presence of γ32-p-ATP. Two non-overlapping fragments, amino acid (aa) 608–670 and aa 770–793, were phosphorylated by both CK1 and GRK (Figure 2A, lanes 3, 5; Figure 2B, lanes 3, 5), suggesting that they harbor CK1 and GRK sites. GRK family kinases tend to phosphorylate S/T in an acidic environment [37]. aa 608–670 contains three sequences (EPS615ADVS619S620A, QDVS642VT, and EIS666PELE) and aa 770–793 contains one sequence (DADS791DF) that match GRK consensus sites (Figure 2C). Indeed, mutating S615, S619 and S620 (SA1), S642 (SA2), or S666 (SA3) reduced and their combined mutations (SA123) abolished phosphorylation of aa 608–670 (Figure 2B, lanes 8–12; Figure S2A and S2C, lanes 3–9), whereas mutating S791 abolished phosphorylation of aa 770–793 by GRK (Figure 2B, lane 14), suggesting that these Ser residues are GRK sites.
CK1 phosphorylation sites conform to the consensus: D/E/S/T(P)X1–3S/T [38]. Site-directed mutagenesis revealed that S615, S619, and S620 mediated CK1 phosphorylation of aa 608–670 (Figure 2A, lanes 8–12; Figure S2A and Figure S2B, lanes 3–9), whereas S774, S777, and S791 mediated CK1 phosphorylation of aa 770–793 (Figure 2A, lanes 14–16). aa 608–670SA1, which has S615, S619, and S620 mutated to Ala but contains intact S642 and S666, was not phosphorylated by CK1 (Figure 2A, lane 8), suggesting that S642 and S666 are not CK1 sites. In addition, we found that aa 581–612 was phosphorylated by CK1 but not by GRK (Figure 2A, lane 18; unpublished data). This region contains a sequence matching CK1 consensus sites: ELS592FS594MHT597VS599. Indeed, mutating S592, S594, T597, and S599 to Ala (aa 581–612SA0) abolished CK1 phosphorylation of aa 581–612 (Figure 2A, lane 19).
For simplicity, we referred to S592, S594, T597, and S599 collectively as S0; S615, S619, and S620 as S1; S642 as S2; S666 as S3; S774 and S777 as S4; and S791 as S5 (Figure 2C). Thus, S1 and S5 are phosphorylation sites for both CK1 and GRK, whereas S0/S4 and S2/S3 are selectively phosphorylated by CK1 and GRK, respectively. Sequence alignment indicates that these phosphorylation sites are conserved among vertebrate Smo proteins (Figure 2C).
To determine if the CK1/GRK sites identified in vitro mediate Shh-induced Smo phosphorylation in vivo, we mutated S0–S5 to Ala in Smo-Myc (SA0–5, Figure 2C). We found that the SA0–5 mutation abolished Shh, CK1α, or GRK2-induced Smo mobility shift (Figure 2D, lanes 5–12). Furthermore, CK1α and GRK2 neither activated SA0–5 nor induced its conformational change (Figure S2D–E).
To further characterize Smo phosphorylation in vivo, we attempted to generate phospho-specific antibodies against phosphorylated CK1/GRK sites and succeeded in obtaining an antibody (PS1) that specifically recognizes phosphorylated S1 (pS615, pS619, and pS620, Figure S2F). To monitor phosphorylation at S1, NIH3T3 cells were transfected with Smo-Myc and stimulated with or without Shh-conditioned medium. In the absence of Shh, Smo-Myc exhibited a weak PS1 signal likely due to basal phosphorylation (Figure 2E, lane 1). Shh induced a clear increase in the intensity of the PS1 signal (Figure 2E, lane 3). Coexpression of CK1α, GRK2, or both also increased the PS1 signal (Figure 2E, lanes 5, 7, and 9). On the other hand, Shh-stimulated S1 phosphorylation was diminished by treating cells with the CK1 and/or GRK2 kinase inhibitors (Figure 2E, lanes 11–13). Furthermore, CK1α or GRK2 shRNA reduced whereas combined CK1α/GRK2 shRNA nearly abolished S1 phosphorylation (Figure 2F, lanes 4, 6, and 8; Figure S2G). In contrast, GRK5 shRNA did not affect S1 phosphorylation (Figure 2F, lane 10; Figure S2G), consistent with its lack of effect on Shh pathway activity. These results demonstrate that Shh induces S1 phosphorylation by CK1α and GRK2.
Hh signaling strength depends on the level of Hh ligand [2]. To determine if the level of Shh pathway activity correlates with the level of Smo phosphorylation, NIH3T3 cells transfected with Smo-Myc were treated with different levels of Shh, followed by mobility shift assay on the phospho-tag gel or western blot with PS1. We found that increasing levels of Shh induced a progressive increase in the degree of Smo-Myc mobility shift (Figure 2G), suggesting that Smo-Myc was phosphorylated at more sites in response to higher levels of Shh. In addition, we found that increasing levels of Shh resulted in a gradual increase in the PS1 signal (Figure 2G), suggesting that the frequency of S1 phosphorylation increases with increasing Shh concentration.
Several oncogenic mutations in human Smo have been identified, including M1 and M2 [39]. The M2 mutation occurs in the seventh transmembrane domain whose murine counterpart is the A1 mutation [20],[39]. Previous studies suggest that SmoA1 exhibits constitutive activity, accumulates at primary cilia, and adopts an open conformation [16],[20],[22]. We found that SmoA1 exhibited slower mobility and elevated PS1 signal intensity regardless of Shh treatment (Figure 2H, lanes 4–5; Figure S2H) and that A1-induced PS1 signal and mobility shift were abolished by the S1–5 mutation (A1SA1–5) (Figure 2H, lanes 7–8; Figure S2H), suggesting that the oncogenic mutation mimics Shh stimulation to induce Smo phosphorylation at CK1/GRK sites. We also observed that M1 increased Smo phosphorylation (see below).
Previous studies demonstrated that small molecules including SAG and 20α-hydroxycholesterol (20-OHC) promote whereas cyclopamine blocks Smo activation [20],[21],[40],[41]. We found that SAG and 20-OHC induced whereas cyclopamine blocked Smo phosphorylation at CK1/GRK sites (Figure 2I), suggesting that these small molecules regulate Shh signaling at the level of Smo phosphorylation.
To determine the functional significance of Smo phosphorylation, CK1/GRK sites were mutated to Ala individually or in different combinations (referred to as SA mutation; Figure 2C), and the effect of SA mutations on Smo activity was determined by the Gli-luc reporter assay in smo−/− MEFs. Mutating S0 (SA0) or S1 (SA1) slightly reduced whereas their combined mutations (S01) markedly inhibited Shh-induced Smo activity (Figure 3A). While mutating S2 to S5 either individually (SA2, SA3, SA5) or in combinations (SA23, SA45) had little if any effect on Smo activation (Figure 3A), combined mutations of these sites with S1/S0 (SA12, SA13, SA123, SA0–3, SA1–5, SA0–5) resulted in a progressive decrease in Shh-induced Smo activity (Figure 3A). Finally, simultaneously mutating all CK1/GRK sites (SA0–5) completely abolished Shh-induced Smo activation. These results suggest: 1) phosphorylation at CK1/GRK sites is essential for Smo activation; 2) S0 and S1 are the major sites while S2 to S5 may play a fine-tuning role; and 3) the level of Smo activity correlates with its level of phosphorylation.
To determine whether phosphorylation renders constitutive Smo activity, we converted individual or different combinations of CK1/GRK sites to Asp (referred to as SD mutations) to mimic different levels of phosphorylation (Figure 2C). SD mutations of individual sites (SD0, SD1, SD2, SD3) or several combinations (SD23, SD12, SD13) caused little if any increase in the basal activity of Smo (Figure 3B); however, other combinations (SD01, SD123, SD0–3, SD1–5, SD0–5) resulted in a clear elevation of Smo basal activity and the level of basal activity correlates with the number of altered sites (Figure 3B). Nevertheless, the constitutive activities of SmoSD variants are lower compared with that of SmoA1 (Figure 3B). Furthermore, the SD variants were further stimulated by Shh, whereas SmoA1 was no longer regulated by Shh (Figure 3B). Thus, phosphorylation at CK1/GRK sites increases Smo activity in a dose dependent manner but does not confer full activation.
To determine whether the oncogenic mutation activates Smo through its phosphorylation, we mutated several CK1/GRK sites to Ala in SmoA1. Mutating S2/3 (A1SA2, A1SA3, A1SA23) had little if any effect on SmoA1 activity (Figure 3C). In contrast, S1 mutation (A1SA1) or combined mutations of S1 with other sites (A1SA12, A1SA13, A1SA123, A1SA1–5) greatly reduced or nearly abolished SmoA1 activity (Figure 3C), suggesting that S1 phosphorylation is critical for the oncogenic mutation to activate Smo. Of note, the SA1 mutation had a more profound effect on the activity of SmoA1 than that of wild type Smo in the presence of Shh (compare SmoA1SA1 with SmoSA1+Shh). The reason for this difference is unclear, but it is possible that the oncogenic mutation may not fully mimic Shh stimulation so that SmoA1 relies on S1 phosphorylation for its activation more than wild type Smo.
We also found that mutating SA0–5 abrogated SAG-induced Smo activation, whereas SD0–5 exhibited resistance to cyclopamine inhibition (Figure S2I), suggesting that SAG and cyclopamine regulate Smo activity by influencing its phosphorylation.
We next used chick neural tubes to determine the role of Smo phosphorylation in Shh signaling in living organisms. CFP-tagged constructs expressing wild type (WT) or mutant forms of Smo were electroporated into one side of the neural tube, leaving the other side as an internal control, followed by immunostaining to visualize the expression of various Hh responsive genes. Electroporation of SmoWT or Smo variants that mimic low-level phosphorylation (SD1, SD12) did not significantly alter the expression of the marker genes (Figure 3D and Figure S3A); however, electroporation of Smo variants that mimic high-level phosphorylation (SD123, SD1–5, SD0–5) resulted in a dorsal expansion of several ventral markers, including Nkx2.2, Olig2, Nkx6.1, and Islet1 (Figure 3D and Figure S3D). Furthermore, SD123 and SD0–5 but not SmoWT restored the expression of ventral markers suppressed by a dominant form of Ptc, Ptc1Δloop2 (PtcΔ2), as well as prevented the derepression of the dorsal marker Pax7 (Figure 3E) [42]. These results suggest that phosphorylation at CK1/GRK sites increased the basal activity of Smo in the chick neural tubes.
In line with tissue culture experiments, SmoA1 is more potent than SmoSD variants in inducing ectopic expression of ventral marker genes in chick neural tubes (Figure 3D, Figure S3B). Mutating S1 (A1SA1) or combination of S1 with other sites to Ala (A1SA12, A1SA13, A1SA123, A1SA1–5) diminished or completely abolished A1-mediated ectopic activation of ventral markers or suppression of Pax7, whereas mutating S2 and 3 (A1S23) had little if any effect on SmoA1 activity (Figure 3D, Figure S3B), suggesting that phosphorylation at S1 is critical for the oncogenic mutation to activate Smo in the chick neural tube.
Shh induces ciliary accumulation of Smo that correlates with pathway activation, but the underlying mechanism is poorly understood [16],[17]. We determined whether Shh promotes Smo ciliary localization by inducing its phosphorylation at CK1/GRK sites by examining ciliary localization of CFP-tagged wild type or phosphorylation site mutant forms of Smo in MEF cells treated with or without Shh-conditioned medium. As overexpression by transient transfection caused high basal ciliary localization of Smo, we used retroviral infection to express low levels of exogenous Smo. In these conditions, SmoWT was found in less than 5% of cilia in the absence of Shh but accumulated in ∼70% of cilia in response to Shh treatment (Figure 4A–B). We found that SA mutations inhibited Shh-induced whereas SD mutations promoted basal ciliary accumulation of Smo in a dose-dependent manner (Figure 4A–B). In addition, constitutive ciliary localization of SmoA1 was inhibited by the SA1–5 mutation (A1SA1–5, Figure 4A–B). Thus, phosphorylation at CK1/GRK sites is both necessary and sufficient for the ciliary localization of Smo.
A recent study suggested that β-arrestins mediate Smo ciliary localization by binding to Smo and facilitating its interaction with the kinesin-II motor [43]. We hypothesized that Shh-induced Smo phosphorylation promotes its ciliary localization by recruiting β-arrestins. To test this possibility, we transfected NIH 3T3 cells with a YFP-tagged β-arrestin2 (β-arr2-YFP) together with Myc-tagged wild type or mutant forms of Smo. As shown in Figure 4, both Shh and the A1 mutation increased the amount of β-arr2 coimmunoprecipitated with Smo (Figure 4C, lanes 2, 7). The SA mutations nearly abolished Shh- or A1-stimulated interaction (Figure 4C, lanes 4,10), whereas SD0–5 promoted Smo/β-arr2 interaction (Figure 4C, lane 5).
We also confirmed that phosphorylation regulates Smo/β-arr2 association using FRET assay. We found that Shh and A1 increased the FRET between Smo-CFP and β-arr2-YFP, and this increase was abolished by the SA0–5 mutation (Figure 4D). Conversely, SD0–5 increased the basal FRET between Smo-CFP and β-arr2-YFP. Thus, Shh-induced phosphorylation at CK1/GRK sites increases the association between Smo and β-arr2, which may account for the increased ciliary localization of Smo.
To determine whether phosphorylation at CK1/GRK sites regulates Smo conformation, we mutated individual or combination of CK1/GRK sites to either Ala or Asp in C-terminally CFP/YFP-tagged Smo and carried out FRET analysis in NIH3T3 cells. SA0 or SA1 slightly reduced Shh-induced FRETC, whereas individual mutations at other sites (SA2, SA3, SA5) had no effect (Figure 5A). S0 and S1 double mutation (S01) or combined mutation of S0/1with other sites (SA12, SA13, SA123, SA1–5, SA0–5) greatly reduced or nearly abolished Shh-induced FRETC (Figure 5A). On the other hand, the SD mutations resulted in a dose-dependent increase in the basal FRETC (Figure 5B). Overall, the effects of SA or SD mutations on Shh-induced FRETC correlated with their effects on Shh-induced Smo activation.
The SA mutations also diminished A1-induced FRETC (Figure 5C). Furthermore, SA0–5 abolished SAG-induced FRETC, whereas SD0–5 conferred high basal FRETC even in the presence of cyclopamine (Figure 5D). Thus, phosphorylation at CK1/GRK sites induced by Shh, A1, and SAG causes a conformational switch in Smo C-tail, leading to its dimerization, whereas cyclopamine locks Smo in the closed conformation by blocking its phosphorylation.
To examine the spatial and temporal regulation of Smo phosphorylation, we carried out immunohistochemistry experiments using the PS1 antibody that recognizes phosphorylated S1. Because PS1 failed to detect endogenous Smo, we generated NIH3T3 cells stably expressing low levels of Smo-CFP (NIH3T3Smo-CFP). NIH3T3Smo-CFP did not exhibit significant basal ciliary localization of Smo-CFP or ciliary PS1 signal but accumulated both signals in the cilia upon stimulation with Shh, SAG, or 20-OHC (Figure 6A, Figure S4A). Cyclopamine induced ciliary accumulation of Smo-CFP but not PS1 (Figure 6A, Figure S4A). Furthermore, cyclopamine blocked Shh or 20-OHC but not SAG-induced ciliary PS1 signals (Figure 6A; Figure S4A). Thus, Shh, SAG, and 20-OHC induced ciliary accumulation of phosphorylated Smo, whereas cyclopamine trapped unphosphorylated Smo in the cilia. The difference in the sensitivity of SAG and 20-OHC to cyclopamine could be due to different mechanisms of action employed by these small molecules to regulate Smo.
To examine the dynamics of Smo phosphorylation, we treated NIH3T3SmoCFP cells with Shh-conditioned medium or SAG for different periods of time (1, 2, 4, and 24 h). In line with a previous report [17], both Shh and SAG induced a rapid ciliary accumulation of Smo-CFP, and the percentage of Smo-CFP positive cilia as well as the mean intensity of SmoCFP signal increased over time (Figure 6B–D). Importantly, we observed a similar kinetics for PS1 accumulation in the primary cilia (Figure 6B–D). Furthermore, the ratio of PS1 versus Smo-CFP signal intensity in primary cilia remained relatively constant over time.
We also monitored Smo phosphorylation in whole cells by western blot using the PS1 and GFP antibodies. We found that the ratio of PS1 versus Smo-CFP signal intensity was lower at early time points and gradually increased over time (Figure 6E). Thus, Smo phosphorylation exhibited faster kinetics in primary cilia than in whole cells, implying that Smo could be preferentially phosphorylated near or in the primary cilia in response to Shh or SAG, leading to its rapid accumulation in the cilia.
To investigate whether primary cilia regulate Smo phosphorylation, we disrupted the cilia using a dominant negative form of Kif3b (DN-Kif3b), a subunit of the kinesin-II motor required for cilia formation [44]. We found that DN-Kif3b diminished but did not completely block Shh-induced PS1 signal associated with either Smo-Myc or SmoA1-Myc (Figure 6F, lanes 3, 6), suggesting that efficient phosphorylation at S1 depends on the kinesin-II ciliary motor.
We also analyzed whether the primary cilium is required for Shh-induced Smo conformational change by measuring FRETC in the wild type or Kif3a−/− MEFs transfected with wild type or mutant forms of Smo-CFPC/YFPC. We found that Shh or A1-induced FRETC was dramatically reduced in Kif3a−/− MEFs compared with WT MEFs (Figure S4B). In contrast, SmoSD0–5-CFPC/YFPC exhibited high FRETC in both WT and Kif3a−/− MEFs (Figure S4B). Thus, in the absence of primary cilia, Shh and A1 failed to induce the active Smo conformation because of compromised Smo phosphorylation, but an open conformation can be restored by phospho-mimetic mutations.
Although SmoSD0–5 adopts an open conformation in Kif3a−/− MEFs, it failed to induce any Gli-luc expression in the absence of primary cilia (Figure S4C). In contrast, overexpression of Gli1 in Kif3a−/− MEFs activated the Gli-luc reporter. These observations suggest that the primary cilium is not only required for Smo activation but is also essential for signal transduction downstream of activated Smo.
Finally, we investigated how Shh induces Smo phosphorylation by testing the possibility that Shh promotes the accessibility of Smo to its kinases. By immunoprecipitation assay, we found that Shh markedly increased the association between Smo-Myc and endogenous CK1α and GRK2 in NIH3T3 cells (Figure 7B, lanes 1–2; Figure 7C). In addition, Shh induced accumulation of CK1α in primary cilia (Figure 7D). The binding of CK1α/GRK2 to Smo-Myc is specific because we did not detect association between Smo-Myc and endogenous CK1ε or GRK5 under the same condition (unpublished data).
To further explore the interactions between Smo and CK1α/GRK2 and their regulation, we generated several N- or C-terminally truncated forms of Smo (Figure 7A). As shown in Figure 7E, CK1α and GRK2 coimmunoprecipitated with HA-tagged Smo C-tail from aa 544 to aa 793 (SmoCT). Deletion of aa 544–565 from the Smo C-tail (SmoCT2) did not affect CK1α/GRK2 binding; however, further deletion of aa 565–588 (SmoCT3) abolished CK1α binding but did not affect GRK2 binding, suggesting that the membrane proximal region of Smo C-tail between aa 565 and 588 mediates CK1α binding, whereas the distal region between aa 588 and 793 binds GRK2. Consistent with this, we found that SmoΔC588 but not SmoΔC565 pulled down CK1α and neither SmoΔC565 nor SmoΔC588 pulled down GRK2 (Figure 7G, lanes 3–6). Thus, the CK1α binding pocket is located N-terminal to the phosphorylation sites.
Interestingly, SmoΔC588 exhibited increased basal binding to CK1α (Figure 7G, compare lanes 1 and 5), suggesting that the distal region of Smo C-tail inhibits CK1α binding in the absence of Shh. We hypothesized that unphosphorylated Smo C-tail adopts a closed conformation that could mask the membrane proximal CK1α binding domain (Figure 7J). Indeed, the SA0–5 mutation, which locked Smo C-tail in its closed conformation, diminished Shh-stimulated CK1α binding, whereas the SD0–5 mutation, which locked the Smo C-tail in its open conformation, increased the basal CK1α binding (Figure 7B, lanes 3–6; Figure 7C). However, the SA0–5 and SD0–5 mutations in the context of Smo C-tail (SmoCT-SA and SmoCT-SD) did not significantly alter CK1α binding (Figure 7F), unlike their effect in the context of full-length Smo. Thus, instead of directly altering the CK1α binding site, phosphorylation may regulate CK1α binding by influencing the conformation of Smo C-tail and thus controlling the accessibility of the CK1α binding pocket. In contrast, the SD0–5 mutation dramatically increased GRK2 binding in the context of both SmoCT and full-length (Figure 7B and 7F), suggesting that phosphorylation may increase the affinity of a GRK2 binding site(s) in the Smo C-tail.
Although kinase binding to Smo is influenced by phosphorylation, we found that Shh still enhanced the binding of CK1α to SmoSA0–5 and SmoSD0–5 (Figure 7B, lanes 3–6; Figure 7C). Furthermore, CK1α binding to SmoΔC588, which lacks all the CK1/GRK phosphorylate sites, was also upregulated by Shh (Figure 7G, compare lanes 5 and 6; Figure S5). These results demonstrate that Shh can stimulate CK1α binding through a phosphorylation-independent mechanism. In contrast, GRK2 binding to SmoSD0–5 or SmoSA0–5 was no longer regulated by Shh (Figure 7B, lanes 3–6), suggesting that Shh promotes GRK2 binding mainly through the phosphorylation-dependent mechanism. Taken together, these data suggest that Shh may regulate CK1α/GRK2 binding in two steps: 1) Shh stimulates CK1α binding to Smo prior to its phosphorylation, which may provide a mechanism to initiate Smo phosphorylation, and 2) phosphorylation of Smo C-tail releases its inhibition on CK1α binding and at the same time increases its binding affinity for GRK2, leading to amplification of Smo phosphorylation (Figure 7J).
To establish the relationship between kinase association and Smo phosphorylation, we examined how gain- or loss-of-function Smo mutations affect CK1α binding, including two oncogenic mutations (A1 and M1) and three loss-of-function mutations in or near the CK1α binding pocket identified by previous studies (Figure 7A) [28],[39]. We found that both A1 and M1 resulted in a constitutive CK1α/GRK2 binding and Smo phosphorylation with A1 being more potent than M1 (Figure 7H, lanes 3 and 5; Figure S5). In addition, Shh further increased the binding of CK1α to and phosphorylation of SmoM1 but not SmoA1 (Figure 7H, lanes 4 and 6; Figure S5). In contrast, the loss-of-function mutations L430A and S570A blocked Shh-induced CK1α/GRK2 binding and Smo phosphorylation (Figure 7H, lanes 7–10; Figure S5). Another loss-of-function mutation, I573A, which mainly affected Smo stability [28], slightly reduced Shh-stimulated CK1α binding and Smo phosphorylation (Figure 7H, lanes 11–12; Figure S5).
If L430A and S570A affect Smo phosphorylation because they interfere with the accessibility of Smo to its kinases, one would expect that increasing the levels of Smo kinases might rescue the phosphorylation defect. Indeed, cotransfection of CK1α with SmoL430A or SmoS570A resulted in their efficient phosphorylation (Figure 7I).
Finally, we found that A1 mimicked Shh stimulation to enhanced CK1α binding to SmoΔ588 (Figure 7G, compare lanes 7–8 with 5–6; Figure S5), whereas L430A blocked both the basal and Shh-stimulated binding of CK1α to SmoΔ588 (Figure 7G, lanes 11–12, Figure S5), suggesting that these mutations affect the phosphorylation-independent mechanism that regulates CK1α binding (Figure 7J). It is possible that the third intracellular loop may also contribute to CK1α binding and this is disrupted by L430A.
In contrast, the M1 and S570 mutations did not affect either the basal or Shh-stimulated binding of CK1α to SmoΔ588 (Figure 7G, lanes 9–10 and 13–14; Figure S5). Thus, M1 and S570 affect CK1α binding only in the context of full-length Smo and may act mainly by regulating the release of C-tail inhibition (Figure 7J).
Smo is a central component of the Hh signal transduction cascade and an important cancer drug target, but the molecular mechanism by which Smo is activated has remained poorly understood. In this study, we demonstrate that Smo is activated by multi-site phosphorylation mediated by CK1α and GRK2, and phosphorylation promotes both ciliary localization and active conformation of Smo. We provide evidence that graded Shh signals induce increasing levels of Smo phosphorylation that fine-tune Smo activity. In addition, we demonstrate that oncogenic mutations and small molecule Hh pathway modulators including SAG, oxysterols, and cyclopamine regulate Smo through CK1α/GRK2-mediated phosphorylation. We provide evidence that Shh promotes Smo phosphorylation by regulating its accessibility to CK1α/GRK2 and effective Smo phosphorylation depends on the primary cilium. The CK1α/GRK2 sites we identified are conserved among vertebrate Smo proteins; thus, the mechanism we uncover here is likely to be conserved in other vertebrate species.
It has been well established that Drosophila Smo is hyperphosphorylated by multiple kinases in response to Hh stimulation [22],[23],[25]–[27]; however, sequence divergence between Drosophila and vertebrate Smo proteins makes it unclear whether vertebrate Smo proteins are similarly phosphorylated in response to Hh. Using the phospho-tag gel and a phospho-specific antibody, we provide the first evidence that Shh induces hyperphosphorylation of Smo, which is mediated by CK1α and GRK2. Several lines of evidence suggest that CK1α and GRK2 are bona fide Smo kinases. First, our in vitro kinase assay with purified Smo fragments and recombinant kinases demonstrated that both CK1 and GRK phosphorylate multiple sites in Smo C-tail. Second, mutating the CK1/GRK sites in the Smo C-tail abolished Shh-stimulated Smo phosphorylation in vivo. Third, using a phospho-specific antibody that recognized an overlapping CK1/GRK site (S1), we demonstrated that Shh induced phosphorylation at this site through CK1α and GRK2.
We identified a total of six CK1α/GRK2 phosphorylation regions, which we named S0 to S5. S0 and S1 contain multiple phospho-acceptor Ser/Thr residues. Our functional study suggests that S0 and S1 play a major role while other sites play a fine-tuning role in Smo regulation. The employment of multi-site phosphorylation may allow graded Hh morphogens to induce different levels of Smo activity through differential phosphorylation. Indeed, we found that increasing levels of Shh induced a progressive increase in the level of Smo phosphorylation. Furthermore, increasing the number of SA mutations gradually decreased the level of Shh-induced Smo activity, whereas increasing the number of phospho-mimetic mutations progressively increased the level of basal Smo activity.
Although phospho-mimetic mutations increase the basal activity of Smo both in vitro and in vivo, they do not confer full activation of Smo, which is in contrast to the A1/M2 oncogenic mutation. One possibility is that the SD mutations may not fully mimic phosphorylation and may even lock Smo in a less optimal conformation for activation. However, we think this is unlikely because the SmoSD variants can be further stimulated by Shh to reach their full activity. In addition, phospho-mimetic mutations did not affect SmoA1 activity (unpublished data). These observations suggest that Shh and A1/M2 may stimulate an additional mechanism(s) that acts in conjunction with CK1α/GRK2-mediated phosphorylation to fully activate Smo. The proposed paralleled mechanisms could be phosphorylation-independent and/or could involve additional kinase(s). Furthermore, although our in vitro and in vivo assays suggest that phosphorylation at S0–S5 is mediated by CK1α/GRK2, we cannot rule out the possibility that some of these sites might also be phosphorylated by other kinases.
A prevalent view regarding Smo activation is that Hh activates Smo by inducing its ciliary localization [16],[17]. However, this view has been challenged by more recent studies showing that the Smo inhibitor cyclopamine promotes instead of blocks ciliary localization of Smo [18],[19],[45], suggesting that ciliary localization of Smo is insufficient for its activation. Our previous and current studies demonstrate that Shh induces a conformational switch in Smo that is also induced by the A1 mutation and SAG but is blocked by Smo inhibitors including cyclopamine [22],[46],[47], suggesting that Hh-induced Smo conformational switch may represent an additional step for Smo activation. How Smo conformational switch and ciliary localization are regulated remained unknown. Here we demonstrate that both events are governed by CK1α/GRK2-mediated phosphorylation of Smo C-tail. CK1α/GRK2 phosphorylation-deficient forms of Smo are locked in a closed conformation and fail to accumulate in primary cilia in response to Shh stimulation, whereas phospho-mimetic forms adopt an open conformation and accumulate in the primary cilia independent of Shh.
In the absence of Hh, Smo may move in and out of the primary cilium with the exit rate far exceeding the entry rate, resulting in a low steady state level of Smo in the primary cilium. However, Hh-induced Smo phosphorylation and conformational change could tilt the balance by increasing the entry rate and/or decreasing the exit rate. In support of this model, we found that Hh-induced phosphorylation promoted the binding of β-arr2 to Smo. A recent study demonstrated that β-arrestins mediate the interaction between Smo and the anterior-grade trafficking motor kinesin-II [43]. Thus, Hh-induced phosphorylation may promote Smo ciliary accumulation by facilitating its anterior grade trafficking through recruiting β-arr2. It is also possible that phosphorylation may impede the retrograde trafficking of Smo or may stabilize Smo protein in the primary cilium.
Our data suggested that Shh stimulates Smo phosphorylation, at least in part by regulating the accessibility of Smo to its kinases. Our deletion analyses revealed that CK1α and GRK2 bind Smo through the membrane proximal and distal regions of Smo C-tail, respectively. We provided evidence that Smo C-tail in its closed conformation inhibits CK1α binding likely by masking the membrane proximal CK1α binding pocket through steric hindrance, and this inhibition is released by phosphorylation that promotes the open conformation of Smo C-tail. Furthermore, we demonstrate that Shh stimulates the binding of CK1α to the membrane proximal region of Smo C-tail through a mechanism that parallels with the phosphorylation-dependent mechanism. We propose a two-step mechanism for Shh-regulated kinase association and Smo phosphorylation (Figure 7J). In the first step (referred to as the initiation step), Shh stimulates CK1α binding to Smo prior to its phosphorylation, likely by inducing a local conformational change near the membrane proximal region that either optimizes the CK1α binding pocket or makes it more accessible to CK1α. This step may contribute to the initiation of Smo phosphorylation and is promoted by the A1 mutation but is blocked by the L430A mutation. In the second step (referred to as the amplification step), CK1α-initiated phosphorylation further increases CK1α binding by promoting the open conformation of Smo C-tail. Furthermore, phosphorylation of Smo C-tail increases its binding affinity for GKR2. Increased binding of CK1α/GRK2 forms a feedback loop to further increase the level of Smo phosphorylation.
There could be a basal association of CK1α/GRK2 with Smo in quiescent cells, and Shh could induce a change in Smo that makes the CK1/GRK sites more accessible to the bound kinases, which may also contribute to the initiation of Smo phosphorylation. Finally, Smo phosphorylation is likely to be counteracted by a phosphatase(s), which could be essential for keeping basal Smo phosphorylation low. Therefore, Shh could regulate the activity or accessibility of a Smo phosphatase(s) in addition to regulating the Smo kinases.
Our time course study revealed that phosphorylation of Smo occurred more rapidly in the primary cilia compared with the whole cell (Figure 6). In addition, expression of a dominant negative form of Kif3b, which blocks ciliogenesis, attenuated Shh- or A1-induced Smo phosphorylation. These observations suggest that Smo phosphorylation occurs more efficiently in the primary cilia. Interestingly, we found that CK1α is accumulated in primary cilia in response to Shh stimulation (Figure 7D). The increase in the local concentration of CK1α may explain, at least in part, why phosphorylation of Smo is more effective in the primary cilium. It is also possible that Shh-mediated inhibition of Ptc is more effective in the primary cilium.
Despite the profound difference in the primary sequence between Drosophila and vertebrate Smo, our study suggests that their activation mechanisms are remarkably similar (Figure 8). In both cases, Hh induces Smo phosphorylation at multiple sites (although by distinct sets of kinases) that fine-tune Smo activity, and phosphorylation activates Smo by inducing its active conformation and regulating its subcellular localization (cell surface accumulation for dSmo and ciliary accumulation for mSmo). Hh-stimulated phosphorylation induces dSmo conformation change by antagonizing multiple Arg clusters in its C-tail [22]. As the inactive conformation of mSmo is also maintained by a long stretch of basic cluster in its C-tail [22], multisite phosphorylation may promote mSmo conformational change through a similar mechanism. A recent study has demonstrated that GRK2 regulates dSmo by both kinase-dependent and kinase-independent mechanisms [48]. The observation that Shh induces mSmo/GRK2 complex formation raises an interesting possibility that GRK2 may also function as a molecular scaffold to promote mSmo activation.
pGE-Smo-CFPC, pGE-Smo-YFPC, and pGE-Smo-CFPL2YFPC have been described previously [22]. SmoSA and SmoSD substitutions were generated by site-directed PCR mutagenesis. To generate GST-Smo fusion constructs, DNA fragments encoding Smo C-terminal regions with wild type or mutated phosphorylation sites were amplified by PCR and inserted between SalI and NotI sites of the pGEX-4T-3 vector. To construct XZ201-Smo-CFP retrovirus, Smo-CFP variants were PCR out and subcloned between HpaI and SalI sites in the MSCV retroviral vector (XZ201, gift from Dr. Alec Zhang's lab in UTSW). The bovine source of kinase-expressing constructs used in the shRNA rescue experiments were generated by PCR amplification and cloned into pCDNA3.1(+) vector, the dominant negative form of bovine GRK2 (bGRK2-K220R) was generated by site-directed PCR mutagenesis strategy, and the pCS2(+)-CK1α and pCS2(+)-DN-CK1α are gifts from John Graff's Lab [49]. LMP/shRNA against kinase: CK1α, GRK2, or GRK5 were constructed by inserting indicated shRNA fragments into LMP vector (Open Biosystems) containing a PGK-puromycin resistance-IRES-GFP cassette. To generate HA-tagged wild type, SA0–5 or SD0–5 versions of Smo C-tail, wild type, or mutant DNA fragments were amplified by PCR and inserted between NotI and XbaI sites in the HA-pUAST vectors [50], and the HA-tagged constructs were subcloned into pCDNA3.1(+) vector with EcoRI and XbaI sites. All the constructs were sequence verified. DN-Kif3b constructs were kindly provided by Dr. Pao-Tien Chuang [44].
CK1/GRK in vitro kinase assay was performed according to the manufacturer's instruction (Upstate Biotechnologies, 14-714). Briefly, GST-fusion proteins, 0.1 mM ATP containing 10 mCi of γ-32p-ATP and kinases: CK1δ (New England Biolabs), GRK5 (Upstate Biotechnologies, 14-714), were mixed well and incubated at 30°C for 1.5 h in reaction buffer (20 mM Tris-HCl, pH 8.0, 2 mM EDTA, 10 mM MgCl2, 1 mM DTT); the reactions were stopped by adding 4× SDS loading buffer and boiled at 100°C for 5 min; and the phosphorylation of GST-fusion proteins were analyzed by autoradiography after SDS-PAGE.
Unless otherwise noted, all the mammalian cell lines were cultured in DMEM, supplemented with 10% fetal bovine serum (FBS), L-glutamine, 1 mM sodium pyruvate, and penicillin. NIH 3T3 cells were obtained from ATCC. smo−/− and Kif3a−/− mouse embryonic fibroblasts were kindly provided by Dr. Pao-Tien Chuang [44]. Wild type MEFs were derived from wild type mice embryos at 9.5 dpc, embryos were dissected to pieces and transferred to 10 cm dishes for adherence, regular DMEM medium were slowly added, fibroblasts cells that migrated from the embryos were collected by trypsinization after 3∼5 d, and expanded wild-type MEFs were aliquot and frozen for further use. Reagents were used in the following concentrations unless otherwise noted: Recombinant Mouse Sonic Hedgehog N-terminus (ShhNp, R&D systems, Cat #464-SH), 293-Shh-conditioned medium (1∶6 v/v; [40]), SAG (200 nM), cyclopamine (1 µM), CKI-7 (10 µM; Sigma), and Heparin (1 µM; Sigma). SAG and cyclopamine are gifts from Dr. James Chen at Stanford University. The kinase inhibitors were added into the medium the night before collecting the samples, and for heparin treatment, 5 µg/ml Lipofectin (Invitrogen) were mixed together with the medium to facilitate their entry into the cells.
For protein expression, cells were transfected with FuGENE 6 transfection reagent (Roche) according to the manufacturer's instructions, harvested and lysed in RIPA buffer (50 mM Tris-Cl at pH 7.9, 150 mM NaCl, 5 mM EDTA), 1% NP-40 supplemented with protease inhibitors (Roche), and lysates were frozen and thaw 2∼3 times. Immunoprecipitation experiments were performed as previously described [51]. The Phos tag-conjugated SDS-PAGE analysis was performed according to the standard protocols [34]. Phos tag-conjugated acrylamide was purchased from the NARD Institute in Japan. First and secondary antibodies used in this study: mouse anti-Myc (1∶5,000; Sigma), rabbit anti-Myc (A-14; Santa Cruz Biotechnologies), mouse anti-HA (1∶10,000; Santa Cruz Biotechnologies), mouse anti-Flag (1∶10,000; Santa Cruz Biotechnologies), rabbit anti-CK1α (Santa Cruz Biotechnologies), rabbit anti-GRK2 (Santa Cruz Biotechnologies), rabbit anti-GRK5 (Santa Cruz Biotechnologies), rabbit phospho-specific antibodies against S1 (PS1, 1∶50), monoclonal anti-Acetylated tubulin (1∶1,000; Sigam#T7451), Goat anti-mouse IgG HRP (1∶10,000), and Goat anti-rabbit IgG HRP (1∶10,000). PS1 antibody was generated by Genemed Synthesis Inc., phosphorylated peptide EP(pS)ADV(pSpS)AWAQHVTC was injected into rabbit, the serum was affinity-purified by antigen, and the flow-through from the affinity-purification was also kept as control antibody S1 against non-phosphorylated peptide. For immunofluorescence, cells were seeded on ploy-D Lysine coated LAB-TEK chamber slides and were transfected with indicated constructs, followed by treating with indicated reagents for indicated time. Cells were washed 2 times with 1XPBS and fixed with 4% PFA, permeabilized, stained, and mounted for observation with Zeiss LSM510 confocal microscope. FRET assays were performed essentially as previously described [22]. Briefly, CFP was exited at 458 nM wavelength and YFP at 514 nM wavelength. CFP signals were collected once before photobleaching (BP) and once after photobleaching (AP) of YFP. YFP was photobleached with full power of the 514 nM laser line for 1∼2 min at the top half of the cells, leaving the bottom half as an internal control. The CFP signals from the bleached half (both membrane and cytoplasmic signals) were used for FRET calculating, and the efficiency of FRET was calculated with the formula: FRET% = [(CFPAP − CFPBP)/CFPAP] ×100.
The day before transfection, different cell lines were seeded at a density of 1∼2×105 cells/ml in 24-well plates, and cells were transfected with 8XGliBS-luciferase and pRL-TK at 4∶1 ratio, and 5% w/w of pGE-Smo constructs with Fugene 6 (Roche) according to the manufacturer's instructions. After 2 d of transfection, cells were changed to low serum medium (DMEM supplemented with 0.5% calf serum) with or without Shh-conditioned medium combined with additional treatments as indicated, and cells were harvested and luciferase activities were determined using the Dual Luciferase Reporter Assay System (Promega) and FLUOstar OPTIMA (BMGLABTCH). Each sample was performed in triplicate and the assays were repeated for at least 3 times.
All constructs were electroporated into the neural tube of HH st11–12 chick embryos [52]. Embryos were harvested 48 h after electroporation, fixed, and processed for immunohistochemistry as previously described [53]. The following antibodies were used: mouse Pax7, Nkx6.1, Nkx2.2 (from DSHB), rabbit Olig2 (Chemicon), rabbit Islet1/2 (a gift from Dr. T. Jessell), and GFP (Biogenesis). Anterior thoracic levels were analyzed in all cases.
Stable NIH 3T3/shRNA cell lines against kinase CK1α, GRK2, or GRK5 were generated by retroviral infection and selected with 3 µg/ml of puromycin.
HEK 293T cells were transfected with XZ201 retrovirus vectors encoding variant Smo cDNAs and pCL-Eco packaging vector, and supernatants were collected 72 h post-transfection, filtered through a 0.45 µM syringe filter, and added to 50∼70% confluent wild type MEFs with 8 µg/ml polybrene (Sigma) overnight.
|
10.1371/journal.pntd.0002192 | Relationship between Serum Antibodies and Taenia solium Larvae Burden in Pigs Raised in Field Conditions | Serological tests have been used for the diagnosis of Taenia solium infection in pigs. However, those serological results do not necessarily correlate with the actual infection burden after performing pig necropsy. This study aimed to evaluate the Electro Immuno Transfer Blot (EITB) seropositivity with infection burden in naturally infected pigs.
In an endemic area of Peru, 476 pigs were sampled. Seroprevalence was 60.5±4.5% with a statistically higher proportion of positive older pigs (>8 months) than young pigs. The logistic model showed that pigs >8 month of age were 2.5 times more likely to be EITB-positive than ≤8 months. A subset of 84 seropositive pigs were necropsied, with 45.2% (38/84) positive to 1–2 bands, 46.4% (39/84) to 3 bands, and 8.3% (7/84) to 4+ bands. 41 out of 84 positive pigs were negative to necropsy (48.8%) and 43 (51%) had one or more cysts (positive predictive value). Older pigs showed more moderate and heavy infection burdens compared to younger pigs. In general, regardless of the age of the pig, the probability of having more cysts (parasite burden) increases proportionally with the number of EITB bands.
The probability of being necropsy-positive increased with the number of bands, and age. Therefore, the EITB is a measure of exposure rather than a test to determine the real prevalence of cysticercosis infection.
| Taenia solium is a parasite that infects humans. The parasite eggs are released into the environment with human feces in villages with inadequate sanitation. Pigs might ingest the parasite eggs and develop the larval stage named cysticercosis (cysts), mainly in the muscles and heart. If a human accidentally ingests the parasite eggs, then the cysts develop principally in the central nervous system, a serious public health problem. Serological results in pigs do not always correlate with what is observed in the necropsy; in fact, some seropositive pigs are negative or have a few cysts. This study aimed to investigate the Electroimmunotransfer Blot (EITB) with the actual pig infection burden. We selected an endemic area of Peru, and sampled 476 pigs. Seroprevalence was 60.5% and there were more old positive pigs than young ones. A subset of 84 seropositive pigs were necropsied and only 43 (51%) had one or more cysts in the whole body. Older pigs also showed more moderate and heavy infection burdens compared to younger pigs. In general, the probability of having cysts increases with the number of EITB bands. This serological assay might be a measure of exposure rather than a test to detect the real disease.
| Taenia solium/cysticercosis infection is an endemic parasitic disease in less developed countries where pigs are raised as a food source [1]. The life cycle of Taenia solium includes the pig as the normal intermediate host, harboring the larval vesicles or cysticerci, and the human as the definitive host, harboring the adult form of the tapeworm. There are widespread economic losses due to the larval stage (cysticercosis) infection of pigs which affects the quality and safety of the pork [2]. In Mexico, for example, porcine cysticercosis caused the loss of more than half the national investment in swine production, and the losses occasioned by the destruction of meat was estimated at $43 million per year [3], [4].
Humans can also develop cysticercosis in the central nervous system (neurocysticercosis or NCC) which affects mainly older children and adults, and the economic consequences due to chronic disability are heavy [3]. In 1988, the cost was estimated in US$15 million per year only for hospital admission of new diagnosed cases of NCC in Mexico [3]. In addition, a recent study in the same country calculated a total of 25,341 (95% CR: 12,569–46,640) DALYs estimated to be lost due to the clinical manifestations of NCC [5]. In Peru, the total cost of NCC during the first 2 years of treatment (healthcare-related costs and productivity losses) was estimated in $966 per patient. This translates into 54% of a minimum wage salary during the first year of treatment and 16% during the second one. Besides, two-thirds of wage-earners lost their jobs owing to NCC and only 61% were able to re-engage in wage-earning activities [6]. However, the real costs might be underestimated since patients with calcified cysticercosis can have seizures or other neurological manifestations persisting for years even after they have been apparently effectively treated to kill cysticerci [3], [4].
The rates of porcine infection are variable but in highly endemic region over 20% to 42% of pigs may be infected [7]. Figures obtained from slaughterhouse inspection generally demonstrate lower levels of infection due to the poor sensitivity of the examination and also because infected pigs are not brought to the abattoir for slaughter since they are often confiscated without payment [2].
Infection by T. solium in pigs under field conditions can be diagnosed by one of three methods: necropsy, detection of cysts in the tongue, and by means of serological assays that would detect either antibodies or circulating antigens. Necropsy, or for that matter, veterinary inspection, is not particularly useful. It can only assess exposed carcass surfaces or a few exploratory incisions, and is usually by-passed as most pigs are killed clandestinely [2]. Tongue examination, although specific, is only moderately sensitive, requires highly trained personnel, is time-consuming, and entails the risk of being bitten [8] which can influence compliance. Immunological assays appear to be best suited for field surveys as pigs can be bled rapidly from the anterior cava vein and it is less dangerous for the examiner than examination of the tongue [9]. A number of assays have been developed for the detection of antibodies, including Enzyme Linked Immuno Sorbent Assay (ELISA) tests with secretory/excretory antigens [10] and fluid antigens [11], and indirect ELISA using heterologous antigens from T. crassiceps [12], [13] and Cysticercus longicolis [14].
The Electro Immuno Transfer Blot test (EITB) in combination with purified antigens [15] is highly specific and more sensitive than either ELISA or tongue examination for the detection of T. solium infection [8]. However, the presence of circulating antibodies against T. solium does not necessarily correlate with presence of cysts at necropsy, and might return positive results when necropsy is negative [16]. It is difficult, if not impossible, to distinguish antibodies found in current cyst infections from those found due to passive immunity [17] (transfer of maternal antibodies via milk), previous exposures to eggs with no established infection, or aborted infections.
However, a number of sero-surveys have shown that reaction to the seven diagnostic glycoproteins in the EITB assay does not occur randomly, but rather in distinct reaction patterns [18]. Other experiments performed mostly to study treatment alternatives have shown that some of the band pattern combinations (more than 4 bands) were more common than others in heavily infected pigs [8], [19], [20]. There have been previous studies where the cyst burden was associated with the antibody level and antigen detection. For instance, Sciutto et al (1998) found a high specificity, sensitivity and positive predictive value (PPV) in experimentally infected pigs by using an ELISA test for the detection of antibodies and antigens. However, the ELISA performance was lower in a small group of naturally infected pigs (rural conditions) for both ELISA and Immunoelectrotransference [21]. On the other hand, pigs experimentally infected with different doses of T. solium eggs developed a heterogeneous response; the level of serum antibodies and antigens varied with the intensity of infection, and pigs with only a few caseous cysts in muscles and/or vesicular ones in brains had no detectable antibodies [22], [23]. It has been also reported that the EITB was able to detect animals with vesicular forms of the cyst, while those with no larvae or only colloidal or caseous ones were negative to EITB in experimentally inoculated animals [24]. The present study was designed to investigate the relationship between reaction to the EITB assay and the burden of infection on examination by necropsy in naturally infected pigs from a hyperendemic area in Peru.
This study was revised and approved by the Ethical Committee of Animal Welfare of the School of Veterinary Medicine, National University of San Marcos (Lima, Peru) which adheres to the guidelines of the Council for International Organizations of Medical Sciences (World Health Organization). Firstly, a meeting was organized with the communities to invite them to voluntarily participate in this study. An oral consent was obtained from owners (normally the Head of household) who were informed of the objective of the study as well as the minimal discomfort to the animals at the time of drawing the blood sample. Oral consent was used since most of the villagers were not able to read Spanish. After receiving permission we were allowed to work with the pigs either outside or inside the house. A registry record was created for each consenting family and their pigs. Pigs were identified by using numbered ear tags.
A serosurvey for porcine cysticercosis was conducted in 10 villages of the district of Quilcas and San Pedro de Saño, a highly endemic area in Huancayo, in the Central Peruvian Highlands. Presence of antibodies against T. solium was determined using EITB as previously described by Tsang et al [15]. Prevalence was estimated from the serology results with 95% confidence interval. Serology was performed within 24 hours and a subset of the EITB-positive pigs was then selected according to age and number of bands and purchased for the necropsy study. Most of the houses with positive pigs were visited, and owners were asked to sell their pigs to us. Pigs were euthanized and necropsied either at a nearby research station in Huancayo (n = 50) or transported to Lima (n = 34). The latter group was euthanized in our animal facilities at the Veterinary School, National University of San Marcos in Lima. Infection burdens were registered for every pig. An ordinal logistic regression model was fit on the necropsied pigs to assess the age, and age/band combinations that predict the infection burden in pigs. To adjust for correlated necropsy responses within households, we used Generalized Estimating Equations (GEEs) to estimate the model.
All pigs except piglets younger than 2 months of age and pregnant sows were sampled. Blood was obtained using the vacutainer system by approaching the anterior cava vein or the jugular vein. The procedure was applied with minor discomfort in the animals. The samples were centrifuged at 3000 rpm for 5 minutes and then sera was separated and frozen until tested. Pig sera were tested for cysticercosis using the EITB. The age, sex and the EITB results were registered for each of the sampled pigs.
The EITB assay was performed as originally described by Tsang et al [15]. Briefly, this assay uses seven purified T. solium glycoprotein antigens (diagnostic bands GP50, GP42-39, GP24, GP21, GP18, GP14 and GP13, where GP stands for glycoprotein and the number refers to the molecular weight of each antigen expressed in kilo Daltons) in an immunoblot format to detect infection-specific antibodies in the serum of cases of cysticercosis in pigs. The sensitivity of the assay was claimed to be >95%, with 100% specificity initially, and reactions to one or more bands are considered positive [15].
A number of EITB-positive pigs (n = 84) were purchased, euthanized and necropsied. The total number of infected pigs purchased was representative of the array of age and band pattern combinations. The price was set according to current pork market prices.
The euthanasia was performed by injecting an overdose of sodium pentobarbital (100 mg/Kg) intravenously in the pigs. In the necropsy, pigs were carefully examined for the presence of cysts in the brain and muscle tissue, including heart and tongue. Pork was carefully chopped using fine cuts (less than 0.5 cm). Cysticercosis burden was classified as follows: negative (no cysts), light if one to 10 cysts were found in the whole carcass, including the brain; moderate for 10 to 100 cysts; and heavy for those with more than 100 cysts in the whole pig. Overall a pig was considered positive if at least one cyst was found in the whole carcass.
Data was analyzed using the statistical package Stata 9 (StataCorp. College Station, Texas, USA). EITB results were arbitrarily divided in four categories (negative, 1–2 bands, 3 bands, 4–7 bands). The univariate analysis was carried out by gender, location, and age. To evaluate the potential effect of maternal antibodies, pigs were classified into two groups: less or equal than 8 months, and older than 8 months. We determined the probabilities of observing different necropsy scores in pigs given different EITB levels identified, controlling for age and sex of the pig. Necropsy has 4 levels (0 to 3 categories): negative, light (1–10 cysts), moderate (11–100 cysts), and heavy infection burdens (more than 100 cysts).
An ordinal logistic regression analysis was then performed to estimate the odds ratio of being positive using the data from the prevalence study. An ordinal model was fit on the necropsied pigs since we have naturally ordered categories for necropsy results (0 through 3). The model accounted for household clustering by using Generalized Estimating Equations (GEEs) to estimate the model since we had to purchase more than one pig in some houses. Additionally, nine EITB-negative pigs were bought from the same place and were included in order to run the model to assess the probability of necropsy outcome adjusted by this category of pigs. The 95% Confidence Intervals and statistical significance were estimated, with the level of significance set at 0.05. An ordinal logistic regression model is written similarly as a logistic regression model; however, we assume that the probability distribution of falling into one of k = 4 categories follows a multinomial distribution. Therefore, assuming (pi0,…pi3) represents the probabilities of each necropsy result for pig i with the cumulative probability defined as and Pi3 = 1, we write the ordinal logistic regression model as:Where r ranges from 0 to 3, q represents the number of independent (predictor) variables, and g(.) is the cumulative logit link function.
A total of 476 pigs were sampled in 10 villages from Quilcas and San Pedro de Saño Districts, in the Peruvian Central Highlands. Of those, 245 (51.5%) were females and 231 (48.5%) were males. The average age was 9.3 months (SD = 6.9). Pigs were divided in two age-groups, ≤8 months old (n = 267), and >8 months old (n = 209). The proportion of males and females for both age groups did not differ from each other statistically.
The overall seroprevalence for swine cysticercosis was calculated as 60.5% (288/476, 95% confidence interval: 55.9%–64.9%). The range of the prevalences among the ten villages varied from 38.4% in Centro, and as high as 90% in Colpar (Table 1). The proportion of EITB-positive pigs was statistically higher for older (>8 months) than younger pigs (73.2% vs 50.6% respectively, p<0.01) (Table 2). Among the seropositive pigs, the 1–2-bands category was observed in greater percentage for pigs ≤8 months old (30.7%) than older pigs (23.4%), though no statistical difference was demonstrated. However, the 3− and 4+ bands categories were statistically greater for pigs older than 8 months than young pigs (≤8 months) (3 bands: 36.8% vs 14.2%, p<0.01; 4+ bands: 12.9%vs 5.7%, p<0.05, respectively) (Table 3). Overall, it was observed that the number and proportion of seropositive pigs decreased with the increase in the number of reactive EITB bands; there was higher proportion of 1–2 EITB-band pigs (27.5%, 131/476) than 4+-EITB band pigs (8.8%, 42/476) (Table 3).
The logistic regression model showed that a pig older than 8 months was 2.5 times more likely to be EITB-positive than those less or equal to 8 months after controlling for sex and village effects (OR = 2.5, 95% CI: 1.7 to 3.7). Using the village with the lowest prevalence (Centro = 38%) as the reference level, the rest of the other villages had statistically significant higher prevalences with the exception of Canchayllo, Llacta, Progreso, and Santa Cruz, after accounting for sex and age categories. For instance, Colpar showed the highest risk for porcine cysticercosis with a 15 times greater likelihood of being seropositive as compared with Centro after adjusting for age and sex (OR = 15, 95% CI: 2.9 to 83.1). Sex was not a significant risk factor for porcine cysticercosis (Table 4). The goodness of fit test for the logistic regression model was adequate (p = 0.1075).
A total of 84 EITB-positive pigs were purchased from the serosurveyed pigs described above, from which 34 were ≤8 months old, and 50 were >8 months. Among these pigs, 45.2% (n = 38) were positive to 1–2 bands category, 46.4% (n = 39) were positive to 3 bands only, and 8.3% (n = 7) showed 4 or more bands. Chi square computation for EITB categories and necropsy status demonstrated statistical association (p = 0.001) (Table 5).
Forty-one out of 84 EITB-positive pigs were negative to necropsy (48.8%) and 43 (51.2%) had one or more cysts, either healthy or degenerated lodged in muscular tissue (positive predictive value). Among these 43 infected pigs, 22, 7, and 14 had light, moderate and heavy infection burdens, respectively. Moderate and heavy infections were observed in the majority of older pigs (>8 months) as compared with young pigs, though no statistical difference was reached (see subtotal columns in Table 5).
The Generalized Estimating Equations (GEE) for an ordinal logistic regression model adequately fit the data according to a Score Test for the Proportional Odds Assumption (p = 0.072) accounting for household cluster. For this analysis, 9 more EITB-negative pigs were included to model the probability of necropsy outcome. EITB-negative and EITB-1 to 2 bands significantly predict the necropsy outcome compared to the reference category of EITB-4+ bands. The predicted cumulative probabilities of necropsy being observed given EITB levels can be observed in Tables 6 and 7; the probabilities are cumulative since the model is cumulative.
For instance, the probability of observing a light infection burden or negative pig when they are EITB-negative is 0.96 (96%) with a 95% confidence interval of 77% to 99%, significantly higher than the probability of 0.32 (32%) when they have 4 or more EITB bands with a 95% CI of 9% to 71% for pigs older than 8 months. In the same context, the probability of having light infection or being negative when a pig has 1–2 EITB-bands was 0.87 (87%) with 95% confidence interval of 72% to 95% for the same age-category pigs (Table 6).
For pigs ≤8 months, the probability of being a light infection burden or negative pig when it is EITB negative is 0.98 (98%) with a 95% confidence interval of 87% to 99%, significantly higher than the probability of 0.49 (49%) when they have 4 or more EITB bands with a 95% CI of 16% to 83%. In contrast, the probability of having light infection or being negative when a pig has 1–2 EITB-bands was 0.93 (93%) with 95% confidence interval of 83% to 97% for the same age-category pigs (Table 7).
From those computations, we calculated the predicted probabilities, not the cumulative ones, for pigs to have cysticercosis according to the number of bands (Table 8). For example, for an EITB-negative-≤8 months old pig, the probability of being necropsy-negative is 93% while the probability of being heavily infected is only 1%. In the same context, a ≤8-months old-pig with 4 or more EITB bands has a probability of 36.8% of being classified as heavily infected (more than 100 cysts). The probability of being heavily infected increases up to 53.9% for a >8 month-old pig when it reacts to 4 or more EITB bands. In general, regardless of the age of the pig, the probability of having more cysts (parasite burden) increases proportionally with the number of EITB bands (Table 8), while the probability of being necropsy negative reduces as EITB bands increases.
We evaluated the performance of the EITB test to diagnose actual swine cysticercosis in pigs from rural communities where the disease has been declared hyperendemic (Central Peruvian Highlands). The meat inspection and the tongue examination to detect the parasite are not practical and almost impossible to occur in the study. The main limitation of veterinary inspection, an insensitive procedure, is that pigs with low cysticercercal burden might not be detected; therefore, lightly infected carcasses would remain in the chain food and the parasite transmission would persist in the population [21]. Thus, immunological assays appeared to be best suited for this type of field surveys. Pigs were bled rapidly from the jugular or anterior cava vein and sera was then analyzed using the EITB test. Among the serological tests available worldwide, EITB has become one of the most common one to study human and pig cysticercosis in endemic countries such as Peru.
Prior to the development of the EITB, serological diagnosis of porcine cysticercosis was hampered by the lack of a reliable test to establish previous exposure to T. solium eggs. However, the presence of T. solium specific antibodies detected by EITB, or any diagnostic test for that matter, does not always correlate with the detection of parasites at necropsy; in fact, some serologically positive pigs have had subsequent negative necropsy results [10], [21], [23], [25]. The interpretation of serological tests may vary not only with the tests (e.g. EITB, ELISA) but also with the type of infection. For instance, it has been shown that the ELISA specificity was very high for the detection of both antibodies and antigens in pigs from commercial farms; this test was, furthermore, found to be highly sensitive and specific in experimentally infected pigs. However, the sensitivity, specificity and PPV did not even reach 50% for ELISA-antigen detection and was very low for antibody detection when evaluating pigs under rural conditions [21].
The initial sensitivity and specificity of the EITB was initially claimed to be >95%, and 100% [8], [15]. Although we were not able to estimate sensitivity and specificity in this study, we found a low PPV of 51.2% even though the study area was a hyperendemic for swine cysticercosis. A positive serological result in the face of a negative necropsy could occur because of the transfer of maternal antibodies to the offspring which remain for several months, prior effective treatment, past infection that has been cleared (degenerated or caseous cysts), exposure to T. solium eggs without development of observable cysts (not enough time for the cyst to develop at the time of the necropsy), or the ingestion of non-viable or infective eggs among other possible explanations. In addition, it is described the effect of “secondary transmission” where seropositive pigs (1–2 EITB bands) were negative at necropsy while a 3-EITB-band pig had very few degenerated and healthy cysts in the whole carcass [26]. Similar low PPV has been reported from pigs maintained under rural conditions in Mexico [21]. These findings may confirm a disadvantage of the EITB test, particularly in pigs from rural areas and natural exposure. It may be interesting to consider the EITB test as an indicator of environmental contamination by detecting seropositive pigs rather than only a tool to assess actual infection status.
There is another aspect related with the EITB result interpretation. We would like to point out that the comparison of the EITB test between different research groups may not be appropriate since the technique and its standardization are more complicated than other serological tests (e.g. ELISA). Different antigen sources and antigen preparation might, additionally, influence the final results of the EITB.
In this study, we are reporting that there is an increasing trend of being necropsy positive with the number of EITB bands regardless the age of the pigs. Interestingly, the predictive probabilities of having cysts were higher in pigs that reacting to more diagnostic bands (3 and 4+) than pigs having a few bands (1–2 bands). The EITB tests was moderate in assessing the heavy infection burden for pig cysticercosis; that probability was observed in older pigs (>8 months of age) with 4+ EITB bands (53.9%) while the same infection status had a probability of 25.2 when pigs reacted to 3 bands. Thus, there are two important factors in order to estimate the actual infection status; first is the pig age, the older ones have higher probability of being infected. Secondly, the number of EITB bands; the probability of having viable cysts (active infection) increases directly with the number of reacting EITB bands. Similar observation was also described by other researchers, where the number of EITB bands increased with the infective dose and it was directly correlated with the duration of the infection in experimentally infected pigs [27].
Although the presence of antibodies in necropsy negative pigs may, in some ways, limit the use of EITB, these results suggest that diagnostic patterns did not happen at random and that these results are related to the final infection outcome. In fact, we have shown a higher proportion of seropositives in older (>8 months old) than younger (≤8 months old) pigs (73.2% vs 50.6%, respectively). A study in Mozambique also demonstrated that the prevalence of swine cysticercosis increased with the age of the animals [28]. Similarly, Garcia et al [29] reported that pig cysticercosis prevalence increased with age. Older pigs might have had more chance to get exposure to T. solium eggs than younger ones, and more time for cyst to develop and trigger the production of circulating antibodies. Older pigs were more than 2 times likely to be seropositive as compared with younger ones. Besides, it could be possible that younger pigs are protected during their first months of life against parasite infection, perhaps due to the presence of maternal cysticercus-antibodies (passive immunity) and become susceptible later after the slow clearance of those antibodies. Passively transferred antibodies for cysticercosis are detected by EITB several months after piglets are born from naturally infected sows [17].
Knowing the true infection status of a pig in an endemic area is crucial for the implementation of a control program. Seropositive pigs might not be necessarily confiscated if reacting with one or two EITB bands, while pigs reacting to more EITB bands might be either eliminated, treated, or may help to identify T. solium hotspot areas. As demonstrated for humans, the seroprevalence defined by the presence of three or more positive EITB bands increased 13% each time distance to the nearest tapeworm carrier halved [30]. Monitoring seroprevalence of pigs and identifying those reacting to more EITB bands (cyst carriers) might help to locate tapeworm carriers (humans) and hotspots in the environment.
Pig cysticercosis seroprevalence is hyperendemic in the Central Highlands of Peru. In this study eight out of the ten villages had more than 50% of their pigs seropositive to cysticercosis, likely one of the highest rates as compared with other endemic countries. For instance, the apparent prevalence of cysticercosis for antibodies or antigen detection ranged from 10–35% in Mozambique [31], and from 24.6–32.2% in Cameroon [32]. However, those serological high rates may represent either disease (presence of cysts) or only exposure to T. solium eggs with no disease at all but detection of circulating antibodies. Although the 10 villages were alike in terms of socioeconomic, educational level and demographic characteristics, there were villages with significant higher risk to have EITB-positive pigs as compared with Centro (the lowest prevalence). There might be other related factors that we did not evaluate at the time of this study (e.g. latrine presence, nearby tapeworm carriers, etc). Serological results, therefore, need careful interpretation since there are often more antibody-positive pigs than pigs harboring cysts, thus the positive predictive value of EITB could be low [33]
This experiment also suggests that the burden of infection was aggregated. The final burden of infection in an infected pig is related to the infective dose and to a series of events that happen before, during and after the infection challenge. Usually, the population distributions of helminthes indicate a tendency towards aggregation, meaning that the majority of parasites are harbored by a minority of host [34]. Aggregation is generally recognized as an important factor in the dynamics of host-macroparasite interactions, and it has been found to be relevant in stabilizing population dynamics in a coexisting equilibrium [34], [35]. Aggregation tends to influence the interactions that regulate parasite numbers, such that the interactions influence a larger proportion of the parasite population. The impact of macroparasites upon host populations is critically dependent upon parasite frequency distributions [36] (Gonzalez, personal communication).
We demonstrated that EITB serology for the diagnosis of cysticercosis in pigs has to be interpreted carefully. Pigs reacting to 4 or more EITB bands have higher probabilities to be infected, harboring more cysts, than pigs with less than 4 bands. In the same context, younger pigs showed less probability of having cysts than older pigs (≥8 months old). There is a proportional trend between the number of EITB bands and the cyst burden; the greater number of bands the higher probability of harboring cysts in the muscles. Transient cysticercosis-antibodies might be the result of passive immunity in piglets, unsuccessful T. solium egg infection (e.g. immunity of the host, ingestion of non-viable eggs), and antiparasitic treatment in pigs. Further studies might be necessary to standardize or refine the current serological test to increase its sensitivity in detecting the true disease (cyst presence).
Other Members of the Cysticercosis Working Group in Peru are Silvia Rodriguez, Luis Gomez, Guillermo Lescano, Viterbo Ayvar, Hermes Escalante, Juan Jimenez and Guillermo Gonzalvez
|
10.1371/journal.pcbi.1002630 | Predicting the Extension of Biomedical Ontologies | Developing and extending a biomedical ontology is a very demanding task that can never be considered complete given our ever-evolving understanding of the life sciences. Extension in particular can benefit from the automation of some of its steps, thus releasing experts to focus on harder tasks. Here we present a strategy to support the automation of change capturing within ontology extension where the need for new concepts or relations is identified. Our strategy is based on predicting areas of an ontology that will undergo extension in a future version by applying supervised learning over features of previous ontology versions. We used the Gene Ontology as our test bed and obtained encouraging results with average f-measure reaching 0.79 for a subset of biological process terms. Our strategy was also able to outperform state of the art change capturing methods. In addition we have identified several issues concerning prediction of ontology evolution, and have delineated a general framework for ontology extension prediction. Our strategy can be applied to any biomedical ontology with versioning, to help focus either manual or semi-automated extension methods on areas of the ontology that need extension.
| Biomedical knowledge is complex and in constant evolution and growth, making it difficult for researchers to keep up with novel discoveries. Ontologies have become essential to help with this issue since they provide a standardized format to describe knowledge that facilitates its storing, sharing and computational analysis. However, the effort to keep a biomedical ontology up-to-date is a demanding and costly task involving several experts. Much of this effort is dedicated to the addition of new elements to extend the ontology to cover new areas of knowledge. We have developed an automated methodology to identify areas of the ontology that need extension based on past versions of the ontology as well as external data such as references in scientific literature and ontology usage. This can be a valuable help to semi-automated ontology extension systems, since they can focus on the subdomains of the identified ontology areas thus reducing the amount of information to process, which in turn releases ontology developers to focus on more complex ontology evolution tasks. By contributing to a faster rate of ontology evolution, we hope to positively impact ontology-based applications such as natural language processing, computer reasoning, information integration or semantic querying of heterogenous data.
| Despite the last decade's efforts to structure and organize the deluge of biomedical data brought on by high throughput techniques, there are still many issues that challenge biomedical knowledge discovery and management [1].
On one hand, most scientific knowledge is still present only in natural language text in the form of scientific publications, whose number grows nearly exponentially making it necessary to employ text mining techniques if we are ever to aspire at keeping up. However, the natural ambiguity and subjectivity of natural language hinders the automated processing of scientific publications. On the other hand, although there is a large number of databases to store biomedical data, the effort to achieve interoperability between them is still lagging behind, given that most resources, particularly the older ones, were developed in a completely independent fashion, and the efforts to connect them to other resources are still insufficient.
One very important breakthrough for both areas, was the development of biomedical ontologies (bio-ontologies). They support both issues, by providing unequivocal and structured models of specific domains, which is fundamental to resolve semantic ambiguities in text mining and also to serve as a common background to biomedical databases.
The development of a biomedical ontology, or other domain ontologies, is a very demanding process that requires both expertise in the domain to model, as well as in ontology design. This means that people from different backgrounds, such as biology, philosophy and computer science should be involved in the process of creating an ontology. However, specific biomedical ontologies are usually built by small teams of life sciences researchers, with little experience in ontology design. They are responsible for first, agreeing on the precise limits of the domain to model; second, defining the structure and complexity of the model; and finally, building the ontology itself by creating the concepts, relations and other axioms it might contain [2].
Several methodologies have been developed to help build ontologies [2]–[5], with the most well-known ontology editors in the biomedical ontologies community being Protégé [6] and OBO-Edit [7]. Nevertheless, ontology development remains a mostly manual and labor-intensive task, which is magnified if the domain to model is as dynamic and complex as the life sciences. Biomedical ontologies can never be considered complete, always having to adapt to our new understanding of biological knowledge. This forces biomedical ontology development to be an iterative process [8], [9] in order to keep up with the dynamic and evolving domain. In fact, one of the tenets of the Open Biological and Biomedical Ontologies (OBO) Foundry, an initiative that establishes a set of principles for ontology development in the biomedical domain, is that an ontology should be maintained in light of scientific advance [10].
This ontology evolution [11] is a continuous effort, requiring large investments of both time and resources with each new version that is produced. Moreover, many biomedical ontologies cover large and complex domains which magnifies the effort required, even when considering highly successful ontologies, such as the Gene Ontology [12], where a large community is engaged in its creation. These challenges create the need for semi-automated systems that are able to support ontology engineers in the task of ontology evolution. However, a significant majority of efforts in this area is not concerned with evolving an existing ontology, but rather in learning a new ontology from scratch, usually from textual resources [13]–[19]. Nevertheless, they can in principle be used for ontology extension as well. These approaches usually depend on either a manually selected corpus of texts to be used as input to narrow down the domain of interest, or process large corpus with generic domains.
A relevant process of ontology evolution is the addition of new elements, i.e. ontology extension. Ontology extension is particularly relevant in fast growing domains such as biomedicine, where new knowledge is created everyday. The first step in this is to identify the changes that need to be performed: change capturing. This is vitally different from a general ontology learning process that handles the whole domain at once, in that it is focused on specific areas within the domain of the ontology to be extended.
In this paper we present a methodology that addresses change capturing by predicting ontology extension. The fact that these changes can in principle be semi-automatically discovered by analyzing the ontology data and its usage motivated the present work. It is a supervised learning based strategy that predicts the areas of the ontology that will undergo extension in a future version, based on previous versions of the ontology. By pinpointing which areas of the ontology are more likely to undergo extension, this methodology can be integrated into ontology extension approaches, both manual and semi-automated, to provide a focus for extension efforts and thus contributing to ease the burden of keeping an ontology up-to-date.
The primary goal of our methodology is to function as a first step in automated ontology learning or extension systems. Ontology learning systems, usually rely on the analysis of a manually constructed corpus of documents pertaining to the domain of interest and their performance is closely coupled to the relevance of these documents. The challenge of focusing the ontology given an heterogenous corpus in ontology learning has been identified [20], a challenge that is amplified when it comes to ontology extension of large ontologies, as is the case of many biomedical ontologies. A comprehensive corpus for these ontologies would be quite large and building and then processing it would be cumbersome. By applying our strategy, ontology developers can identify subdomains to extend, create tailored corpus for them, and then run the learning systems over them, reducing the amount of data they have to process to identify new concepts. Another option for ontology extension is based on ontology matching, which can be used to support the integration of elements from other ontologies. Our strategy can also be interesting in this case, since by pinpointing the areas to extend, it can help to narrow down on specific ontologies to match.
Our main contribution for ontology developers lies in the speeding of the process of extension in these areas, thus releasing the experts to focus on more complex ontology evolution issues. We have chosen to evaluate our approach using the Gene Ontology, since it provides many versions spanning a number of years, and is perhaps one of the best known and widely used biomedical ontologies.
In the remainder of this section we will introduce some basic concepts, present related work and describe the Gene Ontology.
Ontology evolution can be defined as the process of modifying an ontology in response to a certain change in the domain or its conceptualization [21]. These include (1) changes in the portion of the real world they model, (2) a reassessment of the relevance of some element to the ontology, (3) the uncovering of information previously unavailable, or (4) a need to correct previous mistakes [22]. In general, the evolution of biomedical ontologies is mainly concerned with the third and fourth types, given the dynamic nature of biological knowledge production, everyday new discoveries are published, rendering some facts obsolete and bringing new knowledge to light.
Ontology evolution comprises several different processes, based on the type of change transformations they employ over ontology elements: add, remove or modify. While adding new elements is mostly employed in response to a change of the first or third type, removing elements is often related to the first, second and fourth types. Modifying existing elements can belong to any of the four kinds and ultimately be seen as a compound change of removing one element and adding a slightly different new one. In this work we are only concerned with change transformations that add new elements to the ontology, thereby extending it.
Although [21] and [23] provide an exhaustive terminology for ontology change, some finer grained aspects of ontology evolution remained confusing, with several terms being used in an ambiguous fashion. In a previous work [24] we defined and distinguished three terms related to ontology changes concerned with the addition of new elements: ontology extension, ontology refinement and ontology enrichment. Although ontology extension is often used interchangeably with both refinement and enrichment, we defined them as follows:
Ontology extension is the process by which new single elements are added to an existing ontology.
Thus, ontology extension is concerned with elementary changes of the addition type. Many reasons can motivate such a change, such as new discoveries, access to previously unavailable information sources, a change in the viewpoint or usage of the ontology, a change in the level of refinement of the ontology, etc, but they all rely on the finding of new knowledge. Ontology extension encompasses both ontology refinement and ontology enrichment.
Ontology refinement is the addition of new concepts to an ontology, where a new subsumption relation is established between an existing concept and the new concept. For instance, the addition of the concept “mitochondrial fusion" as a subconcept of “organelle fusion" in the biological process branch of the Gene Ontology.
Ontology enrichment is the process by which non-taxonomical relations or other axioms are added to an existing ontology. For instance, the addition of the relation “regulates" between the GO concepts “regulation of mitochondrial translation" and “mitochondrial translation".
Before these changes are actually performed, the need for the change must be identified. This is the first step of any ontology evolution process, the change capturing phase, and it can be based on explicit or implicit requirements [15]. Explicit requirements correspond to those made by the ontology developers or to requests made by end-users. Implicit requirements correspond to those that can be uncovered by change discovery. Stojanovic [25] lists a series of guidelines for change capturing, organized into three types according to the kind of data they exploit, to which Castaño et al. [26] add a fourth:
structure-driven: which are derived from the structure of the ontology, e.g. ‘A class with a single subclass should be merged with its subclass’.
data-driven: which correspond to implicit changes in the domain and are discovered through the analysis of the instances belonging to the ontology, e.g. ‘A class with many instances is a candidate for being split into subclasses and its instances distributed among newly generated classes’.
usage-driven: which are deduced from the usage patterns of the ontology in the knowledge management system e.g. classes that have not been retrieved in a long time might be out of date.
discovery-driven: which is applied when a new instance cannot be described by the ontology classes, and new classes are identified using external resources.
Although there is a large body of work on ontology evolution (for a review see [27]), there are few works on the change capturing phase. Stojanovic at al. [28] proposed an approach to ontology evolution that is based on optimizing the ontology according to the end-users needs. They track end-users interactions with an ontology-based application to collect useful information that can be used to assess what the main interests of the end-users are. Their approach is then a usage-driven change discovery, which focuses on discovering anomalies in the design of an ontology, whose repairing improves the usability of this ontology. This uses several measures, based on querying and browsing of an ontology-based application.
Browsing-based measures are based on the user's browsing of links between ontology concepts. They define the usage of two concepts and as the number of times the link between them has been browsed, where the is a subconcept of a concept . This concept is used in four measures for estimating the uniformity (balance) of the usage of a link regarding the link neighborhood: (1) SiblingUniformity represents the ratio between the usage of a link and the usage of all links, which have the common source node with that link (the so-called sibling links); (2) ChildrenUniformity stands for the ratio between the sum of the usage of all the links whose source node is the given node and the sum of the usage of a node through all incoming links into this node. (3) ParentUniformity is the ratio between the usage of a link and the usage of all links which have the common destination node with that link, and (4) UpDownUniformity characterizes the ratio between the usage of a link in two opposite directions, i.e. in browsing down and browsing up through a hierarchy.
Another usage-driven strategy was proposed by [29] in the context of the evolution of multiple personal ontologies, which is based on a user's ratings of concepts and axioms.
Also relevant for our work is the investigation of ontology evolution in biomedical ontologies.
In [30], the author applied a previously proposed strategy, Evolutionary Terminology Auditing (ETA) [22] to assess the quality of GO using reality as benchmark. This strategy can be used not so much to demonstrate how good an individual version of a terminology is, but rather to measure how much it has been improved (or believed to have been improved) as compared to its predecessor. This is based on matches and mismatches between ontology versions, and their motivations, which are expressed by 17 possible configurations split into four groups, denoting, respectively, the presence or absence of a term and whether the presence or absence of a term in a terminology is justified or unjustified. Of these 17 configurations only two correspond to a need for extension, in which an entity is missing and it is real and relevant for the ontology.
[31] proposes an approach to automatically discover evolving or stable regions of ontologies. This approach is based on a cost model for changes between ontology versions and is able to identify regions that have been undergoing (or not) extensive changes.
On a previous study we delineated a framework to analyze ontology extension and used it as a background for investigating the feasibility of predicting ontology extension based on a set of rules [24]. In predicting ontology evolution we were aiming at developing a methodology for change capturing. We based our set of rules on the guidelines proposed by [25] following [8] for ontology development, namely:
Based on these we created a set of rules for predicting the extension of the Gene Ontology:
Rule 1: A class with less subclasses than its siblings is a candidate for refinement
Rule 2: A class with more total annotations than its siblings is a candidate for refinement
Rule 3: A class with more manual annotations than its siblings is a candidate for refinement
Application of these rules to several versions of the Gene Ontology yielded very poor prediction results, highlighting the need for more complex approaches to model this issue.
The strategy we present here is unlike previously described works, since we use metrics of previous ontology versions to support prediction, whereas change capturing approaches are based on manually derived rules and ontology evolution approaches analyze evolution of existing ontology versions.
The Gene Ontology (GO) is currently the most successful case of ontology application in bioinformatics and provides a controlled vocabulary to describe functional aspects of gene products under three distinct ontologies: biological process, molecular function and cellular component. GO terms are structured in a directed acyclic graph with its hierarchical backbone being composed of and relations.
GO is used to annotated gene products, and these annotations are compiled by the Gene Ontology Annotation project (GOA). GO annotations are assigned an evidence code which identifies the kind of evidence supporting the annotation. Although over a dozen evidence codes exist, the most relevant distinction between them is whether they are manually assigned by a curator or inferred electronically. Electronic annotations are generally considered to be of lower quality than manual ones, but compose the vast majority of present GO annotations (over 97%). Another relevant aspect of annotations is whether they can be considered direct, i.e. the annotation was made precisely to that GO term; or indirect, i.e. the annotation was made to a subconcept of that GO term, from which we can deduce that there is also an annotation to all of its superconcepts.
GO also provides a cut-down version of the GO ontologies, GO Slims, which contain a subset of the terms in the whole GO to give a broad overview of the ontology content without the detail of the specific fine grained terms.
There are about one hundred contributors to GO between the GO Consortium and GO Associates, and they are expected to contribute regularly towards the content of GO. Other GO users can also contribute by suggesting new terms via Sourceforge.net, however the majority of content requests are made by GO team members [32]. GO team experts base their decision to change the ontology on the following precepts:
Although some steps have been taken in the direction of automatizing some aspects of GO evolution, namely the extension of GO with computable logical definitions including cross-references to other ontologies [33] and a new method to optimize the distribution of the information within the GO structure [34], the evolution of GO remains challenging given the complex decision-making processes involved [35].
Following our previous work [24], we used 15 versions of the Gene Ontology spanning a period of seven years. Table 1 identifies these versions, and describes a few general statistics about them. The versions have a six-month interval between them or as close to that as possible, since not all versions have a full database available from the Gene Ontology archive.
The intuition behind our proposed strategy is that information encoded in the ontology or its annotation resources can be used to support the prediction of ontology areas that will be extended in a future version. This notion is inspired by change capturing strategies that are based on implicit requirements. However in the existing change capturing approaches, these requirements are manually defined based on expert knowledge. Our system attempts to go beyond this, by trying to learn these requirements based on previous extension events using supervised learning.
In our test case using GO, we use as attributes for learning a series of ontology features based on structural, annotation or citation data. These are calculated for each GO term and then used to train a model able to capture whether a term would be extended in a following version of GO.
Structural features give information on the position of a term and the surrounding structure of the ontology, such as height (i.e. distance to a leaf term), number of sibling or children terms. A term is considered to be direct child if it is connected to its parent by an is_a or part_of relation, but the total of children of a term encompasses all descendants regardless of the number of links between them. Annotation features are based on the number of annotations a term has, according to distinct views (direct vs indirect, manual vs all). Direct annotations are annotations made specifically to the term, whereas indirect annotations are annotations made to a parent of the term, and thus inherited by the term. Manual annotations correspond to those made with evidence codes that reflect a manual intervention in the evidence supporting the annotation, while the full set of annotations also includes electronic annotations. Citation features are based on citation of ontology terms based on external resources, in our case PubMed. Finally hybrid features combine some of the previous features into one single value. These features can be mapped onto the change discovery types: structural features belong to their homonymous change discovery type; annotations features can be seen as both data and usage based, since they can be interpreted as both ontology instances and ontology usage; and citation features correspond to the discovery-driven change, since they are derived from external sources. In total we defined 14 features, which we grouped into five sets (see Table 2): all, structure, annotations, uniformity, direct, indirect, and . The first three sets are self-explanatory. Uniformity set features were based on [25], where we considered annotations to represent usage. The direct set joins direct features of terms, in terms of children and annotations, whereas the indirect set joins the same kind of features in their indirect versions. The best sets were based on the best features found after running the prediction algorithm for individual features.
Due to the complexity of ontology extension, we have established a framework for the outlining of ontology extension in an applicational scenario. This framework defines the following parameters:
By clearly describing the ontology extension process according to this framework, we are able to accurately circumscribe our ontology extension prediction efforts.
The datasets used for classification were then composed of vectors of attributes followed by a boolean class value, that corresponded to extension in the version to be predicted, according to the used parameters. To compose the datasets we need not only the parameters but also an initial set of ontology versions to be used to calculate features and the ontology version to calculate the extension outcome (i.e. class labels). So given a set of sequential ontology versions , we need to choose one ontology version to predict extension, , and then based on time parameters and FC, select the set of ontologies to be used to calculate features. For example, for a set of ontologies , if we chose to predict extension, along with and FC = 2, the set of ontologies to calculate features will be .
We tested several supervised learning algorithms, namely Decision Tables, Naive Bayes, SVM, Neural Networks and Bayesian Networks, using their WEKA implementations [36]. For Support Vector Machines, we used the LibSVM implementation with an RBF kernel and optimized the cost and gamma parameters through a coarse grid search. For Neural Networks we used the Multilayer Perceptron implementation, with the number of hidden layers equal to , a training time of 500 epochs, and we performed a coarse grid search to optimize the learning rate. Regarding Bayesian Networks, we estimated probabilities directly from the data, and focused on testing different search algorithms, namely Simulated Annealing, K2, and Hill Climbing. Furthermore we had to take into consideration that there are many more terms that are not extended than terms that are, between two sequential ontology versions, which creates unbalanced training sets. To address this issue we used the SMOTE algorithm [37]. SMOTE (synthetic minority over-sampling technique), is a technique that handles unbalanced datasets by over-sampling the minority class and under-sampling the majority class that has been shown to support better classification results for the minority class.
To evaluate our Ontology Extension Prediction strategy we employed a simple approach: compare our predictions to the actual extension of the Gene Ontology in a future version. To this end we employ another time parameter:
This time parameter is used to create the test set, by shifting the ontology versions according to TT. So for instance, given a set of ontologies and using FC = TT, the training and test sets would correspond to the those in Figure 1. Although there may be an overlap in the ontology versions used in a particular training/testing setup, the ontology versions used to determine the class values are always distinct, ensuring that our setup in unbiased.
This approach allows us to compare the set of proposed extensions to real ones that actually took place in a future version of the ontology. We can calculate precision, recall and f-measure metrics, by using the real extension events observed in the more recent ontology version as our test case. These metrics are based on the number of true positives, false positives, true negatives and false negatives. A true positive is an ontology class that our supervised learning strategy identified as a target for extension, and that was indeed extended in the test set, whereas a false positive although having also been identified as a target for extension, was not actually extended. Likewise, a false negative is an ontology class which was not identified as a target for extension, but was in fact extended in reality, whereas a true negative was neither identified as a target nor was it extended in the test set. Precision corresponds to the fraction of classes identified as extension targets that have actually been extended, while recall is the fraction of classes identified as extension targets out of all real extensions. F-measure is a measure of a test's accuracy that considers both precision and recall.(1)(2)(3)
When trying to predict ontology extension we are not just focusing on which features are best predictors, but also on how to design the learning process to best support the prediction. Consequently, we are not only trying to find the best prediction set up in terms of features and machine learning algorithms, but also in terms of our strategy's parameters.
A first step in our experiments was to determine the best term set to use, and to investigate if this was influenced by different parameters. To this end, we tested the following term sets within each GO ontology: all terms, all terms with a depth of 3, 4 and 5, all GO Slim general terms, all GO Slim general leaf terms, all terms at a depth of 1 from the GO Slim general leaf terms, under the same sets of parameters (see Table 3).
To provide a simple basis for our first analysis we focused on the biological process hierarchy and chose a single feature and WEKA's Decision Table algorithm with attribute selection using BestFirst. Results are presented (unless otherwise specified) using the average f-measure obtained using all possible setups derived from the 15 GO versions available, since we are analyzing a large number of combination of different parameters. So for instance, when using , FC and TT, we get a total of ten runs for our prediction evaluation, whereas using = 3, FC and TT we get only six runs.
Before comparing term sets, we need to analyze the trends between parameter sets. First we focused on extension types and modes (see Table 4). The first clear trend to emerge is that indirect extension is predicted with much more success (0.49–0.86) than direct extension (0.1–0.27). Furthermore, in regards to comparing refinement to enrichment and generic extension, enrichment is poorly predicted, with a performance around 0.20–0.30. The performance for indirect refinement and extension in term sets derived from depth performance is comparable (0.63–0.78), whereas in GO Slim sets refinement is better predicted (0.65–0.86 vs. 0.62–0.65).
To clarify this difference, we calculated the average extended proportion for each extension type (see Table 5 for the values for the term set at depth = 4), i.e. the average proportion of extended terms for all GO versions. We verified that the proportion of extended terms is higher for biological process, independently of extension type, followed by cellular component and molecular function, and that the proportion of refined terms is higher than enriched terms, independently of GO term type. This can have an impact on training since there are fewer examples of enrichment.
As for the time parameters (see Table 6) and using indirect extension and refinement, the differences are less marked. An increase in the number of versions () used to calculate the feature values from one to three does not significantly alter the results, and when we extend the interval between versions for feature extraction and extension, we observed an increase in overall performance of about 0.02–0.06.
In general, when comparing term sets considering the best sets of parameters (, and , see Tables 4 and 6), it is clear that smaller term sets show a better overall performance. For the remainder of our analysis we will focus on two term sets, depth = 4 and GO Slim leaves depth = 1, which will be referred to as depth and GOSlim respectively. These sets were chosen to cover both term set strategies and provide a reasonable size set without sacrificing too much performance. We will also from now on focus on refinement and indirect extension, since they represent the primary goal of finding areas of the ontology to extend. Considering time parameters we will use the best overall performers (setup : = 3, FC, tTT).
The next step in our experiment was to compare different features and feature sets. Table 7 presents the average and standard deviation f-measure values for all features and feature sets using our standard setup.
When using single features, the best performers are , and , with average f-measure values around 0.74 in the set and 0.69 in the set. When using sets of features, in the set the top performers are indirect, and , with values between 0.75 and 0.76, whereas in they are , and , with values between 0.77 and 0.78. Using feature sets insetad of single features has a positive impact on performance in the set, which is not noticeable in the set.
So far we have focused on predicting refinement within the biological process ontology. Tables 8 and 9 summarize the results obtained for the molecular function and cellular component hierarchies, showing the top three features and feature sets for each term set. For molecular function we show only results for the term set based on since there is no subset.
Although average f-measure is generally lower for both molecular function and cellular component, than for biological process, and continue to be among the best features. Furthermore, for cellular component the set shows a worse overall performance than the set, in disagreement with what happens in biological process.
In addition to Decision Tables, chosen due to their simplicity, we also tested several other commonly used supervised learning algorithms, namely Naive Bayes, SVM, Neural Networks (Multilayer Perceptron) and Bayesian Networks, using their WEKA implementations. Figure 2 shows a plot for precision and recall for the best feature sets using these algorithms.
When applying different learning algorithms, we still see that overall biological process has the best performance, followed by molecular function and cellular component. Likewise, the general performance in the term set is better than the one in the depth term set for biological process, whereas it is the reverse for cellular component.
Looking in with more detail at the biological process results, the difference between feature sets is small, so we will not distinguish between them in our analysis. Naive Bayes gives the top precision values (0.87–0.90) but the lowest recall (0.48–0.57), whereas Bayesian Networks have the highest recall (0.78–0.79) with precision values between 0.74 and 0.79, which correspond to average f-measures between 0.76 and 0.79. SVM, Decision Tables and Multilayer Perceptron have performances in between these with both recall and precision values clustered around 0.70.
In molecular function, the highest precision is given by Multilayer Perceptron at 0.70 for , and Multilayer Perceptron, SVM and Naive Bayes for at 0.66–0.67. The highest recall is found in by Bayesian Networks at 0.83. Best average f-measure is achieved by SVM at 0.66 for both and .
In cellular component, there is a marked difference between the performance in the depth term set and in the set, with the latter having in general a much lower recall, around 0.40, except when using Bayesian Networks, where recall rises to around 0.7, but at the cost of precision. There is also a visible difference between term sets, with having in general a lower precision for the set, which is not apparent in the depth term set. In the depth term set the best performing algorithms are Decision Tables and Bayesian Networks, with recall around 0.8 and precision above 0.6. Decision Tables achieves the top performance with an average f-measure of 0.72 for .
To provide a basis for comparison, we implemented Stojanovic's browsing uniformity measures [25] and evaluated them on predicting ontology evolution for GO. For link usage we used annotation frequency. Since this strategy does not identify targets for extension, but rather ranks classes according to their uniformity, we evaluated this strategy plotting precision-recall curves for all ontology versions used. Figure 3 shows precision/recall plots for children uniformity, using one version of the ontology to calculate uniformity and predicting refinement for a following version in our dataset, alongside the plots for our prediction strategy best configuration (, ). For both cases we used the term set based on a depth of 4 and a distance between training and testing of two versions.
For plotting our strategy instead of relying on the binary labels output by the classifier, we used the probabilities for each instance to be true (i.e. refined), so that the generated plots are more directly comparable to those produced by the uniformity strategy, allowing a more granular calculation of precision at different recalls to allow for a threshold based evaluation. Consequently, the presentation of the results of our strategy in these plots differs from the presentation in previous tables.
The prediction results for all ontologies were combined together in the plotting of the Precision/Recall plots to provide a better visualization of results. As it is patent in the plots, our strategy has a considerably improved performance in all three GO ontologies, with curves closer to the top right corner, which are indicative of both higher precision and recall. The uniformity strategy performed worse in all cases, except at higher recall values in molecular function.
The other uniformity strategies (parents and siblings) have an even lower performance than that of children uniformity.
Change capturing through prediction of ontology extension is a complex issue, due to the inherently complex nature of ontology extension itself. Ontology extension can be motivated by implicit or explicit requirements, which have very different mechanisms. Implicit requirements are in principle easier to predict since they do not change between ontology versions, whereas explicit requirements, which are created by experts to adapt the ontology to a novel conceptualization or change in the domain, are much harder to predict. Our strategy, by virtue of being based on learning using past extension events, cannot distinguish between these two types, and thus attempts to predict extension regardless of it being motivated by implicit or explicit requirements. To capture both kinds of requirements we use a set of ontology based features that not only contemplate intrinsic features, such as structural ones, but also extrinsic ones, such as annotations and citations.
The assumption that extension can be predicted based on existing knowledge, either in the form of the ontology itself or its usage, is acceptable regarding the more common extension events, but is not applicable to extension events that are the result of deep restructuring or revision of existing knowledge. These extension events are part of a complex ontology change that also includes deletions and modifications. As such, these more complex changes are not the object of our strategy. In fact, one of our strategy's goals is to speed up the process of accomplishing the simpler extensions, to give experts more time and resources to focus on the more complex events.
One very relevant aspect of our evaluation strategy is that we compare our results to the real extension events that occurred in more recent versions of the ontology. This means that although some of our predictions are conceptually correct, they may not have yet been included in the ontology version used for testing and will thus be considered incorrect. This will have an impact on precision values, since we might be capturing needed but still unperformed extensions, and then be considering them to be incorrect in our evaluation. Due to this line of thought, we might then give preference to strategies that increase recall even if at the cost of precision. However, this could have the negative effect of including many incorrect predictions in our output, which is not desirable in a semi-automated ontology extension system. As such we have chosen to base our evaluation on f-measure, to provide balanced precision and recall.
A basic requirement of our strategy is to be able to access several versions of the ontology to consider. The minimum set of ontology versions it requires is two: one which will be used to calculate the features, and a second one, more recent than the first, from which we will extract the class labels to train the model. It then becomes crucial to define the interval between the versions to use. In our test case using the Gene Ontology we decided on versions with an interval of at least 6 months, based on the intuition that a smaller interval would not provide us with sufficient extension examples to be able to train a model. This intuition was shown to be a good approximation, since as seen in Text S1 and Table S1 in Text S1, when using monthly versions we do in fact have a very low number of positive examples.
Due to the complexity of ontology extension, particularly in such a large ontology as the Gene Ontology, our prediction strategy has to account for several parameters that help circumscribe our effort. One such parameter, extension type, was designed to capture the different types of extension: refinement and enrichment. We have found that refinement is considerably easier to predict than enrichment, with refinement having a greater average f-measure by between 0.3 and 0.7. There are two likely explanations for this difference: on one hand, there are many more refinement events between ontology versions than there are enrichment events (see Table 5), which will provide a better support for supervised learning; on the other, the features used may be better correlated to refinement than to enrichment.
Another parameter related to extension, is its mode, direct or indirect. Predicting direct extension, i.e. exactly which terms will be extended in a future version, should be the ultimate goal of an ontology extension prediction strategy. However this was proven to be a difficult task, which is unsurprising given the multitude of different processes that can lead to extension, and also the fact that on average new terms correspond to about 5% of all terms in an ontology version (see Table 1). This follows the trend found in our previous work [24], where we analyzed the extension of GO and found that insertions of new terms often occur together.
To address this issue we focused our prediction efforts in slices of the ontology, and defined the extension that happens within the subgraphs rooted in terms within these slices as indirect extension. Focusing only on the term sets thus defined greatly improved the performance of our strategy (Table 4), with average f-measures for the prediction of refinement of biological process increasing from 0.49 to 0.65–0.86 depending on the term set considered.
Predicting for a subset of the ontology is supported by our previous finding [24] that extension frequently happens by branches and that introducing terms closer to the root has a large impact on the overall structure of the ontology. Consequently, determining which term sets to use must be a compromise between enough specificity to be useful, but enough generality to provide a good enough balance of positive and negative examples. We determined six such subsets, following two distinct approaches: based on distance to root and based on GO Slim general.
We chose distance to root for its simplicity in creating a middle layer of GO terms. However, since terms at the same distance to the root do not always have the same degree of specificity, we also used GO Slim general as a basis for our other strategy. By using GO Slim general we were attempting to capture a similar degree of specificity among terms, specific enough to provide a useful prediction and general enough to allow for branch extension prediction. We tested three different sets within each approach, each yielding different term set sizes. Since molecular function does not have a GO Slim general, we only tested distance to root () based sets.
For both approaches, the smaller the data set the better the results. This can in did be due to the fact that in smaller data sets there is a better balance of positive and negative instances, which despite our use of SMOTE to balance the training sets, still has an impact on training the models. However, we are not interested in very small term sets, since they would not provide enough specificity to change capturing for ontology extension. Considering this we focused on the term set defined by terms at a distance of one from GO Slim leaf terms, which corresponds to an average term set size of 1189 for biological process and 758 for cellular component, and on the term set defined by terms at a distance of four to the root, which corresponds to sizes around 370, 460 and 100, for biological process, molecular function and cellular component respectively.
The final parameters in our strategy are those related with time: , FC and TT. We found that the influence of the number of versions used to derive the features was minimal. Regarding the intervals between versions for feature and class, and for training and testing, we found that increasing those intervals from six months to one year resulted in an increase in performance (about 0.03 to 0.06), which is likely due to the fact that the number of positive examples is larger when considering a larger interval between versions. Considering these results, we focused on the setup of = 3, FC and TT.
Although the parameters previously discussed represent the basis of our strategy, by defining exactly on what the prediction is focusing, it is the features used to support prediction that are essential to be able to capture extension events. Using the best parameter setup we investigated a set of thirteen single features, also arranged into eight sets, and found some interesting trends. In the term set, the single features and were among the top performers for the three GO hierarchies. But in the for biological process feature sets performed better than single features, whereas in cellular component this difference was not apparent. However, the feature sets composed of the best single features ( and ) were shown to provide the better performances across the board, with the exception of the set in cellular component. It is interesting to note that although using just structural or just annotation based features can provide in most cases a performance comparable to combining them, which can simplify our strategy, using a combination of the best single features can in some cases improve performance.
One of the most obvious patterns we get from these results is that terms with a lot of children terms or a lot of total annotations tend to be extended. It is arguable that for larger subgraphs, the probability of an extension event occurring is greater, given that there are more terms in it. However, to support the theory that the only factor involved is indeed the number of terms in the subgraph (i.e allChildren), we would have to consider that the probability of extension for any given term is equal. Intuitively, this does not appear to be a valid assumption, since it would mean that the extension of GO does not follow any particular direction. Nevertheless, we investigated this possibility by comparing the distribution of real refinement events for intervals, with the probability density function of a binomial distribution for at least one success for the same intervals. Figure S3 shows that the two distributions are significantly different, thus supporting the notion that although the number of children has an influence in the refinement probability, the probability of refinement is not the same for all terms. From these results we can hypothesize that the number of children a term has is related to its probability of refinement, because it reflects an increased interest in that area of the ontology.
Furthermore, the total number of annotations is influenced by the total number of children, since the annotations of the children contribute to the total number of annotations of the parent. To take this into account, we created the feature ratioAll to mitigate the influence of the number of children on the annotation data. Although this resulted in a decrease in f-measure of around 6%, compared to either feature separately, it is still a better performance than most other features. This gives further support to the notion that areas which attract a larger interest (in this case patent in the number of annotations) become the object of more refinement events.
Although these simple notions appear quite intuitive, and we could in principle derive a simple generic rule based on the number of children, in order to support automated change capturing, we need to establish the best separation possible between targets and non-targets for refinement, which is best achieved by employing supervised learning.
The results discussed so far were all based in Decision Tables, a simple supervised learning algorithm. We also tested other algorithms, but realized that although other algorithms such as SVM, Neural Networks and Bayesian Networks were capable of providing a better performance, and specifically in the case of SVM and Neural Networks of being parametrized to privilege either precision or recall, Decision Tables was still able to provide generally good results comparatively, without requiring parameter optimization.
We were particularly interested in the performance of Bayesian Networks, since our attributes are not independent, but in fact are temporally related when we consider multiple ontology version for feature extraction. For instance the value of in one version depends on its value in the previous one. However, we did not find a marked difference between Bayesian Networks and other approaches, so this dependency appears to not be very relevant for our current strategy.
Another particularly interesting aspect is that most machine learning algorithms, including the ones that were used, assume that instances are all independent and identically distributed. However, the dataset instances correspond to GO terms which are hierarchically related through the GO structure. Although the inclusion of features that describe the neighboring area tried to capture this aspect (e.g. siblings, and all the uniformity features), we still believe it was not properly contemplated by the proposed setup. The hierarchical relations between instances may be affecting the experiments considering the full set of terms, since they are not being captured by the representation. In the subset of terms dataset, their influence would not be as strong, since there are fewer hierarchical relations between instances.
To complete our evaluation, we compared our strategy to the one proposed by Stojanovic et al. [28] based on uniformity. In general, the uniformity based strategy performed worse than our own. This however is a consequence of Stojanovics approach having been designed to support the manual extension of an ontology that adapts to user's needs, whereas in our setting we have an ontology that models knowledge about a domain, whose extension is caused by many different aspects. Curiously, when transforming the uniformity metrics into features for classification, we achieve a better performance (Table 7) than when using them as intended by the authors, as a simple criteria for ranking.
The output of our extension prediction methodology is a list of ontology classes, which are the roots of subgraphs that correspond to ontology areas which have been predicted as good candidates for extension. Our methodology is applicable to the most simple yet most frequent type of ontology change, the addition of new elements. It is not suited to predict more complex changes such as a reorganization of an entire branch of the ontology. As such, the ontology extension prediction can be used to speed up the process of extension in these simpler cases, by allowing ontology developers and/or ontology learning systems to focus on smaller areas of the domain. This frees the experts to spend more time focusing on the more complex changes that cannot be predicted.
Automated ontology learning systems can also use the list to focus their efforts on the identified areas. For instance, most ontology learning systems employ a corpus of scientific texts as input, and their performance is tightly coupled to the quality of such corpora. If our candidate list is used to guide the creation of specific corpora for the areas to extend, it can have a positive impact on the performance of such strategies.
We have chosen to highlight three examples of the results given by our ontology extension prediction system, two successful ones (Figure 4 and Figure 5), where the predicted areas were in fact extended in the version for which extension was predicted, and one indirectly successful one (Figure 6), where although the extension did not occur when predicted, it did in fact happen at later versions of the ontology.
In Figure 4, extension was predicted for the subgraph rooted in “macromolecular complex assembly". Since we are predicting indirect extension, the addition of new subclasses can occur at any point in the subgraph. In this case, the GO term has four direct subclasses, and all of them gained new subclasses in the future version for which we were predicting. In Figure 5, extension was predicted for the area of “cell pole". In the version used to train the model, “cell pole" had two subclasses “apical pole of neuron" and “basal pole of neuron" but in the version for which extension was predicted, “cell pole" gained a whole new branch rooted on a new subclass for “cell tip". These two examples showcase two different extension patterns: in the first, extension occurs throughout the subgraph, whereas in the second it corresponds to the addition of a single but large branch.
In Figure 6 extension was predicted in the subgraph of “lipid transporter activity" for the version of January 2010, but no extension took place. However in later versions of July 2010 and January 2011, extension did occur by the addition of two new sub-subclasses. This is an example of how our evaluation strategy may be too stringent when considering these cases false positives, since they can eventually undergo extension at later versions.
Ontologies are crucial to handle the challenges of an increasingly data-driven world. However, ontologies themselves face this challenge, since the effort to keep them updated in face of the new knowledge that is produced on a daily-basis is never complete. To support this effort, some of the processes involved in ontology evolution can be automated, in order to reduce the time and resource investment made by expert curators.
In this work we present such a strategy for the first step of ontology evolution: change capturing. Our strategy is based on predicting areas of the ontology that will undergo extension in a future version, by applying supervised learning over features of previous ontology versions. We applied our strategy to the Gene Ontology, where we obtained encouraging results with average f-measure reaching 0.79 for prediction of refinement for a subset of relevant biological process GO terms.
In addition we defined a framework to better define extension in an applicational context, that can be applied to ontologies with versioning, as is the case of OBO ontologies and many of its candidates. This framework is crucial to provide a better understanding of the various nuances of ontology extension, and as such support ontology extension prediction efforts.
We find that two particular characteristics of our strategy can be improved, namely the selection of ontology versions to use and the selection of the term set. Both of these can benefit from recent works on ontology evolution [31], [38] from which we can gather useful information to guide the selection process. For the ontology versions, as we have discussed above, there is a need for a minimum of changes between versions to allow for the training, and by using these works we can pinpoint ontology versions that have enough changes between them. In what concerns the term set, we can benefit from the identification of stable and evolving regions of the ontology, and thus dynamically define distance to root based on this criteria, i.e. for stable regions we predict for terms further away from the root, whereas for evolving regions we stay closer to the root.
Although we applied our strategy to the Gene Ontology, it is applicable to any ontology with multiple versions available, which is becoming increasingly prevalent, as ontologies in biomedicine mature. The performance of our strategy on other ontologies is still to be tested and the next logical testing ground for the proposed methodology are smaller ontologies which lack the maturity and funding of larger ontologies such as GO. Several ontologies would be interesting to explore, such as the Pathway Ontology or the Ontology of Physics for Biology, which provide several versions but are much more recent and quite smaller than GO. The success of our strategy on GO using simple structural data is encouraging, since most ontologies lack such a rich annotation corpus as GO's, but all provide structural data which can be explored.
Predicting the extension of an ontology can have a positive impact in ontology evolution processes, be they manual or automated, by focusing efforts and reducing the amount of new information that needs to be processed. Moreover, OBO's principles of maintenance and orthogonality strongly advocate for the existence of a single ontology for each domain that is progressively enhanced, rather than a myriad of niche ontologies. Consequently, strategies that aid in the evolution of existing ontologies, as the one proposed here, present themselves as relevant contributions to the end goal of ontologies in biomedicine.
|
10.1371/journal.pntd.0006984 | Novel mass spectrometry based detection and identification of variants of rabies virus nucleoprotein in infected brain tissues | Human rabies is an encephalitic disease transmitted by animals infected with lyssaviruses. The most common lyssavirus that causes human infection is rabies virus (RABV), the prototypic member of the genus. The incubation period of RABV in humans varies from few weeks to several months in some instances. During this prodromal period, neither antibodies nor virus is detected. Antibodies, antigen and nucleic acids are detectable only after the onset of encephalitic symptoms, at which point the outcome of the disease is nearly 100% fatal. Hence, the primary intervention for human RABV exposure and subsequent post-exposure prophylaxis relies on testing animals suspected of having rabies. The most widely used diagnostic tests in animals focus on antigen detection, RABV-encoded nucleoprotein (N protein) in brain tissues. N protein accumulates in the cytoplasm of infected cells as large and granular inclusions, which are visualized in infected brain tissues by immuno-microscopy using anti-N protein antibodies. In this study, we explored a mass spectrometry (MS) based method for N protein detection without the need for any specific antibody reagents or microscopy. The MS-based method described here is unbiased, label-free, requires no amplification and determines any previously sequenced N protein available in the database. The results demonstrate the ability of MS/MS based method for N protein detection and amino acid sequence determination in animal diagnostic samples to obtain RABV variant information. This study demonstrates a potential for future developments of rabies diagnostic tests based on MS platforms.
| Although rabies is almost always fatal after the symptom onset phase, it can be prevented by timely administration of post-exposure prophylaxis (PEP), which involves passive antibody transfer and vaccination. One of the primary laboratory confirmatory tests for RABV infection is antigen detection, directed against the RABV encoded N protein using anti-N protein specific antibodies, in central nervous system (CNS) tissue samples of animals. This immuno-microscopy based detection utilizes either fluorescent tags (direct detection) or chromogenic substrates (indirect) in brain impressions from animals in which rabies is suspected. In this study, we explored the detection of N protein by a novel mass spectrometry (MS) based method that is label-free and does not require target amplification. The MS method specifically detected N protein in brain tissue and identified RABV variants based on amino acid sequence information. To our knowledge, this is the first report of an N protein detection method that does not utilize either antibodies or microscopy. This method provides an alternative platform for the development of future rabies diagnostic tests.
| Mass spectrometry (MS) based proteomics is a collection of rapidly evolving techniques being utilized as a diagnostic tool for infectious diseases. MALDI-TOF MS (matrix-assisted laser desorption/ionization–time-of-flight) is being used in clinical microbiology laboratories to identify an impressive spectrum of bacteria and fungi from infected specimens (generally enriched by culturing prior to detection) [1, 2]. At present, there is no comparable MS-based approach for identification of viruses. As the majority of viruses encode only a few proteins, most of the proteome in diagnostic samples like plasma or tissue samples will be dominated by the host [3]. However, if viral proteins accumulate at high concentration or there was shutdown of host protein synthesis, it might be feasible for direct detection of viral proteins from diagnostic samples. As the primary diagnosis in post-mortem animals for rabies relies on viral protein detection in infected samples, we explored MS as a diagnostic option. In this study, we explored LC-ESI-MS/MS (liquid chromatography–electrospray ionization tandem mass spectrometry) for peptide mass and amino acid sequence determination.
Rabies is an ancient disease known to humanity for over 5000 years. Initially, clinical signs exhibited by animals such as excessive salivation, aggression, choking or gagging were used to diagnose if the animal was rabid [4]. Once the causative agent of the disease was determined to be a viral infection, diagnosis relied on the ability of brain homogenate from suspect animal to cause infection in naïve animals or cell cultures [5–8]. Both methods required several days to weeks to confirm a RABV infection. The major breakthrough in rabies diagnosis was achieved after the invention of microscopic methods and histostaining, although initial methods still depended on non-specific staining reagents like Sellers stain to identify and differentiate Negri bodies, intra-cytoplasmic inclusions in the brain tissue of infected animals [9]. The inclusions were later characterized as a ribonucleoprotein complex (RNP), predominantly comprised of nucleoprotein (N protein) expressed from the viral genome. RABV, the causative agent of rabies, is a bullet-shaped virus belonging to the Lyssavirus genus. Lyssaviruses are in the order Mononegavirales and the family Rhabdoviridae, characterized by a 12 kilobase unsegmented negative-sense RNA genome. The genome encodes five proteins starting with the N protein closest to the 3’ end, followed by the phosphoprotein (P protein), the matrix protein (M protein), the glycoprotein (G protein) and the large RNA dependent RNA polymerase (L protein) [10]. These proteins are differentially expressed, the genes closer to the 5’ end of positive sense RNA (or 3’ end of negative sense genome) are expressed at higher levels compared to downstream genes (farther from 5’ end). The N protein mRNA transcripts are transcribed at higher levels to make it the most abundant viral protein synthesized after RABV infection [11]. N protein along with P and L proteins coat the viral genome to form the RNP, which accumulates in the cytoplasm of infected cells as large or granular inclusions [12].
Once immunological methods were developed, antibodies generated against RABV or RNP, comprised predominantly against the N protein, were utilized for specific detection of RABV antigens (proteins expressed from RABV genome). The most widely used rabies diagnostic method, the direct fluorescent antibody (DFA) test or fluorescent antibody test, detects the N protein in brain tissue impressions with polyclonal or monoclonal antibodies (mAbs) directly conjugated to fluorescent compounds [13–15]. The DFA test is considered the gold standard in rabies diagnostics for the detection of RABV antigen in animal brain tissues suspected of rabies [15, 16]. Alternatively, modified chromogenic-based detection methods, such as the direct rapid immunohistochemistry test (DRIT), have also been developed for rabies diagnosis using anti-N mAbs or polyclonal antibodies without the need for fluorescent microscopy [17, 18]. Non-microscopy based protein detection techniques for rapid, point of care diagnostics, such as the rabies immunochromatographic diagnostic (RID) tests, commonly known as lateral flow assay (LFA) are available. LFAs have provided mixed results with concerns on specificity and sensitivity of N protein detection [19–21]. The assay still relies on using a combination of antibodies specific to N protein and the ability to capture and detect protein in infected tissue lysates, which can be visualized by a colored band on LFA strips. Currently, these tests are not yet approved for regular diagnostics by the World Health Organization, but are helpful in countries where surveillance is lacking and for epidemiological studies [19].
In this study, we explored MS as an alternative method for rabies N protein detection. MS is an unbiased proteomics approach, non-amplifying, non-sequence specific technique and does not require specific reagents for protein detection [3, 22]. We demonstrate detection of all RABV encoded proteins in purified or crude infected cell lysates by MS. Additionally, N protein was detected and the amino acid sequence was determined by MS/MS-based peptide fragment mass information from animal diagnostic samples for several RABV variants circulating in the United States (U.S.).
BSR (a clone of Baby Hamster Kidney 21 cells) or mouse neuroblastoma (MNA) cells (CDC collection) were cultured in E-MEM supplemented with 10% FBS (Fetal bovine serum) containing antibiotics (Penicillin and Streptomycin) and antimycotic (Amphotericin B) essential vitamins and L-glutamine. BSR cells were infected with the RABV ERA (Evelyn Rokitnicki Abelseth) virus strain at 0.01 multiplicity of infection for 2–5 days. The media supernatant was harvested, subjected to low speed centrifugation, followed by sucrose density gradient centrifugation to purify ERA virus particles. MNA cells (T75 flask) were either mock-infected or infected with RABV CVS-11 (challenge virus strain, 10 X TC ID50) for 24 h, washed with PBS and harvested by centrifugation. The cell pellet was resuspended in 1X NuPAGE LDS sample buffer containing reducing agent (ThermoFisher).
Brain samples from animals (CDC collection) were submitted by the state public health, state veterinary, and US Department of Agriculture rabies laboratories for confirmatory testing and antigenic typing. Acetone-fixed brain impressions were tested by the standard DFA using the pre-calculated optimal working dilutions of two FITC anti-rabies mAb conjugates (Millipore Sigma Light Diagnostics and Fujirebio Diagnostics) as per the National Standard Protocol for rabies diagnosis (https://www.cdc.gov/rabies/pdf/rabiesdfaspv2.pdf). The two mAbs cocktails have different epitope recognition and affinity/avidity differences are required for DFA confirmatory testing. Similarly, a non-rabies antibody FITC conjugate, negative control reagent (Millipore Sigma Light Diagnostics), which contains the same IgG isotypes as the rabies specific antibody, is used for specificity. All controls (positive and negative) demonstrated the expected results. Samples demonstrating the presence of rabies-specific antigen in brain impressions (typical 4+ sparkling apple-green fluorescent inclusions with both anti-rabies conjugates and no specific fluorescence demonstrated with the non-rabies conjugate) were reported as positive. Based on the level of N protein specific staining, the samples are classified as 1+ to 4+ distribution, where 4+ demonstrate maximum staining (https://www.cdc.gov/rabies/pdf/rabiesdfaspv2.pdf). While most samples have either 3+ or 4+ antigen levels, around 5% demonstrate lower levels of N protein staining by DFA. For validation with MS assay, we included samples from 1+ to 4+ antigen distribution. If no specific rabies fluorescence was observed in the impressions, the samples were reported as negative.
All the RABV positive samples used in this study were subjected to antigenic typing to obtain variant information. Antigenic typing was performed on DFA positive brain samples by indirect fluorescent antibody tests using a panel of twenty mAbs against the RABV N protein (anti-N mAbs). RABV variants were determined by the demonstration of established reaction patterns of terrestrial and bat RABVs based on the recognition of N-protein epitopes by a panel of twenty CDC anti-N mAbs as previously established [23]. For MS assay, based on the availability, samples with different levels of antigen distribution, infected with different RABV variants that circulate in the U.S., and from different host species were selected.
Purified RABV and lysates from control and RABV infected cells were lysed in 1X NuPAGE LDS sample buffer with reducing agent and boiled at 95°C for 10 min. Infected and uninfected brain tissues (26 samples) were homogenized in 1x PBS (100 mg in 150 μl) followed by boiling with LDS sample buffer at 95°C for 10 min. Proteins were separated in a NuPAGE 4% - 12% Bis-Tris protein gels (proteins are separated under denaturing and reducing conditions) using NuPAGE MES running buffer and stained with Imperial protein stain (ThermoFisher). Either entire lane (12 slices) or specific portions of the gel were excised and processed for in-gel tryptic digestion. The amount of tissue homogenates subjected to electrophoresis corresponds to about 0.6 μg to 2 μg (1 μl– 4 μl) of the samples.
Gel slices were processed as follows. They were cut into 1 mm x 1 mm cubes followed by three washes of 50% acetonitrile, 10 mM ammonium bicarbonate and dried in a SpeedVac concentrator. Gel pieces were reduced for 60 min at 37°C using 10 mM 1,4-dithiothreitol in 10 mM ammonium bicarbonate and alkylated at room temperature in the dark using 55 mM iodoacetamide in 10 mM ammonium bicarbonate. Gel pieces were again washed three times with 50% acetonitrile, 10 mM ammonium bicarbonate and dried in a SpeedVac concentrator. Samples were rehydrated with sequencing grade modified trypsin (Promega) in 10 mM ammonium bicarbonate and allowed to digest over night at 37°C. The supernatant was collected and gel slices were washed three times with 60% acetonitrile, 10 mM ammonium bicarbonate to extract tryptic peptides. The washes and supernatant were collected, combined, and were dried in a SpeedVac concentrator. About 10%– 25% of tryptic digests obtained from gel slices were subjected for MS analysis.
Electrospray ionization (ESI) mass spectrometric analysis was performed using a Bruker model maXis ESI-Q-TOF instrument interfaced with a CaptiveSpray ESI spray source (Bruker Daltonics) to perform liquid chromatography–tandem mass spectrometry (LC-MS/MS) using a U3000 RSLCnano HPLC configured for nl/min flows. The Dionex U-3000 RSLCnano nanobore HPLC was configured with a binary nanoflow ultra-high pressure pump and a ternary high pressure microbore pump. The system used a pulled-loop autosampler configured with a 20 μl sample loop. A desalting trap column (0.3 mm x 5 mm, 5 μm C18 PepMap 120 Å, Dionex) was used and the analytical column used was a C18 PepMap (0.075 mm x 250 mm, 2 μm, 120 Å, Dionex). The solvents used were 0.1% formic acid in water (A) and 80% acetonitrile / 0.1% formic acid (B). The gradient was 2%– 55% B in 90 min. The eluent from the analytical column was introduced into the maXis using the Bruker CaptiveSpray source. The source was operated at a spray voltage of 1200 V with a drying gas of nitrogen flowing at 4 liters per min. The capillary temperature was set to 150°C. The mass spectrometer was set to acquire spectra of m/z 50 to 2500. MS/MS data was acquired in an automated fashion using a dynamic precursor ion selection based on the MS scan with precursor ion active exclusion for 60 s after at least 1 spectrum was acquired for each precursor ion. MS data was acquired at a scan speed of 10 Hz and MS/MS data was acquired at a scan speed of 2 Hz– 10 Hz depending on the intensity of the precursor ion. Total cycle time for acquisition of both MS and MS/MS scans was limited to 2.2 s. MS internal calibration was achieved by the use of a lock mass (HP-1222, Agilent Technologies).
The collected data was processed by DataAnalysis (Bruker Daltonics) to produce deconvoluted and internally calibrated data and saved as an xml peaklist, which was uploaded to our Proteinscape database (Bruker Daltonics). Proteinscape automatically submitted the peaklist to our in-house MASCOT server for searches against either the Swiss-Prot curated protein Fasta file or a taxonomic filtered data from NCBI’s RefSeq database. The taxonomic filters applied were human and virus.
N protein sequences from RABV variants utilized for this study were obtained from Genbank. The sequences were aligned using Clustal Omega software (https://www.ebi.ac.uk/Tools/msa/clustalo/) to obtain consensus. The peptide fragments sequenced by mass spectrometry were highlighted to demonstrate the conservation or differences observed across various RABV variants.
To evaluate the potential for mass spectrometric analysis to characterize RABV, purified RABV ERA variant was separated on a protein gel, stained with Imperial protein stain, and subjected to in-gel protein digestion followed by spectrometry (Fig 1A). In the first MS step, based on the experimental peptide mass detected, potential proteins (with theoretical tryptic peptide mass) were identified. In the second MS step, certain precursor peptide ions were subjected to low energy collision and fragmentation into product ions, namely “b” and “y” ions (representing N- and C- terminal fragments respectively) and masses are determined (Fig 1C). As multiple fragments were generated, by comparison of a panel of “b” and “y” ion fragment masses, it was possible to determine amino acid sequence. Based on the peptide mass fingerprint and fragment mass based amino acid sequence of peptides by MS/MS, all five RABV encoded proteins were identified by MS at expected molecular weight positions in the gel slices (Fig 1A). In addition, based on the amino acid sequence determined peptides (denoted in red), N protein present in the sample was clearly identified as RABV ERA variant (Fig 1B). The amino acid sequence coverage by MS/MS for other proteins and the number of unique peptides and percentage of coverage are provided (S1A–S1E Fig). Thus, with sufficient peptide concentrations, fragment mass guided amino acid sequence determination by MS/MS can differentiate and identify the RABV variant present in the sample. MS has the advantage to detect and perform sequence determination for potential variant typing in a single test.
To determine the specificity of MS for RABV detection, MNA cell lysates from mock or RABV infected samples were separated on SDS-PAGE, stained by Imperial protein stain, and the entire lane was excised into 12 slices, subjected to in-gel tryptic digestion and analyzed by MS (Fig 2). All five RABV proteins were detected specifically in infected cell lysates at expected molecular weight by MS (as indicated in the gel slices in Fig 2). None of the RABV encoded proteins were detected in any of the 12 slices from the uninfected cell lysate. N protein detection was the highest based on the number of unique peptides and percentage of amino acid sequence coverage, followed by P, M, G and L proteins (S2A–S2E Fig). In addition, based on amino acid sequence determination by MS/MS fragmentation, N, P and G proteins were identified as RABV CVS-11 variant. The number of unique peptides and amino acid sequence coverage are presented in S2F Fig.
Rabies diagnosis primarily focuses on detection of antigen (N protein) in CNS tissues of animals suspected for rabies. For DFA, touch impressions of brain tissues on slides are tested with anti-N protein specific mAbs or polyclonal Abs. For mass spectrometric detection, tissue homogenates were heat inactivated and separated in a 4%– 12% Bis-Tris protein gel. To limit the number of samples subjected to MS, only the portion of gel corresponding to either N or G proteins (based on the mobility of N and G protein bands in purified RABV ERA) were sliced after staining the gel (Fig 3A). N protein was detected by MS in samples 5 and 6, previously determined positive by DFA. All negative or inconclusive samples were also negative for RABV N protein by MS (Fig 3B). In addition, based on amino acid sequence of unique peptide fragments, both positive samples were correctly identified as either Eastern Raccoon (E. Raccoon) or Tadarida brasiliensis (bat) variants of RABV (Fig 3C, described in bioinformatic analysis results section). Although gel slices corresponding to the position of G protein mobility were subjected to MS, the G protein was not detected.
The limit of detection of RABV N protein was determined by spiking uninfected cell lysate with purified RABV ERA virus. Different amounts of purified ERA virus at 1.0 X 107 focus forming unit (ffu) / ml (from 4 μl to 0.01 μl) were added to an equal volume of uninfected cell lysate, and were separated by protein gel followed by staining with Imperial protein stain (Fig 4A). The gel position corresponding to N protein was sliced and subjected to MS. N protein was detected in five subsequent dilutions (or equivalent to 5.0 X 102 ffu viral particles) based on at least one peptide above the background values (Fig 4B). At higher dilutions, N protein variant information was not obtained by MS/MS fragmentation analysis probably due to low concentration of peptides. This demonstrates the concentration dependence of MS/MS method for detection of N protein and determination of RABV variants.
To compare sensitivity of MS for N protein detection with DFA, samples containing different amounts of antigen were tested. Although, DFA is not quantitative, the relative amount of antigen in tissue impressions can be obtained based on the fluorescence intensity and distribution after observation of all microscopic fields. For example, DFA results grade samples from a minimum (1) to maximum (4), with 4+/4+ being maximum fluorescent intensity and distribution, respectively. Generally, anti-N mAbs are titrated to obtain a maximum fluorescent intensity of 4+. Based on the distribution of antigen in microscopic fields, relative levels of N protein in samples are graded from 1+ to 4+. To assess the utility of MS to detect N protein from samples with different RABV variants observed in the U.S. from different species of infected animals and different amounts of antigen, a panel of samples previously analyzed by DFA were tested with the MS assay (S3A and S3B Fig). In general, MS positively identified samples that had higher N protein concentrations as determined by DFA (3+ or 4+ distribution), while samples with lower concentrations (1+ or 2+) were not detected with the instrument used (Table 1). In addition, from the amino acid sequence information obtained by MS/MS fragmentation, E. Raccoon and North Central Skunk (NC Skunk) variants were correctly identified by general database search in two samples (Table 1). Of all the samples tested, G protein was detected only in sample 17 by the MS method.
Sample–source denotes the CNS tissues of host animals, while RABV variant corresponds to samples from DFA positive CNS tissues determined by antigenic typing. E Raccoon–RABV Raccoon variant observed in Eastern U.S.; NC Skunk and SC Skunk–corresponds to RABV variants circulating in skunks from either North Central or South Central U.S., respectively. AZ gray fox and Arctic fox variants of RABV observed in Arizona and Alaska, respectively.
DFA–Intensity / distribution–corresponds to the semi-quantitative classification of N protein levels based on distribution of anti-N mAb-FITC conjugate on tissue impressions.
Mass spectrometry–Results demonstrate detection of N protein by MS and RABV variant corresponds to identification of RABV variants based on amino acid sequence determination and analysis.
MS/MS fragment mass based amino acid sequence determination has the potential to differentiate RABV variants if peptides containing unique sequences for RABV variants are determined. The list of peptides that contained amino acid residues unique to RABV variants are provided in S1 Spreadsheet. Even though several peptides were identified in both ERA and CVS-11 samples due to higher concentration of N protein in purified virus or experimentally infected samples, only certain peptides could differentiate RABV variants. For ERA, in the peptide VNNQVVSLKPEIIVDQHEYK, H was unique (bold and underlined) compared to other variants (Fig 5). Similarly, in the peptide TDVDGNWALTGGMELTR, aspartic acid (D) was unique to CVS-11 N protein sequence. In the subsequent analysis, RABV variants were determined from infected brain tissues corresponding to E. Raccoon, T. brasiliensis (Fig 3C) and NC Skunk RABV variants (Table 1), while variant information for other positive samples by MS was not obtained by default search results. To improve variant identification, we performed a manual search of peptide fragments based on the Clustal alignment of different N protein sequences for RABV variants used in the study (Fig 5). As the amino acid sequence identity ranged from 94%– 98% (Fig 6), unless peptides containing differences in amino acid sequences are ionized and detected, RABV variant information could not be determined. Identification of RABV E. Raccoon variant in sample 5 was based on amino acid sequence derived from the peptide “DPTIPEHASLVGLLLSYLYR” (Fig 7, S4 Fig). In this peptide, amino acid residues in position 4 ("I”) and position 5 (“P”) were unique to E. Raccoon variant. Two additional unique peptides “ELQDYEAAELTK” and “KPSISLGK” for E. Raccoon variant were identified in sample 11, in which “D” in amino acid position 4 and “S” in amino acid positions 3 and 5 were present only in this N protein sequence. Interestingly, in one of the peptides derived from NC Skunk variant, “QINLTAGEAILYFFHK”, the highlighted “G” in amino acid position 7 is unique and it replaces either a “K” or a “R” in other variants. Since the peptides are generated by proteolytic cleavage with trypsin which cleaves after”K” or “R” residues, this peptide is truncated in other variants leading to this longer peptide only being detected in the NC Skunk variant. Based on the peptide analysis from MS/MS results, South Central (SC) Skunk and Arctic Fox RABV variants were identified. Details of amino acid sequences of peptides utilized for RABV variant identification are provided in S1 Spreadsheet. Only the Arizona (AZ) gray fox variant was not resolved by MS/MS due to the absence of unique peptides. The list of all peptides derived from each sample is provided in a spreadsheet (S2 Spreadsheet). Based on this list, a set of 10 peptides (Fig 5, underlined and boxed) detected in tissue samples and frequency of detection (not considering MS results from purified ERA and infected cell lysate CVS-11) were presented in Table 2. Although, not directly comparable, one of the peptides at the N terminus “VNNQVVSLKPEIIVDQHEYK” encompasses the target for the recently validated LN34 real time RT-PCR [24].
Rabies is a zoonotic disease transmitted by the bite of a lyssavirus infected animal. Due to the lack of specific anti-virals or therapeutics, rabies is considered to have one of the highest case fatality rates for any of the infectious agents after symptom onset [16]. Unfortunately, human rabies diagnosis is not available during the pre-symptomatic phase, which encompasses the time from virus exposure to the establishment of viral replication in the brain [25]. The testing of suspected rabid animals for RABV infection is important to initiate post-exposure prophylaxis for rabies disease prevention in humans who have been bitten. Current rabies diagnostic reagents have been selected by comparison studies to be broadly reactive to highly conserved lyssavirus epitopes, and can be used to detect all of the RABV and lyssaviruses to date albeit at different levels of sensitivity. With new lyssaviruses being discovered, it will be necessary to have reagents that can detect N protein without compromising the specificity of detection. In addition, since the assay depends on the use of antibodies, variations in reagent batches and lots due to purification and conjugation procedures might affect the functionality of the assay. Due to the high expression level of N protein and the characteristic inclusions formed in the cytoplasm of infected cells, microscopy based methods are extremely reliable for rabies diagnosis. Specifically, the DFA is considered the gold standard for rabies diagnostics in animals [16]. Several modifications to detection methods, primary antibodies, and experimental procedures have been developed over the years to detect N protein. Two studies have attempted to determine metabolomics changes as a measure during the initial stages of rabies, however, it still requires additional data to be considered for routine diagnosis [26, 27].
Non-immunomicroscopy based detection of N protein has utilized rapid, point-of-care platforms, including lateral flow assays [19, 21, 28]. These assays rely on two different antibodies that react with N protein, one for antigen capture and the other for detection in diagnostics. Although several products are currently in the market, the results are mixed. The specificity and sensitivity varies with different kits including the limits of detection. The variability in sensitivity raises an important concern with rapid tests and the potential incidence of false negatives [19], that could result in inappropriate recommendations regarding the need for PEP. Although some reagents and protocols demonstrated a high level of concordance with the reference technique DFA [21] and could be utilized for resource limited areas, these assays requires additional confirmatory testing.
Direct detection and sequencing of proteins as a diagnostic for viral infection has not been widely attempted. As each protein has unique amino acid composition and peptide mass fingerprint, it can be used for the specific detection of target proteins. In addition, amino acid sequence of peptides can be determined by MS/MS fragment mass analysis, which further improves the confidence level for protein identifications. In this study, we first demonstrated the ability to detect and obtain RABV variant information for all five RABV encoded proteins by MS in purified RABV particles (Fig 1 and S1 Fig). The specificity of MS was next demonstrated using in vitro infected or uninfected cell culture lysates. RABV expressed proteins were detected only from the infected lysate and three of the five proteins had unique sequence information to classify as CVS-11 variant (Fig 2 and S2 Fig). RABV encompasses different variants, primarily based on the adaptation for efficient replication in host reservoir species. These RABV variants have distinct differences in amino acid sequences of encoded proteins as identified in seven major terrestrial wildlife hosts in the U.S. and territories. These include raccoons (E. Raccoon, in East??), Skunks (in NC and SC variants), foxes (AZ Gray, Texas Gray and Arctic [Alaska]), mongoose (in Puerto Rico), and at least 14 RABV variants associated with different bat species which are ubiquitous in the U.S. In this study, we demonstrate detection of N protein in clinical specimens from major RABV variants observed in the U.S.
The two major observations obtained from this study are, the ability (1) to detect N protein from crude tissue preparation without any amplification or N protein specific reagents and (2) to obtain amino acid sequence information for further confirmation and identification of RABV variants. Samples that were classified as 3–4+ antigen distribution by DFA, were predominantly positive by MS, while samples with lower N protein content were not sensitive enough for detection using the Bruker maXis QTOF instrument. Of the 18 DFA positive samples tested, N protein was detected by MS in 11 samples (61% sensitivity). As majority of samples have higher levels of antigen, the MS assay described in this study would have higher levels of sensitivity in actual field samples. In addition to confirmation of N protein in tissue samples, all but one of the RABV variants were determined based on amino acid sequence information. The current N protein detection method by immunomicroscopy requires two separate tests for RABV variant determination. Once RABV N proteins are confirmed by DFA, a panel of 20 anti-N mAbs that bind differentially to RABV variants are tested by IFA. Based on the pattern of reactivity of these anti-N mAbs and comparison with established terrestrial and bat variant patterns, the RABV variant is identified. In spite of high level of sequence identity exhibited by N protein, amino acid sequencing of unique peptides were able to differentiate RABV variants (Fig 5). This is particularly significant as only changes in nucleotides that results in amino acid change (non-synonymous substitutions) are detected in MS, compared to both synonymous and non-synonymous substitutions obtained by DNA sequencing. Based on MS/MS results, high-resolution peptide sequence analysis that can differentiate E. Raccoon, T. brasiliensis, NC Skunk, SC Skunk and Arctic Fox RABV variants were identified (Fig 5, Fig 7, S1 Spreadsheet). Thus, our results demonstrate for the first time N protein detection and sequencing by MS/MS directly from RABV infected animal CNS tissue samples. Although we also attempted to detect G protein in infected tissues, except for one sample, it was not detected by MS. The detection of N and not G protein, further substantiates the current diagnostic methods focus on N protein detection due to higher level of expression followed by RABV infection.
The MS method was successful in identifying several major RABV variants circulating in the U.S. and from several different host animal species. The results from this study provide information on peptides derived from the N protein that are readily ionizable and detectable by MS. Peptides conserved across all RABV variants or containing unique sequences for differentiation of variants could be employed for either species- or variant- specific MS-based assays (S1 Spreadsheet). Furthermore the development of a PRM (parallel reaction monitoring) type target based assay, would allow for multiplexing since the MS instrumentation can be programmed to scan for dozens of potential targets and for the selective quantification of proteins in samples [29, 30]. While this study focused on RABV variants that circulate in the U.S., MS analysis of RABV variants or non-rabies lyssaviruses observed in other countries will be helpful to generate a peptide database for species and variant specific detections.
In this study, we have explored MS as an alternative method and obtained preliminary data on the feasibility for rabies N protein detection. We will further explore the possibility to detect N protein directly from brain suspensions, instead of separating by protein gels using PRM methods to improve the sensitivity of detection compared to DFA. Although, MS-based assays may not be cost effective at present, it could be an additional antigen/protein detection method relying not on antibodies and microscopy at reference laboratories. With advancement in MS and increased use in clinical laboratories, it offers a next generation technology platform for exploring rabies antigen detection. Another major advantage of recent “-omics” technologies are identification of multiple pathogens or pathogen discovery from a single assay. As the MS-based assay is unbiased and does not need specific reagents, it can be employed for pathogen identification for samples with unknown etiology (negative for RABV, but exhibiting neurological symptoms) by “characterization of proteomics” similar to “metagenomics” approach.
|
10.1371/journal.ppat.1005420 | The Adenovirus E4orf4 Protein Provides a Novel Mechanism for Inhibition of the DNA Damage Response | The DNA damage response (DDR) is a conglomerate of pathways designed to detect DNA damage and signal its presence to cell cycle checkpoints and to the repair machinery, allowing the cell to pause and mend the damage, or if the damage is too severe, to trigger apoptosis or senescence. Various DDR branches are regulated by kinases of the phosphatidylinositol 3-kinase-like protein kinase family, including ataxia-telangiectasia mutated (ATM) and ATM- and Rad3-related (ATR). Replication intermediates and linear double-stranded genomes of DNA viruses are perceived by the cell as DNA damage and activate the DDR. If allowed to operate, the DDR will stimulate ligation of viral genomes and will inhibit virus replication. To prevent this outcome, many DNA viruses evolved ways to limit the DDR. As part of its attack on the DDR, adenovirus utilizes various viral proteins to cause degradation of DDR proteins and to sequester the MRN damage sensor outside virus replication centers. Here we show that adenovirus evolved yet another novel mechanism to inhibit the DDR. The E4orf4 protein, together with its cellular partner PP2A, reduces phosphorylation of ATM and ATR substrates in virus-infected cells and in cells treated with DNA damaging drugs, and causes accumulation of damaged DNA in the drug-treated cells. ATM and ATR are not mutually required for inhibition of their signaling pathways by E4orf4. ATM and ATR deficiency as well as E4orf4 expression enhance infection efficiency. Furthermore, E4orf4, previously reported to induce cancer-specific cell death when expressed alone, sensitizes cells to killing by sub-lethal concentrations of DNA damaging drugs, likely because it inhibits DNA damage repair. These findings provide one explanation for the cancer-specificity of E4orf4-induced cell death as many cancers have DDR deficiencies leading to increased reliance on the remaining intact DDR pathways and to enhanced susceptibility to DDR inhibitors such as E4orf4. Thus DDR inhibition by E4orf4 contributes both to the efficiency of adenovirus replication and to the ability of E4orf4 to kill cancer cells.
| The cellular DNA damage response (DDR) network interprets the presence of replicating viral DNA genomes as DNA damage and strives to repair it, leading to inhibition of virus replication. Many DNA viruses, including adenovirus, evolved mechanisms to inhibit the DDR, thus increasing the efficiency of virus replication. In this study we identify a novel mechanism used by adenovirus to inhibit the DDR. The viral E4orf4 protein, together with its cellular partner, the PP2A phosphatase, inhibits damage signaling by reducing phosphorylation of proteins belonging to different DDR branches. As a result, E4orf4 causes accumulation of DNA damage in the cells. Inhibition of the DDR regulators ATM and ATR, as well as expression of E4orf4, enhances infection efficiency. Moreover, E4orf4 sensitizes cells to killing by sub-lethal concentrations of DNA damaging drugs, likely because it inhibits DNA repair. These findings could provide one explanation for the previously reported ability of E4orf4 to induce cancer-specific cell death, as many cancers have DDR deficiencies leading to their increased reliance on the remaining intact DDR pathways and to enhanced susceptibility to DDR inhibitors such as E4orf4. Thus, inhibition of the DDR by E4orf4 contributes both to viral replication efficiency and to E4orf4-induced cancer cell killing.
| Genome integrity is constantly challenged by endogenous and exogenous agents that cause different kinds of DNA lesions. The cells have evolved a DNA damage response (DDR) which includes several mechanisms to detect and signal the presence of damaged DNA or replication stress, resulting in checkpoint activation and DNA repair, or if the damage is too extensive, resulting in senescence or cell death [1, 2]. Formation of DNA lesions is recognized by sensor proteins such as Poly (ADP-ribose) polymerase 1 (PARP-1) [3, 4], KU proteins [5], or the MRN complex consisting of the Mre11, Rad50 and Nbs1 proteins [6–9]. The sensors recruit proteins that transduce the signal to chromatin, to cellular checkpoints and to the repair machinery [10]. Signal transducers include the phosphatidylinositol 3-kinase-like protein kinase (PIKK) family, including ataxia-telangiectasia mutated (ATM), ATM- and Rad3-related (ATR), and DNA-PK (reviewed in [11, 12]).
Protein phosphatase 2A (PP2A) is composed of three subunits: the catalytic C subunit, a scaffolding A subunit, and one of several regulatory B subunits encoded by at least four unrelated gene families: PR55/B55/B, PR61/B56/B’, B”, and B”‘[13], which dictate substrate specificity of the PP2A holoenzyme [13]. PP2A plays an important role in the DDR by regulating the activity of PIKKs [14–18] and their substrates Chk1, Chk2, or γH2AX, which are part of the DDR signaling pathways [19–24].
During infection with DNA viruses, stretches of single-stranded DNA in replication intermediates and the ends of the double stranded viral DNA are perceived by the cell as damaged DNA, inducing the DDR (reviewed in [25, 26]). Some viruses may exploit the DDR to their advantage [27–29], however, the end result of DDR activation may lead to repair of viral genomes through their concatenation [30–32] and is detrimental to virus propagation. Therefore, many viruses, including adenovirus (Ad), have evolved mechanisms to inhibit the DDR. It has been previously reported that various types of Ad inhibit the DDR by ubiquitin-mediated degradation of DDR proteins including components of the MRN complex [6, 31, 33], DNA ligase IV [33, 34], p53 [33, 35, 36], TOPBP1 [33, 37], and others, or by removing MRN from viral replication foci [31, 38, 39]. These Ad effects are promoted by the E4orf3, E4orf6, and E1B-55K proteins. The cellular response to Ad genomes was recently reported to be biphasic, including a MRN-ATM-dependent DDR that is activated by early replicating virus genomes, is localized at viral replication centers and must be inactivated by the E1B-55K and E4orf3 proteins to allow viral replication; and a global MRN-independent ATM activation induced by viral nuclear domains that does not affect virus replication [40]. At early times after infection, the incoming viral genome is protected from the DDR by the adenoviral core protein VII [41]. The E4orf6 protein was also reported to inhibit double strand breaks (DSB) repair signaling through inhibition of PP2A leading to prolonged presence of γH2AX and PARP hyperactivation and resulting in enhanced apoptosis [42].
The Ad E4orf4 protein is a multifunctional viral regulator. Within the context of the virus E4orf4 contributes to temporal regulation of the progression of viral infection by down-regulating early viral genes and cellular genes affecting Ad replication, controlling alternative splicing of viral RNAs, and influencing protein translation [43–50]. When expressed alone E4orf4 induces cell death which is p53- and caspase-independent but frequently maintains crosstalk with classical caspase-dependent apoptosis [51–54]. The E4orf4 signaling network is highly conserved in evolution from yeast through Drosophila to mammalian cells [55–59], underscoring its importance to cell regulation. Moreover, E4orf4-induced cell death is more efficient in oncogene-transformed cells in tissue culture than in normal cells [60], indicating that study of E4orf4 signaling may have implications for cancer therapy. The basis for the cancer specificity of E4orf4-induced cell death will be better understood once knowledge of the underlying mechanisms is improved [61].
Studies of the mechanisms underlying E4orf4 action have revealed several E4orf4 partners, including the B55 subunit of protein phosphatase 2A (PP2A) [62]. PP2A is a major E4orf4 partner, and its interaction with E4orf4 was shown to contribute to all E4orf4 functions known to date [50, 60–64].
Here we report that E4orf4 provides a novel mechanism to inhibit DNA damage signaling during Ad infection. We show that E4orf4 expression leads to decreased phosphorylation of several proteins that participate in the DDR, in both ATM- and ATR-regulated pathways and causes accumulation of DNA damage. This novel E4orf4 function requires its interaction with the B55 subunit of PP2A. Inhibition of the DDR by E4orf4 has biological significance to Ad infection and to E4orf4-induced cell death.
It has been reported that Ad interferes with the DDR as part of its protection against host defense mechanisms, using the E1B-55K, E4orf6, and E4orf3 viral proteins [6, 31, 33, 39]. Since viruses usually use several different mechanisms to achieve a single aim, and since it has been shown that PP2A participates in turning off the DDR by dephosphorylating DDR proteins such as ATM, Chk1, and Chk2 [15, 17, 20–23], we set out to determine whether the Ad E4orf4 protein, a PP2A partner, contributes to down-regulation of virus-induced DDR. HeLa cells were either mock-infected or infected with the Ad mutant dl366*, lacking the E4 region, or with dl366*+E4orf4, lacking all E4 orfs except E4orf4 [65]. These mutants were used to facilitate the examination of E4orf4 effects when the MRN complex is unharmed by other E4 gene products and the DDR is activated [6, 31, 33, 39]. Protein extracts were prepared at various times after infection and subjected to Western blot analysis. The uniformity of infection was confirmed by finding similar levels of the E1B-55K protein in cells infected with the two E4 mutant viruses (Fig 1A). It should be noted that the E4orf4 protein was reported to reduce the levels of early Ad mRNAs in HeLa cells [44, 45], but E1B levels were affected less than those of other early RNAs [44]. Fig 1A demonstrates that infection with the mutant dl366* led to phosphorylation of several DDR proteins, including ATM, 53BP1, Smc1, Nbs1, Chk1, and Chk2, while these proteins were not phosphorylated in mock-infected cells. In contrast, infection with a similar multiplicity of infection (MOI) of dl366*+E4orf4 mutant Ad resulted in decreased phosphorylation levels of all tested DDR proteins belonging to both ATM- and ATR-regulated pathways. The reduced phosphorylation was confirmed by quantification of the protein bands and normalization to total protein levels. The E4orf4-induced decrease in DDR protein phosphorylation did not represent a global reduction of cellular phosphorylation as E4orf4 did not reduce Akt-S473 phosphorylation, but rather enhanced it. Thus the combined results suggest that E4orf4 reduces DDR induction by an E4-deficient mutant Ad. This conclusion was further validated by similar experiments carried out in non-transformed, primary HUVEC cells (Fig 1B and 1C), indicating that E4orf4 inhibited the DDR in both transformed and non-transformed cells. It should be noted that although phosphorylated H2AX (γH2AX) is a characteristic marker of DSBs, WT Ad infection was reported to result in high levels of γH2AX [66], despite the degradation of the MRN sensor complex and the reduced phosphorylation of downstream substrates. We therefore did not investigate γH2AX in the presence of E4orf4.
To further confirm the results indicating that E4orf4 reduced DDR protein phosphorylation, HeLa cells were infected with mutant viruses as described above and subjected to immunofluorescence staining with antibodies to the 53BP1 protein phosphorylated on S1778. Infected cells were identified by staining of the Ad DBP protein. Fig 1D demonstrates that cells infected with dl366* contained foci of phosphorylated 53BP1 in 75% of infected cells and these p53BP1-S1778-stained foci surrounded nuclear viral compartments in the infected cells (Fig 1E), similarly to other DDR proteins described previously [6, 31, 40]. In contrast, infection with dl366*+E4orf4 did not result in accumulation of phospho-53BP1 and foci containing p53BP1-S1778 were observed in only 6% of infected cells. Thus the results presented in Fig 1 support the conclusion that E4orf4 inhibits DNA damage signaling.
An interaction of E4orf4 with PP2A, mediated by regulatory B subunits, has been reported to contribute to all E4orf4 functions known to date [43, 47, 49, 60, 62–64]. Since E4orf4 appeared to reduce DDR protein phosphorylation (Fig 1), we examined whether this function required the E4orf4-PP2A interaction. Two experimental systems were examined: virus infection of cells in which the PP2A-B55 subunit could be inducibly knocked down, or drug treatment of cells expressing WT E4orf4 or a mutant that cannot bind PP2A. In the first approach, dl366* and dl366*+E4orf4 mutant viruses were used to infect the L11 cell line in which a PP2A-B55 shRNA can be transiently expressed by doxycycline (Dox) induction. Protein extracts were prepared 24 hrs post-infection and were subjected to Western blot analysis. Similarly to our previous observations in HeLa and HUVEC cells (Fig 1), infection of L11 cells with dl366* in the absence of Dox resulted in DDR activation manifested by enhanced phosphorylation of three sites in the Chk1 protein (S317, S345, and S296) that were barely phosphorylated in mock-infected cells. Infection with a similar MOI of dl366*+E4orf4 resulted in decreased phosphorylation of these sites without a similar reduction in Chk1 protein levels (Fig 2A, lanes 1–3). However, when Dox was added to the cells 72 hrs before infection, PP2A-B55 protein levels were dramatically reduced, and Chk1 phosphorylation was not affected by E4orf4 (Fig 2A, lanes 4–6). To ascertain that the reduced effect of E4orf4 on DDR activation was indeed the result of PP2A-B55 knockdown, a plasmid encoding HA-tagged PP2A-B55 which contained silent mutations rendering it resistant to the shRNA was transfected into control and Dox-treated L11 cells one day before infection. In the presence of exogenously expressed PP2A-B55, E4orf4 reduced Chk1 phosphorylation much better both in the presence and absence of Dox (Fig 2A, lanes 7–12), indicating that PP2A-B55 knockdown was indeed the cause of the diminished ability of E4orf4 to reduce Chk1 phosphorylation. Because the HEK293-derived L11 cells contain endogenous E1A and E1B proteins, E1B-55K staining could not be used in these cells to test the similarity of infection. However, a parallel infection of W162 cells with the same MOI revealed similar levels of E1B-55K protein in both dl366* and dl366*+E4orf4 virus infections (S1 Fig), confirming the uniformity of infection. The effect of E4orf4 on ATM phosphorylation was very small in the HEK293-derived L11 cells, and was therefore not studied in these cells.
An additional approach was utilized to confirm the findings indicating that E4orf4 must bind PP2A to reduce DDR protein phosphorylation. Cells containing Dox-inducible WT E4orf4 (clone 13) or the R81F84A E4orf4 mutant that does not bind PP2A (clone 3) were induced with Dox for three hrs, or were left uninduced, and were then treated with hydroxyurea (HU), which induces replication stress or with neocarzinostatin (NCS), a drug that induces DSBs. Protein extracts were prepared and subjected to Western blot analysis. As seen in Fig 2B, HU or NCS treatment caused Chk1 phosphorylation in both clone 13 (lanes 3,5) and clone 3 cells (lanes 9,11). However, expression of WT E4orf4 led to reduced Chk1 phosphorylation in clone 13 cells, especially on S317, and following NCS treatment on S296 (lanes 4,6) whereas in clone 3 cells expression of the R81F84A mutant did not alter Chk1 phosphorylation on S317 and altered S296 phosphorylation much less efficiently (lanes 10,12). The dissimilarities in DDR activation in the presence of the WT and mutant E4orf4 proteins did not result from different expression levels of these proteins. Thus the results support the conclusion that the interaction between E4orf4 and PP2A is required for the reduction in DDR protein phosphorylation which accompanies E4orf4 expression. The results also indicate that E4orf4 can inhibit the response to DNA damage outside the context of virus infection.
Based on our findings that E4orf4 obstructed DNA damage signaling, we expected E4orf4 to inhibit DNA damage repair. A direct evidence of a repair defect in E4orf4-expressing cells was obtained using the alkaline comet assay, which monitors the presence of DNA damage in single cells by microscopic detection of DNA migration in gel [67]. Clone 13 cells induced by Dox for 2.5 hrs to stimulate E4orf4 expression or uninduced cells were treated with H2O2 for 30 min or were left untreated and average comet tail moment was determined. The results demonstrate a significant increase in the number of comets seen in H2O2-treated cells in the presence of E4orf4, indicating a defect in damage repair in cells expressing the viral protein (Fig 3).
Because E4orf4 inhibited phosphorylation of both ATM and ATR targets (Fig 1), and as it was reported that ATM and ATR signaling were not dependent on each other during Ad infection [38], we set out to determine whether E4orf4 reduced phosphorylation of ATM and ATR targets independently of each other, using mutant cells or kinase inhibitors. A-T cells, lacking ATM, and the corresponding WT cells were infected with dl366* or with dl366*+E4orf4. As seen in Fig 4A, activation of both ATM and Chk1 phosphorylation occurred in WT cells infected with dl366*, whereas upon infection with dl366*+E4orf4, ATM phosphorylation at S1981 and Chk1 phosphorylation at three phospho sites were reduced. In A-T cells, ATM was indeed missing, as expected. Phosphorylation of the ATR substrate Chk1 was induced in these cells by dl366* at slightly lower levels compared to WT cells. This may result from the somewhat reduced Chk1 levels in the infected A-T samples, or from loss of a partial indirect contribution of the ATM pathway to ATR activation through end resection [68, 69]. However, a reduction in phosphorylation of all tested Chk1 phospho sites was observed upon infection with dl366*+E4orf4, suggesting that ATM was not required for the reduced phosphorylation of ATR substrates induced by E4orf4. To further confirm this result, clone 13 cells were either induced for E4orf4 expression or left uninduced and were then treated with NCS in the presence or absence of an ATM inhibitor (KU60019). Fig 4B demonstrates that the ATM inhibitor, which diminished phosphorylation of the ATM substrate Smc1-S1778, did not prevent E4orf4 from reducing Chk1 phosphorylation. Thus the results indicate that reducing the phosphorylation of ATR substrates by E4orf4 does not require ATM. Interestingly, E1B-55K levels in A-T cells infected with both dl366* and dl366*+E4orf4 viruses were higher than in WT cells (Fig 4A), suggesting that ATM deficiency may improve virus replication, as discussed below.
To investigate the contribution of ATR to inhibition of DNA damage signaling by E4orf4, HeLa cells were infected with the dl366* and dl366*+E4orf4 mutant Ad viruses in the presence or absence of an ATR inhibitor (ETP46464, [70]). Addition of the ATR inhibitor to the infected cells for 22 hrs did not diminish ATM phosphorylation while resulting in complete loss of Chk1 phosphorylation. When the ability of E4orf4 to reduce ATM phosphorylation was compared in the presence or absence of the ATR inhibitor, no significant changes were observed and a 2.1–2.6 fold reduction occurred in both cases (Fig 4C and 4D). Thus, inhibition of ATR did not diminish the ability of E4orf4 to reduce ATM phosphorylation. Taken together the results indicate that the ATM and ATR signaling pathways can be inhibited by E4orf4 independently of each other.
Since E4orf4 appears to inhibit the ATM- and ATR-regulated DDR (Figs 1, 2 and 4) and DDR was suggested to be detrimental to Ad infection, we examined the contribution of E4orf4 and the DDR components ATM and ATR to the efficiency of virus infection. ATM-deficient A-T cells reconstituted with an empty vector or a vector expressing WT ATM (S2 Fig, [71]) were infected with dl366* and dl366*+E4orf4 at 30 ffu/cell, in the presence or absence of an ATR inhibitor. Morphological changes in the cells and production of infectious viruses were assayed 24 hrs after infection. The cytopathic effects (CPE) that accompany Ad infection, including aggregation, rounding up, and detachment of cells were quantified in two ways. First, the ratio of area covered with adherent cells to the total area was determined (ce), indicating changes in cell number or a clustering of cells. Second, the ratio of area covered with cell clusters larger than a predetermined size to total cell area was calculated (cl), indicating the appearance of large cell clusters. As seen in Fig 5A, CPE were detected more clearly during dl366* infection of A-T cells reconstituted with the empty vector, which lack ATM, than in A-T cells expressing WT ATM (Fig 5A, compare squares b3 and b1). Moreover, the presence of E4orf4 significantly increased the CPE in A-T cells lacking ATM (Fig 5A, compare b3 and c3) and slightly increased CPE in A-T cells expressing ATM (Fig 5A, compare b1 and c1). Addition of an ATR inhibitor dramatically increased the CPE in dl366*- as well as in dl366*+E4orf4-infected cells lacking ATM (Fig 5A, b4 and c4) and somewhat increased CPE in dl366*+E4orf4-infected cells expressing ATM (Fig 5A, c2). However, during infection with dl366*, no difference in CPE was observed in A-T cells reconstituted with WT ATM and treated with the ATR inhibitor compared with no ATRi treatment (Fig 5A, b1 and b2). Efficiency of virus replication was quantified by collecting viruses generated in the cell samples shown in Fig 5A and assaying virus titers. The graph shown in Fig 5B and S1 Table illustrate numerically the enhanced efficiency of virus propagation in the absence of active ATM and ATR and in cells infected with dl366*+E4orf4 viruses compared with dl366* viruses. Thus, ATM deficiency as well as inhibition of ATR increased production of dl366* virus (10.5-fold (ln(fold change) = 2.44) and 5.68-fold (ln(fold change) = 1.76) respectively) and the combination of both led to more than 270-fold increase in dl366* virus propagation (ln(fold change) = 6). Addition of E4orf4 to the E4-deficient virus further enhanced virus multiplication. The temporal progression of infection with dl366* and dl366*+E4orf4 mutant viruses in A-T cells versus matching WT cells was also examined and is demonstrated in Fig 6. CPE was observed earliest in A-T cells infected with dl366*+E4orf4 viruses (24 hrs), and then gradually appeared in A-T cells infected with dl366* and in WT cells infected with dl366*+E4orf4. No CPE was observed in WT cells infected with dl366* viruses up to 48 hrs p.i. (Fig 6A). Virus infection appeared to progress slower in these cells compared with the A-T cells reconstituted with WT ATM or the empty vector described in Fig 5. Virus propagation was measured by a titer assay and the graph shown in Fig 6B together with S2 Table illustrate the enhanced efficiency of virus infection in the absence of ATM or in the presence of E4orf4. Thus at 42 hrs post infection, dl366* produced 11-fold more progeny virus in A-T cells than in WT cells (ln(fold change = 2.46) and dl366*+E4orf4 produced 17-fold more progeny virus (ln(fold change = 3), indicating that ATM inhibits virus replication as was recently reported [40, 78]. Furthermore, dl366*+E4orf4 produced significantly more progeny virus than dl366* both in WT and in A-T cells at late times of infection (2.44-fold and 3.73-fold change respectively (ln(fold change) = 0.83 and 1.39)). Taken together, the results indicate that deficiency in both ATM- and ATR-regulated DDR as well as expression of E4orf4, which inhibits both these DDR pathways, led to a more efficient progression of Ad replication.
Because the results described above suggest that ATM and ATR inhibit Ad propagation and that E4orf4 acts to antagonize their effects, we set out to determine which stages in viral infection are affected by these activities. ATM-deficient A-T cells reconstituted with an empty vector or a vector expressing WT ATM were infected with similar MOIs of dl366* and dl366*+E4orf4, in the presence or absence of an ATR inhibitor (ATRi), as described above (Fig 5). Protein extracts were prepared 24 hrs post-infection and subjected to Western blot analysis. Protein bands from appropriate blot exposures were quantified by densitometry and normalized to Tubulin levels. The normalized protein levels in A-T + VECTOR cells infected with dl366* were defined as 1 and relative levels of E1B-55K and capsid proteins pII, pIV, and pIX in the different infections are shown. As demonstrated in Fig 7, expression levels of E1B-55K were somewhat elevated in the absence of ATM, but ATR inhibition did not increase E1B-55K levels (Fig 7A). E4orf4 levels were altered similarly to E1B-55K (Fig 7A). In contrast, the levels of Ad capsid proteins were dramatically increased when ATM was absent and ATR inhibition further elevated the levels of most capsid proteins except hexon (pII). When both ATM and ATR were deficient, E4orf4 could still increase the levels of some capsid proteins but not of others (compare fiber (pIV) and pIX for example). E4orf4 increased hexon (pII) levels in WT cells (A-T + WT ATM), regardless of ATR activity. Viral DNA levels were also dramatically higher in the absence of ATM than in WT cells but ATR inactivation or E4orf4 expression did not appear to affect them significantly (Fig 7B). These results indicate that ATM inhibition in the cells used here was very important to the virus life cycle starting at the stage of viral DNA replication, whereas ATR inhibition impacted mostly late protein expression and progeny virus production and did not affect earlier stages of the Ad life cycle. Under the experimental conditions used here, the contribution of ATR inhibition to late protein expression was more obvious when ATM was absent, consistent with the changes in virus titer shown in Fig 5.
Since E4orf4 reduced DDR activation and inhibited DNA repair, we examined whether these events resulted in sensitization of E4orf4-expressing cells to death induced by DNA damaging drugs. Clone 13 cells were induced by Dox for two hrs to stimulate E4orf4 expression or were left uninduced, and were then treated with 4 μM HU or 2.5 ng/ml NCS for three hrs or were left untreated. Cell survival was measured using a clonogenic assay at various cell dilutions, without further Dox addition. The number of colonies was counted two weeks later, and relative survival was calculated. As seen in Fig 8, the concentrations of HU and NCS used in these experiments were sub-lethal and did not reduce cell viability by themselves. However, when drug treatment was performed in combination with E4orf4 expression, cell viability was significantly reduced to less than 60% of the viability of E4orf4-expressing cells. These results suggest that E4orf4 sensitized the cells to treatment with genotoxic drugs or drugs causing replication stress.
It has been reported that Ad utilizes several mechanisms to inhibit the DDR. The Ad E1B-55K and E4orf6 proteins cause degradation of various DDR proteins and E4orf3 expression leads to removal of the MRN DNA damage sensor complex from viral replication centers [33, 38, 39]. Moreover, while Ad was shown to induce parylation, the intracellular distribution of PAR-modified proteins was altered by the E1B-55K and E4orf3 proteins, possibly resulting in further modulation of the DDR [73]. The virus also utilizes its core protein VII to protect the incoming genome from the DDR [41]. The exploitation of several mechanisms for DDR attenuation by Ad suggests the importance of this process to virus replication. Our current work demonstrates that the Ad E4orf4 protein contributes a novel mechanism for DDR inhibition. In collaboration with its major cellular partner, PP2A, E4orf4 reduced phosphorylation of several DDR proteins belonging to both the ATM- and ATR-regulated pathways. This inhibition of DNA damage signaling occurred during virus infection as well as when E4orf4 was expressed alone, and was observed both in transformed and non-transformed cells (Figs 1 and 2). E4orf4 reduced phosphorylation of DDR proteins without affecting total levels of the MRN component Nbs1 (Fig 1) which is known to be degraded when the Ad E1B-55K and E4orf6 proteins are present [6, 31]. Disruption of the DDR by E4orf4 resulted in increased accumulation of DNA damage, as demonstrated by a comet assay (Fig 3). This assay was performed in HEK293-derived clone 13 cells which contain the Ad E1 proteins and therefore we cannot rule out some involvement of E1A or E1B in the observed inhibition of DNA repair. However, E4orf4 induced a significant inhibition of repair in this genetic background and thus its role in this process must be critical. Furthermore, E4orf4 inhibited DNA damage signaling in various types of cells, including non-transformed cells, suggesting that it has an important role in inhibition of DNA damage repair. Our results further demonstrate that ATM and ATR are not acting downstream of each other in the signaling pathway initiated by E4orf4 to reduce their activity (Fig 4). PP2A was reported to have a profound impact on the DDR by regulating activity of the primary (ATM, ATR and DNA-PK) and secondary (CHK1 and CHK2) kinases involved in the signaling cascade. It also dephosphorylates downstream targets, such as γ-H2AX and broadly influences repair of DSBs [74]. Thus it is possible that E4orf4 targets the ATM and ATR branches of the DDR separately, by recruiting PP2A to different upstream regulators. Alternatively, it is possible that E4orf4 recruits PP2A to one upstream regulator that affects both these DDR branches. Potential examples of E4orf4-PP2A targets that could affect both ATM- and ATR-regulated pathways include the MRN complex which accumulates at DNA damage foci and initiates signaling involving activation of ATM and ATR kinases [6, 8, 75], PARP-1 which interacts with both ATM and ATR and parylates them [76, 77], or other upstream DDR regulators. Future research will address the identification of direct targets of the E4orf4-PP2A complex in the DDR.
Our results indicate that ATM and ATR inhibition, as well as E4orf4 expression, increase the efficiency of Ad replication to various extents (Figs 5–7). In A-T+WT ATM cells, E4orf4 enhanced virus progeny production only slightly (2-fold, ln(fold change) = 0.76), and we could not detect an E4orf4 effect on viral DNA replication in these cells whereas E4orf4 did increase early and late protein expression to some extent (Fig 7). Because ATM deficiency increased Ad DNA replication (Fig 7B), these results could possibly reflect a low ability of E4orf4 to inhibit ATM activity in the A-T+WT ATM cells. The finding that E4orf4 could enhance progeny virus production without increasing viral DNA levels may reflect the possibility that DNA levels are not the limiting factor in Ad propagation in these cells and therefore the increase in late protein expression is responsible for generation of more infectious viruses. It is also theoretically possible that E4orf4 does not increase total virus DNA levels, but increases the amount of functional DNA, for example by reducing concatenation. However, it was previously reported that ATM and ATR knockdown did not affect mutant virus DNA concatenation in HeLa cells [78]. ATM deficiency led to a 10.5-fold increase in dl366* replication and ATR inactivation increased it by 5.7-fold (Fig 5 and S1 Table). When both ATM and ATR were deficient, the efficiency of dl366* infection was increased by 271-fold (ln(fold) = 6) and E4orf4 further increased replication efficiency by close to 3.3-fold (ln(fold) = 1.11). These results suggest that inhibition of both ATM and ATR by Ad proteins contributes to the efficiency of virus replication. However, besides its contribution to ATM and ATR inhibition, E4orf4 has additional functions that are beneficial for virus replication, such as down-regulation of early gene expression, regulation of splicing of viral RNAs and control of protein translation [44–49]. Execution of these functions may account for the enhancement of the efficiency of virus replication when both ATM and ATR are inhibited. It is also possible that E4orf4 inhibits additional DDR pathways.
It was previously reported that ATM interfered with DNA replication and late protein expression of an E4 mutant Ad in HeLa cells, whereas ATR was not found to interfere with late protein expression in these cells [78]. ATR inhibition was similarly not found to affect replication of a virus lacking E1B-55K and E4orf3 in A549 cells, whereas ATM inhibition significantly increased replication of this virus [40]. Another report presented results showing that neither ATM nor ATR depletion contributed to mutant Ad DNA replication in various cell lines derived from HeLa, U2OS and A-T cells [79]. The reasons for the inconsistent results are not known, but may include the use of different types of cells and different Ad E4 mutants (dl366* [65] vs. dl1004 or dl1007 [80], and ΔE1B-55K-ΔE4orf3 [40]). The significant effect of the ATR inhibitor used here on virus titer (Fig 5) was accompanied by abolishment of Chk1 phosphorylation while no effect on ATM phosphorylation was observed (Fig 4). This inhibitor was also reported to be deficient in inhibiting DNA-PK autophosphorylation [70], indicating its specificity. We found the effect of ATR inhibition to be more prominent when ATM was also absent (Fig 5 and S1 Table) and it may thus be easier to detect under these conditions. Interestingly, whereas ATM inhibition contributed to progression of the Ad replication cycle starting at the DNA replication stage, ATR appeared to exert its effect only later, at the time of late protein expression (Fig 7). We also observed that ATM activation by dl366* infection of HeLa cells occurred earlier than ATR activation (represented by Chk1 phosphorylation) (Fig 1), suggesting that ATR is activated and could exert its effect only at the later stages of infection. The mechanisms by which ATR can inhibit late Ad protein expression are not known, but may include an effect on alternative splicing [81] that could critically impact late gene expression. E4orf4 is known to regulate some alternative splicing events of viral mRNAs [48] and may work, at least in part, via inactivation of ATR. E4orf4 increased fiber protein (pIV) levels both in the presence and absence of an active ATR, whereas it enhanced pIX levels when an active ATR was present but not when ATR was inhibited. These findings could indicate different mechanisms of E4orf4-mediated effects on late protein expression, some of which may depend on ATR inhibition. Our results together with the lengths to which Ad goes to inhibit ATM and ATR signaling, strongly suggest that both ATM and ATR can inhibit Ad replication, at least under some conditions, and that their elimination is therefore important for efficient virus infection.
The results presented here indicate that E4orf4 increases the accumulation of DNA damage following treatment of cells with DNA damaging drugs (Fig 3), resulting in sensitization of the cells to killing induced by sub-lethal concentrations of DNA damaging drugs and drugs inducing replication stress (Fig 8). Because the experiments presented in Fig 8 were performed in cells derived from HEK293 cells which express the Ad E1A and E1B proteins and these proteins have their own effect on cell survival, we cannot rule out some influence of the E1A and E1B proteins. However, E4orf4 further increased the susceptibility of the cells to drug-induced cell death indicating that it has an important contribution to sensitization of these cells to DNA damaging drugs.
It has been reported that when E4orf4 is expressed alone, it induces a unique mode of cancer-specific cell death [60, 61]. Inhibition of the DDR may contribute to this process. Previous reports have indicated that deficiencies in DDR mechanisms are contributing factors in many stages of tumor development [82, 83]. Many malignant tumors show functional loss or deregulation of key proteins involved in the DDR, including p53, ATM, Mre11, BRCA1/2 and Smc1. Such mutations may allow pre-cancerous cells to evade the proliferation barrier created by the DDR, thus allowing the progression of pre-malignant lesions to full malignancy [84]. While defects in DDR components may confer a growth advantage on cancer cells, allowing them to survive and proliferate despite oncogene-induced replication stress and genomic instability, they may also cause cancer cells to rely on the remaining DDR pathways in order to survive DNA damage. Targeting of the remaining pathways by a DDR inhibitor such as E4orf4 may therefore be selectively toxic to cancer cells with mutations in DDR genes.
In summary, E4orf4 employs a novel mechanism to inhibit the DDR, which improves Ad replication and may contribute to induction of cancer-specific cell death by the viral protein. Investigation of this novel mechanism may provide a better understanding of critical DDR nodes that are targeted by E4orf4 and are required for successful application of a combinatorial treatment of cancer. Moreover, our results are important for the understanding of Ad-host cell interactions. The lengths to which Ad goes to inhibit the DDR indicate that DDR inhibition is central to the virus life cycle and our findings enhance this conclusion. Ads have been identified in recent years as significant pathogens in immunocompromised patients [85]. As there is no virus-specific therapy for Ad infection, it has become very important to understand the host responses to Ad infection and the viral strategies used to inhibit these responses in order to promote the development of antiviral therapies. Thus understanding the E4orf4 role in inhibition of the DDR may contribute to the development of new anti-viral and anti-cancer treatments.
HeLa cells (American Type Culture Collection) were cultured in Dulbecco's Modified Eagle's medium (DMEM) supplemented with 10% fetal calf serum (FCS). Clone 13 cells containing tetracycline-inducible E4orf4, clone 3 cells containing a tetracycline-inducible E4orf4 mutant that does not bind PP2A, and clone L11 cells containing a tetracycline-inducible PP2A-B55 shRNA were generated by introducing a tetracycline-inducible pSUPERIOR vector (OligoEngine) containing cDNAs for WT E4orf4 [43], the E4orf4 mutant R81F84A that does not bind PP2A [63], or a PP2A-B55alpha shRNA [86] into T-REx-293 Cells (Invitrogen by Life Technologies). The resulting cell lines were propagated in DMEM supplemented with 10% FCS guaranteed to be tetracycline-free (BD Bioscience), 5 μg/ml blasticidine (Invitrogen), and 200 μg/ml zeocin (Invitrogen). Doxycycline induction was done by addition of 1 μg/ml doxycycline, whereas uninduced cells received an equal amount of the solvent. A-T cells and matching WT fibroblasts were from Y. Shiloh (Tel Aviv University) and were grown in DMEM supplemented with 10% FCS. A-T cells reconstituted with empty vector (GM16666) or the vector expressing WT ATM (GM16667) were from the Coriell Institute and were grown in DMEM supplemented with 15% FCS and 100 μg/ml hygromycin. Human Umbilical Vein Endothelial Cells (HUVEC) were a gift from G. Neufeld and O. Kessler (Technion). These cells were cultured on gelatin-coated dishes in M-199 medium containing 20% FCS, 2 mM glutamine, antibiotics, and 2 ng/ml bFGF, which was added every other day to the cells [87].
A PP2A-B55alpha shRNA-expressing pSUPER plasmid was a gift from S. Strack [86] and the shRNA sequence was subcloned from this plasmid into a tetracycline-inducible pSUPERIOR vector (OligoEngine) for use in the generation of the L11 cell line. The PP2A-B55alpha mutant resistant to the PP2A-B55alpha shRNA was previously described [43].
Adenoviral mutants dl366*, lacking the complete E4 region, and dl366*+E4orf4, lacking all E4 open reading frames (orfs) except E4orf4, were previously described [65] and were propagated on W162 cells (from T. Shenk, Princeton University, [88]). W162 cells were also used for determination of virus titer. Virus infections were performed at a multiplicity of 20–30 fluorescent forming units (ffu) in medium supplemented with 2% FCS at 37°C for 2 hrs, after which additional serum was added to a total of 10%, 15%, or 20%, as required for the specific cell line. Kinase inhibitors were added to infected cells 2 hrs post infection at the following concentrations: the ATM inhibitor KU60019 (Tocris Bioscience): 5μM, the ATR inhibitor ETP46464 [70]: 1 μM. An identical ATM inhibitor quantity and 50% of the original ATR inhibitor were added again 9 hrs later. Untreated cells received equal quantities of the solvent. Infected cells were harvested for protein extraction or virus collection at the indicated times.
Relative levels of viral DNA were monitored by quantitative PCR, as previously described [89]. Total infected cell extracts were prepared in RIPA buffer (150 mM NaCl, 50 mM Tris, pH 8.0, 1.0% IGEPAL CA-630, 0.5% sodium deoxycholate, 0.1% SDS), sonicated, and treated with proteinase K. Quantitative real-time PCR was performed using hexon-specific primers: hexon-qPCR-fw: 5′-CGCT GGACATGACTTTTGAG-3′; hexon-qPCR-rev: 5′-GAACGGTGTGCGCAGGTA-3′. Results were normalized to levels of the cellular PRPH2 gene determined by a parallel quantitative PCR with the primers: TM684-fw: 5’-CTGAAGCCGTACCTGGCTATC-3’; TM685-rev: GTGTCCCGGTAGTACTTCATGC.
Whole cell extracts were prepared in SDS sample buffer (62.5 mM Tris-HCl pH 6.8, 2% SDS, 50 mM DTT, 10% glycerol) and viscosity was reduced by passing the extracts several times through a 27G needle after three min incubation at 95°C. Proteins were analyzed by Western blots using the indicated antibodies. Blot images were scanned with an Epson Photo 4990 scanner and were processed using Adobe PhotoShop 5.0 or 7.0. Band intensities were quantified using the TotalLab software.
Antibodies specific for the following proteins were used in this work: E4orf4 [54], DBP (clone B6) [90], E1B-55K (clone 2A6) [91], PP2A-B55 [43], HA (Covance), pATM-S1981 and ATM (mouse) (Epitomics), pChk1- S296 or S317 or S345, pChk2-T68, p53BP1-S1778, pNbs1-S343 and Nbs1, pAkt-S473 and Akt (Cell signaling), pSmc1-S966 (Bethyl), ATM (rabbit), Chk1, Chk2 (Santa Cruz).
HeLa cells grown on glass coverslips were infected at a multiplicity of infection of 30 ffu/cell. After 24 hrs, the cells were washed, fixed, stained with the indicated antibodies and counter-stained with 4',6-diamidino-2-phenylindole (DAPI) (Sigma). Immunostaining was visualized using an Axiocam camera linked to a Zeiss Axioskop at the indicated magnification.
Clone 13 cells were treated with Dox for 2 hrs to induce E4orf4 expression or were left uninduced and were then treated with hydroxyurea (HU) (Sigma, 4 μM) or neocarzinostatin (NCS) (Sigma, 2.5 ng per ml) for three hrs or left untreated. The cells were then harvested, counted, and plated in several decimal dilutions. Medium (without Dox) was changed every 3–4 days and colonies were counted after two weeks.
Clone 13 cells were induced with Dox for 2.5 hours or were left uninduced and were then treated with 100 μM H2O2 for 30 minutes. An alkaline comet assay was performed as described before [92]. Briefly, cells were harvested and mixed with low-melting agarose. After lysis, the cells were incubated in an alkaline buffer for 30 min and subjected to electrophoresis performed at 1 V/cm for 30 min. The slides were then stained with ethidium bromide dye for 20 min, washed in PBS, dried, and images of at least 100 randomly selected cells per sample were captured by a Zeiss fluorescent microscope at a magnification of 200. Digital fluorescent images were obtained using the AxioVision software. In this assay, the length and intensity of DNA tails relative to heads is proportional to the amount of DNA damage in individual nuclei. These parameters were determined by measurements of comet tail moment using the TriTek Comet Score software (TriTek Corp., Sumerduck, VA).
Box plots were generated using R software (version 3.1.2) and package ‘ggplot2’ (version 1.0.0). Data for box plot analysis was subjected to natural log transformation. Statistical significance was calculated using t-test.
|
10.1371/journal.pbio.1002183 | Glypican Is a Modulator of Netrin-Mediated Axon Guidance | Netrin is a key axon guidance cue that orients axon growth during neural circuit formation. However, the mechanisms regulating netrin and its receptors in the extracellular milieu are largely unknown. Here we demonstrate that in Caenorhabditis elegans, LON-2/glypican, a heparan sulfate proteoglycan, modulates UNC-6/netrin signaling and may do this through interactions with the UNC-40/DCC receptor. We show that developing axons misorient in the absence of LON-2/glypican when the SLT-1/slit guidance pathway is compromised and that LON-2/glypican functions in both the attractive and repulsive UNC-6/netrin pathways. We find that the core LON-2/glypican protein, lacking its heparan sulfate chains, and secreted forms of LON-2/glypican are functional in axon guidance. We also find that LON-2/glypican functions from the epidermal substrate cells to guide axons, and we provide evidence that LON-2/glypican associates with UNC-40/DCC receptor–expressing cells. We propose that LON-2/glypican acts as a modulator of UNC-40/DCC-mediated guidance to fine-tune axonal responses to UNC-6/netrin signals during migration.
| During the development of the nervous system, migrating axons are guided as they navigate through complex environments to reach their target destinations. These directed migrations are essential to ensure the proper wiring and function of the nervous system and are instructed by guidance cues and receptors. There is a remarkably small set of guidance cues and receptors relative to the large number of neuronal migrations, suggesting that the actions of these guidance cues might be diversified by regulatory mechanisms. We have addressed this question in the genetically tractable nematode Caenorhabditis elegans. We identify that the response of migrating neurons to a key guidance cue, UNC-6/netrin, is modulated by a specific proteoglycan, LON-2/glypican. We show that LON-2/glypican may carry out this regulation by interacting with the UNC-40/DCC netrin receptor on the cell surface. We propose that LON-2/glypican acts as an extracellular modulator of UNC-40/DCC-mediated guidance to fine-tune axonal responses to UNC-6/netrin signals during migration.
| Directed migrations of developing axons are essential for the proper wiring of the nervous system. A host of guidance cues and their receptors instruct axon guidance decisions. However, how these cues and the growth cone’s responses to them are spatially and temporally regulated in vivo remains largely unknown. Answering this question is central to our understanding of how growing axons navigate in complex environments to reach their targets during development and regeneration.
UNC-6/netrin is a highly conserved secreted guidance cue with structural similarity to the extracellular matrix protein laminin [1–3]. UNC-6/netrin directs attractive guidance through receptors of the UNC-40/DCC family and repulsive guidance through both UNC-40/DCC and UNC-5/UNC5 receptors [4–6]. Notably, whereas netrin receptors and downstream transduction pathways have been well characterized, how netrin signals are regulated extracellularly remains largely unknown. UNC-6/netrin was identified through genetic analysis in Caenorhabditis elegans [1] and biochemically purified and cloned from vertebrate embryos [2]. A second biochemical component that synergized with netrin to elicit axon outgrowth was termed “netrin synergizing activity” (NSA) [3] and remains unidentified. Vertebrate netrin-1 and its receptor DCC can bind heparin, a fully sulfated version of heparan sulfate (HS), in vitro [3,7,8], and a general disruption of HS chain synthesis is detrimental to netrin-1-mediated axon outgrowth in vitro [9,10]. While heparan sulfate proteoglycans (HSPGs) might be intriguing candidates for NSA, it is not yet known whether a specific HSPG is required for netrin signaling or how interactions with HSPGs might regulate netrin signals to direct axons during nervous system development.
We addressed these questions using the nematode C. elegans, which has been instrumental for discovering major conserved axon guidance pathways. During larval development, the axon of the mechanosensory neuron AVM migrates ventrally as its growth cone integrates signals from two complementary guidance cues (Fig 1A) [1,4–6, 11–13]: (1) UNC-6/netrin is secreted at the ventral midline and attracts the growth cone ventrally via the receptor UNC-40/DCC [5,14], and (2) SLT-1/Slit is secreted by the dorsal muscles and repels the growth cone away from the dorsal side via the receptor SAX-3/Robo [12,13]. Animals null for the guidance cues unc-6/netrin or slt-1/Slit exhibit partial AVM ventral axon guidance defects, and loss of both cues in unc-6 slt-1 double mutants results in fully penetrant guidance defects (S1 Fig, [13]). AVM axons defective in guidance fail to extend ventrally and instead migrate laterally in the anterior direction (Fig 1). In this study, we use the AVM axon as a model to elucidate mechanisms that regulate UNC-6/netrin signaling.
Here we provide a missing link in understanding the modulation of UNC-6/netrin signaling in the extracellular milieu. We demonstrate that LON-2/glypican, a HSPG secreted from epidermal cells, acts as a modulator of the UNC-6/netrin signaling pathways to guide migrating cells and axons. We show that LON-2/glypican modulates UNC-6/netrin signaling in both attractive guidance mediated by the UNC-40/DCC receptor and repulsive guidance mediated by the UNC-40/DCC and UNC-5/UNC5 receptors. We provide evidence that LON-2/glypican associates with UNC-40/DCC-receptor-expressing cells. We show that the N-terminal globular region of LON-2/glypican, lacking the three HS chain attachment sites, is functional in UNC-6/netrin-mediated guidance. Our studies unravel a novel mechanism by which LON-2/glypican is produced by substrate epidermal cells and released from the membrane to likely associate with UNC-40/DCC-expressing neurons, enabling the modulation of their responses to UNC-6/netrin during axon migrations.
To address whether a specific HSPG interacts with the netrin signaling system to guide axons, we first examined axon guidance in mutants lacking core HSPGs. HSPGs are composed of a core protein with covalently attached long unbranched HS chains [15]. HSPGs can be associated with the plasma membrane through either a transmembrane domain (e.g., syndecans) or a glycerophosphatidylinositide (GPI) anchor (e.g., glypicans) or be secreted into the extracellular milieu (e.g., perlecans and agrins). We examined the axon morphology of AVM in single, double, and triple mutants for several core HSPG proteins (see S1 Table for alleles). These included the sole C. elegans syndecan (sdn-1), the two glypicans (lon-2 and gpn-1), perlecan (unc-52), and agrin (agr-1). We found that the mild AVM axon guidance defects of sdn-1/syndecan mutants, including a null, [16] were enhanced by the complete loss of lon-2/glypican in double mutants lon-2 sdn-1 (Fig 1B), revealing a role for lon-2/glypican in AVM axon guidance. Similarly, loss of lon-2/glypican enhances sdn-1/syndecan mutants in motorneuron guidance [17]. Although the C. elegans genome encodes two glypicans, loss of function of the second glypican, gpn-1, using two likely null mutant alleles (see S2 Fig), did not enhance the defects of lon-2/glypican or sdn-1/syndecan null mutants in double or triple mutants (Fig 1B). Moreover, we did not observe abnormal phenotypes in the single mutants for agr-1/agrin or unc-52/perlecan. These observations highlight the specificity of lon-2/glypican function in this axon guidance process and raise the possibility that lon-2/glypican might be a component of the pathways guiding the AVM axon towards the ventral midline.
Considering that AVM axon guidance occurs via the unc-6/netrin and slt-1/Slit pathways, mutations in genes such as lon-2/glypican and sdn-1/syndecan that affect AVM axon guidance may point towards interactions with either of these two guidance systems. Since the AVM axon guidance defects in lon-2 sdn-1 double mutants are qualitatively similar to those of mutants lacking unc-6/netrin or slt-1/Slit, we determined how lon-2/glypican and sdn-1/syndecan impact unc-6/netrin and slt-1/Slit signaling. In animals that completely lack slt-1/Slit function, the complete loss of a gene functioning independently of slt-1/Slit is expected to enhance the AVM guidance defects, such as in the double null mutants unc-6/netrin slt-1/Slit (see S1 Fig). We tested the interactions of lon-2/glypican with the slt-1/Slit pathway in AVM axon guidance and found that the complete loss of lon-2/glypican enhanced a presumptive null allele of slt-1/Slit in lon-2 slt-1 double mutants (Fig 1C), suggesting that lon-2/glypican functions in a pathway separate from slt-1/Slit. Loss of lon-2/glypican also enhanced guidance defects when signaling through sax-3/Robo, the slt-1/Slit receptor, was disrupted in lon-2 sax-3 double null mutants, providing further evidence that lon-2/glypican functions in a pathway separate from that of slt-1/Slit (Fig 1C). As an additional method to investigate the impact of lacking lon-2/glypican function when slt-1/Slit signaling is perturbed, we used a transgene that ectopically expresses slt-1/Slit from both ventral and dorsal body wall muscles (using Pmyo-3::slt-1) and misguides the axon of AVM [18]. Loss of lon-2/glypican enhanced the defects caused by slt-1/Slit misexpression (Fig 1C), consistent with the above findings that lon-2/glypican mediates its axon guidance effects independently of slt-1/Slit.
The unc-6/netrin pathway functions independently of slt-1/Slit to guide AVM. To address whether lon-2/glypican functions in the unc-6/netrin axon guidance pathway, we examined the AVM axon in double mutants of lon-2/glypican and unc-6/netrin. In animals that completely lack unc-6/netrin function, the complete loss of a gene functioning in the same unc-6/netrin pathway is expected to not enhance the AVM guidance defects, such as in the double null mutants unc-6; unc-40 (see S1 Fig). We found that the complete loss of lon-2/glypican did not enhance the guidance defects displayed by unc-6/netrin null mutants ev400 (Fig 1D). Given that loss of lon-2 enhances the defects of other guidance mutants (see doubles with sdn-1, slt-1, sax-3, and Pmyo-3::slt-1 in Fig 1B and 1C and sqv-5 in S3 Fig), the lack of enhancement when combined with the unc-6 null mutation suggests that lon-2/glypican functions in the same pathway as unc-6/netrin. Consistent with this idea, we also found that complete loss of lon-2/glypican did not enhance the AVM guidance defects of two null mutant alleles of the netrin receptor unc-40/DCC in the double mutants unc-40; lon-2 (Fig 1D), suggesting that lon-2/glypican functions in the same pathway as unc-40/DCC in AVM ventral guidance. These observations raise the interesting possibility that lon-2/glypican may be the HSPG dedicated to modulate unc-6/netrin signaling through unc-40/DCC during axon guidance.
Since lon-2/glypican functions independently of slt-1/Slit (Fig 1C) and partly separate from sdn-1/syndecan (Fig 1B), we tested whether sdn-1/syndecan and slt-1/Slit function together to guide the axon of AVM. We found that defects in slt-1 sdn-1 double null mutants were not enhanced compared to the single mutants (Fig 1E), consistent with findings in Drosophila [19,20] and C. elegans [16]. We also found that double null mutants for sdn-1/syndecan and the slt-1/Slit receptor sax-3/Robo were not enhanced compared to the single mutants (Fig 1E). Our results support the notion that sdn-1/syndecan acts in the same genetic pathway as slt-1/Slit to guide AVM. Consistent with this, we found that the double null mutants for sdn-1/syndecan and the netrin receptor unc-40/DCC were enhanced, indicating that sdn-1/syndecan functions in a pathway separate from unc-6/netrin. The analysis of axon guidance in double mutants of unc-6/netrin and sdn-1/syndecan was precluded by their lethality. Our results are consistent with the notion that unc-6/netrin and sdn-1/syndecan act in different pathways of axon guidance.
In addition to unc-6/netrin acting as an attractive cue for cells expressing the unc-40/DCC receptor in ventral guidance, unc-6/netrin also acts as a repulsive cue for cells expressing both the unc-5/UNC5 and unc-40/DCC receptors, which together mediate dorsal guidance away from unc-6/netrin [4–6]. To address whether lon-2/glypican functions in unc-6/netrin-mediated repulsive guidance as well, we examined the dorsal migration of the distal tip cells (DTCs) and of the GABAergic motorneuron axons [4,11]. We found that lon-2/glypican single null mutants are defective in dorsal DTC migrations (Fig 2A and 2B) and that the complete loss of lon-2/glypican did not enhance the dorsal DTC migration defects of unc-6/netrin, unc-40/DCC, or unc-5/UNC5 null mutants (Fig 2B), indicating that lon-2/glypican functions in the unc-6/netrin-repulsive guidance pathway as well. Similarly, complete loss of lon-2/glypican did not enhance the defects of unc-40/DCC mutants in the dorsal guidance of motorneuron axons (Fig 2C). Given that loss of lon-2/glypican enhances the motorneuron axon guidance defects of sdn-1 mutants as shown in [17], lon-2/glypican plays a role in the dorsal guidance of motorneuron axons. The lack of enhancement of the defects in the dorsal guidance of motorneuron axons of unc-40/DCC mutants by loss of lon-2/glypican further supports that lon-2/glypican functions in the unc-6/netrin pathway mediating dorsal guidance. Thus, lon-2/glypican may modulate unc-6/netrin signaling not only during attractive guidance but also during repulsive guidance.
To complement the above loss-of-function approach, we next used a gain-of-function strategy to test the model that lon-2/glypican functions in the unc-6/netrin signaling pathway. We focused on the axon of the PVM neuron instead of AVM, because it could reliably be identified (AVM cannot be distinguished from ALMR in these experiments). In wild-type animals, PVM, like AVM, expresses the receptor unc-40/DCC, and its axon grows ventrally towards unc-6/netrin (Fig 3A). In mutants lacking unc-6/netrin signaling, PVM axons that fail to extend ventrally instead extend anteriorly (never dorsally, see S4 Table). The PVM axon normally does not express the receptor unc-5/UNC5 that mediates repulsive guidance away from ventral unc-6/netrin [6], but misexpression of the receptor unc-5/UNC5 (using transgene Pmec-7::unc-5 [21]) in PVM forces its axon to extend dorsally in an unc-6/netrin- and unc-40/DCC-dependent manner (Fig 3A and 3B, [21]). We used this unc-6/netrin-dependent unc-5/UNC5-mediated abnormal dorsal migration to further investigate the function of lon-2/glypican in netrin signaling. By analyzing lon-2/glypican mutants carrying Pmec-7::unc-5, we found that compete loss of lon-2/glypican function significantly suppressed the unc-6/netrin-dependent unc-5-mediated abnormal dorsal migration of the PVM axon, indicating that unc-6/netrin signaling is lon-2/glypican dependent (Fig 3B). In contrast, the complete loss of sdn-1/syndecan, of slt-1/Slit, or of sax-3/Robo function did not suppress these PVM abnormal dorsal migrations (Fig 3B, see S4 Table), highlighting the specificity of lon-2/glypican action on unc-6/netrin signaling. As expected, lon-2 sdn-1 double mutants lacking both lon-2/glypican and sdn-1/syndecan and expressing unc-5/UNC5 in PVM did not further suppress the abnormal unc-5/UNC-5-mediated dorsal migration of PVM as compared to lon-2 single mutants, further supporting the specificity of lon-2/glypican on unc-6/netrin signaling.
To investigate whether lon-2/glypican functions in the same genetic pathway as known downstream mediators of unc-6/netrin signaling, we tested for genetic interactions between lon-2/glypican and unc-34/enabled. unc-34/enabled is a regulator of actin polymerization for axonal filopodia outgrowth [18,22–26], and its role in both unc-6/netrin and slt-1/Slit guidance pathways renders the analysis of genetic interactions in the context of normal AVM axon guidance challenging. Therefore, we used the unc-6/netrin-specific gain-of-function approach as above, in which the dorsal migration of the PVM axon upon ectopic expression of unc-5/UNC5 is unc-34/enabled dependent (Fig 3B, [21,27]). We asked whether loss of lon-2/glypican could enhance the extent of suppression of PVM dorsal migration induced by loss of unc-34/enabled. We found that the PVM dorsal migration was suppressed to the same degree in the double null mutants lon-2; unc-34 and the single mutant unc-34/enabled upon expression of unc-5/UNC5 in PVM (Pmec-7::unc-5, Fig 3B). These results support that lon-2/glypican functions with unc-6/netrin and unc-34/enabled during axon guidance.
The AVM growth cone extends along a basement membrane that is located between the epidermis, which is referred to as the hypodermis, and body wall muscles [11]. lon-2/glypican is expressed in the hypodermis and the intestine [28]. We asked in which cell type lon-2/glypican needs to be produced to guide AVM. We found that wild-type lon-2(+) transgenes expressed under the heterologous epidermal promoters Pdpy-7 and Pelt-3 (that drive expression in the hypodermis underlying the AVM growth cone, hyp7) rescued lon-2 slt-1 double mutants back to slt-1 single mutant levels, as efficiently as when expressed under the endogenous promoter Plon-2 (Fig 4A, S3 Table). Rescue was not observed when we expressed lon-2/glypican in other epidermal cells (seam cells, Pgrd-10), in the migrating neuron itself (Pmec-7), in the intestine (Pelt-2), or in body wall muscles (Pmyo-3) (Fig 4A, S3 Table). Our results suggest that lon-2/glypican is produced by the hypodermis underlying the growth cone of AVM to function in axon guidance.
We found that expressing wild-type copies of sdn-1(+) in the AVM neuron (using the heterologous promoter Pmec-7) rescued axon defects of lon-2 sdn-1 double mutants (Fig 4B). Accordingly, our examination of a transgene reporting sdn-1/syndecan expression (sdn-1::gfp [16]) revealed that sdn-1/syndecan is indeed expressed in the AVM neuron (S4 Fig), at the time of its ventral migration during the first larval stage. Thus, sdn-1/syndecan appears to function in the migrating neuron in the slt-1/Slit-sax-3/Robo guidance pathway, whereas lon-2/glypican appears to function nonautonomously, as it is produced by the hypodermis underlying the migrating neuron to modulate the unc-6/netrin guidance pathway. Consistent with this, we found that sdn-1(+) cannot replace the function of lon-2/glypican; expressing sdn-1/syndecan in either the cells that normally express lon-2/glypican (using Plon-2::sdn-1) or the migrating neuron itself (Pmec-7::sdn-1) did not rescue the loss of lon-2/glypican (S5 Fig), supporting that lon-2/glypican and sdn-1/syndecan have specific roles in axon guidance.
Glypicans are composed of a core protein moiety with covalently linked HS chains attached via a tetrasaccharide linker at specific Serine residues (Fig 5A, [15]). Prior studies on the role of HSPGs in other developmental pathways indicate that both the identity of the HSPG core proteins and the heterogeneity of their HS chains modified by epimerization and sulfations [15] contribute to the specificity of the interactions between particular HSPGs and the proteins that they bind [15,29,30].
To address the importance of the HS chains linked to LON-2/glypican during axon guidance, we tested whether a mutated form of LON-2/glypican lacking its HS chains could still function in axon guidance. For this experiment, the three Serine residues serving as HS chain attachment sites were mutated to Alanine residues, generating the mutant LON-2ΔGAG [31]. Western blot analysis confirmed that LON-2ΔGAG severely reduced HS chains associated with LON-2, in both worms and S2 cells (Fig 5B and S6 Fig). We then expressed LON-2ΔGAG under the Plon-2 endogenous promoter and found that the AVM guidance defects of lon-2 slt-1 double mutants were rescued back to the level of slt-1 single mutants (Fig 5C). Similarly, the DTC migration defects of lon-2/glypican mutants were rescued by LON-2ΔGAG expression (Fig 5D). Our results indicate that LON-2/glypican devoid of its HS-chain attachment sites can function in unc-6/netrin-mediated guidance, suggesting that the core protein is the critical part of LON-2/glypican for its function in unc-6/netrin-mediated guidance of cell and axon migrations.
Our above observations provide evidence that the HSPG lon-2/glypican functions in the same genetic pathway as unc-6/netrin to guide migrating axons. It has been shown in several models that HSPGs play multifaceted roles across various signaling pathways, such as facilitating ligand-receptor interactions and transporting morphogens, as well as localizing and stabilizing ligands [32,33]. We asked if the LON-2/glypican molecules might interact with either UNC-6/netrin or its receptor UNC-40/DCC, suggesting a potential mechanism of action for LON-2/glypican in unc-6/netrin-mediated guidance. To test these interactions, we generated epitope-tagged versions of LON-2/glypican, UNC-6/netrin, and UNC-40/DCC proteins, with human influenza hemagglutinin (HA), superfolder-GFP (SfGFP), and FLAG, respectively (Fig 6A), and used cell-mixing experiments. We independently expressed each of these labeled proteins in separate populations of Drosophila S2 cells for 2 d, then cocultured them overnight, and detected the tagged proteins by western blot analysis (see S7 Fig) and by immunostaining (Fig 6A).
We observed that the HA::LON-2 signal filled the cytoplasm of HA::LON-2 producing cells (indicated by white asterisks in Fig 6B experiment 1 and S8 Fig). Notably, HA::LON-2 was also found decorating the outline of UNC-40::FLAG-expressing cells (Fig 6B and 6C experiments 1, 6, 7, and 8). This observation suggests that LON-2/glypican is released from the cells that produce it, diffuses in the extracellular medium, and associates with UNC-40/DCC-expressing cells. In contrast, HA::LON-2/glypican did not bind to cells expressing SfGFP::UNC-6 (Fig 6B and 6C experiments 4, 6, and 7) or to cells expressing an unrelated type I transmembrane receptor, Evi (see S9 Fig), or to untransfected cells (Fig 6B and 6C experiments 1–8). Furthermore, we found that another HSPG, SDN-1/syndecan, did not bind UNC-40-expressing cells (see S9 Fig). These findings provide evidence for a specific interaction between LON-2/glypican and UNC-40-expressing cells.
We tested whether the HS chains of LON-2/glypican were necessary for its association with UNC-40-expressing cells. We used a mutated form of LON-2/glypican lacking its three HS chain attachment sites, HA::LON-2ΔGAG (see S6 Fig, [31]). Western blot analysis confirmed that LON-2ΔGAG severely reduced HS chains associated with LON-2/glypican (S6 Fig). We found that LON-2ΔGAG associated with UNC-40/DCC-expressing cells (Fig 6B and 6C experiment 2), suggesting that the association of LON-2/glypican with UNC-40/DCC-expressing cells is HS-chain independent.
The HA::LON-2 signal outlined the UNC-40/DCC-expressing cells (Fig 6B, experiments 1, 6, 7, and 8) suggesting a potential interaction at the cell surface. To further support this idea, we asked whether LON-2/glypican would associate with cells expressing a mutated form of UNC-40/DCC that lacks the extracellular domain and contains only the intracellular and transmembrane domains (UNC-40ΔNt::FLAG). We found that HA::LON-2 did not associate with cells expressing the UNC-40ΔNt::FLAG (Fig 6B and 6C experiment 3), indicating that the extracellular domain of UNC-40/DCC is required for LON-2/glypican to associate, as would be predicted if LON-2/glypican and UNC-40/DCC interact, directly or indirectly, at the cell surface.
Interestingly, HA::LON-2 was absent from cells expressing SfGFP::UNC-6 (Fig 6B and 6C experiments 4, 6, and 7), indicating that while LON-2/glypican interacts with cells expressing UNC-40/DCC, it does not bind to UNC-6/netrin-expressing cells in this assay. Moreover, the presence of SfGFP::UNC-6 did not reduce the ability of HA::LON-2 to associate with UNC-40/DCC-expressing cells in experiments in which the three singly transfected cell populations were mixed (Fig 6B and 6C experiment 6). These results suggest that if LON-2/glypican interacted directly or indirectly with UNC-40/DCC, then the interactions of LON-2/glypican and UNC-6/netrin would occur with different regions of UNC-40/DCC. Consistent with this possibility, we found that LON-2/glypican still associated with cells expressing UNC-40ΔFn4/5::FLAG, a mutated form of UNC-40/DCC that lacks the UNC-6/netrin-binding sites (FnIII domains 4 and 5) (Fig 6B and 6C experiment 7, [34,35]). Our results indicate that for LON-2/glypican to associate with UNC-40/DCC-expressing cells, the FnIII domains 4 and 5 of UNC-40/DCC are dispensable and UNC-6/netrin does not need to be bound to UNC-40/DCC.
Previous work has suggested that overexpression of DCC in cells overactivates DCC downstream signaling pathways, leading to cytoskeletal rearrangements that result in increased membrane extensions and cell surface area [36]. Similarly, expression of UNC-40/DCC leads to changes in cellular morphology in our cell assays (Fig 6D). To test whether the association of LON-2/glypican with UNC-40/DCC-expressing cells results in an activation of signaling downstream of UNC-40/DCC, we examined the impact of LON-2/glypican on the morphology of UNC-40/DCC-expressing cells. For these experiments, we mixed mCherry-expressing cells with either untransfected control cells or LON-2/glypican-expressing cells, and we also mixed UNC-40/mCherry-expressing cells with either untransfected control cells or LON-2/glypican-expressing cells. Examination of the morphology of these cells 1 d after mixing revealed that UNC-40/mCherry-expressing cells mixed with LON-2/glypican exhibited an increased frequency of irregular shapes and membrane extensions, compared to UNC-40/mCherry cells mixed with control cells (Fig 6D). Thus, consistent with a model in which LON-2/glypican functions in the UNC-6/netrin signaling pathway to guide developing axons, the association of LON-2/glypican with UNC-40/DCC-expressing cells leads to increased membrane extensions, suggestive of increased signaling downstream of the UNC-40/DCC receptor.
While LON-2/glypican possesses a signature GPI anchor that mediates its attachment to plasma membranes (Fig 5A), our experiments indicate that LON-2/glypican is released into the extracellular milieu through cleavage where it can diffuse to associate with UNC-40/DCC-expressing cells. This is consistent with prior work demonstrating that many glypicans are shed or cleaved into a soluble form [37]. To verify that LON-2/glypican is indeed released into the extracellular medium, we collected cell-free media from HA::LON-2 cultures (HA::LON-2-conditioned medium) and added it to cells expressing UNC-40::FLAG. We found that HA::LON-2-conditioned medium contained HA::LON-2 that associated with UNC-40::FLAG-expressing cells. As above, this interaction was specific, as no HA::LON-2 signal was found on adjacent untransfected cells (Fig 6B and 6C experiment 8). This result provides compelling evidence that LON-2/glypican can be released from the membrane of LON-2/glypican-expressing cells, diffuses, and associates with UNC-40/DCC-expressing cells. We propose that using a similar mechanism, LON-2/glypican may be shed from epidermal cells and may interact with migrating axons that express UNC-40/DCC. This is consistent with our finding that LON-2/glypican is produced by the hypodermis to function nonautonomously in unc-6/netrin-mediated AVM axon guidance.
To provide evidence for the model that LON-2/glypican can function in axon guidance when detached from the plasma membrane, we used a form of LON-2/glypican lacking the GPI anchor, LON-2ΔGPI, which should be directly secreted into the extracellular milieu [31]. LON-2ΔGPI rescued the AVM guidance defects of lon-2 slt-1 double mutants back to the level of slt-1 single mutants (Fig 5C). We also used a truncated form of LON-2/glypican (N-LON-2) containing the N-terminal globular domain, but lacking the C-terminal region, thus removing the three HS attachment sites and the GPI membrane anchor. N-LON-2 also rescued the AVM guidance defects of lon-2 slt-1 double mutants back to the level of slt-1 single mutants (Fig 5C). In contrast, a reciprocal construct containing only the C-terminus with the three HS attachment sites and the GPI anchor (C-LON-2) did not rescue the AVM axon guidance defects of lon-2 slt-1, consistent with the model that the N-terminal globular domain of LON-2/glypican is the key functional domain during guidance (Fig 5C). A secreted form of LON-2/glypican is also functional in DTC guidance, as we found that DTC guidance defects of lon-2/glypican mutants could be rescued by expression of N-LON-2, containing only the N-terminal globular domain (Fig 5D). These findings also support the hypothesis that LON-2/glypican may normally be released from the hypodermis to interact with the unc-6/netrin pathway to direct the migrating growth cone during development (Fig 7).
Growth cone responses to guidance cues require precise regulation as developing axons traverse complex extracellular environments in order to reach their targets. The mechanisms by which guidance cue signals are regulated in the extracellular milieu are still poorly understood [38]. Here, we demonstrate that the unc-6/netrin-unc-40/DCC guidance system is modulated by the HSPG lon-2/glypican.
Our studies identify the HSPG LON-2/glypican as a component of the unc-6/netrin attractive and repulsive signaling pathways that guide axons during development. We show that LON-2/glypican specifically acts on unc-6/netrin signaling independently of slt-1/Slit. We demonstrate that lon-2/glypican functions from the hypodermis, the epidermal cells that secrete the substrate along which growth cones extend [11], and that a secreted form of LON-2/glypican, containing only its N-terminal globular region and lacking its HS chains, guides cells and axons in vivo. In addition, we provide evidence that LON-2/glypican is released from cells producing it and associates with cells expressing UNC-40/DCC receptors. Taken together, our observations support a hypothetical model in which GPI-linked LON-2/glypican is produced by substrate epidermal cells, is released into the extracellular milieu, and binds growth cones expressing UNC-40/DCC receptors to regulate attractive and repulsive responses of the growth cone to UNC-6/netrin.
The impact of lon-2/glypican on the unc-6/netrin signaling pathway is highly specific. First, loss of lon-2/glypican, but not of sdn-1/syndecan, suppresses the guidance phenotypes elicited by the gain-of-function condition in which unc-5/UNC5 was misexpressed. Second, the complete loss of lon-2/glypican does not enhance the guidance defects observed in null mutants for unc-6/netrin or its receptors unc-40/DCC and unc-5/UNC5, whereas it does enhance the defects of several other axon guidance mutants, including sdn-1/syndecan, slt-1/Slit, misexpressed slt-1/Slit (Pmyo-3::slt-1), sax-3/Robo, and sqv-5, suggesting that lon-2/glypican functions specifically in the unc-6/netrin pathway. Third, sdn-1/syndecan, cannot replace lon-2/glypican function, highlighting a requirement for lon-2/glypican that cannot be achieved by any HSPG. Given that the core protein of LON-2/glypican, devoid of its HS chains, is fully functional in guidance, the specificity of action of LON-2/glypican in netrin-mediated guidance appears to reside in the core protein itself. As a note, whereas lon-2/glypican mutants are defective in DTC migration, the lon-2/glypican mutant by itself does not show drastic alterations in AVM axon guidance as is observed with other modulators [32]. It is possible that in the absence of lon-2/glypican, another HSPG may provide compensation or that our scoring of strong alterations in pathfinding did not include more subtle phenotypes, as could be expected from a modulator of the signal [32].
We show that the LON-2/glypican core protein, devoid of HS attachment sites, is able to associate with UNC-40-expressing cells and is functional in unc-6/netrin-mediated guidance. Thus, the core protein is the critical region of LON-2/glypican for netrin-mediated axon guidance. This is in line with previous studies showing a contextual dependence of HS chains for glypican function. For instance, the core protein of C. elegans LON-2/glypican and of Drosophila glypican Dally do not require HS chains to function in the transforming growth factor beta (TGFβ) pathway [31,39]. Similarly, Drosophila glypican Dally-like interacts with Wg and Hh through their protein core in a HS-independent manner [33,40,41], and mammalian Glypican-3 does not require HS chains for its role in Wnt and Hh signaling [42–45]. While the HS chains are not critical for the role of LON-2/glypican in guidance, a contribution of HS chains to modulate functionality, as observed for other glypicans in the context of BMP4, Wnt3, Wg, and Hh signaling [33,39,41,42], cannot be ruled out. For instance, it is conceivable that the normal endogenous HS chains of LON-2/glypican may impact its trafficking, levels, release from the membrane, recruitment of binding partners, or recycling.
LON-2/glypican is predicted to localize at the cell surface via its GPI anchor [31]. However, in our cell culture studies, we demonstrate that LON-2/glypican can be released as a soluble molecule from producing cells. We also show that two truncated forms of LON-2/glypican, LON-2ΔGPI and N-LON-2, which are no longer associated with the plasma membrane and are secreted into the extracellular milieu, can function to guide axons in vivo. This indicates that LON-2/glypican is likely released from the epidermal cells to reach the growth cone to modulate its guidance. This finding raises the question of how LON-2/glypican is released from the cell membrane and how this process might be regulated during development. The release of LON-2/glypican from the surface of cells could involve phospholipases that cleave the GPI anchor and/or proteases that cleave its extracellular domain, such as at a predicted furin-cleavage site (Fig 5A, [31]).
Glypican cleavage by lipases and proteases has been demonstrated to occur and to be functionally important in other contexts, such as in regulating fibroblast growth factor (FGF) and Wnt signaling during morphogenesis [37,46]. For instance, the Drosophila glypican Dally-like protein is cleaved at the GPI anchor by the lipase Notum, to negatively regulate Wnts [47]. Similarly, several mammalian glypicans, including glypican-3, are cleaved by Notum [48]. The functional importance of glypican proteolytic cleavage is illustrated by the processing of glypican-3 by a furin-like convertase to modulate Wnt signaling in zebrafish [49]. In addition, glypican-1 and glypican-4 are proteolytically cleaved to stimulate long-range FGF signaling in the Xenopus embryo [50] and increase the efficiency of myogenic differentiation in the presence of FGF in mammalian cells [51], respectively. Our studies show that glypican processing also functions during axon guidance.
We demonstrate that LON-2/glypican is secreted into the extracellular medium and decorates the outline of UNC-40/DCC-expressing cells. Deleting the extracellular domain of UNC-40 (UNC-40ΔNt) abrogated the association of LON-2/glypican with UNC-40/DCC-expressing cells, indicating that LON-2/glypican may interact with UNC-40/DCC at the cell surface. The association of LON-2/glypican with UNC-40/DCC may be direct or indirect through interactions with other molecules (Fig 7). Our experiments demonstrate that UNC-6/netrin binding to UNC-40/DCC was undisturbed by the association of UNC-40/DCC with LON-2/glypican, suggesting that the possible interaction of LON-2/glypican with UNC-40/DCC likely involves a region of UNC-40/DCC other that the netrin binding sites. Indeed, we found that LON-2/glypican associates with UNC-40/DCC-expressing cells even when the UNC-40/DCC receptors lack the UNC-6/netrin binding domains.
We found that LON-2/glypican leads to increased irregular morphology of UNC-40/DCC-expressing cells. Ectopic expression of DCC in mammalian cells activates downstream signaling via Cdc42 and Rac1, producing cytoskeletal rearrangements that lead to filopodia outgrowth and cell surface extensions [36]. Our finding that the presence of LON-2/glypican enhances the UNC-40/DCC-induced irregular cell morphology and filopodia-like extensions suggests that the association of LON-2/glypican with UNC-40/DCC-expressing cells may increase signaling downstream of UNC-40/DCC. Consistent with this notion, we show that lon-2/glypican functions in the same signaling pathway as the UNC-40/DCC downstream mediator unc-34/enabled during axon guidance.
Our results suggest a possible regulatory mechanism in the extracellular space in which secreted LON-2/glypican modulates the activity of the receptor UNC-40/DCC. LON-2/glypican may directly bind UNC-40/DCC, or alternatively, LON-2/glypican may instead interact with other molecules to impact UNC-40/DCC to modulate its stability, distribution, or activity. Alternatively, LON-2/glypican could potentially function as a co-receptor for UNC-6/netrin, where it may facilitate the formation of UNC-6/netrin-UNC-40/DCC-LON-2/glypican signaling complexes, similar to the situation in FGF signaling [15]. It is also conceivable that LON-2/glypican could bind UNC-6/netrin directly as well, even if undetected in our assays, as netrin has been found to bind heparin in vitro [3,52,53]. Previous studies have also documented the binding of DCC to heparin in vitro [7,8], and while we have found that the core protein is the critical portion of LON-2/glypican in netrin-mediated axon guidance, it remains possible that the endogenous HS chains contribute to the function of LON-2/glypican in axon guidance.
In summary, our studies uncover a novel mechanism by which UNC-6/netrin signaling through its UNC-40/DCC receptor is modulated by the HSPG LON-2/glypican during axon pathfinding. Given the evolutionary conservation of the UNC-6/netrin pathway components (UNC-6/netrin and its receptors UNC-40/DCC and UNC-5/UNC5) and of glypicans (LON-2 is most similar to mammalian glypican-3) and that synthesis of HS chains is required for mammalian axons to respond to netrin-1 in vitro [9,10], glypicans are likely to play a role in netrin-mediated axon pathfinding in mammals as well. Our findings provide a general mechanism for the extracellular regulation of growth cone responses to netrin during the development of nervous systems.
Nematode cultures were maintained at 20°C on NGM plates seeded with OP50 bacteria as described [54]. Strains were constructed using standard genetic procedures and are all listed in S8 Table. Genotypes were confirmed by genotyping PCR or by sequencing when needed, using primers listed in S9 Table.
Total RNA was extracted from worm samples using Trizol (Invitrogen) according to the manufacturer’s instructions. 500 ng RNA was used to reverse transcribe using the High Capacity cDNA Reverse Transcription Kit (Applied Biosystems) and random primers. PCR reactions were carried out with first-strand cDNA template, primers oCB834 (ATCAAGACCGAGTGATAGTG) and oCB1321 (TGGCGAGTATTCCCGTTTAG) were used for gpn-1 cDNA amplification, and primers oCB992 (TCGCTTCAAATCAGTTCAGC) and oCB993 (GCGAGCATTGAACAGTGAAG) were used for the control gene Y45F10D.4 [55] cDNA amplification.
Animals were mounted on agarose pads, anaesthetized with 100 mM sodium azide, and examined under a Zeiss Axio Scope.A1 or a Zeiss Axioskop 2 Plus.
All inserts of finalized clones were verified by sequencing.
Mixed-stage wild-type (N2), GFP control (lqIs4), LON-2::GFP (TLG257), and LON-2ΔGAG::GFP (TLG199) worms were collected in buffer and protease inhibitors (Roche). Worm pellets were subjected to repeated freeze-thaw cycles. Protein concentration was measured using the Pierce 660 nm Protein Assay on a Nanodrop. 70 μg of samples mixed with 2x Laemmli sample buffer (Bio-Rad) were boiled, separated by SDS-PAGE on a 4%–20% Mini-Protean TGX gel (Bio-Rad), and transferred to PVDF membrane. Membranes were incubated in 1:3000 anti-GFP primary antibody (Millipore #AB3080) and 1:9000 goat anti-rabbit HRP secondary antibody (Bio-Rad #166-2408EDU). For the loading control, membranes were incubated in 1:5000 anti-HSP90 antibody (CST #4874) and 1:10000 goat anti-rabbit HRP secondary antibody (Bio-Rad #166-2408EDU). Signal was revealed using Clarity Western ECL Substrate (Bio-Rad) and imaged using film (LabScientific).
All inserts of finalized clones were verified by sequencing.
S2 cells were transfected with HA::LON-2::myc (pCB313) and HA::LON-2ΔGAG::myc (pCB330) constructs. Cells were washed once with 1X Phosphate Buffered Saline and lysed for 30 min at 4°C in 1X Phosphate Buffered Saline, 0.5% Triton X-100, and 1X Protease Inhibitor Cocktail (Roche). Samples of supernatant and cell lysates were each mixed with 2X Laemmli sample buffer (BioRad). Proteins were separated by SDS-PAGE and transferred to PVDF membrane. Membranes were incubated with rabbit anti-HA (Life Technologies #715500) and rabbit anti-myc (Santa Cruz #sc-789) primary antibodies and HRP-linked goat anti-rabbit (Bio-Rad #166-2408EDU) secondary antibody. Signals were revealed by chemiluminescence with Clarity Western ECL Substrate (BioRad) and imaged using the ChemiDoc System (BioRad).
S2 cells were independently transfected with HA::LON-2::myc (pCB313), UNC-40::FLAG (pCB301), or SfGFP::UNC-6 (pCB292) constructs. 48 h after transfection, old culture medium was removed, and new medium was added to resuspend the cells. Equal volumes of resuspended cells that had been transfected with individual constructs were mixed and cocultured overnight. Cells were harvested, centrifuged, and combined with their corresponding supernatant from each of these cell mixes. 100 μL of supernatant of each mixture was saved and kept on ice. Cell pellets were washed once with 1X Phosphate Buffered Saline and lysed for 30 min at 4°C in 100 μL of ice-cold RIPA buffer (50 mM Tris HCl pH 7.5, 150 mM NaCl, 1% Triton-X100, 0.5% sodium deoxycholate, 0.1% SDS, and 1mM EDTA pH 8.0) supplemented with Protease Inhibitor Cocktail (Roche) and PMSF. Cell lysates were combined with their corresponding supernatant and mixed with 2X Laemmli sample buffer (BioRad). Each sample was split into three in order to run three protein gels in parallel. Proteins were separated by SDS-PAGE and transferred to PVDF membrane. Membranes were incubated with rabbit anti-myc (Santa Cruz #sc-789), mouse anti-FLAG (Sigma #F3165), and rabbit anti-GFP (Millipore AB3080) primary antibodies as well as HRP-linked goat anti-rabbit (Bio-Rad #166-2408EDU) and HRP-linked horse anti-mouse (Vector Labs PI-2000) secondary antibodies. Signals were revealed by chemiluminescence with Clarity Western ECL Substrate (BioRad) and imaged using the ChemiDoc System (BioRad).
S2 cells were maintained in SFX Insect Media (HyClone) containing 10% Fetal Bovine Serum (HyClone) and Penicillin-Streptomycin (50 units-50 μg/mL) (Sigma). 70%–90% confluent S2 cells were transfected with 500 ng of each construct using Effectene (Qiagen) according to the manufacturer’s protocol. 48 h after transfection, old culture medium was removed and new medium was added to resuspend the cells. Equal volumes of resuspended cells that had been transfected with individual constructs were plated onto coverslips and cocultured overnight. Cells were then fixed with 4% paraformaldehyde and immunostained with rabbit anti-HA (Life Technologies #715500) and mouse anti-FLAG (Sigma #F3165) primary antibodies and Alexa594 donkey anti-rabbit (Life Technologies #R37119) and Alexa647 goat anti-mouse (Life Technologies #A21235) secondary antibodies. Confocal analysis was performed on a Zeiss LSM 5 Pascal confocal microscope. Confocal images were processed using ImageJ. Each experiment was repeated at least three times.
For the experiment in which we use HA::LON-2-conditioned medium (supernatant) of cells expressing HA::LON-2, the culture medium was also changed 48 h after transfection, fresh medium was added, and the cells were incubated for another 48 h. This medium was collected and centrifuged at 1,500 rpm to remove cells and debris. This supernatant was added onto cells expressing UNC-40::FLAG, incubated overnight, and as above, fixed, stained, and imaged.
Independent populations of S2 cells were transfected with (1) 450 ng of pActin5.1::mCherry alone, (2) 50 ng of the UNC-40::FLAG construct plus 450 ng of the cotransfection marker pActin5.1::mCherry, or (3) 500 ng of HA::LON-2. The medium was changed and cells were mixed 48 h after transfection. Control mCherry-expressing cells were mixed with untransfected cells or with HA::LON-2-expressing cells. Similarly, UNC-40::FLAG/mCherry-expressing cells were mixed with untransfected cells or with HA::LON-2-expressing cells. To maintain the total number of cells constant in our different mixes, one volume of UNC-40::FLAG/mCherry cells was mixed with either (a) one volume of control/untransfected cells or (b) one volume of LON-2-transfected cells. Cell mixes were cocultured overnight. Cells were then fixed with 4% paraformaldehyde and examined under a Zeiss LSM 5 Pascal confocal microscope. Control mCherry-expressing cells or UNC-40::FLAG/mCherry-expressing cells were identified by the cotransfection marker mCherry. 20 fields of ~300 cells each per mix per were photographed for each of three independent experiments. Cells were categorized as having the typical S2 cell round and smooth shape, irregular edges, and/or extensions protruding from the cell.
|
10.1371/journal.pgen.1004715 | Genetic Analysis of a Novel Tubulin Mutation That Redirects Synaptic Vesicle Targeting and Causes Neurite Degeneration in C. elegans | Neuronal cargos are differentially targeted to either axons or dendrites, and this polarized cargo targeting critically depends on the interaction between microtubules and molecular motors. From a forward mutagenesis screen, we identified a gain-of-function mutation in the C. elegans α-tubulin gene mec-12 that triggered synaptic vesicle mistargeting, neurite swelling and neurodegeneration in the touch receptor neurons. This missense mutation replaced an absolutely conserved glycine in the H12 helix with glutamic acid, resulting in increased negative charges at the C-terminus of α-tubulin. Synaptic vesicle mistargeting in the mutant neurons was suppressed by reducing dynein function, suggesting that aberrantly high dynein activity mistargeted synaptic vesicles. We demonstrated that dynein showed preference towards binding mutant microtubules over wild-type in microtubule sedimentation assay. By contrast, neurite swelling and neurodegeneration were independent of dynein and could be ameliorated by genetic paralysis of the animal. This suggests that mutant microtubules render the neurons susceptible to recurrent mechanical stress induced by muscle activity, which is consistent with the observation that microtubule network was disorganized under electron microscopy. Our work provides insights into how microtubule-dynein interaction instructs synaptic vesicle targeting and the importance of microtubule in the maintenance of neuronal structures against constant mechanical stress.
| Axons and dendrites are two classes of neuronal process that differ in their functions and molecular compositions. Proteins important for synaptic functions are mostly synthesized in the cell body and sorted differentially into the axon or dendrites. Microtubules in the axon and dendrite maintain their structural integrity and regulate polarized protein transport into these compartments. We identified a novel α-tubulin mutation in C. elegans that caused mistargeting of synaptic vesicles and induced progressive neurite swelling, which resulted in late-onset neurodegeneration. We showed that this tubulin mutation weakened microtubule network and abnormally increased microtubule affinity for dynein, a motor protein responsible for cargo sorting to the dendrite. This enhanced microtubule-dynein affinity is due to augmented negative charge at the carboxyl terminus of α-tubulin. Neurite swelling and neurodegeneration could be ameliorated by reduced physical activity, suggesting that recurrent mechanical strain from muscle contraction jeopardized neurite integrity in the long run. Mutations in α- and β-tubulins are found in human neurological diseases; our findings therefore contribute to understanding the pathogenic mechanism of human neurological diseases associated with tubulin mutations.
| Microtubule and molecular motors mediate polarized transport of neuronal proteins to either axons or dendrites [1]. Microtubules are oriented uniformly with their plus ends towards the distal end of the axon, which facilitates kinesin-dependent targeting of presynaptic proteins [2]. By contrast, targeting of postsynaptic molecules to the dendrite, such as glutamate receptors, requires the minus end-oriented dynein motors [3], consistent with the fact that many microtubules in the dendrites orient the minus end distally [4]. Mislocalization of presynaptic proteins to the dendrite occurs when this polarized pattern of axon-dendritic microtubule arrays is disrupted [5], [6], when kinesin function is compromised [7], [8], or when dynein activity is inadvertently increased [3], [7], [8]. Synaptic vesicle (SV) precursors are generated in the neuronal cell body and transported to the synapses by the unidirectional motor Kinesin 3/KIF1A [1]. On the other hand, the dynein motor complex mediates retrograde SV transport in the axon [1], [9]. Since SVs are cargos for both KIF1A and dynein, it is intriguing that they are exclusively targeted to the axon and prevented from entering the dendrites.
Previous biochemical and structural studies suggest that kinesin and dynein share an overlapping binding region at the C-terminus of α-tubulin [10]. The N-terminus of the H12 helix of the α-tubulin contains a stretch of absolutely conserved acidic residues (414EEGE, equivalent to 415EEGE in the yeast α-tubulin) and interacts with ATP-bound KIF1A [11]. A recent study on dynein structures also implicates this region in the interaction between microtubule and the microtubule-binding domain (MTBD) of dynein [12], although validation of this model in the context of in vivo, eukaryotic system is still lacking. Mutations of any of the three glutamic acids in the yeast α-tubulin to alanine dramatically reduced the frequency of kinesin binding to the microtubules [13]. Mutations of several conserved, acidic residues in the H12 helix of β-tubulin (E410, E412, D417) to alanine similarly reduced microtubule affinity for kinesins. Interestingly, E410K, D417H and D417N in the human β-tubulin TUBB3, among other point mutations, had been found in patients with congenital neurological syndrome with ophthalmoparesis and peripheral neuropathy [14]. In particular, TUBB3(E410K) and TUBB3(D417H), but not other disease-related TUBB3 mutations, were shown to impair the bindings of Kinesin 1/KIF5, Kinesin 3/KIF1A and KIF21 when expressed in cultured mammalian neurons [15]. The interaction between dynein and microtubule was not affected by these TUBB3 mutants [15]. These studies established the critical importance of negative charge in the H12 helix of the α- and β-tubulins in mediating microtubule-kinesin interaction, but the molecular mechanisms governing microtubule-dynein interaction and its physiological significance remain unexplored.
Here we describe a novel mutation of G416 in the α-tubulin MEC-12 of C. elegans to glutamic acid (G416E). Homologous mutations at this site of α-tubulin had not been reported in human diseases or tested in genetic model organisms. In C. elegans, mec-12 is highly expressed in the six touch receptor neurons that detect gentle mechanical stimulation on the worm cuticle [16], [17]. MEC-12 and the touch neuron-specific β-tubulin MEC-7 are required to form the unusual, 15-protofilament giant microtubules in these neurons [16]. These giant microtubules had been implicated in the transduction of mechanosensation, although the mechanisms remain enigmatic [18]. We show that this gain-of-function G416E mutation redirects SVs to non-axon compartment in the C. elegans mechanosensory neuron PLM, and it does so by increasing microtubule affinity for dynein.
In wild-type C. elegans, the bilaterally symmetric touch receptor neurons ALM and PLM develop a single anterior process that forms synapses in the nerve ring and in the ventral nerve cord, respectively (Figure 1A). Touch neuron synapses are enriched in RAB-3-(+) synaptic vesicles (SVs), the active zone protein SYD-2/Liprin-α, and mitochondria (Figure 1B, 1D, and S1) [19], [20]. The PLM neurons also have a short posterior process that does not form synapses. In an EMS mutagenesis screen (see Materials and Methods), we identified gm379, a mutant with prominent SV phenotypes in the touch neurons (Figure 1B′-1G′). gm379 animals lacked RAB-3-(+) SVs at the touch neuron synapses, and instead SVs accumulated in the neuronal soma (Figure 1B′-1E′, 1H). We refer to these as SV transport defects for the rest of the paper. Surprisingly, SVs were also redirected to the PLM posterior process, a phenotype that we call SV mistargeting (Figure 1E′, 1I). These results were confirmed using two other SV reporters, jsIs37(Pmec-7::SNB-1::GFP) that marked the SV membrane protein synaptobrevin/SNB-1, and jsIs219(Psng-1::SNG-1::GFP) labeling another SV protein synaptogyrin/SNG-1, with GFP (Figure 1F-1G′). These ectopic SVs showed very limited motility, and many were stationary (Figure 1J). We followed gm379 mutants through development, and confirmed that SV transport defects and mistargeting were present at early larval stages and progressively worsened (Figure 1H). The transport and targeting of synaptic active zone protein SYD-2 was affected to a much milder degree, and surprisingly, SYD-2 failed to mistarget to the PLM posterior process (Figure S1). The dissociation in the mutant phenotypes of SV and active zone proteins indicates that the gm379 mutation caused relatively specific defects in SV targeting rather than generally impaired axon transport or induced ectopic synapse formation.
In addition to SV transport defects, the gm379 mutant touch neurons had progressive neurite swelling and misshapen soma (Figure 2A and S2A). Neurite defects evolved from small beadings at early larval stages into triangular-shaped swellings in L4 and adult animals, and mitochondria were frequently found to be present at the swellings (Figure 2A, 2B and S2A). These swellings were dynamic in morphology, as movements often induced reversible buckling of the neurite and changed the width and height of the swellings (Figure 2C and Video S1), which was similar to what had been described earlier for the tubulin acetyltransferase mutant mec-17 [21]. Neurite buckling or swelling were never seen in the wild type even under maximal muscle contraction induced by levamisole. This observation suggests that the gm379 mutation rendered the touch neurite susceptible to deformation under mechanical strain, a phenotype that was also seen when the membrane skeleton protein UNC-70/β-spectrin was lost [22]. We therefore test whether genetic paralysis of the animals suppresses neurite defects of the gm379 mutant. Mutation in the muscle myosin gene unc-54 almost completely paralyzed the animals, and it significantly reduced the number of neurite swellings in the gm379 mutant touch neurons (Figure 2D and 2E). This result implies that the gm379 mutation compromises the ability of the touch neurites to cope with mechanical stress.
We were curious whether massive neurite swellings in the gm379 mutant predispose touch neurons to degeneration. We first performed longitudinal imaging of individual ALM and PLM neurons through adulthood, and found that touch neurites in the gm379 mutant underwent progressive disorganization (Figure 2F). Touch neuron degeneration, characterized by swelling of neuronal soma, neurite interruption and extensive beadings, began to emerge in the gm379 mutant at D9 and progressively increased (Figure 2G and 2H). Touch neuron degeneration was extremely rare in the wild type at comparable age (Figure 2G and 2H) [23]. Interestingly, touch neuron degeneration of the gm379 mutant was suppressed by the unc-54 mutation (Figure 2G and 2H). These results indicate that recurrent mechanical strain imposed on the touch neurons during locomotion is an important precipitating factor for late-onset neurodegeneration in the gm379 mutant.
Under serial thin-section electron microscopy, we found that the characteristic 15-protofilament microtubules of C. elegans touch neurons were preserved in the gm379 mutant, including those in the PLM posterior process (Figure 3A-C, S2B-D). Mitochondria could be found where touch neuron processes swelled abnormally (Figure 3B), consistent with our light microscopic observation (Figure 2B). In longitudinal sections, in contrast to the wild type, where neuronal microtubules formed long straight bundles, touch neuron microtubules in the gm379 animals curved focally at sites of organelle accumulation (Figure 3D). We observed bending and splitting of neuronal microtubule bundles at sites of mitochondria accumulation in focal axonal swellings (Figure 3D4). These ultrastructural studies suggest that the microtubule network of the touch neurons is abnormal in the gm379 mutant.
We cloned gm379 by single nucleotide polymorphism (SNP) mapping, complementation test and DNA sequencing, and found that it contained a missense mutation of the touch neuron-specific α-tubulin mec-12 that alters an absolutely conserved C-terminal glycine residue to glutamate (G416E) [19], [17] (Figure 4A and 4B). Interestingly, gm379 animals were touch-insensitive (30% touch-sensitive, compared to wild type, 89%; and mec-12(e1607) null, 12%, n>35). We found another intron mutation in gm379 that was distant to exon-intron junctions (nucleotide 828 of unspliced transcript, G to A mutation). Two null mutants of mec-12, e1607 and tm5083, also had SV transport defects, but not SV mistargeting in the PLM or axon swelling in the touch neurons (Figure 4C, S3). RNAi against mec-12 in the gm379 mutant almost completely abolished axon swelling or SV mistargeting, indicating that these two phenotypes were neomorphic (Figure 4C, 4D). Expression of the MEC-12(G416E) mutant tubulin in the mec-12 null mutants recapitulated the SV mistargeting phenotypes of the mec-12(gm379) mutant (67% of transgenic animals showed SV mistargeting, n = 21 v.s. 6.5% of array-loss siblings showing SV mistargeting, n = 77), confirming that SV mistargeting and neurite swelling were indeed caused by the mec-12(gm379) rather than other unidentified mutations in the background.
To gain further insights into the genetic nature of the mec-12 phenotypes, we analyzed various heterozygous mec-12 mutants as well as trans-heterozygotes between these alleles (Figure 4E). Heterozygous animals containing the e1607, tm5083 or gm379 mutations all showed moderate SV transport defects that were less severe than the homozygous animals. These observations suggest that loss of mec-12 functions causes semi-dominant SV transport defects due to haploinsufficiency. Heterozygous gm379 mutants did not display neurite swelling or SV mistargeting. By contrast, mec-12(gm379)/mec-12(e1607) and mec-12(gm379)/mec-12(tm5083) trans-heterozygous mutants had SV mistargeting without neurite swelling, implying that the presence of wild-type MEC-12 somehow prevents MEC-12(G416E) from mistargeting SVs or disrupting microtubule organization. In conclusion, our genetic analysis indicates that gm379 causes both neomorphic (neurite swelling and SV mistargeting) and semi-dominant loss-of-function (SV transport defects) phenotypes of mec-12.
Stable microtubules formed in the gm379 touch neurons, based on the EM data and the preserved lysine 40 (K40) acetylation of α-tubulin [17], [24], [25] (Figure S4A). Moreover, a mec-7/β-tubulin null mutation completely suppressed neurite swelling and dramatically reduced SV mistargeting of mec-12(gm379) (Figure S5), suggesting that SV mistargeting and neurite swelling phenotypes require intact microtubules. Labeling touch neuron microtubules with the plus end-binding protein EBP-2::GFP showed that in the wild type, microtubules oriented plus-end distally in the anterior PLM process (Figure S6) [26]. In the PLM posterior process, microtubules showed mixed polarity (Figure S6). These patterns of microtubule polarity were preserved in the mec-12(gm379) mutant (Figure S6).
A recent study reported that mutations in the tubulin acetyltransferase mec-17 caused neurite degeneration with SV mislocalization in the touch neurons [27]. To test whether altered microtubule posttranslational modifications are responsible for SV mistargeting in the mec-12(gm379) mutant, we performed immunostaining experiments, but did not observe gross difference in microtubule acetylation or tyrosination in the touch neurons between the wild type and the mec-12(gm379) mutant (Figure S4A). Tubulin polyglutamylation signal was restricted to the amphid and phasmid sensory cilia as in the wild type, and was not ectopically expressed in the mutant touch neurons (Figure S4B). Moreover, we performed feeding RNAi to knock down the tubulin polyglutamylase ttll-4, the tubulin deglutamylase ccpp-6, and a few genes (ttll-5, ttll-12, ttll-15 and ccpp-1) that bear sequence homology to human tubulin amino acid ligases (Wormbase at http://www.wormbase.org) [28], [29], in both wild type and the mec-12(gm379) mutant. None of these RNAi resulted in SV mistargeting in the wild type or suppressed SV mistargeting in the mec-12(gm379) animals. Based on these results, we conclude that SV mistargeting in the mec-12(gm379) is not a consequence of altered microtubule posttranslational modifications.
It is possible that the interaction between KIF1A and microtubule was altered by the G416E mutation. We found that the strong loss-of-function unc-104(rh43)/KIF1A mutation caused completely penetrant SV transport defects and, surprisingly, low percentage of SV mistargeting in the PLM (Figure 5A and 5B). Moreover, this unc-104 mutation enhanced SV mistargeting of mec-12(gm379) rather than suppressing the phenotype, with more SVs mistargeted to the PLM posterior process and distributed more distally (Figure 5C, 5D). This result suggests that SV mistargeting in the mec-12(gm379) mutant is not caused by aberrant UNC-104 activity. Furthermore, overexpression of UNC-104 significantly rescued SV transport defects and mistargeting in the mec-12(gm379) mutant, with more SVs reaching synapses in the nerve ring or entering the anterior ALM and PLM processes (Figure 5E-5H, Figure S7A, S7B). While these data are consistent with the interpretation that UNC-104 activity was reduced in the mec-12(gm379) mutant, resulting in severe SV transport defects, they also indicate that SV mistargeting in the mutant requires the activity of an unknown molecule.
We wondered whether increased activity of the minus end motor dynein is responsible for SV targeting to the PLM posterior process in the mutant, based on the presence of minus end-out microtubules in the PLM posterior process and the unc-104 effects. dhc-1 encodes the heavy chain for cytoplasmic dynein in C. elegans [30]. If enhanced dynein activity is responsible for SV mistargeting in the mutant, elimination of dynein function should suppress it. We could observe SV mistargeted to the PLM posterior process as early as 2-3 fold embryos, before the animal hatched. With the available dhc-1 mutant alleles, it was not possible to lose DHC-1 functions at such early stages without compromising animals' viability. To eliminate DHC-1 functions as early as possible, and to circumvent lethality due to widespread DHC-1 loss, we specifically knocked down dhc-1 in the touch neurons, but not in other somatic tissues, by simultaneously expressing sense and antisense dhc-1 from the mec-7 promoter, which we named transgenic dhc-1 RNAi. Strikingly, transgenic dhc-1 RNAi significantly suppressed SV mistargeting of the mec-12(gm379) mutant, with about one third of the transgenic animals completely devoid of mistargeted SVs (Figure 6A, 6B). This result was confirmed by another independently generated dhc-1 RNAi array (Figure S8A). Transgenic dhc-1 RNAi also significantly reduced SV mistargeting in the unc-104; mec-12(gm379) mutant (Figure 6C). In the wild type, transgenic dhc-1 RNAi had little effects on the intensity of GFP::RAB-3 or SNB-1::GFP in the PLM soma or synapses (Figure S8B). These data indicate that SV mistargeting in the mec-12(gm379) mutant is mediated by the dynein motor. The neurite swelling phenotypes of the mutant, by contrast, were not changed by dhc-1 RNAi, suggesting that SV mistargeting and neurite defects are mechanistically distinct.
We noted that eliminating DHC-1 functions in the mec-12(gm379) mutant significantly restored synaptic targeting and anterograde transport of SVs, with concomitant reduction of SV accumulation in the touch neuron soma (Figure 6A and 6D). These effects were similar to those caused by excess UNC-104 (Figure 5E-H, S7A, S7B). Indeed, synaptic targeting and anterograde transport of SVs were completely abolished in the dhc-1; unc-104; mec-12(gm379) triple mutant, suggesting that the phenotypic rescue caused by the dhc-1 mutation requires UNC-104 (Figure 6D). Although significantly increased compared to those in the mec-12(gm379) mutant, SV signals in the nerve ring synapses of the dhc-1; mec-12(gm379) were still much weaker than the wild type. Together with the severe reduction of presynaptic SVs in the mec-12(gm379) mutant, these observations suggest that UNC-104 was not able to display a full range of activity on the mutant microtubule scaffolds.
To test whether the MEC-12(G416E) mutant microtubules have increased affinity for dynein, we performed microtubule sedimentation in transgenic strains expressing MEC-12 and MEC-7 pan-neuronally, and assayed for the amount of DHC-1 or UNC-104 associated with microtubules by blotting with anti-DHC-1 and anti-UNC-104 antibodies, respectively. The presence of MEC-12-containing, stable microtubules was verified by detecting 6-11B-1 antibody immunoreactivity in neurons other than the touch receptors (). After normalization to tubulin, the amount of DHC-1 co-sedimented with MEC-12(G416E)-containing microtubules dramatically increased, compared to that co-sedimented with wild-type microtubules (Figure 6E). In support of this conclusion, we found that GFP::DHC-1, expressed from the low-copy integrated transgene orIs17(Pdhc-1::gfp::dhc-1), which drives DHC-1 expression from the endogenous promoter, was accumulated in the PLM posterior process in the mec-12(gm379), but not in the wild type (Figure 6F). Together these results indicate that mutant microtubules have increased affinity for DHC-1. To our surprise, we did not detect a change in the amount of UNC-104 co-sedimented with mutant microtubules (Figure 6E) or a change of UNC-104 protein localization in the PLM neuron. This implies that UNC-104 can still associate with the mutant microtubules, although its function is somehow disrupted.
The aforementioned data indicate that G416 of MEC-12 plays a critical role in determining the relative affinity of microtubules for dynein. To further decipher the mechanisms that instruct microtubule-dynein affinity, we systemically replaced G416 with acidic (aspartic acid/D) or basic (lysine/K, arginine/R) residues, as well as alanine (A) and glutamine (Q), the latter being similar to glutamic acid in side chain length but did not carry charges (Figure 7A, 7B). These MEC-12 species were expressed in the touch neurons of the mec-12(e1607) null mutant. SV mistargeting was seen only with the expression MEC-12(G416D), but not other G416 substitutions (Figure 7A, 7B). These results suggest that SV mistargeting in the gm379 mutant was caused by the increased negative charges at the EEGE cluster of MEC-12.
Is the arrangement of the acidic residues important for dynein affinity of microtubules? To answer this question, we moved the glycine to residue 414, 415 or 417, with reciprocal glutamic acid substitution at 416 (E414G/G416E, E415G/G416E, G416E/E417G, referred as GEEE, EGEE and EEEG, respectively; Figure 7A, 7B), so that the arrangement, but not the sum, of negative charges was altered, and asked whether this manipulation affects SV targeting. While expression of MEC-12(GEEE) or MEC-12(EGEE) did not result in significant SV mistargeting, expression of MEC-12(EEEG) triggered SV mistargeting in about 30% of the mec-12(e1607) animals (Figure 7A, 7B). These observations indicate that SV targeting critically depends on the magnitude and the spatial arrangement of negative charges in the EEGE cluster of the H12 helix.
Previous structural studies suggest that dynein binds the H12 helix of α-tubulin [10], [12]. Redwine et al. proposed that E3378 and R3382 of the dynein microtubule-binding domain (MTBD) form intramolecular salt bridge, and upon approaching the tubulin dimer, negative charges of the α-tubulin H12 disrupt this salt bridge by attracting R3382, which carries positive charges (Figure 8C). An E3378K mutation of MTBD disrupted this intramolecular salt bridge and increased both the affinity and the run length of dynein on the microtubules [12]. We speculate that the E3378K mutation facilitates the electrostatic interaction between MTBD and the negative charges of the α-tubulin H12 domain. To test this, we expressed and purified a fragment of C. elegans DHC-1 MTBD, and showed that this MTBD precipitated with microtubules synthesized from purified bovine tubulin in the in vitro sedimentation experiment (Figure 8A and S9). Moreover, D3323K mutation, which is equivalent to E3378K mutation of the yeast dynein MTBD, enhanced MTBD-microtubule interaction across a range of tested concentrations (Figure 8A, 8B and S9). This result supports our hypothesis that exaggerated microtubule affinity for dynein critically depends on the electrostatic interaction between the EEGE-containing H12 helix of α-tubulin and the dynein MTBD (Figure 8C).
In the present study, we characterized a gain-of-function mec-12/tubulin mutant that displayed synaptic vesicle mistargeting and neurite swelling phenotypes, which were absent in the mec-12 null mutant. In this mutant, single amino acid substitution augments microtubule-dynein interaction, thereby mistargeting synaptic vesicles to non-axon compartments. This observation extends previous structural studies on the charged cluster of the α-tubulin H12 helix and provides a biological context for such charge-based coupling between microtubule and dynein. The neurite swelling and degeneration phenotypes could bear important implications for human neurological diseases associated with missense tubulin mutations, as discussed below.
Previous structural studies suggest that dynein binds the H12 helix of α-tubulin, and the interaction between dynein and microtubule is not as strong as that between kinesins and microtubule [10], [12]. One advantage for this suboptimal dynein-microtubule affinity is flexibility and dynamic range for dynein processivity on the microtubule scaffolds. The hypothesis that dynein-microtubule interaction is not optimized also suggests that it is possible to further enhance dynein-microtubule affinity by mutating residues at this interface. Here, we show that this is indeed the case: by increasing negative charges of the α-tubulin H12 helix or augmenting positive charges of the dynein MTBD, microtubule-dynein interaction was strengthened. More importantly, we show that this aberrantly high affinity of microtubule for dynein redistributed synaptic vesicles to ectopic compartments in the neurons. We speculate that the suboptimal affinity of microtubule for dynein prevents initial SV entry into the dendrite, yet allows for retrograde SV transport within the axon. The G416E mutation of MEC-12 increased microtubule affinity for dynein probably at the expense of dynein processivity: we found that motility of the mistargeted SVs in the PLM posterior process was profoundly compromised (Figure 1J). Consistent with this view, Redwine et al. showed that E3378K mutation of MTBD caused about 40% reduction in the velocity of dynein movements on microtubules. By contrast, none of the G416 substitutions restored the UNC-104/KIF1A-dependent, anterograde SV transport in the mec-12(e1607) null mutant. This is also consistent with the observation by Redwine et al. that microtubule-kinesin interaction was molecularly optimized, therefore changes to the tubulin residues at the microtubule-kinesin interface will only compromise but not compensate for or even further improve this interaction.
Rather than merely serving as a track, mounting evidence indicates that microtubules play an active role in directing axon transport [2]. One such example is the various posttranslational modifications on microtubules, which regulate axon transport by restricting different kinesin motors to discrete subcellular compartments [31], [32]. The intrinsic organization of microtubule is another critical factor regulating polarized axon transport. UNC-33 and UNC-44, the C. elegans homologs for the Collapsin Response Mediator Protein 2 (CRMP2) and neuronal ankyrin, respectively [33], [34], regulate axon transport by maintaining uniform microtubule polarity in the axon and the dendrite [5]. In C. elegans head neurons, microtubules are uniformly oriented with their minus-ends towards the distal of the dendrite. In the unc-33 and the unc-44 mutants, dendritic microtubules showed mixed polarity, which resulted in aberrant sorting of axonal proteins into the dendrite [5]. In these mutants, the dendritic localization of SVs requires UNC-104. Of note, mistargeting of axonal proteins in the unc-33 and unc-44 mutants occurred to the SVs and multiple active zone components. By contrast, SYD-2 was not mistargeted in the mec-12(gm379) mutant, and we found that SV transport in the touch neurons was not affected in unc-33 and unc-44 mutants. These observations indicate that while perturbation to the gross architecture of microtubule causes extensive mistargeting of multiple axonal cargos, changes at a restricted, yet functionally critical site of microtubule could lead to targeting defects of specific axonal proteins.
Many tubulin mutations associated with human diseases are point mutations that alter protein function rather than eliminating protein products [35]–[37]. Mutant tubulins are still incorporated into microtubule polymers; phenotypes presumably arise from altered microtubule dynamics or disrupted interactions with molecular motors or microtubule-associated proteins [14], [15], [38]. The G416E mutation of MEC-12 caused extensive neurite swellings that eventually led to degeneration of the touch neurons. Ultrastructurally, there were unbundling of microtubules and accumulation of mitochondria at focal neurite swellings. Interestingly, these neurite defects were independent of dynein activity, but could be suppressed by genetic paralysis of locomotion. In light of two recent studies that showed touch neurite buckling or swelling when unc-70/β-spectrin or mec-17/tubulin acetyltransferase was mutated [21], [22], this suggests that microtubules confer resistance of touch neurites to deformation imposed by constant muscle activity. When this structural resilience is compromised, touch neurons are likely to degenerate presumably in a wear-and-tear fashion induced by the animal's constant movements. Supporting this notion, commissural axons in unc-70 mutants also increasingly break as the animals grow and move, and paralysis of the animals prevents axon interruption [39]. Unlike the mec-17 mutant, in which defective tubulin acetylation was associated with abnormal microtubule protofilament number, tubulin acetylation and microtubule protofilaments were unaffected in the mec-12(gm379) mutant. We hypothesize that the G416E substitution may alter the association of microtubule-binding proteins with microtubules, changing the stability or structure of microtubule lattice and rendering the neurites susceptible to mechanical strain. It would be interesting to test whether any of the tubulin mutations found in human diseases generates similar axon defects and could also be ameliorated by reduced activity of neighboring musculature. The molecular mechanism by which G416E mutant microtubules cause axon degeneration awaits future investigation.
Strains were cultured as described [40]. The following alleles were used in this study: N2 (Bristol strain), CB4856, LG I: dhc-1(or283ts), unc-54(e190); LG II: unc-104(rh43); LG III: mec-12(e1607) (a gift from Martin Chalfie, Columbia University), mec-12(tm5083), mec-12(gm379); LG V: sid-1(pk3321); him-5(e1490); LG X: mec-7(ok2152). Transgenes used in the current study are: jsIs37(Pmec-7::SNB-1::GFP)/IV, jsIs219(Psng-1::SNG-1::GFP)/II, jsIs821(Pmec-7::GFP::RAB-3)/X, jsIs973(Pmec-7:mRFP)/III, jsIs1111(Pmec-4::UNC-104::GFP), jsIs1238(Pmec-7::SYD-2::GFP) (jsIs821, jsIs973, jsIs1111, and jsIs1238 are gifts from Michael Nonet, Washington University), juIs76(Punc-25::GFP)/II, Punc-17::RFP/V (a gift from Joshua Kaplan, Massachusetts General Hospital), orIs17[Pdhc-1::GFP::DHC-1, unc-119(+)] (a gift from Bruce Bowerman, University of Oregon), otIs118(Punc-33::GFP)/IV, uIs71(Pmec-18::SID-1, Pmyo-2::mCherry), zdIs5[Pmec-4::GFP, lin-15(+)]/I, twnEx8(Pmec-7::TOMM20::mCherry, Pmyo-2::gfp) (“Pmec-7::mito::mCherry”), twnEx40(Pmec-7::GFP::EBP-2, Pdpy-30::dsRed), twnEx42[Pmec-7::dhc-1(RNAi), Pmyo-2::GFP], twnEx55[Pmec-7::MEC-12(G416E, E417G), Pdpy-30:: NLS::dsRed], twnEx73[Pmec-7::MEC-12(G416E) 5 ng/µl, Pdpy-30::NLS::dsRed], twnEx74[Pmec-7::MEC-12(G416D), Pdpy-30:: NLS::dsRed], twnEx75[Pmec-7::MEC-12(G416Q), Pdpy-30:: NLS::dsRed], twnEx76[Pmec-7::MEC-12(G416A), Pdpy-30:: NLS::dsRed], twnEx77[Pmec-7::MEC-12(G416K), Pdpy-30:: NLS::dsRed], twnEx78[Pmec-7::MEC-12(G416R), Pdpy-30:: NLS::dsRed], twnEx79[Pmec-7::MEC-12(E414G, G416E), Pdpy-30:: NLS::dsRed], twnEx80[Pmec-7::MEC-12(E415G, G416E), Pdpy-30:: NLS::dsRed], twnEx88[Pmec-7::MEC-12(G416E) 10 ng/µl, Pdpy-30::NLS::dsRed], twnEx89[Pmec-7::dhc-1(RNAi), Pdpy-30:: NLS::dsRed], twnEx98[Punc-119::MEC-12, Punc-119::MEC-7, unc-119(+)], twnEx99[Punc-119::MEC-12(G416E), Punc-119::MEC-7, unc-119(+)], Ex(Punc-104::UNC-104::GFP), Ex(Punc-104::UNC-104::mRFP) (both from Oliver Wagner, National Tsing-Hua University, Taiwan). mec-12(e1607) is a G to A point mutation at nucleotide 430 of mec-12 cDNA, resulting in glycine to serine mutation at amino acid 144. Germ line transformation was performed by microinjection of purified DNA of interest as described [41].
Initially, an EMS mutagenesis screen was performed in the zdIs5; cwn-1(ok546) animals to identify mutations that cause synthetic polarity defects of the touch neurons [42]–[44]. While causing no defects in neuronal polarity, gm379 was recovered due to its prominent axonal defects. The cwn-1 mutation was then removed from the mutant. The SV transport defects, SV mistargeting and axonal swellings were all independent of the cwn-1(ok546) mutation in the background, and the cwn-1 mutant displayed none of the gm379 phenotypes.
Single nucleotide polymorphism mapping was performed as described [45]. zdIs5 was included in this mapping to assist the identification of the homozygous gm379 mutants. In brief, male animals of the Hawaiian strain CB4856 were crossed to zdIs5; gm379, and gm379 homozygotes were later recovered from F2 progeny. F3 animals from individually cloned F2 animals were washed off plates and genomic DNA extracted by proteinase K treatment. To map gm379, we selected 48 SNPs from the 5 autosomes and the sex chromosome, and semi-quantitatively determined the ratio of N2/Hawaiian SNP for each locus using the restriction enzyme DraI. Our SNP mapping located gm379 to a region between -12 and +7 MU of Chromosome III, a region that contains the mec-12 locus.
Feeding RNAi was performed as described [46], with 1 mM IPTG pre-induction for 2 hours. For touch neuron-specific RNAi, we used jsIs973(Pmec-7::mRFP) mec-12(gm379); sid-1(pk3321); jsIs821(Pmec-7::GFP::RAB-3); uIs71(Pmec-18::SID-1, Pmyo-2::mCherry) animals [47]. The only RNAi-sensitive cells in the sid-1; uIs71 background are the six touch receptor neurons. Five L4 animals were placed on the RNAi plates and cultured at 20°C. The F1 progeny of were then transferred to another freshly prepared RNAi plates at L4, and their progeny (F2) scored for axon and synaptic vesicle phenotypes. Each RNAi experiment was repeated three times to confirm the results. Efficiency of feeding RNAi against neuronal genes in this genotype was confirmed by mec-12 RNAi, which resulted in 55% animals losing the PLM branch (n = 22), with control RNAi having no effects (0%, n = 27). This penetrance was similar to what was observed in the mec-12(e1607) null mutant (58%, n = 60). Additional RNAi control included mec-7 and rho-1 and all showed results comparable with mutant analysis.
Cloning and construction of plasmids were performed with standard molecular biology techniques. All expression constructs in the twnEx series transgenes were in the pPD95.77 Fire vector backbone, which contains the unc-54 3′-UTR for optimized expression in C. elegans. Primer sequence information is available upon request.
Neurite swelling of ALM and PLM was scored in live animals with the integrated GFP reporter zdIs5(Pmec-4::GFP), which is expressed in the six mechanosensory neurons: ALMs, PLMs, AVM and PVM. Beading is defined as oval or round swelling along primary axons. Neurite swelling is defined as triangular protrusion or looping of axonal membrane. Neurodegeneration is define as swelling and round-up of the neuronal soma with neurite interruption, thinning and large beading formation. To characterize the evolution of neurite defects in mutants, wild type and gm379 animals were synchronized by hatching and arresting early L1 in M9 at 20°C. Animals were then allowed to feed on regular E. coli plates with axon morphology scored at different time points (6 hr, 12 hr, 24 hr, 36 hr, 48 hr, and 60 hr post hatching) that correspond to distinct larval and adult stages. Because unc-54 animals are defective in egg laying and die from progeny hatched inside their bodies, 5-fluoro-2′-deoxyuridine (FUdR) was added to the plate at the final concentration of 50 µM to stop progeny production. FUdR was applied to the mec-12(gm379) mutant and the wild type in experiments where unc-54 was also tested.
Synaptic vesicles in the touch neurons were visualized with the integrated GFP reporter jsIs821(Pmec-7::GFP::RAB-3) which labels synaptic vesicles in the six touch neurons. The authenticity of synaptic vesicle defects in the gm379 mutant was confirmed with another GFP reporter, jsIs37(Pmec-7::SNB-1::GFP). Touch neurons were simultaneously labeled by the RFP reporter jsIs973(Pmec7::mRFP). Animals were synchronized and jsIs821-labeled synaptic vesicles quantified at distinct developmental stages. Images were acquired using the 63x Carl Zeiss Apochromat objective and the Zeiss AxioImager M2 imaging system. Because the posterior PLM process was very thin (less than 0.5 µm, see Figure S2B and S2C), we did not take confocal z-axis image stacks for pixel quantification. Pixel density was derived using ImageJ by quantifying total pixel number divided by the area marked by the neuronal marker jsIs973. We excluded the neuronal nucleus when quantifying pixel density of the soma. For Figure 5G, because the UNC-104-overexpression array carries mCherry fused to UNC-104, which may complicate determination of neurite area by the jsIs973 marker, we decided to quantify total pixel number of fluorescence on the entire PLM posterior process. Mistargerted SVs often formed GFP::RAB-3 aggregates of variable size, and we therefore did not quantify GFP punctum number or individual punctual intensity. The length of the PLM posterior process was measured by the software Axio Vision Rel. 4.8. The distance of synaptic vesicle distribution was determined as the fraction of the PLM posterior process marked with synaptic vesicle GFP. All image quantification was done blind to avoid bias.
Worms were high pressure frozen in either a Bal-Tec HPM 010 (Bal-Tec AG, Liechenstein) or Leica HMP 100 (Leica Microsystems, Vienna) high pressure freezer and freeze substituted in 1% osmium tetroxide and 0.1% uranyl acetate in acetone over a period of 2 hours by the SQFS method of McDonald and Webb [45]. Infiltration of Epon epoxy resin was carried out by 15 minute incubations in 25, 50, and 75% acetone-resin mixtures on a rocker, then three 15 minute incubations in pure resin. Polymerization of resin was for 2 hours in a 100°C oven. Sections of 70 nm thickness were post-stained with 2% uranyl acetate in 70% methanol for 4 minutes and lead citrate (Reynolds, 1963) for 2 minutes. Images were viewed on a Tecnai 12 (FEI Inc., Hillsboro, OR, USA) transmission electron microscope operating at 120 kV, and images recorded with a Gatan Ultrascan 1000 CCD camera (Gatan Inc., Pleasanton, CA, USA). Some high magnification views of microtubule were taken out of focus in order to highlight protofilament patterns [48], [49].
EBP-2 comets were barely visible in wild type touch neurons, which could be attributed to the very stable microtubule structures in these cells. Therefore we devised an assay in which low-dose (0.125 mM) colchicine was applied to the worms to generate a moderate level of microtubule perturbation. L2 worms with twnEx40(Pmec-7::EBP-2::GFP, Pmyo-2::GFP) were grown on colchicine-containing NGM plates for 8 hours, picked off the plates and imaged one hour later. Under such treatment, a significant percentage of touch neurons displayed variable degree of microtubule growth with EBP-2 comets. Imaging acquisition was performed with the Zeiss AxioImager M2 imaging system.
Worm immunostaining was performed as described [50]. Briefly, mixed-stage animals were flash-frozen in liquid nitrogen and fixed in 2% paraformaldehyde on ice for at least 4 hours, permeabilized by Tris-Triton, and subjected to series of reduction and oxidation by sequential β-mercaptoethanol, dithiothreitol (DTT) and hydrogen peroxide treatment in 1% borate base buffer, and stained with primary antibodies in PBST-A. The following primary antibodies were used in this study: 6-11B-1 (mouse monoclonal anti-K40 acetylated α-tubulin, 1∶500, Santa Cruz Biotech), GT335 (mouse monoclonal anti-polyglutamylated tubulin, 1∶100, Enzo Life Sciences), rabbit polyclonal anti-detyrosinated tubulin (1∶200, Millipore), YL1/2 (rat monoclonal anti-tyrosinated tubulin, 1∶200, Santa Cruz Biotech), and rabbit polyclonal anti-GFP (1∶250, Santa Cruz Biotech). Secondary antibodies are goat anti-rabbit, goat anti-rat or goat anti-mouse IgG conjugated with Alexa488 or Alexa568 used at 1∶100 (Molecular Probes). Animals were counterstained with DAPI at 1∶1000 diluation in 2% n-propylgallate (NPG) and observed with the Zeiss AxioImager M2 imaging system. For fluorescence confocal microscopy, L4 to young adult hermaphrodite animals were anesthetized with 1% sodium azide, mounted on agar pad, and observed under Zeiss LSM700 confocal imaging system.
A fragment of the microtubule-binding domain (MTBD, amino acids 3207-3372) of C. elegans DHC-1 was cloned into the KpnI site of the pET30α vector with the primers: 5′ GGTACCCTCGCAGAGCAGCTGAAG 3′ (forward) and 5′ GGTACCTTATTCCTGGGTCTTCTTTGCAGC 3′ (reverse), and tagged at the N-terminus with 6xHis. Expression in E. coli was induced by 0.5 mM IPTG at 16°C for 4 hours, for avoiding protein aggregate formation in subsequent steps of purification. Bacterial pellet was collected by centrifugation and resuspended in the lysis buffer containing 50 mM NaH2PO4, 500 mM NaCl, 10 mM imidazole, 0.1% lysozyme, protease inhibitor cocktail and were homogenized by sonication. Cell extract was centrifuged at 15,000 g for 30 min at 4°C and MTBD was purified by passing the supernatant through HisPur™ Ni-NTA resin (Thermo Fisher Scientific, Walthem, USA). Purified MTBD was dialyzed in HEPES buffer (80 mM HEPES pH 7.0, 2 mM MgCl2, 0.5 EGTA) for microtubule binding protein spin-down assay.
The microtubule sedimentation assay was performed as described with modifications [51]. In brief, unc-119; twnEx98[Punc-119::MEC-12, Punc-119::MEC-7, unc-119(+)] and unc-119; twnEx99[Punc-119::MEC-12(G416E), Punc-119::MEC-7, unc-119(+)] transgenic animals were grown to gravid adults on standard NGM plates. Animals were collected by washing and centrifugation in 0.1 M PIPES (pH 6.94), 4.0 mM MgCl2, 5 mM EGTA, 0.1 mM EDTA, 0.9 M glycerol, 1 mM PMSF, and 1 mM DTT (PMEG) at 4°C and resuspension in cold PMEG with protease inhibitor cocktail. Worms were then manually homogenized and centrifuged at 20,000xg for 45 min and the pellet discarded. The supernatant was centrifuged at 150,000xg for 60 min and the pellet discarded. The translucent supernatant was then supplemented with 2 mM GTP, 10 pM taxol, 1 U/ml hexokinase, 50 mM glucose, and 25–50 nm AMP-PNP and incubated on ice for 90 min for microtubule polymerization. Microtubule polymers were sedimented through 20% sucrose cushion made in PMEG with10 pM taxol by centrifugation at 20,000xg for 90 min. The pellet was resuspended in 1 ml PMEG containing 10 nM taxol and 50 mM NaCl. Microtubules and associated proteins were pelleted at 20,000xg for 40 min and dissolved in water for further analysis.
In vitro microtubule sedimentation assay was performed by using microtubule binding protein spin-down assay kit (Cytoskeleton, Denver, USA). In brief, microtubules were synthesized from purified bovine tubulins and incubated with purified MTBD at room temperature for 30 minutes. Microtubule-associated proteins were pelleted in sucrose cushion by centrifugation at 20000xg, resuspended and analyzed by SDS-PAGE. Coomassie blue-stained signals of tubulins and MTBD were quantified using ImageJ. The signal intensity of MTBD was normalized to respective tubulin signals.
Protein lysate from microtubule sedimentation assay was boiled with SDS lysis buffer and separated by SDS-PAGE (6% acrylamide). Proteins were transferred to a nitrocellulose membrane and probed with mouse anti-UNC-104 monoclonal antibody (1∶60, a gift from Dr. S. Koushika) [52], rabbit anti-DHC-1 polyclonal antibody (1∶200, a gift from Dr. P. Gonczy) [30] or 6-11B-1 (1∶1000, Santa Cruz Biotech) for acetylated α-tubulin, followed by HRP-based chemiluminescence detection. Signal intensity was quantified using ImageJ. All experiments were repeated at least three times. To quantify the amount of DHC-1 or UNC-104 co-sedimented with microtubules, the pixel intensity of individual bands was first quantified using ImageJ. We first normalized the amount of bound DHC-1 or UNC-104 relative to sedimented microtubules in respective experiments. Next, we normalized the DHC-1/microtubule values to the averaged UNC-104/microtubule value, and data from five independent experiments were expressed as a fold change relative to the UNC-104/microtubule ratio.
|
10.1371/journal.pntd.0006883 | Integrating evidence, models and maps to enhance Chagas disease vector surveillance | Until recently, the Chagas disease vector, Triatoma infestans, was widespread in Arequipa, Perú, but as a result of a decades-long campaign in which over 70,000 houses were treated with insecticides, infestation prevalence is now greatly reduced. To monitor for T. infestans resurgence, the city is currently in a surveillance phase in which a sample of houses is selected for inspection each year. Despite extensive data from the control campaign that could be used to inform surveillance, the selection of houses to inspect is often carried out haphazardly or by convenience. Therefore, we asked, how can we enhance efforts toward preventing T. infestans resurgence by creating the opportunity for vector surveillance to be informed by data?
To this end, we developed a mobile app that provides vector infestation risk maps generated with data from the control campaign run in a predictive model. The app is intended to enhance vector surveillance activities by giving inspectors the opportunity to incorporate the infestation risk information into their surveillance activities, but it does not dictate which houses to surveil. Therefore, a critical question becomes, will inspectors use the risk information? To answer this question, we ran a pilot study in which we compared surveillance using the app to the current practice (paper maps). We hypothesized that inspectors would use the risk information provided by the app, as measured by the frequency of higher risk houses visited, and qualitative analyses of inspector movement patterns in the field. We also compared the efficiency of both mediums to identify factors that might discourage risk information use. Over the course of ten days (five with each medium), 1,081 houses were visited using the paper maps, of which 366 (34%) were inspected, while 1,038 houses were visited using the app, with 401 (39%) inspected. Five out of eight inspectors (62.5%) visited more higher risk houses when using the app (Fisher’s exact test, p < 0.001). Among all inspectors, there was an upward shift in proportional visits to higher risk houses when using the app (Mantel-Haenszel test, common odds ratio (OR) = 2.42, 95% CI 2.00–2.92), and in a second analysis using generalized linear mixed models, app use increased the odds of visiting a higher risk house 2.73-fold (95% CI 2.24–3.32), suggesting that the risk information provided by the app was used by most inspectors. Qualitative analyses of inspector movement revealed indications of risk information use in seven out of eight (87.5%) inspectors. There was no difference between the app and paper maps in the number of houses visited (paired t-test, p = 0.67) or inspected (p = 0.17), suggesting that app use did not reduce surveillance efficiency.
Without staying vigilant to remaining and re-emerging vector foci following a vector control campaign, disease transmission eventually returns and progress achieved is reversed. Our results suggest that, when provided the opportunity, most inspectors will use risk information to direct their surveillance activities, at least over the short term. The study is an initial, but key, step toward evidence-based vector surveillance.
| Chagas disease is a serious infection that is spread by blood-sucking insects called ‘kissing bugs.’ These bugs live in and around human homes, and until recently, they infested thousands of human homes throughout Arequipa, the second largest city in Perú. However, a decades-long control campaign drastically reduced the number of infested houses, and the city is now in a stage where health personnel annually inspect a sample of houses throughout the city for kissing bug reinfestation. A large amount of information was collected during the control campaign that could be used to help identify the houses at highest risk for re-infestation, so we developed a cell phone app to provide this information to health personnel in the form of interactive, user-friendly risk maps. We carried out a pilot study to see if health personnel would use these maps to select houses to inspect for re-infestation, and we found that most inspectors did use the information. We also observed that using the app did not slow the inspectors down, which can be an issue when introducing new technology. Our results suggest that the app could be a useful tool for monitoring diseases spread by insects in cities.
| Chagas disease is a neglected tropical disease (NTD) endemic to the Americas with a current estimated prevalence of six to nine million people worldwide [1,2] and 70 million more at risk [3]. An estimated 30% of those with Chagas disease will develop serious cardiac and/or gastrointestinal problems for which there is no vaccine or cure [4,5]. The etiological agent of Chagas disease, Trypanosoma cruzi, is a parasite of mammals that is transmitted between vertebrate hosts by triatomine bugs [6], and vector control is at the core of large-scale Chagas disease control efforts [7–9].
Historically, Chagas disease was considered to be a rural problem [5,10,11] associated with homes made of rudimentary materials [12–15], the presence of domestic animals in and around the domicile [16,17], and/or in close proximity to less disturbed landscapes that serve as habitat for sylvatic mammal reservoirs of T. cruzi and T. cruzi vector foci [18,19]. Disease control efforts in the past were designed accordingly, to accommodate the characteristics of rural areas. However, Chagas disease is now known to be established in several urban settings, creating a new epidemiological challenge for prevention [20–28].
In Arequipa, Perú, with a population approaching one million people, Chagas disease is an urban problem due to widespread domestic infestation by the triatomine bug species Triatoma infestans [23,28–37]. In 2002, a vector control campaign targeting T. infestans was implemented in Arequipa, and today the bug is nearly eliminated from the city. The campaign is now in its 16th year; over 70,000 households in 16 out of 18 target districts have been treated with insecticides in what was called the ‘attack’ phase of the campaign. These houses are now in the surveillance phase of the campaign, in which the highly challenging task of monitoring for vector resurgence is carried out through annual inspections of a fluctuating proportion of houses in each district. Although the 'attack' phase of the campaign generated a great amount of data relevant to the risk of subsequent vector infestation [30], these data are rarely used to inform the selection of houses to visit in the surveillance phase. Rather, the selection of houses is often carried out haphazardly or by convenience. Therefore, we asked, how can we harness the extensive data collected during the attack phase to enhance vector surveillance, and continue the considerable progress made toward the elimination of T. infestans from Arequipa?
To this end, we developed a cloud-based, open-source mobile app that provides vector infestation risk maps for use by health inspectors. The app, which we call ‘VectorPoint,’ is intended to enhance vector surveillance by giving inspectors the opportunity to incorporate infestation risk information into their process of selecting houses to inspect for T. infestans. Risk information is generated by a predictive model that calculates infestation risk estimates using data from the attack phase of the control campaign, in combination with new data collected during the surveillance phase. The app also provides a data entry function to collect new surveillance data. Upon collection, new data are sent directly to a virtual server, and then incorporated into the next run of the model, after which they are immediately visualized in the risk maps.
There are currently several apps for disease surveillance in resource limited settings, the most common being SMS-based apps (FrontlineSMS [38–40], RapidSMS [41,42], U-Report [43,44], Ushahidi [40,45], CycleTel [46,47], Geochat [48], among others [49]; see [50] for a thorough review of SMS apps for disease surveillance), and generic software and tool collections that offer mobile device-based data collection as their primary function, and some combination of basic data analysis, visualization and/or mapping as secondary functions (SAGES [51], Open data kit [52–54], Epicollect [54–56], eMOCHA [57,58], Medic mobile [59], Magpi [formerly Episurveyor, 60–62], DataWinners [63], and PhiCollect [64], among others [65–69]). A small number of apps have been developed for vector surveillance (Dengue Chat [70], CHAAK [71], eMOCHA [72]), with the primary features being data collection based in social networking and community based surveillance [70], and data collection using electronic forms and/or SMS [71,72].
VectorPoint is unique in that it supports independent decision making by the individual collecting the data, and it does not dictate a path to the user or mandate which houses to surveil. Rather, it provides the opportunity to integrate risk information into the inspector’s decision making process. Collaborative approaches that give control to the end-user have been shown to contribute to the sustainability of new technologies in resource-limited settings [73], and this is an important feature of VectorPoint. However, an inherent challenge with implementing a technology that supports independent front-end user decision-making is that the user can decide not to use it. As such, the potential for the app to enhance vector surveillance lies in the hands of the front-end user, and a critical question becomes, will inspectors use the information it provides?
To answer this question, we carried out a pilot trial comparing surveillance using the VectorPoint app to the current practice of surveillance using paper maps. We hypothesized that inspectors would use the risk information provided by the app, as measured by the frequency of higher risk houses visited and inspected with the app compared to the paper maps. We also looked for qualitative evidence of risk information use by analyzing daily and weekly maps of the inspector’s movement patterns throughout the search zones when using the app and paper maps. Finally, we compared measures of productivity between the app and the paper maps to ensure that the app was not hindering inspector progress, which might also discourage its use.
All health inspectors in the field study described below participated in the study under a written informed consent approved under University of Pennsylvania IRB protocol number 824603 and Universidad Peruana Cayetano Heredia IRB protocol number 66427.
The front end (i.e., what the user sees and interacts with) of VectorPoint is a neighborhood map that displays T. infestans infestation risk at the individual house level, and a data entry tool for collecting data resulting from home inspections (Fig 1). The back end of VectorPoint (Fig 2) is composed of a spatio-temporal Gaussian field model that generates the infestation risk estimates visualized in the maps, and a relational system of cloud-based databases and servers that are used to store and send data between the predictive model and the platform that visualizes the model-generated risk estimates in the maps. Below is a more detailed description of each component of VectorPoint.
The principal feature of VectorPoint is the risk map (Fig 1), which displays T. infestans relative risk estimates at the household level that are generated by a statistical model (detailed below). The map is intended to be used by health inspectors carrying out house-to-house T. infestans surveillance. The output from the model is presented in a simple and user-friendly format in which risk estimates are divided into five quantiles, ranging from lowest to highest infestation risk, and each quantile is then represented in the map by a color. We use the five-class, multi-hue color scheme, ‘YlOrRd,’ developed by Brewer [74] for cartography data visualization [75,76], which is color-blind friendly. The color scheme progresses from light yellow to dark red, (a color progression found to be associated with increasing hazard among Spanish speakers [77]) with saturation increasing with infestation risk, and orange representing intermediate risk. Each house is displayed in the map as a dot that is colored as one of the five colors that corresponds to the its infestation risk estimate. A legend in the corner of the map presents the colors accompanied by a one or two word description in Spanish of their corresponding infestation risk category, which translate into English as, “lowest,” “low,” “medium,” “high,” and “highest” infestation risk. Each risk category is represented in the map equally. The maps are set to display relative risk (i.e., a house’s risk of infestation relative to all other houses in the neighborhood), but they can be adjusted from the back end to display data divided into any number of quantiles, or to display absolute, instead of relative, risk estimates.
The second feature of VectorPoint is its data collection functionality (Fig 2). Inspectors can enter the data resulting from individual home visits and inspections directly into a data entry form in the app. The form is designed to collect the same data as the paper forms used by the Ministry of Health for T. infestans surveillance: date, house code (explanation below), areas of the home inspected (inside, outside, or both), number of inhabitants, number and type of domestic animals, and whether T. infestans or signs thereof (generally, eggs, feces or exuviae, grouped together as ‘rastros,’ meaning ‘traces’) were found. Radio buttons and drop-down menus are provided whenever possible for consistency, and to avoid typographical errors. After the data entry form is completed, data are encrypted and transmitted from the app to a SQL database, eliminating the step of digitizing data from paper forms.
It should be noted that T. infestans surveillance data in Arequipa are organized with four tuple identification (ID) codes assigned to each home by the Peruvian Ministry of Health at the beginning of the vector-control campaign. The four tuple consists of: province/district/locality/house. (VectorPoint is designed to be used for house to house surveillance at the locality level, which are neighborhoods ranging from 30–2000 households.) We have maintained the four tuple ID system in VectorPoint, and throughout the manuscript, we refer to the four tuple house IDs as ‘house codes.’
The model in VectorPoint is designed to estimate the relative probability of T. infestans infestation of sites (primarily households) in an urban landscape. The model incorporates three types of information: (i) site covariates; (ii) the results of any previous inspections for T. infestans; and (iii) infestation history in neighboring sites. For each site, we include one covariate that is an indicator of participation in the attack phase of the vector control campaign, during which insecticide was applied to all participating households, as previous studies have shown that houses that did not participate are more likely to be infested [30]. We did not include other finer-scale risk factors for T. infestans infestation, such as guinea pig husbandry [23], because data were not available at the scale required for app.
Concretely, let the probability of vector presence, i, at time, t, be given by πi. We model the probability using a logistic model with intercept, β0, covariate information ,β1, and separable spatio-temporal random effects, uitandvit:
logit(πi)=β0+β1+uit+vit
where uit is a realization of the Gaussian field with a Matérn covariance structure [78,79]. The Gaussian field functions such that any adjustment to the estimate for one house affects all other houses in a given area, with a greater effect on those nearby. The Matérn function is a versatile model of covariance that includes Gaussian covariance as a special case [80]. The term vit is a first order autoregressive discrete time random effect.
As mentioned earlier, the model takes into account the inspection history of each household/site. We currently include four discrete time periods (Fig 3). We selected our time periods in reference to the phase of the vector control campaign in each area. The earliest point reflects the 'attack' phase of the campaign, which occurred between January 11th, 1997 and January 6th, 2014, depending on the district. The second time period is the early surveillance period, and it includes all inspection data collected between January 7th, 2014 and January 6th, 2016. The third time period includes inspection data collected between January 7th, 2016 and January 4th, 2018. The final time point reflects the current calendar year, currently set to (at the time of this publication) January 5th, 2018—present. The later time periods can be adjusted if needed. The predicted probability of infestation for the most recent time point is visualized in the app.
We fit the model with integrated nested Laplace approximations (INLA) using the R-package, “INLA” [81,82]. To account for effects of streets as semi-permeable barriers to the spread of T. infestans [29], we used an extension of a Gaussian Field model in which we stretched the city map so the geographic center (i.e., the statistical mean of the coordinates) of each city block is at a multiple (1.5;[83]) of the true distance. We maintain the within-block structure, so only the distance between blocks is stretched [83]. We set strong priors (mean = 1.17 and standard deviation = 0.01) on the covariate of not participating in the original insecticide application campaign, based on our previous analysis of this factor [30]. We set the prior on the intercept term to correspond to an expected baseline infestation prevalence of approximately 1 in 1000, with the precision matrix set to 50. This value reflects our best estimate of T. infestans infestation prevalence in Arequipa based on recent results from both passive surveillance (i.e., reports of T. infestans infestation from community members that are later confirmed by health personnel) and active surveillance (house to house surveillance conducted by our team and the Ministry of Health).
Infestation risk estimates generated by the model are sent to a cloud-based database (Amazon Relational Database Service from Amazon Web Services) through the RMySQL package [84]. These data are then sent to the Shiny [85] server, which graphically renders the risk estimates in the app. Inversely, new data collected with the app are sent back to the SQL database, and incorporated into the next run of the model. We present a diagram of the VectorPoint workflow in Fig 4.
To address connectivity issues, VectorPoint has a caching function that stores partial inspection results, and retrieves them when connectivity is restored. VectorPoint supports multiple model result tables, and all operations are computationally parallel, allowing the app to be used by multiple inspectors simultaneously without speed degradation.
We constructed VectorPoint using open source software throughout to enable sharing and extensions. We built it using the Shiny package for R [85], and we implemented it entirely in the open source R programming language [82]. We mapped vector infestation predictions using the Leaflet package for R [86], and we overlaid these data on top of street data from OpenStreetMap.org. Open source code and related tools for VectorPoint can be downloaded from https://github.com/chirimacha/VectorPoint, including a link to a fully-functional installation of the app, which is available in the ‘README’ section.
VectorPoint is located on a web server that can be accessed using a web browser on any desktop or mobile device, regardless of platform (Android, iOS, Windows, OSX, Linux, etc). Upon loading the VectorPoint web page, the inspector is presented with an authentication form in which they enter a username and password. All connections are encrypted, and risk maps can be accessed only by the study team and authorized health personnel. After user authentication, the inspector selects the locality or group of localities where they will carry out surveillance that day. The app retrieves the data for the locality from the database, and loads the corresponding risk map, zoomed out (Fig 1A). This view provides the inspector with a high-level view of the houses and their relative risk levels of infestation.
From there, the inspector zooms in on the map (Fig 1B), and selects a house to potentially visit by clicking on the corresponding dot on the map. A dialogue box will open up containing the house code, the date that the house was last visited, and whether or not the house was inspected at that time. If the inspector decides to visit that house, they can load a data entry form with the house code and date auto-filled in. If the inspector receives permission to inspect the house for T. infestans, data from the inspection is entered into the data entry form. If the inspector does not receive permission to inspect the house, it is recorded in the data entry form as one of four alternative outcomes: “interview,” “closed,” “refused,” or “return.” ‘Interview’ means that the inspector spoke with someone at the door about T. infestans infestation, but did not receive permission to enter the house and inspect it; ‘closed’ means that no one answered the door; ‘refused’ means that inspection was directly refused; and ‘return’ means that the inspector was asked to return at a later time. After each house visit, data are sent from the app to the database, regardless of visit outcome. In cases of data outages or other internet connectivity issues, inspection data can be saved and sent to the database at a later time, as described above. This process is repeated for each home visited by the inspector in a given day. At the end of the day, all data collected with the app are pulled from the database and run in the model to generate new predictions. The predictions are then pushed back to the database and visualized in the map.
In the study comparing T. infestans surveillance with the app to surveillance under the current practice of using hand-drawn paper maps (Fig 5), eight members of our field team previously trained to carry out home inspections for T. infestans carried out vector surveillance in Arequipa for a total of two work weeks (10 days). At the beginning of each week, inspectors were randomly assigned (i) a zone to surveil, and (ii) if they would use the app or the paper map in that zone. Only one inspector was assigned to each search zone, which were all in the same city district. Each search zone met the following five criteria: (i) it was located in a developed area (i.e., all roads paved) within the central portion of the district, (which is safer than peripheral, less developed areas); (ii) it contained a minimum of 400 houses, and no more than 1.25 times the number of houses in the zone with the fewest houses; (iii) its area was a minimum of 0.1 km2, and could not be greater than twice that of the zone with the smallest area; (iv) house density was at least 2000 houses per km2; and (v) the search zone was in a locality where at least one house had been found positive for T. infestans during the attack phase of the control campaign. These criteria resulted in 16 search zones with 416–514 homes, areas of 0.12–0.20 km2, perimeters ranging between 1.62–2.37 km, and house densities ranging from 2,570 to 3,623 houses per km2. Half of the inspectors used the app in the first week and paper maps in the second week, and the other half used paper maps in the first week and the app in the second week. All inspectors used the same cell phone model and operating system when using the app (Samsung Galaxy J7, with Android version 7.0), to control for variation between devices. Inspectors received training in app use with these phones before starting the experiment.
Surveillance was carried out daily during normal working days (Monday-Friday) and hours (7am-1pm). Inspectors were told to carry out T. infestans surveillance as they would normally, and that their objective was simply to find T. infestans, in order to avoid perceptions that they needed to fulfill a quota of visiting a certain number of houses. The protocol for T. Infestans inspections is stipulated by the Peruvian Ministry of Health, and consists of systematic searches in all areas of the domicile and peri-domicile (pending permission by the resident), including spaces occupied by humans and live animals. Searches last approximately one person-hour; flexibility is allowed to account for the heterogeneity in the size of houses. During the search, the inspector looks for live T. infestans, in addition to T. infestans eggs, exuviae, and feces.
Inspectors using the paper maps did sometimes have access to two pieces of infestation risk information, as each house code (which is painted on the outside of the house) contains indicators of participation in the attack phase of the vector control campaign, and a '+' sign at the end of the code if the house was ever found to be infested with T. infestans. This information is available only when standing directly in front of a house looking at the house code; it is not shown in the paper maps, meaning that inspectors using the paper maps could not see the spatial distribution of houses with these risk factors. In addition, many of the house codes have been painted over by homeowners in the years since the attack phase of the campaign, so all houses with these risk indicators are not identifiable.
To measure the effect of using the app in the field on inspector productivity we compared the total houses visited and total houses inspected between the app and paper maps. We also compared the proportion of total houses visited that ended up being inspected between the two mediums. We selected these metrics to test if the app was slowing the inspectors down or constricting their surveillance activities in some way, due to technical difficulties or otherwise. As mentioned above, we did not compare the number of infested houses found, since the prevalence of infestation is currently very low [29].
To investigate if inspectors used the risk information provided in the app to select houses to visit, we compared the proportion of houses visited that were higher risk houses (top two risk levels) when using the app and when using paper maps. To further investigate risk information use, we compared the proportion of total houses visited that resulted in inspection between the app and paper maps among just the higher risk houses visited (houses presented in the app as ‘highest’ and ‘high risk’) and just the lower risk houses visited (houses presented in the app as ‘medium,’ ‘low,’ and ‘lowest’ risk). We were interested in how the possession of information about a house’s estimated risk level might influence the visit outcome (i.e., if a house was inspected, closed, inspection refused, an interview took place, or the inspector was asked to return at a later time).
Finally, we looked for qualitative evidence of risk information use by comparing maps of inspector movement patterns throughout the search zones when using the app and paper maps. The daily maps for each inspector are found in the VectorPoint repository (https://github.com/chirimacha/VectorPoint). We examined movement patterns on a smaller scale, such as changes in direction and the tendency to visit every neighboring house versus skipping houses. In addition we looked at patterns on a larger scale, such as the tendency to visit all houses in one area of a zone versus visiting a few houses across several areas, as well as cumulative movement throughout the total search zone across all five days, which we refer to as ‘spatial coverage’ of the zone.
To compare the total number of houses visited and houses inspected between the app and paper maps we used a paired t-test. For analyses involving house risk level, we split the houses visited into binary categories of higher and lower risk, consisting of the top two risk levels and the bottom three risk levels, respectively. For analyses involving visit outcome, we classified visit outcomes into the binary categories, ‘inspection,’ and ‘other.’ For all metrics except total houses visited and total house inspected, we carried out a preliminary analysis using Fisher's exact test to test for differences between the app and paper maps for each individual inspector, and a Mantel-Haenszel chi-square test with continuity correction to test for an overall shift in one direction among all inspectors. We carried out a second analysis using binomial Generalized Linear Mixed Models with "Inspector ID" random intercepts to test whether the inclusion of an "app" fixed effect helped to explain any variance. In this analysis, we compared the results from a model run with only inspector ID random intercepts with results from a model run with inspector ID random intercepts plus an app fixed effect. We used the BIC scores [87] as a metric to assess if the addition of the app fixed effect improved the performance of the models, with smaller BIC scores suggesting model improvement. In addition, we evaluated the odds ratios in the app fixed effects models to understand if the app increased the odds of the outcome occurring. Models were fit by maximum likelihood (Laplace Approximation) using the ‘glmr’ function in the lme4 package [88] for R. All data analyses were carried out in the R statistical computing environment [82].
Over the course of ten days, eight inspectors visited a total of 2,119 houses, of which 767 were inspected for T. infestans (Table 1). In the five days using the paper maps, 1,081 houses were visited, resulting in 366 inspections (33.9%). In five days with the app, 1,038 houses were visited, resulting in 401 inspections (38.6%).
There was no difference between the app and paper maps in the total number of houses visited (paired t-test, p = 0.67, Table 1), the total number of houses inspected (paired t-test, p = 0.17, Table 1), or the proportion of total visits that resulted in inspection (Mantel-Haenszel test, chi square = 2.63, p = 0.105, odds ratio (OR) = 1.18, 95% CI = 0.97–1.42; Fig 6), further suggesting that the app did not reduce productivity in the field. When looking at inspectors individually, no one inspected proportionally more houses with paper maps. Two inspectors, C and H, inspected a significantly higher proportion of houses when using the app as based on test p-values (Fisher’s exact test, Inspector C: p = 0.046, OR = 1.91, 95% CI = 0.99–3.76; Inspector H: p = 0.026, OR = 1.80, 95% CI = 1.05–3.08, Fig 6).
In the generalized linear mixed models (GLMMs), the odds ratio of the app fixed effect was 1.18 (95% CI = 0.97–1.42, Table 2), suggesting little effect of the app on the odds of a visited house being inspected. The standard deviation in inspector random effects was high regardless of whether or not the model included an app effect (SD = 0.76, 95% CI = 0.49–1.39 with the app; SD = 0.77, 95% CI = 0.49–1.40 without the app). The addition of the app fixed effect did not improve the model, with similar BIC scores for both models (138.71 and 138.80 for the model with and without the app, respectively). Results from the GLMMs are presented in Table 2.
Five out of eight inspectors (B,D,E,F, and H; 62.5%, Fig 7), visited proportionally more higher risk houses (top two risk levels) when using the app than when using paper maps (Fisher’s exact test, Inspector B: p = 4.14e-14, OR = 25.38, 95% CI = 7.80–131.06; Inspector D: p = 1.81e-14, OR = 40.71, 95% CI = 9.94–360.00; Inspector E: p = 4.82e-07, OR = 4.42, 95% CI = 2.37–8.51; Inspector F: p < 2.2e-16, OR = 10.10, 95% CI = 5.58–18.80; Inspector H: p < 2.2e-16, OR = 14.86, 95% CI = 6.80–37.13). Two inspectors (25%), A and C, showed no difference in the proportion of total visits that were to higher risk houses when using the app (Fisher’s exact test, Inspector A: p = 0.65, OR = 1.19, 95% CI = 0.63–2.25; Inspector C: p = 0.12, OR = 1.43, 95% CI = 0.90–2.26; Fig 7), and and one inspector, G, visited proportionally more higher risk houses when using the paper map (Fisher’s exact test, p < 2.2e-16, OR = 0.0, 95% CI = 0.0–0.02; Fig 7). Overall, there was a significant shift upward in the risk level of the houses visited from the paper maps to the app (Mantel-Haenszel test, chi-square = 104.44, p < 2.2e-16, common OR = 2.42, 95% CI = 2.00–2.92; Fig 7), suggesting that inspectors did incorporate the risk information provided in the app into their selection of houses to visit.
In the GLMM, the app effect was a significant predictor of a higher risk house being visited (p < 2e-16), and it increased the odds of visiting a higher risk house by 2.73 (95% CI = 2.24–3.32, Table 2). Standard deviation in inspector ID random effects was again high in both models (0.76, 95% CI = 0.48–1.37 with the app, and 0.77, 95% CI = 0.50–1.20 without the app). BIC scores suggested that the app improved the model, with scores of 442.9 and 544.3 for the models with and without the app, respectively.
We did not observe any patterns associated with arm order (i.e, which medium was used in the first week and which was used in the second week, Fig 7). Out of the five inspectors who visited proportionally more higher risk houses using the app, two used the app first and three used the paper maps first. For the two inspectors who showed no difference between the app and paper maps, one inspector started with the app, and the other started with the paper map. The inspector who visited more higher risk houses with paper maps used the app first.
We also found no difference between the app and paper maps in the proportion of visits to higher risk houses or lower risk houses that resulted in inspection, suggesting that there is no association with knowledge of a house’s infestation risk and obtaining permission to inspect it (Fisher’s exact test, p > 0.05 for all inspectors for higher and lower risk houses; overall: Mantel-Haenszel test, higher risk houses: chi square = 1.71, p = 0.19, common OR = 1.19, 95% CI = 0.93–1.51; lower risk houses: chi square = 0.44, p = 0.51, common OR = 1.18, 95% CI = 0.78–1.78).
We observed two predominant movement patterns throughout the study. The first movement pattern was more individual house focused, while the second pattern was more focused on coverage of the total search zone. (As mentioned above, daily maps for each inspector are available in the VectorPoint respository, https://github.com/chirimacha/VectorPoint.) While most inspectors consistently exhibited one pattern or the other, some inspectors displayed a mixture of both throughout the week. In the individual-house-focused movement pattern, inspectors tended to travel short distances (1–3 blocks/day), and visit every house on a block. Movement patterns appeared systematic and focused on covering all households in a small part of the zone, resulting in low spatial coverage of the total search zone. In the pattern that was more focused on search zone coverage, inspectors tended to travel longer distances, visiting a cluster of houses in one area, and then moving on to another part of the zone. Movement appeared less systematic than the individual-house-focused pattern, with a more holistic focus on the entire search zone, resulting in higher spatial coverage. There are many practical constraints in the field that will influence inspector movements, but larger movements across the zone were much more common in inspectors that tended toward the broader coverage of the search zone.
Seven out of eight inspectors (A, B,C,D,E,F and H) displayed at least one indication of risk information use in the context of the predominant movement patterns. In inspectors with individual house-focused movement patterns, risk information use was indicated when they skipped lower risk houses in between higher risk houses, or made abrupt changes in direction to avoid a cluster of lower risk houses, which sometimes resulted in increased spatial coverage of the search zone. While using the paper maps, these inspectors rarely skipped houses or changed direction, and their spatial coverage of the zone was low. Inspectors who were more focused on coverage of the total zone tended to indicate risk information use with movement that was more systematic and directed toward higher risk houses when using the app. When using the app in areas with clusters of higher risk houses, some of these inspectors adjusted their usual broader coverage movement pattern to be more individual house focused, resulting in visits to every house on a higher risk block.
We observed indications of risk information use in inspectors that did visit significantly more higher risk houses with the app (B,D,E,F,H) and in those that did not (A and C). Inspector C did not visit significantly more higher risk houses, but did skip clusters of lower risk houses on some days, suggesting they were ‘trying out’ the risk information. Inspector A greatly increased spatial coverage when using the app as compared to the paper map, which was due to a movement pattern that appeared to ‘sample’ a new risk level each day. Over the week, the inspector progressed in order of risk level from visiting lowest risk houses on day one to visiting highest risk on day five. This inspector did not visit significantly more higher risk houses using the app, but did appear to be paying attention to the risk information in the app. Indeed, by the end of the week, the inspector was skipping houses that were low and lowest risk.
Only one inspector, G, showed no qualitative or quantitative indications of using the risk information in the app to select houses to visit. In fact, this inspector unexpectedly visited significantly more higher risk houses when using the paper map, which we attribute to clustering of houses of the same risk level in the search zone. Based on examination of this inspector’s movement patterns when using the paper map, it seems that they visited more higher risk houses with the paper map because the inspector displayed a strong tendency toward fine-scale (individual house-focused) movement. They selected their day one starting point with the paper map based on house code (lowest codes first) and proceeded through the zone in numerical order. Coincidentally, the starting house was located in the beginning of a large cluster of higher risk houses, resulting in only higher risk houses being visited with the paper map.
Here we present a mobile app designed to provide the opportunity to incorporate data-based risk information into field surveillance for the Chagas disease vector species Triatoma infestans in the city of Arequipa, Perú. In our study comparing surveillance using the app to the current practice of surveillance using paper maps, we observed multiple indications that the risk information provided in the app was used in the selection of houses to inspect, suggesting that the app is a feasible tool to enhance vector surveillance and support decision making in the field.
mHealth tools, defined as mobile and wireless technologies for health-related objectives [89] are being introduced at a rapid-fire pace [90]. While the majority of mHealth apps are for personal uses related to aspects of individual health, the use of apps for disease surveillance in resource-limited settings has been growing steadily. The large number of data collection apps available for disease surveillance has provided the opportunity to replace often slow and cumbersome paper-based data collection systems with mobile data collection systems in which data are sent directly to a central database at the time of collection. Ideally, the shift from paper to technology will lead to increased data completeness and coverage in resource-limited settings, a critical step for achieving disease surveillance goals [91]. (Of course, new technologies bring new challenges, some of which are detailed below.) The next step is now integrating data collection functionalities with more complex tools to support ongoing decision making in the field in real or close to real time. While several data collection apps offer basic data visualization capabilities such as simple maps or bar graphs showing raw data distributions, these functionalities are often not immediately available to the individuals collecting data in the field, and the usefulness of raw data for on the ground decision making can be limited. The VectorPoint app is among the first in its marriage of data collection with predictive modeling and spatial data visualization, all of which are intended to support the individual collecting the data.
As mentioned previously, for VectorPoint to meet the objective of integrating data into vector surveillance activities, inspectors with years of experience and well-established routines must adapt their decision-making processes to include at least some of the risk information in the app. Model-generated estimates are not necessarily more valuable than knowledge derived from experience, and VectorPoint is meant to complement an inspector’s experience, not replace it. In the study, we observed multiple indications that risk information was used by the inspectors. Namely, we observed a significant upward shift among all inspectors in the number of high risk houses visited when using the app compared to paper maps, and a majority of the inspectors visited significantly more higher risk houses when using the app. These results were then confirmed in a second analysis comparing generalized linear mixed models run with and without the app, in which the app was a highly significant predictor of a higher risk house being visited. In the maps of inspector movement throughout the search zones, we also observed changes in predominant movement patterns when using the app that indicated risk information use. Even inspectors who did not visit significantly more higher risk houses with the app appeared to at least ‘try out’ targeting higher risk houses or explore the different risk levels, suggesting that over time inspectors may gradually introduce the risk information into their surveillance activities.
While our results are encouraging, one inspector did not show any signs of using the risk information, and we cannot assume adoption would be universal. Indeed, our more rigorous analyses using generalized linear mixed models confirmed that random among-inspector variation is high regardless of the outcome tested. Willingness to use the information in the app may be associated with factors specific to the individual, such as age or experience with technology [89], and we may need to take into account inspector demographic information when analyzing app use in future studies in order to develop strategies for increasing risk information use. In addition, motivation to use the app may increase as more data are collected, allowing us to compare infestation indices of houses inspected while using the app versus paper maps. As mentioned above, this study was not scaled to formally compare infestation indices; only one infested house was found in the study, and the inspector was using a paper map at the time. Specialized training or more creative solutions such as incentives [92–94] or games [95,96] may be necessary in cases where inspectors have trouble engaging with the app, or if enthusiasm for the new technology wanes over time.
Although the tendency to travel or not tended to be a fixed characteristic between both arms of the study, in a few cases we observed qualitative changes in spatial coverage of the search zone when using the app. While the app is not currently designed to directly address the issue of spatial coverage, the model is expected to redistribute the relative risk estimates throughout the search zone every time it is run with new inspection data. In other words, if a house is surveyed and not found to be infested with T. infestans, infestation risk should decrease for all households within a certain distance of the uninfested house. In this way, inspectors who use the risk information in the app to target high risk houses may also display even spatial coverage of their search zone. In the study, we observed this effect only occasionally, and in some cases, inspectors targeting high risk houses displayed reduced spatial coverage of their search zone when higher risk houses were clustered. This outcome is likely due to the relatively small amount of data collected by each inspector, and the varying amounts of information available in each search area prior to the study. As detailed above, we expect risk level clustering to be dynamic, and, in the absence of a positive house, risk distribution should even out as more data are collected. The utility of the model behind the app, like all Bayesian models, increases as more data are accumulated.
Extensions to the app, or the model itself, may be useful in improving spatial coverage of surveillance, and for exploring more complex questions revolving around the benefits or drawbacks of using the app. In particular, we do not know if a trade off occurs between exploration and exploitation when using the app; by focusing on areas that are known or predicted to have a higher risk of infestation (i.e., exploitation), will we fail to detect new high risk foci (i.e., exploration)?
The use of mHealth tools in resource-limited settings presents inherent challenges [50,55,73] due to the financial and technical requirements of cell-phone based systems. An advantage of VectorPoint is that it uses entirely free open-source technologies, which allows large-scale deployment at little cost. The remaining operational costs involve the acquisition of mobile devices and their corresponding cellular network subscriptions. For cloud-based and web-based apps such as VectorPoint, these networks must provide data coverage of at least 2G speeds. In our field tests of VectorPoint, we found that frequent interruption and reloading of the maps were avoided when data speeds were 3G or higher. Fortunately, Peru has a robust mobile phone culture, with 117 mobile phone subscriptions per 100 people in 2016 [97], higher than the world average of 101.5. There are several networks available throughout Peru, all offering high-speed data plans. In cases where network problems do occur, VectorPoint has a caching functionality, as mentioned above, that allows data entry, although some internet connectivity is still needed for efficient use of the app. We are working towards a completely offline-capable version of VectorPoint to overcome this limitation. In our study, there was no difference in the number of houses visited between paper maps and the app, suggesting that any connectivity issues encountered did not significantly slow surveillance activities.
Being a cloud-based computing app means that VectorPoint can be run on any device without any configuration or installation, but it does require human expertise to oversee and problem-solve its backend components. Fortunately, there is a growing culture and acceptance of electronic health tools in Peru, which has been found to contribute to the sustainability of health information systems in resource-limited settings [73]. Gozzer Infante [98] reviewed 38 electronic health systems that were introduced in Peru between 2002–2010, and 66% of them were still being used in 2015. These systems include Alerta, an large multi-disease top down surveillance system [99]; Netlab, a laboratory support system for HIV treatment [100]; and Magpi, a data collection app being used by researchers to study HPV prevalence [61]. As of 2018, several mHealth apps were confirmed to be in use at the national level, (Leonardo Rojas, Peruvian Institute of Health, personal communication), such as Guardianes de la Salud [101], a mobile phone based disease surveillance system for use during the early 2018 Papal visit. In the case of VectorPoint, we collaborated with local personnel in the development of the app from the start, and they currently control both its operational use and its engineering. However, it should be mentioned that although the VectorPoint app is flexible, the predictive model must still be adapted to fit the local epidemiological and ecological characteristics of each disease surveillance scenario to which it is applied, which could present challenges in cases where expertise is limited.
Our findings suggest that the VectorPoint app could be a useful tool to integrate evidence and models into epidemiological surveillance in cities. The app was designed to be used for T. infestans surveillance, but its components are cloud-based, open-source, and ready to be adapted to other appropriate scenarios, although the availability of sufficient data and/or resources may be a hurdle to overcome in some cases. VectorPoint is simple to use, but critical in function: without staying vigilant to remaining and re-emerging vector foci following a vector control campaign, disease transmission inevitably returns and progress achieved is reversed [102].
|
10.1371/journal.pcbi.1004187 | A Bayesian Ensemble Approach for Epidemiological Projections | Mathematical models are powerful tools for epidemiology and can be used to compare control actions. However, different models and model parameterizations may provide different prediction of outcomes. In other fields of research, ensemble modeling has been used to combine multiple projections. We explore the possibility of applying such methods to epidemiology by adapting Bayesian techniques developed for climate forecasting. We exemplify the implementation with single model ensembles based on different parameterizations of the Warwick model run for the 2001 United Kingdom foot and mouth disease outbreak and compare the efficacy of different control actions. This allows us to investigate the effect that discrepancy among projections based on different modeling assumptions has on the ensemble prediction. A sensitivity analysis showed that the choice of prior can have a pronounced effect on the posterior estimates of quantities of interest, in particular for ensembles with large discrepancy among projections. However, by using a hierarchical extension of the method we show that prior sensitivity can be circumvented. We further extend the method to include a priori beliefs about different modeling assumptions and demonstrate that the effect of this can have different consequences depending on the discrepancy among projections. We propose that the method is a promising analytical tool for ensemble modeling of disease outbreaks.
| Policy decisions in response to emergent disease outbreaks use simulation models to inform the efficiency of different control actions. However, different projections may be made, depending on the choice of models and parameterizations. Ensemble modeling offers the ability to combine multiple projections and has been used successfully within other fields of research. A central issue in ensemble modeling is how to weight the projections when they are combined. For this purpose, we here adapt and extend a weighting method used in climate forecasting such that it can be used for epidemiological considerations. We investigate how the method performs by applying it to ensembles of projections for the UK foot and mouth disease outbreak in UK, 2001. We conclude that the method is a promising analytical tool for ensemble modeling of disease outbreaks.
| Epidemiological forecasting is inherently challenging because the outcome often depends on largely unpredictable characteristics of hosts and pathogens as well as contact structure and pathways that mediate transmission [1]. Faced with such uncertainty, policy makers must still make decisions with high stakes, both in terms of health and economics. For instance, global annual malaria mortality was recently estimated at around 1.1 million [2] and to optimize control efforts, policy makers must make seasonal predictions about spatiotemporal patterns [3]. The prospect of an emergent pandemic influenza outbreak remains a global threat and emergency preparedness must evaluate the costs and benefits of control measures such as border control, closing of workplaces and/or schools as well as different vaccination strategies [4]. Livestock diseases are major concerns for both animal welfare and economics. As an example, the United Kingdom (UK) 2001 outbreak of foot and mouth disease (FMD) involved culling of approximately 7 million animals, either in an effort to control the disease or for welfare reasons, and the total cost has been estimated at £8 billion [5]. To minimize the size and duration of future outbreaks, various strategies for culling and vaccination must be compared [6–8]. As a tool to address these challenging tasks, mathematical models offer the possibility to explore different scenarios, thereby informing emergency preparedness and response to epidemics [1,9–12].
The predictive focus of epidemiological models can either be classified as forecasting or projecting [13]. Forecasting aims at estimating what will happen and can be used for example to predict seasonal peaks of outbreaks [3,14] or to identify geographical areas of particular concern [15]. Projecting, which is the main focus of this study, instead aims at comparing different scenarios and exploring what would happen under various assumptions of transmission, e.g. comparing the effectiveness of different control actions [7,16–19].
Whilst analytical models clearly provide important insight into observed dynamics and a theoretical understanding of epidemiology [20–22], there has been a shift in recent years towards stochastic simulation models for predictive purposes [1]. Typically, dynamic models are constructed and outbreaks are simulated repeatedly, thus generating predictive distributions of outcomes [1,17,18,23]. This variability in outcomes caused by the mere stochasticity of the transmission process includes one level of uncertainty, but still only relies on a single set of assumptions about the underlying disease transmission process. However, multiple assumptions can often be justified, leading to further uncertainty in the predictions. For instance, different models may have different projections because of different assumptions about transmission or because they incorporate different levels of detail. It may also be informative to explore different projections in terms of different parameterizations of a single model, for example corresponding to worst or best case scenarios. Faced with a set of projections, an important issue is how to combine these in a manner such that they can be used to inform policy.
The issue of multiple projections is not unique to the field of epidemiology, and various techniques of ensemble modeling have been used to merge projections based on different modeling assumptions. The key concept is that rather than relying on a single set of assumptions, a range of projections is used for predictive purposes. For instance, climate forecasting has employed ensemble techniques to account for uncertainty about initial conditions, parameter values and structure of the model design when predicting climate change [24,25]. Weather forecasting has been improved by combining the results of multiple models [26,27]. Similarly, hydrological model ensembles have been demonstrated to increase reliability of catchment forecasting [28] and have been used to assess the risk of flooding events [29]. Ensemble methods have also proven to be a powerful decision tool for medical diagnostics [30,31] and ecological considerations including management [32] and prediction of future species distribution [33].
Ensemble modeling has not yet been extensively used in epidemiology. However a few implementations exist, commonly by feeding climate or weather ensembles into disease models. Daszak et al. [34] coupled a set of climate projections to an environmental niche model of Nipah virus to predict future range distribution of the virus under climate change. Similarly, Guis et al. [35] investigated the effect of climate change on Bluetongue emergence in Europe by simulating outbreaks under different climate change scenarios. Focusing on a shorter time scale, Thomson et al. [3] used an ensemble of seasonal forecasts to predict the spatiotemporal pattern of within seasonal variation in malaria incidence. These studies all used a single disease model projection, coupled to an ensemble of climate or weather forecasts and the use of structurally different epidemiological models are to our knowledge still rare. However, Smith et al. [36] compared different malaria vaccination strategies by implementing an ensemble approach with different alterations of a base model. Also, in order to estimate global malaria mortality, Murray et al. [2] used weighted averages of different predictive models.
Given the success of ensemble methods in other fields, we expect that epidemiological implementations will increase. For that purpose however, there is a need for methods that combine multiple projections. A central issue in ensemble modeling is how to weight different projections, and we envisage four main procedures for this. Firstly, all models can be given equal weights. For instance, the IPCC 2001 report on climate change [37] used a set of climate models and gave the range of probable scenarios by averaging over different models and uncertainty by envelopes that included all scenarios. Gårdmark et al. [32] used seven ecological models for cod stock and argued that in order to prevent underestimation of uncertainty, weighted model averages are not to be used and when communicating with policy makers, it is preferable to present all included projections as well as the underlying assumptions behind them. A similar approach was also used by Smith et al. [36], who presented the prevalence of malaria under different vaccination strategies by medians of individual models and the range of the whole ensemble.
Secondly, expert opinions can be used to weight models. To our knowledge, no ensemble study has implemented weights based exclusively on expert opinion, but Bayesian model averaging can incorporate expert opinion as a subjective prior on model probabilities [38]. This approach relies on engaging stakeholders and communicating the underlying assumptions of the projections.
Thirdly, models can be weighted by agreement with other models. This approach was implemented by Räisänen and Palmer [39], who used cross-validation to weight climate models. As a more informal approach to the use of model consensus, the third IPCC report excluded two models because these predicted much higher global warming than the rest of the ensemble, thus acting as outliers [24].
Fourthly, weights can be determined by the models’ ability to replicate data. If all models are fitted to the same data using likelihood based methods, weights can be given directly by Akaike or Bayesian Information Criterion (AIC or BIC) [40,41]. In the FMD context, this may be a suitable approach if all included models are data driven kernel models that estimate parameters from outbreak data, such as those proposed by Jewell et al. [42] or Tildesley et al. [43]. However, such weighting schemes would be unfeasible when including detailed simulation models that rely on a large number of parameters, that are determined by expert opinion or lab experiment, such as AusSpread [44], NAADSM [45] and InterSpread Plus [46]. We propose that the future of ensemble modeling for epidemiology will benefit from combining structurally different model types, and methods of weighting need to handle both kernel type models as well as detailed simulation models.
Thus, bias assessment is often confined to the ability of models to replicate observed summary statistics of interest, in particular when the resolution of data observation is on a courser scale than the model prediction [47]. Such methods have been developed within the field of climate forecasting. Giorgi and Mearns [48] introduced a formal framework in which model weights were assessed based on model bias compared to observed data as well as convergence, i.e. agreement with the model consensus. Tebaldi et al. [47] extended the approach to a Bayesian framework. This approach is appealing because it provides probability distributions of quantities of interest, hence uncertainty about the projected outcomes may be provided to policy makers. As such, it would be a suitable approach also for epidemiological predictions.
However, methods developed in one field might not be directly transferable to another. Tebaldi et al. [47] points out that lack of data at fine scale resolution is a limiting factor for climate forecasting. Yet, at courser resolution climate researchers have access to long time series of climate variables to assess model bias. Comparable data may be available for endemic diseases, such as malaria [36] or tuberculosis [49], or seasonally recurrent outbreaks, such as influenza [14] or measles [50]. However, for emerging diseases, long time series would rarely be available, making the lack of data an even bigger issue for epidemiology.
In this methodological paper we aim to explore the potential of using ensemble methods based on the approach presented by Tebaldi et al. [47] for epidemiological projections. The Tebaldi et al. methodology focus on ensembles where projections are made with different models and our aim is to provide a corresponding framework for disease outbreak projections. To investigate the potential of the framework for epidemiology, we here use variations of a single model as a proxy for different models, thus allowing us to investigate how the methodology performs under different levels of discrepancy among projections in the ensemble. We exemplify the implementation by using the UK 2001 FMD outbreak and projections modelled by different parameterizations of the Warwick model [7,9].
In the 2001 UK FMD outbreak, livestock on all infected premises (IPs) were culled. In addition, livestock on farms that were identified to be at high risk of infection were culled as either traditional dangerous contacts (DCs) or contiguous premises (CPs). CPs were defined as “a category of dangerous contact where animals may have been exposed to infection on neighboring infected premises” [5,8]. We start by focusing on ensemble prediction of epidemic duration under the control action employed during the 2001 outbreak compared with an alternative action that excludes CP culling. We investigate sensitivity to priors and explore a hierarchical Bayesian extension of the method to circumvent potential problems with prior sensitivity. We also explore the potential of including subjective a priori trust in the different modeling assumptions and extend the methodology further to allow incorporation of multiple epidemic quantities, here exemplified by adding number of infected and culled farms to the analysis. Through a simulation study, we finally explore the capacity and limitations of the proposed ensemble method, pinpointing some important features of ensemble modeling
We apply a terminology such that control actions refers to different strategies for disease control. In the ensemble, each action is simulated under different modeling assumptions about the underlying process, expressed as either different models or, as in the example described here, different parameterizations of the same model. We refer to the combination of control action and modeling assumption as a projection. Each projection is further simulated with several replicates, which produce different outcomes merely due to the stochasticity of simulation models. We are also interested in how discrepancy among projections influences the performance of the weighting method and refer to sets of different projections as different ensembles with small and large discrepancy. A flow chart that demonstrates the relationship between different concepts and weighting schemes are presented in Fig 1.
We focus on projections of FMD made by the Warwick model [7,9]. This model was developed in the early stages of the 2001 FMD outbreak by Keeling and coworkers to determine the potential for disease spread and the impact of intervention strategies [9]. Here, we utilized a modified version of the model used in 2001, and we briefly describe relevant aspects of the Warwick model with regard to ensemble modeling. Full details of the model can be found in [7,9]. The rate at which an infectious farm I infects a susceptible farm J is given by:
RateIJ=SusJ×TransI×K(dIJ)
(1)
where
SusJ=([Zsheep,J]ps,SSsheep+[Zcow,J]pc,SScow)
(2)
is the susceptibility of farm J and
TransI=([Zsheep,I]ps,TTsheep+[Zcow,I]pc,TTcow)
(3)
is the transmissibility of farm I and K(dIJ) is the distance-dependent transmission kernel, estimated from contact tracing [9]. In this model Zs, I is the number of livestock species s (sheep or cow) recorded as being on farm I, Ss and Ts measure the region and species-specific susceptibility and transmissibility, dIJ is the distance between farms I and J and K(dIJ) is the distance dependent transmission kernel. The parameters, ps, S, pc, S, ps, T and pc, T, are power law parameters that account for a non-linear increase in susceptibility and transmissibility as animal numbers on a farm increase. Previous work has indicated that a model with power laws provides a closer fit to the 2001 data than when these powers are set to unity [43,51,52].
This version of the model has previously been parameterized to fit to the 2001 FMD outbreak [43]. Region-specific transmissibility and susceptibility parameters (and associated power laws) capture specific epidemiological characteristics and policy measures used in the main hot spots of Cumbria, Devon and the Welsh and Scottish borders. The model is therefore fitted to five regions: Cumbria, Devon, Wales, Scotland and the rest of England (excluding Cumbria and Devon). A table listing all the parameter values used in this model is given in Tildesley et al. [43].
In order to obtain multiple modeling assumptions for ensemble modeling, we specified different transmission rates, i as
RateIJi=SusJ×k1iTransI×K(k2idIJ)
(4)
where k1i and k2i are constants, specific for each modeling assumption, that scale the transmissibility and the spatial kernel, respectively. k1i = k2i = 1, follow the parameterizations of Tildesley et al. [43] and by decreasing or increasing these constants, we obtain parameterizations that correspond to best or worst case expectations about the transmissibility and spatial range of transmission. We are interested in how the level of discrepancy among modeling assumptions influences the performance of the ensemble method and we therefore created two different ensembles with different k1i and k2i, as listed in Table 1. We refer to these as the large and small discrepancy ensemble, corresponding to large and small differences, respectively, in the underlying modeling assumptions used for projections.
DCs in our model were determined based upon both prior infection by an IP and future risk of infection in the same way as in previous work [8]. CPs were defined as farms that share a common boundary and were determined on an individual basis. The model was seeded with the farms that were predicted to be infected prior to the introduction of movement restrictions on the 23rd February. For each modeling assumption i and control action, 200 replicates were simulated and each simulation progressed until the epidemic died out.
To demonstrate concepts and explore the potential of using the Tebaldi et al. [47] approach for epidemiological considerations we initially focus on outbreak duration. This is often considered to be the most costly aspect of FMD outbreaks due to its implication for trade [53]. In section 2.7 we extend the methodology to multiple epidemic quantities. However, the outbreak duration example offers transparent transition from the original climate analysis of Tebaldi et al. [47] that considers the ensemble estimated difference between current and future mean temperatures. In order to introduce the framework to epidemiology, we consider the difference between the implemented and an alternative control action, attempting to show whether the inclusion of CP culling was an appropriate choice given the conditions at the start of the outbreak. As this is a post outbreak analysis, we know the final outbreak duration of the observed outbreak, but that is just a single realization and due to the stochastic nature of disease transmission, the exact outcome may be quite variable. We also have no observed outbreak under the alternative control action to compare with the implemented control. Under these conditions, the most appropriate quantities to compare are the mean duration of a large number of outbreaks under the two control actions, something that can only be achieved through epidemic modeling.
We are interested in comparing projections under the implemented control action to the observed data in order to estimate model weights. Using the Bayesian method of Tebaldi et al. [47], weights as well as statistics of the outbreak, like duration, are considered unknown random variables, and we denote the mean outbreak duration under the implemented and the alternative control action as μ and v, respectively, corresponding to the mean current and future temperature, respectively, for the climate application. In order to fit with the normal assumptions of the method, we consider log-duration in the analysis.
Weights are expressed through precision, λ = λ1, λ2,…, λn, with λi denoting the precision of modeling assumption i. The projection specific parameter xi indicates the mean of all replicates under the implemented control action (analog of current climate) for modeling assumption i. For the UK 2001 outbreak this included culling of IPs, DCs and CPs. The corresponding projection for the alternative control action (analog of future climate), that included culling of IPs and DCs is denoted yi. The relationship between projections and ensemble parameters is expressed as
xi~Normal(μ,λi-1)
(5)
yi~Normal(ν+β(xi-μ),(θλi)-1),
(6)
with Normal(μ,λi-1) denoting the normal distribution with mean μ and variance λi-1. Parameter θ is included to allow for difference in overall precision of the modeling assumptions under implemented and alternative control actions. However, since projections xi and yi are based on simulations, it is fair to assume that modeling assumption i that has a high precision for the observed control action also will do well for the unobserved action. This is incorporated by the λi term in both Eqs 5 and 6. For the same reason, we may expect that a projection of a large xi also corresponds to a large value for yi and thus β is included to allow for correlation between corresponding projections for the two control actions; a projection that e.g. over-predicts duration of the outbreak for the observed control action can be expected to also over-predict the alternative control action.
The analysis of Tebaldi et al. [47] also assessed bias of projections by their ability to reproduce observed current climate by incorporating the relationship between observed current climate x0, an unobserved true mean climate variable (μ) and the precision of natural variability τ0 through
x0~Normal(μ,τ0-1).
(7)
In climate modeling, it is a fair assumption that τ0 is a known, fixed parameter because it can be assessed through historical data. That would rarely be the case for the corresponding epidemic considerations, at least for emerging diseases. Using a single outbreak to evaluate bias, we clearly have no way of assessing variability in outcomes. We therefore include τ0 as an unknown, random variable that is estimated in the analysis as described in the following section.
To aid the interpretation and transfer from the climate to the epidemiological interpretation, we have included Table 2 that lists the variables used in the analysis.
Our main interest in terms of outcome under the implemented control action is μ rather than x0. However, it is clear that in addition to the mean duration of the outbreak, the uncertainty about the process also results in some variability in the outcomes that we need to consider. The stochastic simulations used to generate projections provide not only a mean simulated outbreak quantity, but also a range of outcomes that projects the variability. In the absence of repeated outbreaks to evaluate variability of outcomes, an obvious choice would be to use this information to inform the variability τ0. Defining the variability τi as the precision of projections under the implemented action for modeling assumptions i = 1,2,…,n we include a hierarchical structure in the analysis so that for i = 0,1,2,…,n
τi~Gamma(aτ,bτ),
(8)
where Gamma(aτ, bτ) indicates the gamma distribution with shape parameter aτ and rate bτ both of which are unknown parameters and are estimated in the analysis. Thus, as it would be cumbersome to elicit a fixed prior for τ0 based on our prior expectations about variability, we instead assume that τ0 comes from some, unknown distribution, and make use of τ1, τ2,…, τn to inform what this distribution should be.
Similarly, we need to model the variability of projections under the alternative control action, and denoting this φi we specify
The parameters aφ and bφ are conditionally independent from all other parameters in the analysis and can be modelled separately in the Bayesian analysis. As xi, yi, τi and φi are calculated from a finite number of realization with each modeling assumption and control action, there is some uncertainty related to this. Tebaldi et al. [47] however points out that while it is certainly possible to construct a Bayesian model that takes this uncertainty into account, the effect is minimal if the number of replicates is large. With the R = 200 replicates preformed here, the uncertainty of the mean will in practice have very little effect, and we have included xi, yi, τi and φi as fixed observations.
Priors for aτ and bτ were specified as a gamma distribution with shapes Aaτ and Abτ, respectively, and rates Baτ and Bbτ, respectively. Similarly, the priors for aφ and bφ, were specified as a gamma distribution with shapes Aaφ and Abφ, respectively, and rates Baφ and Bbφ, respectively. We explored different parameter choices for the hyperpriors and found that the results were insensitive to the choice of prior for a wide range of values. In the analysis presented, we used Aaτ = Abτ = Baτ = Bbτ = Aaϕ Abϕ = Baϕ = Bbϕ = 0.001. This corresponds to prior distributions with a mean of one and a variance of 1000, thus allowing for a wide range of plausible values.
Bayesian analysis requires the specification of prior parameters. We follow Tebaldi et al. [47] with priors specified as uniform on the real line for μ, ν, and β, and λi~Gamma(aλ, bλ) for i = 1,2,…, n and θ~Gamma(aθ, bθ). We also need to specify hyperpriors for aτ and bτ, and we implement aτ~Gamma(Aaτ, Baτ) and bτ~Gamma(Abτ, Bbτ). Denoting x = x1, x2,…, xn and y = y1, y2,…, yn, the full posterior distribution under these priors is given by
P(μ,ν,β,λ,θ,τ0|x0,x,y,τ1,τ2,…)∝∏i=1n(λiaλ−1e−bλλiλiθ1/2exp{ −λi2[ (xi−μ)2+θ(yi−ν−β(xi−μ))2 ] })θaθ−1e−bθθτ01/2exp{ τ02(x0−μ)2 }∏i=0n(τiaτ−1e−bττi)aτAaτ−1e−BaτaτbτAbτ−1e−Bbτbτ
(10)
This posterior only differs from the one defined by Tebaldi et al. in that we include τ0 as an unknown variable and use a hierarchical structure for its prior. Using Markov Chain Monte Carlo (MCMC) techniques as described in 2.9, we first performed the analysis with priors as specified by Tebaldi et al. [47] where applicable, i.e. aλ = bλ = aθ = bθ = 0.001, because they argue that this ensures that the prior contributes little to the posteriors.
However, we propose that this argument is not necessarily always valid. In particular λi could be expected to be sensitive to priors because it is essentially only fitted to two data points, xi and yi. Yet, based on approximations of conditional distributions, Tebaldi et al. argued that the gamma distribution with aλ = bλ = 0.001 is appropriately vague for the analysis. For transparency we here follow their approach and investigate the effect of the prior for the simplified model where β = 0. The mean of the conditional distribution of λi is then approximated by
E(λi|X0,X,Y)≅aλ+1bλ+12[ (xi−μ˜)2+θ(yi−ν˜)2 ],
(11)
where μ~ is the conditional mean of the distribution of μ, given by
μ~=(∑i=1nλiXi+τ0x0)/(∑i=1nλi+τ0)
(12)
and ν~ the corresponding value for v, given by
ν~=(∑i=1nλiyi)/(∑i=1nλi).
(13)
We stress that Σλi need not sum to one, as might be intuitive when using weights. As given by Eqs 11 and 12, the mean of μ and ν only depends on the relative values of λi, but the absolute values influence the width of the distribution, with the variance of the conditional distributions increasing with lower absolute values of λi (Table 3).
While a low value of aλ certainly ensures little contribution to the numerator in Eq (11), it is less evident that a low value for bλ contributes little to the denominator because if xi→μ~ and y→ν~, the denominator actually approaches bλ. Hence, to ensure that a low value of bλ can be considered vague such that our posterior is informed primarily by x0, x and y, we must conclude that |xi-μ~| or |yi-ν~| is clearly separated from zero. However, if λi≫λj for all i ≠ j and λi≫τ0, then μ~≈xi and ν~≈yi and nothing in the model prevents this relationship. In fact, if we consider the gamma prior with shape and scale parameters set to 0.001, the distribution has most of its density near zero, however with a fat tail (yet exponentially bounded) that allows for high values. In the current analysis, this corresponds to the prior belief that the majority of modeling assumptions will have very low precision while a few will have very high. Under this prior belief, it is expected that for some model i, λi≫λj for all i ≠ j. In the instance where instead τ0≫λi for all i, then μ~≈x0 and the approximation would hold, but we cannot expect that relationship.
As we cannot a priori be sure that the choice of aλ and bλ does not influence our posterior as long as they are arbitrarily small, we performed a prior sensitivity analysis and re-ran the analysis with aλ = bλ = 0.01 and aλ = bλ = 0.0001. We could expect that the sensitivity to priors depends on the difference among modeling assumptions, and we investigate this by analyzing ensembles with little and large discrepancy between assumptions in the ensemble as given by Table 1.
We refer to this as the Non Hierarchical Weighting (NHW) method.
If we cannot ensure that the analysis is insensitive to the choice of prior, it implies that our prior beliefs will influence how much different projections contribute to ensemble predictions with the current method. Using prior beliefs is sometimes desirable, and in section 2.6 we consider the situation where we trust some modeling assumptions more than others. However, it would rarely be the case that we would have substantial expectations that could inform the shape, aλ, and scale, bλ, of the prior for λ.
A potential solution might be to extend the model to include hierarchical priors such that the prior for λi is estimated in the model rather than a priori fixed. We make a slight change to the parameterization of the prior distribution such that
λi~Gamma(aλ,aλ/mλ),
(14)
i.e. specifying the distribution by its mean mλ and shape aλ, which are estimated in the model. In that way, we move our uncertainty up a level and express our beliefs about the distribution of mλ and aλ, rather than λ. Using mλ rather than bλ facilitates the specification of a prior for the mean precision parameter that corresponds to the priors previously specified on individual λi. This parameterization further aids the use of prior beliefs about weights in section 2.6.
While we can never be strictly uninformative in Bayesian analysis, the hierarchical prior can allow for a wide range of plausible mλ and aλ whereas the model presented in section 2.4 requires these to be specified explicitly. This also allows for the concept of “borrowing strength” [54], such that the distribution of each λi can be indirectly informed by all other precisions via the hierarchical distribution. This is often beneficial in situations where individual parameters are fitted to a small amount of data [55,56], which clearly is the case for λi here. To extend Eq (10) to a hierarchical model, we include hyperpriors such that
aλ~Gamma(Aaλ,Baλ)
(15)
and
mλ~Gamma(Amλ,Bmλ).
(16)
We performed the corresponding sensitivity analysis for the hierarchical ensemble prediction by applying hyperpriors Aaλ = Baλ = Amλ = Bmλ set to 0.01, 0.001 and 0.0001. We refer to this as hierarchical sensitivity set-up one. Secondly, we performed a sensitivity analysis, hierarchical sensitivity set-up two, where we fixed the shape parameters Aaλ = Amλ = 0.001 and only varied Baλ = Bmλ, again set to either 0.01, 0.001 or 0.0001.
We refer to this as model as the Standard Hierarchical Weighting (SHW) method.
Using expert opinions may substantially improve predictions [57], and there are several instances where incorporating prior beliefs that reflect the “trust” in different modeling assumptions could be useful. For instance, a policymaker might have more trust in one model type over another, and rather than excluding the models that are considered less reliable (i.e. giving them a priori zero weigh), it could be useful to include them, yet with less contribution to the ensemble.
In the case considered here, where modeling assumptions represent most likely, best and worst case in terms of parameterization, we might want to give the “most likely” modeling assumption higher weight. For the analysis with fixed aλ and bλ, described in section 2.4, we could merely elicit a different scale parameter bλ for each λi, such that modeling assumptions with high trust are given a low value. However, with the shape parameter aλ set to a low value (“vague” shape), the prior may have little effect on the posterior λi. Eliciting a high value of aλ would instead result in a posterior that is merely the results of our prior beliefs and we have no foundation for which to elicit some intermediate value.
In order to combine the hierarchical approach with informative priors, we propose a modification of the analysis presented in section 2.5, where the assumption of exchangeability is relaxed in the hierarchical structure with
λi~Gamma(aλ,aλ/m^λi),
(17)
where m^λi=wimλ and wi indicates the a priori trust in modeling assumption i. With wi = kwj, the prior distribution of λi will have a mean that is k times that of λj and from Eqs (12) and (13) the relationship also implies that before λ is estimated (i.e. involving the data x0, x and y), the outputs of modeling assumption i will contribute k times as much to μ and v as does assumption j.
To demonstrate the effect that a priori trust in different modeling assumptions can have on the posterior estimates, we consider the case where the best, most likely and worst case scenarios for each of the two varied parameters corresponds to percentile 2.5, 50 and 97.5, respectively, of a normal distribution. Given that the density at percentiles 2.5 and 97.5 then is 0.15 of that at the mode, we specify wi = 0.15 for i = 2, 3, 4 and 7, i.e. for modeling assumptions where one of the varied parameters follows the most likely scenario, whereas the other one is set to either worst or best case. With the same rationality, we specify wi = 0.021 for i = 5, 6, 8 and 9, i.e. modeling assumptions where both parameters follow either best or worst case expectations.
We also investigate the case where a high weight is given to a projection xi further away from the observed data x0. Consistently, modeling assumptions i = 5 predicted the shortest duration for all actions and ensembles. We therefore also performed the analysis with w5 = 1 and w1 = 0.021, and all other weights are as above. This allows us to investigate the performance of the informative weighting method when an outlier is up-weighted.
We refer to this method as the Informative Hierarchical Weighting (IHW) method.
In the above examples, we focused on a single epidemic quantity, allowing for transparent transition from the original Tebaldi et al. work [47] that focused on temperature. For epidemiology, it may however be useful to consider multiple epidemic quantities. This could be done in different ways, but here we offer a straightforward multi-quantity extension of the Bayesian model for the single epidemic quantity, based on the supposition that the relative weights are equal for all quantities. As such, we implement a single weighting parameter λ, shared among all quantities. For other parameters, we use a similar notation as for the single quantity analysis, but give many of the parameters an additional index q, indicating that the parameter is quantity specific. We expand the Bayesian model by defining
xi,q~Normal(μq,θμ,qλi-1),
(18)
yi,q~Normal(νq+βq(xi,q-μq),(θν,qλi)-1),
(19)
x0,q~Normal(μq,τ0,q-1),
(20)
where xi, q and yi, q are the mean projections of modeling assumption i for epidemic quantity q for the implemented and alternative control action, respectively, and x0, q is the corresponding observed value. As for the single epidemic quantity example, μqand Vqare the expected values of quantity q, and because we cannot expect to have the same correlation between control actions for all quantities, βq is included as unique for each q. Parameters θμ, q and θv, q scales the precision of models between actions and quantities and the parameters of the Bayesian model are identifiable by defining θμ, 1 = 1. Similarly, we specify quantity specific parameters
τi,q~Gamma(aτ,q,bτ,q),i=0,1,2,…,n,φi,q~Gamma(aφ,q,bφ,q),i=1,2,…,n.
(21)
The conditional distributions for the multi-quantity extension are provided in Table 4. We denote the total number of quantities in an analysis as Q.
The above examples focus on the UK 2001 FMD outbreak and show how the introduced framework can be applied to actual outbreak data. However, a limitation to this approach is that we are confined to investigating the behavior of the ensemble methodology for that particular outbreak. To further investigate the potential and limitations of the proposed methods, we also performed analysis of simulated outbreak data. With simulated data, we have “true” estimates of μ and v, and we want to explore the ability of the ensemble to predict these under two different conditions; when the true values lies within the range of X and Y predicted by the individual models of the ensemble and when it does not. For multi-model ensembles, this corresponds to the situation where the true behavior of the outbreak is encapsulated within the range of underlying assumptions of the individual models and when it is not.
Here we explore the outcome of these conditions by first simulating outbreaks with the parameterizations of modeling assumption 1 (k1 = k2 = 1), i.e. located in the center of both the small and large discrepancy ensemble. This simulates outbreaks where the true behavior of the outbreak is encapsulated within the range of underlying assumptions of the individual projections for both ensembles. We also simulate outbreaks with a parameterization where both k1 and k2 are set to 0.9. This produces outbreaks where the true behavior is outside of the assumptions of the projections for the small discrepancy ensemble, yet inside the range of the large discrepancy ensemble.
The exact behavior of the ensemble depends on the actual realization of the individual outbreak, because the observed values x0 are different due to the stochastic disease transmission process. We therefore apply both the small and large discrepancy ensembles to ten realizations of each of the simulation parameterizations. We implement both the single and multiple epidemic quantity analysis, thus further highlighting the effect of using multiple quantities.
We use Markov Chain Monte Carlo (MCMC) techniques to obtain samples from the full posterior distribution of the proposed Bayesian models (NHW, SHW and IHW). For many parameters, the conditional distribution belongs to a standard parametric family, thus allowing for Gibbs sampling. We list these conditional distributions in Table 3 for single quantity analysis and Table 4 for multiple quantities.
We also rely on Metropolis-Hastings (M-H) updates, and with the computation used for multi-quantity analysis being a straightforward extension of that used for the single quantity, we start by describing the update scheme for the single quantity analysis. The conditional distribution of bτ has a known form, P(bτ|…)=Gamma(Abτ+(N+1)aτ,Bbτ+∑i=0Nτi), that would allow for Gibbs sampling of bτ, whereas M-H updates need to be implemented for aτ. We however found strong correlation between the marginal posterior estimates of aτ and bτ, and mixing was improved by performing joint M-H updates of these parameters by multivariate Random Walk (RW) proposals. Mixing can be further improved by updating parameters on a transform that resembles a Gaussian distribution, and we therefore performed updates on the log-transform, i.e. based on current values of aτ and bτ We proposed candidate parameters [log(aτ*),log(bτ*)] from MVN([log(aτ),log(bτ)],Στ). Here MVN indicates the multivariate normal distribution and Στ the covariance matrix. Candidate values are accepted with the probability
min(1,Gamma(aτ*|Aaτ,Baτ)Gamma(bτ*|Abτ,Bbτ)∏i=0NGamma(τi|aτ*,bτ*)Gamma(aτ|Aaτ,Baτ)Gamma(bτ|Abτ,Bbτ)∏i=0NGamma(τi|aτ,bτ)|Jτ|),
(22)
where |Jτ|=aτ*bτ*(aτbτ)-1 indicates the Jacobian determinant of the log-transform.
Mixing can be improved if the covariance matrix Στ is proportional to the covariance of the marginal posterior of [log(aτ),log(bτ)], here indicated as Φ. However, this is not known prior to the analysis. We therefore implement an optimized method of the Robbins-Monroe search process as presented by Garthwaite et al. [58]. This estimates the covariance during the MCMC and finds the scaling parameter ρ such that Στ = ρΦ provides a chosen long term acceptance rate, here set to 0.234 based on suggestions by Roberts et al. [59]. The method has been demonstrated to be appropriate also for RW on transformed scales of the parameters [60].
The corresponding updates of aφ and bφ were also performed with M-H updates and we proposed candidate parameters [log(aφ*),log(bφ*)] from MVN([log(aϕ),log(bϕ)],Σϕ). and accepted them with probability
min(1,Gamma(aφ*|Aaφ,Baφ)Gamma(bφ*|Abφ,Bbφ)∏i=1NGamma(φi|aφ*,bφ*)Gamma(aφ|Aaφ,Baφ)Gamma(bφ|Abφ,Bbφ)∏i=1NGamma(φi|aφ,bφ)|Jφ|).
(23)
We used a similar approach for updates of aλ and mλ in the hierarchical methods (SHW and IHW) and proposed [log(aλ*),log(mλ*)] from MVN([log(aλ),log(bλ)],Σλ). Candidate parameters were accepted with probability
min(1,Gamma(aλ*|Aaλ,Baλ)Gamma(mλ*|Abλ,Bbλ)∏i=1NGamma(λi|aλ*,b^λi*)Gamma(aλ|Aaλ,Baλ)Gamma(mλ|Abλ,Bbλ)∏i=1NGamma(λi|aλ,b^λi)|Jλ|),
(24)
where b^λi=aλ/mλ for all modeling assumptions i in the SHW method and b^λi=aλ/m^λi with m^λi=wimλ in the IHW method. As above, we used the method of Garthwaite et al. [58] to determine Σλ to obtain a long term acceptance rate of 0.234.
We also found strong correlation between μ and ν. In order to improve mixing, we repeated the updates of these parameters five times for each iteration of the MCMC.
The same update scheme was used for the multi-quantity consideration, yet with a separate Στ, q, Σϕ, q and Σλ, q adaptively estimated for each quantity q.
The algorithm was implemented in MATLAB (The MathWorks, Inc., Natick, Massachusetts, United States) and code is available as supplementary information (S1 File).
We start by presenting the results for the single quantity analysis, highlighting the behavior of the method for the NHW, SHW and IHW schemes. Fig 2, panels A and B show the estimates of outbreak duration for the two control actions for the large discrepancy ensembles using the NHW method and reveals rather large prior sensitivity. Note that we plot marginal posteriors of M = eμ and N = ev, respectively. As such, the posteriors represent the geometric mean outbreak duration. The corresponding arithmetic mean can be calculated as eμ+1/(2τ0) and ev+1/(2ϕ0), respectively, yet we use the geometric mean as it more clearly shows the relationship with individual projections, here presented by xi = exi and Yi = eyi, respectively. For aλ = bλ = 0.0001 (solid gray lines), the distributions are multimodal with peaks at individual model predictions, whereas a more smooth shape is obtained for aλ = bλ = 0.01 (dashed black lines) and aλ = bλ = 0.001 (solid black line) yields an intermediate result.
With the SHW method, we instead obtain posteriors that are largely insensitive to the choice of hyperprior. Fig 2, panels E and F present the result of sensitivity set-up one, showing near identical posterior estimates when hyperparameters Aaλ, Baλ, Amλ and Bmλ are set to 0.01, 0.001 or 0.0001. Sensitivity set-up two produced results that were visually indistinguishable from panels E and F and are not presented. This further indicates that the hierarchical method is robust to the choice of hyperpriors.
Within epidemiology, there is clearly an interest in not just the expected outbreak duration, but also other statistics such as the probability of large outbreaks occurring. We therefore consider the posterior predictive distributions of individual outbreak durations under the two control actions. For the non-hierarchical model (Fig 2, panels C and D), there is an obvious effect of the choice of prior with higher probability of long outbreaks for lower values of aλ and bλ. For the hierarchical model (Fig 2, panels G and H), there is again little difference among posteriors corresponding to different priors.
When evaluating the efficiency of control actions, the difference N-M is of particular interest. In the example presented here, this estimates how much longer the outbreak would have been if culling of CPs had been excluded from the control. As shown in Fig 3, the estimates are again sensitive to the choice of prior with the NHW method, yet insensitive with the SHW method. The range of the posterior under the NHW method is less sensitive to the prior for the low discrepancy ensemble (panel B) than for the large discrepancy ensemble (panel A), where higher probability of less difference is estimated with aλ = bλ = 0.01 than for aλ = bλ = 0.0001. However, the multimodal behavior of the NHW method with low values of aλ and bλ is obtained also for the low discrepancy ensemble.
Fig 4 demonstrates the effect that a priori beliefs about the weights have on the predicted outbreak duration under large and small discrepancy ensembles. When using a priori higher weights for the most likely scenarios (modeling assumption one; black dotted lines), the posterior estimates are shifted and become more centered on projections of that particular modeling assumption compared to the case where a priori weights are equal (black solid line). The outcome of up-weighting the outlier (modeling assumption five; solid gray lines) is however different between the two ensembles. For the small discrepancy ensemble (panels A and B), similar results are found as for the up-weighting of the most likely scenarios; posteriors are shifted towards the projection with a priori high weight. For the high discrepancy ensemble (panels C and D), the posterior estimates of outbreak duration instead become wider for both control actions, indicating larger uncertainty about the expected duration of outbreaks.
Fig 5 shows the marginal posterior estimates of individual weights λi under different discrepancy among projections and informative weighting schemes. When using a priori equal weights, there is little difference in the estimates for the small discrepancy ensemble (top left panel) whereas moderate differences are obtained for large discrepancy (bottom left panel). Note that while the error bars are overlapping, the mean estimate of the most likely scenarios (modeling assumption one) is approximately 1.7 times as large as that of the outlier(modeling assumption five), meaning that the former will contribute approximately 1.7 times as much to the posterior means of μ and v than the latter (Eqs (12) and (13)).
When giving a priori highest weight to the most likely scenario (modeling assumption one; middle column panels), the posterior estimate of λ1 is consistently shifted upwards, meaning that the most likely scenarios (modeling assumption one) will contribute more to the posteriors of μ and ν than other projections. For the up-weighting of the outlier, projections corresponding to modeling assumption five, the same is found when there is little discrepancy among projections (lower right panel). This is however not found for the high discrepancy ensemble (top right panel), where the main effect is that compared to equal a priori weights (top left panel), the error bars are wider; this indicates larger uncertainty about weights and consequently about the contribution of individual projections to the posterior estimates of outbreak durations.
The proposed multi-quantity method can be implemented with either NHW, SHW or IHW schemes. Here we aim to illustrate the effect of using multiple quantities and focus on the SHW scheme. Fig 6 plots the marginal posterior density of mean outbreak duration under the two control actions as estimated for the multiple quantity analysis (solid) together with the corresponding estimates for the single quantity analysis (dashed). The figure illustrates how inclusion of multiple quantities in the analysis leads to tighter distributions, centered on projections for i = 1. The multi-quantity analysis produces a probability distribution of all considered quantities, and Fig 7 further illustrates how the marginal posterior densities are located above zero for all three considered quantities.
To illustrate the performance of the method under different conditions, we also analyzed simulated outbreaks. Fig 8 shows the posteriors of mean duration for outbreaks simulated with the k1 = k2 = 1 parameterization and applying the small (triangles) and large (circles) discrepancy ensembles, represented by the median values and error bars indicating the 95% central credibility interval. Note that individual realizations, indicated by stars, are expected to frequently be outside of the credibility envelopes. Error bars are inclusive of the true mean outbreak duration (dashed lines) for all ten analyzed realizations for both the implemented and alternative control actions. However, the credibility envelopes are tighter and medians closer to the true value for the multi-quantity analysis. This indicates that the ensemble prediction is improved by including multiple quantities.
When applying the analysis to outbreaks simulated with the k1 = k2 = 0.9 parameterization (Fig 9), the large discrepancy ensemble error bars are still consistently inclusive of the true value. As with Fig 8, credibility envelopes are tighter for the multi-quantity analysis. The error bars of the small discrepancy ensemble that all rely on simulations with parameterizations with higher k1 and k2 than the true value, are not inclusive of the true value, indicating that the small discrepancy ensemble fails in predicting the true values of the outbreak.
Ensemble modeling is appealing because it offers the possibility to combine multiple projections. Within weather forecasting, the approach has given more robust predictions, and we could expect that to be the case for epidemiology as well. However, there is a need for the development of methods describing how to combine several epidemiological projections. The aim of this study was to investigate the possibility of using the Bayesian framework introduced by Tebaldi et al. [47]. We find that it is a promising approach, for primarily three reasons.
Firstly, when the methodology is implemented in a hierarchical Bayesian framework, it provides an appealing interpretation of model exchangeability. Essentially, projections and their underlying modeling assumptions are treated as random draws from a population of possible projections. By estimating the hierarchical parameters aλ and mλ jointly with individual precisions (weights) λi, the characteristics of this hypothetical population are estimated. Smith et al. [61] used a similar approach for climate ensembles and pointed out that this reduces the impact of which models are included or excluded in the ensemble. That is, we should expect to get similar results when using a different set of model assumptions if they are chosen independently. We stress that this interpretation is more valid for multi-model ensembles, however. Also, the term “random draws” should not be interpreted as arbitrary. Rather, the interpretation is that models should come from a population of well-informed, reasonable models. The analysis treats the outputs of the performed simulations under different assumptions as data (Eq 1, Table 1), and as such they are used to inform the quantities of interest (μ and v). This may seem counterintuitive, yet it only serves as a formal means to combine the results of multiple projections, and by Eqs (7) and (20), these are combined with available outbreak data.
Secondly, the framework can handle several different weighting schemes simultaneously. The original methods introduced by Tebaldi et al. [47] used convergence and bias to assess weights. Here, we further extend the framework such that informative priors can be included to inform the weights, thus relaxing the supposition that all modeling assumptions are a priori exchangeable. Epidemiological predictions suffer from lack of available data to assess model bias, and we propose that expert opinions will play a larger role than in other fields of research. With the analytical tool proposed here, a policymaker can choose to include a range of projections based on different modeling assumptions, yet give them different weights, rather than including one or a few (given a weight of one) and excluding others (given a weight of zero). When using different mechanistic models, subjective trust in the different models can be incorporated by using methods of prior elicitation based on expert opinion [38]. Importantly, our methods can incorporate these subjective beliefs in the hierarchical framework, requiring only the specification of the a priori relative confidence in the underlying assumptions of the projections. Definition of an individual, fixed prior would undoubtedly be cumbersome to elicit from expert opinion; it would not be feasible to ask policymakers to define an individual gamma prior for each modeling assumption.
Here we used ensembles based on projections of the same model with different parameterizations, demonstrating the possibility to explore parameter space, yet with unequal probabilities of different parameterizations. Uncertainty about parameters will be an issue for most epidemiological models, and we propose that multi-model ensembles should incorporate projections with different models and different parameterizations. Thus, different mechanistic assumptions as well as parameter uncertainty would be incorporated in the ensemble.
Thirdly, the framework produces easily interpretable probability distributions. It is important that communication with policymakers include uncertainties about prediction rather than just the most likely outcome [16]. In the ensemble context, these uncertainties take into account different assumptions about the transmission process. Gårdmark et al. [32] suggested that uncertainty should be communicated with policymakers by presenting the full range of predicted outcomes. However, that would give equal weights to all included projections and would require that the results be communicated with a detailed description of all assumptions made, thus allowing the policymaker to decide how much to trust each modeling assumption. This would be a cumbersome task, particularly for detailed simulation models that rely on a large number of parameters. We therefore argue that it is beneficial to communicate the aggregated and weighted result as easily interpretable probability distributions. With further modifications of the methodology, we propose that the approach could also be used as a forecasting tool during an outbreak, e.g. by letting xi and yi denote current and future numbers of infected farms. In such a situation, there is a great need for rapid and clear communication of model results to aid policy decisions. The visual manner in which uncertainty is presented using probability distributions makes them easy to understand and communicate [62].
We here show that these distributions are sensitive to the choice of priors when using the NHW method (Fig 2, panels A-D, Fig 3, panels A, B). However, the impact of the prior is heavily reduced when using the hierarchical framework (Fig 2, panels E-H, Fig 3, panels C, D). Thus, our results demonstrate that the hierarchical approach is preferred for ensemble modeling and using the non-hierarchical approach can lead to spurious conclusions. We argue that this would also be the case for other fields, such as climate ensembles, but it is likely to be a larger concern for epidemiology where data to modify the prior are fewer. Considering Eq (11), we could ensure that b has little contribution to the denominator if τ0≫λi for all i, ensuring that the prior has little contribution to the posterior. For climate considerations, we envisage that the precision of natural variability, τ0, would be large relative to each λi if bias is assessed by comparing model simulations to long time series of climate data. For epidemiological considerations, this would however rarely be the case. In the proposed method, we instead inform τ0 largely by the simulation outputs, letting the projections of the ensemble determine how variable outcomes are.
Climate modeling, from which the proposed method is adapted, is primarily concerned with differences between current and future mean climate variables [24]. Epidemiology is not only concerned with mean projections but also with other quantities such as the probability of very large or long outbreaks occurring. Fig 2, panels G and H illustrates the probability of a given epidemic duration occurring for a single outbreak under the two control actions with the preferred SHW method. Comparing the posterior predictive distribution to the density of merely lumping the results of all simulations, as illustrated by the colored bars, the posterior predictive distribution of the ensemble method has a lower probability of both very long and short outbreaks. This is because projections of such outbreaks are down-weighted when their bias is assessed in the analysis; the observed outbreak duration would be unlikely under the modeling assumptions that produce these projections. Thus, ensemble methods that give equal weights to all projections can overestimate the uncertainty about outbreaks, preventing the models from informing appropriate policy decisions.
We have further extended the methodology to allow for informative priors on the weights. Compared to climate models, epidemiology often has far less available data to assess model bias. As such, expert opinion will often play a larger role within this field. Fig 4 illustrates the behavior of the ensemble prediction under such informative priors. When up-weighting projections for i = 1, which is also likely under the observed outbreak duration, the posteriors are shifted towards these projections and produce tighter distributions. This is also found when up-weighting the outlier, i = 5, in the small discrepancy ensemble (Fig 4, panels A and B), in which no projection is particularly unlikely for the observed duration. Projection x5 is however unlikely in the large discrepancy ensemble. As a result, the effect of up-weighting the underlying modeling assumptions of this projection primarily makes the distribution wider, resulting from a larger uncertainty about individual weights (Fig 5). This is to be interpreted such that if expert opinions a priori determine that a modeling assumption that is unlikely to predict the observed data is better than other assumptions, the conclusions should be that there is less information in the ensemble as whole. However, when expert opinions are well informed and do not contradict with observed data, they can lead to more precise predictions.
It should be stressed that discrepancy among projections in the ensembles should be viewed as relative to τ0, the estimated variability in outbreak duration given the initial conditions. A crucial difference between the original method applied to climate change and the epidemiological consideration presented here is that τ0 is unknown for the latter and therefore needs to be estimated. We argue that in the absence of multiple outbreaks, it is sensible to inform this by the model simulations. Stochastic simulations are often used to estimate the range of outcomes for non-ensemble projections [1,17,18,23], and we propose that when extending the use of models to the ensemble context, they can be used to estimate this feature as well. We have therefore chosen a Bayesian model structure where τ0 is informed largely by the within projection variability, τ1, τ2,…, τn, via the Gamma(aτ, bτ) distribution in Eq (8). All projections of the implemented control action contribute equally to this distribution in the method presented here, thus we are giving equal weights to all modeling assumptions in terms of informing τ0. Estimation of different weights in terms of informing τ0 based on a single outbreak, analogous to the estimation of λ, would not be conceivable. However, if policymakers believe that some modeling assumptions are more reliable in terms of capturing the variability of outcomes, we envisage that the Bayesian model structure can be altered to include this. If applied to endemic disease, τ0 could be informed similarly to the natural variability of temperature in climate application, and the algorithm we supply is set up to handle this situation. Also, data from multiple outbreaks could be used to inform τ0 when available. Yet, data quality will rarely be comparable to climate data, which highlights one of the major challenges for epidemiological modeling.
We also provide a multi-quantity extension of the Bayesian ensemble framework. Fig 5 shows that when adding number of infected and culled farms to the analysis, the marginal posteriors of outbreak duration become narrower and centered on x1 and y1, i.e. the projections based on the most likely scenario. This illustrates that predictions can be improved by incorporating multiple quantities when assessing the weights.
The main scope of this study is to introduce ensemble methods to the field of epidemiology rather than to produce inference about the 2001 FMD outbreak. However, Fig 7 illustrates the types of conclusions the method can provide. The three quantities we include in the multi-quantity analysis are all of great concern to policy makers when assessing the impact of control actions. The probability distributions represent the ensemble predicted difference in the outcome of the outbreak if the control action had excluded culling of CPs. The distributions all have most of the density above zeros, indicating that excluding culling of CPs would most likely have resulted in a prolonged and larger outbreak. We should however point out that these results are based on a single model. To make more robust predictions, we propose that the same type of analysis be made with projections of different models.
We also analyzed simulated data to provide a more general depiction of the performance of the method under different conditions. Fig 8 shows the result of analysis of ten simulated outbreaks with the parameterization in the center of the ensemble, i.e. k1 = k2 = 1. As this is in agreement with both the small and large discrepancy ensemble, the true values (dashed lines) consistently lie within the 95% credibility intervals. However, when using the k1 = k2 = 0.9 parameterization, the assumptions of the model used to simulate the outbreak is only inclusive of the large discrepancy ensemble, and consequently only the large discrepancy ensemble error bars are inclusive of the true values. Noting that we primarily use the different parameterization as a proxy for different models, this simple simulation example illustrates some obvious but essential points. Ensemble modeling should not be interpreted as a remedy for models based on poor assumptions about the modeled process. It offers the ability to combine multiple assumptions, thus integrating uncertainty with regards to this in the predictions. However, if all models are based on similar but inaccurate assumptions, ensemble modeling will not improve predictions. Intentionally making models similar to each other increases this risk and should be avoided if the models are to be used for ensemble purposes.
Accepting these limitations, we argue that the ensemble approach will be beneficial to epidemiological risk assessment because rather than choosing a single model for the purpose, it offers the possibility to combine projections from models that make mechanistically different assumptions about the transmission process. Thus, uncertainty with regard to this is incorporated in the predictions, which is important as projections of different models have been reported to deviate [63–65]. The use of multi-model ensembles would rely on collaboration of modeling teams, as well as overcoming confidentiality constraints in accessing outbreak data and population demographics. The current development in FMD modeling is seeing encouraging development in that area. The Quadrilateral Epiteam [19] has compared simulation of several outbreak scenarios in a subset of the UK demographics with five different models: NAADSM [45], Netherlands CVI [66], InterSpreadPlus [46], AusSpread [44] and ExoDis [67].
This demonstrates that potential obstacles for multi-model ensembles can be overcome and we envisage that epidemiology will see a shift towards multi-model ensembles to inform policy decisions, as has been seen in climate research [24,25] and weather forecasting [26,27]. Combining the results of multiple models however requires means of weighting these. We conclude that the presented framework is a promising approach because it provides easily interpretable probability distributions of quantities of interest. It also offers an appealing interpretation of model exchangeability, while at the same time combining several different weighting schemes, including a priori beliefs when such are available.
In this study, we introduced this framework by applying it to a simple question: how would exclusion of contiguous premises culling from the control action have affected the outcome of the UK 2001 outbreak? The aim of the study has been to introduce the methodological framework to epidemiology and solve some key issues associated with this transfer, including prior sensitivity, informing weights by expert opinion, using models to inform the variability in the outcome of individual outbreaks and extension to consider multiple epidemic quantities. We have purposely chosen the simple example because it allows for a straightforward transfer from the original climate implementation, and at the same time lets us demonstrate essential concepts and the potential of the framework. Models are however used to answer a range of different questions in epidemiology, and combining multiple projections has the potential to improve the way models are used to inform policy. We argue that the framework we introduce here has great potential, and foresee that many of the questions addressed in epidemiological modeling would require further developments of the Bayesian model, structured to fit with the specific problem. To facilitate this, we have supplied the algorithm (S1 File) and hope that it will aid further development of ensemble methods for epidemiology.
|
10.1371/journal.ppat.0030080 | A Molecular Link between Malaria and Epstein–Barr Virus Reactivation | Although malaria and Epstein–Barr (EBV) infection are recognized cofactors in the genesis of endemic Burkitt lymphoma (BL), their relative contribution is not understood. BL, the most common paediatric cancer in equatorial Africa, is a high-grade B cell lymphoma characterized by c-myc translocation. EBV is a ubiquitous B lymphotropic virus that persists in a latent state after primary infection, and in Africa, most children have sero-converted by 3 y of age. Malaria infection profoundly affects the B cell compartment, inducing polyclonal activation and hyper-gammaglobulinemia. We recently identified the cystein-rich inter-domain region 1α (CIDR1α) of the Plasmodium falciparum membrane protein 1 as a polyclonal B cell activator that preferentially activates the memory compartment, where EBV is known to persist. Here, we have addressed the mechanisms of interaction between CIDR1α and EBV in the context of B cells. We show that CIDR1α binds to the EBV-positive B cell line Akata and increases the number of cells switching to the viral lytic cycle as measured by green fluorescent protein (GFP) expression driven by a lytic promoter. The virus production in CIDR1α-exposed cultures was directly proportional to the number of GFP-positive Akata cells (lytic EBV) and to the increased expression of the EBV lytic promoter BZLF1. Furthermore, CIDR1α stimulated the production of EBV in peripheral blood mononuclear cells derived from healthy donors and children with BL. Our results suggest that P. falciparum antigens such as CIDR1α can directly induce EBV reactivation during malaria infection that may increase the risk of BL development for children living in malaria-endemic areas. To our knowledge, this is the first report to show that a microbial protein can drive a latently infected B cell into EBV replication.
| Malaria and Epstein–Barr virus (EBV) infections are recognized cofactors in the genesis of endemic Burkitt lymphoma, the most common paediatric cancer in equatorial Africa. EBV is a ubiquitous virus residing in B lymphocytes that establishes a lifelong persistence in the host after primary infection. EBV has two lifestyles: latent infection (non-productive), and lytic replication (productive). Children living in malaria-endemic areas exhibit an elevated viral load, and acute malaria infection increases the levels of circulating EBV. The mechanisms leading to viral reactivation during Plasmodium falciparum malaria infection are not well understood. Cystein-rich inter-domain region 1α (CIDR1α) is a domain of a large protein expressed at the surface of P. falciparum–infected red blood cells. Based on previous findings showing that CIDR1α activates and expands the B cells compartment where EBV persists, we assessed the impact of CIDR1α on viral reactivation. Here, we identify CIDR1α as the first microbial protein able to drive a latently EBV-infected B cell (no virus production) into lytic replication (virus production). Our results suggest that P. falciparum–derived proteins can lead to a direct reactivation of EBV during acute malaria infection, increasing the risk of Burkitt lymphoma development for children living in malaria-endemic areas.
| Epstein–Barr virus (EBV) is a human γ-herpes virus that establishes a persistent infection in >90% of the world's population. Like other herpes viruses, EBV has two alternative lifestyles: latent (non-productive) infection, and lytic (productive) replication. Following primary infection, EBV persists within memory B lymphocytes in a latent state for the life of the host. A low level of reactivation into lytic replication allows viral shedding into the saliva and transmission of the virus in vivo [1]. The lifelong persistent infection established by EBV is harmless in almost every host and rarely causes disease, unless the host–virus equilibrium is upset. Thus, viral persistence represents a balance between viral latency, viral replication, and host immune responses.
The lytic phase of viral replication can be triggered in vitro by a variety of reagents and stimuli, including halogenated pyrimidine [2], phorbol esters [3], calcium ionophores [4], transforming growth factor β [5], butyrate [6], and triggering of the B cell receptor (BCR) with anti-immunoglobulin (anti-Ig) antibody (Ab) [7]. Less is known about the physiological stimuli that control activation of the virus productive cycle in vivo, although replication seems to occur following plasma cell differentiation [8].
It has been well documented that EBV is causally associated with various malignancies, including endemic Burkitt lymphoma (BL), nasopharyngeal carcinoma, and B cell lymphoma, in immunocompromised hosts [9]. Both EBV infection and intense exposure to Plasmodium falciparum malaria (holoendemic malaria) are recognized cofactors in the pathogenesis of BL, which is the most common paediatric cancer in equatorial Africa, accounting for up to 74% of childhood malignant disorders [10]. Development of BL, a B cell malignancy, is heralded by high Ab titers to replicative antigens indicative of EBV reactivation [11]. Recent reports indicate that the impact of malaria infection on EBV persistence is reflected by an increased viral replication. Children living in malaria-endemic areas have an elevated EBV load [12,13], and acute malaria infection leads to increased levels of circulating EBV that are cleared following anti-malaria treatment [14].
The mechanisms that may lead to viral reactivation during P. falciparum malaria are not well understood. The identification of a polyclonal B cell activator and Ig-binding protein in P. falciparum is of particular relevance in this context. We demonstrated that the cystein-rich inter-domain region 1α (CIDR1α) of the P. falciparum erythrocyte membrane protein 1 (PfEMP1) induces proliferation and activation of B cells, preferentially of the memory subset, where EBV is known to reside [15,16]. To understand the relative contribution of malarial antigens on EBV reactivation, we used the prototype EBV-positive BL cell line Akata as a model to determine whether CIDR1α could induce reactivation of the EBV lytic cycle in latently infected B cells. Furthermore, we analyzed the effect of the CIDR1α on freshly isolated peripheral blood mononuclear cells (PBMCs) from EBV-positive healthy donors and from children with BL living in malaria-endemic areas. The results support the hypothesis that CIDR1α is one of the molecules involved in EBV reactivation during the course of malaria infection. Our data provide new insights into how malaria infection may contribute to BL development.
During the blood stage of P. falciparum malaria, infected red blood cells (iRBCs) express high levels of PfEMP1, reaching their maximum at the trophozoite stage (28–32 h post-invasion). The CIDR1α domain of PfEMP1 (clone FCR3S1.2-var1) has a multi-adhesive phenotype and binds to different cell surface receptors, such as CD36, PECAM-1 (CD31), and immunoglobulins (Igs) [17]. CIDR1α also binds to isolated B cells via an interaction that involves surface Igs [15].
To establish whether iRBCs and the soluble form of CIDR1α interact with EBV-carrying B cells, we used the EBV-positive BL cell line Akata as a model. Akata cells stained with PKH67 (green) were co-incubated with PKH26 (red)-stained uninfected red blood cells (RBCs), or with enriched iRBCs at the trophozoite stage (28 h post-invasion, 75%–80% final parasitemia), at a ratio of 1:2, respectively. RBCs did not bind to Akata (Figure 1A), but co-incubation with iRBCs led to the formation of conjugates that varied in size but frequently involved two to five iRBCs/Akata cell (Figure 1B). A higher magnification of the conjugates showed a polarization of the iRBC, where the parasites were found at the proximity of the membrane's tight junction between the two cell types (Figure 1C).
Being that CIDR1α is a domain expressed on iRBCs with multi-adhesive phenotypes [17], we investigated its ability to bind Akata cells. A soluble form of CIDR1α was produced as a glutathione-S-transferase (GST) fusion protein, and the GST protein alone was used as control. Immunofluorescence studies with anti-GST fluorescent Abs demonstrated that CIDR1α, but not the GST control, binds to the membrane of Akata cells (Figure 1D and 1E). Flow cytometry analysis showed a peak shift representing an increased mean fluorescence intensity (MFI) as compared to the MFI values obtained when Akata cells were incubated with GST control protein or with the isotype control Ab (Figure 1F). Thus, both iRBCs and the recombinant CIDR1α domain of PfEMP1 bind to the EBV-carrying B cell line Akata.
In contrast to the variety of reagents and signals able to induce EBV lytic production in vitro [18], the physiological signals involved in the activation of the virus productive cycle have not been well characterized yet, although plasma cell differentiation seems to represent one such trigger [8]. Anti-Ig treatment, which leads to BCR signalling [19], has served as a relevant in vitro model of reactivation for inducing virus replication in some EBV-carrying BL cell lines, including Akata [7,20].
Because CIDR1α is an Ig-binding protein [17], we analyzed its functional impact on the reactivation of lytic virus production and used the well-characterized Akata cell line model as a read-out system. First, we analyzed whether stimulation of Akata cells with CIDR1α would affect the number of viral DNA copies produced. Cells were cultured with increasing concentrations of CIDR1α, GST (range 0,5–2 μM, corresponding to 25–100 μg/mL), or in medium alone. After 48 h, we quantified the EBV viral DNA copy number in the cultures (cells + supernatant) by monitoring the EBV LMP1 gene, which is present as a single copy in the virus genome. As shown in Figure 2A and 2B, stimulation with CIDR1α increased the viral DNA load in a dose-dependent manner. Cells incubated with GST contained numbers of EBV genomes comparable to that of cells cultured in medium alone. In two out of four independent experiments, there was a statistical significance in relative increase of EBV load between the concentrations of 0,5 μM and 2 μM (p = 0,03).
We have recently demonstrated that CIDR1α induces B cell proliferation and protects B cells from apoptosis [16]. It could be then argued that the augmented viral load might simply result from a net increase in the number of cells in the culture. Cell cycle analysis performed by propidium iodide staining after 24 and 48 h of incubation did not reveal any significant change in the proportion of dead (<G0/G1) or cycling cells (S-G2/M) between cultures containing CIDR1α, GST, or medium alone (unpublished data). Moreover, experiments carried out in the presence of z-VAD, a pan-specific caspase inhibitor that blocks apoptosis, did not significantly affect the extent of EBV DNA increase in cultures containing GST or CIDR1α (Figure 2B), ruling out the involvement of apoptosis. In conclusion, stimulation of the EBV-carrying B cell line Akata with CIDR1α leads to an increased number of viral genomes that is not dependent on apoptosis or increased proliferation.
The interaction between iRBCs and B cells is partially mediated by CIDR1α [15] that is expressed at the iRBC surface along with a variety of other antigens [21]. Therefore, it became of interest to see whether the interaction of Akata cells with iRBCs would lead to viral reactivation. Intact iRBCs could not be used for this purpose, as detection of viral production requires 48 h of co-incubation with Akata, and during this time the iRBCs burst, leading to toxicity and cell death. The erythrocytic parasite cycle from invasion to merozoite release takes 48 h and the purification of synchronized iRBCs requires a high cellular content of the paramagnetic pigment hemozoin that is reached 28 h post-invasion; i.e., the rupture would occur in the middle of the test. To overcome the iRBC burst and the related cytotoxicity, we used crude extracts from RBCs and iRBCs obtained 28 h post–parasite invasion. Incubation with iRBC extracts resulted in increased viral production as compared to exposure to RBC extracts (Figure 2C). The above results suggest that iRBC-derived molecules can induce viral production, although the role played by CIDR1α in this context is difficult to assess.
To investigate whether the increased number of EBV genomes resulted from lytic cycle reactivation, we used an Akata cell line–based system in which induction of the lytic cycle is accompanied by increased expression of green fluorescent protein (GFP). Upon BCR cross-linking with anti-Ig, 20%–50% of cells typically enter the replicative cycle. Cells that support lytic replication up-regulate GFP (10- to 100-fold), whereas cells that remain in latency express little GFP. The kinetics of GFP expression follows that of early lytic genes. The GFP expression persists throughout the lytic cycle [22], and is therefore used as a quick and simple read-out for viral replication.
Akata-GFP cells were incubated with increasing concentrations of CIDR1α and GST for 48 h, at which time the extent of cells expressing GFP was measured by flow cytometry. Cells induced to lytic replication, either with CIDR1α or anti-Ig stimulation, and therefore expressing high levels of GFP, had a higher granularity (side scatter) and a minor increase in size (forward scatter) as compared to the GFP-negative population. These cellular changes reflect an activated status. Analysis of the data, gating on the region containing highly granular cells, showed that the increase in GFP-positive cells was 1,4 to 2,4 times higher among the CIDR1α-stimulated cells than in the GST control cultures. This is accounted by 12%–19% of cells in replication when cultured with 1 μM CIDR1α versus 5%–9% for cells cultured with GST (Figure 3A).
The increased proportion of cells in lytic cycle (GFP-positive) induced by CIDR1α treatment correlated with a rise in the number of viral genomes as measured by quantitative PCR. The rise in the proportion of GFP-positive cells after CIDR1α stimulation was directly proportional to the increase of EBV load (R2 = 0.90; p < 0,001) (Figure 3B).
Activation of lytic production by CIDR1α was further confirmed by analysis of the BZLF1 protein (also known as Z, Zebra, Zta, EB1) expression. BZLF1, an immediate early protein which acts as a transcriptional activator and disrupts EBV latency, is essential for full expression of lytic genes and viral DNA replication [23]. The kinetics of BZLF1 expression vary according to the infected cell type and the replication-inducing agent used. As early as 6 h after induction by surface Ig cross-linking, Akata cells start expressing BZLF1 and continue doing so throughout the lytic cycle [22]. Stimulation with CIDR1α led to an increased expression of BZLF1 in a CIDR1α dose-dependent manner as revealed by Western blot analysis 48 h after stimulation (Figure 3C). These results support the hypothesis that CIDR1α activation induces viral reactivation in EBV-carrying B cells. Taken together, the data indicate that the increased viral load induced by CIDR1α in Akata cells results from a switch in EBV from latency to the replicative lytic cycle.
B cells derived from the mucosal lymphoid tissue of the Waldeyer ring (tonsils) have a ten times higher frequency of virus-infected cells as compared to peripheral blood B cells [24]. To investigate the possible physiological impact that malaria infection has on EBV persistence, we analyzed the effect of CIDR1α on PBMCs and B cells derived from healthy EBV-positive donors with regard to quantitative changes in EBV DNA. PBMCs were co-incubated with CIDR1α, and after 48 h, the EBV genome copy number was quantified by real-time PCR and compared to the one in the GST/medium control cultures. This analysis was extended to PBMCs derived from children with BL living in malaria-endemic area.
Stimulation of tonsil B cells with CIDR1α increased by 3-fold the number of EBV genomes as compared to the control antigen (p = 0,04) (Figure 4A). In PBMCs, where the B cell number and the frequency of EBV-positive cells is lower, stimulation with CIDR1α led to a 1,4- and 2,5-fold increase in EBV genomes (Figure 4B). This increase was statistically significant (pDonor1 = 0,04; pDonor2 = 0,001). The two healthy donors used for this study have an elevated number of EBV-carrying B cells as reflected by a high frequency of spontaneous outgrowth of EBV-positive B cells when their PBMCs are cultured in vitro. The absolute levels of EBV genomes in PBMCs are lower than the ones in tonsil B cells, a fact that may be the result of a lower frequency of EBV-carrying B cells in PBMCs.
Although EBV infection is usually harmless, EBV is linked to several human cancers and it is recognized as a cofactor in the development of the endemic form of BL, in which virtually all tumour cells are EBV-positive. We have previously shown that the overall EBV DNA load is elevated in serum from children with BL [14]; therefore, it can be assumed that PBMCs from BL patients have a high frequency of EBV-positive B cells. Thus, we tested the effect of CIDR1α on PBMCs derived from seven Ugandan patients with BL. In four out of seven patients, stimulation with CIDR1α (2 μM) for 48 h resulted in an increased median viral load as compared to that of control cultures. The difference was statistically significant (p ≤ 0,03) (Figure 4C). Although the overall median value was comparable in CIDR1α- and GST-treated PBMCs in two of the three other patients, there were few replicate wells with higher EBV load compared to respective GST control (for both patients). Consequently, we conclude that CIDR1α stimulation increases the viral load in EBV-carrying B cells derived from EBV-positive healthy donors and in patients with BL.
Despite the well-established link between malaria and EBV infection with BL, little is known about the interaction between the two pathogens and the mechanisms responsible for the elevated EBV load observed in children living in malaria-endemic areas [12,25]. In this report, we identify, to our knowledge for the first time, a microbial protein from P. falciparum, that can drive a latently infected B cell into viral replication. Previous studies by Minoura-Etoh et al. demonstrated that monochloramine (NH2Cl), a Helicobacter pylori–associated oxidant, induces viral production in epithelial cells [26].
Following primary infection, EBV is kept latent within memory B cells. In vivo, terminal differentiation of B cells into plasma cells triggers the switch from latency into the lytic replicative cycle [8]. Signalling of BCR is initiated upon binding of the antigen to the membrane-bound Igs, and this activation contributes to cell differentiation and Ab production. The EBV reactivation in Akata cells that results from BCR cross-linking with anti-Ig may reflect physiological mechanisms that operate as latently infected memory B cells proceed through the germinal center reaction and/or undergo plasma cell differentiation. Given the fact that both CIDR1α and iRBCs bind non-immune Igs [15,27], it could be assumed that the induction of virus replication in Akata cells mediated by CIDR1α and by iRBC extracts may involve similar signalling pathways. We cannot ascertain that CIDR1α is the only molecule responsible for the viral production induced by iRBC crude extracts, because in addition to CIDR1α, iRBCs express other malarial antigens that could play a role in viral reactivation.
Children living in malaria-endemic areas have an elevated EBV load that is cleared after anti-malaria therapy [14] and is directly related to the level of endemicity [12]. The rise in viral load may partly be a consequence of the interaction between malarial antigens able to bind Igs and latently infected B lymphocytes. The interaction between B cells and iRBCs is of particular relevance; we have reported that non-immune B cells bind iRBCs, and that this interaction is partially mediated by CIDR1α on the iRBC membrane and Igs on the B cell surface [15]. Here, we show that the iRBCs and the malaria parasite protein CIDR1α bind to the EBV-positive B cell line Akata. Most likely, the Ig-binding capacity of CIDR1α [15] and potentially other P. falciparum antigens present on iRBCs could lead to viral replication in a manner similar to the way anti-Ig induces viral production in Akata cells. This assumption is based on the fact that CIDR1α induces B cell activation and Ab production, and preferentially activates memory B cells, where the virus persists [15,16]. The role of malaria infection in enhancing viral replication is further supported by our previous observation that acute malaria patients have high levels of Abs against the early viral lytic protein BZLF1 [14].
As a corollary of the in vitro experiments, it was important to analyze the impact of CIDR1α on PBMCs and tonsil B cells derived from healthy EBV-positive donors, and from patients with BL. As judged by quantification of EBV DNA, CIDR1α increased viral production. In healthy donors, the effect was more pronounced in tonsil B cells than in PBMCs. We believe that this observation reflects inherent differences in the frequency of memory B cells, as EBV-positive B cells are more frequent among tonsil B cells than in PBMC preparations [24].
EBV reactivation is suspected to play a role in BL lymphomagenesis and is indicated by the presence of high Ab titers to viral lytic components, which characteristically herald the onset of BL [11]. It has been previously suggested that BL may rise as a consequence of the combination of multiple mechanisms that involve chronic stimulation of the B cell compartment, increased viral production, and suppressed EBV-specific responses (reviewed by [28,29]). Our previous and present studies indicate that malaria has inherent effects on EBV–host balance that contribute to the abnormal viral load present in children from malaria-endemic areas and likely result in an increased risk for BL. In addition, the polyclonal B cell–activating capacity of malaria may increase the proliferation of EBV-positive cells [15]. The present data demonstrate that iRBCs and antigens related to the infectious cycle of malaria (such as CIDR1α) can trigger lytic production in EBV-carrying B cells. In vivo, the viral reactivation mediated by CIDR1α and other malaria-derived antigens might take place in the spleen, where the probability of trapped iRBCs to encounter EBV-carrying B cells is high, as over 40% of the splenocytes are B cells. Notably, splenomegaly is often observed in children living in highly endemic areas where BL occurs [30].
In conclusion, interactions between P. falciparum malaria and EBV could play a direct role in promoting the emergence of BL.
Akata, a human BL-derived cell line carrying EBV, was maintained in RPMI 1640 (GIBCO-BRL, http://www.invitrogen.com) supplemented with 10% fetal bovine serum, 100 U/mL penicillin, and 2 mM glutamine, hereafter referred to as complete medium. The Akata GFP was previously generated by infecting the EBV-negative variant of Akata line with a recombinant EBV virus, carrying a cassette that encodes for both neomycin resistance gene under control of a thymidine kinase promoter and a modified GFP gene under control of the cytomegalovirus immediate early promoter [31]. With the recombinant EBV, the neomycin resistance gene and the modified GFP gene are inserted into the open reading frame of the non-essential EBV thymidine kinase gene (BXLF1) [31]. Akata cells were maintained in complete medium, and Akata-GFP cells in complete medium supplemented with 500 μg/ml geneticin (G418; Sigma-Aldrich, http://www.sigmaaldrich.com).
BL patients were recruited from the Ugandan Cancer Institute, Mulago Hill, Kampala, Uganda. Blood was collected from children with BL (range 4–7 y; mean age 5 y). BL was diagnosed on the basis of clinical symptoms, presence of a tumour mass, and histological analysis. The study protocol was approved by the higher degrees of the university, the research committee and ethical committee of Makerere University Faculty of Medicine, the Uganda National Council of Science and Technology, and the Karolinska Institutet Research Ethical Committee. Written, informed consent was obtained from guardians of study participants.
Blood samples collected from two EBV-positive healthy volunteers (15–20 mL) and from BL patients (2–5 mL) before chemotherapy were processed the same day. PBMCs were isolated by centrifugation over Ficoll Plaque Plus (Amersham, http://www.amersham.com), washed in PBS, and resuspended in complete medium. Tonsils were obtained from patients undergoing routine tonsillectomy at Karolinska University Hospital in Stockholm, Sweden. Lymphocyte cell suspensions were prepared by mincing the tonsils and suspending the cells in complete medium. Isolated mononuclear cells were depleted of T cells by two rounds of rosette formation with ethyl isothiouronium bromide–treated sheep RBCs on ice. Rosettes were removed by centrifugation over lymphoprep (Nycomed Pharma, http://www.nycomed.com) [32]. The tonsil B cell purity was >95% as revealed by cytofluorimetric (FACS) analysis after staining with the pan–T cell Ab CD3 and the monocyte marker CD14.
The highly rosetting and auto-agglutinating P. falciparum parasite clone FCR3S1.2 was obtained through cloning by micromanipulation of the mother clone FCR3 [33] and cultured according to standard methods. Parasites were maintained and expanded in RBCs (O Rh+) at 5% haematocrit. iRBCs were cultured in RPMI 1640 medium supplemented with 10% B+ human serum, 0,6% HEPES (GIBCO-BRL), 25 μg/mL gentamycin (GIBCO-BRL), 0,25% sodium bicarbonate (GIBCO-BRL), and 2 mM L-glutamine (GIBCO-BRL). Trophozoites were enriched by magnetic-assisted cell sorting, using a Vario-MACS (Miltenyi Biotec, http://www.miltenyibiotec.com). Synchronous P. falciparum parasite cultures, grown 24–28 h post-invasion, were washed twice in RPMI 1640 and resuspended in 5–10 mL of phosphate-buffered saline (PBS) with 2% bovine serum albumin (BSA). Rosettes were disrupted mechanically by repeated passage through a 0,6-mm-thick injection needle. The material was slowly added to a MACS separation column mounted in the magnet and rinsed with 50 mL of 2% BSA in PBS. The iRBCs were eluted in 50 mL of 2% BSA in PBS after removing the column from the magnet, spun down at 500g, and resuspended in 1 mL of RPMI 1640. Following enrichment, the parasitemia was evaluated by fluorescence microscopy after addition of one drop of acridine orange (10 μg/mL).
Parasite extracts were prepared as previously described [34]. Briefly, parasites at trophozoite stage (iRBCs, 28 h post-invasion) or RBCs were washed, resuspended in PBS, and sonicated (25w) on ice at short intervals for 2 min. The extracts were then centrifuged at 500g for 10 min at 4 °C and filter-sterilized. After determination of protein concentration by Bradford Assay (Bio-Rad, http://www.bio-rad.com), the extracts were diluted with PBS.
The sequence of the CIDR1α domain of the PfEMP1 from clone FCR3S1.2-var1 was optimized for codon adaptation in Escherichia coli, re-synthesized chemically (GeneArt, http://www.geneart.com), cloned into the pGEX4T-1 vector (Amersham Pharmacia, http://www.gelifesciences.com), and expressed in E. coli BL21 CodonPlus-RIL (Stratagene, http://www.stratagene.com) as fusion protein with the GST. Non-recombinant pGEX4T-1 was used to produce GST as control protein. Protein purification was carried out according to an optimized protocol [35], and the recombinant protein was purified on glutathione-sepharose column GSTrap FF (Amersham Pharmacia) according to the manufacturer's instructions. Throughout the paper, we refer to the recombinant CIDR1α-GST fusion protein as CIDR1α, while GST is referred to as control protein.
Akata iRBCs: Akata cells and enriched iRBCs (trophozoite stage) were stained with PKH67 (green) and PKH26 (red) (Sigma-Aldrich), respectively, according to manufacturer's instructions. The binding was evaluated by fluorescence microscopy after co-incubation in PBS at room temperature (RT) for 1 h at a 1:2 ratio (Akata:iRBC).
Akata CIDR1α: Cells incubated for 1 h at RT in PBS containing GST or CIDR1α (1 μM) were washed twice in PBS and incubated for 30 min at RT with anti-GST Abs (Sigma-Aldrich) diluted 1:500 in PBS. After two washes with PBS, anti-mouse IgG Alexa-488-conjugated Ab (Molecular Probes, http://probes.invitrogen.com) diluted 1:100 in PBS was added for 30 min at 4 °C. Cell binding was analyzed by fluorescence microscopy, and the fluorescence intensity was measured with a FACSCalibur flow cytometer and analyzed with Cell Quest Pro software (Becton Dickinson, http://www.bd.com).
Twenty-four hours before experiments, Akata and Akata-GFP cells were suspended at a concentration of 106/mL in complete medium and complete medium containing G418 (500 μg/mL), respectively. Fresh PBMCs were washed with PBS and resuspended in complete medium at a concentration of 105 cells per ml. The cells were then seeded in 96-well plates and cultured in medium alone or with increasing concentrations of CIDR1α, GST (range 0–4 μM, corresponding to 0–200 μg/mL), or with anti-Ig (10 μg/mL) (Jackson ImmunoResearch, http://www.jacksonimmuno.com). After 48 h, cells were harvested for analysis. All tests were set up in multiple replicates.
DNA was extracted from cells and supernatants using the QIAamp Blood kit (Qiagen, http://www.qiagen.com) according to manufacturer's instructions and eluted in 50 μL of DEPC-treated water (Ambion, http://www.ambion.com). Purity and DNA concentration were evaluated using a NanoDrop ND-1000 spectrophotometer (http://www.nanodrop.com). The PCR primers and probe used for the quantification of EBV genomes were selected from the LMP-1 gene as previously described [13,14]. The primers used were the EBV-LMP1 forward primer 5′-AAGGTCAAAGAACAAGGCCAAG-3′ and the EBV-LMP1 reverse primer 5′-GCATCGGAGTCGG-3′. The fluorogenic probe (PE Applied Biosystems, http://www.appliedbiosystems.com) was synthesized using a FAM reporter molecule attached to the 5′ end and a TAMRA quench-er linked to the 3′ end (5′-AGGAGCGTGTCCCCGTGGAGG-3′). A standard curve was prepared using serial dilutions of DNA derived from the EBV-positive BL line Namalwa that contains two copies of EBV genome per cell. Detection was performed using an ABI Prism 7700 Sequence detection System (PE Applied Biosystems). Briefly, cycling parameters were 50 °C for 2 min, 95 °C for 10 min, 45 cycles 95 °C for 15 s, and 60 °C for 1 min. The EBV DNA copy number was calculated as mean of triplicates.
Cells were washed twice in ice-cold PBS and resuspended in 100 μL of PBS containing 50 μg propidium iodide/mL and 0.1% (v/v) triton X-100. After incubation at 4 °C for 8 h, cell DNA analysis was performed by flow cytometry (FACSCalibur, Becton Dickinson) using Cell Quest Pro software (Becton Dickinson). p-Values ≤ 0,05 were regarded as statistically significant.
Expression of BZLF1, the early lytic antigen, was assessed using a specific mouse Ab (DAKO, http://www.dako.com). Total extracts from 106 cells were separated in 10% SDS polyacrylamide gels, blotted onto nitrocellulose filters probed with 1:500 dilution of the DAKO monoclonal Ab for 1 h at RT, and then reacted with horseradish-peroxidase–labelled anti-mouse Abs (Amersham). Immunocomplexes were visualized by enhanced chemiluminescence according to the manufacturer's instructions (Amersham). To control for equal loading, the total protein concentration of each sample was checked using Bradford assay.
Computations were performed with the Prism 4 package (GraphPad Software, http://www.graphpad.com). For parametric data, differences between groups were analyzed with Student t-test; for nonparametric data, differences between groups were analyzed with the Wilcoxon rank sum test. Results are expressed as mean ± SEM, and unless otherwise stated, considered statistically significant at p < 0,05. |
10.1371/journal.ppat.1004048 | IFITM3 Restricts Influenza A Virus Entry by Blocking the Formation of Fusion Pores following Virus-Endosome Hemifusion | Interferon-induced transmembrane proteins (IFITMs) inhibit infection of diverse enveloped viruses, including the influenza A virus (IAV) which is thought to enter from late endosomes. Recent evidence suggests that IFITMs block virus hemifusion (lipid mixing in the absence of viral content release) by altering the properties of cell membranes. Consistent with this mechanism, excess cholesterol in late endosomes of IFITM-expressing cells has been reported to inhibit IAV entry. Here, we examined IAV restriction by IFITM3 protein using direct virus-cell fusion assay and single virus imaging in live cells. IFITM3 over-expression did not inhibit lipid mixing, but abrogated the release of viral content into the cytoplasm. Although late endosomes of IFITM3-expressing cells accumulated cholesterol, other interventions leading to aberrantly high levels of this lipid did not inhibit virus fusion. These results imply that excess cholesterol in late endosomes is not the mechanism by which IFITM3 inhibits the transition from hemifusion to full fusion. The IFITM3's ability to block fusion pore formation at a post-hemifusion stage shows that this protein stabilizes the cytoplasmic leaflet of endosomal membranes without adversely affecting the lumenal leaflet. We propose that IFITM3 interferes with pore formation either directly, through partitioning into the cytoplasmic leaflet of a hemifusion intermediate, or indirectly, by modulating the lipid/protein composition of this leaflet. Alternatively, IFITM3 may redirect IAV fusion to a non-productive pathway, perhaps by promoting fusion with intralumenal vesicles within multivesicular bodies/late endosomes.
| Interferon-induced transmembrane proteins (IFITMs) block infection of many enveloped viruses, including the influenza A virus (IAV) that enters from late endosomes. IFITMs are thought to prevent virus hemifusion (merger of contacting leaflets without formation of a fusion pore) by altering the properties of cell membranes. Here we performed single IAV imaging and found that IFITM3 did not interfere with hemifusion, but prevented complete fusion. Also, contrary to a current view that excess cholesterol in late endosomes of IFITM3-expressing cells inhibits IAV entry, we show that cholesterol-laden endosomes are permissive for virus fusion. The ability of IFITM3 to block the formation of fusion pores implies that this protein stabilizes the cytoplasmic leaflet of endosomal membranes, either directly or indirectly, through altering its physical properties. IFITM3 may also redirect IAV to a non-productive pathway by promoting fusion with intralumenal vesicles of late endosomes instead of their limiting membrane.
| The recently identified interferon-induced transmembrane proteins (IFITMs) inhibit infection of diverse enveloped viruses [1]–[3]. Ectopic expression of IFITM1, -2 and -3 restricts a growing number of unrelated viruses, including IAV [1], [2], [4]–[7]. IFITM3 has been shown to potently restrict infection by IAV and the Respiratory Syncytial Virus in vivo [8]–[10]. In contrast, arenaviruses and some retroviruses, such as murine leukemia virus (MLV), are resistant to IFITM restriction [2], [6]. The IFITMs have been reported to inhibit HIV-1 entry, albeit less potently than IAV and apparently in a cell type-dependent manner [11]–[13].
The mechanism by which IFITMs inhibit infection of diverse viruses is not fully understood. IFITM2 and -3 are predominantly found in late endosomes (LE) and lysosomes [13], [14], whereas IFITM1 is also found at the cell periphery [4], [15]. Different membrane topologies of IFITMs have been proposed [16], but recent data suggests that IFITM3 is a type II transmembrane protein [17]. Accumulating evidence implies that IFITMs may interfere with virus-endosome fusion [1], [2], [5], [13], [14]. The fact that IFITMs seem to expand acidic intracellular compartments [13] indicates that the fusion block is downstream of the low pH trigger. Effective restriction of viruses that enter from the LE, such as IAV, Ebola virus (EBOV) and SARS coronavirus seems consistent with the cellular localization of IFITM2 and -3 proteins. However, these proteins also restrict Vesicular Stomatitis Virus (VSV) that appears to fuse with early endosomes [18].
IFITMs have been reported to curtail viral infection by modifying properties of cellular membranes, such as fluidity and spontaneous curvature [3], [5], [14]. These effects could be related, in part, to the accumulation of cholesterol in LE as a result of IFITM-mediated disruption of the interaction between the vesicle-membrane-protein-associated protein A (VAPA) and oxysterol-binding protein (OSBP) [14]. Since lipids play an important role in membrane fusion, these findings offer an attractive paradigm for a broad antiviral defense mechanism that involves altering the lipid composition of cellular membranes. The recent finding that amphotericin B, which forms complexes with sterols [19], rescues IAV infection in IFITM2- and IFITM3-expressing cells [20] is in line with the notion that cholesterol may be directly or indirectly involved in IAV restriction. However, lipid composition-based models do not readily explain the lack of restriction of amphotropic MLV and arenaviruses, which enter cells via distinct endocytic routes [21], [22]. These findings indicate that IFITMs may restrict virus entry from a subset of intracellular compartments. In order to define the mechanism of IFITM restriction, it is important to identify the viral entry step(s) targeted by these proteins, define compartments in which restriction occurs, and elucidate potential changes in intracellular membranes that may be responsible for this phenotype.
Here, we examined the mechanism of IFITM3 restriction of IAV using single particle imaging and a direct virus-cell fusion assay. Our results show that IFITM3 does not inhibit the lipid mixing stage of IAV fusion but blocks the release of viral contents into the cytosol, and that this phenotype does not correlate with cholesterol accumulation in intracellular compartments. Specifically, IFITM3 inhibits the conversion of hemifusion to fusion through a mechanism that does not rely on cholesterol accumulation. Together these findings reveal a previously unappreciated view of IFITM-mediated restriction and suggest new avenues of investigation to delineate the mechanism by which these proteins block infection.
We chose to focus on IFITM3 to study the mechanism of IAV restriction because this protein potently inhibits infection in vitro and in vivo [8]–[10]. Since published data suggest that IFITM3 likely inhibits the viral fusion step, a direct virus-cell fusion assay was employed to evaluate the extent of restriction in different cell lines [23]. HIV-1 particles carrying the β-lactamase-Vpr (BlaM-Vpr) chimera and pseudotyped with the influenza HA and NA proteins from the H1N1 A/WSN/33 strain (referred to as IAVpp) were allowed to fuse with cells transduced with an empty vector or with an IFITM3-expressing vector. The resulting cytosolic BlaM activity was measured as previously described [24]. Out of several cell lines tested, A549 and MDCK cells over-expressing IFITM3 were least permissive to IAVpp fusion (Fig. 1A). In agreement with the previous reports [2], [13], we found that IFITM3 over-expression partially inhibited VSV G glycoprotein-mediated fusion of pseudoviruses (VSVpp) carrying the BlaM-Vpr chimera (Fig. 1A). Similar to inhibition of IAVpp fusion, the IFITM3-mediated restriction of VSVpp was most potent in A549 and MDCK cells. As expected, fusion of particles pseudotyped with the Lassa fever virus glycoprotein (LASVpp), which directs virus entry through an IFITM3-resistant pathway [2], [6], was not considerably affected by IFITM3 over-expression.
We next checked if the strong suppression of virus fusion in A549 and MDCK cells was related to the level of IFITM3 expression. Immunostaining for IFITM3 in these and CHO cells which exhibited modest restriction of viral fusion (Fig. 1A) did not reveal a clear correlation between IFITM3 expression and inhibition of IAVpp or VSVpp fusion (Fig. 1B). Of note, potent IAV restriction in A549 and MDCK cells was not related to the usage of HIV-1 core-based pseudoviruses. Influenza virus-like particles containing the IAV BlaM-M1 chimera [25] also failed to efficiently fuse with A549-IFITM3 and MDCK-IFITM3 cells while fusing well with vector-transduced cells (Fig. 1C). We also found that both vector-transduced A549 and MDCK cells were highly susceptible to IAV infection, as determined by virus titration (see Materials and Methods). These two cell lines were therefore chosen for studies of IFITM3-mediated restriction described below.
IFITM-based restriction has been studied using a cell-cell fusion model, as well as by forcing viral fusion with the plasma membrane by lowering the pH [5], [20]. Since fusion with the plasma membrane is more amenable to mechanistic studies than endocytic entry, we asked whether IFITM3 can restrict forced IAV fusion. Exposure to acidic buffer induced IAVpp fusion with A549-Vector cells pretreated with Bafilomycin A1 (BafA1), which blocked low pH-dependent entry from endosomes (Fig. 1D). The extent of forced fusion was lower compared to the conventional entry route. By contrast, forced IAVpp fusion with A549-IFITM3 cells was ∼3-fold more efficient than endocytic fusion with cells not treated with low pH or BafA1, showing that IFITM3 does not restrict IAVpp fusion at the cell surface. Interestingly, IFITM1 suppressed IAVpp-plasma membrane fusion at low pH (Fig. 1D), in agreement with the Jaagsiekte sheep retrovirus (JSRV) and IAV fusion data [5], [20]. The inability of IFITM3 to block IAV fusion with the plasma membrane is consistent with its lower abundance at the cell surface [13], [14], [20] and shows that the mechanism of restriction must be studied in intracellular compartments.
Preponderance of evidence implies that hemifusion is a universal intermediate (reviewed in [26], [27]) that precedes the formation of a fusion pore. Having shown that IFITM3 over-expression inhibits viral fusion (Fig. 1A, C), we asked whether this protein also blocks the upstream hemifusion step. This was accomplished by labeling the A/PR/8/34 virus membrane with a self-quenching concentration of vybrant DiD (vDiD), using a modification of the previously published protocol [28]. Incorporation of self-quenching quantities of a lipophilic dye enables the visualization of single lipid mixing events based on the marked increase in fluorescence upon dye redistribution to an endosomal membrane (see for example [28], [29]).
Significantly, to control for fluctuations in the vDiD fluorescence caused by deviation from a focal plane, the viral surface proteins were labeled with the amine-reactive AlexaFluor-488 (AF488) dye. The relatively steady AF488 signal before and after hemifusion is allowed correcting for the vDiD intensity fluctuations due to moving in and out of focus. The vDiD/AF488 co-labeling protocol only modestly (<2-fold) reduced IAV infectivity compared to the mock-labeled viruses (Fig. S1A). Immunofluorescence staining of AF488-labeled virions with anti-HA antibodies revealed an excellent co-localization of the two signals (Fig. S1B, C), thus supporting the notion that AF488/vDiD-labeled particles are bona fide virions.
Labeled viruses were allowed to enter A549-Vector cells, and the resulting lipid mixing activity was examined by single particle tracking. A fraction of virions exhibited a marked increase in the vDiD signal (Fig. 2A, B). Redistribution of vDiD was mediated by low pH-dependent conformational changes in the IAV HA glycoprotein, as evidenced by potent inhibition of lipid mixing by anti-HA antibodies (Fig. 2C) and by NH4Cl (Fig. 3A). Without simultaneous monitoring of the viral content release into the cytoplasm, vDiD dequenching does not discriminate between hemifusion (operationally defined as lipid mixing without content transfer [30]) and full fusion. To avoid over-interpreting dequenching events, we will refer to these events as lipid mixing or hemifusion. A similar vDiD dequenching pattern was observed in MDCK cells transduced with an empty vector (data not shown). Analysis of lipid mixing showed that 2.2±0.4% and 5.6±0.6% of cell-bound particles released vDiD in A549 and MDCK cells, respectively (Fig. 3A). By comparison, a much greater fraction of virions (38.3±0.6%) hemifused with CHO cells (data not shown), in agreement with the previously reported data [28].
Importantly, IAV lipid mixing was readily detected in IFITM3+ A549 and MDCK cells (Figs. 2D–G and 3A). Not only was lipid mixing not inhibited in A549-IFITM3 cells, but a >3-fold greater fraction of particles released vDiD in these cells compared to control cells (Fig. 3A, P<0.001). By comparison, IFITM3 over-expression in MDCK cells did not significantly promote vDiD dequenching (Fig. 3A). Thus, contrary to the cell-cell fusion results [5], IFITM3 does not inhibit and can even promote IAV lipid mixing, consistent with the block of virus entry at a post-hemifusion stage. Accordingly, the addition of oleic acid, which augments hemifusion by altering spontaneous membrane curvature, did not rescue IAVpp or VSVpp fusion with A549-IFITM3 cells (Fig. S2). This is in agreement with the recent infectivity results [20], but in contrast with the rescue of fusion between JSRV Env- and IFITM-expressing cells by this fatty acid [5].
The higher frequency of vDiD dequenching in A549-IFITM3 cells could be caused by the increased endosome acidity compared to control cells [13]. However, the distribution of waiting times to the onset of lipid mixing was independent of IFITM3 expression or the type of target cells (A549 vs. MDCK, Fig. 3B, P = 0.37). The fact that the kinetic curves do not reach plateau indicates that IAV entry into A549 and MDCK cells is not completed within the first hour. Our results thus demonstrate that IFITM3 restricts the IAV fusion at a post-hemifusion step, most likely at the point of fusion pore opening, as evidenced by the dramatic decrease of the BlaM signal in A549 and MDCK cells expressing this protein (Fig. 1A).
Under our conditions, vDiD dequenching was typically completed within a few minutes for both control and IFITM3+ cells (Fig. 2). This dequenching rate is much slower than sudden increases in fluorescence of the IAV membrane markers described previously [28], [31]. While a portion of vDiD dequenching could be completed within seconds (Fig. S3), these fast events were not common. Slow dequenching was also typical with the vDiD/AF488-labeled X31 virus, as well as with the X31 virus labeled with a 15-fold excess of DiD, using the published protocol for single virus imaging [28] (data not shown).
Slow vDiD dequenching during the first hour of virus-cell co-incubation did not appear to result from IAV degradation in LE/lysosomes, since the surface-exposed AF488 label persisted long after vDiD dequenching was completed and because anti-HA antibodies blocked vDiD dequenching (Fig. 2). In addition, we did not detect any correlation between the lag before the onset of lipid mixing and the vDiD dequenching slope (Fig. S4A). This result reinforces the notion that late lipid mixing events are mediated by HA and not by virus degradation. Control experiments, in which samples were not exposed to laser light during the first 30 min at 37°C, did not reveal fast dequenching events reaching completion in less than 1 min (data not shown). This control argues against phototoxicity-related attenuation of virus fusogenicity as the cause for sluggish lipid redistribution.
Since free vDiD diffusion between a virus and a small endosome should be completed in less than a second [32], [33], an initial membrane connection between IAV and an endosome must severely impair lipid movement. To assess whether early fusion intermediates in control and IFITM3+ cells restrict vDiD diffusion to the same extent, we examined the rate of vDiD dequenching. Single particle analysis revealed that, in A549 cells, the average vDiD dequenching profile (Fig. 3C) was independent of IFITM3 expression, as were the initial slopes of vDiD dequenching (Fig. S4B, P>0.5). These results indicate that IFITM3 over-expression does not affect the properties of fusion intermediates responsible for vDiD redistribution, such as the size and/or architecture of a hemifusion site (e.g., [34], [35]). We then asked whether the rate of vDiD dequenching varied depending on the cell type. The average rate of vDiD fluorescence increase in MDCK cells was ∼2-fold greater than in A549 cells (Figs. 3C and S4B, P<0.02). This demonstrates our ability to detect changes in the rate of vDiD transfer and shows that lipid transfer lasts several minutes irrespective of the cell type.
We also examined the final extent of vDiD dequenching, which is proportional to the surface area of a target membrane over which it redistributes. This parameter was not significantly affected by IFITM3 expression in A549 cells or by the cell type (MDCK vs. A549 cells, Fig. 3D). Together, similar kinetics and extents of viral lipid dilution in control and IFITM3+ cells suggest that neither the size/architecture of early fusion intermediates nor the surface area of target endosomes is considerably affected by IFITM3 expression.
To investigate the relationship between lipid mixing and productive IAV infection, we compared the fraction of cells “receiving” at least one vDiD dequenching event in live cell imaging experiments to the fraction of cells that got infected under the same conditions. The only difference was that virus imaging was not continued beyond 1 h after initiation of fusion, whereas infection proceeded overnight. We found that one or more vDiD dequenching events occurred in 15% of A549 cells while 44% of cells got infected (Fig. S5). Under the same conditions, 20% of MDCK cells “hosted” one or more dequenching events and 36% were infected. The greater fraction of infected cells compared to those permissive to hemifusion is likely due to the shorter time widow for single virus imaging, which is likely to miss late vDiD dequencing events (Fig. 3B). The lower apparent fraction of cells supporting vDiD dequenching could also be caused by the presence of viruses that did not incorporate self-quenching amounts of vDiD. Importantly, the comparable efficiencies of lipid mixing and infection, indicate that the former events likely culminate in productive infection.
To determine whether IFITM3 impairs the IAV's ability to form small fusion pores, we attempted to load the virus with a content marker by soaking in a concentrated solution of sulforhodamine B, as described in [36]. However, only a small fraction of AF488-labeled particles stained with sulforhodamine, and the retained dye was lost in live cell experiments under conditions that blocked IAV fusion (data not shown). We therefore resorted to using HIV pseudoviruses bearing A/WSN/33 HA and NA glycoproteins and co-labeled with the capsid marker, YFP-Vpr, and the content marker, Gag-iCherry [24], [37]. Upon virus maturation, the “internal” mCherry is proteolytically cleaved off the HIV-1 Gag-iCherry and released through a fusion pore, as manifested by the loss of the red signal (Fig. 4 and [37]). The YFP-Vpr signal, which remained associated with the viral core after fusion, provided a reference signal for single particle tracking.
Under our conditions ∼1% of double-labeled pseudoviruses entering A549-Vector cells lost their content marker, while approximately 2% fused with MDCK-Vector cells. In sharp contrast, the mCherry release in IFITM3+ A549 and MDCK cells or in vector-transduced cells in the presence of NH4Cl could not be detected (Fig. 4E, P<0.001). Thus, IFITM3 does not adversely affect IAV hemifusion but severely inhibits viral content release into the cytoplasm. Together these findings suggest that the mechanism of IFITM3-mediated restriction arises from the entrapment of viruses at a hemifusion intermediate prior to fusion pore formation.
A recent study has shown that, through disrupting the interaction between VAPA and OSBP, IFITM3 causes cholesterol accumulation in LE [14]. Based on this finding, the authors proposed that high levels of endosomal cholesterol may inhibit IAV fusion and/or the release of nucleocapsid. Staining with filipin revealed that IFITM3+ A549 cells exhibited increased levels of intracellular cholesterol (Fig. 5A). However, the filipin signal was still primarily associated with the plasma membrane and the total cellular cholesterol was not elevated in IFITM3+ cells (Fig. S6). In addition, the overall intensity of intracellular cholesterol poorly correlated with the level of IFITM3 expression (Fig. 5C).
By comparison, pretreatment of A549-Vector cells with U18666A, which inhibits transport of LDL-derived cholesterol from LE/lysosomes (reviewed in [38]), resulted in a dramatic shift in the filipin staining pattern from the plasma membrane to endosomes (Fig. 5B). Aberrant accumulation of cholesterol in LE is also known to occur in cells lacking the functional NPC1 cholesterol transporter [39]. We therefore knocked down NPC1 expression in A549 cells using shRNA (shNPC1, Fig. 5D) and examined the resulting cholesterol distribution (Fig. 5B). Reduced NPC1 expression correlated with excess cholesterol in intracellular compartments, which was also much more pronounced than endosomal filipin staining in A549-IFITM3 cells.
We next asked whether the cholesterol accumulation induced by U18666A pretreatment or by down regulation of NPC1 can phenocopy the IFITM3-mediated restriction of viral fusion. Neither IAV lipid mixing (vDiD dequenching) nor fusion (BlaM signal) was inhibited by silencing NPC1 in A549 cells (Fig. 5E, F). VSVpp also fused with shNPC1-transduced cells as efficiently as with control cells (Fig. 5E). These results show that excess cholesterol does not inhibit viral fusion or hemifusion. In control experiments, silencing the NPC1 expression potently suppressed fusion of Ebola GP-pseudotyped particles (EBOVpp, Fig. 5E), which use NPC1 as a receptor [40], [41]. Similar to the NPC1 knockdown phenotype, pretreatment of A549 cells with 10 µM U18666A, which caused cholesterol buildup in endosomes (Fig. 5B), did not inhibit fusion of IAVpp or VSVpp (Fig. 5G). As will be shown below for MDCK cells, higher doses of U18666A can inhibit viral fusion (Fig. 5G), but this effect is due to elevation of endosomal pH as opposed to cholesterol accumulation in endosomes.
To generalize the effects of excess cholesterol in A549 cells, we tested whether endosomal cholesterol can inhibit viral fusion in MDCK cells. As in A549 cells, IFITM3 over-expression in MDCK cells caused moderate accumulation of cholesterol in endosomes (Fig. 6A), while pre-treatment with U18666A caused a much more dramatic buildup of intracellular cholesterol (Fig. 6B). However, unlike A549 cells, IAVpp and VSVpp fusion was significantly inhibited in U18666A-treated MDCK cells (Fig. 6C). Since prolonged exposure to U18666A has been reported to raise endosomal pH [42], we sought to determine if insufficiently acidic pH could prevent IAV hemifusion/fusion with pretreated MDCK cells.
The pH in IAV-carrying endosomes was measured using virions co-labeled with the pH-insensitive AF488 (green) and CypHer5E (red), which fluoresces brighter at acidic pH [28] (Fig. S7A). Cells were incubated with viruses for 45 min, and the red/green signal ratio from individual particles was measured (Fig. S7B). The average pH in virus-containing endosomes of MDCK-IFITM3 cells was slightly less acidic than in control cells: 5.38±0.03 (n = 498) vs. 4.98±0.04 (n = 242), respectively (Fig. 6D and F, P<0.001). Interestingly, as shown in Figure 6E, endosomal pH in U18666A-treated MDCK cells was markedly shifted to neutral values (6.44±0.05, n = 160, P<0.001). Since the pH threshold for triggering A/PR/8/34 fusion is reported to be around 5.6 [43], elevation of endosomal pH in U18666A-treated MDCK cells is the likely cause of inhibition of viral fusion. Together our results imply that U18666A most likely attenuates IAV fusion with MDCK cells by raising endosomal pH and not through inducing cholesterol accumulation.
We also took advantage of the available CHO cell line that does not express NPC1 [44] to further ascertain the role of endosomal cholesterol in IAV fusion. These cells (designated CHO-NPC1−) exhibited exaggerated endosomal cholesterol staining, in sharp contrast to a peripheral staining pattern in parental CHO cells (Fig. 7A). In spite of the high endosomal cholesterol content in CHO-NPC1− cells and of the elevated level of total cholesterol (Fig. S6), IAVpp fused with these cells as efficiently as with parental cells (Fig. 7C). The NPC1-null cells also supported IAV lipid mixing, albeit at somewhat reduced level compared to control (Figs. 7D and S8). Pretreatment of CHO cells with U18666A also trapped cholesterol in endosomes and raised the total cholesterol content (Figs. 7B and S6), but only modestly diminished the extent of IAVpp or VSVpp fusion (Fig. 7E). Interestingly, in contrast to the decreased endosome acidity in MDCK cells, endosomes in U18666A-treated CHO cells were more acidic than in control cells (Fig. S9). In control experiments, both the lack of NPC1 expression and U18666A pretreatment blocked EBOVpp fusion (Fig. 7C, E), consistent with its reliance on NPC1 receptor and high sensitivity to disruptions of cholesterol transport [45].
Together, our results show that the cholesterol accumulation achieved through two different interventions – U18666A pretreatment and NPC1 silencing – does not phenocopy IFITM3-mediated restriction of viral fusion. This implies that (i) elevated levels of endosomal cholesterol do not generally confer resistance to viral fusion, and (ii) the mechanism by which IFITM3 blocks transition from hemifusion to full fusion is not through the mislocalization of cholesterol.
The IFITMs restrict the cellular entry of multiple pathogenic enveloped viruses. Recent studies lead to a model that IFITMs inhibit virus-host hemifusion [5] and that the membrane-rigidifying properties of cholesterol may contribute to antiviral actions [14]. In contrast to these studies, our results now demonstrate that IFITM3 prevents the release of viral genomes into the cytosol by inhibiting viral entry after hemifusion but prior to fusion pore formation (Fig. 8). Moreover, we found that IFITM3 can promote hemifusion in some cells, perhaps secondary to its acidifying the endosomal pathway. IFITM3 therefore does not negatively regulate the properties of contacting leaflets involved in hemifusion, but stabilizes the cytoplasmic leaflet of the endosomal membrane, thereby disfavoring the formation of fusion pores [35]. In one potential scenario IFITM3 is located directly at the site of arrested hemifusion, perhaps “toughening” the endosomal membrane to create a barrier to viral entry (Pathway 1). A considerable colocalization of IFITM3 with internalized IAV ([3] and Fig. S10) is consistent with Pathway 1's direct mechanism of inhibition. Alternatively, IFITM3 might arrest hemifusion through an indirect mechanism, perhaps involving modulation of lipid and/or protein composition of the cytoplasmic leaflet (Pathway 2). Recent findings that changes in global membrane properties interfere with productive entry would appear to support an indirect mechanism [5], [14].
Lipids, such as unsaturated fatty acids and cholesterol that confer negative spontaneous curvature to membranes can promote hemifusion (a net negative curvature structure) and disfavor a fusion pore (a net positive curvature intermediate), as has been previously shown for oleic acid [35]. Although this prediction is consistent with efficient lipid mixing in endosomes of IFITM3+ cells observed in our imaging experiments, several studies [20], [46]–[48] and our own results do not support cholesterol accumulation as playing a role in fusion inhibition. We found that cholesterol-laden endosomes in cells pretreated with U18666A or expressing undetectable/low levels of NPC1 supported efficient viral fusion. It is thus possible that IFITM3 interferes with cellular functions of VAPA other than the interaction with OSBP, such as regulation of SNAREs and modulation of lateral mobility of membrane proteins (reviewed in [49]).
IFITM3 appears to induce the formation multivesicular bodies and increase the number of ILVs [13], [14]. One can therefore envision that IFITM3 may inhibit infection by redirecting viruses to a non-productive pathway, perhaps involving fusion with ILVs instead of the limiting membrane of LE (Fig. 8, Pathway 3). If, as suggested in [14], IFITM3 disallows back fusion of ILVs with the limiting membrane, then virus-ILV fusion products will likely be degraded. Indeed, back fusion has been implicated in the VSV core release into the cytosol following the virus-ILV fusion [50]. It should be stressed that this “fusion decoy” model does not explain the ability of IFITM1 to interfere with fusion at the cell surface ([5] and Fig. 1D). It is also not clear why the Old World arenaviruses, which have been reported to enter from MVBs [51], are not restricted by IFITMs.
The indistinguishable extents of vDiD dequenching in control and IFITM3+ cells (Fig. 3D) indicate that target endosomes have similar sizes. While this appears to argue against redirection of IAV fusion to small ILVs, the lack of a post-hemifusion decay of vDiD fluorescence in A549 and MDCK cells (Figs. 2 and S3) is consistent with IAV fusion with abundant ILVs in endosomes of IFITM3+ cells. This is because a lipophilic dye in the limiting membrane of an endosome should be quickly removed through membrane trafficking [24], [31], [52]. Because post-dequenching decay was not observed irrespective of the level of IFITM3 expression, it is possible that IAV may infect several cell lines by fusing with small intralumenal vesicles followed by the nucleocapsid release through back fusion (Fig. 8, dashed black arrows). This pathway could explain the similar extents and rates of vDiD dequenching in control and IFITM3-expressing cells, which are indicative of similar lipid intermediates and of the size of a target membrane, respectively.
As discussed above, slow vDiD dequenching observed by single IAV imaging can be rationalized in the context of fusion with the limiting membrane of endosomes (Pathways 1 and 2), as well as in the context of fusion with ILVs (Pathway 3). Slow dilution of this dye in Pathway 3 could occur through multiple rounds of IAV fusion with small ILVs, whereas Pathways 1 and 2 would predict restricted lipid diffusion through early fusion intermediates formed at the limiting membrane. Although the latter notion is in agreement with the reported restriction of lipid movement through hemifusion sites and small fusion pores [34], [35], [53], [54], these intermediates are usually short-lived under physiological conditions and tend to resolve into larger structures that do not impair lipid movement [28], [32], [35]. Clearly, more detailed studies of virus-endosome hemifusion and fusion are needed to understand the nature of slow lipid redistribution between IAV and endosomes.
The IFITMs may now arguably be one of the most broadly acting and clinically relevant restriction factor families [1], [3]. While both IFITM3's membrane-associated topology and its localization to the site of viral attenuation suggest it acts to restrict viral entry via a direct mechanism, additional work remains to be done to fully elucidate its actions. Nonetheless, as the primary effector of IFN's anti-IAV actions, IFITM3 represents a previously unappreciated class of restriction factor that prevents viral entry by stabilizing a hemifusion intermediate, likely comprised of an invading virus fatally tethered to the interior of the endosome's limiting membrane. Future single virus experiments combining the detection of both viral lipid and content release events (see for example [52]) should provide further insights into IAV entry pathways and the mechanism of IFITM3-mediated restriction. Indeed, such efforts may also bring to light unknown viral countermeasures, which are perhaps employed by the IFITM-resistant New and Old World arenaviruses.
HEK 293T/17 cells and human lung epithelial A549 cells were obtained from ATCC (Manassas, VA) and grown as previously described [55]. Wild-type CHO cells and CHO-NPC1− cells, a gift from Dr. L. Liscum (Tufts University) [44], were grown in Alpha-MEM (Quality Biological Inc, Gaithersburg, MD) supplemented with 10% FBS and penicillin-streptomycin. The A549, MDCK, HeLaH1 and CHO cells stably expressing IFITM3 or IFITM1 were obtained by transducing with VSV-G-pseudotyped viruses encoding wild-type IFITM3 and IFITM1 or with the vector pQCXIP (Clontech) and selecting with puromycin, as described previously [2].
The pR8ΔEnv, BlaM-Vpr, pcRev, HIV-1 Gag-iCherryΔEnv and pMDG VSV G expression vectors were described previously [37], [55]. The YFP-Vpr was a gift from Dr. T. Hope (Northwestern University). The pCAGGS vectors encoding influenza H1N1 WSN HA and NA were provided by Donna Tscerne and Peter Palese, and the pCAGGS BlaM1 (WSN) plasmid was a gift from Dr. A. Garcia-Sastre (Mount Sinai). Vectors expressing phCMV-GPc Lassa and pcDNA3.1-Ebola GP (Zaire) were gifts from Dr. F.-L. Cosset (Université de Lyon, France) [56] and Dr. L. Rong (University of Illinois) [57], respectively.
U18666A was from Tocris Bioscience (Bristol, UK). Poly-L-lysine, filipin, sulphorhodamine B Bafilomycin A1 and the Cholesterol Kit were from Sigma-Aldrich. AlexaFluor-488 amine-reactive carboxylic acid, vybrant-DiD (vDiD, 1,1′-dioctadecyl-3,3,3′,3′-tetramethylindodicarbocyanine,4-chlorobenzenesulfonate salt), Hoechst-33342 and Live Cell Imaging buffer were purchased from Life Technologies (Grand Island, NY). CypHer5E Mono NHS Ester was from GE Healthcare (Pittsburgh, PA). Antibodies used were rabbit anti-IFITM3 (to N-terminus) from Abgent (San Diego, CA), mouse anti-IAV-NP and goat anti-IAV-polyclonal antibodies from Millipore (Billerica, MA), rat anti-mouse-IgG-FITC from eBioscience (San Diego, CA), and goat anti-rabbit-Cy5 from Jackson Immunoresearch (West Grove, PA).
Pseudovirus production and titration were described previously [58]. Pseudoviruses were produced by transfecting HEK293T/17 cells using JetPRIME transfection reagent (Polyplus-transfection SA, NY). For LASV and EBOV pseudoviruses, 5 µg of the phCMV-GPc Lassa or 5 µg of the pcDNA3.1-Ebola GP was included in the transfection mixture. Fluorescently labeled influenza pseudoviruses were produced using 1 µg of pR8ΔEnv, 2 µg of HIV-1 Gag-iCherryΔEnv [37], 2 µg of YFP-Vpr, 1 µg of pcRev, and 2 µg of each WSN HA- and NA-expressing vectors. Ebola GP pseudoviruses were concentrated 10×, using Lenti-X™ Concentrator (Clontech, Mountain View, CA). To generate influenza BlaM1 VLPs, HEK293T cells were transfected with pCAGGS-BlaM1 (5 µg) and 2.5 µg of each pCAGGS-WSN HA and pCAGGS-WSN NA. After 12 h, the transfection reagent was removed, and cells were further cultivated in phenol red-free growth medium.
The influenza virus surface proteins and the lipid membrane were labeled with AF488 and vDiD, respectively. A hundred µg of influenza virus from the purified H1N1 A/PR/8/34 stock (2 mg/ml, Charles River, CT) was diluted in 95 µl of sodium bicarbonate buffer (pH 9.0) supplemented with 50 µM AF488. The mixture was incubated for 30 min at room temperature, after which time, 5 µl of vDiD (from 1 mM stock in DMSO) was added followed by an additional incubation for 90 min in the dark at room temperature with mild agitation. The labeled viruses were purified through a NAP-5 gel filtration column (GE Healthcare) in 145 mM NaCl solution buffered with 50 mM HEPES, pH 7.4. Approximately 50% of AF488-labeled particles incorporated detectable amounts of vDiD with minimal contamination by free dye aggregates.
The infectious IAV titer was determined in MDCK or A549 cells after incubation with serially diluted inoculum for 15 h at 37°C. Cells were fixed, permeabilized, blocked and incubated with rabbit R2376 anti-WSN HA antibody (a gift from Dr. D. Steinhauer, Emory University) for 2 h at room temperature. Cells were then washed and incubated with secondary Cy5-conjugated goat anti-rabbit antibodies (Jackson ImmunoResearch, PA) in 10% FBS-containing buffer supplemented with 10 µg/ml Hoechst-33342 for 1 h. The number of infected cells per image field was determined by fluorescence microscopy and normalized to the total number of cells (stained nuclei). The infectious titer (IU/ml) was calculated by taking into account the ratio of the area of well and the image area and correcting for dilution and volume of viral inoculum.
The β-lactamase (BlaM) assay for virus-cell fusion was carried out as described previously ([24] and Methods S1). Briefly, pseudoviruses bearing β-lactamase-Vpr chimera (BlaM-Vpr) were bound to target cells by centrifugation at 4°C for 30 min at 1550×g. Unbound viruses were removed by washing, and fusion was initiated by shifting to 37°C for 90 min, after which time cells were placed on ice and loaded with the CCF4-AM substrate (Life Technologies). The cytoplasmic BlaM activity (ratio of blue to green fluorescence) was measured after an overnight incubation at 12°C, using the Synergy HT fluorescence microplate reader (Bio-Tek, Germany).
IAV was pre-bound to A549-IFITM3 cells in the cold, followed by incubation at 37°C for 90 min and immunostaining with mouse anti-IAV-NP (Millipore, Billerica, MA) (when applicable) and rabbit anti-IFITM3 antibody (N-terminus, Abgent, San Diego, CA), as described in [13]. Rat anti-mouse-IgG-FITC (eBioscience, San Diego, CA) and goat anti-rabbit-Cy5 antibodies were used for secondary staining. Cellular distribution of cholesterol was examined by incubation with 0.25 mg/ml filipin added during the incubation with secondary antibodies. Images were collected on a LSM 780 laser scanning microscope (Carl Zeiss, Germany) using a 63× oil immersion objective. All staining methods involved fixation with 2% paraformaldehyde, permeabilization with 0.25% Triton-X100, blocking in with 10% FBS and dilution in phosphate buffered saline (with calcium and magnesium), and sequential incubation with primary and secondary antibodies for 2 h and 1 h, respectively.
To silence the NPC1 gene, A549 cells were transduced with five shRNAs encoded by pLK0.1 lentiviral vector (Sigma) and selected with puromycin. The samples for Western blotting were processed as described in [24]. The NPC1 protein band was detected with rabbit anti-NPC1 (Abcam, Cambridge, MA) and horseradish peroxidase-conjugated Protein G (Bio-Rad, Hercules, CA), using a chemiluminescence reagent from GE Healthcare.
Cells grown on glass-bottom Petri dishes (MatTek, MA) were chilled on ice and washed with cold Hank's balanced salt solution (HBSS). Predetermined amount of viral suspension (MOI∼0.01) was added to the cells and spinoculated at 4°C for 20 min. The cells were then washed twice with cold HBSS and placed on the stage of an LSM 780 confocal microscope. Virus entry was initiated by adding 2.5 ml of pre-warmed imaging buffer and imaged at 37°C using a C-Apo 40×/1.2NA water-immersion objective. Three Z-stacks separated by ∼2 µm were acquired every 7–8 s through the MultiTime macro (Carl Zeiss). To block IAV hemifusion and fusion, experiments where performed in HBSS supplemented with 50 mM HEPES/70 mM NH4Cl (pH 7.6) or containing 200 nM of BafA1. The time lapse images were first visually inspected to identify vDiD dequenching or loss of mCherry events. The number of relevant events in each experiment was independently determined by two trained individuals. Particle trajectories and their mean/total fluorescence intensities were obtained using Volocity (PerkinElmer, MA). The onset of lipid mixing and the initial slope of vDiD dequenching were determined by fitting to a pair of straight lines (Fig. S11).
IAV particles were co-labeled with the AF488 dye (pH-insensitive) and CypHer5E, which fluoresces brighter at acidic pH. The ratios of the CypHer5E and AF488 signals were converted to pH values using a calibration curve obtained by exposing coverslip-immobilized viruses to citrate-phosphate buffers of different acidity (Fig. S7). Images were collected from 3 different fields, and sum of single-particle fluorescence was calculated. The mean ratios of CypHer5E to AF488 signals as a function of pH were used for the calibration curve. Cells were inoculated with labeled viruses for 45 min at 37°C, as described above. Images were collected from at least 10 different fields, and single particle-based ratio of fluorescence signals was calculated. Outliers with a near-background CypHer5E signal were rejected to reduce the uncertainty in pH measurements.
Statistical significance was assessed using the pairwise t-test or rank sum test. Single-particle fusion events in control and IFITM3 expressing cells were compared by the z-test.
|
10.1371/journal.ppat.1002469 | Allelic Variation on Murine Chromosome 11 Modifies Host Inflammatory Responses and Resistance to Bacillus anthracis | Anthrax is a potentially fatal disease resulting from infection with Bacillus anthracis. The outcome of infection is influenced by pathogen-encoded virulence factors such as lethal toxin (LT), as well as by genetic variation within the host. To identify host genes controlling susceptibility to anthrax, a library of congenic mice consisting of strains with homozygous chromosomal segments from the LT-responsive CAST/Ei strain introgressed on a LT-resistant C57BL/6 (B6) background was screened for response to LT. Three congenic strains containing CAST/Ei regions of chromosome 11 were identified that displayed a rapid inflammatory response to LT similar to, but more severe than that driven by a LT-responsive allele of the inflammasome constituent NRLP1B. Importantly, increased response to LT in congenic mice correlated with greater resistance to infection by the Sterne strain of B. anthracis. The genomic region controlling the inflammatory response to LT was mapped to 66.36–74.67 Mb on chromosome 11, a region that encodes the LT-responsive CAST/Ei allele of Nlrp1b. However, known downstream effects of NLRP1B activation, including macrophage pyroptosis, cytokine release, and leukocyte infiltration could not fully explain the response to LT or the resistance to B. anthracis Sterne in congenic mice. Further, the exacerbated response in congenic mice is inherited in a recessive manner while the Nlrp1b-mediated response to LT is dominant. Finally, congenic mice displayed increased responsiveness in a model of sepsis compared with B6 mice. In total, these data suggest that allelic variation of one or more chromosome 11 genes in addition to Nlrp1b controls the severity of host response to multiple inflammatory stimuli and contributes to resistance to B. anthracis Sterne. Expression quantitative trait locus analysis revealed 25 genes within this region as high priority candidates for contributing to the host response to LT.
| We show that genetic variation within an 8.3 Mb region on mouse chromosome 11 controls host response to anthrax lethal toxin (LT) and resistance to infection by the Sterne strain of Bacillus anthracis. Specifically, congenic C57BL/6 mice in which this region of chromosome 11 is derived from a genetically divergent CAST/Ei strain presented with a rapid and strong innate immune response to LT and displayed increased survival following infection with Sterne spores. CAST/Ei chromosome 11 encodes a dominant LT-responsive allele of Nlrp1b that may partially account for the severe response to LT. However, the strength of this response was attenuated in mice with only one copy of chromosome 11 derived from CAST/Ei indicating the existence of a recessive modifier of the inflammatory response to LT. In addition, congenic mice displayed a pronounced immune response using an experimental model of sepsis, indicating that one or more genes within the chromosome 11 region control host response to multiple inflammatory stimuli. Analyzing the influence of allelic variation on gene expression identified 25 genes as candidates for controlling these responses. In summary, we report a genetic model to study inflammatory responses beneficial to the host during anthrax.
| Microbial pathogens have evolved various mechanisms to block host immune responses and thereby increase virulence. The MAP kinase (MAPK) signaling pathways have a central role in innate immune responses mounted by both plants and animals, and are common targets that are inactivated by a variety of bacterial toxins and effector molecules [1], [2]. B. anthracis produces anthrax lethal toxin (LT), a bipartite toxin that contributes to immunosuppression and pathology in the host [3]. The catalytic moiety of anthrax LT, lethal factor (LF), is a zinc-dependent metalloprotease that cleaves the N-termini of MAPK kinases (MKKs). By inactivating MKKs, LT blocks production of proinflammatory chemokines and cytokines such as TNF-α and inhibits survival signals activated via downstream MAPKs [4]–[11]. Thus, LT-mediated cleavage of MKKs leads to the silencing of a pro-inflammatory response, effectively repressing host immunity and favoring bacterial survival [9], [12], [13].
In response to such pathogenic mechanisms, eukaryotic hosts have evolved means to detect and counter pathogen encoded virulence factors that target intracellular signaling pathways. Specifically, nucleotide-binding domain leucine-rich repeat (NLR) proteins sense bacterial products or host cell-derived danger signals to initiate defense pathways. NLR-mediated responses can function locally through induction of cell death and/or distally through production and release of antimicrobial products and signaling molecules. Allelic variation at the NLR gene, Nlrp1b, in rodents is one mechanism that controls the host cellular response to LT and subsequent sensitivity to B. anthracis infection [14]–[16]. Specifically, LT-responsive alleles of Nlrp1b drive caspase-1 mediated proinflammatory cell death, termed pyroptosis, of macrophages and dendritic cells. Increased resistance to B. anthracis is correlated with LT activation of the NLRP1B inflammasome, resulting in IL-1β release and pyroptosis [15], [17].
Sensitivity of multiple animal species to anthrax varies inversely with sensitivity to injection of purified LT [18]. This inverse relationship holds true when comparing inbred strains of mice [19]. Therefore, genetic comparison of mouse strains is predicted to reveal mechanisms of the host response to B. anthracis. Indeed, differential sensitivity of mouse strains to B. anthracis infection is known to be influenced by allelic variations in at least two genes: Nlrp1b and Hc encoding complement C5 [15], [17], [19]–[23]. Allelic variation of these genes does not, however, fully account for differential sensitivity to infection or to intoxication by LT [17], [24]. Therefore, we hypothesized that additional genes may contribute to host susceptibility to anthrax. Due to the critical role of LT as a virulence determinant for B. anthracis, a genome-wide collection of congenic mouse strains was screened for altered responses to this toxin. Here we report the identification of a quantitative trait locus (QTL) on chromosome 11 that influences host response to multiple inflammatory stimuli including LT, resulting in increased resistance to B. anthracis Sterne infection.
To identify chromosomal regions affecting response to LT, a library of congenic mice consisting of homozygous CAST/Ei segments on a B6 background was screened. The genome coverage of this library spans roughly 80% of the autosomal chromosomes and consists of approximately three strains per chromosome in which CAST/Ei segments are introgressed onto the C57BL/6J (B6) background in an overlapping manner [25]. Three strains, B6.CAST.11 medial (B6.CAST.11M), B6.CAST.11 proximal medial (B6.CAST.11PM), and B6.CAST.11 complete (B6.CAST.11C), all harboring CAST/Ei segments on chromosome 11, displayed a rapid and transitory response following LT injection similar to the early response phenotype (ERP) previously observed in LT-challenged B6Nlrp1b(129S1) transgenic mice [15]. The latter express a 129S1/SvImJ(129S1)-derived LT-responsive allele of Nlrp1b on an otherwise LT-resistant B6 background. Akin to the ERP of LT-injected B6Nlrp1b(129S1) transgenic mice, chromosome 11 B6.CAST mice presented with ataxia (Figure 1A), hypothermia (Figure 1B) and one or more of the following: bloat, dilated vessels on pinnae, loose/watery feces, labored abdominal breathing (not shown). This response developed as early as 30 min post LT injection, and all animals presented by 4 h (not shown). Importantly, the ERP displayed by chromosome 11 B6.CAST mice was significantly more pronounced compared to that in B6Nlrp1b(129S1) animals as evidenced by a more severe ataxia score (Figure 1A) as well as a more severe hypothermic state (Figure 1B). Other congenic strains and B6 mice did not display clinical signs associated with the ERP (data not shown). Upon careful observation, CAST/Ei and BALB/c strains displayed a very mild, inconsistent version of these early signs (data not shown), indicating that the mixture of CAST/Ei alleles of chromosome 11 genes with B6 alleles accounting for the rest of the genome likely resulted in increased expressivity of a toxin-responsive phenotype present in LT-responsive strains. The response to LT was independent of route of toxin administration and was present when mice were administered toxin intravenously (i.v.) (not shown). Interestingly, i.v. LT challenge produced an accelerated presentation of the ERP that was observed as quickly as 18 min post challenge (not shown). Following the ERP, chromosome 11 B6.CAST mice typically recovered to normal behavior within 4–25 h post LT injection and subsequently relapsed into a second round of clinical signs, eventually succumbing to moribund state and/or death within the same timeframe as parental B6 mice (Figure 1C, and not shown). Endotoxin contamination of protective antigen (PA), the host cell-binding moiety of LT, or LF was not responsible for the ERP or ultimate lethality, as no response was detected following injection of a 2X dose of individual toxin components (data not shown).
Anthrax LT induces a rapid pyroptotic cell death in macrophages derived from mice with LT-responsive alleles of Nlrp1b. Five murine alleles of Nlrp1b have been described [14]. Allele 2, encoded by B6 mice, and alleles 3 and 4 do not respond to LT, while allele 1, encoded by 129S1 and Balb/C mice, and allele 5, encoded by CAST/Ei mice, are LT-responsive. Responsiveness to LT is fully dominant and macrophages from heterozygous mice with one LT-responsive and one LT-resistant allele display sensitivity to LT indistinguishable from macrophages encoding two LT-responsive alleles ([26] and data not shown). The kinetics of ERP in mice is consistent with timing of macrophage and DC pyroptosis treated with LT ex vivo. Therefore, we sought to determine whether allelic variation of Nlrp1b influenced sensitivity of macrophages to LT. Bone marrow derived macrophages (BMDMs) from B6Nlrp1b(129S1) and B6.CAST.11M mice, as well as those from BALB/c mice encoding an LT-sensitive allele of Nlrp1b (allele 1) displayed similar sensitivity to LT ex vivo (Figure 1D). Further, BMDMs from B6Nlrp1b(129S1) and B6.CAST.11M mice displayed similar IL-1β responses to LT (Figure 1F). Together, these data demonstrate that the more severe ataxia and hypothermia observed in B6.CAST.11M mice did not result from alterations in LT-induced macrophage pyroptosis.
We next tested whether B6.CAST.11M mice have an altered response to additional inflammatory stimuli. Preliminary studies indicated that B6.CAST.11M mice, but not B6 controls, display ataxia and clinical signs associated with inflammation following challenge with recombinant IL-1β (not shown). To further text LT-independent inflammatory responses, an established model of sepsis was employed whereby mice were injected i.p. with muramyl dipeptide (MDP) and lipopolysaccharide (LPS), resulting in a rapid TNFα-dependent hypothermia [27]. Using this model, B6.CAST.11M mice displayed exacerbated ataxia (not shown) and hypothermia (Figure 1E) compared to B6 control animals. This response was not due to alterations in macrophage responsiveness to MDP/LPS as determined by TNFα and IL-1β release (Figure 1F). These results are consistent with a heightened responsiveness of CAST/Ei chromosome 11 alleles to multiple inflammatory stimuli.
Previously, we reported that the LT-induced ERP in B6Nlrp1b(129S1) mice was associated with release of proinflammatory cytokines [15]. To determine whether a proinflammatory cytokine response also accompanies the LT-induced ERP phenotype in chromosome 11 B6.CAST mice, sera from LT-challenged animals were analyzed. As predicted, proinflammatory cytokines previously identified as induced in B6Nlrp1b(129S1) mice were also induced in B6.CAST.11M mice following LT challenge (Figure 2A). Interestingly, these cytokines were induced to a similar level and with similar kinetics in both strains despite the exacerbated ERP displayed by B6.CAST.11M mice compared to B6Nlrp1b(129S1) [15]. This finding is consistent with the observation that BMDMs derived from these strains displayed similar pyroptotic responses to LT (Figure 1D, F). Endotoxin contamination of PA or LF was not responsible for the cytokine induction observed, as no response was detected following injection of a 2X dose of individual toxin components (not shown). To further test whether alterations in cytokine responses could explain the altered phenotype severity, a panel of additional cytokines and chemokines were assayed following LT challenge (Figure 2B–D). A total of 27 cytokines were induced in B6.CAST.11M and/or B6Nlrp1b(129S1) but not B6 mice (Figure 2A –C), while five cytokines showed no response in any strain (Figure 2D). Only four out of 27 cytokines that responded to LT were differentially induced in the sera from B6Nlrp1b(129S1) mice compared to sera from B6.CAST.11M mice (Figure 2C). Interestingly, all four of these cytokines were preferentially induced in B6Nlrp1b(129S1) mice compared to B6.CAST.11M mice. Of the cytokines differentially induced, three function as pro-inflammatory mediators while one cytokine, IL-4, exhibits both pro- and anti-inflammatory properties [28].
B6.CAST congenic strains displaying the LT-mediated ERP share a CAST/Ei derived critical region between 43–107 Mb on chromosome 11 (Figure 3A). A single cytokine, IL-4, preferentially induced in B6Nlrp1b(129S1) mice during the LT-induced ERP relative to B6.CAST.11M mice (Figure 2C) maps to this critical region and is encoded at 53.4 Mb on chromosome 11. Interestingly, IL-4 functions as a Th2 cytokine, and elevated expression of IL-4 has been linked to reduced inflammation during sepsis in humans [29]. To determine the role of IL-4 in the inflammatory response to LT, mice deficient in IL-4 but expressing a LT-responsive allele of Nlrp1b were generated and tested for their response to LT. If reduced IL-4 levels in B6.CAST.11M mice are responsible for the exacerbated LT-induced ERP, then B6Nlrp1b(129S1);IL4−/− mice would be predicted to display a strong ERP akin to that in B6.CAST.11M mice. However, B6Nlrp1b(129S1);IL4−/− mice displayed an ERP equal in strength to B6Nlrp1b(129S1) mice following LT injection (Figure 3B), indicating that the absence of IL-4 does not affect ERP severity following LT challenge.
Nlrp1b inflammasome activation results in ASC-dependent maturation of cytokines including IL-1β, as well as ASC-independent cell lysis [30]. The ERP is associated with high serum concentrations of multiple cytokines, of which IL-1β is one of the earliest detectable (Figure 2A). To determine whether IL-1β is required for the ERP, B6.CAST.11M, and B6Nlrp1b(129S1) mice were pretreated with a blocking antibody to IL-1β, then challenged with LT. Notably, this treatment had no significant effect on ERP as determined by ataxia scoring (not shown). However, it remained possible that the anti IL-1β antibody treatment was not sufficient to block the LT-induced release of this cytokine. Therefore, mice deficient in the type 1 IL-1 receptor (IL-1R) and expressing the LT-responsive 129S1 allele of Nlrp1b were tested for response to LT. Consistent with the antibody studies, IL-1R deficient B6Nlrp1b(129S1) animals displayed ataxia (not shown) and hypothermic responses indistinguishable from B6Nlrp1b(129S1) mice (Figure 3C).
Presentation of the ERP in response to LT correlates with an increased resistance to B. anthracis Sterne infection [15]. To determine whether the more severe ERP displayed by chromosome 11 B6.CAST mice correlates with an increased resistance to B. anthracis spore challenge, B6.CAST.11M, B6Nlrp1b(129S1), and B6 mice were challenged i.p. with 3×107 Sterne strain spores. At this dose, all B6.CAST.11M and B6Nlrp1b(129S1) mice survived to the experimental endpoint whereas the majority of B6 animals succumbed to the infection (Figure 4A). Using a ∼13 fold higher dose of 4×108 Sterne spores revealed that B6.CAST.11M mice were significantly more resistant to infection compared to B6Nlrp1b(129S1) mice (Figure 4B). Therefore, a more robust ERP correlates with increased protection from B. anthracis Sterne infection. Further, allelic variation of a chromosome 11– encoded gene(s) contributes to the increased ability of B6.CAST.11M congenic mice to limit B. anthracis Stern infection.
To determine the cellular mediators providing infection resistance, mice were challenged with Sterne spores i.p. and peritoneal exudates were collected and analyzed at various time points following challenge. A significant increase in the number of PMNs was observed in both B6Nlrp1b(129S1) and B6.CAST.11M mice at earlier time points following spore challenge compared to nontransgenic (i.e. B6) control mice (Figure 4C). These results correspond to previously reported data indicating a role for PMNs in bacterial clearance following B. anthracis spore challenge [12], [15], [21]. However, no significant differences were observed in the PMN response between B6Nlrp1b(129S1) and B6.CAST.11M mice that could explain their differential resistance to B. anthracis. Indeed, at 4 h, neutrophil influx in B6Nlrp1b(129S1) animals was significantly greater than that in B6.CAST.11M mice (p = 0.04). Similarly, monocytic infiltration could not explain the increased resistance seen in B6.CAST.11M mice (Figure 4D).
Given that leukocyte infiltration did not account for the difference in susceptibility to B. anthracis, we next considered whether the differential cytokine response could explain the increased resistance in B6.CAST.11M animals. Although IL-4 did not contribute to hypothermia (Figure 3B) or ataxia (not shown) following LT challenge, it remained possible that this cytokine still affected resistance to spore challenge. Indeed, as a Th2 cytokine, IL-4 can function by altering PMN and macrophage activity [31], [32]. Alteration of phagocyte function may contribute to differential responses in long-term studies such as resistance to spore challenge, but not contribute to immediate phenotypes such as ataxia and hypothermia following LT challenge (Figure 3B). To test the role of IL-4 in resistance to B. anthracis, B6Nlrp1b(129S1);IL4−/− mice were challenged i.p. with Sterne strain spores (Figure 4E). IL-4 deficiency did not affect host susceptibility to spore challenge in the presence of the LT-responsive allele of Nlrp1b, excluding this gene as a candidate. Interestingly, in the absence of an LT-responsive allele of Nlrp1b, the loss of IL-4 resulted in a slightly higher resistance to anthrax (Figure 4F).
Next, the role of IL-1β in Nlrp1b-mediated resistance to B. anthracis Sterne was tested. LT dampens the host cytokine response in the absence of LT-responsive Nlrp1b [9], [10], [33], [34]. However, this immunosuppression is not absolute and IL-1β contributes to resistance to B. anthracis even in the absence of LT-responsive Nlrp1b; animals expressing LT-resistant alleles of Nlrp1b and lacking IL-1R or MyD88 (required for TLR and IL-1R signaling) show increased sensitivity to infection by B. anthracis [35]–[38]. A critical role for IL-1R was further validated in mice expressing LT-responsive alleles of Nlrp1b [17]. However, it is still unknown whether increased resistance to B. anthracis mediated by LT-responsive alleles of Nlrp1b requires IL-1β. Of note, inflammasome-mediated resistance to Francisella tularensis is mediated by both IL-1β and IL-18, and mice deficient in either cytokine are resistant to tularemia, while those deficient in both IL-1β and IL-18 are sensitive [39]. To address the mechanism by which Nlrp1b mediates resistance to B. anthracis, B6Nlrp1b(129S1);IL-1R−/− mice were challenged with Sterne strain spores and viability was compared to that of spore-challenged B6IL-1R−/− mice. Both strains showed similar susceptibility to B. anthracis (Figure 4G), indicating that Nlrp1b-mediated protection from anthrax requires IL-1β signaling.
We previously demonstrated that a LT-responsive 129S1 allele of Nlrp1b is sufficient to drive the ERP in B6Nlrp1b(129S1) mice [15]. The CAST/Ei allele of Nlrp1b is also LT-responsive, though distinct from the 129S1 allele [14]. To determine the contribution of the CAST/Ei allele of Nlrp1b in the B6.CAST.11 response to LT, B6.CAST.11M mice were crossed to B6 animals, and the resulting [B6.CAST.11M x B6] F1 mice were subsequently challenged with LT. LT-responsive alleles of Nlrp1b behave in a fully dominant and penetrant manner in controlling activation of caspase-1 and resulting macrophage pyroptosis [14], [26]. We therefore predicted that if allelic variation of Nlrp1b was responsible for the exacerbated ERP displayed by B6.CAST.11 mice, [B6.CAST.11M x B6] F1 offspring expressing one CAST/Ei allele of Nlrp1b would display an ERP equal in strength to the ERP presented by B6.CAST.11M mice. Strikingly, F1 animals displayed the ERP at a much weaker strength (Figure 5). Thus, the ERP observed in B6.CAST.11 mice is controlled by a gene or genes that behave in a recessive fashion.
To refine the critical region controlling the increased inflammatory response to LT, we undertook a positional cloning strategy. B6.CAST.11M mice were crossed to B6, and the resulting [B6.CAST.11M x B6] F1 mice were backcrossed to B6.CAST.11M animals. Backcross progeny (n = 139) were intoxicated and scored for ERP presentation based on ataxia and body temperature (Figure 6A). To correlate phenotype with genotype within the chromosome 11 critical region, backcross progeny were genotyped at multiple positions along the critical region using chromosome 11 microsatellite markers (Figure 6B). A subset of mice displaying a robust ERP following LT challenge, as evidenced by a body temperature less than 30°C and severe walking ataxia score, was heterozygous (i.e. CAST/Ei and B6) at multiple microsatellite markers localized within the original critical region on Chromosome 11 (Figure 6B). Conversely, several mice displayed a weak ERP, as evidenced by a body temperature greater than 34°C and mild walking ataxia score, despite retaining homozygous CAST/Ei alleles over a significant fraction of the critical region (Figure 6B). Comparison of the genotypes of these groups of mice revealed that a gene(s) that modifies host response to LT lies within the 66.36–74.67 Mb region on chromosome 11.
To identify candidate genes that contribute to host response to LT from the 273 genes in the current critical region, we employed expression quantitative trait locus (eQTL) analysis [40]–[42]. This approach relies on the hypothesis that differential gene expression is linked to phenotypic changes [40]–[43]. Transcript levels can be analyzed as genetic traits in the same way that phenotypes such as LT sensitivity can be analyzed [44], [45]. Using a single, large cross between two strains of mice, genetic loci, i.e. eQTL, that control differential transcript levels can be mapped. Importantly, data from a single experiment can be analyzed iteratively to identify a unique eQTL associated with every gene with an altered transcript level. If the transcript levels are controlled by structural variation of a gene that influences its rate of transcription or the maturation or stability of the transcript, the eQTL would be expected to map directly over the gene in question. Such an eQTL is termed a cis-acting eQTL. The approach of examining the segregation of transcript levels in a genetic cross should be distinguished from a study in which transcript levels are simply compared between two different strains. In the latter study, differences in transcript levels can be identified, but it is not possible to determine from such data whether any differences observed are the result of cis-acting genetic differences or trans-acting differences. In contrast, the combined genetic/gene expression approach makes it possible to identify genes and pathways that are perturbed by the cis- acting genetic variations.
Analysis of a dataset consisting of over 1,600 microarray experiments performed on 442F2 (B6 x CAST/EiJ) mice identified 81 cis-acting eQTL with a LOD score >4.3 within the refined critical region (Table S1). Comparison of genomic sequences of CAST/EiJ and B6 revealed multiple single nucleotide polymorphisms (SNPs) within the predicted mature mRNA sequences of 74 of the 81 eQTL, including non-synonymous changes in the coding regions (not shown). Further refinement based on reported gene function and expression pattern resulted in a list of 25 high-priority candidate genes (i.e., genes known to participate in immune or inflammatory processes) (Table S1).
We previously reported a role for a LT-sensitive Nlrp1b allele in controlling an inflammatory response to LT that we termed the ERP [15]. This response was characterized by the release of proinflammatory cytokines, noticeable ataxia and a reduction in body temperature that were not present in B6 animals encoding only LT-resistant alleles of Nlrp1b [15]. The ERP reported here in multiple B6.CAST.11 strains was more severe compared to the response exhibited by B6Nlrp1b(129S1) animals (Figure 1A, B), and this correlated with increased resistance to B. anthracis (Figure 4B). Two models exist that may explain the altered response to LT between B6.CAST.11 and B6Nlrp1b(129S1) animals. First, variation between CAST/Ei and 129S1 alleles of Nlrp1b may directly contribute to ERP severity. Indeed, the current critical region contains Nlrp1b, and CAST/Ei alleles of this gene are LT-responsive yet genetically distinct from other LT-responsive alleles [14]. However, this explanation would require that CAST/Ei alleles of Nlrp1b contribute to phenotypes, such as response to MDP + LPS (Figure 1E, F), not currently associated with the Nlrp1b inflammasome. Further, two established mechanisms by which inflammasomes drive innate immune responses, macrophage pyroptosis (Figure 1D) and cytokine responses (Figs. 1F, 2), cannot account for altered severity of host response to LT in B6.CAST.11 mice. Finally, [B6.CAST.11M x C57BL/6] F1 mice showed loss of the severe LT-induced ERP (Figure 5), indicating a recessive mode of inheritance that is inconsistent with the well-established dominance of LT-sensitive alleles of Nlrp1b [14], [26]. Interestingly, the ataxia scores in F1 mice match those seen in C57BL/6Nlrp1b(129S1) transgenic mice, consistent with retention of one dominant LT-responsive allele of Nlrp1b, and further suggesting that both CAST/Ei and 129S1 alleles do in fact function in a similar, dominant manner to drive a less severe ERP. Therefore, we propose a second model, wherein at least one additional gene in the critical region of chromosome 11 is required for full expressivity of the ERP and high-level resistance to spore challenge. While we cannot exclude a novel role in inflammation and/or mechanism of action for the CAST/Ei allele of Nlrp1b, the findings presented here are more consistent with contributions by an additional gene. Of note, this model does not rule out the possibility that multiple or different genes contribute to the host response to MDP + LPS, LT, and/or spores. Indeed, the critical region contains numerous genes with established or proposed roles in inflammation (Table S1).
Three murine QTL, Lethal toxin sensitivity 1-3 (Ltxs1-3), and one gene, Nlrp1b (within Ltxs1), have been reported to control macrophage and/or whole animal sensitivity to LT [14], [46], [47]. Although backcross mapping data presented here eliminates a role for Ltxs3 contributing to the ERP, it is possible that one or more genes within Ltxs1 and/or 2 contribute to the severity of this phenotype. Notably, while the gene at Ltxs1 responsible for macrophage pyroptosis was identified as Nlrp1b [14], a dominantly inherited resistance to LT was also mapped to this region [46], indicating that more than one overlapping QTL may exist at Ltxs1. Indeed, in contrast to LT-resistance phenotypes reported for Ltxs1 and Ltxs2 [46], allelic variation in B6.CAST.11 animals does not alter susceptibility to LT, defined as time to moribund behavior, compared to B6 mice (Figure 1C). Therefore, we propose that additional QTL influencing host response to LT reside within this critical region.
In order to identify candidate genes within the critical region that potentially contribute to ERP and resistance to B. anthracis Sterne, we mined a gene expression dataset from a B6 x CAST/EiJ F2 cross for cis-acting eQTL. Variation in basal transcription levels of genes within a disease QTL has been used previously to identify candidate genes controlling phenotypic variation [42], [48]. Given the rapid response to LT, as quickly as 18 minutes post i.v. challenge, we reasoned that alterations in basal gene expression levels and/or protein activity could account for altered ERP strength between B6.CAST.11M and C57BL/6Nlrp1b(129S1) transgenic mice. We further refined the selection criteria for candidate genes to include the presence of single nucleotide polymorphisms within the coding region and known association of candidate genes with inflammatory responses. Finally, we considered other genes previously identified as candidates for host response to LT. The gene encoding inducible nitric oxide synthase (iNOS/NOS2) was previously studied with respect to host response to lethal toxin and was identified as overlapping with Ltxs3 [24], [36], [46], [49]. However, Nos2 falls outside the critical region reported here and was therefore excluded. Within the current critical region, one gene, Mgl1/Clec10a, and one QTL controlling susceptibility to Salmonella infection, Ity2, were proposed as candidates for controlling host response at Ltxs2 [46]. Mgl1/Clec10a, and the neighboring Mgl2, were identified here as high-priority candidates as they encode galactose/N-acetyl-galactosamine binding lectins expressed on alternatively activated macrophages [50]. Alternatively activated macrophages modulate inflammatory responses and Mgl1−/− mice displayed more severe inflammation in a model of experimental colitis [51].
From the list of high priority eQTL candidates, it is noteworthy that several play a role in inflammatory responses by acting through lipid mediators. Alox8, Alox12e and Alox15 belong to a family of arachidonate lipoxygenases responsible for production of anti-inflammatory lipoxins from arachidonic acid. Lipoxins are predicted to suppress vascular changes induced by inflammatory mediators [52]. Similarly, proteins encoded by phospholipase D2, Pld2, phospholipid scramblase 3, Plscr3, phosphoinositide-3-kinase, regulatory subunit 6 Pik3r6, and spinster homolog 2, Spns2, are involved in phospholipid synthesis and/or signaling events associated with inflammatory responses including leukocyte migration, phagocytosis, oxidative burst, and vascular permeability.
β-arrestin 2 (Arrb2) and netrin 1 (Ntn1) are two additional high priority candidates, and are involved in signaling in response to inflammatory stimuli [53]–[56]. Indeed, Arrb2 regulates LPS-induced inflammatory response and endotoxemia [54], [55], while Ntn1 can minimize inflammatory damage associated with ischemia-reperfusion injury [56]. Additional eQTLs were identified in mediators of host innate immunity, including complement C1q binding protein, C1qBP, and the chemokine CXCL16. CXCL16 is elevated in patients with inflammatory bowel disease, and Cxcl16−/− mice display less inflammation in a murine model of enterocolitis [57]. Platelet-activating factor acetylhydrolase, isoform 1b, subunit 1 (Pafah1b1), another high-priority candidate, was implicated in susceptibility to necrotizing enterocolitis in humans, and Pafah1b1 deficiency in mice led to increased susceptibility to this disease [58]. Finally, two eQTLs were identified that have opposite functions in protein metabolism. Eukaryotic initiation factor 4A-I, Eif4a1, is a component of the ribosome involved in protein translation, while Psmb6 (proteasome (prosome, macropain) subunit, beta type 6) is a component of the proteasome involved in protein degradation. Interestingly, PSMB6 is replaced by an alternative proteasome subunit, LMP2, to form an “immunoproteasome” in response to interferon signaling, and has been identified as a candidate gene contributing to autoimmune type-1 diabetes in mice [59].
While differential gene expression has been used extensively with positional cloning efforts to identify candidate genes [42], [59], [60], it is possible that genes responsible for ERP and resistance to B. anthracis Sterne infection are not differentially regulated at the level of transcription. For example, the critical region encodes two paralogues of Nlrp1b, Nlrp1a and Nlrp1c, that were not identified as cis-acting eQTL. Very little is known regarding these paralogues, and it is possible that they have a role in the increased severity of ERP seen in B6.CAST.11 animals.
NLRP1B inflammasome formation allows for the processing and extracellular release of proinflammatory cytokines. The LT-induced ERP is coupled with the induction of several cytokines, of which IL-1β is one of the earliest detectable in the sera of intoxicated mice (Figure 2). IL-1β activity was not required for the ERP, as LT-challenged B6Nlrp1b(129S1);IL-1R−/− animals showed ataxia and hypothermic responses that were indistinguishable from B6Nlrp1b(129S1) mice (Figure 3C). Interestingly, B6.CAST.11M, but not B6 mice challenged i.p. with recombinant IL-1β displayed ataxia (J.K.T. and S.M.L., unpublished observation), suggesting that this cytokine is sufficient but not necessary to bestow the ERP. In contrast, IL-1R deficient B6Nlrp1b(129S1) mice were equally susceptible to infection compared to IL-1R deficient mice expressing only B6 alleles of Nlrp1b, demonstrating that this cytokine is required for Nlrp1b-mediated resistance to B. anthracis. In total, these data indicate that IL-1β plays different roles, i.e., for the ERP it is sufficient but not necessary, but for resistance to anthrax it is necessary. One interpretation is that NLRP1B-mediated pyroptosis results in release of IL-1β that can induce multiple other inflammatory cytokines and/or inflammatory mediators that contribute, in a redundant manner, to ataxia, hypothermia, and other ERP-associated clinical signs, but that only a few are critical for resistance to B. anthracis. Indeed, we did not observe TNFα release from LT-treated BMDMs (Figure 1F), indicating that other cell types may be responsible for the TNFα response, and possibly hypothermia, in LT-treated mice (Figures 1B, 2). This may occur either directly, in response to LT, or indirectly, in response to IL-1β. Conversely, TNFα, but not IL-1β was released from BMDMs derived from all three mouse strains in response to MDP/LPS under the conditions tested (Figure 1F), consistent with a Nlrp1b-independent role for TNFα in LPS-induced hypothermia. Whether the exacerbated response of B6.CAST.11M mice to LT and MDP/LPS derive from a common mediator, or whether these congenic mice are more responsive to multiple mediators is not currently known. The identities and mechanisms of action of such mediators are currently under investigation.
Our understanding of the role of cytokines in response to LT challenge and B. anthracis infection has recently changed. Studies presented here and [15] indicate that cytokines are not a major influence of mortality in response to purified LT injection as once thought, and we find no evidence of alterations in time-to-death as a result of Nlrp1b allele status. In contrast, a modest influence of LT-responsive alleles of Nlrp1b has been reported in a different murine intoxication model [17], [24]. Regardless of the role of cytokines in response to LT challenge, it is clear that IL-1β is critical for resistance during infection with B. anthracis Sterne [17], [36], [37], [61]. Importantly, we extend these prior findings and demonstrate here for the first time that IL-1β is required for protection from infection bestowed by LT-responsive alleles of Nlrp1b.
Septic shock and autoinflammatory diseases such as vitiligo, Crohn's disease, Muckle-Wells syndrome, and gout, among others, result from overactive release of proinflammatory cytokines including IL-1β [62]. Based on differential inflammatory response to disparate stimuli such as LT and LPS/MDP, we propose that B6.CAST.11 and B6Nlrp1b(129S1) mice may provide a unique system to further analyze inflammatory syndromes. Systemic Inflammatory Response Syndrome (SIRS) is a term used for the generalized inflammatory reaction that occurs in patients undergoing sepsis or a non-septic trauma, such as pancreatitis, hemorrhagic shock, thermal injury, or severe surgery [63]. Accordingly, SIRS is defined by the presentation of two or more of the following clinical markers that match LT-induced ERP: a body temperature higher than 38°C or lower than 36°C, endothelial dysfunction and increased microvascular permeability, and platelet sludging causing maldistribution of blood flow [64], [65]. A second inflammatory syndrome, Compensatory Anti-inflammatory Response Syndrome (CARS), was more recently defined as a counter-regulatory response to the overzealous inflammatory reaction that occurs in SIRS [29]. This adaptive response is characterized by the increased induction of anti-inflammatory molecules IL-10 and IL-4 [29]. Although attenuation of inflammation may be beneficial in some instances, CARS patients are often more susceptible to secondary bacterial infections [29], [66]. Interestingly, B6Nlrp1b(129S1) transgenic mice display a weaker ERP (Figure 1), are more susceptible to B. anthracis infection (Figure 4), and release higher levels of IL-4 than B6.CAST.11M mice in response to LT (Figure 2). Because the gene encoding IL-4 resides at 53.4 Mb on chromosome 11 (part of the original critical region), we tested the role of this cytokine in the differential response to LT. B6Nlrp1b(129S1);IL4−/− mice showed no alteration in the ERP or resistance to B. anthracis compared to B6Nlrp1b(129S1) animals, indicating this cytokine cannot, on its own, account for differences in hypothermia, ataxia, or resistance to B. anthracis between B6Nlrp1b(129S1) and B6.CAST.11M animals. However, mice lacking both an LT-responsive allele of Nlrp1b and the IL4 gene were slightly more resistant to B. anthracis spores compared to B6 animals (Figure 4F), suggesting a negative role for Th2 responses in the outcome of B. anthracis infection. Further mapping confirmed that IL4 lies outside the current critical region (Figure 6), eliminating it as a candidate involved in the regulation of the ERP. Notably, IL-10 levels were equivalent between B6Nlrp1b(129S1) and B6.CAST.11M mice following LT challenge (Figure 2). Therefore, we hypothesize allelic variation of gene(s) in the chromosome 11 critical region represent a novel genetic control mechanism for a SIRS/CARS like inflammatory response to LT. B6.CAST.11M mice further showed increased responsiveness to a MDP + LPS model of sepsis compared with B6 controls (Figure 1E), indicating that gene(s) within the critical region may play a role in multiple inflammatory syndromes including sepsis. Currently, no single treatment for sepsis exists [67], and there is little understanding of mechanisms driving resolution of shock/sepsis [68], [69]. We predict that identification of genetic and molecular mechanisms controlling severity of LT-induced ERP and/or host response to MDP + LPS will provide insight into novel intervention strategies for sepsis and other inflammatory diseases.
All studies involving the use of mice were conducted in compliance with the Animal Welfare Act and other federal statutes and regulations relating to animals and experiments involving animals and adheres to principles stated in the Guide for the Care and Use of Laboratory Animals, National Research Council, 1996. All studies involving the use of mice were approved by the University of California Animal Research Committee and/or the USAMRIID Animal Care and Use Committee (Permit numbers: 2005-122 and 2007-019). UCLA and USAMRIID are fully accredited by the Association for the Assessment and Accreditation of Laboratory Animal Care International.
B6 and CAST/EiJ mice were purchased from the Jackson laboratory (Bar Harbor, ME). Transgenic mice expressing a 129S1/SvImJ (129S1)-derived LTS allele of Nlrp1b/Nalp1b on a LT-resistant (LTR) B6 background (B6Nlrp1b(129S1)) and backcrossed to B6 for seven generations were obtained from Drs. E. Boyden and W. Dietrich (Harvard Medical School) [14], [15]. B6Nlrp1b(129S1) mice both heterozygous and homozygous for the Nlrp1b transgene were used for experiments reported here as no difference in response to LT or Sterne spores was observed. The library of congenic mice consisting of introgressed segments of CAST/Ei DNA on a B6 background (B6.CAST) has been described [25]. B6.CAST.11M mice were crossed to B6 to obtain [B6.CAST.11M x B6] F1 offspring used for intoxication experiments and for determining the mode of inheritance. Backcross progeny were generated by crossing B6.CAST.11M mice to B6 mice, and backcrossing the [B6.CAST.11M x B6] F1 mice to B6.CAST.11M mice.
Genomic DNA was isolated from tail biopsies using Qiagen DNeasy blood and tissue kit. Presence of Nlrp1b(129S1) transgene was monitored as previously described [15]. Mutant and wild type alleles of Il1r1 were identified using PCR primers oIMR000160, oIMR0161, oIMR7898, and oIMR7899 as per supplier's protocol (Jackson Laboratory, Bar Harbor, ME). Mutant and wild type alleles of Il4 were identified using PCR primers oIMR0077, oIMR0078, and oIMR0079 (Jackson Laboratory, Bar Harbor, ME). [B6.CAST.11M x B6] F1 x B6.CAST.11M backcross progeny were genotyped at multiple positions within the region of interest using chromosome 11 microsatellite markers that distinguish CAST/Ei and B6 alleles (Mouse Genome Informatics (MGI) database (http://www.informatics.jax.org)). All 139 backcross progeny were genotyped at D11MIT190 (47.6 Mbp), D11MIT131 (55.9 Mpb), D11MIT260 (61.6 Mbp), D11Die36 (70.9 Mbp), D11MIT357 (90.1 Mbp), and D11MIT199 (101.7 Mbp) by PCR and gel electrophoresis. Backcross progeny that displayed genetic recombination within this region and/or presented with a strong ERP were further genotyped at the polymorphic microsatellite markers indicated in Figure 6B.
LT components were expressed and purified as previously described [15]. Specifically, the PA expression plasmid, PA-pET22b was provided by Dr. John Collier (Harvard Medical School) and transformed into Escherichia coli BL21 DE3 cells. A fresh colony was inoculated into a 20 mL starter culture of Luria Bertani (LB) Lennox media (EMD Biosciences, Inc.) with 100 mg/mL ampicillin and grown overnight at 37°C. The following day a 1∶50 dilution was made into a 2 L baffled Erlenmeyer flask of LB Lennox supplemented with 100 mg/mL ampicillin. The culture was grown at 37°C and shaking at 250 rpm until an optical density of 1.0 was reached. The culture was then induced with a final concentration of 1 mM isopropyl β-D-1-thiogalactopyranoside and allowed to grow at 30°C and shaking at 250 rpm for 4 h. PA was isolated from the periplasm and purified over a Macro-Prep HighQ (BioRad) column. LF expressed and purified from B. megaterium was obtained from Dr. Jeremy Mogridge (University of Toronto), and resuspended in pharmaceutical grade saline for all animal experiments. A dose of 5 µg PA and 2.5 µg LF per g body weight was diluted in pharmaceutical grade saline and injected i.p. Alternatively, PA and LF were purified from B. anthracis strain BH450 as described [70]. LF produced from strain BH450 displayed 3-fold lower activity [71], and consequently a dose of 15 µg PA and 7.5 µg LF per g body weight was used to achieve a similar mortality rate [15]. Endotoxin was removed from PA and LF protein preparations using the Detoxi-Gel Endotoxin Removing Gel (Pierce). Purified proteins were assayed for endotoxin using the Limulus Amebocyte Lysate kit (BioWhittaker/Lonza Bioscientific), which detects a minimum of 0.03 endotoxin units/mL.
A total of 36 B6.CAST congenic strains were screened for LT response. For each strain, ∼5 mice of mixed gender and ∼8 weeks of age were administered LT by i.p. injection at a dose that induced mortality in B6 control animals with a mean time to death of ∼60 h (see above). Larger-scale secondary screens were performed using 5–10 additional animals of each candidate or control strain. Upon injection, mice were closely monitored for ∼2 h and then every 2–3 h for clinical signs consisting of ataxia, bloat, lethargy, loose feces and/or hunched posture. Ataxia was measured using a grading system in which mice were scored as displaying a mild, moderate, or severe phenotype. A mild phenotype was defined as reduced exploratory behavior or rearing on hindlimbs, a slower and/or less steady gait, but free ambulation throughout the cage. A moderate score was defined as a preferred sedentary state, but the mouse was able to generate a slow, unsteady (e.g. wobbly) gait for up to ∼7 sec before resting. A severe score was defined as a stationary state, but upon stimulation the mouse could generate a few unstable steps (e.g. severe wobble and/or tremor) before stopping.
For experiments focused on the acute presentation of clinical signs, mice were closely monitored for ∼6–8 hours then euthanized. Mouse body temperatures were measured following LT injection using a rectal thermometer probe. Baseline temperatures were determined prior to LT injection and no differences were observed between animal groups (not shown). Temperatures were recorded hourly for up to 7 h following LT injection. Some mice were injected i.p. with recombinant murine IL-1β (R & D Systems) at a dose of 100 ng per mouse immediately prior to LT injection and observed as described above.
In vitro BMDM intoxication studies were performed as previously described [72]. Briefly, bone marrow was flushed from femur and tibia bones of 8-week old mice using DMEM supplemented with 10% fetal bovine serum (FBS) and 1% penicillin-streptomycin-glutamine (PSG) cocktail (Invitrogen). Marrow was then briefly centrifuged at 1500 rpm and resuspended in 1x ACK Lysis Buffer (150 mM NH4Cl, 1 mM KHCO3, 100 µM Na2-EDTA, pH 7.2) for 5 min. Cells were briefly centrifuged at 1500 rpm and resuspended in DMEM supplemented with 10% FBS, PSG, and 2% conditioned medium of CMG14-12 cells as a source of M-CSF. Ten million cells were seeded per 15 cm plate and incubated at 37°C, 5% CO2 for 6–7 days, then harvested and seeded at 5×104 cells per well in a 96 well plate in DMEM supplemented with 10% FBS, PS cocktail (Invitrogen), 25 mM HEPES, and Glutamax (Invitrogen). PA and LF were titrated in a final volume of 100 µL / well and incubated for 4 h prior to addition of ATPlite 1-step reagent (PerkinElmer, Waltham, Massachusetts). Luminescence intensity was measured using Victor 3V (PerkinElmer) plate reader. Alternatively, BMDMs were exposed to 0.1 µg/mL MDP and 0.1 ng/mL LPS for 8 hr [27], or 250 ng/mL PA and 250 ng/mL LF for 3 hr prior to collecting cell culture supernatant for cytokine analysis.
For cytokine analysis, mice were injected i.p. with LT at a dose of 15 µg PA and 7.5 µg LF per g body weight. Blood was collected via cardiac puncture and allowed to coagulate for ∼1 h. Samples were centrifuged, and sera was collected and stored at −80°C. Cytokines were detected using the Millipore Milliplex MAP Mouse Cytokine Kit per manufacturer's instructions and read on a Luminex 100 IS or BioRad Bioplex 200 instrument. Serum from each mouse was analyzed in duplicate and average values from these independent measurements were used to calculate a mean for each animal group (n = 5). For BMDM experiments, at least two independent mice per strain were used, and cells from each mouse were tested in duplicate. All cytokine data were analyzed using MILLIPLEX Analyst software (EMD-Millipore) with five parameter logistic curve fitting.
B6.CAST.11M, B6Nlrp1b(129S1), and non-transgenic littermate/cagemate (i.e. B6) mice were challenged with unencapsulated, toxigenic Sterne strain (7702) by i.p. injection as indicated in figure legends and monitored daily for 14 days. B6.CAST.11M, B6Nlrp1b(129S1), and B6 mice were also used for peritoneal cellular infiltration determination following spore challenge. For cellular analysis, mice were infected i.p. with ∼1.6×107 Sterne spores and euthanized at 4, 28, 52, 76, and 135 h post infection. Uninfected mice were used to determine baseline cell populations in the peritoneal cavity of each strain (t = 0). Peritoneal exudates from infected mice were harvested by injecting 7 ml sterile HBSS and 3 ml air into the peritoneal cavity. The fluid was agitated within the cavity and then extracted. The fluids were analyzed as follows from three mice per group at each time point, except where indicated. Total cell counts were determined microscopically by using a hemocytometer; four fields for each mouse were counted and averaged. The cells from an aliquot of sample were then collected onto slides with a Cytospin centrifuge (Shandon, Inc., Pittsburgh, PA). The slides were fixed in methanol, stained with Diff-Quik (Harleco, Philadelphia, PA), and then evaluated microscopically to determine the percentages of cell types (i.e., percentages of monocytes, PMN). The average % of each cell type per mouse group was calculated, and the total number of each cell type was determined by multiplying the mean proportion of each cell type by the mean hemocytometer count for each mouse group. These values were plotted including standard deviation for each mouse group.
Comprehensive mapping of gene expression in adipose, brain, liver and muscle of 442 F2 progeny of a cross between C57BL/6J and CAST/Ei mice was previously described [73]. From this dataset, we selected genes with strong (LOD >4.3) cis-acting expression-QTLs located in the region between 66.36 and 74.67 on Chromosome 11. Peak LOD scores for each tissue are reported on Table S1. Gene names, positions, and functions were compared to NCBI build 37.1 and MGI annotation datasets. Genes were prioritized by biological function as determined by MGI genome analysis tools (www.informatics.jax.org/tools.shtml) combined with manual curation (www.ncbi.nlm.nih.gov/gene). B6 and CAST/Ei nucleotide sequences were compared using fully sequenced genomes from each strain available at http://www.sanger.ac.uk/cgi-bin/modelorgs/mousegenomes/snps.pl to identify single nucleotide polymorphisms (SNPs). Consistent with genetic divergence between these strains, 64,102 SNPs were identified between 63.36–74.67 Mbp. The search was refined to analyze only predicted mRNA sequences, resulting in 2,421 SNPs that mapped to 74 out of the 81 eQTL.
|
10.1371/journal.pntd.0007621 | Insights into the genetic variation profile of tprK in Treponema pallidum during the development of natural human syphilis infection | Although the tprK gene of Treponema pallidum are thought to play a critical role in the pathogenesis of syphilis, the profile of variations in tprK during the development of human syphilis infection have remained unclear.
Through next-generation sequencing, we compared the tprK gene of 14 secondary syphilis patients with that of 14 primary syphilis patients, and the results showed an increased number of variants within the seven V regions of the tprK gene in the secondary syphilis samples. The length of the sequences within each V region also presented a 3-bp changing pattern. Interestingly, the frequencies of predominant sequences within the V regions in the secondary syphilis samples were generally decreased compared with those found in the primary syphilis samples, particularly in the V7 region, where a frequency below 60% was found in up to 57% (8/14) of all secondary samples compared with 7% (1/14) of all primary samples. Moreover, the number of minor variants distributed between frequencies of 10 and 49.9% was increased. The alignment of all amino acid sequences within each V region of the primary and secondary syphilis samples revealed that some amino acid sequences, particularly the amino acid sequences IASDGGAIKH and IASEDGSAGNLKH in V1, were highly stable. Additionally, the amino acid sequences in V6 also exhibited notable intrastrain heterogeneity and were likely to form a strain-specific pattern at the interstrain level.
The identification of different profiles of the tprK gene in primary and secondary syphilis patients indicated that the tprK gene of T. pallidum undergoes constant variation to result in the best adaptation to the host. The highly stable peptides found in V1 are likely promising potential vaccine components. The highly heterogenetic regions (e.g., V6) could help to understand the role of tprK in immune evasion.
| Antigenic variation of the TprK antigen has been acknowledged to explain the persistence of Treponema pallidum in the host, however, the profile of variations in tprK during the development of human syphilis infection has not been well characterized. Here, we performed next-generation sequencing to compare the variations in tprK between primary and secondary syphilis samples. The profiles of tprK in the samples at different stages showed differences. A higher amount of pool variants within seven V regions was found in the secondary syphilis samples, and the frequencies of their predominant sequences generally decreased with increases in the number of minor variants with frequencies in the range of 10 to 49.9%. However, the length of variable sequences within the V regions of tprK in the secondary syphilis samples also presented a 3-bp changing pattern. Notably, the amino acid sequences IASDGGAIKH and IASEDGSAGNLKH in V1 not only presented a high proportion of interstrain sharing but also were found at a relatively high frequency (above 80%) in the populations. The sequences in V6 of the samples demonstrated substantial variability at the intra- and interstrain levels. These findings could provide insights into the potential syphilis vaccine components and the role of TprK in immune evasion.
| Syphilis is a complex chronic disease caused by infection with Treponema pallidum subsp. pallidum (T. pallidum). Specifically, the disease has a series of highly distinct clinical stages [1], which usually includes the localized chancre primary stage, the disseminated secondary stage, and the late tertiary stage in untreated individuals [2]. This pattern of successive episodes during infection evokes other similar chronic diseases in which antigenic variation explains this characteristic long-term infection [3, 4]. Previous studies have indicated that antigenic variation in outer membrane antigens is a hallmark of many multistage infectious diseases [5–7].
Investigations of tprK from a 12-member gene family (tpr) of the T. pallidum have revealed that tprK is highly heterogeneous at both inter- and intrastrain levels. The sequence diversity of this gene is restricted to seven discrete variable (V) regions (V1–V7), which are separated by conserved sequences [8–10]. Although further investigations are needed to determine whether TprK is an outer membrane antigen [11–13], researchers have found that infection-induced antibody responses are directly related to V regions of TprK, where sequence variations would abrogate specific antibody binding [14, 15]. Therefore, it has been hypothesized that antigenic variations in TprK would facilitate T. pallidum to escape immune clearance and thereby allow the pathogen to persist in the host. Remarkable results using rabbit models support this hypothesis [14–16]. In our previous study [17], we employed a more sensitive and reliable approach, next-generation sequencing (NGS), to explore the tprK gene of T. pallidum directly from primary syphilis patient samples instead of rabbit-derived samples and found that variations in the V regions of tprK generally exhibited a sequence pool containing a high-proportion sequence (frequency above 80%) and many low-frequency minor variants (frequency below 20%). Only specific V region sequences appeared at frequencies of 20–80%.
Based on these findings, we were interested on the variations in the tprK gene in secondary syphilis samples. Comparisons of the variations between primary and secondary syphilis infection could provide notable information on the association of genetic variations in tprK with disease progression, help researchers gain insights into the processes underlying immune evasion by the pathogen and aid the identification of potential vaccine components for human immunology study.
The subjects included in this study were adults, and all of the subjects provided written informed consent in accordance with the institutional guidelines prior to the study. This study was approved by the Institutional Ethics Committee of Zhongshan Hospital, School of Medicine, Xiamen University, and complied with national legislations and the Declaration of Helsinki guidelines.
25 skin lesion samples (erythema or condylomata lata) were collected from patients with secondary syphilis. The clinical diagnosis of syphilis was based on the US Centers for Disease Control and Prevention (CDC) [18] and the European CDC (ECDC) guidelines [19]. The lesions were placed into a sterile Petri dish containing 1 mL of saline (containing 20% normal rabbit serum), minced into very small pieces and squeezed into the liquid [20]. The samples were then examined by dark field microscopy, and the positive samples were used for subsequent DNA extraction.
DNA extraction was performed using the QIAamp DNA Mini Kit (Qiagen, Inc., Valencia, CA, USA) as previously described [21]. qPCR targeting tp0574 was performed to determine whether each DNA sample contained treponemal DNA. 14 secondary syphilis samples (S-1~14) were ultimately included in this study. The samples were then subjected to molecular typing using the Enhance CDC system [22] and amplification of tp0136 to determine whether they belonged to the Nichols-like group or SS14-like group [23].
Segmented amplification of the tprK gene was conducted as described previously [17]. Briefly, the extracted DNA was directly used for amplification of the tprK gene open reading frame (ORF), and the amplicons were gel purified. The purified tprK amplicons (diluted 100×) were used as the segmented amplification template for partial amplification of four fragments of 400–500 bp (overlapping by at least 20 bp) covering the tprK ORF. All the products were verified by 2% agarose gel electrophoresis and gel purified. A high-fidelity PCR polymerase, KOD FX Neo polymerase (Toyobo, Osaka, Japan), was used for the amplification, and the amplification primers are shown in S1 Table. The four subfragment amplicons corresponding to each sample were mixed in equimolar amounts into one pool to produce a separate library, and a barcode was used to distinguish each sample.
Library construction and sequencing were performed by Sangon Biotech Company (Shanghai, China) using the MiSeq platform (Illumina, San Diego, CA, USA) in the paired-end sequencing (2×300 bp) mode. The FastQC and FASTX tools were applied to check and improve the quality of the raw sequence data, respectively. The final reads of tprK were compared with the tprK of the Seattle Nichols strain (GenBank Accession Number AF194369.1) using Bowtie 2 (version 2.1.0) to estimate the sequencing depth and coverage.
Based on a previously described principle for the extraction of sequence data [17], an in-house Perl script was applied to specifically capture DNA sequences within seven regions of the tprK gene from the raw data, both forward and reverse. Thus, the exact number of distinct sequences within seven variable regions of the tprK gene from each sample was acquired. The intrastrain heterogeneous sequences were valid if the following conditions were simultaneously verified: (1) supported by at least fifty reads and (2) with a frequency above 1%. The relative frequency of the sequences within each variable region was then calculated. To systemically present the variation characteristics of tprK at different clinical stages, we included previous data for tprK in primary syphilis patients (X-1~14) for comparison purposes [17].
All statistical analyses were performed using SPSS version 22.0 (SPSS, Chicago, IL, USA). To compare the frequencies below 60% in V7 in secondary versus primary samples, odds ratios were estimated by logistic regression. The Chi-square test was used to identify differences in the amount of variants captured in seven variable regions of the tprK gene between primary syphilis samples and secondary syphilis samples, and the distribution of minor variants between the samples at two different stages comprised three ranges (1–5%, 5–10% and 10%-49.9%). A two-sided P value < 0.05 was considered statistically significant.
The raw data of tprK obtained in this study were deposited in the SRA database (BioProject ID: PRJNA512914) under the following BioSample accession numbers: SAMN10690826- SAMN10690839 for S-1~14. The data of tprK in previous studies were deposited in the SRA database (BioProject ID: PRJNA498982) under the following BioSample accession numbers: SAMN10340238-SAMN10340251 for X-1~14.
The 14 secondary syphilis samples (S-1~14) were collected at Zhongshan Hospital, Xiamen University. The clinical information for all 14 patients is shown in Table 1. The data obtained from the qPCR analysis of the tp0574 gene showed that each DNA sample contained a certain amount of treponemal DNA for amplification of full-length tprK. Molecular typing using the ECDC system detected seven different genotypes, and genotype 16d/f was the most prevalent in these 14 samples (S2 Table). Based on the sequencing data for the tp0136 gene, most strains belonged to the SS14-like group, and only three strains (S-7, S-9 and S-12 strains) belonged to the Nichols-like group. The median sequencing depth of the tprK segment samples ranged from 9810.91 to 52366.84, and the coverage ranged from 99.34% to 99.61%, indicating high identity with the tprK gene of the Seattle Nichols strain (S2 Table). To more clearly present the results, background data for the 14 primary syphilis samples (X-1~14) obtained in previous studies were also included in Table 1.
Using the extraction strategy, distinct nucleotide sequences in the individual V regions of tprK were captured from each sample, and 491 sequences were obtained for the 14 secondary syphilis samples (Fig 1). Calculation of the relative frequencies of distinct sequences within each V region in a single strain revealed that the tprK gene in secondary syphilis samples also contained a predominant sequence within the regions (Fig 2). However, the distribution of the sequences within seven V regions presented some dispersity: the predominant variants had broader frequency spectra (almost between 20–80%), and the minor variants reached higher frequencies (above 20%).
Compared with the sequences captured within each V region from primary syphilis samples in previous studies (335 in total) (S3 Table), the secondary syphilis samples presented a higher number of variants within each V region. The Chi-square test was used to identify differences in the amounts of variants captured within each V region of the tprK gene. The trend found for the sequence variability in the V regions of tprK showed no significant differences between the samples at the two different stages (P = 0.767), and the highest and lowest sequence variability was found in V6 and V1, respectively (Fig 1). Compared to the frequencies of predominant sequences (frequency almost above 80%) within the V regions among primary syphilis samples, the frequencies of the sequences within the V regions among secondary syphilis samples were generally lower and this finding was particularly true for the V7 region, where a frequency below 60% was found in up to 57% (8/14) of the secondary samples compared with 7% (1/14) of the primary samples. We used logistic regression to estimate the odds radio for frequencies below 60% appearing in V7 among secondary versus primary samples. The odds ratio for frequencies below 60% appearing in V7 among secondary samples were 17.3-fold higher than those found among primary samples (OR = 17.3 [95% confidence interval, 1.75 to 171.78]; P = 0.015). Notably, the frequencies of predominant sequences in V1 among all 28 samples remained almost above 80%.
In the secondary syphilis samples, tprK still contained a pool of minor variants within each V region. As shown in Fig 3, most of the minor variants were concentrated in the frequency range of 1–5% in both groups, and the proportions in the other two frequency ranges (5–10% and 10–49.9%) among the secondary syphilis samples were reversed relative to the distribution pattern in primary syphilis samples (9.4% and 14.0%, 14.3% and 9.3%, respectively). However, the Chi-square test was used to investigate the distribution of minor variants in these three ranges, and no significant difference was found between the primary and secondary syphilis samples (P = 0.053).
Additionally, the length of variable sequences within the V regions in secondary syphilis samples corroborated the finding that the length of these variants within the V regions differed in multiples of 3 bp (S4 Table). Compared with the lengths within the regions in primary syphilis samples, V3 and V5 in the secondary syphilis samples maintained the same forms, but the other regions showed the appearance of some new lengths (V1, V4, V6 and V7) or the disappearance of some lengths (V2 and V7).
We translated the variable nucleotide sequences within each V region in silico. No early terminations or changes in the reading frames in tprK were found among the secondary samples, and synonymous sequences were rare and also only found in V2 and V5 (S5 Table). Similar to the phenomenon found among the primary syphilis samples, substantial interstrain sequence redundancy was found in each V region. Altogether, V1, V2 and V4 showed strong shared sequence ability, and V6 showed the least shared ability region (Table 2).
Furthermore, we determined whether a specific V region sequence found in the secondary syphilis samples also presented in the primary syphilis samples. After aligning the amino acid sequences that were unique to the samples at one of the two stages, we found that a number of sequences in V1, V2 and V4 that were specific to the secondary samples were also found in the primary samples (Fig 4A). Notably, the predominant sequences of the V regions also presented overlapping (Fig 4B). Among the seven V regions, V2 and V5 showed more identical predominant sequences between samples at the two different stages. However, the sequences were only found in a few samples. The analysis of the sequences in V1 and V4 showed that although V1 and V4 only presented two identical predominant sequences (IASDGGAIKH and IASEDGSAGNLKH in V1 and DVGHKKENAANVNGTVGA and DVGRKKDGAQGTVGA in V4), the identical sequences showed high interstrain sharing. In addition, the frequencies of the two shared sequences in V1 reached 80% in the strains.
As previously described, V6 was the most variable region of tprK. We further corroborated this feature in the context of primary and secondary syphilis infection. As shown in Fig 4A, only eight identical sequences were found in the samples at the two different stages, and the proportions of these overlapping sequences relative to the unique sequences in the primary and secondary samples were 12.7% (8/71) and 7.8% (8/110), respectively. Moreover, none of the predominant sequences were identical (Fig 4B). The levels of nucleotide diversity in V6 between each sample (Dxy) were calculated using DnaSP v.6.12.01. The Dxy nucleotide diversity in V6 across each sample was almost above 0.15 (S1 Fig), which was in agreement with the proposed view that V6 presents high diversity among most T. pallidum strains.
With the identification of a 12-member gene family (tpr) in the Nichols strain of T. pallidum [24], the antigen-coding tprK had been extensively studied because of its highly variable antigenic profile [9, 10, 14, 25, 26]. Similar to known mechanisms through which many pathogens undergo antigenic variation to evade the immune system and establish chronic infection in the host [5, 27], tprK is believed to play an essential role in the pathogenesis of syphilis [15, 16]. Hence, efforts to understand tprK diversity in the context of human infection, particularly at different clinical stages, would be beneficial to the clinical elucidation of the role of tprK in successive episodes of this chronic infection and would contribute to a more in-depth understanding of the pathogenesis of syphilis.
In this study, NGS was used in combination with an in-house Perl script to confirm the features characterizing the diversity of the tprK gene during natural human infection: tprK contained a predominant sequence and numerous minor variants within each V region, and most variants were found at low frequency in the range of 1 to 5%. Interestingly, in primary syphilis samples, the frequencies of predominant variants were almost above 80%, and those of minor variants were almost below 20%. However, the predominant variants in secondary syphilis samples had broader frequency spectra (almost between 20 and 80%), and more minor variants reached higher frequencies with a broader range of frequencies (above 20%). Combining these two different profiles of the tprK gene, it seems that the variants within the V regions in secondary syphilis samples fill in the middle zone which is almost empty in primary syphilis pattern. This finding suggested that the variations in the tprK gene might follow a logical fitness-based evolution. An analysis of T. pallidum infection at the primary stage showed that the sequences within each V region of the tprK gene presented a two-level distribution (above 80% and below 20%), suggesting the high frequency sequences may be better associated with the avoidance of immune recognition. Changes in the immune environment (development into secondary syphilis infection) could cause the original fitted sequences within each V region to no longer facilitate the survival of T. pallidum. The original predominant sequences in V regions need to change to obtain a new better TprK epitope for T. pallidum. At present, the frequencies of the predominant sequences are lower in the populations, and certain minor variants might be selected and exhibit higher frequencies in the populations. As a result, the original sequences might disappear, and new advantageous sequences would emerge [17, 28]. Due to the continuous evolution of the sequences of tprK, the TprK antigen in the infection process becomes increasingly diverse, which would enable T. pallidum to successively evade the antibody response and thereby establish chronic infection [15, 26].
In addition, we demonstrated that V6 might be the first region to change in primary syphilis samples [17]. In this study, we noted that the predominant sequences in V7 among secondary syphilis samples appeared at frequencies almost below 60% (P = 0.015), which might suggest that variations in V7 evolved following V6 and that the region might be important for the development of secondary syphilis infection [29]. Additionally, a strict 3-bp changing pattern in each variable region was further confirmed in the secondary syphilis samples, and no frame shifts have been found [9, 10, 29], which demonstrates the existence of an elaborate system for the regulation of tprK sequence variation.
Substantial interstrain sequence redundancy was observed in tprK among the samples at the two different stages. Among all V regions, the amino sequences IASDGGAIKH and IASEDGSAGNLKH in V1 and the amino sequences DVGHKKENAANVNGTVGA and DVGRKKDGAQGTVGA in V4 showed strong interstrain sharing ability across 28 clinical strains. Moreover, the sequences in V1 were those that presented a relatively high frequency (above 80%) in the populations. In fact, the two sequences were also found to be the most stable amino acid sequences among the samples in the investigation of Pinto et al. [29]. As described previously [17], a high-frequency amino acid sequence for antigen-coding tprK highly reflects the immune response of the host. Moreover, this sequence presented high-level interstrain sharing, that is, the sequence was found in several syphilitic patients, which indicates that TprK has a better-fitted epitope profile for allowing T. pallidum to adapt to its host. Among the seven V regions in tprK, V1 was found to be relatively stable among the samples at the two different stages, and the two sequences of V1 were most stable among the 28 clinical strains, which suggested that maintaining V1 relatively stable would be essential for the pathogen and that the stable peptides in V1 would be a promising vaccine component for future research [30]. Notably, the promising vaccine peptides found in this study might target the majority but not all of the strains. This problem should be considered further to explore the function of these peptides.
Additionally, we confirmed high heterogeneity at the intrastrain level in V6 throughout the infection process. The existence of a highly diverse region in this antigen-coding gene of an isolate of T. pallidum might greatly enable the pathogen to resist binding by existing opsonic antibodies and might make the pathogen less likely to be recognized by activated macrophages [15]. The amino acid sequences of V6 also presented high diversity at the interstrain level, showing a strain-specific pattern for the sequences, which might explain why the protection of TprK was compromised and a lack of heterologous protection [26]. Currently, it is very difficult to distinguish between treatment failure (relapse) and reinfection in clinical practice. Myint et al. [31] used a molecular method by analyzing tprK sequences to distinguish relapse from reinfection in a patient with recurrent secondary syphilis. Based on the results, whether there is speculation that the sequences in V6 retain a high homology in a relapse case, but the sequences are highly strain-specific in a reinfection case? This speculation requires further supporting evidence from additional future experiments.
Finally, the limitations of our study should be discussed. First, the limited sample size did not support us to draw further definitive conclusions, and the study did not explore the function of the sequences within each V region. A future study could investigate the different peptides in tprK regions observed between primary and secondary syphilis to explore the potential importance of these differences in host interactions and immune evasion. Moreover, the potential promising vaccine components identified in this study could be synthesizes to investigate the immune function of these peptides and thereby lay a foundation for vaccine development. Second, the study provided information on individual V regions instead of information on a single tprK ORF. Using a novel PacBio sequencing pipeline to obtain full length of the tprK sequence could be optimal. And the data might provide useful insights into the structure and function of TprK. Third, because the samples used in this study were from one lesion rather than different lesions, we cannot completely exclude the possibility that the individual patient was infected with different strains resulting in the initial diversity in tprK, even though we verified that the samples presented a single genetic background, as demonstrated by molecular typing (ECDC system and sequencing of tp0136 locus). Additionally, the same genetic background of the tested samples may require similar studies to explore the potential relationship between the genetic background and the variations in tprK.
In this study, we revealed that the characteristic profiles of tprK in the context of primary and secondary infection were different, which indicated that throughout the development of the disease, T. pallidum might constantly undergo variations in its tprK gene to achieve its best adaptation to the host. Interestingly, tprK maintains a contradictory scenario during the course of infection, that is, having a relatively conserved region (V1) and a highly diverse region (V6). The stable sequences in V1 and the highly heterogeneous sequences in V6 could provide important information for exploring promising potential vaccine components and the role of tprK in persistent syphilis infection.
|
10.1371/journal.pcbi.1000390 | Computational Models of the Notch Network Elucidate Mechanisms of
Context-dependent Signaling | The Notch signaling pathway controls numerous cell fate decisions during
development and adulthood through diverse mechanisms. Thus, whereas it functions
as an oscillator during somitogenesis, it can mediate an all-or-none cell fate
switch to influence pattern formation in various tissues during development.
Furthermore, while in some contexts continuous Notch signaling is required, in
others a transient Notch signal is sufficient to influence cell fate decisions.
However, the signaling mechanisms that underlie these diverse behaviors in
different cellular contexts have not been understood. Notch1
along with two downstream transcription factors hes1 and
RBP-Jk forms an intricate network of positive and negative
feedback loops, and we have implemented a systems biology approach to
computationally study this gene regulation network. Our results indicate that
the system exhibits bistability and is capable of switching states at a critical
level of Notch signaling initiated by its ligand Delta in a particular range of
parameter values. In this mode, transient activation of Delta is also capable of
inducing prolonged high expression of Hes1, mimicking the
“ON” state depending on the intensity and duration of the
signal. Furthermore, this system is highly sensitive to certain model parameters
and can transition from functioning as a bistable switch to an oscillator by
tuning a single parameter value. This parameter, the transcriptional repression
constant of hes1, can thus qualitatively govern the behavior of
the signaling network. In addition, we find that the system is able to dampen
and reduce the effects of biological noise that arise from stochastic effects in
gene expression for systems that respond quickly to Notch signaling.
This work thus helps our understanding of an important cell fate control system
and begins to elucidate how this context dependent signaling system can be
modulated in different cellular settings to exhibit entirely different
behaviors.
| The Notch signaling pathway is an evolutionarily conserved signaling system that
is involved in various cell fate decisions, both during development of an
organism and during adulthood. While the same core circuit functions in various
different cellular contexts, it has experimentally been shown to elicit varied
behaviors and responses. On the one hand, it functions as a cellular oscillator
critical for somitogenesis, whereas in other situations, it can function as a
cell fate switch to pattern developing tissue, for example in the
Drosophila eye. Furthermore, malfunctioning of Notch
signaling is implicated in various cancers. To better understand the underlying
mechanisms that allow the network to function distinctly in different contexts,
we have mathematically modeled the behavior of the Notch network, encompassing
the Notch gene along with two of its downstream effector transcription factors,
which together form a network of positive and negative feedback loops. Our
results indicate that the qualitative and quantitative behavior of the system
can readily be tuned based on key parameters to reflect its multiple roles.
Furthermore, our results provide insights into alterations in the signaling
system that lead to malfunction and hence disease, which could be used to
identify potential drug targets for therapy.
| Cells continuously receive signals from their microenvironments – including
factors present in the extracellular matrix, soluble media, and surrounding cells
– which collectively influence cell function and behavior via activating
intracellular signal transduction and gene regulation networks. These networks
generally involve complex, nonlinear interactions of proteins, such as
phosphorylation cascades (reviewed in [1]) and second messenger
signaling systems [2], whose structures feature positive and negative
feedback loops, feed-forward interactions, signal amplification, and cross-talk with
other pathways [3]. Mathematical models of these interactions are
therefore very insightful or even necessary avenues to analyze and understand the
regulation of cell behavior, as the properties of these networks can exceed an
intuitive understanding [4]–[6].
Notch is a signaling system required for numerous critical cell fate specification
events during the development of the nervous system, hematopoietic system, eye, and
skin [7]–[11]. The receptor for
this pathway is the single pass transmembrane protein Notch that, when bound by its
ligands Delta or Jagged, undergoes a series of cleavage events to release its
intracellular domain (NICD) [9],[12].
This NICD then translocates into the nucleus and acts as a transcriptional
upregulator of target genes, including members of the hes family,
through its interaction with the transcription factor RBP-Jκ [13]. In mammals
there are four different Notch proteins (Notch1-4) and 5 ligands (Delta 1, 3, and 4
and Jagged 1 and 2). For this study, we have focused primarily on the Notch1
signaling pathway.
In its role as a critical regulator of cell fate [7]–[11],
Notch has been known to function via lateral inhibition and induction mechanisms to
create fine-grained patterns in undifferentiated cells, a process required for
proper boundary formation and differentiation of various tissues [14],[15]. It
can also function as a binary cell fate switch, for example during differentiation
of the epidermis [16] and endodermal epithelium of the gut [17], to
promote differentiation of one cell type from precursor cells at the expense of
another. Furthermore, in some cases continuous Notch activity is not required for
cell fate specification. For example, transient Delta-Notch signaling has been shown
to be sufficient to induce T-cell [18] and NK cell differentiation [19] from their respective
precursor cells, and can induce an irreversible switch to gliogenesis in neural
crest stem cells [20]. Notch signaling also occurs only transiently in
many instances during the development of Drosophila
[21],
zebrafish [22],[23], and mice [24]. It was also
recently shown that human embryonic stem cells (hESCs) require activation of Notch
signaling to form the progeny of all three embryonic germ layers, and subsequent
transient Notch signaling enhanced generation of hematopoietic cells from committed
hESCs [25].
The mechanisms by which a short Notch signaling pulse can permanently switch cell
fate are not elucidated.
The Notch system has also been shown to function as an oscillator. Specifically, the
expression levels of members of the hes family, a group of
downstream Notch target genes [26], have been shown to oscillate with a 2 hour
periodicity in some systems during development, which for example aids in
somitogenesis (i.e. the patterning of somites) [27]–[29]. Hes1
protein and mRNA concentrations have also been observed to oscillate with an
approximate 2 hr time period upon serum starvation in various cultured cell lines
including myoblasts, fibroblasts, and neuroblastoma cells [30]. Furthermore,
oscillations in the Notch network have been proposed to be important in maintaining
neural progenitor cells in an undifferentiated state [31]. Finally, there is
evidence that such oscillations may also afford cells the opportunity to repeatedly
test for the continued existence of a signal [32], thereby increasing
cellular response sensitivity and flexibility by allowing the cell to integrate the
results of many periodical evaluations of the signal before making an ultimate cell
fate decision.
The Delta-Notch signaling system has been previously modeled to elucidate its role in
fine-grained pattern formation through the action of lateral inhibition and
induction [33]–[35]. Collier et
al. developed a simple 2-parameter model that focuses on pattern formation
due to feedback inhibition between adjacent cells via Delta-Notch signaling [33]. Other
models build upon this simple model by adding more molecular detail at the
intercellular level [34],[35]. In addition, several studies have focused on
trying to understand the underlying mechanism of Notch system oscillations [32],[36], where a
Hes1 negative feedback loop composed of Hes1 protein repressing
hes1 transcription, likely plays a central role [37].
Delays related to transcription and translation were also proposed to be important
for the observed oscillations [38]. However, while several models have thus been
proposed and have yielded important insights into this system [30], [36], [38]–[40], they
have focused exclusively on Hes1 and not analyzed its interactions with other
signaling proteins in the Notch system. Additionally, all these models focus on a
particular aspect or mode of Notch signaling (e.g. lateral inhibition or
oscillation) but do not yet address how complex, alternative behaviors could arise
from the same network.
Here we mathematically model the Notch signaling system to analyze how the same
network is capable of functioning as a cell fate switch or an oscillator in
different biological contexts. This model, which includes the regulation of the
notch1-RBP-Jk-hes1 gene circuit, predicts that the Notch1-Hes1
system acts as a bistable switch in certain regions of parameter space, where Hes1
levels can change by 1–2 orders of magnitude as a function of the input
Delta signal. In addition, it predicts that a transient pulse of a high level of
Delta is capable of inducing high Hes1 expression levels for a duration that would
be sufficient to induce a cell fate switch. Moreover, the model elucidates how the
network can be ‘tuned’ to function in different regimes, either
as an oscillator or a cell fate switch, by changing a key parameter. Finally, low
numbers of reactants can lead to significant statistical fluctuations in molecule
numbers and reaction rates, making cells intrinsically noisy biochemical reactors
[41],[42]. Stochastic
simulations of the Notch system, which enable the analysis of the effect of
biological noise in the system arising due to stochastic variations in gene
expression, reveal that for systems that respond quickly to Notch signaling, the
network is able to dampen the effects of this biological noise and function in a
manner similar to what is predicted by the deterministic model. In summary, the
model enables analysis of the different behavioral responses of the Notch signaling
network observed over a broad spectrum of signaling inputs and parameter values and
can be further expanded to study Notch signaling in numerous contexts.
We developed a model of Notch signaling to investigate how this system can function
as either an oscillator or as a simple binary switch capable of responding to steady
state or transient inputs. Brief experimental work revealed that the
notch1 promoter is positively upregulated by its gene product and
is downregulated by Hes1 (Text S1, Fig. S1). We thus examined the behavior of the
notch1, RBP-Jκ, and
hes1 genes, which form a complex set of regulatory feedback loops
(Fig. 1). A deterministic
model composed of a system of differential equations was developed to analyze
dynamic changes in the levels of the network constituents. However, since the
concentrations of some of the species were low, stochastic simulations were also
conducted to examine whether noise in the levels of the network components could
significantly impact system behavior, as noise has the potential to undermine the
fidelity of cell fate choices [41],[43].
A set of differential equations was developed to track changes in the
concentrations of various species in the nucleus and cytoplasm of a cell as a
function of time following activation of Notch by its ligand. The cell is
modeled as a 10 µm diameter sphere with a 5 µm diameter
nucleus. Numerous processes were modeled as terms in the differential equation
system, including transcription, translation, transport, degradation or - in the
case of Notch - receptor cleavage (Fig. 1A, Text S1). As examples, the three equations
tracking the Hes1 cytoplasmic mRNA, Hes1 cytoplasmic protein and Hes1 nuclear
protein concentrations are given by:The rate of change (in units of ) of the cytoplasmic mRNA concentration of Hes1 is given by the
difference in the rates of it transcription and degradation. RfHcm is
the transcription rate of hes1 mRNA in the nucleus. We assume
instantaneous export of mRNA to the cytoplasm. A factor of 7 is included to take
into account the dilution due to export to the cytoplasm (Text S1).
kdHcm, kdHcp, and kdHnp denote the
degradation constants for the hes1 mRNA (Hcm), cytoplasmic
protein (Hcp), and nuclear protein (Hnp), respectively, which are assumed to
undergo first order degradation kinetics. ktrHc denotes the
translation constant (min−1) for conversion of cytoplasmic
hes1 mRNA into cytoplasmic protein. Transcriptional and
translational delay times are incorporated into the model, as these are
processes that inherently involve delays between initiation and the production
of a molecule of mRNA or protein, as previously described [38],[44].
Thus, the translation of Hes1 protein is based on the delayed
hes1 mRNA concentration HcmD (delayed by time TpHc,
the average time for translation of Hes1), which is the concentration of mRNA
present when the process of translation was initiated instead of the
concentration at the present time. kniHcp denotes the nuclear import
rate in units of min−1. A dilution factor of 7 is again
used to incorporate differences in nuclear and cytoplasmic volumes.
The transcription rates for notch1, hes1, and
RBP-Jκ are based on the states of their respective
promoters. Previous promoter analysis has been complemented with Genomatix Suite
Gene2Promoter transcription factor (TF) binding site prediction software to
identify potential TF binding sites in the promoters of the three genes in the
model.
Takebayashi et al. [37] observed that hes1
transcription is repressed by its own gene product through Hes1 protein binding
to sites in the hes1 promoter termed N-boxes. Through a series
of binding and transcriptional activity assays, the study determined that Hes1
bound strongly to three N-boxes found upstream of the transcriptional start site
and repressed transcription of the hes1 mRNA up to 40-fold.
Also, while the work concluded that there was a synergistic rather than an
additive effect of the N-box binding dependent repression of gene expression,
further mathematical analysis has indicated that there is no or very weak
synergy among the different binding sites [45]. Several positive
regulatory regions were also found in the hes1 promoter, and it
was also shown to have two adjacent RBP-Jκ binding sites [46],[47]. Thus, we have
modeled the hes1 promoter to have three equivalent N-boxes
where the Hes1 protein can bind and repress transcription, as well as two
equivalent RBP-Jκ sites. The presence of all other positive regulators
of transcription is lumped into a constant basal rate of transcription.
As it has not been extensively investigated, the notch1 promoter
sequence was analyzed in the Gene2Promoter software. One putative Hes1 site
(N-box) and two putative RBP-Jκ sites were found in the ∼1 kb
notch1 promoter analyzed. This may imply that
notch1 is both positively and negatively regulated by its
own gene product. To test this, a transcriptional activity experiment using the
dual luciferase assay system was conducted. The promoter of murine
notch1
[48]
was used to drive expression of hRluc cDNA (Renilla luciferase). Co-transfection
studies with plasmids expressing RBP-Jκ, NICD, Hes1, and dNHes1 (a
dominant negative form of Hes1) indeed demonstrated that the
notch1 promoter is regulated negatively by Hes1 and
RBP-Jκ in the absence of NICD but is positively regulated by NICD in the
presence of RBP-Jκ (Text S1, Fig. S1).
The notch1 promoter was modeled with two RBP-Jκ sites
and one N-box.
A 418 bp sequence upstream of the RBP-Jκ gene as
characterized by Amakawa et al. [49] was analyzed in
the Gene2Promoter software for TF binding sites of interest. Three potential
Hes1 binding sites and three potential RBP-Jκ sites were found. Thus,
the RBP-Jκ gene also potentially undergoes
autoregulation under Notch signaling, and a three N-box, three RBP-Jκ
site model was utilized.
As discussed above, all three promoters have one or more binding sites for both
Hes1 (the N-box) and RBP-Jκ. It is assumed that Hes1 can bind and
repress transcription of the corresponding promoter only in its homodimer form,
and the dimerization reaction is assumed to be at steady state over timescales
of protein transcription, translation and import, driven by mass action kinetics
such that the concentration of the dimer is given by:Where, KaHp is the association equilibrium constant
for the dimerization reaction. Similarly, the time scales of transcription
factor binding to and dissociation from the promoter elements are also assumed
to be much faster than those of gene transcription and protein synthesis, such
that binding to the promoter is at pseudo steady state. In addition, it is
assumed that NICD can bind only when an RBP-Jk protein is bound to its site on
the promoter, and that this NICD binding converts RBP-Jκ from a
transcriptional repressor to an activator [13].
The level of promoter activation (i.e. rate of mRNA synthesis) is modeled by an
approach termed BEWARE [50],[51], in which the probabilities of a promoter
being in any one of its many possible states are calculated based on the
relative concentrations of the three transcription factors (Hes1, RBP-Jκ
and NICD), and their respective DNA binding affinities, using equilibrium
binding equations. The level of activation of the promoter is then given by: , where, P[Pi] is the probability
of the promoter being in state i, and vi is the activation rate of
gene transcription associated to the promoter being in that state i. When the
promoter is empty, the gene activation rate is assumed to be the basal
transcription rate (Vb) for that promoter. When a Hes1 dimer is bound
to an N-box, the rate is reduced by a factor rN that takes into
account the repressive effect of the Hes1 transcription factor, and when
RBP-Jκ is bound, the rate is reduced by a factor rR.
Furthermore, when the promoter is in its maximally activated state with the NICD
bound to the RBP-Jκ and no Hes1 dimers bound, the activation rate is
assumed to be at its maximum and is given by
(Vmax+Vb). In the case of multiple RBP-Jκ
binding sites, an additional factor tc (<1) is used to account
for states where not all RBP-Jκ sites bind NICD to represent the
decrease from the maximum possible activation rate. For a detailed expression of
transcription rates please refer to Supplemental Materials (Text S1).
Although explicit parameters have been included to account for cooperative
binding for Hes1 dimers to multiple N-boxes and for RBP-Jκ binding
(cooperativity factors Cn, Cr and Cnr - please
refer to Table 1 for model
parameters), they have been set to 1 for these simulations, as recent work
suggests there is very little if any cooperative effect in Hes1 binding to
N-boxes [45]. Finally, it is assumed that each mRNA produces a
fixed number of proteins, i.e. mRNA dynamics have been neglected [50].
Experimentally determined values for half-lives of proteins and mRNA, association
and dissociation constants of proteins to their respective DNA binding sites,
dimerization constants, and protein translation and transcription rates have
been used when possible (Table
1). These values are often not available for the exact species of
interest; however, the best available estimates based on similar protein classes
are used wherever applicable as the starting point. The time delays for
transcription and translation for each of the three genes are calculated as
previously described [52] and are detailed in the Supplemental
Materials (Text
S1). 4.5 transcripts per minute [45] and 20 transcripts
per minute [53] were used as initial estimates for
hes1 basal and maximum transcription rates respectively.
The transcription rates for RBP-Jκ and
notch1 were then determined from these estimates and the
estimates of their minimum transcription times (Text S1).
The degradation rates for the Hes1 protein and mRNA were determined
experimentally by Hirata et al. in fibroblasts [30]. They observed
similar values in other cultured cell types including myoblasts, neuroblastomas,
and teratocarcinomas. Pulse chase experiments of Logeat et al. [54]
were used to assess the degradation rates for the full-length Notch1 protein,
and an estimate of Notch1 protein half-life of ∼40 minutes was derived.
GSK3β has been shown to affect the stability of NICD [55].
Although there are conflicting results as to whether GSK3β helps to
stabilize [55] or destabilize the cleaved NICD [56],
our experimental results show that GSK3β is essential for the NICD
regulation of neural stem cell differentiation into astrocytes (Agrawal, Ngai,
and Schaffer, manuscript in preparation). Furthermore, we show that Notch1
signaling upregulates the expression of GSK3β in these cells. Thus, the
effect of GSK3β is incorporated into the model by increasing the
half-life of NICD from 3 to 8 hrs [55] above a threshold
concentration of Hes1 (which is assumed to directly or indirectly regulate the
expression of GSK3β). This increased NICD half-life does not however
change the qualitative behavior of the Hes1 switch (Fig. S3A).
The repression constant of Hes1 dimer bound to an N-box (rNbox) is
estimated from the results of Takebayashi et al. [37] that show that
in the presence of three N-boxes, transcription is repressed by ∼40
fold. This yields a repression value of ∼0.3 per N-box (Please refer to
Supplemental Materials (Text S1) for details). Since there are no
reliable estimates of the NICD generation constant upon Delta binding
(kfNcp), a lumped parameter of this constant with the Delta
concentration is used to report the strength of the Delta signal
(kfNcp*Delp). The initial parameters for which the
experimentally determined values are not accurately available were later
subjected to sensitivity analysis (See results).
The differential equations described in the model were solved (with parameter
values given in Table 1)
using Berkeley Madonna 8.3.11 software (www.berkeleymadonna.com) with the Runge Kutta 4 module at a step
size of 1 min. To arrive at realistic initial conditions for the model, the
initial concentrations of all species were set to 0 with zero Delta signal, and
the simulations were run until the various species attained steady state
concentration levels. These steady state values (listed in Table 2) were then used as
the initial conditions for subsequent simulations. For the various experiments,
the system was run for 750 minutes without stimulation with the Delta ligand to
attain a basal steady state, and the Delta concentration was then increased to
different levels to initiate Notch1 signaling. Simulations were run either with
a constant Delta signal throughout or with varying duration pulses of the Delta
signal. The system was simulated for a duration of 5,000–10,000
minutes (∼3.5–7 days), as neural progenitor stem cells have
been previously shown to undergo differentiation upon Notch activation in
3–5 days ([57]). Longer simulations up to 50,000 minutes
were conducted when required to confirm Hes1 had reached steady state levels.
Since the levels of several protein species in the deterministic model
simulations were very low (Table
2), at the level of tens of molecules per cell, assumptions of mass
action kinetics and pseudo steady state may not hold true, and stochastic
effects may play an important role in the dynamics of the signaling network
[41],[58]. To analyze whether
noise in protein and mRNA concentrations would impact the dynamics of the
system, a stochastic simulation of the model using the Gillespie algorithm [43]
was implemented in C++ (code available upon request). To relax
the assumptions of mass action kinetics and pseudo steady state, we explicitly
simulated every reaction step, making a total of 299 reactions. For example,
every interaction between a transcription factor and a promoter was modeled as a
discrete reaction in the simulation. The τ-leap method [59]
was also incorporated into the algorithm to accelerate the stochastic
simulations and increase their efficiency.
The response of the Notch1-RBP-Jκ-Hes1 system to a step change in an
input Delta signal was analyzed. Simulations were initiated using the steady
state levels of the different species in the absence of any external Delta (also
listed in Table 2), and at
t = 750 minutes a Delta signal was applied.
Fig. 2A demonstrates
that when a low input Delta stimulus is applied, the Hes1 concentration settles
to a correspondingly low steady state value. However, when the input Delta
signal was increased (10-fold), Hes1 shows a rapid increase to a new, 20-fold
higher steady state value. Further steady state analysis at a range of input
Delta levels and initial conditions reveals that the system exhibits
bistability. At low levels of Delta signal, basal levels of Hes1 are maintained
in the cell (“OFF” state), but as the Delta signal strength
is increased beyond a threshold level, it stimulates the production of Hes1,
which is then maintained at high levels (“ON” state) through
the concerted regulation of the Notch1-RBP-Jκ-Hes1 network (Fig. 2B). Bistability
– which has previously been proposed as an advantageous mechanism to
mediate an unambiguous cell fate switch, including in stem cells [51],[60]
– is evident within an intermediate range of Delta signal values
(Fig. 2B).
The initial numbers of some protein and mRNA species in the system were in the
range of tens of molecules per cell (Table 2), such that stochastic fluctuations
in individual species may impact the dynamics of the network. In particular,
intracellular noise inherent in systems with small numbers of molecules and/or
slow biochemical reactions can randomize or undermine the
“accuracy” of cell fate choices [41],[58]. To analyze such
behavior, stochastic simulations based on the Gillespie algorithm [43],
distinct from the deterministic model, were developed. Steady state analysis
shows that at low, constant Delta signals, the Hes1 levels fluctuate about a low
mean value corresponding to the “OFF” state, as expected
(data not shown). However, if the Delta signal is increased to a level just
below the concentration at which the deterministic model would predict a switch
in state (Fig. 2B),
stochastic simulations reveal that noise in the network can induce some
trajectories to spontaneously switch states (Fig. 3A). Analogous to results previously
observed in other systems [51],[61],[62],
noise thus undermines the bistable switch and induces spontaneous flipping
between states. Analysis of the time it takes the system to initially pass from
the lower to the upper state reveals that as the strength of the input signal is
increased, this average first passage time (FPT) decreases, and the percentage
of trajectories that change state increases (Fig. 3B). However, this
“uncertainty” occurs within a narrow range of intermediate
Delta signal levels, and if this intermediate window is avoided, the system
effectively behaves deterministically.
In addition, “ON” to “OFF” transitions
were simulated by first stimulating with a high Delta signal for 4000 minutes to
induce high Hes1 expression levels. When Delta was then reduced to levels that
were in the predicted bistable region based on the deterministic model, the
system maintained high expression levels of Hes1 (Fig. 4A), as anticipated from the
deterministic results (Fig.
2B). Contrary to what was expected based on the deterministic model,
however, when the Delta signal was instead reduced to zero, some trajectories
remained in the high Hes1 expression (“ON”) state (Fig. 4B). This indicates the
role of stochastics in potentiating high Hes1 expression levels even in the
absence of continued signal.
It has been shown for neural crest stem cells [20] that a transient
Notch signal is sufficient to induce cell differentiation. Also, there are
numerous situations where transient Notch-Delta signaling determines the fates
of immature cells, both in tissue culture [18],[19]
and during organismal development [21]–[24]. Under continuous Delta stimulation, the system
can attain high steady-state Hes1 expression levels, thus acting as a switch,
but we next wanted to examine whether transient Delta activation was also
capable of eliciting high Hes1 expression. We thus examined the dynamic response
of the system to transient activation of the Notch1 pathway upon variation in
the strength and duration of an applied Delta signal.
When the system is stimulated for a short duration (10 minutes) with a moderate
strength Delta signal, the deterministic model predicts a transient peak in the
Hes1 expression that eventually decays to its low steady state value (Fig. 5A). However, the peak
expression of Hes1 continually increases with increasing input signal duration
up to ∼800 minutes, beyond which the maximum expression levels of Hes1
attained remain the same but the duration of prolonged high expression levels
progressively increases (Fig.
5A). Similarly, as the input Delta signal strength is increased for a
constant pulse duration, the peak Hes1 concentrations attained also increase up
to a maximum value, after which a further increase in the signal strength only
increases the duration of high Hes1 levels (Fig. 5B). The cell is thus able to attain
high Hes1 expression either under prolonged low intensity Delta signaling or a
short burst of high intensity Delta signaling.
We also examined the effect of stochastics on transient activation of the
network. Simulations were run using the parameter values as in the deterministic
model for various Delta pulse durations ranging from 10 minutes to 3000 minutes
and >40 trajectories per input duration value were analyzed. For Delta
pulse durations of less than 500 minutes, the stochastic simulations followed
the prediction of the deterministic model (data not shown). However, for a
500-minute Delta pulse, even though the deterministic model predicts a transient
Hes1 peak that does not attain the maximum possible expression level, a small
percentage of the stochastic trajectories in fact did switch to the
“ON” state (corresponding to high Hes1 expression levels)
(data not shown). Also, as the duration of the Delta pulse is increased, the
percentage of trajectories that remain in the “ON” state for
the simulated 15,000 minutes progressively increases even though the
deterministic model predicts that the system would revert back to the
“OFF” state within that time. Furthermore, the average first
passage time (FPT) of the trajectories that do switch state increases as the
Delta pulse duration increases (Fig. 6). It is likely that for shorter Delta pulse durations, if the
system is to undergo the spontaneous “OFF” to
“ON” transition, it does so early, soon after the
application of the Delta signal. However, in the case of longer duration input
signals, the continued presence of the signal allows trajectories to switch
state even much later in the simulation, resulting in an apparently longer first
passage time. Collectively, these results imply that even for very short signal
pulse, a small fraction of a population of cells receiving a pulse of Delta
signal could switch their state due to stochastic effects.
A number of parameters in the model have not been directly experimentally
measured and were estimated from data available for similar protein classes in
different contexts, and we thus performed sensitivity analysis for all such
parameters by varying them individually through a broad range of values in the
deterministic model (Table
3, Fig.
S2). Although in most cases the qualitative behavior of the system
remained unchanged, the system did exhibit considerable sensitivity to specific
parameters, which were then subjected to further analysis. These include: the
half-life of NICD, the equilibrium binding constant of NICD with RBP-Jκ
(Ka), the maximal transcription rates (Vmax), and the
repression constant of Hes1 (rNbox). NICD has a long half-life of a
few hours under normal physiological conditions [55]. However, our model
indicates that if the NICD half-life is drastically reduced, the system fails to
function as a switch and cannot express high levels of Hes1 (Fig. S3).
In addition, the equilibrium binding constant (Ka) of NICD to
RBP-Jκ in the model is 108 M−1, but as
Ka increases – denoting stronger interactions of NICD
with the promoter – bifurcation analysis demonstrates that the OFF-ON
transition occurs at accordingly lower values of the Delta signal
(kDelp) (Fig.
7A). Similarly, increasing the maximal transcription rate of Hes1
(Vmaxh) to indicate a stronger promoter shifts the OFF-ON
transitions to lower Delta signal strengths (Fig. 7B).
Interestingly, the response of the deterministic model was most sensitive to the
extent to which Hes1 binding reduced or repressed expression of target genes
(rNbox). As the Hes1 repression constant (rNbox) is
progressively decreased (or the repressive strength of Hes1 progressively
increased) from 0.3 to 0.1, the final steady state concentrations of Hes1
progressively decrease for a given level of Delta signaling (Fig. S4),
but the system continues to exhibit bistability. Intriguingly, as the value of
rNbox is further decreased below 0.1, there is a dramatic
qualitative change in the response of the system. Specifically, the system
undergoes a bifurcation or transition from bistable to monostable behavior and
at such high repressive strengths is unable to attain high steady state Hes1
expression levels. Finally at very low values of rNbox
(<0.03), it once again undergoes a transition to a stable oscillatory
response where the Hes1 levels in the cell oscillate about a low mean steady
state value (Fig. 8A). A
phase plot of the response of the system with variable rNbox (Fig. 8B) demonstrates how the
same gene network can transition from behaving as a bistable switch to being an
oscillator. The model thus elucidates the versatility of the system, where
tuning of a single key parameter can convert its behavior from a switch to a
clock. Previous hes1 models showing sustained oscillations have
focused exclusively on the low rNbox region (i.e.
rNbox = 0) of such a phase plot
[30],[36],[38],[39].
The Notch signaling system is an evolutionarily conserved network that functions in
multiple organs to orchestrate cell fate specification [63]–[65] in a
context dependent manner. In some cases, it can function as a binary cell fate
switch at the individual cell level [16],[17], whereas in other
situations cell-cell contact dependent Notch signaling can result in pattern
formation in an array of cells [14],[15], and in yet other contexts it can function as a
biological clock to govern pattern formation and differentiation during
somitogenesis [27]–[29]. Although several
additional components such as Fringe, Numb, and Presenilin can feed into and
modulate the Notch signaling cascade, the core of the signaling pathway is
relatively simple, where Notch acts as a membrane bound transcription factor that is
activated by ligand binding and induces transcription of target hes
genes via its interaction with the RBP-Jk transcription factor
[10].
However, the system can exhibit complex inter-regulation of its components. A better
understanding of the functioning and regulation of this signaling system –
and in particular how it exhibits diverse behaviors in different contexts
– is valuable from a basic biology standpoint, in understanding how
misregulation of the Notch signaling pathway can underlie disease, and from
regenerative medicine viewpoint in therapeutic applications of stem cells.
Mathematical modeling can provide valuable insights into the behavior of this gene
regulatory circuit. Previous models have focused either on the level of cell-cell
interactions to simulate the levels of Notch and Delta within adjacent cells and
thereby analyze pattern formation based on levels of Delta and Notch levels in an
array of cells [33]–[35], or on the autoregulation
of the hes genes in isolation to examine the oscillatory behavior
of the gene circuit [30],[36],[39],[40],[44],[45],[66],[67]. Here we have developed an integrative model that
takes into account the intracellular signaling network downstream of Notch
activation through its ligand Delta, leading to the activation of the
hes1 gene via interaction with RBP-Jκ. These three genes
potentially regulate the transcription of one another (Text S1, Fig S1) [37],[46],[47], forming
a network of positive and negative feedback loops (Fig. 1B). Our model begins to elucidate how a
cell can potentially tune key system parameters in the resulting Notch1-Hes1 gene
circuit to elicit diverse responses.
The behavior of the system was most sensitive to the repression constant of Hes1,
rNbox. The degree of Hes1 repression of a transcriptional target can
be modulated by the presence of co-factors. For example, whereas Groucho can act as
a transcriptional co-repressor for Hes1, Runx2 can act as a negative regulator of
the repressive activity of Hes1 by interfering with the interaction of Hes1 with the
TLE corepressors [68]. The repressive activity of Hes1 can also be
further potentiated by its interaction with the winged-helix protein brain factor 1
[69].
Therefore, because different cells can express these factors to different extents,
which can thereby modulate the value of rNbox, the same gene circuit can
be tuned to transduce an input Delta signal into qualitatively different responses
– oscillation vs. switching.
The model predicts that for low repressive strengths of Hes1
(0.1<rNbox<0.3), the Hes1 expression level functions as a
bistable switch in response to varying the strength of the Delta signal, thereby
providing an unambiguous fate switch that is insensitive to the presence of small
fluctuations in input signal (Fig.
2). Hysteresis has been previously observed experimentally in other
biological systems including the JNK signaling cascade [70],[71] and the Cdc2 cell
cycle regulation [72]. Parameters such as Ka (the
association binding constant of NICD to RBP-Jκ) and Vmax (the
maximal transcription rates) can shift the region of bistability, thus changing the
sensitivity of the system to the Delta signal, but the qualitative nature of the
gene network remains the same for a broad range of these parameter values. Positive
feedback loops with nonlinearity can yield bistability [51], and both Notch1
autoregulation and NICD-mediated conversion of RBP-Jκ into a transcriptional
activator that in turn upregulates Notch1 expression constitute positive feedback
loops that can drive this behavior.
Since the numbers of some protein and mRNA species in the model were low (Table 2), we developed a
stochastic model to examine the effect of biological noise and cell-to-cell
variability on the bistable response of the system to Delta signaling. Spontaneous
OFF to ON switching of states was observed even in regions not predicted by the
deterministic model. For example, as the Delta values are increased through the
bistable range, the percentage of trajectories switching to the ON state increases,
and the average FPT for these trajectories decreases (Fig. 3B). These results are consistent with
observations in other bistable systems [73], and computationally in
other signaling systems [51], where noise has been shown to cause spontaneous
switching of states. However, since the timescale of a system's downstream
response to the Notch network's state varies from a few hours (for example
during somitogenesis) [74] to a few days (for example during stem cell
differentiation) ([57]), the impact of stochastic noise on the fate
switch will also be different in different contexts. Thus, for very low Delta
signals, the average FPT is sufficiently high (>110 hrs) such that the cell
remains in the OFF state for prolonged periods of time and would be non-responsive
to Delta signaling over timescales of a few hours, whereas in the case of a
population of cells experiencing Notch signaling over a period of 4–6
days, spontaneous switching could undermine the genetic switch and cause some cells
to change fate at these low Delta input signals.
While the system can behave as a switch in a particular range of parameters at steady
state, there are also many situations in which Notch signaling is transient, yet is
sufficient to induce a switch in cell fate [18]–[24].
To simulate this, the model behavior was analyzed under transient Delta activation.
The network response to a transient Delta stimulus was a strong function of both the
signal intensity and duration, and either a high intensity signal for a short
duration or a low intensity signal for a prolonged duration was capable of inducing
transient increase in Hes1 expression levels for up to 2.5 days after withdrawal of
the signal (Fig. 5), a time
sufficient to initiate a biological response [57].
This prolonged expression of Hes1 upon transient Delta activation is due to the long
half-life of NICD [55]. The bistable switch is thus sensitive to the
degradation constant of NICD. If the NICD half-life were for example drastically
reduced, the model would predict that the system would fail to express high levels
of Hes1 regardless of Delta levels (Fig. S3). Hes1 is a repressive transcription
factor that in some systems plays a crucial role in suppressing the activation of
oncogenes. For example, in breast cancer cells, Hes1 can inhibit both estrogen- and
heregulin-beta1-stimulated growth via downregulation of E2F-1 expression [75]. Thus,
a malfunction in the Notch system, such as a reduction in NICD half-life, could
contribute to cell transformation. Indeed, aberrant Notch signaling is implicated in
many cancers (reviewed in [76]). For example, integrin-linked kinase (ILK),
which is either activated or overexpressed in many types of cancers including breast
cancer [77], can remarkably reduce the protein stability of
Notch1 and thus decrease its half-life drastically [78]. Interestingly, high ILK
and low NICD levels are detected in basal cell carcinoma and melanoma patients [78].
By increasing the repressive strength of the Hes1 dimer by 10-fold
(rNbox<0.03), the cell can transition from being a bistable
system, to a brief region of monostability, and finally to an oscillator (Fig. 8B). Oscillations occur with
a time period of approximately 2 hrs, similar to what Hirata et al. observed in cell
culture [30]. This value also compares well with the various
models that have been developed (for the Hes system in isolation) to explain
oscillations in the hes family of genes and their homologues. These
models assume complete repression in the presence of even a single Hes homodimer
bound to the promoter region [36],[39],[45],[66]. This corresponds to an rNbox value of
0, in which case there would be no difference between the repressive strength of
promoters with 1, 2 or 3 N-boxes. From the experimental observations of Takebayashi
et al. [37], where the repressive strength of the promoter did in
fact increase with the number of N-boxes, the estimated value of rNbox is
0.3. However, during somitogenesis, the factors expressed in the presomitic mesoderm
(PSM) may enhance the repression due to Hes1 such that the value of rNbox
is very low.
This current model represents the Notch signaling network core in a single cell, and
it can readily be extended to a field of cells to analyze the role of Notch in
patterning tissue formation [60]. In addition, there are numerous cell-specific
mechanisms and factors that feed into this important signaling core [79]–[81]. Additional molecular
species can be added to this model framework, or the parameter values of the current
model can readily be modulated for example to simulate changes in DNA binding
affinities, repressive constants, or the protein and mRNA stabilities as a function
of cell-specific factors. This simple but versatile model can therefore be expanded
by incorporation of additional molecular mechanism, specific to particular cell
types, to make predictions on the role of Notch signaling in diverse cells and
tissues.
In summary, we have theoretically and computationally analyzed the
Notch1-RBP-Jκ-Hes1 signaling network, which is responsible for cell fate
specification in numerous contexts. Our results indicate that the network,
consisting of both positive and negative feedback mechanisms, can be tuned to
function either as a bistable cell fate switch or an oscillator based on relatively
small changes in a key parameter value. Furthermore, the duration and strength of
the Delta signal regulate either the peak or the final steady state levels of Hes1
attained. Therefore, cells can readily tune the Notch system to regulate a variety
downstream cell fates and functions.
|
10.1371/journal.ppat.1002412 | Sequential Adaptive Mutations Enhance Efficient Vector Switching by Chikungunya Virus and Its Epidemic Emergence | The adaptation of Chikungunya virus (CHIKV) to a new vector, the Aedes albopictus mosquito, is a major factor contributing to its ongoing re-emergence in a series of large-scale epidemics of arthritic disease in many parts of the world since 2004. Although the initial step of CHIKV adaptation to A. albopictus was determined to involve an A226V amino acid substitution in the E1 envelope glycoprotein that first arose in 2005, little attention has been paid to subsequent CHIKV evolution after this adaptive mutation was convergently selected in several geographic locations. To determine whether selection of second-step adaptive mutations in CHIKV or other arthropod-borne viruses occurs in nature, we tested the effect of an additional envelope glycoprotein amino acid change identified in Kerala, India in 2009. This substitution, E2-L210Q, caused a significant increase in the ability of CHIKV to develop a disseminated infection in A. albopictus, but had no effect on CHIKV fitness in the alternative mosquito vector, A. aegypti, or in vertebrate cell lines. Using infectious viruses or virus-like replicon particles expressing the E2-210Q and E2-210L residues, we determined that E2-L210Q acts primarily at the level of infection of A. albopictus midgut epithelial cells. In addition, we observed that the initial adaptive substitution, E1-A226V, had a significantly stronger effect on CHIKV fitness in A. albopictus than E2-L210Q, thus explaining the observed time differences required for selective sweeps of these mutations in nature. These results indicate that the continuous CHIKV circulation in an A. albopictus-human cycle since 2005 has resulted in the selection of an additional, second-step mutation that may facilitate even more efficient virus circulation and persistence in endemic areas, further increasing the risk of more severe and expanded CHIK epidemics.
| Since 2004, chikungunya virus (CHIKV) has caused a series of devastating outbreaks in Asia, Africa and Europe that resulted in up to 6.5 million cases of arthritic disease and have been associated with several thousand human deaths. Although the initial step of CHIKV adaptation to A. albopictus mosquitoes, which promoted re-emergence of the virus, was determined to involve an E1-A226V amino acid substitution, little attention has been paid to subsequent CHIKV evolution after this adaptive mutation was convergently selected in several geographic locations. Here we showed that novel substitution, E2-L210Q identified in Kerala, India in 2009, caused a significant increase in the ability of CHIKV to infect and develop a disseminated infection in A. albopictus. This may facilitate even more efficient virus circulation and persistence in endemic areas, further increasing the risk of more severe and expanded CHIK epidemics. Our findings represent some of the first evidence supporting the hypothesis that adaptation of CHIKV (and possible other arboviruses) to new niches is a sequential multistep process that involves selections of at least two adaptive mutations.
| The potential of RNA viruses to emerge into new environments often depends on their ability to efficiently adapt to new hosts. These adaptations sometimes comprise a stepwise process that includes 1) initial viral introduction/establishment in the recipient species, followed by 2) finite adjustment/optimization of the virus replication and transmission strategies for specific environments associated with a new host [1], [2]. This process has been well documented for several single-host viruses such as pandemic influenza A virus, the SARS coronavirus and canine parvovirus (reviewed in [3], [4]) that do not rely on alternating infection of disparate hosts for their maintenance in nature. However, much less is known about the adaptive processes that mediate cross-species jumps for double-host viruses such as arthropod-borne viruses (arboviruses). Several recent studies documented that the acquisition of a single mutation in an arbovirus genome can mediate their cross-species transfer [step (1)] [5]–[8]. However, in none of these cases have subsequent, additional adaptive mutations been detected, posing the question of whether selection of second-step adaptive mutations is possible or necessary for these viruses to persist in nature. This information is critical for understanding and predicting the long-term consequences of pathogen emergence and maintenance in affected areas, which in turn could influence the development and success of targeted intervention strategies for managing outbreaks.
A new lineage of Chikungunya virus (CHIKV) [arbovirus in family Alphavirus, genus Togaviridae] emerged in 2004 in Kenya and subsequently spread into many countries in the Indian Ocean basin [hence the name: Indian Ocean lineage (IOL)], causing devastating outbreaks of arthritic disease [9]. In India, IOL strains were first detected in December 2005 followed by extensive geographic expansion during 2006–2011 into 19 Indian states with a total number of human cases estimated in 2007 at between 1.4 and 6.5 million [10], [11]. During 2006, the states most affected by CHIKV were Karnataka and Maharashtra, with a subsequent shift to Kerala, Coastal Karnataka and West Bengal [12], [13]. Several hypothetical factors may have contributed to the CHIKV emergence/spread on the Indian subcontinent [14], including: 1) the use of immunologically naïve human populations for maintenance, amplification and virus dispersal among localities, 2) reliance on peridomestic and anthropophilic mosquitoes as vectors, and 3) the IOL-specific genetic predisposition for rapid adaptation to Aedes (A.) albopictus, which was previously considered only a secondary CHIKV vector [9].
The mode of CHIKV maintenance in nature is complex and appears to be region-specific. In Africa, CHIKV is maintained in enzootic cycles involving transmission between non-human primates and canopy-dwelling, primatophilic Aedes mosquitoes, primarily A. furcifer, A. taylori, A. africanus, A. luteocephalus and A. neoafricanus [15]–[19]. In contrast, CHIKV transmission in Asia is believed to rely on humans alone as reservoir/amplification hosts, with the domestic A. aegypti and to lesser extent the peridomestic A. albopictus serving as primary urban mosquito vectors [19], [20]. Recent evidence, however, suggests the possibility of additional sylvatic, zoonotic transmission cycles [21], [22].
In India, both urban CHIKV vectors are present, although their distributions differ, and their epidemiologic significance for CHIKV transmission probably varies locally. A. aegypti was considered to be the most important during the early phase of the CHIK epidemic in 2006 [23]. However, in subsequent years (2007–2009), the involvement of A. albopictus as the principal vector was documented at least in the states of Kerala and Coastal Karnataka [24]–[27]. Interestingly, CHIKV transmission by A. albopictus was shown to be associated with the acquisition of the A226V amino acid substitution in the E1 envelope glycoprotein [24], [28]–[32] (Figure S1) that is responsible for alphavirus virion assembly and virus fusion in endosomes of target cells [33]–[35]. The role of the E1-A226V substitution on CHIKV adaptation to A. albopictus was directly demonstrated in laboratory studies, including those using reverse genetics, showing that this mutation is directly responsible for increased CHIKV infection, dissemination and transmission by this vector species [6], [36]. In India, evidence that CHIKV was undergoing genetic adaptation to A. albopictus via the E1-A226V substitution first came from Kerala State. During 2006, only the E1-226A variant was recovered there; however, during subsequent years (2007–2008), all isolates sequenced possessed the E1-226V residue [24] (Figure S1). In 2008 the E1-A226V substitution was also found among the majority of CHIKV isolates from Coastal Karnataka, adjacent to Kerala [37], suggesting introduction from the latter state.
In a follow-up study conducted in the state of Kerala, a novel substitution in the E2 envelope glycoprotein, L210Q, was discovered in all human and mosquito CHIKV isolates collected during 2009 [27] (Figure S1). The E2 protein is located on the tips of alphavirus spikes and interacts with host cell receptors as well as with neutralizing antibodies [38], [39]. The L210Q substitution has not been reported in any other CHIKV strains, including those isolated in Kerala State during 2006–2008. This suggests that E2-L210Q substitution was selected as a result of CHIKV adaptation to specific ecological conditions present in Kerala State. Position E2-210 is located in the domain B of the E2 glycoprotein [39], and several earlier studies demonstrated that mutations in this domain mediate host specificity of several alphaviruses [5], [7], [40]–[42] as well as the selection of escape mutants by neutralizing antibodies [43]–[45]. Moreover, we recently demonstrated that epistatic interactions between mutations at positions E1-226 and E2-211 of CHIKV influence the penetrance of the E1-226V substitution for fitness in A. albopictus [46]. The E2-I211T substitution was probably acquired by IOL CHIKV strains around 2004–2005 [47], and provides a suitable background to allow CHIKV adaptation to A. albopictus via the subsequent E1-A226V substitution.
Considering that A. albopictus was a principal CHIKV vector in the state of Kerala in 2009, it was hypothesized that the novel substitution E2-L210Q provided an additional selective advantage for CHIKV transmission by this mosquito [27]. To test this hypothesis we undertook a comprehensive reverse genetic analysis of the effects of E2-L210Q in various CHIKV hosts. Our observations demonstrate that the E2-L210Q substitution mediates a significant increase in CHIKV dissemination in A. albopictus by increasing initial infectivity for midgut epithelial cells. In addition, we show that the E1-A226V substitution has a significantly stronger effect on CHIKV fitness in A. albopictus than E2-L210Q, probably explaining the observed time differences required for selective sweeps of these mutations in nature.
To investigate the effect of the E2-L210Q substitution on CHIKV fitness in A. albopictus mosquitoes, we employed a reverse genetics approach based on the SL-CK1 strain of CHIKV (hereafter abbreviated SL07), isolated in 2007 in Sri Lanka [9]. Previous phylogenetic studies indicated that SL07 evolved from the Indian subgroup of IOL and represents one of the most closely related isolates to strains responsible for CHIKV outbreaks in India (including the Kerala state) [9], [48]. The SL07 isolate was passed only twice on Vero cells since its isolation from a febrile patient, thus limiting the potential for cell culture-adaptive mutations that can artificially influence alphavirus fitness in vertebrate and/or mosquito hosts. The SL-07 strain has an alanine residue at E1 position 226 and a leucine residue at E2-210, corresponding to prototype IOL strain introduced into India in late 2005. Since the E2-L210Q substitution was only detected in CHIKV strains form Kerala that had previously acquired the A. albopictus-adaptive E1-A226V substitution [24], single E1-A226V and double (E1-A226V and E2-L210Q) substitutions were introduced into an infectious clone (i.c.), generated from the SL07 strain using site-directed mutagenesis. In addition, a clone with the single E1-A226V substitution (SL07-226V) was genetically marked by introducing a synonymous mutation 6454A→C that creates an ApaI restrictase site (SL07-226V-Apa). Previously we demonstrated that the 6454A→C substitution does not influence CHIKV fitness in vitro or in vivo [6]. The infectious viruses SL07-226V-Apa and SL07-226V-210Q were rescued by electroporation of in vitro-transcribed RNA into Vero cells. The specific infectivity and viral titers in cell culture supernatants were almost identical for all constructs (Table S1), indicating that the introduced mutations did not attenuate CHIKV in Vero cells.
Although a variety of mechanisms may be involved, adaptation of arboviruses for enhanced transmission by mosquitoes is typically expected to result in an increased ability to develop a disseminated infection leading to salivary gland infection. To investigate the effect of the E2-L210Q substitution on CHIKV fitness in A. albopictus mosquitoes, direct competition experiments were performed using SL07-226V-Apa and SL07-226V-210Q viruses (Figure 1). For these experiments, A. albopictus (Thailand strain) was presented with blood meals containing a mix of 5x105 plaque forming units (pfu)/mL of SL07-226V-Apa and 5x105 pfu/mL of SL07-226V-210Q viruses (combined titer 106 pfu/mL) and 10 days post-infection (dpi), the presence of disseminated viral infection as sampled from individual mosquito legs and heads was analyzed as described in the Materials and Methods. The dissemination of the SL07-226V-210Q in the Thailand colony of A. albopictus was 4.3 times more efficient compared to SL07-226V-Apa (Figure 1A)(p = 0.021), supporting the hypothesis that glutamine at E2-210 was selected in CHIKV population in Kerala State due to its positive effect on CHIKV transmission. To corroborate these findings, the ApaI site was introduced into the backbone of SL07-226V-210Q, and the resultant virus (SL07-226V-210Q-Apa) was tested in direct competition in A. albopictus (Thailand colony) against SL07-226V that was produced by reverting the ApaI site in SL07-226V-Apa to the wild-type (w.t.) nucleotide sequence (Figure 1B). The dissemination of SL07-226V-210Q-Apa was 3.4 times higher than that of SL07-226V (p = 0.017), indicating that the genetic marker was not responsible for the competition outcome, and supporting the role of the E2-L210Q substitution in increased CHIKV dissemination in A. albopictus. To demonstrate that the outcome of competition experiments was not affected by CHIKV propagation in Vero cells (which were used to identify mosquitoes with disseminated infection, prior to CHIKV genotype analysis), these cells were infected at a multiplicity of infection of 0.1 pfu/cell in triplicate with 1∶1 mixtures of viruses that were used in mosquito competition experiments. At 2 dpi, cell culture supernatants were collected for viral RNA extraction and processed as described in the Materials and Methods. No detectable differences in viral fitness (changes in the ratios of the 2 viruses) were observed after Vero cell passage, indicating that E2-L210Q substitution does not affect CHIKV fitness in these cells (Figure S2).
Early studies of CHIKV competence to infect A. albopictus demonstrated significant variation in susceptibility among different geographic strains of this mosquito [49]. To demonstrate that the effect of the E2-L210Q substitution on CHIKV dissemination in A. albopictus was not limited to a particular geographic strain, we also compared dissemination efficiency of the SL07-226V-Apa versus SL07-226V-210Q viruses in mosquitoes derived from Galveston, USA. Similar to the results with Thailand mosquitoes, the E2-L210Q substitution provided a mean 4.5-fold increase in dissemination efficiency of CHIKV (p = 0.003) (Figure 1A). These data suggest that the E2-L210Q substitution would likely have a similar effect on CHIKV fitness in A. albopictus from Kerala State, India, and from other parts of the world.
To investigate if the fitness advantage associated with the E2-L210Q substitution is sufficient for selection of mutant viruses in a w.t. CHIKV population, the SL07-226V-210Q was serially passaged in the presence of a 100-fold excess of SL07-226V-Apa (surrogate “wild-type” virus) in an alternating cycle between A. albopictus mosquitoes and Vero cells. To initiate the cycle, A. albopictus (Galveston colony) were presented with blood meals containing 5x105 pfu/mL SL07-226V-Apa and 5×103 pfu/mL SL07-226V-210Q viruses (100-fold excess of SL07-226V-Apa). After three consecutive passages, heads and legs of individual mosquitoes were processed as described above to determine if selection of virus with E2-L210Q substitution had occurred (Figure 2A). Despite being present in 100-fold lower quantity in the initial virus population, the SL07-226V-210Q virus alone was detected in 31.6% of mosquitoes after 3 alternating mosquito-Vero passages, whereas SL07-226V-Apa (w.t.) alone was found in 52.6% of mosquitoes, while 15.8% of mosquitoes had both competitors in their heads and legs (Figure 2B) (p = 0.227 one-tailed McNemar test). These data indicate that the E2-L210Q substitution has the potential to be selected in CHIKV populations in locations where A. albopictus serves as the primary vector. The 31-fold increase over 3 artificial transmission cycles in the relative frequency of SL07-226V-210Q over its initial ratio in blood meals corresponds to a ca. 3-fold increase per cycle, which is in agreement with fitness advantage of the E2-L210Q substitution observed earlier in direct competition experiments (Figure 1).
Historically, A. aegypti mosquitoes were the primary vector of CHIKV in Asia [19], [20], and this species still plays a significant role in CHIKV transmission in India [23], [50]–[52]. To investigate if the E2-L210Q substitution also affects CHIKV fitness in A. aegypti, we analyzed the effect of E2-L210Q on CHIKV dissemination in this vector using competition experiments as described above. Since CHIKV transmission by A. aegypti has never been associated with the E1-A226V substitution, we first used the w.t. genetic background of the SL07 strain (E1-226A and E2-210L) to introduce the E2-L210Q substitution. Also, because A. aegypti is less susceptible to CHIKV than A. albopictus for strains with E1-226V, we used higher total oral doses up to 2.4×107 pfu/mL (Figure 3A, 3B). The dissemination efficiency of SL07 and SL07-210Q-Apa viruses in A. aegypti (Thailand strain) were almost identical (p = 0.395) (Figure 3A). Similarly, no difference in the dissemination efficiency between SL07 and SL07-210Q-Apa viruses was detected in Galveston A. aegypti mosquitoes (Figure S3). Additionally, competition between SL07-226V-210Q-Apa and SL07-226V viruses, which express E2-210Q and E2-210L residues in the background of E1-226V, respectively, was also analyzed in Thailand A. aegypti, again revealing no statistically significant differences in dissemination efficiency [p = 0.402] (Figure 3B). These data indicate that it is unlikely that the polymorphism at E2-210 affects CHIKV transmission by A. aegypti.
Alternatively, E2-L210Q could have been selected as a result of CHIKV adaptation to a vertebrate host in India, probably humans. Although we did not observe any fitness change associated with this mutation during propagation of CHIKV in Vero cells (derived from African green monkey kidneys), to extend our analysis we repeated competition experiments using the human-derived cell line 293 (embryonic human kidney) because earlier studies showed that CHIKV can infect and replicate in various primary human cell lines including epithelial, endothelial, fibroblast, muscle satellite and macrophages (reviewed in [11]). No detectable difference in fitness resulting from the E2-L210Q substitution was observed in this cell culture model (Figure 3B, 3C). Although cell lines are not ideal surrogates for in vivo infections, our data further support the conclusion that the E2-L210Q substitution was most likely selected only by A. albopictus.
Previous studies determined that the A. albopictus-adaptive E1-A226V substitution acts primarily at the level of midgut infectivity. It was suggested that efficient CHIKV infection of and replication in midgut cells promotes more active CHIKV dissemination and transmission by this vector [6], [36], [46], thus allowing the selection of A. albopictus-adapted CHIKV strains in nature. To explore which step during CHIKV infection of A. albopictus mosquitoes is affected by the E2-L210Q mutation, we first compared the relative ratios of SL07-226V-Apa and SL07-226V-210Q RNAs in whole mosquitoes, mosquito midguts and mosquito carcasses (bodies without midguts) after oral infection (Figure 4A). We observed a marked increase in the relative amount of E2-210Q RNA in all samples analyzed, including midguts at 7 dpi. Furthermore, similar increases in the relative amount of E2-210Q RNA in mosquito midguts were observed as early as day 1 post infection, regardless of which of the two competitors was marked by the ApaI site (Figure 4B, 4C). In contrast, no difference in the relative amounts of E2-210Q versus E2-210L RNAs were observed 2 days after intrathoracic infection of A. albopictus, when CHIKV titers peak (Figure 4D). When injected intrathoracically, CHIKV does not require infection of and replication within mosquito midguts to disseminate to other organs and tissues, suggesting that the initial infection/replication of the midgut epithelium is a major site of selection of the E2-L210Q substitution in A. albopictus.
To further test the hypothesis that the E2-L210Q substitution affects CHIKV fitness only during initial infection of the A. albopictus midgut, we first compared infection rates of mosquitoes presented orally with serial dilutions of the viruses expressing either E2-210L or E2-210Q residues in the backbone of the SL07 strain that has the E1-A226V substitution. The E2-210Q residue was associated with significantly higher infectivity (p = 0.006 and p = 0.034, Fishes exact test) for A. albopictus (Thailand) at the blood meal titers of 3.5 and 2.5 Log10(pfu)/mL, and the oral infectious dose 50% (OID50) value calculated for SL07-226V-210Q was 8.9 times lower (higher infectivity) than that for SL07-226V (Figure 5A, 5B). The lack of a significant difference in infection rates after the highest dose (4.54 Log10(pfu)/mL) probably reflected the oral dose nearing saturation. By way of comparison, earlier studies, including those using the SL07 strain, determined that the well-established A. albopictus-adaptive substitution E1-A226V mediates a much greater, ∼100-fold decrease in OID50 values [6], [9].
To directly study the effect of E2-L210Q substitution on initial CHIKV infection of A. albopictus midgut cells, we developed a replicon/helper system for the SL07 strain. Sub-genomic replicons of alphaviruses can be packaged into virus-like particles (VLPs) by co-transfection of replicon and helper RNAs into susceptible cells [53]. The helper RNA provides the structural genes that package replicon RNA into VLPs, but the helper RNA itself is not packaged into the VLPs. Therefore, the VLPs are capable of primary infection and replicon RNA replication within cells, but cannot spread to neighboring cells due to the lack of the structural genes in the replicon. Thus, replicon VLPs allowed us to investigate the effect of mutations of interest on initial infection of midgut cells.
Since transfection efficiency of viral RNA is critical in determining the efficiency of VLP production, we switched to BHK-21 cells that have superior RNA susceptibility compared to Vero cells. Earlier, we observed that CHIKV isolates that have not been passaged in rodent-derived cells lines (including SL07) are impaired in their ability to replicate in BHK-21 cells (KT, SCW, unpublished). Therefore, to ensure efficient recovery of CHIKV VLPs from BHK-21 cells, double BHK-adaptive substitutions (nsP1-L407P and nsP3-T348A) were introduced into the SL07 i.c. (see Materials and Methods for details). Although these substitutions increase replication capacity, rather than electroporation efficiency, of CHIKV in BHK-21 cells, they have no effect on mosquito infection (data not shown). The modified SL07 i.c (contains nsP1-L407P and nsP3-T348A substitutions) was subsequently used to generate all CHIKV replicons used in the mosquito infectivity study.
The SL07 replicon expressing green fluorescent protein (GFP) was packaged into VLPs using w.t. SL07 helper (with E2-210L and E1-226A residues) or using a modified helper encoding E2-L210Q and E1-A226V substitutions. The SL07 replicon expressing cherry fluorescent protein (CFP) was packaged into VLPs using a helper encoding E2-210L and E1-226V residues (Figure 6A). In addition, the ApaI marker was introduced into the GFP-expressing replicon. The infectious titers of all recovered VLPs, as determined by titration on Vero cells, were identical (Figure 6A). Infection of Vero and C6/36 cells with 1∶1 mixtures of GFP and CFP expressing VLPs [based on infectious unit (i.u.) titers] yielded equal number of cells expressing these fluorescent proteins (data not shown).
For mosquito experiments, GFP- and CFP-expressing VLPs were mixed 1∶1 (based on i.u. titers) and presented in blood meals to A. albopictus as shown in (Figure 6A). At 1 and 2 dpi, midguts of individual mosquitoes were dissected and analyzed by fluorescent microscopy to determine a number of cells expressing GFP and CFP in the same fields of vision (Figure 6B). We found that, on average, midgut cells were 4–5 times more likely to become infected with VLPs expressing the E2-210Q residue as compared with VLPs expressing E2-210L (Figure 6D, 6E). Similarly, 4–5 fold increases in relative amounts of E2-210Q RNA were observed after an ApaI digestion of RT-PCR amplicons derived from VLP-infected midguts (Figure S4). Infectious viruses were not recovered after infecting Vero cells with homogenates of 30 mosquitoes infected with VLPs mixes (see Materials and Methods for details), indicating that the hypothetical formation of full-length viruses via recombination between helper and replicon RNAs, which could confound the interpretation of this experiment, did not occur. Altogether, these data demonstrate that the E2-L210Q substitution acts specifically by increasing initial CHIKV infectivity for midgut cells of A. albopictus.
In a parallel experiment using VLPs, we also compared the effect of the previously characterized E1-A226V substitution on CHIKV infectivity for midguts of A. albopictus (Figure 6A). The CFP-expressing replicon packaged using a helper encoding E2-210L and E1-226V residues (CFP/E2-210L/E1-226V) was competed against GFP-expressing replicon packaged using w.t. SL07 helper encoding E2-210L and E1-226A residues (GFP/E2-210L/E1-226A). In contrast to the polymorphism at E2-210, the valine residue at position E1-226 provided a far greater (41-43 fold) increase in a midgut cell infection compared to the alanine residue at the same position (Figure 6C, 6D, 6E), which agrees with previous results using infectious viruses [6], [36]. These data also indicate that the results of experiments using VLPs with 2 different fluorescent reporter proteins (GFP and CFP) cloned into replicons RNAs are not influenced by those reporter proteins themselves. The significant difference of ∼10-fold between the effects of the polymorphisms at positions E1-226 versus E2-210 on CHIKV infectivity (p = 0.026 and p = 0.005 for 1 and 2 dpi respectively) (Figure 6D, 6C) indicates that the E1-A226V substitution exerts significantly stronger selection compared to E2-L210Q, and thus would be expected to be selected faster during CHIKV transmission by A. albopictus.
To corroborate these findings we also analyzed effect of the E2-L210Q substitution on CHIKV infectivity for midgut cells of A. albopictus when this substitution is expressed in the background of w.t. CHIKV with the E1-226A residue. For this experiment, a GFP-expressing replicon was packaged using a w.t. SL07 helper encoding E2-210L and E1-226A residues (GFP/E2-210L/E1-226A), and was competed against a CFP-expressing replicon packaged using a helper encoding E2-210Q and E1-226A residues (GFP/E2-210Q/E1-226A). The E2-L210Q substitution caused a 2.3–2.4-fold increase in CHIKV infectivity for A. albopictus midgut cells (Figure 7), which was about 17.5 times weaker than the effect of the E1-A226V substitution in the same genetic background. Similarly, using direct competition experiments between infectious viruses SL07 and SL07-210Q-Apa (both have the E1-226A residue) we observed that the E2-L210Q substitution provided a mean 2.0-fold increase in dissemination efficiency of CHIKV (p = 0.022) (Figure S5) in the Thailand strain of A. albopictus. These data indicate that the E2-L210Q substitution would be selected more efficiently in CHIKV strains that previously acquired the E1-226V mutation.
In this study we showed that an E2-L210Q substitution recently identified in CHIKV populations of Kerala State, India, when expressed in the background of the initial adaptive E1-226V substitution, confers a selective advantage by increasing initial infection of A. albopictus midgut epithelial cells. Efficient infection of midguts promotes subsequent CHIKV dissemination into the hemocoel and transmission by this vector. However, the E2-L210Q substitution has no apparent effect on CHIKV fitness in the other primary mosquito vector, A. aegypti, or on fitness in cell culture models for primate infection (Vero and 293 cells). These results as well as surveillance data indicating that CHIKV was transmitted primarily by A. albopictus in Kerala state of India when the E2-L210Q substitution was first detected [24]–[27], provide a comprehensive evolutionary explanation for its appearance in 2009. These results also indicate that adaptation of CHIKV to A. albopictus mosquitoes mediated by the previously characterized E1-A226V substitution was probably just a beginning of multi-step adaptive process that included the selection of a second (E2-L210Q) and possibly additional, future mutational steps by IOL strains now circulating in urban areas. These mutations, which have no deleterious effect on transmission by A. aegypti, will enable CHIKV to even more efficiently exploit urban transmission in environments populated by A. albopictus, but also to maintain the ability to utilize A. aegypti, which tends to occur in major urban centers [54]. Thus, our findings regarding the continued adaptation of CHIKV to A. albopictus raise serious public health concerns that even more efficient transmission may exacerbate the already devastating CHIK epidemics in India and Southeast Asia. Furthermore, the introduction of the E2-L210Q strain into new areas like Italy and France, where autochthonous cases have already occurred [55]–[57], could spread epidemics into temperate climates where A. albopictus thrives. Considering the broad global distribution of A. albopictus, including nearly throughout the Americas, the E2-L210Q substitution may significantly increase the risk of CHIKV becoming endemic in additional locations.
Interestingly, Niyas et al. (2010) demonstrated that CHIKV strains with the E2-L210Q substitution can be isolated from adult A. albopictus mosquitoes that were reared from wild-caught larvae collected in Kerala State, suggesting that transovarial transmission (TOT) may also play a role in CHIKV maintenance, especially during dry seasons [27], [58]. Also, evidence suggests that TOT occurred in a small percentage of wild mosquitoes during recent CHIK outbreaks on Reunion Island, Madagascar, and in Thailand [59]–[61]. Although we did not attempt to study the effect of the E2-L210Q substitution on TOT, and at least one laboratory study failed to demonstrate TOT in A. albopictus of CHIKV strains with the E1-A226V substitution [58], so the possibility that CHIKV mutations could influence rates of TOT warrants a thorough investigation.
The molecular mechanism explaining the effects of the E2-L210Q substitution on CHIKV infectivity for A. albopictus midgut cells remains unknown. Earlier, we hypothesized that the E2 region around position 211 could be directly involved in interactions with a specific cell surface receptor [46]. We showed that the E2-211T residue mediates a significant increase in infectivity for A. albopictus in concert with the E1-A226V substitution, and that residue E2-211I, which is common among CHIKV strains, blocks this effect. Moreover, using virus overlay protein binding assays (VOPBA) to study CHIKV binding to the proteins associated with the brush border membrane fraction of A. albopictus midguts, we demonstrated that the E2-T211I substitution dramatically alters CHIKV interactions with as yet unidentified proteins (KT unpublished). The recently determined crystal structure of the CHIKV E2 glycoprotein [39] provides additional insights into the possible involvement of residues E2-211 and E2-210 in interactions with a putative mosquito receptor (Figure 8). Both positions are located at the C'B sheet of the E2 protein, which is exposed on the virion surface on the lateral side of domain B, suggesting that these positions could be involved in interactions with cellular proteins. Substitutions of the aliphatic moieties with polar residues in this region may therefore be directly responsible for changing CHIKV affinities to as yet unidentified receptor(s). Interestingly, positions E2-207, E2-213 and E2-218, which have been shown to be involved in VEEV adaptation to equine and mosquito hosts [5], [7], [42], are also located in the same lateral surface of domain B, further supporting the hypothesis that E2-L210Q enables CHIKV to interact with a particular protein expressed on the surface of midgut cells. The studies to identify these protein(s) are underway.
In the study by Niyas et al (2010) that discovered the E2-L210Q substitution in CHIKV strains from Kerala, only limited portions of CHIKV genomes including the nsP2, E2 and E1 genes were sequenced [27]. Since we did not have an access to these isolates or to complete sequence of these strains, we cannot rule out the possibility that other genome regions could be influencing CHIKV evolution in Kerala State. Epistatic mutations in different genome positions can dramatically affect CHIKV infection of A. albopictus [9]. For example, the recently determined, lineage-specific epistatic interactions between positions E1-226 and E1-98 probably limited for at least 60 years the emergence and establishment of new CHIKV strains in Asian regions inhabited by A. albopictus [9]. This suggests that Kerala strains of CHIKV might have acquired adaptive substitutions in addition to E2-L210Q that promote efficient transmission in the human-A. albopictus cycle, and indicates the need for a more detailed, continuous molecular characterization of CHIKV strains from throughout its distribution.
We also investigated if residue E1-226 has an epistatic effect on amino acid E2-210. The E2-L210Q substitution was detected only in CHIKV strains that already acquired the E1-A226V substitution. We observed that the E2-L210Q substitution mediates a 4–5-fold increase in A. albopictus midgut infectivity when expressed in the background of E1-226V, whereas the same substitution caused only a 2.3–2.4-fold increase when expressed in the background of E1-226A (Figure 6 and 7). These results indicate that selection of this mutation would have been even less efficient if it had occurred in a CHIKV strain that did not yet acquire E1-A226V change. Interestingly, our data show that, with regard to CHIKV infectivity of A. albopictus midgut cells, E2-L210Q has an approximately 17-fold (E1-226A background) or 10-fold (E1-226V background) weaker effect compared with E1-A226V (Figure 6 and 7). This could explain why E1-A226V was selected convergently by unrelated CHIKV strains on at least 4 well documented occasions, while selection of E2-L210Q has thus far been observed only once in Kerala State (Figure S1). The stronger fitness effect of E1-A226V is consistent with its historically faster selection, which resulted in a selective sweep in parts of the Indian Ocean, India and Southeast Asia, compared with E2-L210Q, which has predominance in only one location. After CHIKV introduction into a region with large A. albopictus populations, the E1-A226V substitution has consistently taken about 0.5-1 year to appear [24], [28], whereas the E2-L210Q change was observed after at least 3 years of circulation in Kerala State [27] (Figure S1). More studies are needed to determine the precise dynamics of the selective sweeps associated with both mutations.
Another interesting observation is that both A. albopictus-adaptive substitutions exert their effect on CHIKV fitness primarily at the level of midgut infectivity (Figure 4 and 6). The overall increase in the number of midgut cells infected with CHIKV VLPs expressing E2-210Q correlates with the increase in dissemination efficiency observed for infectious viruses. Also, the relative increase in the amount of E2-210Q RNA in midguts infected with VLPs is almost indistinguishable from the relative increase in amount of E2-210Q RNA in midguts exposed to infectious viruses (Figure 4 and S4). Although we did not examine replication in a comprehensive set of mosquito tissues, these results suggest that, after establishing an initial infection from the midgut lumen, the subsequent spread of viruses among neighboring cells is not influenced by position E2-210. Moreover, no differences were observed in CHIKV replication in A. albopictus bodies after intrathoracic infection, indicating that replication of CHIKV in secondary mosquito organs also is unaffected by residue E2-210 (Figure 4). Similar observations were provided earlier for position E1-226 [62]. Experimental studies of epizootic versus enzootic VEEV VLP interactions with the epidemic vector, A. taeniorhynchus, also indicated that midgut epithelia is the target organ for VEEV adaptation to this vector [63]. These findings suggest that adaptation of alphaviruses to a mosquito vector primarily occurs at the level of midgut infection.
In summary, we demonstrated that adaptation of CHIKV to a new mosquito vector can be a multistep process that, since 2005, has involved at least 2 amino acid substitutions in the envelope glycoproteins. The substitution that provides the strongest selective advantage, E1-A226V, was followed by second adaptive mutation (E2-L210Q) that has resulted in a strain circulating in India with the fittest phenotype detected yet for transmission by A. albopictus. We hypothesize that this sequential adaptation will facilitate even more efficient circulation and persistence of the A. albopictus-adapted strains in endemic areas and will further increase the risk of expanded and more severe CHIK epidemics in new geographic ranges. This underscores the need for continued surveillance and studies of ongoing CHIKV evolution, as well as the molecular mechanisms that govern CHIKV adaptation to new environments.
The SL07 (SL-CK1) strain of CHIKV was isolated in 2007 from a human in Sri-Lanka (GenBank Acc. No. HM045801.1). This strain belongs to Indian subgroup of the IOL [9] and was obtained from the World Reference Center for Emerging Viruses and Arboviruses (WRCEVA) at the University of Texas Medical Branch, Galveston, TX after its generous submission by Aravinda de Silva of the University of North Carolina. Since its isolation the strain was passed twice on Vero cells before being used for i.c. construction. Viral RNA was extracted from lyophilized virus stock using TRIzol reagent (Invitrogen, Carlsbad, CA), reverse-transcribed using Superscript III (Invitrogen, Carlsbad, CA) and cDNA was amplified using Pfu DNA polymerase (Stratagene, La Jolla, CA) and PCR. To assemble the i.c., overlapping RT-PCR amplicons were cloned into modified pSinRep5 vector (Invitrogen) under the control of an SP6 promoter using a strategy described previously for strain LR2006 OPY1 [64]. Point mutations 10670C→T (E1-A226V), 9170T→A (E2-L210Q) and 6454A→C (synonymous, ApaI marker) were introduced in various combinations into the i.c. of SL07 using conventional PCR-based cloning methods [65], and the PCR-generated regions were completely sequenced. Plasmids encoding sub-genomic replicons of strain SL07 were generated from the i.c. of the BHK-21 cell-adapted version of this strain [SL07-BHK that contains 1296T→C and 5087A→G (nsP1-L407P and nsP3-T348A) substitutions] which was reported previously [9]. These mutations were identified by electroporation of the SL07 i.c. into BHK-21 cells, followed by sequencing of the recovered, plaque purified viruses. Replicons were generated by replacing the structural gene region of SL07-BHK with the sequence of eGFP or CFP genes utilizing standard techniques [64], [66]. In addition, a point mutation 6454A→C (synonymous, ApaI marker) was introduced into the pRep-GFP construct that allows comparison of the relative RNA quantities in an experimental, mixed infection sample. The helper plasmids were generated by deleting the 373–7270 nt. cDNA fragment from i.c. of SL07 that has mutations of interest at E1-226 and E2-210. Plasmids were propagated using the MC1061 strain of E. coli in 2xYT medium and purified by centrifugation in cesium chloride gradients. Detailed information for all plasmids is available from the authors upon request.
Vero cells (African green monkey kidney) were propagated at 37°C, with 5% CO2, in Minimal Essential Medium (MEM; Invitrogen, Carlsbad, CA) supplemented with 5% fetal bovine serum (FBS). BHK-21(S) [Baby Hamsters Kidney] and 293 (Human Embryonic Kidney) cells were maintained at 37°C with 5% CO2 in MEM-alpha (Invitrogen) supplemented with 10% FBS and 1x MEM vitamin solution (Invitrogen). The Galveston colonies of A. albopictus and A. aegypti mosquitoes were established from the mosquitoes collected in Galveston, TX (USA). Thailand colonies of A. albopictus and A. aegypti mosquitoes were established from mosquito eggs collected in Bangkok, Thailand. All manipulations and handling of mosquitoes were done as described previously [67].
Infectious viruses were generated by electroporation of the in-vitro transcribed RNA into Vero cells. RNA was transcribed from SP6 promoter of the NotI linearized i.c. DNA using the mMESSAGE mMACHINE kit (Ambion, Austin TX). Ten µg of RNA were electroporated into 107 Vero cells using a BTX-Harvard Apparatus ECM 830 Square Wave Electroporator (Harvard Apparatus, Holliston, MS) and 2mm cuvette at the following conditions: 680V, pulse length 99 µs, 5 pulses, with an interval between the pulses of 200ms. Cells were transferred to a 75 cm2 tissue culture flasks with 14 mL of Leibovitz L-15 (L-15) medium supplemented with 10% FBS and 5% tryptose phosphate broth (Sigma-Aldrich, St. Louis, MO). At 3 h post electroporation the cell supernatant was replaced with 14 mL of L-15 medium and maintained at 37 °C without CO2. Cell culture supernatants were collected at 24 and 48 h and stored at −80°C.
To estimate the specific infectivity of electroporated RNAs, an aliquot containing 1x105 electroporated Vero cells was serially ten-fold diluted and cells were allowed to attach to sub-confluent monolayers (1x106 cells/well) of uninfected Vero cells in six-well plates [64]. After 2 h of incubation at 37°C, cells were overlaid with 0.5% agarose in MEM supplemented with 3.3% FBS and incubated for 48 h until plaques developed. The results (specific infectivity values) were expressed as pfu/µg of electroporated RNA (Table S1). Titers of the viruses recovered after electroporation and all experimental samples were determined by titration on Vero cells by plaque assay as previously described [7].
To generate CHIKV VLPs expressing residues of interest in E2 and E1 glycoproteins, BHK-21(S) cells were used, which have superior RNA susceptibility compared to Vero cells. To ensure efficient recovery of CHIKV VLPs from BHK-21 cells, all CHIKV replicons were designed to include BHK-adaptive mutations (nsP1-L407P and nsP3-T348A) identified after rescue of w.t. i.c's in BHK-21(S) cells. Ten micrograms of in-vitro transcribed replicon and helper RNA were mixed and electroporated into 107 BHK-21(S) cells as described above for Vero cells. Cells were maintained in L-15 medium at 37°C, followed by harvesting supernatants at 30 h post-electroporation. The titer of VLPs was determined by titration on Vero cells as described earlier [68]. Briefly, 1×106 Vero cells were seeded in six-well plates and, after a 16 h incubation at 37°C, monolayers were infected with 10-fold dilutions of the samples for 1 h at 37°C, followed by adding 2 mL of MEM. After 24 h of incubation at 37°C the numbers of GFP- or CFP-expressing cells were quantified by fluorescent microscopy and titers were expressed as infectious units (i.u.)/mL.
The role of viral mutations at position E2-210 on CHIKV dissemination in A. albopictus and A. aegypti mosquitoes was analyzed using direct competition experiment as described earlier [6], [9]. A pair of viruses that differed by mutations of interest in the E2 protein was mixed at a 1∶1 ratio, with one of the viruses containing the ApaI marker. Viral mixes were used to prepare infectious blood meals by dilution in an equal volume of the defibrinated sheep blood (Colorado Serum, Denver, CO), then orally presented to 4–5 day old female mosquitoes at 37°C as described previously [6], [67]. Ten days post infection, heads and legs of individual mosquitoes were triturated in 500 µL of MEM media containing 5 µg/mL of Amphotericin B (Fungizone), and 100 µL of clarified supernatant were added to duplicate wells of a 96-well plate containing 5x104 Vero cells/well. At 3 dpi, supernatant from virus-induced CPE (cytopathic effect)-positive wells was used for RNA extraction followed by RT-PCR with 41855ns-F5 (5`-ATATCTAGACATGGTGGAC) and 41855ns-R1 (5`-TATCAAAGGAGGCTATGTC) primers sets using One-Step RT-PCR kit (Qiagen, Valencia, CA). The PCR products were digested with ApaI restrictase (NEB, Ipswich, MA) and separated on 1.5% agarose gels followed by ethidium bromide staining. One PCR band in the digested sample corresponded to disseminated infection for one out of two viruses in the pair; two bands indicated that both viruses disseminated in the same mosquito. Differences in dissemination efficiencies were tested for significance with a one-tailed McNemar test.
Viral competition experiments with serial, alternating CHIKV passaging in A. albopictus and Vero cells were performed as described above with minor modifications. For the first passage, virus SL07-226V-210Q was mixed with 100-fold excess SL07-226V-Apa to generate infectious blood meals containing 5x105 pfu/mL (combined). The blood meal was used for oral infection of A. albopictus (Galveston colony) followed by virus extraction from combined head and leg homogenates derived from 50 individual mosquitoes in 1.5 mL of MEM medium at 10 dpi. Homogenates were filtered and used to infect 75 cm2 flasks of Vero cells. At 2 dpi, cell culture supernatants were diluted 1∶10 in L-15 medium and mixed with equal volumes of defibrinated sheep blood to prepare a blood meal for the second passage. The cycle was repeated a total of 3 times. At 10 dpi of third mosquito passage, heads and legs of individual mosquitoes were processed as described above.
For CHIKV competition experiments in specific body parts of A. albopictus, the mosquitoes were exposed to blood meals containing 1∶1 mixes of [SL07-226V-Apa and SL07-226V-210Q] and [SL07-226V and SL07-226V-210Q-Apa]. Depending on the experiment, at 1, 2, 3 and 7 dpi whole mosquito bodies, mosquito carcasses, or mosquito midguts were collected in pools of ten, and were used for RNA extraction using TRIzol (Invitrogen, Carlsbad, CA). RNA was RT-PCR amplified, followed by ApaI restrictase digestion of amplicons as described above. Gel images were analyzed using TolaLab (version 2.01) and relative fitness for a given virus during competition was determined as the ratio between E2-210L and E2-210Q bands in the sample, divided by the starting ratio of E2-210L and E2-210Q in the blood meal. The results were expressed as an average value of 2 pools of 10 mosquitoes midguts per pool.
For CHIKV competition experiments in intrathoracically infected mosquitoes, 5 pfu of SL07-226V-Apa and 5 pfu of SL07-226V-210Q in 0.5 µL of L-15 media were directly injected into thoraxes of cold-anesthetized A. albopictus (Galveston colony) using capillary needles as described previously [69]. RNA from 2 pools, 5 mosquitoes/pool, was extracted at 1 and 2 dpi and processed as described above.
To investigate the relationship between the blood meal titers and infection rates in A. albopictus, the SL07-226V and SL07-226V-210Q viruses individually were serially 10-fold diluted, mixed with defibrinated sheep blood and presented orally to A. albopictus (Thailand). At 10 dpi individual mosquitoes were triturated in one mL of MEM and used to infect 5x104 Vero cells in duplicate in 96 well plates. CHIKV was detected by observing virus-induced CPE. The difference in the infection rates between SL07-226V and SL07-226V-210Q was tested for significance with a two-tailed Fishes exact test. The oral infectious dose 50% (OID50) values were calculated using the PriProbit program (version 1.63).
For VLP experiments, A. albopictus (Thailand) were infected with 1∶1 mixes (based on i.u. titers) of GFP- or CFP-expressing subgenomic replicons packaged into VLPs using CHIKV helpers that differed by substitutions at positions E1-226 and E2-210 (Figure 6A and 7A). At 1 and 2 dpi, 5–10 mosquito midguts were dissected in PBS, and cut longitudinally to generate monolayers of epithelial cells. These sheets were rinsed in PBS to remove residual blood and gently spread out on a glass slide. A cover slip was applied and the midgut sheets were immediately analyzed by fluorescent microscopy to determine the numbers of cells expressing GFP and CFP in the same field of vision. One or two fields of vision were analyzed for each midgut sheet. In parallel experiment, midguts infected with VLPs packaged using helpers that differ by substitutions at position E2-210 were dissected at 1, 2 and 3 dpi, collected in pools of ten, which were used for RNA extraction using TRIzol (Invitrogen). The RNA was processed as described above.
To demonstrate that replicon and helper RNAs did not recombine to generate infectious virus capable of autonomous replication, 30 mosquitoes were infected with VLPs mixes and at 7 dpi were triturated in 1 mL of MEM, filter sterilized and 300 µL of homogenate was used to infect each of 3 wells of confluent Vero cells in six-well plates. After 1 h of infection at 37°C, 2 mL of MEM was added to each well, followed by incubation at 37°C with 5% CO2. Cells were observed daily for signs of CPE for 5 days.
To investigate the effect of substitutions at E2-210 on CHIKV fitness in Vero and 293 cells, these cells were infected at a multiplicity 0.1 pfu/cell in triplicate with 1∶1 mixtures of [SL07-226V-Apa and SL07-226V-210Q] and [SL07-226V and SL07-226V-210Q-Apa] viruses. Cells were maintained at 37 °C with 5% CO2 in MEM and at 2 dpi, supernatants were collected for RNA extraction and processed as described above.
|
10.1371/journal.pmed.1002572 | Complement-activating donor-specific anti-HLA antibodies and solid organ transplant survival: A systematic review and meta-analysis | Anti-human leukocyte antigen donor-specific antibodies (anti-HLA DSAs) are recognized as a major barrier to patients’ access to organ transplantation and the major cause of graft failure. The capacity of circulating anti-HLA DSAs to activate complement has been suggested as a potential biomarker for optimizing graft allocation and improving the rate of successful transplantations.
To address the clinical relevance of complement-activating anti-HLA DSAs across all solid organ transplant patients, we performed a meta-analysis of their association with transplant outcome through a systematic review, from inception to January 31, 2018. The primary outcome was allograft loss, and the secondary outcome was allograft rejection. A comprehensive search strategy was conducted through several databases (Medline, Embase, Cochrane, and Scopus).
A total of 5,861 eligible citations were identified. A total of 37 studies were included in the meta-analysis. Studies reported on 7,936 patients, including kidney (n = 5,991), liver (n = 1,459), heart (n = 370), and lung recipients (n = 116). Solid organ transplant recipients with circulating complement-activating anti-HLA DSAs experienced an increased risk of allograft loss (pooled HR 3.09; 95% CI 2.55–3.74, P = 0.001; I2 = 29.3%), and allograft rejection (pooled HR 3.75; 95% CI: 2.05–6.87, P = 0.001; I2 = 69.8%) compared to patients without complement-activating anti-HLA DSAs. The association between circulating complement-activating anti-HLA DSAs and allograft failure was consistent across all subgroups and sensitivity analyses. Limitations of the study are the observational and retrospective design of almost all included studies, the higher proportion of kidney recipients compared to other solid organ transplant recipients, and the inclusion of fewer studies investigating allograft rejection.
In this study, we found that circulating complement-activating anti-HLA DSAs had a significant deleterious impact on solid organ transplant survival and risk of rejection. The detection of complement-activating anti-HLA DSAs may add value at an individual patient level for noninvasive biomarker-guided risk stratification.
National Clinical Trial protocol ID: NCT03438058.
| Allograft rejection is a major threat to allografts, with consequences for the patients in terms of mortality and morbidity.
Over the last decade, studies on solid organ transplant patients have reported that complement-activating anti-human leukocyte antigen donor-specific antibodies (anti-HLA DSAs) are highly associated with allograft rejection and failure, with varying magnitudes of effect.
This study was designed to evaluate the clinical relevance of complement-activating anti-HLA DSAs at a population level and across the entire solid organ transplants spectrum (kidney, liver, heart, and lung transplant patients).
The present meta-analysis, including 37 studies and 7,936 patients, provides evidence that circulating complement-activating anti-HLA DSAs are a major determinant of long-term allograft rejection and allograft failure.
These results suggest that circulating complement-activating anti-HLA DSAs are potential noninvasive biomarkers to stratify the risk for allograft failure and rejection.
Further research will be needed to investigate the possibility that the detection of these antibodies might have therapeutic significance and could provide opportunities for a pathogenesis-driven approach to prevention and/or treatment of rejection for solid organ transplant recipients.
| Organ transplantation is the treatment of choice for many patients with end-stage chronic disease, which is an increasing burden on industrialized and newly industrialized countries [1,2]. Despite substantial progress in the development of effective immunosuppressive regimens, thousands of allografts fail every year worldwide due to rejection, with immediate consequences in terms of mortality, morbidity, and billions in extra costs to healthcare systems [3,4]. In the past decade, the role of circulating anti-human leukocyte antigen donor-specific antibodies (anti-HLA DSAs) has been increasingly recognized as a major contributing factor to allograft rejection [5] and long-term allograft failure [6–9] in kidney transplantation [10], with the same important associations more recently appreciated in lung [11], heart [7–12], liver [13], intestinal [14], and pancreas transplants [15].
However, not all antibodies are equal in terms of pathogenicity, and they exert a heterogeneous influence on organ allograft outcomes, ranging from acute forms of rejection leading to immediate allograft dysfunction and early allograft loss to more indolent or subclinical forms leading to progressive allograft deterioration.
The inconsistent effects of anti-HLA antibodies on allograft outcomes, which limit their prognostic value, has recently led to attempts to refine their assessment on the basis of pathogenic characteristics to determine which anti-HLA DSAs carry the highest risk for adverse transplant outcomes. Among the notable characteristics of HLA antibodies, their capacity to activate complement has been suggested as a potential factor directing their pathogenicity in the rejection process [16]. Data support that circulating anti-HLA DSAs have the ability to activate complement by their complement component 1q (C1q), C3d, and C4d complement fraction-binding capacities or by their immunoglobulin G3 (IgG3) subclass component, which are associated with an increased risk of antibody-mediated rejection (ABMR) and allograft loss in solid organ transplant recipients [16–25]. However, prior studies have reported different magnitudes of effect for these antibodies, ranging from strong effects to the absence of associations with allograft outcomes [18,19,26–30], limiting their implementation in clinical practice. Greater precision in predicting allograft outcomes using a mechanistically informed, noninvasive biomarker generalizable to diverse solid organ transplants has been identified as a major goal by professional societies (e.g., the European Society of Organ Transplantation, the American Society for Transplantation, and the American Society of Transplant Surgeons), agencies (e.g., the European Medicine Agency and the Food and Drug Administration) [31], and consortia [32]. These groups have pointed to the need for such biomarkers as vital both to optimizing allocation policy and to better stratifying the risk of long-term allograft failure for individual patients. This meta-analysis aims to evaluate the role of complement-activating anti-HLA DSAs on graft survival and graft rejection across the entire spectrum of solid organ transplants.
This meta-analysis is reported in adherence with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) and the reporting Meta-Analyses of Observational Studies in Epidemiology (MOOSE) [33,34].
A comprehensive search was designed and conducted by an experienced librarian with input from the study investigators. The complete protocol of the research strategy was prespecified and the analysis plan prospectively written (S1 Text). Controlled vocabulary supplemented with keywords was used to search for complement-activating anti-HLA DSAs in human solid organ transplantation in any language. The following databases were included: Ovid MEDLINE In-Process & Other Non-Indexed Citations, Ovid MEDLINE, Ovid EMBASE, Ovid Cochrane Central Register of Controlled Trials, Ovid Cochrane Database of Systematic Reviews, and Scopus. The research was conducted from database inception to January 31, 2018. Complement-activating anti-HLA DSAs were defined by their capacity to activate complement cascade at different levels—C1q [23], C3d [35], C4d [26], or presence of IgG3 subtype [36].
The following keywords were used for the research: “solid organ transplantation,” “kidney transplantation,” “liver transplantation,” “lung transplantation,” “heart transplantation,” “intestines transplantation,” “donor specific anti-HLA antibodies,” “solid-phase assay,” “complement-activating DSA,” “C1q,” “C3d,” “C4d,” “IgG3 subclass,” “outcome,” “graft loss,” “graft survival,” “ABMR,” and “rejection.” For comprehensiveness, we also reviewed all references listed in the full-text publications and reviews on the subject that were not identified by our search criteria. An example of the research strategy in the Ovid database is described in S2 Text.
Studies of any relevant design and in any language on the impact of complement-activating anti-HLA DSAs on long-term graft survival and/or the risk of rejection were initially selected. The eligible studies included all solid organ transplant patients (kidney, liver, lung, heart, and intestinal transplantation), both adult or pediatric patients. Anti-HLA DSAs detected by the Luminex single-antigen bead (SAB) technique were required for the DSA detection technique. Complement-activating anti-HLA DSAs were defined according to their ability to bind C1q, C3d, C4d or their IgG3 subclass. The endpoints of interest for inclusion were either allograft loss for the primary endpoint and/or biopsy-proven rejection as a secondary endpoint. Allograft rejection was labelled either antibody-mediated or mixed-rejection as defined by the Banff international classification for kidney and liver transplants [37,38] or the International Society for Heart and Lung Transplantation (ISHLT) classification for heart and lung transplants [39]. Data on graft loss (hazard ratio [HR]) and/or allograft rejection (HR or odds ratio [OR]) were extracted when available and defined as effect sizes with their 95% confidence intervals (CIs).
The corresponding author of each eligible study was contacted and asked to provide HRs and/or ORs when these were not available in the manuscript. All initial communications with authors were based on a template explaining the study and the data required. Two separate reminders were sent unless we received a definitive response. When no answer was obtained, the study was excluded from the analysis.
We excluded unrelated articles, including those without information on complement-activating anti-HLA DSAs, duplicates, those with nonhuman results or non–solid-organ transplant data, case reports, abstract-only articles, and reviews.
Two reviewers (C Loheac and A Bouquegneau) independently assessed the potential eligibility of each of the titles and abstracts that resulted from the search and then reviewed the full texts of all potentially eligible studies. Chance-adjusted inter-reviewer agreement (kappa statistic) was calculated. All disagreements were resolved by consensus between reviewers and principal investigators (C Lefaucheur and A Loupy).
The collected data included author name, year of publication, study size, mean or median follow-up time, mean age of population, type of complement-activating anti-HLA DSA, comparison used (patients with complement-activating anti-HLA DSAs were either compared to patients without complement-activating anti-HLA DSAs, patients with non-complement activating anti-HLA DSAs detected, or a mixed group of patients without anti-HLA DSAs and with non-complement activating anti-HLA DSAs), effect sizes (HR and/or OR) and their 95% CIs, potential confounding factors, and unadjusted and adjusted estimated risks of graft loss or graft rejection. Adjusted HRs and ORs were used when available; otherwise, univariate effect sizes were used.
We used the Newcastle–Ottawa Scale (NOS) to assess the methodological quality (i.e., risk of bias) of nonrandomized studies [40]. NOS score was calculated on the basis of the following 3 major components: the selection of the study groups and ascertainment of exposure (0 to 4 points), quality of the adjustment for confounding variables (0 to 2 points), and ascertainment of outcomes (0 to 3 points). A high NOS score represents high methodological quality. The only randomized controlled trial was assessed using the Cochrane Risk of Bias tool. Details regarding the NOS scoring system are provided in S3 Text.
Meta-analysis was performed using a random-effects model [41] because of the anticipated heterogeneity across studies. In a random-effects meta-analysis model, the effect sizes from the studies that actually were performed are assumed to represent a random sample from a particular distribution of these effect sizes and take into account both within-study variability (expressed by the CI in each study’s effect sizes) and between-study variability (heterogeneity).
The index group for comparison was patients with complement-activating anti-HLA DSAs, and they were either compared to patients with non–complement-activating anti-HLA DSAs, patients without anti-HLA DSAs detected, or a mixed group of patients without anti-HLA DSAs and with non-complement activating anti-HLA DSAs.
Statistical heterogeneity across the studies was tested with the I2 index [42]. The I2 index describes the percentage of total variation across studies due to heterogeneity rather than chance. A value of 0% indicates no observed heterogeneity; values exceeding 50% may elicit considerable caution and warrant further analysis through subgroup analyses [43]. A low P value of the I2 test (below 0.05) provides evidence of heterogeneity of intervention effects (variation in effect estimates beyond chance). Publication bias was visually assessed using funnel plots and statistically assessed by the Egger’s bias coefficient, which weighted the regression of the intervention effect on its standard error (SE), with weights inversely proportional to the variance of the intervention effect [44]. P < 0.05 (2-sided) was considered statistically significant for the presence of a publication bias.
We investigated the extent to which statistical heterogeneity between results of multiple studies can be related to one or more characteristics of the studies by using metaregression [45]. Metaregression merges meta-analytic techniques with linear regression principles (predicting treatment effects using covariates). Metaregression could also explore possible causes of heterogeneity and ascertain stability of results between subgroup analyses. In the present study, we decided to adjust effect sizes on the following covariates if available: date of publication, mean fluorescence intensity (MFI) for anti-HLA DSAs, number of HLA mismatches, period of inclusion, and mean recipient age. We used the overall model P value to assess whether there is evidence for an association of any of the covariates with the outcome [46].
These analyses were performed to explore potential sources of heterogeneity regarding the primary outcome and to assess the consistency of our results, and the choice of the different subgroup analyses was prespecified prior to any analysis. The following subgroup analyses were considered.
The electronic search identified 5,861 potentially relevant citations. A schematic diagram of the literature search procedure used in the present study is shown in Fig 1. The kappa statistic for study eligibility was 0.9941 between the two reviewers (SE = 0.0949). Finally, 37 studies and 7,936 patients were included in the final meta-analysis, including 24 studies with data on allograft loss, 8 studies with data on rejection, and 5 studies with both primary- and secondary-outcome data. Table 1 summarizes characteristics of the included studies. S1 Table provides a detailed characteristic of included studies.
Overall, 22 (59.5%) studies originated from Europe, 9 (24.3%) originated from North America, 4 (10.8%) originated from the United Kingdom, and 2 (5.4%) originated from Asia. The patients included were kidney recipients (n = 5,991; 75.5%), liver recipients (n = 1,459; 18.4%), heart recipients (n = 370; 4.7%), and lung recipients (n = 116; 1.4%). None of the studies included patients with intestine or pancreas transplantation. Complement-activating anti-HLA DSAs were assessed by their capacity to bind C1q (19 studies), C4d (6 studies), or C3d (4 studies) or by their IgG subclass composition (8 studies). Six studies simultaneously analyzed 2 complement-activating anti-HLA DSA assays [17,20,24,48,49,56]. The mean patient follow-up time post transplantation was 71.2 ± 32.3 months. None of the studies included were sponsored or conducted by diagnostic companies involved in the manufacture or sale of complement-activating antibody assays. Nineteen authors were contacted and asked for supplementary data, and 63% of them provided with the requested information.
The funnel plot presented in Fig 2 demonstrates the absence of a publication bias (Egger’s test P = 0.224). The randomized controlled trial was of moderate quality [70]. The NOS scores for quality assessments of the included studies are presented in S2 Table. The median NOS score was 6 (minimum 3, maximum 9), with 2.8%, 2.8%, 19.4%, 38.9%, 22.2%, 11.1%, and 2.8% of studies having a NOS score of 3, 4, 5, 6, 7, 8, and 9, respectively.
Subgroup and sensitivity analyses were performed on the outcome of graft loss to confirm the consistency of the results and explain some of the heterogeneity found in the overall results. Table 2 summarizes the different effect sizes for the different subgroup analyses.
In the present meta-analysis including 7,936 solid organ transplant patients, we established that complement-activating anti-HLA DSAs represent an important determinant of allograft loss across multiple types of organ transplants without a significant publication bias and with acceptable heterogeneity. Patients with complement-activating anti-HLA DSAs have a 3-fold–increased risk of allograft loss compared with patients without anti-HLA DSAs and/or patients with non–complement-activating anti-HLA DSAs. These associations were consistent regarding long-term allograft loss in high-quality studies, across different solid organ transplant populations (kidney, heart, lung, and liver transplant recipients), across different types of tests used for detecting complement-activating anti-HLA DSAs, and at different times of evaluation for complement-activating anti-HLA DSA status (before and after transplantation). Moreover, beyond the effect on allograft survival, we found that complement-activating anti-HLA DSAs were also strongly associated with an increased risk of allograft rejection. These findings reinforce the robustness of the results and their applicability in different clinical scenarios and transplant programs with different practices and support the possibility of a causal effect between complement-activating antibodies and allograft injury.
One of the major hurdles in the quest to develop personalized medicine in transplantation and improve overall transplant patient outcomes is the lack of valid, mechanistically-informed noninvasive biomarkers for predicting allograft outcomes that can be used for patient risk stratification, clinical trial design, and as surrogate endpoints. The recognition of the dominant role of anti-HLA antibodies in rejection and late failure of kidney [10], heart [12], liver [13], lung [11], or intestinal [14] transplants has been a turning point for transplant medicine in the past decade. However, not all anti-HLA DSAs are equal in terms of pathogenicity and therefore may not be consistently associated with adverse allograft outcomes. Because activation of the complement cascade is an important component of the ABMR process, new approaches have been developed to better characterize anti-HLA DSAs and link their capacity to activate complement to the pathophysiology of transplant rejection. The complement-activating ability of anti-HLA antibodies and/or complement-activating IgG subclasses have been shown to be associated with more severe rejection episodes and diminished long-term graft survival [17,49,50]. However, some groups have reported different results, with varying magnitudes of effects ranging from strong to marginal associations between complement-activating anti-HLA DSAs and allograft loss [19,27].
The results of this meta-analysis were robust across diverse subgroup analyses. First, although kidney transplant patients represented the highest number of patients included in the present meta-analysis, the effects of complement-activating anti-HLA DSAs on allograft loss remained significant in heart, lung, and liver transplant patients. Grouping non-kidney transplant studies together (liver, lung, and heart transplantation) as opposed to kidney transplant studies was based on the larger volume of studies focusing on kidney transplant patients. This mirrors the distribution of solid organ transplants worldwide (84,347 kidney transplantations among the 126,670 total organs transplanted) [72].
Second, the same effect was observed regardless of whether the antibody was preexisting or de novo. Third, we found similar associations regardless of the type of test used for assessing complement-activating anti-HLA DSAs.
In most of the studies included in this meta-analysis, a correlation existed between complement-activating antibody status and anti-HLA DSA level (assessed by MFI). Despite this correlation, 8 studies included in the present meta-analysis with sufficient statistical power to perform multivariable models demonstrated that the association between C1q-, C3d-binding tests or IgG3 test and allograft outcomes was independent of the level of anti-HLA DSA MFI (Fig 5). Moreover, the SAB assays can be falsely low, while the C1q assay is more accurate. Therefore, the SAB assay has limitations that mislead the interpretation in comparing MFI versus C1q, C3d, or C4d assays [73]. In contrast to MFI that was reported in most of the studies in this meta-analysis, anti-HLA DSA level determined by titer of antibody correlated with complement-fixing ability [22,74]. In addition to the requirement of minimum titer of DSAs (>1:16) to be complement fixing, the composition of IgG subtypes may also influence the complement-binding capacity [48,75]. Therefore, C1q, C3d and IgG3 assays provide additional insights beyond the DSA strength/titer. Finally, the cutoffs used for antibody detection and for complement-activating anti-HLA DSAs in the different studies was variable. These different cutoffs and technical issues in anti-HLA DSA detection, such as avoidance of the prozone effect, are beyond the scope of the present study.
The heterogeneity (I2) found in the present study may be explained by (i) different tests and protocols used for screening complement-activating antibodies (C1q, C4d, C3d, and IgG subclass), (ii) different types of transplant cohorts and clinical management, including risk-taking strategies (high versus low immunological risk transplant populations), (iii) the timing of antibody detection before and after transplantation, and (iv) nonoptimal statistical power and statistical methodologies used in some studies. Despite this overall heterogeneity, when subgroup analyses were performed including studies with high methodological quality, the heterogeneity decreased from 29.3% to 3.1%. When patients with kidney transplantation were analyzed, the heterogeneity remained stable. Also, when studies using multivariable models were selected in the main analysis, the heterogeneity dropped to 17.4%. Last, despite the overall heterogeneity, the association between complement-activating antibodies and allograft loss remained highly significant in many different clinical scenarios, transplant populations, and relative to the timing of antibody detection, thereby reinforcing the study conclusions.
The findings of the present study have important clinical implications. The magnitude of the overall association found in the present study further reinforces the possibility of using circulating complement-activating anti-HLA DSAs as a potential prognostic factor for allograft loss in transplant patients. Relative to studies from other medical fields such as oncology or cardiology, well-recognized prognostic biomarkers did not always provide associations as high as the one observed in the present medical scenario [76–79]. Beyond their prognostic ability, the characterization of complement-activating anti-HLA DSA properties may influence the allocation system. The consolidation of the SAB–pan-IgG assay in the detection of preformed anti-HLA antibodies has improved transplantation success. However, its high sensitivity has limited the allograft allocation for sensitized patients. The result from this meta-analysis reveals that not all anti-HLA DSAs detected by SAB–pan-IgG assays are equally pathogenic, supporting that, overall, the neat-serum MFI value alone—which only offers a semiquantitative measurement of antibody level—is not entirely reliable for predicting transplant outcome. While the clinical use of SAB–C1q assay for the identification of unacceptable mismatches would improve wait-listed patient stratification regarding their risk of allograft loss, it might also increase the limited allograft allocation of highly sensitized patients—predefined by the standard SAB–pan-IgG assay but restratified as non–C1q-binding DSAs by the SAB–C1q assay—thereby shortening their waiting time.
Characterization of complement-activating anti-HLA DSAs may also have therapeutic significance, providing opportunities for the prevention and/or treatment of ABMR given the availability of specific drugs targeting complement or inhibiting complement-dependent cytotoxicity [80–82]. The present study provides an important step toward a pathogenesis-based approach for preventing and/or treating ABMR. Compared with the current approach to treatment, which only considers the presence of circulating anti-HLA DSAs, a risk-stratified approach on the basis of the complement-activating capacity of anti-HLA DSAs might significantly improve the response rate to complement-inhibitor drugs. The validity of this approach has recently been suggested in a clinical trial [83] in addition to post hoc analyses of 2 clinical trials (NCT01567085 and NCT01399593) including kidney transplant recipients with preformed anti-HLA DSAs receiving C5 inhibitor (eculizumab) for rejection prophylaxis, showing that the effect of eculizumab on allograft function depends on the complement-activating capacity of anti-HLA DSAs [84]. Further studies are needed for defining whether complement-activating anti-HLA DSAs have the potential to inform therapeutic decision-making for timely intervention and to streamline the use of expensive complement inhibitors in kidney transplantation.
We recognize the following limitations. We first acknowledge the higher proportion of kidney recipients compared to heart, liver, and lung transplant recipients. We also acknowledge that fewer studies regarding allograft rejection are included, which is partly due to the lack of histological phenotyping provided by the allograft biopsy in certain studies. Further studies are required to quantify the magnitude of the effect of complement-activating anti-HLA antibodies on the risk of allograft rejection and the efficacy of ABMR therapies. Third, the timing of anti-HLA detection is also a limitation, and because of the number of studies in the different groups of DSA detection, a comparison between groups was not reliable. Fourth, no data were available from Australian or South American transplant populations or from intestines or pancreas transplantation, limiting the extrapolation of our results to these patient populations. Finally, almost all of the included studies were observational and retrospective. Confounding factors from unknown origin may explain part of the residual heterogeneity observed.
In conclusion, circulating complement-activating anti-HLA DSAs represent a significant determinant of long-term allograft survival and solid organ transplant rejection and may be considered a potential valuable prognostic biomarker for improving the risk stratification for allograft loss.
|
10.1371/journal.pntd.0000085 | Epidemiology and Clinical Features of Patients with Visceral Leishmaniasis Treated by an MSF Clinic in Bakool Region, Somalia, 2004–2006 | There are few reports describing the epidemiology of visceral leishmaniasis (VL) in Somalia. Over the years 2002 to 2005, a yearly average of 140 patients were reported from the Huddur centre in Bakool region, whereas in 2006, this number rose to 1002 patients. Given the limited amount of information on VL and the opportunity to compare features with the studies done in 2000 in this part of Somalia, we describe the epidemiologic and clinical features of patients who presented to the Huddur treatment centre of Bakool region, Somalia, using data routinely collected over a five-year observation period (2002–2006).
Methods used included the analysis of routine data on VL cases treated in the Huddur treatment centre, a retrospective study of records of patients admitted between 2004 and 2006, community leaders interviews, and analysis of blood specimens taken for parasite species identification in Antwerp Institute of Tropical Medicine.
A total of 1671 VL patients were admitted to the Huddur centre from January 2002 until December 2006. Nearly all patients presented spontaneously to the health centre. Since 2002, the average patient load was stable, with an average of 140 admissions per year. By the end of 2005, the number of admissions dramatically increased to reach a 7-fold increase in 2006. The genotype of L. donovani identified in 2006 was similar to the one reported in 2002. 82% of total patients treated for VL originated from two districts of Bakool region, Huddur and Tijelow districts. Clinical recovery rate was 93.2% and case fatality rate 3.9%.
After four years of low but constant VL case findings, a major increase in VL was observed over a 16-month period in the Huddur VL centre. The profile of the patients was pediatric and mortality relatively low. Decentralized treatment centers, targeted active screening, and community sensitization will help decrease morbidity and mortality from VL in this endemic area. The true magnitude of VL in Somalia remains unknown. Further documentation to better understand transmission dynamics and thus define appropriate control measures will depend on the stability of the context and safe access to the Somali population.
| Our paper describes the epidemiological features of visceral leishmaniasis in the Bakool region, South Central Somalia, over the years 2004 to 2006. Since 2000, Médecins Sans Frontières has been providing care for patients suffering from visceral leishmaniasis in Huddur, located in a region endemic for visceral leishmaniasis. By the end of 2005, we witnessed a dramatic increase in the number of patients admitted to the Huddur centre with visceral leishmaniasis. In our paper, we provide a description of the profile of patients admitted, thus giving an insight into the epidemiology of visceral leishmaniasis in a part of the world where relatively little has been documented and where the true magnitude of this neglected disease remains unknown.
| Visceral leishmaniasis (VL) is a vector-borne parasitic disease caused by Leishmania donovani. According to WHO, over the last 15 years, endemic regions have been extending and there has been a sharp increase in the number of recorded cases of the disease. For example, in eastern African countries it has caused epidemic outbreaks like the ones that occurred in Southern Sudan from 1984–1994 [1], in North-eastern Kenya and South-eastern Ethiopia in 2000–1, in eastern Sudan from 1996–97 [2, in Ethiopia and Eritrea in 1997–98 3]. Much of VL is concentrated in East Africa [4] yet little has been reported from the endemic parts of Somalia.
Different profiles of patients with VL and outcomes have been described in Africa. In Ethiopia VL is commonly observed as an opportunistic infection in HIV infected adults with documented mortality rates up to18.5% [5]. In Western Upper Nile, Sudan, the majority of cases reported during a major outbreak from 1984 to 1994 were adults with death rates of 38–57% [1]. In other regions of Sudan and in West Pokot of Uganda it presents mainly as a pediatric problem [6]. In the endemic area of Baringo district in Kenya changing lifestyle has led to a decreasing proportion of new VL cases among men [7].
Areas of Somalia where VL has been reported include the coastal areas in the south of the country [8,9], the area along the Shebelle river in the south of Somalia 10], Lower Juba region (MSF, unpublished report), and Baidoa in Bay region [11]. Information on local vector behaviour and risk factors for infection or disease in Somalia are very limited. In Somalia transmission is thought to be anthroponotic similar to other endemic areas of the region (Uganda, Southern Sudan, Kenya) [12,6]. A study in Kenya revealed that transmission occurs in and around houses [7], but whether this occurs in Somalia is unknown. Termite hills are the favoured breeding and resting sites of P.martini and they are very common in Bakool [13,14].
The turmoil and factional fighting that followed the regime's overthrow in 1991 has left large parts of Somalia without any form of health care. Even in 2006, the majority of health care provided in South Central Somalia is carried out by non-governmental organizations – but with very limited coverage of the Somali population.
Bakol region is located in south-central Somalia, bordering with Hiiraan region to the east, Bay region to the south, Gedo region to the west, and Ethiopia to the north. Médecins Sans Frontières (MSF), a private, non-governmental organization, has been working in Huddur, the capital of Bakool region since 2000, running a primary health care project consisting of outpatient and in-patient departments, a therapeutic feeding centre, a tuberculosis and a VL program. It has been the only treatment centre for VL in Bakool region until 2006. The first report published in the medical literature about VL in the Bakool region in 2000–01 concluded from the pediatric profile of the disease and information obtained from qualitative methods that VL was since long endemic in that region [15]. The infectious agent was confirmed as L.donovani, and entomological studies revealed the presence of potential vectors, Phlebotomus martini and Phlebotomus vansomerenae in Bakool region [16]. In the first year that VL was treated in Huddur - between July 2000 and August 2001- 230 patients with VL were identified and treated [15]. Since 2002 the average caseload was stable at around 140 VL cases per year until September 2005 when an increase in admissions was observed. A total of 1002 patients representing a seven-fold increase compared to previous years average were diagnosed in year 2006.
Given the limited amount of information on VL and the opportunity to compare features with the studies done in 2000 in this part of Somalia, we describe the epidemiologic and clinical features of patients who presented to the Huddur treatment centre of Bakool region, Somalia, using data routinely collected over a five-year observation period (2002–2006).
We describe the profile of VL in the Bakool region using several methods: an analysis of epidemiological and clinical data from VL cases treated in a VL-treatment centre, using a retrospective analysis of case records, and analysis of blood specimens.
The Bakool region consists of 5 districts: Huddur, Tijeglow, Rabdhure, Wajid and El-Berde, with a total estimated population of 245 000. Most of the population of Bakool region have a semi-nomadic lifestyle. The health centre in Huddur run by MSF serves as the primary provider of medical care for the population living in Huddur district. The population described in this report included all patients diagnosed with VL at the Huddur health centre between January 2002 and December 2006. All patients with VL were treated as in-patients in the VL ward of the centre for the entire duration of treatment of one month.
Since the VL treatment centre was established, data from patients was collected by one national clinical officer who has been working continuously on the program. The total number of admissions was obtained from the main registration book of the treatment centre. Demographic and clinical information were collected from individual patient cards. Data from these patient cards was entered routinely in the Access-based Kala Azar Software that was first introduced in the MSF centre the field in year 2004. Clinical records of patients admitted before 2004 were not available for retrospective data entry. A total of 970 patient records were entered into the software from January 2004 until December 2006. This data was then exported to Excel program for data management and analyses. We analyzed the number of cases of VL detected over time, by age, sex, and geographical origin. We compared clinical features between patients admitted during the low case detection period before September 2005 (Jan04–Aug05) and the high case detection period starting in September 2005 when the first increase in admissions was noticed (Sep05–Dec06). Chi-squared test was used to compare proportions. Student t-test was used to compare means.
This outbreak investigation was viewed as a routine operational response. Ethical issues were addressed the following way: we used only routinely collected data in the process of monitoring a treatment program, confidentiality of clinical and laboratory patient information was maintained, patients were explained the reason for taking additional blood samples and were asked for oral consent, and blood samples analyzed in Antwerp Tropical Institute were only used only to detect the parasites and perform species identification. There are no laboratories in the southern Somalia doing parasitological tests. The Ethics Review Board instituted by Médecins Sans Frontières reviewed that relevant ethical issues in this project were well considered.
A total of 1671 VL patients were admitted for treatment in the Huddur Health Centre from Jan 2002 until December 2006. Except for 27 found during outreach activities, all patients presented spontaneously to the health centre. See Table 1 for the number of admissions per year and the number of clinical records available. Clinical records which were entered in the Kala azar Software in 2004–2006 were available for a total of 970 patients. The incomplete data entry in 2005–6 is attributed to the loss of patient cards during the high workload period when record keeping became secondary to patient care.
Figure 1 illustrates the number of VL patients admitted for treatment in Huddur centre across time since January 2002. After a period of relatively low caseload with usually fewer than 20 admissions per month (Jan02 until Oct04), there were 5 months of very low patient admissions due to the absence of the expatriate team (Nov04 until Mar05), during which patients were accepted on an exceptional basis. In September 2005 the monthly case detection started increasing to reach two peaks of over 100 admissions in March–April06 and in September06. Although the number of admissions was dropping after September 2006, the number of admissions in the last quarter of 2006 still remained higher than what was observed in previous years.
Information on place of origin was available for 905/970 patients. Although patients with VL originated from all 5 districts of Bakool region as well as from Baidoa district of Bay region (6%), the majority of patients originated from only 2 districts, Huddur and Tijeglow throughout the years as seen in Figure 2.
Figure 3 shows the age and sex distribution for 969 patients. Median age of patients was 3.8 years (inter-quartile range 2 to 5) and overall boys represented 59.4% of all patients. No adults were diagnosed.
From the 970 clinical series entered in the database, we had laboratory results for 943 patients and out of these 916 (97.1%) were serologically confirmed with either DAT or the Optileish dipstick. Post Kala Azar dermal leishmaniasis was exceptional with only 1 case diagnosed in 2006. The overwhelming majority of patients admitted were new cases with only 6 clinical relapses diagnosed in 2006 and none in 2005. The core clinical features of VL were commonly observed: fever, splenomegaly, weight loss/wasting, and clinical anemia. Cough, epistaxis, and vomiting were frequently reported accompanying symptoms. Diarrhea, jaundice, and lymph node enlargement were infrequent. Duration of illness before first consultation was around 4 months. Variables which showed significant differences in frequencies between patients admitted during the low (Jan04–Aug05) and high case detection period (Sep05–Dec06) included loss of appetite, moderate malnutrition, clinical fever on admission, edema, and average spleen size on admission. Moderate malnutrition, as measured by a Middle Upper Arm Circumference (MUAC) less than 126 mm on admission, was much less prevalent among patients during the low case detection period compared to the high one (42.0% compared to 86.3%; P-value<0.001). The proportion of patients with clinical fever (>37.5°C) measured at admission was higher during the low case detection period than the high case detection period (85.6% compared to 61.2%; P-value<0.001). Average spleen size was higher during low case detection period compared to the high one (8.5 cm compared to 6.4 cm; P-value<0.001). Hepatomegaly was more frequent in low case detection period than the high one (54.8% compared to 22.7%; P-value<0.001) but data was only available for 42 patients during the low case detection period.
Information on treatment outcome was available for 925 of the 970 case series studied (95.4%). A total of 36 deaths were recorded in the case series from January 2004 to December 2006, giving an overall case-fatality rate of 3.9%. Clinical recovery rate was 93.2% (862/925). A total of 27 patients defaulted (2.9%).
DNA of Leishmania spp. was found in 12 of the 17 samples taken from Optileish positive patients under SSG treatment (range of 1 to 9 days of treatment). In 3 of them species identification was reached and confirmed the presence of L.donovani. The 12 samples were negative with the L.infantum-specific PCR. The 9 negative samples with the L.donovani-specific PCR were likely due to a lower sensitivity of the species-specific PCR in comparison with the diagnostic PCR. For 2 of the L.donovani positive samples, cpb PCR-RFLP patterns could be visualised and appeared to be similar to the ones encountered in the L.donovani samples taken in year 2000 [15]. PCR positive samples belonged to patients living in the districts of Huddur, Tijeglow, and Baidoa district of Bay region.
This report describes the pattern of VL admissions over a five-year period and the clinical characteristics and outcomes of 970 patients treated for VL in Huddur centre run by MSF in Somalia over a 3-year period. After four years of low but constant passive VL case finding in the endemic area of Bakool region, a major increase in VL patient admission was observed over a 16-month period in the Huddur centre. Although the reported number of patients treated gives an underestimate of the real prevalence, the trend in case detection clearly shows a sharp increase during the past 16 month period (Sep05–Dec06). The number of patient admissions was not found to be subject to seasonal variation. 82% of total patients treated for VL originated from two districts of Bakool region, Huddur and Tijeglow districts. This could reflect a real clustering of VL as is known to occur within endemic areas. For instance in Baringo district, Kenya, important differences in seroprevalences between villages of the same endemic area were documented [25]. There could be specific environmental and social factors in the group of villages most affected in Tijeglow district that favour transmission, but we have no information on it.
The pediatric profile of patients suggests that adults in the area are immune and that we are dealing with an endemic pattern of VL. The DNA pattern of L.donovani identified in 2006 was similar to the one identified in 2000 [15] suggesting that the same parasite strains remain circulating in the area over the period separating the two sample collections. We cannot exclude the existence of new parasite variants that would have been introduced in the area, but to explore this possibility, further and extensive molecular analysis would be required.
Better awareness amongst the population of the treatment availability, spread by successfully treated patients, may have contributed to the rise of detected cases. But the five-fold increase in the monthly case detection implies that other factors came into play and we cannot exclude a real increase in disease because the prevalence of VL in the area is unknown.
Clinical features observed were typical of VL. The possible increase in the population's knowledge on the disease and on the availability and reputation of treatment in Huddur, may have contributed to the reduced average duration of sickness before presentation: 4 months compared to 19 months in year 2000–1 [15]. This short duration of illness could explain the low case fatality rate (3.9%) as found in Sudan [12]. We cannot explain the observed sex difference among VL patients treated in Huddur centre but suggest the boys had greater exposure to sand fly bites of boys and/or possibly better access to care.
Although, statistically, there was no difference in duration of illness, this may be due to lack of power, since the clinical differences between patients of low and high caseload periods suggest a shorter duration of illness in patients admitted during the high caseload period. However, the difference in hepatomegaly is probably artificial: missing values most likely represent negative clinical findings, which if added to the denominator would reduce the proportion of hepatomegaly during the low caseload period. The fact that malnutrition was more common among patients admitted during the high case detection period compared to the low case detection period would suggest that a worsening in the nutritional status among children in Tijeglow and Huddur districts could have contributed to the increase in cases of of VL. Malnutrition is a well known risk factor in the development of VL disease [26].
Based on our experience certain measures could be implemented for improved care. Lack of transport and the long distances to travel are known to limit the physical access to Huddur centre. Opening other temporary treatment centers especially in the most affected districts could improve access to care for VL patients. Active screening for VL, when doing outreach work such as for vaccination or nutritional screening, would help further increase the number of cases detected and treated. Sadly, the instability in the area due to armed conflict seriously thwarts these efforts. Shorter treatment regimens are needed and would greatly help improve acceptability of treatment and increase the treatment completion rate.
Vector control and other preventive measures have not been implemented but could improve control of VL in the area. For instance the use of treated bed-nets to protect from the P.martini bites, active at night time, which would not only reduce in-household transmission of VL but also of malaria. Targeted information and education of the population to increase awareness could help increase early case detection and limit the use of traditional remedies like abdominal scarification. Insecticidal application to termite mounds could be a measure of targeted control in the most affected villages.
There are a number of significant limitations to our data. One was the lack of full clinical records. Although we cannot exclude bias, there is no reason to believe that the missing patient charts are associated with a particular patient characteristic or outcome as missing data was likely due to reduced documentation of patient charts due to high patient load. The constrained time and access to remote villages did not allow for a prevalence survey in the whole area affected by VL.
As exposure to sandfly varies from area to area, a case-control study to determine the local risk factors of VL would be useful to define targeted control measures. Additional documentation on rain patterns, vector behaviour, and other risk factors for VL like HIV co-infection would be useful in designing adapted interventions to decrease morbidity and mortality.
Our experience suggests that VL is substantially underreported in Bakool region and possibly in neighboring regions of southern Somalia. The true magnitude of the problem of VL in Somalia is likely to remain unknown and documentation and implementation of effective interventions to control VL will be limited as long as there will be no safe access to population and inexistent health care services. |
10.1371/journal.pcbi.1003521 | Exploring the Conformational Transitions of Biomolecular Systems Using a Simple Two-State Anisotropic Network Model | Biomolecular conformational transitions are essential to biological functions. Most experimental methods report on the long-lived functional states of biomolecules, but information about the transition pathways between these stable states is generally scarce. Such transitions involve short-lived conformational states that are difficult to detect experimentally. For this reason, computational methods are needed to produce plausible hypothetical transition pathways that can then be probed experimentally. Here we propose a simple and computationally efficient method, called ANMPathway, for constructing a physically reasonable pathway between two endpoints of a conformational transition. We adopt a coarse-grained representation of the protein and construct a two-state potential by combining two elastic network models (ENMs) representative of the experimental structures resolved for the endpoints. The two-state potential has a cusp hypersurface in the configuration space where the energies from both the ENMs are equal. We first search for the minimum energy structure on the cusp hypersurface and then treat it as the transition state. The continuous pathway is subsequently constructed by following the steepest descent energy minimization trajectories starting from the transition state on each side of the cusp hypersurface. Application to several systems of broad biological interest such as adenylate kinase, ATP-driven calcium pump SERCA, leucine transporter and glutamate transporter shows that ANMPathway yields results in good agreement with those from other similar methods and with data obtained from all-atom molecular dynamics simulations, in support of the utility of this simple and efficient approach. Notably the method provides experimentally testable predictions, including the formation of non-native contacts during the transition which we were able to detect in two of the systems we studied. An open-access web server has been created to deliver ANMPathway results.
| Many biomolecules are like tiny molecular machines that need to change their shapes and visit many states to perform their biological functions. For a complete molecular understanding of a biological process, one needs to have information on the relevant stable states of the system in question, as well as the pathways by which the system travels from one state to another. We report here an efficient computational method that uses the knowledge of experimental structures of a pair of stable states in order to construct an energetically favoravle pathway between them. We adopt a simple representation of the molecular system by replacing the atoms with beads connected by springs and constructing an energy function with two minima around the end-states. We searched for the structure with highest energy that the system is most likely to visit during the transition and created two paths starting from this structure and proceeding toward the end-states. The combined result of these two paths is the minimum energy pathway between the two stable states. We apply this method to study important structural changes in one enzyme and three large proteins that transport small molecules and ions across the cell membrane.
| Complex macromolecular systems such as enzymes, channels, transporters and pumps need to change their shapes and visit many conformational states in order to perform their functions. Experimental data from functional, biochemical, spectroscopic and structural techniques often inform us on the long-lived stable functional states of macromolecular systems. Accordingly, the average structures of thousands of important biomolecules have been determined using X-ray crystallography or NMR. For many well-studied systems, hundreds of structures have been resolved in the presence of different ligands, or under different conditions or functional states. In contrast, for most systems, little or no experimental data are often available on the intermediate structures along the conformational transition pathway associated with a function. To understand the molecular mechanism of a specific biological process, one needs to go beyond the static information and determine how macromolecules change their conformations as a function of time. In practice, however, obtaining direct structural data about a transition pathway is exceedingly difficult, because intermediate conformations are transient and usually short-lived compared to the timescale of the whole process.
Computational methods can help generate physically plausible pathways for conformational transitions, which can then serve as “hypotheses” to be tested and refined experimentally [1]–[3]. The relevance of any in silico pathway lies in its ability to predict the occurrence of intermediates, which can sometimes be detected using X-ray crystallography [1] or indirectly inferred by perturbing the system via site-directed mutagenesis [3]. Computation and experimental validation thus offers a powerful combination to study the mechanisms of complex biomolecular events.
All-atom molecular dynamics (MD) simulation, arguably, provides the most realistic representation of biomolecular dynamics [4], [5]. If one could simulate the system of interest for sufficiently long time-scales then the trajectory could provide the information required to understand a conformational transition (albeit based on a virtual model). However, brute-force MD is often impractical since most large-scale conformational changes take place over timescales ranging from milliseconds to seconds, which are far beyond the reach of the most powerful supercomputers. Special-purpose hardware and software like Anton [6] are pushing the limits of current molecular simulations; however, they still fall short of accessing the relevant time-scales for the cooperative structural changes of large biological systems. Statistical mechanical methods have also been developed specifically to simulate rare dynamical events [7]–[10], though their application to study large-scale conformational transitions in biological macromolecules remains challenging.
One alternative strategy has been to formulate the problem of the conformational reaction pathway as a “chain-of-state”, i.e., a sequence of configurations representing the progress of the system between two known end-states in the multi-dimensional conformation space [11]–[15]. For example, the so-called “string method” based on all-atom MD simulations has been employed successfully to study functionally important conformational transitions in a variety of biological systems, including Src kinases [16], insulin receptor kinase [17], [18], adenylate kinase [19], amyloidogenic isomerization of 2-microglobulin [20], cholesterol flip in membranes [21], myosin VI [22], DNA polymerase [23], and voltage-gated K+ channels [3]. Other notable methods that seek to shed some light on the important intermediate structures between experimentally known intermediates include the weighted ensemble method [24]–[27] and dynamic importance sampling [28]–[30]. Several enhanced sampling methods such as conformational flooding [31], metadynamics [32] and accelerated molecular dynamics [33] have also been used for similar purpose even though these methods are not designed for searching for transition pathways between two known endpoints. A very different approach based on shapes of biomolecules rather than detailed energetics of the system has been used to search for transition pathways in the tCONCOORD method of Seeliger et al. [34]. However, despite these promising advances, the investigation of large-scale transitions of multimeric systems at atomic details remains prohibitively expensive. It is, therefore, desirable to dispose of simpler models and computationally efficient methods to generate a putative pathway with qualitatively reasonable features that can be tested by experiments.
A straightforward way to reduce the computational cost is to simplify the atomistic system by constructing a coarse-grained (CG) model and adopting a simple potential function that uses knowledge of resolved structures. We adopt here a broadly used/tested structure-based CG model, the elastic network model (ENM) [35]–[37], a powerful example of which is the anisotropic network model (ANM) [38].
In the ANM, the protein is represented by a set of CG sites (nodes) placed at the positions of atoms of all the residues and the energy function is a pairwise additive harmonic potential where each site interacts with all the sites within a cut-off distance. ENMs are often used in conjunction with normal mode analysis (NMA) where one diagonalizes the Hessian matrix of the potential constructed around an experimental structure and studies the deformation of the system along the low frequency normal modes. The simplified potential function in the ANM presents the advantage of yielding an analytical expression for the Hessian, directly expressed in terms of the known structure coordinates [38], which is readily decomposed to obtain the ANM (normal) modes. It has been found for a wide variety of large biomolecular systems that collective motions relevant to function occur along the low energy normal modes of motions predicted by ENMs [39]–[45], suggesting that native contact topology accounted for by the network model is a major determinant of accessible modes of function.
Even though ENMs coupled with NMA have been successful in providing insights into important conformational transitions, they explore, by definition, the neighborhood of a given energy minimum and as such they are not adequate for constructing a transition pathway between two endpoints (minima on conformational energy landscape). However, ENM-based approaches have been very influential in the development of a series of methods that aim at providing plausible intermediate structures along a transition. One of the early studies along these lines is that of Jernigan, Chirikjian and coworkers [46], [47] who have used an interpolation technique with distance constraints to avoid steric clashes. They also showed that normal mode calculations could be accelerated by dividing the system into rigid clusters connected by elastic springs [48], and employed cluster-NMA for constructing pathways by successively creating new structures from an end-state [49], [50]. Miyashita et al. [51] started from one stable state, performed successive normal mode calculations and for each new set of normal modes used a small subset based on the overlap with the other end structure to create an intermediate structure. In the plastic network model (PNM), Maragakis and Karplus [52] constructed a two-state elastic network potential by mixing two ENMs, one for each end-state, and then the pathway was constructed in two steps: identification of a saddle point and two steepest descent minimizations. Yang et al. [53] used the same two-state potential to start from both end structures and used well-defined criteria for recruiting small subsets of normal modes to create a series of intermediate conformers via an adaptive ANM (aANM) methodology until the two intermediates merged within a predefined root-mean-square-deviation (RMSD). Hummer and co-workers [54]–[56] also constructed a two-state potential by mixing two ENM surfaces using an exponential mixing rule and constructed the pathway on this surface using saddle point search. Yang and Roux [57] have used a two-state G model [58], [59] and extensive CG simulations in conjunction with clustering methods to investigate pathways in conformational transition of Src-kinase. Chu and Voth [60] used a more complicated two-state potential by representing each pairwise interaction as a double well and used a saddle point search algorithm to construct the pathway. Their double-well network model has more frustration than a two-state elastic network model and captures complexity of a transition that are not present in models with smoother potential energy functions. Franklin et al. [61] used two ENM surfaces in an entirely different way to construct a pathway method. In their MinActionPath method, they developed an algorithm based on the minimization of the Onsager-Machlup action to construct the path with minimum resistance between two stable states. Even though the problem of construction of pathway between two stable states described by simple CG models has attracted a lot of attention, there is still enough room and need for development of new methods that can address the scalability problem in particular, and help efficiently calculate pathways for large systems.
We propose a simple and efficient method, called ANMPathway, and apply the method to understand conformational transitions of several important globular and membrane proteins. We adopt a simple ENM representation for each of the end-states, which accounts for the topology of inter-residue contacts in the structure. We construct a very simple two-state potential by mixing these two ENMs. Our potential has a cusp hypersurface where the energies from both the ENMs are same. We search for a minimum energy structure on the cusp hypersurface and treat it as the transition state. We then start from the transition state and perform two separate steepest descent minimizations to connect the end-states. Conformers collected from two steepest descent paths along with the transition state provide a pathway. Even though the existence of a cusp hypersurface in our potential is somewhat unphysical, we demonstrate, by way of applications to several systems (adenylate kinase (AK), ATP driven calcium pump SERCA, leucine transporter (LeuT) and glutamate transporter ()), that ANMPathway gives physically meaningful pathways and helps generate experimentally testable hypotheses.
The goal of the ANMPathway method is to construct a transition pathway between two end-states of a conformational transition. As in the string method, the pathway is represented by a chain of equidistant states (conformers/images) [11]–[14]. The macromolecular structure is described by a CG model where interaction sites are placed at the positions of atoms, which serve as the set of collective variables for the string. Conceptually, the string is the minimum free energy pathway on the potential of mean force (PMF) of the system with respect to those collective variables [14]. Assuming a Euclidian metric in the cartesian space of the atoms, the equidistant conditions implies that neighboring images along the string are separated by a fixed RMSD. However, in practice, the ENM energy function employed in the CG model is a knowledge-based construct that is only applicable near the experimental structure used to construct the model. For describing a conformational transition between two stable states, we need an approximation to the PMF that is applicable for large distortions from the experimental structures.
In the presence of structural data on the end-states, it is reasonable to construct an effective energy function with two minima centered around the endpoints of the transition. One obvious route to such two-state potentials involves creating two separate energy surfaces that are defined around each of the end-states and then combine these surfaces by an empirical rule. We have adopted this strategy and used two ANMs [38] and a very simple mixing rule to construct an energy function with two minima.
For a protein with N residues, the configuration of the system is denoted by a 3N dimensional vector where, is a three-dimensional vector giving the position of the ith site ( atom of the ith residue). ANM is an elastic network model defined around an experimental structure (e.g. crystal or NMR structure) with the following energy function(1)Here is the distance between nodes i and j, k is the uniform force constant, Cij is an element of the contact matrix defined by(2)(3) is the cut-off distance and is the energy of the system at the reference state. The advantage of including the term is that it allows us to create energy difference between the end-states when more than one ENM are included in the model. In order to construct the potential function we first define two ANM energy functions, and , centered around the end structures and and combine them by the following mixing rule,(4)The energy difference between the end-states of this two-state ENM is .
The two-state potential based on Eq. (4) has a cusp hypersurface in the 3N-dimensional configuration space. Even though potential energy functions developed for real systems are differentiable everywhere, we will show that the simple two-state potential is a reasonable first approximation and is capable of capturing important qualitative features of the conformational transition in question. Both the ANM energy functions are 3N-dimensional harmonic surfaces and the hypersurface where they intersect (i.e. where the energies from both ANM surfaces are same) is another harmonic surface of dimension . We define the transition state as the minimum energy structure on the cusp hypersurface. Given a sequence of conformers that linearly interpolates the Cartesian distance between two conformers that reside on the opposite sides of the cusp hypersurface, it is possible to identify a conformer that has equal energies, within a tolerance, from both the ANM surfaces. This conformer, by construction, resides on the cusp hypersurface. This simple observation allows us to devise an algorithm to search for the energy minimum on the cusp hypersurface i.e. the transition state. Once we have identified the transition state, we can start from there and slide down the harmonic surfaces until we reach the endpoints, by performing two separate steepest descent minimizations. In the end, we collect all the conformers in proper order to construct the transition pathway. The pathway obtained by ANMPathway can be regarded as the minimum energy path between the end structures since it is the combination of two steepest descent paths on two surfaces joined at the transition state which is the minimum energy conformer on the cusp hypersurface. A detailed description of the algorithm is given below.
Several parameters listed above need to be specified before performing the calculation. The two-state potential function is characterized by the force constants and cut-off distances of ANMs. The ANM force constant does not affect the qualitative results (or the shape of conformational change driven by the normal modes), but uniformly scales the absolute size of motions. Our choice of force constants is inconsequential, since the absolute size of the motion is adjusted by the step sizes and . The cut-off distance is usually selected in the range Å and the overall qualitative features of the pathways were found to be quite insensitive to the choice of within this range for all the systems we studied. If desired the force constants can be estimated by fitting the crystallographic B-factors although the B-factors themselves may be biased by the crystallization conditions and crystal contacts. The energy offsets can be tuned if there are experimental information on the relative energies of the end-states. The value of was chosen to be in the range between 10−4 and 10−5 which could be achieved by setting M = 100 (step 1 of the algorithm). The most important parameters for an efficient implementation of the algorithm turned out to be the step-sizes involved in the transition state search on the cusp hypersurface (sA and sB in step 3a). If step-sizes are too large then the resultant movement of the transition state structure on the cusp hypersurface is large and the minimization algorithm does not work. On the other hand, if the chosen values are too small then the convergence becomes slow. For optimal values of step-sizes, short trial runs were performed for several choices, starting from large values and systematically decreasing them at each trial run until the energy of the transition state conformer decreased monotonically for the entire duration of the trial run. The starting values of sA and sB were chosen between 0.8 and 0.4 with ANM force constants set at 0.1 kcal/(mol Å2). These values need to be adjusted if the force constants are changed by maintaining the inverse proportionality between the step-size and the force constant. Our experience shows that a few very short trial runs are sufficient for finding the optimal values of sA and sB and the overall procedure is extremely efficient. The convergence criterion was selected between 10−4 and 10−5. The number of iterations needed for convergence ranged from 200 to 1000 for the systems studied in this paper.
The pathway is constituted of equally spaced structures obtained by the above mentioned algorithm between the two end structures. We have calculated several quantities to analyse the pathway and understand the conformational transition in terms of collective coordinates. The change in energy of the system along the pathway illustrates the shape of the harmonic surfaces used to describe the system. For example, if the structural change involves movements along the low frequency modes, the energy changes are smaller for a given deformation, compared to those involved in movements along high frequency (more local) movements. The cusp region along the pathway, which is easily identified as the place where the system hops from one surface to another, does not necessarily fall in the middle of the two end structures if one endpoint is more compact than the other. In order to understand the importance of the normal modes (ANM modes) of one end structure in describing the conformational transition, we calculated the cumulative correlation cosine, defined below, of few selected structures along the pathway(8)where is the displacement of a selected conformer/image from one of the reference end structures, is the qth normal mode of that structure and m is the total number of modes (starting from the lowest frequency modes) used for evaluating the cumulative correlation cosine.
It is difficult to validate the results of our method with direct experimental observations. Many of the intermediate structures in a pathway are short-lived and may not be amenable to experimental detection. However, it is reasonable to expect that some predictions can be indirectly verified. In order to make closer connections to experiments, we have looked at the possible formation of close non-native contacts along the pathway. The hope is that some of these predictions can be tested in cross-linking experiments. We looked for pairs of residues that are far apart (>10 Å) in both the native states but come close (<5–7 Å) somewhere along the pathway. We were able to find such pairs in two of the four systems we studied.
Adenylate kinase (AK) is an enzyme that catalyzes the transfer of a single phosphoryl group from ATP to AMP via the reversible reaction . The structure of AK consists of three domains: the AMP-binding domain (NMP), the ATP-binding domain (LID) and the CORE domain (Fig. (1A)). The phosphoryl transfer reaction involves a large-scale conformational transition in AK. In the open (O) state, the NMP and LID are farther apart; and in the closed state, they are tightly packed (right and left structures in Fig. (1A)). We have applied ANMPathway on the open (O) to closed (C) transition in AK. The end-states were obtained from the crystal structures (PDB IDs: 4AKE [62] and 1AKE [63] for the O and C states, respectively). The pathway has 100 images with an RMSD of 0.1 Å between two consecutive images and the transition state corresponds to image 89 (Fig. (1B)) which is situated almost at the end of the open to close transition. The transition between the functional substates of AK comprises large scale hinge-like motions of NMP and LID with respect to a rigid CORE. At the initial stage only the LID moves like a rigid body and the rest of the protein is almost unchanged. This motion corresponds to the slow rise in energy (Fig. (1B)). Then the NMP starts to move and the energy rises as local structural rearrangements take place. Finally the CORE domain undergoes some changes and the transition is complete (movie S1 in Supplementary Material (SM)). The overall result is a two step transition mechanism: LID closing followed by NMP closing (or in the reverse direction: NMP opening followed by LID opening).
Because of the functional importance of domain opening/closing, it is natural to expect that the transition can be described by a small number of normal modes, as shown in Fig. (1C). For image 40, only two modes are sufficient to represent 90% of the displacement from the starting (O) structure. As we move away from the reference state, more modes are needed but the number of modes increases slowly. For attaining the other end structure (C, black curve) with a correlation cosine of more than 0.90 no more than 10–15 modes are needed, which is only ∼2% of the total number of available modes. Therefore the normal mode picture is extremely useful for studying this transition.
In recent years, several computational studies have examined the allosteric transition in AK. These studies revealed a multiplicity of pathways, as well as their dependence on the initial conformers. Among them, two types of transitions appear to be consistently observed in independent studies: (i) LID closing followed by NMP closing along the transition [52], [64]–[71] and (ii) LID opening followed by NMP opening along the transition [64], [70]–[72]. Fig. (2) illustrates these pathways in the conformational space defined by LID-CORE and NMP-CORE angles (see Fig. (2) caption for the definition of the angles), with the intermediates obtained by a recently introduced hybrid methodology, coMD [64], and by ANMPathway, as labeled. As can be seen, mechanisms (i) and (ii) are predicted by coMD provided that the starting points are the O and C states, respectively. The pathway (iii), on the other hand, is obtained by conducting two parallel runs, starting from both ends, and generating intermediates until the two paths merge. The initial steps conform to paths (i) and (ii) in this case. Similar transition mechanisms were recently reported by Kidera and coworkers [19] in further support to the LID-movement-first behavior in both directions. This behavior is also in agreement with the free energy surface obtained by Woolf and coworkers [72] as a function of LID-CORE and NMP-CORE angles, although other mechanisms, such as NMP-first closure for unligated AK [67] or monotonous LID/NMP closing [72] have also been reported. ANMPathway yielded a transition state very close to the closed state, i.e. by definition, most of the change in structure (89 steps out of 100) proceeded in the energy well near the open state, hence its consistency with modes accessible to end-state O, or pathway (i).
Neurotransmitter sodium symporters (NSS) are integral membrane proteins responsible for secondary transport of glycine, γ-amino butyric acid and biogenic amines across the plasma membrane. LeuT is a bacterial orthologue of eukaryotic NSS. The protein consists of twelve transmembrane (TM) helices (TM1-12), extracellular (EC) (EL2, EL3, EL4a, El4b) and intracellular (IC) loops (IL1, IL5) and two β-sheets. The crystal structures of LeuT in the outward-facing (OF) occluded (PDB ID: 2A65 [73]) and inward-facing (IF) open (PDB ID: 3TT3 [74]) states have been resolved (left and right structures of Fig. (3A), see figure caption for color codes).
We note that several residues were not resolved in these structures. We have built the ANMs based on the residues commonly resolved in the two structures and used the scaffold region (TM3, TM4, TM8 and TM9) [74] for structural alignment of all conformers along the pathway.
The results are presented in Fig. (3B–C) for the transition from the OF occluded to IF open state. The pathway is composed of 95 images with an RMSD of 0.05 Å between two consecutive images and the transition state is located at image 40 (closer to the OF occluded state). In order to analyze the transition we have looked at the rearrangements of EL4a, EL4b, TM1a, TM1b, TM2, TM5, TM6a, TM6b and TM7 which play important roles in substrate-binding and gating at the EC and IC regions [74]. The motions are much more subtle compared to AK. At the initial stage there seems to be an almost rigid-body rotation in the aforementioned domains (movies S2 and S3 in SM). Then a downward movement of EL4a and EL4b closes the opening at the EC side. Subsequently, concerted motions of TM5 and TM1a occur. At the later phase of the transition, the main event is the movement of TM1a, which, along with other domains, creates an opening into the IC region. Barring the initial stage, various motions involve intra-domain movements and do not follow a strict rigid-body character. TM2 and TM7 move together for the entire duration of the conformational transition.
The complexity of the motions and lack of rigid-body character are reflected on the normal mode projections of various conformers along the pathway (Fig. (3C)). At the initial stage when the motions are rigid-body like, a few normal modes are sufficient to describe the structural change. But as the transition progresses many more normal modes are needed to represent the displacement from the reference (OF occluded) state. For the later stage of the transition and for the end-state almost 500 modes out of possible ∼1550 modes are required to attain a cumulative correlation of 0.8. This is in sharp contrast to the transition in AK where far fewer percentages of modes were sufficient in describing the structural changes.
To gain more insights and validate the analysis, we compared the pathway from ANMPathway to a 235 ns long all-atom MD trajectory using as initial structure an intermediate close to the OF occluded state [75]. The details of the simulation protocols are described in the text S1 of the Supplemental Material (SM). The time evolution of the structure during this transition was probed by monitoring a relevant order parameter, namely the distance between the binding-site residues N21 and S256 shown in Fig. (4A). Notably, a spontaneous transition to IF open state was observed in this conventional (unbiased) full atomistic MD simulation. The MD trajectory thus provides an important data-set for benchmarking the ANMPathway method. Fig. (4B) compares the projection from the ANMPathway (white curve) and the MD trajectory on the space of two order parameters, the N21-S256 distance and the RMSD from the OF occluded state. The MD data are represented as a crude free energy calculated by taking the logarithm of the two-dimensional histogram of the above mentioned two order parameters shown in Fig. (4B). The pathway predicted by ANMPathway is constructed on a smooth potential energy function and has no thermal fluctuations. As such, it is representative of an average pathway of the real system and it should go through the low energy regions of the free energy landscape obtained from the MD simulation of LeuT embedded in fully solvated membrane lipids at finite temperature. This is exactly the behavior we observe in Fig. (4B) for the most parts where all-atom MD data are available. This agreement is satisfactory, given the minimal computational cost required by ANMPathway compared to that (several orders of magnitude larger) required for the full scale all-atom MD simulation.
There is another crystal structure of LeuT which models the outward-facing open state (PDB ID: 3TT1 [74]). The sequence of functional states in the reaction cycle is: OF open (PDB ID: 3TT1)OF occluded (PDB ID: 2A65)IF open (PDB ID: 3TT3). There are important differences in the helical orientations of several TM helices between the OF open and OF occluded structures (Fig. (5A)) even though their overall architectures are quite similar. It is natural to ask whether the ANMPathway could predict the existence of OF closed state along a transition pathway calculated between OF open and IF open states. We indeed found a conformer which is very close to the OF closed state (RMSD from 2A65: ∼1.0 Å) along the pathway between the crystal structures of OF open and IF open states. The detection of the occluded intermediate is studied by monitoring order parameters that describe the instantaneous conformations of the TM helices responsible for gating and binding of ions. The order parameters are the center of mass (COM) distances between the pairs of helices TM1a-TM10, TM6b-TM10 and TM1a-TM6b. The results are shown in Fig. (5B). In all three distance profiles the intermediate is detected as indicated by the yellow arrows. Given the simplicity of the potential energy function, it is quite remarkable that the method is capable of detecting a functionally relevant and experimentally observed intermediate state. This fact highlights the usefulness of the method as well as significance of global modes of motion in facilitating the conformational transition of transporters.
Excitatory amino acid transporters (EAATS) constitute a class of integral membrane proteins that are responsible for secondary active transport of amino acids like glutamate and aspartate across the plasma membrane. The aspartate transporter , an archaeal orthologue of eukaryotic EAAT, is broadly used as a structural prototype, being functionally resolved in multiple states. The protein is a homotrimer. Each protomer consists of eight TM helices (TM1-8) and two helix-turn-helix motifs (HP1 and HP2) at the substrate-binding core [76]. According to the alternating access mechanism that enables the transport of substrate, the trimer alternates between OF and IF states and vice versa, via structural changes in all three protomers, to expose the substrate-binding site to the EC and IC regions, respectively. Crystal structures of in OF (PDB ID: 1XFH [76]), IF (PDB ID: 3KBC [77]) states, as well a mixed intermediate state (iOF) with two protomers in IF conformation and one in an intermediate between OF and IF conformations (PDB ID: 3V8G [78]) have been determined (Fig. (6A)). The iOF state has been suggested to be relevant to uncoupled anion permeation during the transport process [78]. This asymmetric structure closely approximates the intermediate predicted earlier by a combined experimental and computational study [79].
We examined the transition between the iOF and IF states, using ANMPathway. The crystal structures reveal that protomers can be divided into two domains: a trimerization domain (TM1, TM2, TM4 and TM5) that closely maintains its internal conformation during the transition, and a transport domain (TM3, TM6, HP1, TM7, HP2 and TM8) that practically undergoes a downward rigid-body movement (perpendicular to the membrane) relative to the trimerization domain. Our pathway is made of 79 images with an RMSD of 0.1 Å between two consecutive images and the transition state is located at image 47 (Fig. (6B)). In accord with experimental data, the trimerization domain does not exhibit significant change in its internal structure during the entire transition while the transport domain of the reconfiguring protomer (colored green in Fig. (6A)) first moves like a rigid-body and then undergoes intra-domain rearrangements. The short segment of (the broken helix) TM8 moves first, then the lower longer segment and at the end the entire helix moves. Similarly, the upward part of TM7 moves initially followed by the rest of the helix. The movements of HP1 and HP2 are particularly important since they are involved in substrate gating [77]. Along the inward transition of the protomer, HP2 moves first followed by HP1 toward the end, consistent with the higher (initial) mobility of HP2 observed in all-atom MD simulations [80], [81]. See the movies S4 and S5 in SM. The long flexible loop in the extracellular part also undergoes significant movements during the transition. A simpler view of these local rearrangements can be obtained if one constructs two blocks as suggested by Reyes et al. [77] and predicted by ANM analysis to constitute two distinctive substructures subject to anticorrelated motions [82]. Block 1 is composed of HP1 and the lower part of TM7; and block 2 consists of HP2 and the upper part of TM8 up to the point where the helix is broken. Block 2 moves first followed by block 1 (movie S6 in SM).
Fig. (6C) displays the number of normal modes needed for describing the transition to several images/conformers starting from the iOF state. At the initial stage when motions are more rigid-body like, very few modes are sufficient to describe the structural changes. However, similar to LeuT, this number increases as more localized events that do not conform to en bloc movements of low frequency modes become important. These involve flexible regions within the transport domain. However, the total number of modes for a reasonable description of the overall structural change remains significantly smaller than those available (e.g. <100 modes accomplish a cumulative squared cosine of 0.8 with the reference state), supporting the utility of low frequency modes for efficiently mapping the transition pathway.
The complex structural changes result in the formation of eight non-native contacts along the transition. The formation of non-native contact is defined as two residues that are more than 10 Å apart in the end-states but come closer by less than 7 Å during the transition. The non-native contact forming pairs are Val58-Ala358, Val62-Ala353, Leu152-Leu347, Gln220-Met385, Val274-Ala391, Thr275-Gly357, Gly280-Pro356 and Val355-Ile361. All the pairs belong to the protomer that is undergoing the transition to the IF conformation (chain C, colored green in Fig. (6A)). It is worth noting that none of these pairs corresponds to the cysteine cross-link present in the crystal structure of the IF state [77]. Fig. (7) shows the distance profiles for three of these pairs along the transition. Most of these contacts involve pairs where one residue is in the trimerization domain and the other in the transport domain (movies S7 to S9 in the SM). These observations provide a route to test the predictions of the ANMPathway method against experiments.
We have compared the ANMPathway results with the data collected from all-atom simulations. It is very challenging to simulate a spontaneous transition by straightforward conventional MD simulations. In order to perform a qualitative comparison, we have adopted the following protocol. First, a targeted MD (TMD) pathway is generated between the end-states (iOF and IF) with targeting forces acting on the backbone atoms only (see SM text S1 for details). Then, we launched a series of conventional MD runs from various intermediates visited during the TMD. These runs might be expected to follow the local free energy gradient and proceed along the result from ANMPathway provided that the latter offers a reasonable approximation to the actual transition pathway. In order to understand the transition in terms of a simple order parameters we have used, as order parameter, the z component of the distance vector between the COM of the atoms of the transport domain of a conformer and that of the crystal structure of the iOF state. Fig. (8) shows the projection of the predicted pathway and snapshots from unbiased MD runs initiated from various intermediate structures (shown in different colors) on the space spanned by the above-mentioned order parameter and by the RMSD from the end structure (iOF state). Except for one of the trajectories (shown in red points, near the starting point), the MD runs yielded snapshots in accord with the transition pathway predicted by ANMPathway. The iOF state of represents an intermediate between the OF and IF states, it is conceivable that the MD runs starting from this intermediate (shown in red) tend to go back to the more stable OF state, instead of drifting toward the IF state. This is primarily due to the proximity of the initial structure to the deep free energy basin of the more stable (OF) end-state. Overall, these data validate the ability of ANMPathway to provide a meaningful description of the structural changes involved in the global transition of protomers as they reconfigure from OF to IF states.
Calcium transporting pump of sarco/endoplasmic reticulum (SERCA) is an integral membrane protein that pumps ions from calcium-poor cytoplasm of the muscle cell to the calcium-rich lumen of the sarcoplasmic reticulum at the expense of ATP hydrolysis. This process gets rid of the excess ions in the cytoplasm caused by their release from the lumen during muscle contraction and reverts the muscle to relaxed state. The protein is composed of a single polypeptide chain of 994 amino acids that form three cytoplasmic domains (nucleotide-binding domain N, phosphorylation domain P, actuator domain A) and ten TM helices (M1–10) (Fig. (9A)). Extensive structural studies have revealed atomic models of various functionally relevant states in the pumping cycle [83]. We have used the ANMPathway method to explore the transition between the (PDB ID: 1SU4 [84]) and E1.ATP (PDB ID: 1T5S [85]) states. In the state the calcium ions can dissociate from the transmembrane binding sites but, in the E1P state (i.e. phosphorylated state) they are occluded and can not go back to the cytoplasmic side. The architecture of the E1.ATP state is almost identical to that of E1P or the transition state analog E1∼P.ADP state ( RMSD<0.5 Å). Therefore, at the level of a CG model, a transition pathway between the and E1.ATP states can provide important insights into the large scale motions responsible for occlusion of ions in the transmembrane binding sites.
The pathway consists of 106 images/conformers with an RMSD of 0.2 Å between two consecutive images and the transition state is located at image 74 which is closer to the E1.ATP state (Fig. (9B)). Careful examination of the pathway obtained by ANMPathway reveals that at the initial stage only the N domain moves while the rest of the protein remains fixed. This is reflected in the initial slow rise in the energy as shown in Fig. (9B). As the N domain comes closer to the P and A domains, first the P domain undergoes intra-domain changes and then the A domain rotates. This rotation causes upward motions of M1 and M2 and in this process M1 helix adopts a bent conformation. The energy rises quickly as the cytoplasmic domains close and local structural rearrangements take place. The bending of the M1 helix is responsible for shielding the TM calcium-binding sites from the cytoplasmic side and helps to form the occluded state [83], [86] (movie S10 in SM). These observations can be regarded as a crude qualitative picture of the conformational couplings among various parts of SERCA that give rise to the occluded state. The initial motion of the N domain which costs little energy suggests that the protein can exist in alternative conformations where the N domain may be much closer to other cytoplasmic domains compared to the crystal structure. This observation is validated by other simulation and experimental studies [87], [88]. Fig. (9C) shows that only a handful of modes are sufficient for describing the transition. This is due to the initial rigid-body motion of the N domain, in good agreement with the slow modes predicted by the ANM based on the structure. The entire pathway can be projected onto ∼50 modes to take into account more than 80% of the displacement. The normal mode picture is therefore very useful for exploring the transition of this system.
The distance profiles of contact-forming residues along the transition are shown in Fig. (10). The formation of non-native contact is defined as two residues that are more than 10 Å apart in the reference states but come close by less 5 Å during the transition. All contact-forming pairs involve one residue in the M1 helix. The other residue is either on the M4 helix or on the cytoplasmic loop region (see movies S11 to S13 in the SM). These contacts form as the result of bending of the M1 helix which is thought to be responsible for the formation of the occluded state. Therefore the non-native contacts can be probed experimentally to validate the predictions of method as well as to establish structural changes that have functional consequence.
We presented a new computational method, ANMPathway, for constructing the most probable (lowest energy) transition pathway between two stable endpoints of a conformational transition. Conceptually, ANMPathway represents a direct application of the string method to a two-state CG system approximated by ANM energy surfaces. For this reason, the method is simple and efficient; it can produce a completely optimized pathway for a 1000 residue protein in about one hour on a single CPU of a standard desktop computer. We have implemented the method as an open access web server with user-friendly features at http://anmpathway.lcrc.anl.gov/anmpathway.cgi. Although there exist other servers for exploring the transition paths between pairs of known endpoints [53], [55], [61], [89]–[91], ANMPathway is perhaps one of the simplest approaches that efficiently provides a unique solution for the most probable pathway, with a minimal number of parameters and no biased simulations using the ANMs for the two endpoints. The resulting pathway can be interpreted as the minimum energy pathway on the two-state elastic network surface with a cusp hypersurface. The presence of cusp hypersurface does not seem to influence the qualitative nature of the pathway as evident from the comparison of our results on adenylate kinase with those from other methods. Franklin et al. [61] in their MinActionPath method effectively used a similar two-state potential with cusp hypersurface and their results for the AK system are in good agreement with ours. We note that the presence of cusp hypersurface will have noticeable effects on the quantitative details of the pathway, especially near the transition state region. The energetic cost of breaking non-bonded contacts increases as one moves away from the native state which is reflected in the rapid change in energy near the transition state. A smooth surface, like the one adopted in other similar studies [52], [55], [89], will have a better representation of the underlying physics compared to our present two-state potential with a cusp hypersurface. This is the sacrifice we have to make in order to exploit the simplicity and algorithmic efficiency imparted by the two-state potential. However, the influence of this drawback on the pathways produced by ANMPathway was verified to be minimal and localized in the transition state region. This is clearly demonstrated by the refinement of the pathways by systematic smoothing of the two-state potential as described below. The method of Zheng et al. [55] had been implemented as a web server called AD-ENM which allowed us to directly compare our results with that of a similar method based on a more realistic energy surface. The pathways produced by the AD-ENM server had different numbers of images than the corresponding pathways generated by ANMPathway. For the purpose of comparison, we projected the pathways using the same sets of order parameters already used to analyze ANMPathway results in the previous section. For the AK case, it seems that AD-ENM produced a qualitatively different pathway than ANMPathway (SM Fig. S1). The AD-ENM pathway is closer to the scenario (iii) in Fig. (2), whereas ANMPathway result resembles scenario (i) in the same figure. Both methods produced qualitatively similar pathways for LeuT and as illustrated by the projections presented in SM Figs. S2 and S3 respectively. The AD-ENM server makes small modifications to the end-states which can explain the discrepancy found near one of the endpoints in SM Fig. S2. The ANMPathway method is also closely related to the PNM method of Maragakis and Karplus [52]; our two-state potential is a limiting form of their smooth two-state ENM, when the mixing parameter goes to zero. We use a novel and efficient algorithm for searching for the minimum energy structure on the cusp hypersurface, which is equivalent to the saddle point search in the PNM method. If a sufficiently small value of mixing parameter is used, the PNM result will be very similar to ours and this is indeed observed for the AK system.
The effect of the cusp hypersurface on the potential energy surface was further examined by refining the pathway on a smoothed surface based on the following potential function [54], [57],(9)where and are ANM potentials defined in Eq. (1). The parameter determines the amount of mixing and the height of the barrier. In the limit , the potential defined above becomes the two-state potential with a cusp hypersurface defined previously in Eq. (4) i. e. . Therefore, if we refine the pathway after smoothing the potential using Eq. (9) with a high value of , the resulting pathway should not be very different. We tested this hypothesis by performing zero temperature string method [92] calculation on the smoothed energy surface starting from the final path produced by the ANMPathway method. The results for AK are shown in Fig. (11). It is noteworthy that, for , the RMSDs between corresponding images along the two pathways are extremely small (compared to the the resolution of the structures) and the position of the transition state is almost unchanged (the projections of the pathways on the space spanned by the two order parameters described in Fig. (2) are shown in the SM Fig. S4). As the value of is decreased the RMSDs increase and the position of the transition state moves slowly to the right. These observations demonstrate that the algorithm used in ANMPathway has the ability to find a proper transition state on the two-state surface with the cusp hypersurface and also illustrate that presence of cusp hypersurface does not have a significant impact on the over all features of the two-state potential.
A desirable feature of any pathway method is the ability to detect functionally relevant intermediate structures. The two-state potential energy function used in the ANMPathway may be too simple for such purpose. If an intermediate has a topology which is very different from both the end-states then it is unlikely to be detected by this method. However, it is reasonable to expect that an intermediate that maintains the overall fold and much of the secondary structure shared by the two endpoints would show up along the computed transition pathway. This was verified in the transition pathway of LeuT from OF open to IF open state, via the intermediate OF occluded state (see Fig. (5)) consistent with experimental data. We note that Adelman et al. [93] also observed the passage over an occluded intermediate in their weighted-ensemble simulations of the transition of Mhp1 (another transporter that shares LeuT fold) between its OF and IF states. In their case, a significantly more detailed two-state G potential (with several parameters) is used for each endpoint, together with a computationally expensive Monte Carlo sampling algorithm (compared to ANMPathway). The present approach offers the multiple advantages of being simpler and more efficient, while detecting at the same level of resolution an experimentally validated intermediate.
The effective two-state CG potential used by ANMPathway is meant to approximate the true PMF of the system with respect to the coordinates chosen as collective variables [14]. The basic assumption of such structure-based CG models is that the information about the native contact topology is sufficient for describing large scale transitions. Yet, the effective CG potential can produce pathways with various degrees of complexity depending on the system. For AK, the main structural difference between the end-states is rigid movements of large domains and the transition also involves largely rigid-body motions of various domains. But for , even though the inspection of end-states reveals a rigid-body movement of the transport domain with respect to the trimerization domain, the structural changes along the pathway are much more complex. This fact highlights that simple ANM-based surfaces are capable of capturing significant complexity of biomolecular conformational change. For systems involving large-scale rigid-body domain motions, a few ANM modes predicted for the initial structure enable displacements far away from the starting state toward the end-state, as observed in transitions in AK and SERCA. However, if the intermediate structures involve more subtle changes, including localized motions, many more modes may be necessary (e.g. the case of LeuT). Notably, ANM-based surfaces are capable of capturing the complexity of biomolecular conformational change by inclusion of a higher number of modes. These observations validate the conventional picture of correlation between initial direction of structural change and displacements along low frequency normal modes [42]–[44] and will have important ramifications for pathway methods based on subsets of normal modes.
From a computational point of view, ANMPathway can be used as a first approximation to simulate the real transition in all-atom models. Our simple CG model does not have any residue-specific information or side chains and for membrane protein systems we introduce a further simplification by excluding the membrane. Therefore, results from ANMPathway should be interpreted accordingly and should only be used for a quick assessment of a likely path, or for studying relatively large-scale movements such as those of entire domains or subunits. The study demonstrates that ANMPathway can yield very good starting point for building all-atom pathways and refining them with more advanced methods.
Experimental validation of any conformational transition pathway produced via computations is obviously difficult. Owing to their stability, the end-states are often amenable to direct observation by scattering or spectroscopic methods. In contrast, the short-lived intermediate structures occurring transiently along the pathway are much more difficult to detect directly. The most one can hope for is that the predicted pathway could be validated indirectly. In this regard, a particularly interesting possibility is to engineer cross-links between pairs of residues that come in close contact somewhere along the pathway but are otherwise far away from one another in the two end-states. In practice, we were able to find such pairs of residues in and in SERCA which are farther than 10 Å in both end-states, but transiently come within 5–6 Å from one another along the transition pathway. This is encouraging and perhaps experimentally probing pairs of residues satisfying this criterion could become a routine endeavor to validate and test computational pathways.
|
10.1371/journal.pgen.1000810 | Co-Orientation of Replication and Transcription Preserves Genome Integrity | In many bacteria, there is a genome-wide bias towards co-orientation of replication and transcription, with essential and/or highly-expressed genes further enriched co-directionally. We previously found that reversing this bias in the bacterium Bacillus subtilis slows replication elongation, and we proposed that this effect contributes to the evolutionary pressure selecting the transcription-replication co-orientation bias. This selection might have been based purely on selection for speedy replication; alternatively, the slowed replication might actually represent an average of individual replication-disruption events, each of which is counter-selected independently because genome integrity is selected. To differentiate these possibilities and define the precise forces driving this aspect of genome organization, we generated new strains with inversions either over ∼1/4 of the chromosome or at ribosomal RNA (rRNA) operons. Applying mathematical analysis to genomic microarray snapshots, we found that replication rates vary dramatically within the inverted genome. Replication is moderately impeded throughout the inverted region, which results in a small but significant competitive disadvantage in minimal medium. Importantly, replication is strongly obstructed at inverted rRNA loci in rich medium. This obstruction results in disruption of DNA replication, activation of DNA damage responses, loss of genome integrity, and cell death. Our results strongly suggest that preservation of genome integrity drives the evolution of co-orientation of replication and transcription, a conserved feature of genome organization.
| An important feature of genome organization is that transcription and replication are selectively co-oriented. This feature helps to avoid conflicts between head-on replication and transcription. The precise consequences of the conflict and how it affects genome organization remain to be understood. We previously found that reversing the transcription bias slows replication in the Bacillus subtilis genome. Here we engineered new inversions to avoid changes in other aspects of genome organization. We found that the reversed transcription bias is sufficient to decrease replication speed, and it results in lowered fitness of the inversion strains and a competitive disadvantage relative to wild-type cells in minimal medium. Further, by analyzing genomic copy-number snapshots to obtain replication speed as a function of genome position, we found that inversion of the strongly-transcribed rRNA genes obstructs replication during growth in rich medium. This confers a strong growth disadvantage to cells in rich medium, turns on DNA damage responses, and leads to cell death in a subpopulation of cells, while the surviving cells are more sensitive to genotoxic agents. Our results strongly support the hypothesis that evolution has favored co-orientation of transcription with replication, mainly to avoid these effects.
| The fundamental processes of replication and transcription take place on the same template efficiently and accurately, requiring them to be coordinated with each other to avoid potential conflicts. In cells growing rapidly, both replication and transcription of ribosomal RNA (rRNA) genes, and many other genes, are initiated more frequently, further elevating this potential conflict [1]–[4]. Due to the asymmetry of the replisome and the transcription complex, the outcome of their encounter should depend strongly on their relative directionality. RNA polymerase (RNAP) is dislodged by replication in either direction [5],[6]. On the other hand, replication is affected mostly by head-on transcription [6]–[12].
Preventing or resolving this conflict not only requires numerous protein factors [13]–[16] but may also underlie several non-random aspects of genome organization [17],[18]. First, the highly-expressed rRNA and tRNA genes are transcribed almost exclusively co-directionally with replication across numerous species [19],[20]. Chromosomes of the bacteria Bacillus subtilis and Escherichia coli are replicated by bi-directional replication forks initiated from a single origin (oriC), and all rRNA operons are oriented away from oriC [21]–[25]. In yeast, replication fork barriers at the end of ribosomal DNA operons prevent replication from entering head-on into these strongly-transcribed regions [26]. Second, other highly-transcribed genes are also significantly enriched in the leading strand of replication in bacteria, ensuring that their transcription is co-oriented with replication [27]. This feature may be conserved in certain regions of the human genome [28]. Third, longer transcription units are enriched in the leading strand [27],[29]. Fourth, essential genes are enriched to a greater extent than non-essential genes in the leading strand [19]. Finally, there is a general bias for co-directionality of replication and transcription. In B. subtilis and E. coli, this bias is 75% and 55% of all genes, respectively [22]–[24].
Despite a general theme of avoiding head-on transcription and replication, the precise evolutionary forces shaping these inter-connected aspects of genome organization are not understood. The effect of head-on replication on transcription is proposed to impact fitness negatively by interrupting the expression of highly-transcribed genes [27], or in the case of essential genes, by leading to the formation of incomplete transcripts, which subsequently results in toxic truncated polypeptides [19]. However, the effects on replication are also deleterious. In E. coli, replication rate is largely unaffected by co-directional transcription, but is significantly slowed when it occurs head-on to a strong transcription unit [5],[30]. In addition, reversing transcription bias over an extended segment of the B. subtilis genome leads to a significant (30%) decrease of replication rate, extending the time required to replicate the chromosome and potentially impeding the cell cycle [31]. Head-on orientation of replication and transcription has been shown to result in genome instability, which can be due to obstructed replication or disrupted transcription [32]–[35]. It is proposed that the transcription of essential genes is preferentially co-oriented to lower their rate of mutagenesis [30]. Finally, apart from effects on replication and transcription, the transcription bias is also proposed to promote chromosome segregation [36],[37]. Is there a single evolutionary advantage associated with the co-directional bias? Alternatively, is the orientation of each gene selected in its own right? One challenge in understanding the evolutionary bases of orientation biases is dissecting how different aspects of genome organization are important in different circumstances and how they impact cellular fitness.
Here we report that the extent of the impact of head-on transcription on replication differs between genes within the same organism B. subtilis. This was dissected by creating new inversions of either an extensive region of the genome, or a localized region containing strongly-transcribed rRNA genes. Using quantitative genomic approaches, we observed differential rates of replication throughout the genome of the inversion strains—normal replication in intact genomic positions, impedance of replication elongation by ∼30% within the head-on region, and strong blockage of replication at inverted rRNA operons. We further characterized the fitness cost and found that inversion of the oriC-proximal half of a replichore results in a small decrease in growth rate in minimal medium, but is sufficient to confer a significant competitive disadvantage. On the other hand, the replication block at rRNA operons leads to major disruption of replication, induction of the DNA damage response and cell death. We also observed that the rate of mutation of the gene rpoB is increased when it is transcribed head-on to replication within an extended chromosomal inversion, specifically in rich medium. Our results strongly suggest that preservation of genome integrity has contributed to evolution of the genome-wide co-directional bias and its further enrichment in highly-expressed and essential genes.
We previously moved the origin of replication (oriC) away from its endogenous position at 0° (Figure 1A) to 257° (not shown) or 94° (Figure 1B) to reverse the genomic transcription bias in an extended region of the chromosome, and observed that replication elongation was slowed moderately between 0° and the ectopic oriC position due to transcription [31]. This raised the intriguing question: what would be the potential impact of reversed transcription bias on cellular fitness and genome integrity? However, this question cannot be answered using these strains, because other aspects of their genome organization were also altered, including location of oriC and symmetry of the replichores (one spanning ¾ of the chromosome, the other ¼ of the chromosome). Such alterations have been shown to strongly impact cellular fitness in both E. coli and B. subtilis [38]–[40].
To examine exclusively the biological impact of reversed transcription bias, we constructed several new strains. We took advantage of the fact that the strain with repositioned oriC (Figure 1B) has oriC-flanking sequences present both at 0° and the ectopic location (Figure 1C). We reasoned that homologous recombination might occur at these repeats, and screened for such progenies (Figure 1D and 1E). Homologous recombination of repeats upstream of oriC (Figure 1C- repeats marked L, ∼400 bp) repositioned oriC to 0°, with concurrent inversion of the 0°–94° portion of the chromosome. The resulting strain had head-on transcription (HT) between 0° and 94° and equal replichore lengths (Figure 1D). Using the same strategy, we also obtained strains in which homologous recombination had taken place between repeats downstream of oriC (Figure 1C- repeats marked R, ∼300 bp). The resulting chromosomes have oriC positioned at 94° and unequal replichores (UR) but without extended regions of head-on transcription (Figure 1E). Finally, to minimize the possibility of reversion of the inversion, we removed a portion of the remnant homology region at 94°.
Using the HT strain, we evaluated the impact of head-on transcription on fitness by first examining its exponential growth (Figure 1F). The HT strain grew slowly in rich medium (LB) with a doubling time of 70 minutes, compared to 20 minutes for the control. In minimal medium however, the doubling time of the HT strain was similar to that of the control (53 and 52 minutes, respectively). In contrast, the UR strain was sicker than the control in both LB and minimal media (doubling times of 52 and 77 minutes, respectively), indicating a general growth defect due to differing replichore lengths and/or ectopic positioning of oriC. Therefore, unlike the general growth defect introduced by uneven replichores, the growth defect caused by inverting transcription bias over ¼ of the chromosome is nutrient-dependent.
We next examined whether head-on transcription has an effect on replication in the HT strain. We monitored synchronized fork progression in this strain using genomic microarrays (Figure 2A), in minimal medium where no significant loss of growth rate was observed (Figure 1F). Cells were synchronized for their replication cycles using a temperature-sensitive allele of the replication protein DnaB [41],[42]. The gene dosage profile obtained 30 minutes after replication initiation indicates that ∼50% of cells initiated replication. The average position of replication forks can be estimated as the midpoint of the transition between replicated and unreplicated genomic positions [43]. This position is ∼0.68 Mbp from oriC on the left replichore and ∼0.56 Mbp on the inverted right replichore (Figure 2A, blue arrows), indicating that replication forks move slower within the inverted region. We inhibited transcription initiation by adding the drug rifampicin 4 minutes after synchronized replication began, and found that in rifampicin-treated cells replication forks progressed further in the inverted region (∼0.72 Mbp) (Figure 2A, red arrows) compared to untreated cells (∼0.56 Mbp), demonstrating that the reduction in replication fork speed is due to transcription. This reduction in fork movement does not lead to proportionally slower growth likely because B. subtilis has flexible cell division cycles and can compensate for slower fork progression via multifork replication [4].
We next examined whether the decreased replication rate in the inverted region varied depending on nutrient status and genomic position. To this end, we obtained the genomic microarray profile of the HT strain during exponential growth (Figure 2B). Cells are not synchronized and hence the positions of the replication forks would vary from cell to cell (Figure 2B inset). Importantly, by assuming that these cells are in a steady state and their genomic profile is time-invariant on a population basis, we can use this profile to calculate replication speed at every position on the chromosome. This speed is inversely correlated with the local slope of log values of the gene dosage with respect to gene positions (see Materials and Methods).
We first observed that in minimal medium, the genomic profile was smooth but asymmetric (Figure 2B, blue). The rates of replication were similar in the unaltered regions of the chromosome, indicated by similar slopes on the left replichore (172° to 360°, 0.408±0.002/Mbp) and the non-inverted region on the right replichore (94° to 153°, 0.462±0.007/Mbp) (Table 1). In contrast, within the inverted segment on the right replichore (0° to 94°), the slope was 0.646±0.004/Mbp, indicating a ∼30% decrease in fork speed within the head-on region, in agreement with our previous results using an ectopic oriC [31].
Interestingly, when cells were grown in a relatively rich medium (minimal medium supplemented with casamino acids, hereafter referred to as CAA), the gene dosage profile changed sharply at specific locations on the chromosome (Figure 2B, orange). These transition points could be differentiated more clearly when we derived the relative gene dosages between the profiles in CAA vs. minimal medium (Figure 2C), and clearly corresponded to the positions of the rRNA operons within the inverted segment (Figure 2B and 2C, green arrows). The steep slopes at these transitions suggest that replication progression is strongly impeded and even stalled at these locations. Other than at rRNA loci, the genomic profile in CAA was very similar to that of minimal medium including at the intact rRNA operon on the left replichore (Figure 2C, orange arrow). Our results indicate that the genome-wide impedance of replication by reversed transcription bias other than at rRNA loci is largely uniform and unaltered by nutrient conditions. Importantly, we identify rRNA loci as positions of strong, nutrient-dependent obstruction of fork movement when they are transcribed head-on.
We next examined whether the forks obstructed by head-on transcription during growth in CAA were also disrupted. The recombination protein RecA localizes to stalled replication forks as foci only when the forks are disrupted [44]. Hence, we examined the sub-cellular localization of RecA using a recA-gfp fusion construct [45]. We visualized microscopically cells carrying this allele grown in CAA, and found that 97% of the cells of the HT strain had RecA foci compared to 27% in the control (Figure 3A and 3B), indicating that replication forks are disrupted by head-on transcription.
We also noticed that the HT strain had abnormal nucleoid morphology, which appeared filamented and even fragmented in some cases (Figure 3A, lower panels). Cell lengths also significantly increased (not shown) and chain lengths doubled (Table 2), which might explain the ∼2-fold decrease in the number of colony-forming units in strain HT compared to its control (Table 3).
In addition to RecA foci formation, a subpopulation of cells exhibits the SOS DNA damage response (Figure 3C). Using a GFP-fusion reporter of tagC, a member of the SOS regulon [46], we found that the SOS response is induced in greater than 6% of single cells of the HT strain in LB medium, but in less than 1% in minimal medium (Figure 3C). The increase in SOS response in rich medium is accompanied by increased cell death. We performed live/dead staining in which the nucleoids of dead cells with permeable membranes stain with propidium iodide (red), while live cells stain with SYTO9 (cyan) (Figure 3D). There was a marked increase in the fraction of dead cells in the HT strain relative to the isogenic control, again specifically in rich medium (12% versus 1% of cells in LB) (Figure 3E). This suggests that failure to repair replication forks disrupted by strong head-on transcription might lead to failure to complete replication and cell death. In agreement with this hypothesis, we found that HT cells show higher sensitivity to the genotoxic agent mitomycin C in rich medium (Figure 3F), suggesting that exogenous DNA damage adds further demand on their already overwhelmed DNA repair capacity, and dramatically elevates cell death in the population.
Having obtained evidence of disruption of replication by head-on transcription, we next examined whether it also has a consequence on genome stability by measuring the rates of mutations conferring resistance to the drug rifampicin (rifR). In B. subtilis, rifR mutations map to rpoB [47],[48], which is transcribed co-directionally to replication in the wild-type strain but head-on in the HT strain (Figure 4A). Since rpoB encodes a sub-unit of RNA polymerase, its mutation might confer a growth advantage in this scenario. Hence we analyzed the results of the fluctuation test with the P0 method, which measures the rate of mutation independently of its effect on growth rate [49]. We observed that the mutation rate increased ∼3-fold in rich medium in the HT strain compared to an isogenic control with no inversion (Figure 4C). This could be due to a global cellular response to disruption of replication caused by the chromosomal inversion. To examine this possibility, we also monitored the rifR mutation rate in a strain with inversion of ½ of the left replichore, leaving the rpoB region unaltered (Figure 4B). This strain has the same extent of reversed transcription bias as the HT strain, exhibits a similar nutrient-dependent growth defect (Table 4, Figure S1A), and has one inverted rRNA operon located near oriC after chromosome inversion, which causes replication blockage in rich medium, as monitored by microarrays (Figure S1B and S1C). However, there was no increase in the rifR mutation rate in this strain (Figure 4C). Therefore the presence of an inversion alone is not sufficient to cause an increase in rifR mutation rate, rather it is specific to the strain in which rpoB is within the inverted region. In minimal medium all strains had similar rifR mutation rates (Figure 4D).
The most dramatic reduction of replication speed due to head-on transcription occurs at the rRNA operons within the inversion (Figure 2C). This suggests that the nutrient-dependent effect on replication fork progression is mostly due to inversion of the strongly-transcribed rRNA operons. We tested this hypothesis by examining the consequences of specifically inverting rRNA operons. We inverted the rrnIHG cluster (∼17 kbp) that contains 3 rRNA operons and 6 tRNA genes, by inserting two overlapping halves of the neomycin resistance gene (neo) flanking the cluster, and selecting for recombination events that created a complete neo gene, similar to [50] (Figure 5A and 5B). The rrn inversion strain was inviable in LB and had a strong growth defect in CAA compared to the pre-inversion control (doubling times of 44 and 28 minutes, respectively), while their doubling times in minimal medium were similar (44 and 42 minutes, respectively) (Figure 5C). These results indicate that the growth defect of the HT strain was mostly due to inverted rRNA genes.
We next examined asynchronous replication in the rrn inversion strain using genomic microarrays, and observed impedance specifically at the inverted loci, where the gene dosage profile showed sharp discontinuity (Figure 5D and 5E). This indicates that a significant number of replication forks were stalled in this short segment. This effect was much stronger in cells replicating asynchronously in CAA than in minimal medium. Further, the majority of cells with the rrn inversion had RecA-GFP foci/filaments when grown in CAA (Figure 5F and 5G), indicating that strong head-on transcription at this cluster is sufficient to disrupt replication forks. Notably, the nucleoid morphology of the rrn inversion strain was largely normal (Figure 5F, lower panels), indicating that the gross nucleoid defects of the HT strain were not due to disruption of replication by rrn inversion.
Finally, the rrn inversion lowered cell viability. The rrn inversion strain had a much higher fraction of dead cells relative to the pre-inversion control in CAA (10% versus 0.7%, respectively) (Figure 5H and 5I). Thus strong head-on transcription of just rRNA and tRNA operons drastically impacts cell viability through disruption of replication.
Genome organization has evolved to enhance fitness, as evidenced by observations that certain genome rearrangements are not tolerated, or cause growth defects [17], [38], [39], [51]–[56], while others do not and might even be prevalent [57]. One feature of genome organization is that it precludes extensive head-on transcription [5],[17],[18],[27],[31]. In this study we engineered B. subtilis strains with the co-directional bias of replication and transcription reversed over either an extended segment of the genome or at a localized rRNA gene cluster. We employed microarray-based copy number profiling that enabled us to visualize directly the replication status of the entire genome, to identify positions at which there are significant perturbations and to quantify the extent of such perturbations. We found that replication is affected by head-on transcription in at least two ways: one is the apparently uniform deceleration throughout the extended region of reversed transcription bias, and the second and stronger is the disruption of replication at highly-transcribed rRNA loci. Disruption of replication at rRNA genes activates DNA repair pathways and results in sensitivity to genotoxic stress and loss of viability. Together these observations support the hypothesis that head-on collisions between transcription and replication result in loss of genome integrity, and that avoidance of this consequence contributes to the evolution of co-orientation bias in genomes.
Previously we had engineered B. subtilis strains in which the replichores were unequally distributed, and a significant portion of the genome was replicated by forks traveling in the opposite direction to transcription. We found that DNA replication elongation was impeded within the region of reversed transcription bias and these strains had strong growth defects [31]. However it was not possible to attribute the growth defects to reversed transcription bias alone since uneven replichores themselves have been shown in E. coli to lead to strong growth defects and dependence on recombination and/or DNA translocation machineries for viability [38],[39]. Therefore, we have constructed several new strains that separate the alteration of replichore symmetry from that of the transcription-replication bias (Figure 1).
We found that the B. subtilis HT strain with head-on transcription and replication exhibits a strong growth defect only in rich medium, and the UR strain with asymmetric replichores exhibits growth defects regardless of growth medium. The effect of both aspects of genome alteration on fitness is approximately multiplicative (Figure 1F and Table S1) [58],[59]. There are small deviations but this is significant only in minimal medium (p = 0.01). Thus we were able to largely separate two aspects of genome organization–the impact of unequal replichore size, and the impact of colliding transcription and replication forks, on genome integrity and cellular fitness.
We found that the HT strain with the extended inversion also exhibits strong disruption of replication and induction of DNA damage response in rich medium. In addition, it has altered nucleoid morphology and a long and twisted cell shape especially in rich medium (Figure 3A and 3D). We further inverted only an rrn cluster, and demonstrated that it is sufficient to cause disruption of replication, but not the nucleoid morphology defect. The change in morphology in the HT strain can be either associated with head-on transcription at locations other than rRNA operons, or with additional effects of inverting ¼ of the chromosome. The latter could include the alteration of gene positions relative to the origin, which affects their dosage [4],[60], ectopic localization of the parS sites which affect chromosome organization [61]–[63], or defects in chromosome segregation which is proposed to be facilitated by transcription [36],[37].
Using synchronized microarrays, we demonstrate that inverting the transcription bias over ¼ of the chromosome decreases fork progression rate in a transcription-dependent manner (Figure 2A), confirming our previous results obtained using an engineered strain with an ectopic oriC [31]. However, transcription throughout the genome is not uniform. To examine whether different inverted transcription units have different impacts on replication, we obtained asynchronous microarray profiles of exponentially growing cells (Figure 2B). Using mathematical analysis to obtain the rate of replication as a function of genomic position, we confirmed a modest genome-wide impedance of replication throughout the inverted region and showed that it is mostly independent of gene position. An important exception is at rRNA and tRNA operons, evidenced by the punctuated pattern of replication stalling at these clusters. Further, this stalling is strongly potentiated by growth in rich medium.
In general, growth in rich medium results in higher initiation frequencies of both replication and rRNA transcription [1]–[4], either of which could elevate the conflict between transcription and replication, thereby accounting for the observed increase in replication stalling. However, we observed that in the right replichore HT inversion strain, replication is not initiated more frequently in rich medium than in minimal medium. In contrast to wild type cells where gene dosage at oriC is much higher in rich medium than in poor medium, indicating higher rate of replication initiation [31], gene dosage at oriC in the HT strain is similar in CAA and minimal media (Figure 2B). The genomic profiles of the HT strain in the two media are almost identical except at the inverted rrn loci (Figure 2C). It is possible that failure of replication elongation prevents subsequent replication initiation; alternatively, replication initiation frequency could be lower because it is coupled to growth [3],[4], which is slower for the HT strain in rich medium (Figure 1F). Regardless of the reason, it is clear that in the inversion strains, the effect on replication observed at rRNA operons in rich medium is not due to increased replication but is exclusively due to stronger rRNA transcription, which is initiated more frequently because of higher iNTP and lower (p)ppGpp levels [64].
Several models exist to explain why replication is stalled by strong head-on transcription of rRNA operons. The replisome might be capable of bypassing a single head-on RNAP, but the presence of multiple RNAPs on the long and highly-transcribed rrn region could make it harder for the replisome to proceed. In addition, since rRNA operons are highly structured regions, their transcription might obstruct replication forks, as proposed for other unusually structured regions [8],[65]. RNAPs might also stall upon head-on replication to form backed-up RNAP arrays. Backed-up RNAP arrays can create a barrier to replication [15]. Finally, head-on transcription might create RNA-DNA duplexes or supercoiling of DNA that poses a barrier to replication [66]–[68], and this barrier might strengthen to the extent of blocking replication when transcription is sufficiently strong.
Inverting the transcriptional bias of 1/4 of the B. subtilis chromosome slows replication rate within this region by ∼30%. The growth rate of the HT inversion strain in minimal medium is not significantly affected (Table 4). However we discovered that the HT strain indeed has a significant growth disadvantage even in minimal medium when competing with wild type cells, with its relative fitness being 0.92 (+/−0.07) (after factoring in the marker effect) (Table S2, Text S1). This selective effect enables the wild type strain to take over after multiple generations and clearly is sufficient to shape genome evolution.
The impact of inversions on replication and cellular fitness is much stronger when cells are grown in rich media. Inverted rRNA genes in both the HT and rrn strains result in replication blocks (Figure 2 and Figure 5) and likely lead to extensively delayed cell-cycle progression, which explains the dramatic increase of doubling time in rich medium (Figure 1F and Figure 5C). Indeed, blockage of replication elongation has been shown to prevent cell proliferation in E. coli, which can only be reversed upon removal of the barrier [69]. More importantly, we obtained strong evidence that the obstruction created by inverted rRNA transcription also leads to disruption of replication. First, RecA forms foci/filaments in the majority of these cells in rich medium, indicating generation of single-stranded DNA or double-stranded ends (DSEs). Second, there is a significant increase in induction of the SOS DNA damage response [70] in the inversion strains in rich medium (Figure 3C). In B. subtilis the SOS response is not robustly turned on by DSEs due to efficient repair by RecN, and our observation of ∼6% of cells of the HT strain showing SOS induction agrees with the reported value [71]. Third, an increased number of cell deaths occur in the inversion strains especially in rich medium, likely due to failure to repair damaged replication forks. Finally, the inversion renders cells more sensitive to the genotoxic agent mitomycin C especially in rich medium, suggesting that the DNA repair capacity in these cells is highly compromised due to overwhelming demand, leading to detrimental consequences upon challenge by external DNA damage.
It remains unclear whether replication fork collapse [72] takes place soon after forks are stalled by head-on transcription, or only when a second round of replication forks collides with a prior round of stalled replication forks, as was demonstrated previously at the replication terminator sequence in E. coli [73]. It is possible that both types of collisions would lead to disruption of replication, with collision of subsequent replication forks being more costly. This might explain why inverting rrnIHG near the origin of replication results in a particularly strong growth defect, as the second fork encounters the stalled first fork soon after it initiates from the origin of replication. Indeed the oriC-proximal region is the chromosomal location of highly-expressed genes, especially those involved in macromolecular synthesis [74], and thus inversions here might be particularly detrimental.
Disruption of replication has been shown to lead to higher levels of genome instability especially in the vicinity of the disruption [73],[75],[76]. We observed an increase in rpoB mutation rate when the genomic region encoding rpoB is inverted (Figure 4). This increase is only observed when cells are grown in rich medium, implicating a dependence on the level of rpoB transcription or disruption of replication. It is unlikely to be solely due to a global cellular response to disruption of replication such as the SOS response, since inversion of a symmetric half of the other replichore without affecting rpoB, did not elevate its mutation rate despite the fact that replication is clearly inhibited by inverted rrnB transcription in this strain (Figure 4 and Figure S1). One possibility is that disruption of replication at rpoB results in recruitment of an error-prone DNA polymerase via RecA. RecA has been shown to activate directly an error-prone DNA polymerase in E. coli [77]. Further work will be required to differentiate this possibility from other remaining possibilities, e.g., that the increased mutation rate is due to altering rpoB location rather than orientation, or due to disruption of replication at the neighboring rRNA loci, or that the rpoB mutagenesis is due to a threshold level of the SOS response that is only met upon inversion of the right, but not the left replichore.
The impairment of replication by transcription has been shown to result in transcription- associated recombination [32] or deletion [34]. Replication orientation significantly influences the spectrum of point mutations in yeast [35]. This suggests that impairment of genome replication might also contribute to transcription-associated mutagenesis [78]–[81]. Another major mechanism which might be at play is the direct activation of the error-prone translesion polymerase via its interactions with transcription factors such as NusA [82].
Our observations offer strong support for the hypothesis that the effect of transcription on replication is an important driving force in the evolution of genome organization [18]. In addition, our work suggests that the precise cost might vary depending on both the gene type and the growth environment. First, inversion of the strand bias of transcription within an extended segment of the chromosome results in a small growth defect in nutrient-poor medium yet is sufficient to confer a strong competitive disadvantage. Second, the inversion of rRNA operons leads to disruption of DNA replication, which is especially costly if cells are grown in rich medium (Figure 2, Figure 3, and Figure 5). This explains why rRNA operons are all oriented co-directionally. Although the most prominent disruptive effects we observed were at inverted rRNA operons, it is possible that these effects can also be extended to other highly-expressed or long and structured genes. Third, we also observed that mutation rate is higher for the rpoB gene transcribed head-on, supporting a model that co-orientation of transcription and replication of essential genes might have evolved to avoid their mutagenesis [30] (Figure 4). In addition, highly-expressed non-essential genes are also intolerant of mutagenesis since it could result in reduction of their expressivity. Thus minimizing mutagenesis may also underlie the orientation bias of highly-expressed genes.
There are considerable differences in transcription orientation biases among organisms. While low G+C Firmicutes (such as B. subtilis) and the Mycoplasmas have strong transcription orientation biases, other bacteria do not [18]. The widely-studied bacterium E. coli has only 55% co-orientation bias. Interestingly, inversion of long segments of the E. coli replichores including rRNA operons results in no growth defect [38]. What causes such a drastic difference in the penalty of head-on transcription? There are at least three possible explanations: differences in the composition of their replication machineries [24], non-replicative helicases [16] or transcription factors (AT and JDW unpublished). These considerable differences might underlie the differential abilities of organisms to cope with conflict between transcription and replication and thus influence the evolution of their genome organization.
Cells were grown in LB or defined minimal medium (50mM MOPS) [83] with 1% glucose, supplemented with 40 µg/ml tryptophan, 40 µg/ml methionine, 40 µg/ml phenylalanine, 100 µg/ml arginine, or casamino acids (CAA) (Difco) (0.5%). Cells containing inversion of rRNA operons were maintained on 5 µg/ml kanamycin (kan). Chloramphenicol (cm), erythromycin (erm), and spectinomycin (spc) were used as described at 5, 0.5 and 40 µg/ml, respectively.
Standard techniques were used for genetic and molecular biological manipulations [84]. Strains used are listed in Table S3. Primer sequences are listed in Table S4. Strains were constructed in the JH642 background [85], and in the YB886 phage-defective background [86] because phage excision and duplication during stressful conditions often create localized gene copy number alterations. YB886 is cured of phage SPβ, defective for phage PBSX induction [86], and also lacks the transposon-like element ICEBs1 [87].
Strain JDW704 was created by introducing an ectopic oriC at aprE (94°) in YB886 and then deleting the endogenous oriC (0°), using genomic DNA from JDW258 [31] and MMB703 [40], respectively. Progenies of JDW704 were screened for recombination events in sequences flanking oriC by PCR using oJW114/oJW135, oJW115/oJW157, oJW112/oJW75 and oJW113/oJW146. The stabilized inversion strain JDW713 was obtained by transforming JDW704 with linearized plasmid pJW247. The isogenic control was generated by transforming YB886 with linearized plasmid pJW207. The left replichore inversion strain JDW605 was created by screening progenies of MMB703 [40] by PCR using primers oJW112/oJW388, and oJW113/oJW146.
Inversion of rRNA operons was performed similarly to the method described in [50] and [88]. Briefly, two halves of the neomycin resistance gene (neo) overlapping by 583 bp, were inserted flanking the rrnIHG region in strain YB886 using plasmids pJW260 and pJW261, respectively. The strain was plated on kanamycin to select for cells in which a complete neo gene was created by recombination between the two halves. The inversion junctions were tested by PCR using oJW450/oJW442, and oJW452/oJW436. Genomic coordinates [22] of the rrn inversion are 159778 to 176408 in JDW860 and 154793 to 176408 in JDW861.
Plasmid pJW247 was constructed by cloning sequences flanking the inversion junction into the vector pUC18, on either side of the cat gene. The sequences were amplified using oJW206/oJW360 and oJW210/oJW211, and cat was amplified from pGEMcat [89] using oJW208/oJW209. Plasmid pJW207 was constructed similarly, except that the PCR product of oJW204/oJW205 was used instead of oJW206/oJW360. Plasmid pJW260 was constructed by cloning sequences flanking the downstream rrn inversion junction into the vector pUC18erm, on either side of the ‘neo fragment. The sequences were amplified using oJW434/oJW435 and oJW438/439, and ‘neo was amplified from pBEST502 [90] using oJW436/oJW437. Plasmid pJW261 was constructed by cloning sequences flanking the upstream rrn inversion junction into the vector pBEST501 [90], on either side of the neo’ fragment. The sequences were amplified using oJW428/oJW429 and oJW432/433, and cat was amplified from pGEMcat using oJW485/oJW486.
Strains with the dnaB134ts allele [41],[42] were grown in minimal medium at 30°C to OD600 = 0.2. Cells were shifted to 45°C for 60 minutes to prevent new initiation and allow ongoing replication to complete. The temperature was rapidly shifted down to 30°C to allow synchronized initiation of replication, and the culture was split into 2 flasks. 4 minutes after the down-shift, rifampicin was added to one flask to 0.25mg/ml. Cells were collected 30 minutes after the down-shift, mixed with an equal volume of 100% ice-cold methanol and processed for microarray analysis as described [44].
Hybridization was performed according to the Agilent Oligo aCGH protocol using custom 44K oligonucleotide Agilent microarrays. Microarrays were scanned using a GenePix 4000B scanner (Axon Instruments). Cy3 and Cy5 levels were quantified using Agilent's Feature Extraction software. Relative DNA content (log2 ratio of Cy3 to Cy5 levels) was plotted against the gene position on the chromosome, with the origin in the center and the terminus at each end of the x-axis. For the inversion strains, the genomic positions are rearranged to reflect genome reorganization. For synchronized microarrays, the rolling average of gene dosage ratios (log2) for every 200 consecutive positions was calculated from the raw data, and plotted against the mid-point of these positions.
Genomic microarray profiles of cells grown in mid-exponential phase were obtained by hybridizing against a synchronized reference, such that the ratios were proportional to the actual gene dosage. The data for the right replichore were analyzed based on the following (for the left replichore, the equations are identical except with negative signs):
1. During exponential growth, the total number of cells increases exponentially with cell mass doubling. Define T as the mass doubling time, t as the time of measurement, and N(t) as the total number of cells at time t, then(E1)
2. Define x as the position of the gene, and the total gene dosage of the population of cells at position x as f(x, t). There are 4.2Mbp of nucleotides and over 108 cells, so we can approximate f as a continuous function with continuous variables x and t, despite that fact that each cell is undergoing discrete events including replication initiation and cell division.
Assuming that cells are in steady state and their genomic profile is time-invariant on a population basis, we have(E2)
Combining E1 and E2, we have(E3)
3. The rate of replication fork progression is a function of genome position x but not of time t during steady state growth. Therefore we can define the replication rate at position x as v(x). By this definition, for small Δx, we have:(E4)Combining E3 and E4, we haveDefine , thenwhich can be rearranged as(E5)which can also be written in differential form as(E6)
Once we obtain g(x, t), equations E5 and E6 give a precise definition of how to obtain v(x). Our microarray data show that, for a fixed t, g(x, t) can be approximated by a piecewise linear function over x. Let a discrete series (x1, x2….xm) be the connecting points of the piecewise linear function. If x falls between xi and xi+1, and , where ai is independent of t, then we can obtain v(x) as:(E7)
Measurement of rifampicin-resistance (rifR) mutation rates was performed using the fluctuation test as described [49]. 50 parallel cultures of 1–2 ml each were set up for each strain, grown at 37°C to OD600∼0.5 and plated on minimal medium containing rifampicin (5 µg/ml). Serial dilutions were also plated on non-selective medium to count the number of colony-forming units. After incubation at 37°C for 36 hours, the number of plates with no rifR colonies was counted. RifR mutation rate was calculated using the P0 method [49]. Results from 2 or more independent experiments were averaged. Error bars represent the range of data for n = 2, and the standard error for n>2.
The recA::(recA-gfp spc) allele [45] was used to replace the endogenous recA gene in the inversion and control strains. The recA-gfp strains were grown at 30°C in CAA and minimal medium. At OD600 = 0.2–0.6, 250 µl aliquots were labeled with the membrane dye FM4-64 (0.05 µg/ml) and/or DAPI (0.1 µg/ml). Cells were spotted onto thin agarose pads (1% agarose in 1× Spizizen's salts) on multi-well slides, covered with a cover slip and imaged in a Zeiss Axiovert microscope using a 100× oil immersion objective. Images were captured using a Hamamatsu Digital CCD camera, and analyzed using the AxioVision software. The number of cells with RecA-GFP foci or filaments relative to the total number of cells in each image was counted.
SOS induction was monitored similarly in cells containing the tagC-gfp reporter grown at 37°C in LB, CAA and minimal medium. Data were analyzed by calculating SOS induction per cell length.
For estimating the number of dead cells microscopically, the Live/Dead BacLight Bacterial Viability Kit (Molecular Probes) was used, in which live cells are labeled with SYTO9 (green fluorescence, colored in cyan in Figure 3D and Figure 5H) and dead cells with propidium iodide (red fluorescence). Data were analyzed by calculating cell death per cell length.
For all microscopy experiments approximately 1000 cells were counted for each strain and each growth condition. Results from 2 or more independent experiments were averaged. Error bars represent the range of data for n = 2, and the standard error for n>2.
Cells were grown at 37°C to OD600∼0.2–0.6 and serial dilutions were plated on minimal medium. Colonies were counted after 36 hours of incubation at 37°C. Data from 3 independent experiments were averaged.
Cells were grown in either LB or minimal medium to OD600 = 0.3. 5 µl of 1∶10 serial dilutions ranging from 10−2 to 10−6 were spotted correspondingly on LB, Min and plates supplemented with mitomycin C to a concentration of 0.0625 µg/ml. Plates were incubated overnight at 37°C and photographed the next day. The experiment was repeated twice and representative images are shown.
|
10.1371/journal.pbio.1001914 | Mitosis Gives a Brief Window of Opportunity for a Change in Gene Transcription | Cell differentiation is remarkably stable but can be reversed by somatic cell nuclear transfer, cell fusion, and iPS. Nuclear transfer to amphibian oocytes provides a special opportunity to test transcriptional reprogramming without cell division. We show here that, after nuclear transfer to amphibian oocytes, mitotic chromatin is reprogrammed up to 100 times faster than interphase nuclei. We find that, as cells traverse mitosis, their genes pass through a temporary phase of unusually high responsiveness to oocyte reprogramming factors (mitotic advantage). Mitotic advantage is not explained by nuclear penetration, DNA modifications, histone acetylation, phosphorylation, methylation, nor by salt soluble chromosomal proteins. Our results suggest that histone H2A deubiquitination may account, at least in part, for the acquisition of mitotic advantage. They support the general principle that a temporary access of cytoplasmic factors to genes during mitosis may facilitate somatic cell nuclear reprogramming and the acquisition of new cell fates in normal development.
| Cells are dividing very actively at a time in development when new gene expression and new cell lineages arise. At mitosis, most transcription factors are temporarily displaced from chromosomes. We show that, after transplantation to oocytes, somatic cell nuclei that have been synchronized in mitosis can be reprogrammed to pluripotency gene expression up to 100 times faster than interphase nuclei. We find that, as cells traverse mitosis, their genes pass through a temporary phase of unusually high responsiveness to oocyte reprogramming factors (mitotic advantage). Many other genes in the genome have also shown a mitotic advantage, which affects the rate rather than the final level of transcriptional enhancement. This is attributable to a chromatin state rather than to more rapid passage of reprogramming factors through the nuclear membrane. Histone H2A deubiquitination at mitosis is required for the acquisition of mitotic advantage. Our results support the general principle that a temporary access of cytoplasmic factors to genes during mitosis facilitates somatic cell nuclear reprogramming and the acquisition of new cell fates in normal development.
| Normal development, as well as nearly all cases of experimentally induced changes in gene transcription, is accompanied by cell division. It is therefore hard to distinguish those molecular events which prepare cells for, or engage them in, mitosis from those that are required specifically for transcriptional reprogramming. The relationship between the cell cycle and cell fate decisions has for a long time attracted interest [1]. Transition through mitosis is a time when many transcription factors are displaced from chromatin, potentially permitting new transcription factors to occupy chromatin sites on mitotic exit and so direct a postmitotic cell fate change [2]–[5]. Mitotic remodelling has been shown to be of great importance for the efficient replication of erythrocyte nuclei by Xenopus egg extracts [6],[7]. For new transcription, cell division seems to be needed in some cases [8],[9] but not in others [10],[11]. Here we have used nuclear transfer to amphibian oocytes to compare directly the ability of mitotic chromatin or interphase nuclei to be reprogrammed in the absence of cell division.
Germinal vesicle (GV) stage oocytes do not replicate or divide. They therefore provide an opportunity to test whether the cell cycle phase of donor nuclei affects the efficiency of nuclear reprogramming as judged by active transcription of previously silenced genes [12]. To our surprise, we found that a mitotic state of donor nuclei dramatically increases the efficiency of activating certain quiescent pluripotency genes in these nuclei. Our results support an idea that a brief period during mitosis facilitates an exchange of gene regulatory factors on chromatin and that this could be an important mechanism to help cells embarking on new cell lineages during normal development.
Permeabilized mouse C2C12 cells, a cultured myoblast cell line which we have used extensively in our oocyte nuclear transfer experiments, were arrested at specific stages of the cell cycle (Figure S1a) and were injected into the GV of oocytes (Figure 1a). The DNA content of these donor cell populations (Figure 1b) confirmed cell cycle arrest in each of the cell cycle stages. The transcriptional reactivation of three silent genes quiescent in C2C12 cells (Nanog, Oct4, and Sox2) was assessed by RT-qPCR 38 h after nuclear transplantation (Figure 1c). Nuclei at a late stage of the cell cycle (M) show greatly enhanced transcription of each of the genes when compared to unsynchronized nuclei (predominantly G1 and S), whereas an already active gene (c-jun) shows little increase in transcript level. Particularly impressive is the 100-fold enhancement in Sox 2 expression from mitotic donor nuclei when compared to interphase donor nuclei (Figure 1c). In over 50 experiments, donor cells arrested in mitosis or in late G2 always generated more Sox2 transcripts from reactivated genes at 25–48 h after injection to oocytes than unsynchronized donor cells. This difference ranged from a few fold to over 100-fold and is much affected by the exact duration of nocodazole treatment. Sox2 is a gene that is more widely expressed than most others, notably in early embryos, in most stem cells, and in the nervous system [13].
To test whether this result is a peculiarity of this donor cell type (C2C12 myoblasts) or is a nonspecific effect of nocodazole, we repeated these experiments with 10T1/2 donor nuclei (Figure S1b) or prepared mitotic C2C12 donor cells without any inhibitors by a shake-off procedure (Figure S1c). In both cases, enhanced transcription from mitotic donors was observed, although the magnitude of mitotic advantage was lower (particularly in the case of the shake-off samples, many cells of which appeared to be apoptotic by visual inspection). Mitotic donor nuclei were also prepared using another cell synchronization agent (Taxol), and the mitotic advantage was again seen (Figure S1d). When G1/G0 cells were exposed to nocodazole for the same period of time as used to prepare mitotic cells, no enhancement of transcription of the genes was observed (Figure S1e). These results indicate that the observed mitotic advantage is not due to a nonspecific activity of nocodazole nor to a peculiarity of one line of cells (C2C12).
To ask if this mitotic advantage applies more widely in the genome than to the pluripotency genes so far tested, we compared by RNAseq the genes transcribed in injected oocytes by interphase nuclei or mitotic chromatin. We focussed our analysis on genes that were found to be consistently expressed by interphase nuclei after nuclear transfer. One experiment indicated that 617 genes were transcribed in oocytes at least 2-fold more in mitotic nuclei compared to interphase nuclei. Of these mitotically up-regulated genes, Sox2 was 4-fold more transcribed than in interphase nuclear transfers, and over half of the 617 genes were more strongly transcribed than Sox2. The list of these genes is in Table S1.
The enhanced reprogramming from mitotic donor material could be due to an increased rate or to a greater eventual level of reprogramming. To distinguish these ideas and to measure the rate of reprogramming, we measured the incorporation of GFP-tagged histone B4 (an early marker of oocyte reprogramming) [14] and the association of Cherry-labelled histone H2B by live imaging of mixed populations of mitotic and interphase donor cells after injection into oocytes (see Figure S2a for design). Mitotic donor material becomes very rapidly marked with both histone B4 and histone H2B, whereas interphase donor nuclei show a lag in the association of both and particularly of H2B (Figure 2a). In support of a difference in the rate of reprogramming, we find that oocyte-derived TBP2 marks the transplanted mitotic cells more strongly than interphase donor cells (Figure S2b; compare white mitotic with yellow interphase arrows). We then asked if there is a more rapid association and activation of RNA polymerase II with mitotic chromatin. We used immunostaining for the elongating form of RNA polymerase II on a mixed population of mitotic and interphase nuclei injected into oocyte GVs. Mitotic donor material is clearly marked with elongating Pol II before interphase donor material (Figure 2b, compare panels ii and iv for pol II). In view of this difference between the two nuclear types in the onset of global pol II transcription after nuclear transfer, we asked whether reprogrammed genes are activated at a different rate in mitotic donor cells compared to interphase cells or if the magnitude of activation is greater. A time course of reprogramming from oocytes injected with either interphase or mitotic donor cells was assessed by RT-qPCR and revealed that genes from mitotic donor cells are activated more rapidly than the same genes from interphase cells (Figure 2c); the accumulation of transcripts reached by 63 h is similar.
We conclude that the difference in reprogramming between interphase and mitotic donor material giving this mitotic advantage reflects the rate of reprogramming rather than the eventual magnitude of transcript generation from these two types of nucleus.
The most obvious explanation for this mitotic advantage is the absence of a nuclear envelope in the mitotic karyoplasts. We have quantitated this difference in membrane permeability by time course imaging a mixture of injected interphase nuclei and mitotic karyoplasts. We carried out a “double permeabilization,” in which both the cell and nuclear membranes, of interphase or mitotic donor cells, were permeabilized as illustrated in the scheme in Figure 3a. We then compared the rate of oocyte factor uptake with the rate of reprogramming by RT-qPCR. A difference in the amount of B4 and H2B uptake is indeed seen after plasma permeabilization with digitonin (Figure 3b) but is no longer seen after double permeabilization of the nuclear envelope with Triton (Figure 3c). Nevertheless, the mitotic difference between interphase and mitotic chromatin does persist in respect of the transcriptional reprogramming of silenced genes (Figure 3d).
We have confirmed this conclusion using permeabilization by different reagents. Streptolysin 0 (SLO) permeabilizes the plasma membrane but not the nuclear membrane; SLO and Lysolecithin (LL) together permeabilize the plasma membrane and nuclear membrane [15]. Permeabilization was tested using different sizes of dextran (Figure S3a). We then compared transcription from transplanted nuclei, comparing those treated with SLO alone and those treated with SLO and LL. The transcription ratio following these two procedures shows no advantage when the nuclear envelope is permeabilized (Figure S3b).
We conclude that the presence of an intact interphase nuclear envelope does not explain the mitotic advantage.
Because the difference in reprogramming rate between interphase and mitotic donor cells is maintained after extensive permeabilization of the interphase nuclear membrane, we asked if the source of the difference lies in the chromatin of the two donor cell preparations. To answer this, we mildly sonicated both interphase and mitotic donor cell preparations to give fragments of chromatin of similar sizes (Figure 4a and b), injected these preparations in parallel with a permeabilized cell preparation into oocyte GVs, and assessed gene reactivation by RT-qPCR (Figure 4c). It is clear that the difference in the rate of gene reactivation from interphase and mitotic nuclei is maintained when the injected material is sonicated chromatin as opposed to whole nuclei. This suggests that the “mitotic advantage” is present in the chromatin of mitotic cells. This result also confirms that the difference between interphase and mitotic donor cells is not due to the interphase nuclear membrane, nor to any other aspect of nuclear organization that is eliminated by sonication.
The difference between interphase and mitotic reprogramming is, however, abolished when genomic DNA prepared from donor nuclei is injected into oocyte GVs (Figure 4d); this excludes differences at the DNA level (sequence and DNA methylation for example) as possible sources of the difference in reprogramming between interphase and mitotic samples. The possibility of DNA methylation accounting for the mitotic effect was further excluded by bisulphite analysis of specific loci on mitotic and interphase DNAs, as this revealed no mitosis-specific differences (Figure 4e). These two results indicate that whatever accounts for the difference between mitotic and interphase donor cells is not present at the level of genomic DNA itself but is in non-DNA components of chromatin.
As the mitotic advantage is likely to be due either to the loss or gain of chromatin binding factors, we removed most of these from our donor suspension of interphase nuclei by incubating such nuclei in a high-salt Triton buffer. We thereby tested whether a loss of chromatin binding factors at mitotic entry could remove the mitotic advantage. We also largely depleted chromatin binding factors from permeabilized mitotic cells and thus removed many chromatin factors that may be gained by cells entering mitosis. The depletion of chromatin binding factors was achieved with 300 mM salt and Triton, which removed most nonhistone DNA binding factors. A scheme of the cell preparation and examples of the proteins removed are shown in Figure 5a and 5c. It can be seen that the great majority of the nonhistone chromosomal proteins that have been tested and that normally exist in interphase nuclei have been removed from mitotic chromatin by 300 mM salt and Triton. Nevertheless interphase nuclei depleted of salt soluble nuclear protein (300 mM sample) do not acquire the same reprogramming responsiveness as mitotic donor material (Figure 5b). Likewise, extensive protein removal from mitotic donor material (Figure 5c) before nuclear transplantation does not abolish the mitotic advantage (mitotic 300 mM; Figure 5b), indicating that the acquisition of chromosomal proteins by mitotic nuclei does not account for this advantage.
Independently of salt release experiments, we tested topoisomerase II whose activity increases from S phase to the end of G2. The inhibition of topoisomerase II and of its adaptor molecules 14-3-3z and H3S10ph by inhibitors, inhibitory peptides, and antibody injection in transplanted mitotic nuclei did not reduce the mitotic advantage. We also found that salt release removes topoisomerase from mitotic chromatin, as in experiments shown in Figure 5c, but does not change the mitotic advantage.
We conclude that a loss of salt-soluble chromatin binding factors does not account for the mitotic advantage. It is likely therefore that the source(s) of the difference is either a non-salt-soluble factor (gained or lost at mitotic entry), a covalent modification of chromatin, or the spatial arrangement of nucleosomes.
We next considered covalent histone modifications that may be lost or gained on mitotic chromatin compared to interphase chromatin. A large number of histone modifications are associated with mitotic entry [16],[17], as well as changes in nucleosome positioning and in chromatin compaction. We first tested the most striking changes involving global histone deacetylation, phosphorylation, and some small increases in histone H3 lysine 4 and 9 methylation that have been seen on mitotic chromatin [16],[18],[19]. Histone deacetylation in mitotic cells is successfully inhibited during mitotic synchronization by the histone deacetylase inhibitor TSA (Figure 6a). Histone phosphorylation in mitotic cells is inhibited by the Aurora B/JAK inhibitor AT9823 (Figure 6b). Nevertheless, the mitotic advantage persists after both of these treatments (Figure 6c and 6d). Similarly, the removal of mitotic histone phosphorylation from the Sox2 gene by protein phosphatase treatment of mitotic and interphase donor cells before nuclear transplantation also failed to abrogate the mitotic effect (Figure S4a and b). A small (2-fold) local increase in Sox2 locus histone methylation at mitosis (Figure S4c and d), seen by ChIP in mitotic chromatin, is eliminated by the methylation of MTA (not shown), but the mitotic advantage is retained (Figure 6e).
In normal cells, histone ubiquitination (primarily H2AK119Ub and H2BK120Ub) is dramatically reduced at mitotic entry [16]. H2AK119Ub is associated with transcriptional repression [20]. Thus, a reduction in H2A ubiquitination at mitosis is an attractive candidate to explain the enhanced reprogramming of mitotic chromatin. We first tested the effect of increasing ubiquitination of mitotic chromatin, by preparing nuclei for injection in the presence of iodoacetimide (IAA), which inhibits deubiquitinases (Figure 7a) [21]. Under normal conditions, interphase chromatin is at least five times more globally ubiquitinated than mitotic chromatin (Figure 7a,b). The inhibition of mitotic deubiquitination by IAA increases the ubiquitin level in mitotic chromatin, so that it is nearly equal to that of interphase nuclei (Figure 7c). When tested for transcription in oocytes, hyperubiquitinated mitotic chromatin by IAA does not show an advantage over interphase chromatin (Figure 7d), in accord with the idea that the deubiquitinated state of normal mitotic chromatin could account for its special transcriptional advantage.
As further support for this idea, we tried to remove ubiquitin from interphase nuclei with a recombinant deubiquitinase (Ubp-M), and then tested the effect of this by oocyte injection followed by RT-qPCR. Treatment of interphase nuclei with Ubp-M removes histone ubiquitination (Figure 7g). However, unexpectedly, we see that the removal of histone ubiquitination by Ubp-M does not significantly enhance the reprogramming of interphase nuclei, so that they behave the same, in this respect, as mitotic chromatin (Figure 7h). This suggested that deubiquitination itself is not sufficient to confer mitotic advantage. We hypothesized that H2A deubiquitination is a required step in a series of chromatin remodelling events that eventually lead to mitotic advantage. We therefore chose to reduce the ubiquitinated state of interphase nuclei in living cells in order to allow events downstream of ubiquitin-depleted chromatin to take place. The inhibitor MG-132 is thought to lower histone ubiquitination by reducing the pool of free ubiquitin through inhibition of the proteasome. MG-132 treatment of interphase nuclei before injection to oocytes (Figure 7a) gave a partial but significant reduction in H2A ubiquitination on Sox2 (Figure 7e compared to 7b); it also resulted in a substantial enhancement of oocyte-induced transcription from interphase, but not mitotic chromatin (Figure 7f).
The increase in ubiquitination of transplanted mitotic nuclei, coupled with the removal of the mitotic advantage, shows that chromatin ubiquitination contributes to the mitotic advantage. It may, however, not be sufficient to explain the whole phenomenon, because we do not achieve a complete mitotic advantage in interphase nuclei by deubiquitination.
Our results show a substantial effect of the cell cycle stage of donor nuclei in nuclear transfer experiments. The reason why this has not been seen before in some of the older nuclear transfer experiments with eggs is probably for several reasons. One is that the somatic nuclei used as donors were not able to be well synchronized [22]–[24]. Another is that tests have involved the normality of development, rather than gene activity. Third, and most importantly, tests have been carried out on cell dividing eggs, whereas our work has tested gene transcription in the complete absence of DNA replication or cell division. The more recent results of [6] are in agreement with the work described here. The success of the first mammal cloning work was attributed in part to the use of donor cells in G0 [25],[26]; this result was not, however, found by others [27]–[30]. Lemaitre et al. [6] have described a dramatic effect of mitosis on the efficiency of DNA replication by egg cytoplasm, but transcription could not be tested in their extract experiments. Egli et al. [3] have proposed an important role for mitosis in permitting chromosomal protein exchange in mouse nuclear transfer experiments but not necessarily on gene transcription. Our results are therefore in accord with previous work, but reveal a specific effect of mitosis on gene transcription.
An important component responsible for the acquisition of mitotic advantage appears to be the removal of ubiquitin from histone H2A or H2B in mitotic chromatin. Ubiquitination of histones is primarily monoubiquitination, and we assume that this is the modification involved in mitotic advantage. H2A ubiquitination on lysine 119 is associated with transcriptional repression, particularly of lineage-specifying genes in ES cells [31], possibly through its association with members of the polycomb repressive proteins [32]. It could be envisaged that the deubiquitination of chromatin seen at mitotic entry permits this mitotic advantage by removing this inhibitory mark or the associated binding proteins [33]. In keeping with this idea, we have been able to partly simulate the mitotic effect by biochemically removing histone ubiquitination in interphase donor cells.
What could be the significance of the mitotic advantage identified here? In our experiments, the mitotic advantage takes place during the early stages of transcriptional activation, and is no longer seen after 2 d. In normal dividing cells, mitosis lasts for only a few hours. We therefore think that mitosis is a time when cells can most easily change their chromatin state, exchange transcription factors, and embark on a new lineage. When a cell has adopted a new fate, its daughter cells will usually follow the same lineage, unless an exchange of nuclear components takes place. The acceleration of postmitotic transcriptional activation [34] may be an associated phenomenon.
Cells were cultured in DMEM (D5796, Sigma; E15-810, PAA; 41965-062, Invitrogen) with 10% FCS (10270106, Invitrogen), 100 units/ml Penicillin-Streptomycin (15140-122, Invitrogen), and 0.25 µg/ml Fungizone (15240-096, Invitrogen). Inhibitors used for various experiments include the following: 5′-Deoxy-5′-(methylthio)adenosine (MTA)(D5011, Sigma) used at 1 µg/ml, Aphidicolin (A0781, Sigma) at 1 µg/ml, AT9823 (gift from M. Dawson) used at 100 nM; ICRF-90 (I4659, Sigma) at 1 µg/ml, iodoacetamide (IAA) made freshly and used at 10 µM, MG-132 (Sigma) used at 4 µM, Nocodazole (M1404, Sigma) at 75–100 nM; Taxol (T7402, Sigma) at 1 µM, Thymidine (T1895, Sigma), and Trichostatin A (T8552, Sigma) at 1 µg/ml.
Cell synchronization was achieved according to the scheme in Figure S1A. In general, media containing the desired inhibitor were applied to unsynchronized cells a day after seeding for 16–20 h. For mitotic cells, seeded cells were initially arrested in 2 mM thymidine for 16–24 h, washed 3× in PBS, released into fresh media for 6–12 h, and then media replaced with Nocodazole or Taxol containing media for 10–16 h, after which rounded cells were detached by “shake-off” and the culture media harvested for the mitotic cell fraction. G1 arrest was achieved by Serum starvation for 72 h.
Cells in suspension (either from mitotic shake-off or trypsinization of adherent cells) were washed twice in PBS, transferred to SuNaSP, and permeabilized with Digitonin (40–100 µg/ml) for 3 min on ice. The reaction was stopped by addition of and excess of SuNaSP-BSA and the nuclei concentrated to an appropriate volume for GV transfer [12]. Nuclear transplantation to oocyte GVs was performed as described in [12].
The following media were used for permeabilization: SuNaSP, 0.25 M Sucrose, 75 mM NaCl, 0.5 mM Spermidine, 0.15 mM Spermine; SuNaSP-BSA, SuNaSP with 3% (w/v) BSA; and HPRicLS, Final [1×] Hepes 20 mM, KCl 75 mM, MgCl2 1.5 mM.
Cells were permeabilized with Digitonin and incubated for 15 min in prebuffer (20 mM Hepes, 75 mM KCl, 1.5 mM MgCl2, 25 mg/ml Gelatin, 60 mg/ml BSA) and washed twice into permeabilization buffer (20 mM Hepes, 75 or 300 mM KCl, 1.5 mM MgCl2, 0.2% TritonX100, 12 mg/ml Gelatin, 30 mg/ml BSA). Cells were then extensively washed (20 mM Hepes, 75 mM KCl, 1.5 mM MgCl2) and resuspended in a suitable volume of SuNaSP-BSA for nuclear transplantation.
Cells were permeabilized in Digitonin, incubated in “prebuffer,” washed into permeabilization buffer (75 mM salt), washed in SuNaSP, and then transferred into a suitable reaction buffer with or without recombinant enzyme. Dephosphorylation was performed with Protein Phosphatase I (NEB, P0754S) in HPRicLS (20 mM Hepes, 75 mM KCl, 1.5 mM MgCl2) and Deubiquitination performed using recombinant enzyme prepared from insect cells (as described in [33]) in buffer (50 mM NaCl, 50 mM Tri-Cl PH 8.0, 1 mM DTT, 1× Complete EDTA-free Protease inhibitor [Roche]). After enzyme treatment, cells were washed in HPRicLS and SuNaSP and resuspended in a suitable volume of SuNaSP-BSA.
PBS washed cell suspensions were (fixed in ethanol and stained with 50 µg/ml propidium iodide). DNA content analyses were then performed on a FACSCalibur cytometer (BD Bioscience).
The following antibodies were used: αAurura B (AIM1) (Cell Signalling, 3094), αBmi1 (Cell Signalling, 6964), αBRD4 (Cell Signalling, 12183), αCTCF (Cell Signalling, 3418), αCyclinB (Cell Signalling, 4138), αDNMT1 (Abcam, ab92453), αH2AK119Ub (Cell Signalling, 8240), αH2BK120Ub (Cell Signalling, 5546), αH3 (Abcam, ab1791 and Cell Signalling, 4620), αH3K4me3 (Abcam, ab8580), αH3K9ac (Cell Signalling, 9649), αH3K9me2/3 (Cell Signalling, 5327), αH3S10ph (Sigma, H 0412), αH4 (Abcam, ab31830), αHP1α (Cell Signaling, 2616), αphosphoSer2-PolII (Covance, MMS-129R), αRunx2 (Cell Signalling, 8486), and αTBP (Abcam, ab62125).
RT-qPCR was performed as described in [11]. Unless otherwise stated, results are normalized to VegT (correcting for intrasample RNA extraction variation) and G3PDH (correcting for nuclear number differences between injected oocyte samples). Error bars indicate SME or standard deviation, and significance is determined by unpaired Student t test, with p<0.05 being considered significant. All experiments presented were single experiments representative of at least three experimental repeats unless otherwise noted.
Genomic DNA was prepared using DNeasy blood and tissue kits (69504, Qiagen), bisulphite conversion was performed using EpiTect Bisulfite Kit (59104, Qiagen), and primer sequences for DNA preparation were designed using Qiagen. Pyrosequencing was performed on a Qiagen Pyromark Q96 ID using PyroMark Gold Q96 Reagents (972804, Qiagen) and PyroMark PCR Kit (978705, Qiagen), as per the manufacturer's recommendations.
|
10.1371/journal.pntd.0007402 | Identification of Burkholderia thailandensis with novel genotypes in the soil of central Sierra Leone | The soil-dwelling bacillus Burkholderia pseudomallei is the etiological-agent of the neglected and life-threatening emerging infection melioidosis. The distribution of B. pseudomallei in West Africa is unknown. In the present study we aimed to determine whether B. pseudomallei and B. thailandensis are present in the environment of central Sierra Leone.
In June-July 2017, we conducted an environmental surveillance study–designed in accordance with existing consensus guidelines—in central Sierra Leone. A total of 1,000 soil samples (100 per site) were collected and cultured. B. pseudomallei was not identified in the soil, but we identified seven novel B. thailandensis sequence types with multi-locus sequence typing (MLST) and 16S rRNA gene sequence analyses.
The presence of B. pseudomallei was not demonstrated, however, multiple novel B. thailandensis sequence types were identified. More environmental and sequencing studies are needed to further understand the genetic diversity, evolution and virulence of these emerging organisms.
| The environmental bacterium Burkholderia pseudomallei is the cause of melioidosis, an often-fatal but neglected infection prevalent across tropical areas. B. thailandensis is a member of the B. pseudomallei complex, rarely causes disease in humans and is considered a-virulent. Modelling studies have estimated a high prevalence of B. pseudomallei in western Africa. In this study, we performed an environmental surveillance study in the West African country of Sierra Leone. Remarkably, we could not demonstrate the presence of B. pseudomallei in the soil of Sierra Leone. However, by both culture and sequencing methods, we identified multiple B. thailandensis strains and novel genotypes. Patients with debilitating B. thailandensis infection have been occasionally reported in among others Southeast Asia and the US. Environmental and sequencing studies on both B. pseudomallei and B. thailandensis are essential to further understand the genetic diversity and evolution of these neglected but emerging organisms.
| The Gram-negative environmental bacterium Burkholderia pseudomallei is the etiological agent of melioidosis, an emerging but neglected infectious disease. Disease presentations vary from abscess formation to fulminant sepsis [1]. Melioidosis has a mortality up to 50% in low resource settings and is predominantly found in Southeast-Asia and northern Australia [1]. Infection with B. pseudomallei primarily occurs in people who are in regular contact with soil and water [1,2]. B. thailandensis is a member of the B. pseudomallei complex, is considered a-virulent [3,4] and rarely causes disease in humans [5–10]. Knowledge about the global distribution of B. pseudomallei and B. thailandensis, however, is limited.
Patients from sub-Saharan Africa reported with melioidosis are few and isolated (e.g. the The Gambia, Burkina Faso, Nigeria and Gabon), which most probably is the result of under-recognition and under-reporting. These cases may represent the ‘Tip of the Iceberg’ [1,11]. From the West African country of Sierra Leone, only one case of melioidosis has been reported [12]. Modelling studies, however, estimate that in Sierra Leone annually hundreds of patients suffer from melioidosis, of which the vast majority will die [13]. The tropical climate, heavy rains and abundant rice farming in central Sierra Leone all contribute to the high pre-odds likely-hood for the presence of B. pseudomallei and B. thailandensis in its soils [13]. In the present study we aim to determine whether B. pseudomallei and B. thailandensis are present in the soil of central Sierra Leone.
During the rainy season (June-July 2017), an environmental surveillance study–designed in accordance with existing consensus guidelines [14]—was conducted in Tonkolili and Bombali districts, central Sierra Leone (Fig 1A and 1B). Ten different sites were selected based on local maps and consultations with inhabitants on factors known to be associated with B. pseudomallei and B. thailandensis presence (e.g., wet soil such as rice paddies or land use such as goat and cattle farming) [13,14]. Oral informed permission was obtained from landowners and written informed permission from the paramount chief of Yele, Sierra Leone, prior to soil sampling. We used a fixed-interval sampling grid of five meters between soil samples. Thirty grams of soil from a depth of 65 cm [15] was taken for each sample, stored from direct sunlight and kept at room temperature until further processing (for details on geographical features and distribution see Table 1). Culture of suspected Burkholderia isolates from soil samples was done as described previously [14,16]. First, 10 grams of each soil sample was diluted in 10 mL of threonine-basal solution containing colistin at 50 mg/L (TBSS-C50 broth) and crystal violet. This mixture was vortexed and subsequently incubated at 40 °C for 48 hours. Ten μL of the upper layer of enrichment medium was subcultured onto an Ashdown-agar plate and checked for suspected Burkholderia colonies every 24 hours for a period of 7 days. Initial identification of suspected Burkholderia growth was done by colony morphology, positive oxidase test result, antimicrobial drug susceptibility pattern (susceptible to amoxicillin-clavulanic and resistant to gentamicin and colistin) and latex-agglutination tests [14,16] (see S1 Fig for a schematic overview of the culture methodology). All experiments were conducted in a biosafety level 3 laboratory.
All isolates were identified as B. thailandensis by MLST as well as 16S sequence analyses (see S1 Table). B. pseudomallei was not found. Four B. thailandensis isolates corresponded to the earlier described sequencing type 73 (ST73) [17], but the remaining 28 isolates were of seven novel sequence types (ST1677 through ST1683). ST1677 and ST1680 are single locus variants of ST73, while ST1678, ST1679, ST1681 and ST1682 are double locus variants of ST73 and ST1683 differed at three loci from ST73 (see appendix).
Multi-locus sequence typing (MLST) and 16S rRNA gene sequence analyses were used to determine the species of the isolated Burkholderia isolates using the MiSeq platform (Illumina, San Diego, CA, USA) as described previously [16]. Assembly of the reads was performed with help of SPAdes 3.9. MLST data analysis was performed based on partial sequences of seven housekeeping genes (see appendix). Allele numbers and sequence types deduced from MLST allelic profiles were assigned using the BIGSdb B. pseudomallei database (https://pubmlst.org/bpseudomallei/). Cluster analysis was subsequently performed in the MEGA 6.06 using the UPGMA algorithm and the Jukes-Cantor model. Bootstrap test was performed for 500 repetitions.
Oral informed permission was obtained from landowners and written informed permission from the paramount chief of Yele, Sierra Leone, prior to soil sampling.
A total of 1,000 samples (100 per site) were collected at ten sampling sites in the Tonkolili and Bombali District, central Sierra Leone (see Table 1). Initial identification methods [14,16], led to the isolation of 32 Burkholderia strains from 25 soil samples. Four Burkholderia strains showed a negative latex-agglutination test; the rest showed (possible) positive latex-agglutination test results.
Two main clusters are presented in a phylogenetic tree based on the concatenated sequences of the seven household genes of all B. thailandensis STs available in the PubMLST database (Fig 2). Cluster I contains exclusively isolates from Asia and Oceania, while cluster II comprises isolates from all isolates from Sierra Leone and the one from Gabon (ST1126). One isolate with ST537 was an outlier. Interestingly, ST1126, ST696 and ST101 were identified to express a B. pseudomallei-like capsular polysaccharide (BTCV) [18] possibly explaining why many of the isolated B. thailandensis showed cross-reactivity with the B. pseudomallei latex-agglutination test. MLST data and microarray based comparative genomic hybridization revealed earlier that there is a separate subgroup of B. thailandensis isolates (ST696, ST101 and ST73) containing BTCV strains, which are genetically different from the other B. thailandensis isolates [17].
In this study, two clearly separated B. thailandensis clusters (I and II) were observed. Various studies have reported human infections by B. thailandensis belonging to both cluster I (ST77, ST80 and ST345) and cluster II (ST73 and ST101) (https://pubmlst.org/bpseudomallei/) [5–10]. It has been postulated that B. thailandensis isolates within cluster II are more virulent than those in cluster I [17], but evidence has not been reported [6,17]. Clinical characteristics are indistinguishable from B. pseudomallei infection and include soft tissue infection, abscess formation, pneumonia and sepsis [5–10]. The most recently described B. thailandensis case occurred in a 29-year old diabetic woman with an infected wound and swelling of her forearm after a car incident in Arkansas, US [6].
Our study has several limitations, including the lack of standard blood culture services for febrile patients across Sierra Leone. This limits targeted soil sampling studies centred around an index case. In addition, we cannot dismiss the possibility of sampling-error, although consensus guidelines were followed [14]. More knowledge about the exact composition of soils in which B. pseudomallei and B. thailandensis reside could help to determine which sites to study in future environmental surveillance studies. Furthermore, bacteria could have been present in a viable, but not cultivable state. It remains difficult to differentiate B. thailandensis and other members of the B. pseudomallei complex from B. pseudomallei by methods commonly available in clinical labs, even in developed countries, which may result in diagnostic confusion. Therefore, improving the detection and differentiation of members of the B. pseudomallei complex to improve patient care and appropriate public health responses is desired. Taken together, B. pseudomallei was not cultured from the soil of central Sierra Leone, but B. thailandensis with novel genotypes were found. B. thailandensis infection in humans have been sporadically reported in the literature in both the US and Asia [5–10]. B. thailandensis in general is considered a-virulent. As a result, clinical disease attributed to B. thailandensis is important. This also holds true for to the melioidosis research community, because no strict biocontainment conditions for B. thailandensis are required. The true clinical relevance of this soil-dwelling bacillus, however, remains to be elucidated. We encourage further environmental and sequencing studies on both B. pseudomallei and B. thailandensis to further understand the genetic diversity, virulence and evolution of these emerging organisms.
|
10.1371/journal.pntd.0001847 | Deaths from Symptomatically Identifiable Furious Rabies in India: A Nationally Representative Mortality Survey | It is estimated that India has more deaths from rabies than any other country. However, existing estimates are indirect and rely on non-representative studies.
We examined rabies deaths in the ongoing Million Death Study (MDS), a representative survey of over 122,000 deaths in India that uses enhanced types of verbal autopsy. We estimated the age-specific mortality rates of symptomatically identifiable furious rabies and its geographic and demographic distributions. A total of 140 deaths in our sample were caused by rabies, suggesting that in 2005 there were 12,700 (99% CI 10,000 to 15,500) symptomatically identifiable furious rabies deaths in India. Most rabies deaths were in males (62%), in rural areas (91%), and in children below the age of 15 years (50%). The overall rabies mortality rate was 1.1 deaths per 100,000 population (99%CI 0.9 to 1.4). One third of the national rabies deaths were found in Uttar Pradesh (4,300) and nearly three quarters (8,900) were in 7 central and south-eastern states: Chhattisgarh, Uttar Pradesh, Odisha, Andhra Pradesh, Bihar, Assam, and Madhya Pradesh.
Rabies remains an avoidable cause of death in India. As verbal autopsy is not likely to identify atypical or paralytic forms of rabies, our figure of 12,700 deaths due to classic and clinically identifiable furious rabies underestimates the total number of deaths due to this virus. The concentrated geographic distribution of rabies in India suggests that a significant reduction in the number of deaths or potentially even elimination of rabies deaths is possible.
| Rabies, a disease of antiquity, has been partially controlled in many countries and eliminated in a few. However, according to the World Health Organization, rabies continues to kill thousands of people in India each year, more than in any other country. We used an enhanced type of verbal autopsy (a structured interview of the relatives or close associates of the dead by non-medical staff with central medical coding by at least two doctors) to identify the causes of over 122,000 deaths in a large scale, representative sample in India in 2001–03. Using these data, we estimate that in 2005 approximately 12,700 people died from symptomatically identifiable furious rabies. Because verbal autopsy is not able to identify atypical presentations of rabies, our figure underestimates the actual number of rabies deaths in India. The majority of rabies deaths occurred in males, in rural areas, in children below the age of 15 years, and in a few states. The concentrated geographic distribution of rabies in India suggests that targeting with preventive campaigns including vaccination of animals and post exposure vaccination of humans might achieve a significant reduction in the number of deaths or potentially even elimination of deaths from this disease.
| Rabies has been recognized for many millennia in India, long before Aristotle recognized the disease in the Graeco-Roman era [1]. The ancient Vedic text “Sushruta Samhita” contains graphic descriptions of rabies in animals and in humans: “If the patient becomes exceedingly frightened at the sight or mention of the very name of water, he should be understood to have been afflicted with Jala-trsisa (hydrophobia) and be deemed to have been doomed” [2].
Several indirect estimates [3]–[4] have suggested that modern India has more rabid dog bites and human rabies deaths than any other country. In 2002, the World Health Organization (WHO) estimated that rabies caused 30,000 human deaths per year in India, which accounted for approximately 60% of the estimated global total of rabies deaths [5]. A non-representative survey based on case detection of rabies, and verbal autopsies of identified furious rabies cases, estimated about 17,000 human rabies deaths for the whole country [3]. This total was further expanded by 20% to account for paralytic and atypical forms and resulted in the widely quoted final figure of just over 20,000 rabies deaths per year. In 2004, a dog-bite probability model was used to re-evaluate the burden of rabies in Africa and Asia. This method also yielded an estimate of about 20,000 human deaths from rabies in India [4].
All these estimates are much higher than the Government of India's official reported deaths in the range of 244 to 556 per year between 2000 and 2009 [6] based on routine hospital surveillance which is likely to miss many rabies deaths. The official Government of India reports of rabies deaths from hospitals are underestimates for several reasons. Most deaths in India occur at home, in rural areas, outside medical care, and there are very large numbers of stray dogs throughout India which frequently bite humans [7]–[9]. In many states, a lack of community access to education about post-exposure rabies prophylaxis and adherence to traditional beliefs about the disease are likely to increase the risk of developing rabies after exposure. Laboratory confirmation of rabies in humans or animals in India is rarely possible. Typical signs and symptoms of classic “furious” rabies are striking and uniquely characteristic and are therefore well recognized by both medical staff and lay people. However, paralytic “dumb” rabies and atypical presentations may easily be misdiagnosed as other neurological entities [10]–[13].
Effective dog rabies control, and possibly elimination, is achievable in India [14]–[15]; however, data on the prevalence of the disease and its distribution across the states are required to raise public awareness, give direction to control programmes, and to establish a basis against which to measure the success of future efforts to reduce rabies transmission or deaths. Here, we provide an estimate of national and regional human rabies mortality based on a nationally representative direct survey of over 122,000 deaths in India. We focus on understanding the geographical, age, and gender distributions of rabies deaths.
Following each 10-yearly census, the Registrar General of India (RGI) divides India into approximately one million units, each containing about 1,000 people. In 1993, the RGI randomly selected 6,671 of these units from the 1991 census, from all 28 states and 7 union territories of India, to be included in its Sample Registration System (SRS). The SRS is representative of India at the rural/urban stratum for the major states of India. Each unit has about 150 households (totaling 1.1 million households and approximately 6.3 million people), which are monitored for vital events on a monthly basis by a part-time enumerator and every 6 months by a full-time surveyor. The Million Death Study (MDS) seeks to assign causes to all deaths in the selected SRS areas for the period from 2001 to 2014 [16]–[21].
Verbal autopsy is a tool used to ascertain cause of death based on a structured interview with the relatives or close associates of the dead, in areas where medical certification of the cause of death is lacking. As part of the MDS, an enhanced type of verbal autopsy, using both an open-ended narrative and close-ended questions [16], [22] (termed RHIME: Routine, Reliable, Representative and Re-sampled Household Investigation of Mortality with Medical Evaluation), was administered by trained RGI surveyors for each identified death starting from 2001. Two of 130 trained physicians independently reviewed each completed RHIME and assigned a single cause of death using the International Classification of Diseases 10th revision (ICD-10) [23] and specific guidelines developed for the MDS [24]. Differences in coding were resolved by anonymous reconciliation of initial codes, and if needed, by a third, senior physician who adjudicated the final cause of death. Details of the methods, validation and preliminary results for various conditions have been reported elsewhere [16]–[19],[25]. About 5% of deaths in the MDS sample were randomly re-sampled and subsequently independently re-interviewed by teams other than the SRS staff.
From the MDS data available (2001–2003), we identified all deaths in which at least one physician had coded rabies (ICD-10 code A82) or dog bite (ICD-10 code W54) as the cause of death. All non-English narratives were translated into English and data were extracted in a standardized fashion. Based on a preceding history of exposure to a dog [or other mammal] bite combined with symptoms such as altered behavior, hydrophobia, psychosis/delirium/confusion, and fever, the causes of deaths were classified as either rabies or not rabies by the authors.
We further characterized the rabies deaths by gender, age, urban or rural location, and region. To account for sampling design, the age-specific proportions were weighted according to the SRS sampling fractions in the rural and urban parts of each state [18],[20],[26], although such sampling made little difference to the estimated national totals. Using methods described previously, the proportion of deaths coded as rabies was applied to the United Nations (UN) Population Division estimates of deaths in India in 2005 [27] to generate rabies specific death totals and rates for India and its major states.
SRS enrolment is on a voluntary basis, and its confidentiality and consent procedures are defined as part of the Registration of Births and Deaths Act, 1969. Oral consent was obtained in the first SRS sample frame. The new SRS sample obtains written consent at baseline. Families are free to withdraw from the study, but the compliance is close to 100%. The study poses no or minimal risks to enrolled subjects. All personal identifiers present in the raw data are anonymized before analysis. The study has been approved by the review boards of the Post-Graduate Institute of Medical Education and Research, St. Michael's Hospital and the Indian Council of Medical Research.
A total of 95 of the 122,429 surveyed deaths in 2001–3 were coded as rabies by at least one physician. An additional 59 cases were coded as dog bite. Following central review of the details of each of these dog bite deaths, 45 were re-classified as rabies, arriving at a total of 140. The majority of rabies deaths occurred in rural areas (91%) and few occurred in health care facilities (16%) (Table 1). About 97% of rabies deaths were the result of dog bites and the remaining 3% were from cat and wild mammal bites. The median time from a bite to death was 8 weeks (range 1 week to 4 years). Hydrophobia was described in 22% of rabies deaths and other neuropsychiatric symptoms, such as altered behavior (49%), psychosis/delirium/confusion (21%), restlessness (14%), barking/cough (18%), and dysphagia (6%) were also mentioned in the narratives.
Among the treatment histories of patients detected by our survey, 65% (91/140) had not sought any hospital treatment. While we are not able to infer the specific nature of treatment sought, 34% (48/140) received one or more injections after their most recent bite. However, only one patient completed a course of 14 injections, which constitutes complete treatment with the rabies vaccine most commonly used in India at the time of our study. Most of the remaining 47 patients received only 1–3 injections, though 5 patients received 4–10 injections (Table 1).
Projection of the 2001–3 survey deaths from rabies to 2005 UN death totals, yields 12,700 (99% CI 10,000 to 15,500) symptomatically identifiable furious rabies deaths in India (Table 2). Approximately 62% of all rabies deaths in India in 2005 were in males and 50% were in children under 15 years. The overall rabies mortality rate was 1.1 deaths per 100,000 population (99% CI 0.9 to 1.4), with the highest rates being in children under 5 years and in the elderly age 70 years or older.
Rabies deaths were not evenly distributed throughout the country. One third of all rabies deaths were found in Uttar Pradesh (4,300) and nearly three quarters (8,900) were in 7 central and south-eastern states: Chhattisgarh, Uttar Pradesh, Odisha, Andhra Pradesh, Bihar, Assam and Madhya Pradesh. Among larger states, the highest rates of rabies death per 100,000 population were in Chhattisgarh (3.5), Uttar Pradesh (2.3), and Odisha (1.9). (Figure 1 and Table S1). No rabies deaths were reported in study areas from the following states: Kerala, Jammu & Kashmir, Jharkhand, Manipur, Meghalaya, Nagaland, Sikkim, Mizoram, Andaman & Nicobar Islands, Lakshadweep, Chandigarh, Dadra & Nagar Haveli and Daman & Diu. Together, these states represent approximately 7% of India's population.
Of the 5% (n = 3275) MDS sample deaths randomly chosen for independent re-sampling and re-administration and coding of the VA, 2 were originally coded as rabies. Both of these deaths were again identified as rabies in the re-sampling process and there were no other rabies deaths
This study is the first to provide an estimate of deaths from symptomatically identifiable furious rabies based on a representative sample of Indian deaths and to report the geographic, age and gender distributions of these deaths. While the MDS was not designed specifically to identify rabies deaths, its large size, and representative sampling make it suitable for identifying deaths due to relatively rare conditions and subsequently generating reliable estimation of population based rates. Our figure of 12,700 (99% CI 10,000 to 15,500) human deaths from rabies in 2005 is within the uncertainty ranges of a recent indirect estimate by Sudarshan and colleagues of 17,137 (95% CI 14,109–20,165) prior to the addition of 20% to account for paralytic/atypical forms of the disease [3]. While the Sudarshan study also used verbal autopsies, it relied on case finding in communities located near large medical centers followed by interviews of people in the communities in which the cases originated and thus cannot be considered a truly nationally representative sample. Similarly, the derivation of 19,713 (95% CI 4,192–39,733) human deaths using a dog-bite probability model is based on several assumptions [4], most notably that the epidemiology of canine rabies in India, where very few dogs are tested for rabies, is similar to that in Africa. To our knowledge, there have been no nationally representative studies of canine rabies in India. Despite these methodological challenges, the three studies together suggest a range of rabies deaths between 13,000–20,000 deaths. Although we did not report any rabies deaths in a small number of states (which represent less than 7% of India's population and total deaths), routine government hospital data [6] and medically certified causes of death from urban areas [28] from 1998 to 2004, would add only about an additional 100 to 500 rabies deaths from these states (Figure S1). Thus, the inclusion or exclusion of these states does not alter our national estimate of 12,700 deaths and lies well within the 99% confidence range of our estimates (10,000–15,500).
To further compare our rabies mortality estimates with other estimates, we plotted the proportional mortality from rabies for each of the years from 2001–2003 of the MDS and the estimated proportional mortality of rabies from various government surveys and other published studies over a two-decade period (Figure 2). This figure shows that our estimate of proportional mortality for rabies (1.3 per 1000 deaths) is consistent with other data sources and also with the apparent steady decrease in rabies as a cause of death in India starting in the early 1990s. Figure 2 also suggests a crude cyclical pattern of deaths.
The demographic characteristics of our estimates were generally similar to those reported by other epidemiological studies in India. Sixty two per cent were males (compared to 71% [28], 72% [29], and 66% [9]) and 50% were children less than 15 years old (compared to 35% [28], 28% [29], and 54% [9]).
While the MDS was not designed to examine rabies treatment, we were nonetheless able to extract useful information from the narratives. The completely treated cases probably received the Semple-type rabies vaccine that was still widely used in India during the study period (2001–03) [30]. The partially treated cases might have received Semple or cell culture rabies vaccine, tetanus toxoid, antibiotics, another drug, or a traditional remedy. Since treatment information was contained only in the narrative, we are not able to comment on the timing or specific contents of the injections received by the deceased.
The most important limitation in our study is the potential for misclassification of rabies deaths as other causes of death. Some rabies deaths were in fact misclassified as being directly due to dog bite, but central review enabled correction of this misclassification. Death with dramatic neurological symptoms (including the pathognomonic symptom of hydrophobia) occurring weeks or months after a dog bite would seem to be a distinctive event that would readily be detected by verbal autopsy. However, it is well recognized that not all human patients develop typical furious rabies [31]–[32] and some may die after a short illness, before the signs are recognized or the history of an animal bite is elicited and others may have a long incubation period, in exceptional cases up to about 20 years [33]. Verbal autopsy is unlikely to be able to identify such cases. Furthermore, an unknown proportion of human rabies victims in India develop more insidious paralytic or atypical features without hydrophobia or alternating excitation and lucidity, making it unlikely that rabies will be identified as the cause of death by their family, neighbors or medical staff [10]–[11],[34]–[35]. Paralytic rabies most often resembles other encephalomyelitides or Guillain-Barré syndrome/Landry's paralysis, but many other atypical presentations of rabies have been reported [12],[36]–[39]. The proportion of rabies cases presenting with paralytic or atypical symptoms is unknown, although estimates of “less than a fifth” [40] or one third [41] have been suggested but with little, if any supporting evidence. In the MDS, approximately 8.7% of captured deaths were deemed to be due to unspecified or ill-defined causes. We do not believe it likely that deaths due to typical rabies are included in this group. While it is possible that there are atypical cases of rabies included in this group, we believe that this number would be very small. Since verbal autopsy is unlikely to identify paralytic or atypical rabies deaths, our estimates presented in this study are restricted to typical, clinically identifiable classic furious rabies. Furthermore, human rabies cases often cluster geographically around a particular rabid dog that bites multiple people. The SRS was not specifically designed to identify such clustered events, and our results might therefore be under-estimating the true rabies mortality rate.
Finally, the most recent data available for analysis from the MDS is from deaths that occurred in 2001–2003. While it would have been preferable to have utilized more recent data, no other more recent nationally representative source of comparable data exists. MDS data collection is continuing and we will update our analysis, including for time trends, when newer data are available.
We estimate that there were 12,700 deaths due to symptomatically identifiable furious rabies in India in 2005. It is very important to note that this figure underestimates the total number of deaths due to rabies since paralytic and atypical cases would not have been detected by verbal autopsy.
This study is the first to estimate rabies mortality based upon a nationally representative sample of deaths rather than modeling or from extrapolation from selected focal surveillance. Thus we provide previously unavailable regional and demographic information about human rabies deaths that can help to focus both human and canine rabies control programmes in the country and act as a baseline that can be used as comparison for future estimates of rabies mortality. Elimination of the canine reservoir of rabies is not likely in India at anytime in the near future. However, the concentrated geographic distribution of rabies in India suggests that a significant reduction in the number of human deaths or potentially even elimination of rabies deaths is possible and this study serves as a baseline against which future gains may be measured.
|
10.1371/journal.ppat.0040011 | Interaction of Polymerase Subunit PB2 and NP with Importin α1 Is a Determinant of Host Range of Influenza A Virus | We have previously reported that mutations in the polymerase proteins PB1, PB2, PA, and the nucleocapsid protein NP resulting in enhanced transcription and replication activities in mammalian cells are responsible for the conversion of the avian influenza virus SC35 (H7N7) into the mouse-adapted variant SC35M. We show now that adaptive mutations D701N in PB2 and N319K in NP enhance binding of these proteins to importin α1 in mammalian cells. Enhanced binding was paralleled by transient nuclear accumulation and cytoplasmic depletion of importin α1 as well as increased transport of PB2 and NP into the nucleus of mammalian cells. In avian cells, enhancement of importin α1 binding and increased nuclear transport were not observed. These findings demonstrate that adaptation of the viral polymerase to the nuclear import machinery plays an important role in interspecies transmission of influenza virus.
| The natural hosts of influenza A viruses are aquatic birds. On rare occasions these viruses may be transmitted to humans and then give rise to an influenza pandemic. Human influenza is therefore a typical re-emerging infection. Evidence is increasing that the viral polymerase, an enzyme that has to enter into the nucleus of the infected cell in order to promote replication and transcription of the viral genome, is a major determinant of host range. Thus, in a comparative study of an avian influenza strain and its mouse adapted variant we have previously shown that adaptation to mice depended exclusively on mutations in the polymerase proteins. These findings supported the concept that adaptation of the polymerase to host factors is an important mechanism underlying interspecies transmission. In the present study, we have identified importin α1, a component of the nuclear pore complex, as such a host factor. We show that adaptive mutations in polymerase subunits improve binding to importin α1 in mammalian, but not in avian cells. As a result, nuclear transport of these proteins and efficiency of replication are enhanced in mammalian cells. These observations demonstrate that the interaction of the viral polymerase with the nuclear import machinery is an important determinant of host range.
| The natural host of influenza A viruses is waterfowl where these agents occur in large variety defined by 16 hemagglutinin and 9 neuraminidase subtypes. Avian influenza viruses are the source of devastating outbreaks in poultry. Moreover, because of their potential to cross species barriers, to adapt to new hosts, and to cause on rare occasions pandemics, they are also a constant threat to human health [1]. As members of the Orthomyxoviridae family influenza A viruses have a segmented RNA genome of negative polarity. The eight segments are present in enveloped virus particles as ribonucleoprotein (RNP) complexes with the nucleocapsid protein (NP) and the three subunits of the RNA-dependent RNA-polymerase (PB1, PB2, PA). In the infected cell, the polymerase is responsible for transcription and replication of the viral genome [2,3].
Evidence is increasing that the viral polymerase plays a major role in host adaptation. Thus, in a comparative study of the avian strain SC35 (H7N7) and its mouse-adapted variant SC35M we showed that adaptation to mice was the result of seven mutations in the polymerase proteins, six of which (L13P and S678N in PB1, D701N and S714R in PB2, K615N in PA, and N319K in NP) were responsible for enhanced polymerase activity in mammalian cells. In avian cells, replication and transcription of SC35M were reduced, whereas both activities were increased with SC35. Furthermore, the differences in polymerase activity were paralleled by differences in pathogenicity: SC35 was highly pathogenic for chickens and had reduced pathogenicity in mice, whereas the opposite was the case with SC35M. Thus, the efficiency of the viral polymerase is a determinant of both host specificity and pathogenicity [4,5]. PB2 mutation D701N has also been implicated in the adaptation of H5N1 viruses to mammalian hosts [6,7] but other mutations may be involved, too, notably PB2 mutation E627K [8–13]. All of these findings support the concept that adaptation of the polymerase to host factors is an important mechanism underlying interspecies transmission [5].
The influenza virus polymerase is active in the nucleus of the infected cell where cytoplasmicly expressed PB1 and PA appear to be imported as a subcomplex which then assembles with separately imported PB2 [14]. There is evidence that the different polymerase subunits use the classical import pathway of the host cell depending on the recognition of a nuclear localization signal (NLS) of the cargo protein by an importin α/β dimer as well as non-classical pathways that rely on direct interaction of the cargo with an importin β homologue receptor [15]. Thus, PB1-PA dimers enter the nucleus via a non-classical transport pathway by binding to RanBP5 [16]. NP, in contrast, binds to importin α1/α2 [17] indicating that it uses the classical pathway. The same route is used for the nuclear transport of PB2 as demonstated by a recent crystallographic study showing that the C-terminus of this protein forms a complex with importin α5 [18]. Interestingly, the authors of this study hypothesized that PB2 mutation D701N which is also a host range determinant as pointed out above, may affect the binding of importin α5 to PB2. These observations suggest that importins belong to the host factors to which the polymerase adapts during interspecies transmission.
To test this hypothesis we have compared now in avian and mammalian cells the interaction of the polymerase complex of SC35 and SC35M with importin α1, which not only binds to NP, but is also the most abundant importin in a variety of human cells [19] and may function as an interchangeable housekeeping transport factor [20]. We show that adaptative mutations D701N and N319K improve binding of PB2 and NP, respectively, to importin α1 in mammalian, but not in avian cells. As a result, the efficiency of the transport of these proteins into the nucleus of mammalian cells is enhanced. These data support the notion that the interaction of PB2 and NP with importin α1 plays an important role in determining host range and pathogenicity of influenza A viruses.
To test the hypothesis that the polymerase proteins of SC35 and SC35M differ in their interaction with the nuclear import machinery of mammalian and avian cells we have analyzed the binding of PB2 and NP to importin α1 in co-immunoprecipitation experiments. Cultures of human 293T cells and quail embryo cells (CEC-32) were transfected with plasmids encoding PB2 or NP. Importin α1 complexed with PB2 or NP was precipitated from lysates of transfected and mock transfected cells using specific antibodies against PB2 or NP, which were covalently coupled to an amine-reactive gel. Bound proteins were eluted three times from the columns using acidic elution buffer to allow quantitative estimation of interaction partners. All three eluates were used to detect importin α1, PB2 and NP, respectively, by Western blot analysis. In 293T cells transfected with PB2 of SC35M or PB2 D701N, 4 to 7 times more importin α1 was bound than in cells transfected with PB2 of SC35 (Figure 1A and 1C). Likewise, NP derived from SC35M precipitated 2 times more importin α1 from 293T cells than SC35 NP (Figure 1B and 1D). In contrast, SC35 PB2 and SC35M PB2 bound similar amounts of importin α1 when CEC-32 cells were transfected (Figure 2A and 2C). There was also no difference in importin binding efficiency when SC35 NP and SC35M NP were expressed in avian cells (Figure 2B and 2D). These observations indicate that the adaptive mutations PB2 D701N and NP N319K specifically improve the binding of PB2 and NP to mammalian importin α1 adaptor protein.
It was now of interest to find out if increased binding of NP and PB2 to importin α1 affected nuclear transport in mammalian cells. First, we have carried out Western blot experiments to determine importin α1 in nuclear and cytoplasmic fractions prepared from infected cells early and late in the replication cycle. Since replication kinetics were different in avian and mammalian cells, time points with comparable virus titres were selected (6 h and 12 h p.i. in 293 T cells, 9 h and 18 h p.i. in CEC32 cells) (Figure S1). Early after infection of 293T cells with SC35M, we found a distinct accumulation of importin α1 as well as importin ß1 and NP in the nucleus. This phenomenon was not observed with SC35 (Figure 3A). Late in infection nuclear accumulation of these proteins was seen neither with SC35 nor with SC35M (Figure 3B). Nuclear accumulation of importin α1 was also not detected in avian cells early or late after infection with either virus. Furthermore, the expression levels of SC35 and SC35M NP were similar in the nuclear and cytoplasmic fractions of CEC-32 cells at both time points (Figure 3C and 3D). When unfractionated cell lysates were analyzed, no differences in the amounts of importin α1 were detected, irrespective of cell type, virus strain, and time after infection (Figure S2). This observation indicates that the increased amount of importin α1 seen in the nuclei of 293T cells early after infection with SC35M was the result of increased transport into the nucleus and did not reflect up-regulated importin α1 synthesis. We have then determined the kinetics of nuclear accumulation of importin α1. 293T cells were infected with SC35M or SC35, and the intracellular localization of importin α1 was determined at different intervals by immuno-fluorescence analysis (Figure 4). In cells infected with SC35M, importin α1 was gradually shifted from the cytoplasm to the nucleus where it was almost exclusively present 6 h p.i. Later in infection it was detected again in the cytoplasm. In cells infected with SC35, importin α1 was equally distributed between cytoplasm and nucleus throughout the replication cycle. Taken together, these results demonstrate that SC35M induces specifically in mammalian cells a transient accumulation of importin α1 in the nucleus with a peak at 6 h p.i. which presumably coincides with the maximal replication rate of the virus.
To identify the mutations in the polymerase proteins of SC35M responsible for the nuclear accumulation of importin α1, we have first analyzed single gene reassortant (SGR) viruses containing one of the polymerase genes of SC35M in a SC35 background. Cells infected with SGR viruses were separated into nuclear and cytoplasmic fractions, and importin α1 and importin ß1 localization was determined (Figure 5). Only SGR viruses SC35-PB2SC35M and SC35-NPSC35M containing the PB2 or the NP gene of SC35M showed enhanced importin α1 and importin ß1 levels in the nuclear fractions, although the increase was not as distinct as with the parental SC35M virus (Figure 3). With the SGR viruses containing PB1 or PA of SC35M, there was no importin accumulation in the nucleus (Figure 5). Since SC35M NP differs from SC35 NP only by one amino acid substitution (N319K), it can be concluded that this mutation contributes to nuclear accumulation of importin. SC35 PB2 displays two amino acid substitutions contributing to increased mouse virulence (D701N and S714R). To determine which of these mutations controls nuclear transport of importin, we infected cells with single point mutant (SPM) viruses SC35-PB2701N and SC35-PB2714R containing only one of these mutations. Nuclear accumulation of importin α1 was observed in SC35-PB2701N infected, but not in SC35-PB2714R infected cells (Figure 5). These findings indicate that mouse adaptation mutations PB2 D701N and NP N319K promote nuclear accumulation of importin α1 in mammalian cells.
Some of the data described so far suggested already that nuclear accumulation of importins in SC35M infected mammalian cells is accompanied by increased NP transport into the nucleus. To further analyze the effect of the adaptive mutations in NP and PB2 on the intracellular localization of these proteins, we have performed immunofluorescence assays in mammalian cells. While PB2 of SC35M was predominantly located in the nucleus of human lung cells (A549), most of the cells infected with SC35 showed a nuclear and cytoplasmic distribution of PB2 (Figure 6A). This observation indicates that SC35M PB2 is transported into the nucleus more efficiently than SC35 PB2. Increased nuclear transport of SC35M PB2 correlated with the nuclear accumulation of importin α1 as indicated by a distinct co-localization of PB2 and importin α1 in the nucleus. On the other hand, in cells infected with SC35, importin α1 was present throughout cytoplasm and nucleus as was the case with PB2 (Figure 6A). When A549 cells were infected with SGR virus SC35-PB2SC35M and SPM virus SC35-PB2701N, PB2 was also concentrated in the nucleus where it co-localized with importin α1 (Figure 6A). Nuclear accumulation of PB2 was not observed in cells infected with SC35-PB2714R (data not shown). We have then analyzed the nuclear import of NP by immunofluorescence (Figure 6B). NP of SC35M and SC35-NPSC35M was located predominantly in the nucleus, whereas a large part of SC35 NP was found in the cytoplasm. In cells infected with SC35M and SC35-NPSC35M, importin α1 was concentrated in the perinuclear region where it partly co-localized with NP. In SC35 infected cells importin α1 prevailed in the cytoplasm. Similar results have been obtained when the intracellular localization of these proteins was analyzed by immunofluorescence in 293T human kidney cells (Figure S3). Taken together, these observations indicate that, in mammalian cells, mutations PB2 D701N and NP N319K enhance the efficiency of the nuclear import of PB2 and NP as well as importin α1.
Intracellular localization of NP, PB2 and importin α1 has also been analyzed in CEC-32 cells. With both viruses, PB2 was clearly present in the nucleus, but it was also detected in the cytoplasm. Importin α1 co-localized with PB2 in the cytoplasm as well as the nucleus (Figure 7A). NP displayed a similar intracellular distribution pattern. Again, there was co-localization with importin α1 in the nucleus and in the cytoplasm, and there were no differences between SC35 and SC35M (Figure 7B). Thus, it appears that PB2 and NP of the avian and the mouse-adapted virus are transported with the same efficiency into the nucleus of avian cells.
We show here that mutations D701N in PB2 and N319K in NP responsible for mouse adaptation of the avian influenza virus SC35 enhance binding of these proteins to importin α1 of mammalian origin and, thus, improve the efficiency of their transport into the nucleus of mammalian cells. Increased transcription and replication activities in mammalian cells [4,5] are therefore, at least in part, the result of facilitated recruitment of polymerase subunits into the nucleus. In avian cells, neither enhancement of importin binding and nuclear transport nor increased polymerase activity [5] were observed. The differences in the nuclear transport of PB2 and NP were reflected by the efficiency of virus replication. In human lung cells SC35M grew to virus titres 2 logs higher than SC35. The SGR virus SC35-NPSC35M grew as well as SC35M, whereas titres of SGR virus SC35-PB2 SC35M were slightly reduced. In avian cells, parental and SGR viruses grew to equally high titres (Figure S5). These findings demonstrate that adaptation of PB2 and NP to importin α1 plays an important role in interspecies transmission. It has to be pointed out, however, that adaptive mutations different from PB2 D701N and NP N319K have been described in SC35M as well as numerous other viruses [4,6,7,21,22]. Thus, host adaptation is clearly a multifactoral process.
Importin α is the receptor for NLS-bearing cargo proteins. PB2 was first shown to contain NLSs at residues 449–495 and 736–739 [23]. The recent crystallographic analysis by Tarendeau and coworkers (2007) revealed that the latter sequence was part of a classical bipartite NLS comprising residues 738–755 which has to be unfolded by release of a salt bridge between aspartate 701 and arginine 753 to allow binding to importin α5. These authors also suggested that the adaptive mutation D701N observed in PB2 of SC35M might affect binding to importin α5 in different species. By showing that this mutation, indeed, improves importin α binding and nuclear transport of PB2 in mammalian cells we now confirm this concept. It has to be pointed out, however, that importin α5 used for the crystallographic analysis and importin α1 identified as a binding partner of PB2 in the present study differ significantly from each other in tissue specificity and cargo selectivity [24,25]. There are also large variations in amino acid sequence, but tryptophan residues 149, 191, 234 and 360 in importin α5, shown to be directly involved in the binding of the bipartite NLS of PB2 [18], are highly conserved and present in importins α5 as well as α1 (Figure S4) indicating that the crystallographic data obtained with importin α5 are also relevant for importin α1. However, additional structural and functional studies are clearly needed to assess the role of individual members of the importin α group in host adaptation. More detailed structural analysis of PB2 binding to importin α1 may also help to explain why adaptive PB2 mutation S714R did not affect in our study importin α1 binding and nuclear transport, even though it has also been implicated in PB2 binding to importin α5 [18].
Considerably less is known about the mechanism by which mutation N319K improves importin α1 binding and nuclear transport of NP in mammalian cells. Several NLSs have been identified on NP. An unconventional NLS located between amino acid residues 1–13 was shown to be involved in NP binding to importin α1 and importin α2 [17]. The second NLS is a bipartite signal located in the middle of NP between residues 198–216 which contributes less to the nuclear localization of NP than the unconventional NLS at the amino-terminus [26]. Furthermore, a nuclear accumulation site has been attributed to residues 336–345 in microinjection studies employing Xenopus oocytes and cells of rodent and primate origin [27,28]. None of these sites includes amino acid 319. A mutation in this position may therefore modulate one of the NLSs identified on NP by an allosteric mechanism.
Interestingly, we did not see differences in importin α1 binding and nuclear localization of PB2 and NP when we compared SC35 and SC35M in avian cells. Thus, while up-regulating nuclear transport in mammalian cells, the adaptive mutations did not interfere with it in avian cells albeit 82% homology between the avian and the human importin α1 (data not shown). This observation supports the concept that the virus when crossing the species barrier, goes through a phase that allows gradual acquisition of adaptive mutations without loosing fitness for the old host [5].
NLSs have also been identified in PB1 [29] and PA [30], but there is no apparent link between these signals and adaptation mutations L13P and S678N in PB1 and K615N in PA. Furthermore, neither importin binding nor nuclear localization of PB1 and PA have been analyzed in the present study, but the observation that PB1 and PA of SC35M did not induce nuclear accumulation of importin α1 (Figure 5) is compatible with the view that these proteins enter the nucleus via a non-classical pathway [16]. Altogether, however, it remains to be seen whether the adaptive mutations observed in PB1 and PA up-regulate polymerase activity in mammalian cells also by increased nuclear transport or by another mechanism. All cellular proteins using the classical nuclear import pathway have to compete with each other for importin α, and there is evidence that the equilibrium of these transport events which is necessary for the functional integrity of the cell is unbalanced when nuclear accumulation of importin α is excessive. This has been observed under various stress conditions, such as heat shock, UV irradiation and oxidative stress [31,32]. The cytoplasmic depletion of importin α1 as occurring in SC35M infected mammalian cells may therefore interfere with the nuclear transport of cellular proteins needed for the initiation of an antiviral status or for other vital functions and, thus, represent a new pathogenetic mechanism.
293T (human embryonic kidney cells) and A549 (human lung carcinoma) cells were grown in DMEM (Dulbecco's minimal essential medium) supplemented with 10% FCS (fetal calf serum; Gibco). CEC-32 quail embryo fibroblasts [33] were grown in RPMI-1640 (Gibco) supplemented with 5% FCS and 2% chicken serum (Sigma). Influenza A viruses were propagated in 11 day old embryonated chicken eggs. Growth curves were determined in three independent experiments by plaque titration [34]. The recombinant viruses SC35 and SC35M and their mutants have been described before. Briefly, the chicken adapted virus SC35 was originally derived from the seal isolate A/Seal/Mass/1/80 (H7N7) by 35 passages in chicken embryo cells. SC35M was obtained from SC35 by sequential passages in mouse lung [35,36].
Monolayers of 293T, A549 and CEC-32 cells were infected with recombinant virus and incubated for 30 min for absorption at 37°C. Cells were washed twice with phosphate-buffered saline (PBS) and incubated for 6 h, 12 h and 18 h at 37°C in appropriate medium containing 0.2% bovine serum albumin (MP Biomedicals). For transfection, we seeded the cells in 100-mm-dishes and used 10μg pHW2000-SC35-NP, 10 μg pHW2000-SC35M-NP, 30 μg pHW2000-SC35-PB2, or 30 μg pHW2000-SC35M-PB2 plasmid [4] using LipofectamineTM 2000 (Invitrogen) according to the manufacturers protocols. The medium was replaced with fresh growth medium at 6 h posttransfection and incubated for further 48 h for immunoprecipitation assays.
293T, A549 and CEC-32 cells were inoculated with virus at a multiplicity of infection (MOI) of 2 for single cycle replication and at an MOI of 10−4 for multicycle replication. Virus inoculum was removed after 30 min of incubation at 37°C, and cells were washed two times with PBS pH5. Cells were then incubated in the appropriate medium containing 0.2% bovine serum albumin at 37°C. At time points 0, 3, 6, 9, 12, 15, 18, and 24 h for single cycle replication and at time points 0, 24, 48, 72, and 96 h for multicycle replication, we collected supernatants and determined plaque titers on MDCK cells. The growth curves shown are the average result of two independent experiments.
60-mm-dishes of 293T and CEC-32 cells were infected with a MOI of 2. At the indicated time points, cells were washed twice with PBS and fractionated in 1ml PBS by using a Mixermill MM301 homogenizer (Retsch) at 20 Hz for 20 min. Lysates were mixed with 1% NP40 and separated into a nuclear (bottom) and a cytoplasmic (upper) fraction by centrifugation on a 20% sucrose cushion at 3000 rpm for 20 min. Nuclear fractions were sonicated for 15 min. Whole lysates and cell fractions were subjected to SDS-polyacrylamide gel electrophoresis followed by Western blot analysis as described [37]. Monoclonal antibodies specific for importin α1 (BD Biosciences) were used to detect the importin α1 distribution. As an internal standard, β-actin was determined with specific antibodies (Abcam). The results shown are representatives of three independent experiments.
293T, A549 or CEC-32 cells were grown on glass cover slips and infected at a MOI of 2 with recombinant virus. At the indicated time points after infection, cells were fixed with PBS containing 2% paraformaldehyde and permeabilized with PBS containing 0.1% Triton-X-100. After blocking fixation with 3% BSA in PBS, we incubated the cells for 1h with monoclonal antibodies specific for importin α1 (BD Biosciences), PB2 (Santa Cruz) or NP (kindly provided by M. Schwemmle, Freiburg). After incubation with primary antibodies we washed the cells three times with PBS and incubated for further 30 min with either FITC-coupled goat anti-rabbit (1/200; Jackson ImmunoResearch Laboratories), FITC-coupled goat anti-mouse (1/200; Jackson ImmunoResearch Laboratories), Rhodamine-coupled donkey anti-goat (1/200; Jackson ImmunoResearch Laboratories) or Rhodamine-coupled goat anti-mouse (1/200; Jackson ImmunoResearch Laboratories) secondary antibodies. Cells were washed three times with PBS and coverslips were mounted on glass plates. Cells were observed in an Axiovert 200M microscope equipped with an ApoTome device (Zeiss). Localization of importin α1, PB2, and NP in nucleus, nucleus/cytoplasm and cytoplasm was determined using the microscope by counting cells (n = 100) infected with recombinant viruses. The results shown represent three independent experiments.
Transfected cells were washed twice with PBS and collected by centrifugation. Cell pellets were resuspended in 1 ml PBS and sonicated for 15 min. Complexes of importin α1 and viral proteins were determined by co-immunoprecipitation using the ProFoundTM Co-Immunoprecipitation Kit (PIERCE) according to the protocols provided by the manufacturer. Briefly, the antibodies specific for PB2 (Santa Cruz) or NP (polyclonal serum raised against A/FPV/Rostock/34 (H7N7) [37]) were coupled covalently to an amine-reactive gel (PIERCE) using the provided coupling buffer (0.14M sodium chloride, 0.008M sodium phosphate, 0.002 potassium phosphate, 0.01M KCl; pH7.4). Cell lysates were added to the antibody-coupled columns and incubated with gentle end-over-end mixing for 2 h at room temperature. Columns were then washed several times with coupling buffer to remove non-specifically bound material and finally eluted using the provided elution buffer (pH2.8). Protein complexes generally eluted in the first three fractions. Eluted immunoprecipitation complexes were then separated by SDS-polyacrylamide gel electrophoresis followed by Western blot analysis. For detection of the PB2 protein in Western blot we used the monoclonal anti-PB2 antibody which was kindly provided by Juan Ortin, Madrid [38]. For the detection of the NP protein we used polyclonal serum raised against A/FPV/Rostock/34 (H7N7). The results shown represent two independent experiments. |
10.1371/journal.ppat.1007315 | The trypanocidal benzoxaborole AN7973 inhibits trypanosome mRNA processing | Kinetoplastid parasites—trypanosomes and leishmanias—infect millions of humans and cause economically devastating diseases of livestock, and the few existing drugs have serious deficiencies. Benzoxaborole-based compounds are very promising potential novel anti-trypanosomal therapies, with candidates already in human and animal clinical trials. We investigated the mechanism of action of several benzoxaboroles, including AN7973, an early candidate for veterinary trypanosomosis. In all kinetoplastids, transcription is polycistronic. Individual mRNA 5'-ends are created by trans splicing of a short leader sequence, with coupled polyadenylation of the preceding mRNA. Treatment of Trypanosoma brucei with AN7973 inhibited trans splicing within 1h, as judged by loss of the Y-structure splicing intermediate, reduced levels of mRNA, and accumulation of peri-nuclear granules. Methylation of the spliced leader precursor RNA was not affected, but more prolonged AN7973 treatment caused an increase in S-adenosyl methionine and methylated lysine. Together, the results indicate that mRNA processing is a primary target of AN7973. Polyadenylation is required for kinetoplastid trans splicing, and the EC50 for AN7973 in T. brucei was increased three-fold by over-expression of the T. brucei cleavage and polyadenylation factor CPSF3, identifying CPSF3 as a potential molecular target. Molecular modeling results suggested that inhibition of CPSF3 by AN7973 is feasible. Our results thus chemically validate mRNA processing as a viable drug target in trypanosomes. Several other benzoxaboroles showed metabolomic and splicing effects that were similar to those of AN7973, identifying splicing inhibition as a common mode of action and suggesting that it might be linked to subsequent changes in methylated metabolites. Granule formation, splicing inhibition and resistance after CPSF3 expression did not, however, always correlate and prolonged selection of trypanosomes in AN7973 resulted in only 1.5-fold resistance. It is therefore possible that the modes of action of oxaboroles that target trypanosome mRNA processing might extend beyond CPSF3 inhibition.
| Trypanosomes and leishmanias infect millions of humans and cause economically devastating diseases of livestock; the few existing drugs have serious deficiencies. Trypanosomosis of cattle caused by Trypanosoma congolense and Trypanosoma vivax, is a serious problem in Africa because cattle are used not only for food but also for traction, and new drugs are needed. A single injection of the benzoxaborole compound AN7973 cured T. congolense infection in cattle and goats. Although slightly lower effectiveness against T. vivax precludes development of AN7973 as a commercially viable treatment against cattle trypanosomosis, it could still have potential for diseases caused by other trypanosomes. We used a large range of methods to find out how AN7973 kills trypanosomes, and compared it with other benzoxaboroles. AN7973 and some of the other compounds had effects on parasite metabolism that resembled those previously seen for a benzoxaborole that is being tested for human sleeping sickness. The most rapid effect of AN7973, however, was on processing of trypanosome mRNA. As a consequence, amounts of mRNA decreased and synthesis of proteins stopped. We conclude that AN7973 and some other benzoxaboroles kill trypanosomes by stopping gene expression.
| Kinetoplastid protists cause severe human diseases affecting millions of people. Trypanosoma cruzi causes Chagas disease in South America, and various Leishmania species cause a spectrum of diseases throughout the tropics. Salivarian trypanosomes, the subject of this study, cause sleeping sickness in humans and economically important diseases in cattle, horses and camels [1–3]. Approximately 70 million people, living in sub- Saharan Africa, are estimated to be at risk of contracting human African trypanosomosis, which is caused by Trypanosoma brucei subspecies [4, 5]. As a result of sustained international activities to control the disease [4–6], less than 3000 cases were reported in 2016 (http://www.who.int/trypanosomiasis_african/en/). Trypanosomosis in cattle, caused by infection with Trypanosoma congolense, Trypanosoma vivax and, to a lesser extent, T. brucei, is in contrast a major problem, with wide-reaching effects on human well-being: cattle are used not only as a source of milk and meat, but also for traction. Elimination of cattle trypanosomosis could create economic benefits estimated at nearly 2.5 billion US$ per year [2]. Within Africa, trypanosomosis is transmitted by tsetse flies, but outside Africa, variants of T. brucei are transmitted venereally or by biting flies, and T. vivax can also be transmitted non cyclically by non-tsetse biting flies with massive economic losses affecting draught and milk animals from Argentina to the Philippines [7].
Control of cattle trypanosomosis currently relies on reducing the tsetse population by means of traps and insecticidal dips, together with treatment as required. The most popular treatment is with the diamidine diminazene aceturate (Berenil), the alternative being the DNA-intercalating molecule isometamidium [3]. Suramin is also sometimes used [3]. Development of new animal therapeutics is constrained by the need for cure after a single intramuscular injection [3].
In the last ten years, benzoxaboroles have generated considerable excitement for antimicrobial and other applications. Benzoxaboroles have a range of known effects. For example, they are able to bind cis-diols, such as those found in sugars, yielding stable spiro complexes [8]. This activity is the basis of the mode of action of the antifungal drug Tavaborole (AN2690) [9], which binds to the editing site of leucyl tRNA synthetase [10]. Other oxaboroles inhibit bacterial leucyl tRNA synthetases by the same mechanism [11–13]. Other benzoxaborole classes interact with ATP-binding pockets, but with different modes of binding. Crisaborole, approved for the treatment of atopic dermatitis [14, 15], selectively inhibits phosphodiesterase PDE4, with the oxaborale oxygen atoms coordinating the zinc and magnesium ions within the active site [16]. The aminomethylphenoxy benzoxaboroles inhibit Rho-activated protein kinases through hydrogen bond interactions with the hinge region of the protein and the aminomethyl group interacting with the magnesium/ATP-interacting aspartic acid [17].
The first trypanocidal oxaboroles to be described were the oxaborole 6-carboxamides [18, 19]. Acoziborole (trade name of SCYX-7158 or AN5568, 1) [20] is orally available and can cross the blood brain barrier [18]; it is now in phase IIb/III clinical trials for sleeping sickness (see https://www.dndi.org/diseases-projects/portfolio/). AN7973 (Fig 1), the main subject of this paper, is efficacious against T. congolense and T. brucei and was considered as a candidate for treatment of cattle trypanosomosis, but was later replaced by AN11736 (Fig 1), which can achieve single-dose cure of both T. congolense and T. vivax infection in cattle [21].
At present, not enough is known about structure-activity relationships in benzoxaboroles to enable predictions to be made about their modes of action. Jones et al. [22] obtained T. brucei lines with 4-5-fold resistance to AN2965 (Fig 1, named oxaborole-1 in their paper). They observed numerous single nucleotide polymorphisms (SNPs), genome rearrangements and amplifications in the resistant lines. Affinity purification with a benzoxaborole column yielded 14 proteins that bound specifically, but since none of these was affected in the resistant mutants, there was no clear indication as to which might be relevant to benzoxaborole action [22]. Aminomethyl phenoxyl benzoxaboroles such as AN3056 are sequentially activated to an active carboxylic acid form by serum and intracellular enzymes [23].
Various oxaboroles are being considered for treatment of apicomplexan parasites; and AN13762 (Fig 1) is in development [24, 25]. Plasmodium and Toxoplasma that were resistant to AN3661 (Fig 1) had mutations in the cleavage and polyadenylation factor CPSF3 [26, 27], which is implicated in 3' cleavage of mRNA precursors prior to polyadenylation [28]. The two zinc ions of CPSF3 are thought to interact with phosphate [28]. In silico molecular docking of AN3661 suggested that the boron atom occupies the position of the cleavage site phosphate of the mRNA substrate, with one hydroxyl group interacting with a zinc atom in the catalytic site—similar to binding of other benzoxaboroles to the bimetal centers of beta-lactamase and phosphodiesterase-4 [26, 27]. Introduction of resistance mutations—which were all in or near the active site—into susceptible parasites resulted in compound resistance. These results, combined with the loss of transcripts for three trophozolite-expressed genes in treated parasites, suggest that AN3661 inhibits mRNA polyadenylation through its interaction with CPSF3 [26, 27].
In trypanosomes, transcription of mRNAs is polycistronic. Individual mRNA 5'-ends are created co-transcriptionally by trans splicing of a 39nt leader sequence (SL) [29–31]. Trans splicing is spatially and mechanistically coupled to polyadenylation of the preceding mRNA. Polyadenylation sites are dictated by the positions of trans-splicing sites [32–35], RNAi-mediated depletion of polyadenylation factors inhibits trans splicing [36, 37], and disruption of splicing stops polyadenylation [32–35]. The spliced leader precursor RNA (SLRNA) is 139 nt long and is synthesised by RNA polymerase II [38] from approximately 200 tandemly repeated genes [39]. Unlike protein-coding genes, each SLRNA gene has its own promoter [40–42]. The SLRNA cap and the following four nucleotides are methylated [43–46]. Inhibition of SLRNA methylation using S-adenosyl-L-homocysteine or Sinefungin prevents splicing [47, 48]. This results in loss of mRNA from the cells by the normal mechanisms of mRNA turnover [49]. It has recently been shown that an accumulation of numerous metabolites, including S-adenosylmethionine, occurs in trypanosomes treated with Sinefungin, and a similar profile was identified in trypanosomes treated with acoziborole [50].
We here describe studies to discover molecular targets of several anti-trypanosomal benzoxaboroles. Our work concentrated on AN7973 (Fig 1), which is orally active and was the DNDi back-up for SCYX-7158/AN5568 for the treatment of human African trypanosomosis. We undertook a comprehensive analysis, including resistance generation and characterisation of morphology, metabolomes, macromolecular biosynthesis and molecular modeling.
AN7973 (Fig 1) was selected as a candidate veterinary drug from a range of 7-carboxamido-benzoxaboroles structurally similar to AN5568. The selection of AN7973 was based on its in vitro potency against T. congolense (S1 Table, sheet 1) and its ability to cure T. congolense-infected mice with a single 10mg/kg i.p. dose (S1 Table, sheet 3).T. congolense-infected goats were also cured when AN7973 was administered as a single bolus dose injection of 10 mg/kg, but for T. vivax-infected goats, two intramuscular injections of 10 mg/kg were required (S1 Table, sheet 3). Testing in cattle was done using T. vivax and T. congolense isolates that were resistant to maximum dosages of diminazene (7 mg/kg) and isometamidium (1 mg/kg). A single 10 mg/kg intramuscular injection of AN7973 cured 3/3 cattle of T. congolense infection, but two 10 mg/kg injections of AN7973 cured only one out of two T. vivax infections and a single injection failed to cure 3 animals (S1 Table, sheet 3). This meant that AN7973 would be inappropriate for field use [3]. The reduced efficacy of AN7973 against T. vivax might have been a consequence of weaker intrinsic potency: the ex vivo EC50 against T. vivax was 215 nM, as against an in vitro EC50 of 84 nM for T. congolense (S1 Table, sheet 1). The latter value is similar to the results for T. brucei (20–80 nM, S1 Table).
The lack of single-dose cattle efficacy at 10mg/kg i.m. against T. vivax precluded development of AN7973 as a commercially viable treatment against cattle trypanosomosis, but it could still have potential for diseases caused by other salivarian trypanosomes.
One of the most direct strategies to identify the targets of antimicrobial drugs is selection of resistant mutants. We therefore attempted to select parasites resistant to AN7973 using an over-expression library [51]. The level of resistance obtained was very modest—less than 2-fold (S2 Table, sheet 1) and in the two lines that we obtained, the over-expression plasmids did not contain an in-frame coding sequence. We therefore sequenced their genomic DNA. As previously observed for trypanosomes that were mildly resistant to AN2965 [22] we found numerous single nucleotide polymorphisms, amplifications and deletions (summarized in S2 Table, details in S3 and S4 Tables). The large number of genes affected made it impossible to pinpoint any particular pathway as being relevant to resistance.
Examination of parasite morphology can reveal defects in DNA synthesis, cell cycle regulation, cell motility and protein trafficking, each of which can ultimately cause parasite death. For example, Jones et al. saw accumulation of parasites in the G2 phase of the cell cycle after AN2965 treatment [22]. Treatment of trypanosomes with AN7973 with 10x EC50 for 7h caused growth arrest (Fig 2A), but no obvious changes in parasite morphology or motility and no significant changes in the proportions of cells in different stages of the cell cycle (Fig 2B). These results suggest that AN7973 does not interfere directly with DNA synthesis, cell motility, or protein trafficking.
Amino-acyl tRNA synthetase inhibition is a known mode of action of several benzoxaboroles, and inhibition of tRNA synthetase should result in cessation of protein synthesis. We therefore measured this in AN7973-treated trypanosomes by pulse labelling with [35S]-methionine followed by denaturing polyacrylamide gel electrophoresis and autoradiography. Inhibition of protein synthesis (Fig 2C) was clear, but some prominent protein bands were affected more than others (compare bands labeled a and b). The kinetics of the inhibition, combined with apparent selectivity for particular proteins, suggested that protein synthesis was not a primary target of AN7973, but might be inhibited in a secondary fashion.
Loss of protein synthesis could be caused by loss of mRNA. For example, if mRNA synthesis were inhibited, the pattern in Fig 2C would be explained if the mRNA encoding protein "b" were less stable than that encoding protein "a". Loss of mRNA could be caused by inhibition of either RNA transcription or processing. To investigate this possible mechanism, we incubated cells with AN7973, prepared RNA, and hybridised Northern blots with a probe that detects the spliced leader SL. This probe detects all processed mRNAs, as well as the ~139nt precursor, called SLRNA, that donates the SL. Incubation with AN7973 for 9h had little or no effect on the total amount of rRNA, as judged by methylene blue staining (Fig 3A) but caused progressive loss of spliced mRNA (Fig 3B). In contrast, the levels of the SLRNA remained roughly constant (Fig 3B). We therefore suspected that AN7973 was inhibiting mRNA processing. To test this we re-hybridised the blot with a probe that detects beta-tubulin. The tubulin genes are arranged in alternating alpha-beta tandem repeats, and splicing inhibition through heat shock [52] or Sinefungin treatment [47] leads to accumulation of partially processed beta-alpha dimers and multimers. Partially processed tubulin mRNAs were indeed detected within an hour of AN7973 application (Fig 3C). This result showed that after AN7973 addition, transcription of protein-coding genes continued but mRNA processing was impaired. At later time points, the tubulin mRNAs disappeared, suggesting that the failure in processing was complete. (Most measurements suggest that the tubulin mRNAs have half-lives of about half an hour [53].)
As noted above, after 1h AN7973 treatment, the level of SLRNA was not much changed. This is somewhat counter-intuitive, since one might expect processing inhibition to lead to a build-up of SLRNA. However, RNAi that targets splicing or polyadenylation factors rarely causes SLRNA increases of more than 2-fold [37, 54, 55]. The reason is that SLRNA synthesis is balanced by degradation. Thus when Actinomycin D (Act D) is used to inhibit transcription, SLRNA disappears within 30 min even though no new precursors are available for splicing [56]. Fig 3D compares the effects of AN7973 with those of Act D treatment. As expected, 30 min Act D resulted in SLRNA loss whereas AN7973 had no effect. Moreover, if Act D was added after a one-hour AN7973 treatment, SLRNA again disappeared within 30 min—whereas SLRNA was still visible after 9h in the presence of AN7973. We concluded that treatment of the cells with AN7973 for only one hour inhibited mRNA processing but did not prevent synthesis of SLRNA.
Inhibition of procyclic trypanosome mRNA processing by RNAi targeting polyadenylation factors causes growth arrest within 3 days, followed by cell death [37, 54]. These delays in observing effects are explained by the fact that the targeted proteins remain detectable (see e.g. [55]). Processing inhibition can therefore easily explain killing of trypanosomes by AN7973.
The only small molecule currently known to inhibit kinetoplastid mRNA processing in vivo is the S-adenosyl methionine analogue Sinefungin, which is a general methylation inhibitor [57]. Chemical inhibition of spliced leader methylation prevents trans splicing in a permeabilised cell system [48], and Sinefungin does the same in vivo [47]. Since we already knew that AN5568 caused increases in methylated intermediates [50], we wondered whether splicing inhibition by AN7973 was also caused by inhibition of spliced leader methylation. We used primer extension to assess the amount and methylation status of SLRNA. At the same time, we assayed the level of the 2'-5' branched "Y-structure" splicing product (Fig 4A). cDNA synthesis was primed by a 5'-labelled oligonucleotide that is complementary to a region towards the 3' end of the SLRNA (black arrows in Fig 4B). The products from full-length SLRNA form a small ladder, because 5' methylation partially blocks reverse transcriptase, while the Y-structure trans splicing intermediate gives a product of 87nt (Fig 4B). These products are seen in Fig 4C, lane 1. As expected, incubation with the methylation inhibitor Sinefungin for 30 min abolished the multiple bands caused by cap methylation, with cDNA synthesis extending cleanly to the 5' end of the SLRNA (Fig 4C, lane 4, arrowhead); at the same time, the signal from the Y structure was decreased. The concentration of Sinefungin used, 2 μg/mL, is 4000 times the EC50. This is the standard concentration in these types of experiments and was chosen in order to prevent cap methylation within about 10 min [58].
Importantly, AN7973 (10x EC50) had no effect on the pattern of bands from the SLRNA, showing that cap methylation was not affected (Fig 4C). In contrast, Y-structure formation was clearly decreased (Fig 4C, lanes 2 and 3). The combined results so far thus showed that AN7973 acts specifically on splicing, not on production of mature splicing-competent SLRNA. Quantitation of four independent experiments showed that after 30 min, the effect of AN7973 was already significant (p<0.0005, S3A Fig). Remarkably, considering the differences in the doses used, a 2-h incubation with AN7973 inhibited Y-structure formation to the same extent as a 30-min incubation with Sinefungin (Fig 4D).
Although the effect of AN7973 on splicing was quite rapid, we nevertheless had to consider the possibility that it was secondary to inhibition of some other process. Apart from Sinefungin, the only treatment that was previously shown to cause accumulation of tubulin precursors within 30 min was severe heat shock [52, 59], but the mechanism is unknown and there are numerous other effects including general translation arrest [59] and transcription inhibition [60]. To find out whether trans splicing inhibition was a general feature of parasites stressed through addition of trypanocidal or trypanostatic drugs, we measured Y structure abundance in cells treated with DFMO, diminazene aceturate, pentamidine or suramin, all at 10x EC50. No significant inhibition of Y-structure formation was observed (Fig 4E–4H, S3 Fig). This result demonstrated conclusively that splicing inhibition is not a general side-effect of growth inhibition. The effect of AN7973 on splicing therefore indicates a specific mechanism.
Another possible cause of splicing inhibition might be severe metabolic disruption. The fact that the parasites retained normal motility for several hours indicated that there were no immediate effects upon ATP generation, and we already knew that both mRNA transcription and SLRNA modification were unaffected. Analysis of the metabolome after 5h AN7973 treatment revealed, however, that the main effects were—as with AN5568—related to methylation, with substantial increases in S-adenosylmethionine (SAM) methyl thioadenosine (MTA), methyl lysine (ML), dimethyl lysine (2ML), trimethyl lysine (3ML) and acetyl lysine (AL) (Fig 5; S5 Table). SAM is a methyl donor and it has been shown that the methyl groups on the methylated lysines are derived from methionine [50]. It therefore seemed worthwhile to find out whether other anti-trypanosomal benzoxaboroles had similar effects.
We started with a panel of 30 compounds. These were tested in vitro against T. brucei and T. congolense (S2 Fig, S1 Table) and metabolic effects were also assessed (S6 Table). We then selected various compounds based on potency, the time taken to see an effect (S3 Fig), and the effects on methylated amino acids, SAM and MTA (S6 Table). Nearly all of the chosen compounds had EC50s in the low nM range (Fig 6). The notable exception was AN3661 (Fig 1), the anti-malarial lead compound that targets CPSF3 [26, 27]: its EC50 for bloodstream-form trypanosomes was at least 20 times higher than that of AN7973. At a concentration of 5x the EC50 determined at 48h, most of the compounds acted within 6-8h (S3 Fig, S1 Table). The veterinary drug candidate AN11736 kills the parasites more slowly because it requires processing for full activity (Giordani et al., manuscript in preparation). Results for treatment of T. brucei with the selected compounds are shown in Fig 5. There were no discernable structure-activity relationships for EC50, time to kill, or the SAM/MTA effect, but most of the oxaboroles that induced a SAM/MTA pattern were faster killing.
To investigate the possible link between splicing inhibition and the methylated metabolites, we measured the effects of the chosen compounds on splicing. We also included AN2965 (Fig 1), the compound whose mode of action had been investigated previously [22], and the antimalarial candidate AN13762 (Fig 1, compound 46 in [25]). Initial triplicate measurements (Fig 7A, experiment 1) gave reproducible results for most of the compounds; repeat assays for three that had given ambiguous results confirmed that they gave substantially less splicing inhibition than AN7973 (Fig 7, experiment 2 and S4A Fig). The clear processing inhibitors included the antimalarial candidate AN13762 (compound 46 in [25]); the EC50 of this compound for trypanosomes in the 3-day assay was 4 times higher than that for P. falciparum. Overall, there was a remarkably good correlation between splicing inhibition and the average increase in SAM and MTA (Fig 7B); only two compounds did not conform to the overall pattern. A caveat is that AN11736 is slow acting, and was used at low doses given its extreme potency, so it is possible that splicing inhibition by AN11736 might be seen upon more prolonged incubation.
Why could splicing inhibition and S-adenosyl methionine-related metabolites be linked? We could think of two explanations. First, transcription and processing of the SLRNA genes absorbs substantial cellular resources, since the cell needs to make at least 10,000 SLRNAs per hour [53]. It was possible that after 6-8h (the time of most metabolome measurements), there might have been some feedback inhibition of transcription that led to a decreased methylation requirement. Second, loss of mRNA production leads to loss of unstable mRNAs. If the protein encoded by an unstable mRNA has a high turnover rate, then that protein will disappear; and if that protein is a metabolic enzyme, its substrates will accumulate. This too could have led to metabolic changes as we noted.
To address the second hypothesis, we directly measured the effect of mRNA synthesis inhibition on the metabolome, using Actinomycin D (10 μg/mL) for 6h. Interestingly, there were again large increases in the amounts of methylated amino acids and significant (but smaller) increases in SAM and MTA as well (Fig 5). Comparison with the well-characterised effects of AN5568 revealed a clear correlation (Fig 7C, S7 Table). These results suggest that the increases in methylated and acetylated lysine—and perhaps also in SAM and MTA—that were seen after treatment with benzoxaboroles could indeed be an indirect consequence of loss of mRNA.
As an alternative way to find out whether the effect of AN7973 on mRNA processing was direct or indirect, we measured trans splicing in permeabilised procyclic-form trypanosomes [48]. (Equivalent assays are not established for bloodstream forms.) As for most benzoxaboroles that have been tested (Fig 6, S1 Table) the EC50 of AN7973 for procyclic forms was 5–10 times higher than that for bloodstream forms. Treatment with 10x EC50 clearly inhibited splicing in procyclic forms, with 70% loss of Y structure after 2h (S4B and S4C Fig).
To test splicing in vitro, procyclic trypanosomes were permeablised with lysolecithin, pre-incubated with AN7973 or DMSO, then transcription was allowed to proceed for 10 min in the presence of [alpha-32P]-UTP [61]. RNA was separated on 7% polyacrylamide-urea gels and visualized by autoradiography. Under these conditions, the SL intron is visible as a ~100nt species, which disappears if incubation is continued for a further 20 min [61]; it is also not made if cap methylation is inhibited by S-adenosyl homocysteine [48].
We do not know the intracellular concentration of AN7973, and in the in vitro transcription reaction the density of permeabilised parasites is 1000 times higher than in the experiments with cultures. For the in vitro assays we therefore chose to use a concentration of 100 μM (500x the EC50), which gives a AN7973:parasite ratio that is equivalent to the ratio at 5x EC50. This treatment reproducibly prevented formation of the Y structure without preventing transcription of SLRNA or smaller RNAs (Fig 8). An additional band (arrowhead), which appeared in the presence of AN7973 and was slightly shorter than SLRNA, might be a 3' degradation product that was previously described [48]. AN7973 also reproducibly inhibited labeling of RNAs longer than 500nt (indicated by a question mark). These RNAs are thought to include mRNAs and rRNA [62], and the inhibition could either be a consequence of trans splicing inhibition or another effect of AN7973. In future it would be interesting to repeat these studies at a variety of concentrations and with other benzoxaboroles.
When trypanosomes are stressed by heat shock or starvation, their mRNAs accumulate in RNA-protein particles called stress granules containing a helicase, DHH1 [63]. Generally, these are throughout the cytoplasm, but after treatments that inhibit splicing, they transiently cluster around the nuclear envelope [64]. To find out whether AN7973 had the same effect we followed localization of YFP-tagged DHH1 in treated cells (Fig 9; the YFP signal is coloured magenta). A few granules were seen in untreated cells (Fig 9A) but after 30 min Sinefungin treatment, the nuclear periphery granule pattern was clear (Fig 9B). AN7973 also induced formation of perinuclear granules in a subset of cells (Fig 9C). However, when the remaining compounds were tested (S5 Fig, S6 Fig, S7 Fig) we observed no correlation between peri-nuclear granule formation and Y structure inhibition (Fig 9D). This may partly be explained by the transient nature of the perinuclear localization, but the clear perinuclear granule formation after AN11736 treatment also suggests that the pattern can be caused by stresses other than splicing inhibition. We concluded that without careful time-course studies, this assay could not be used to distinguish between specific splicing effects and more general stress responses.
To try to identify possible targets implicated in mRNA processing, we re-examined the genomes of our partially-resistant lines (S2–S4 Tables). Since loss of polyadenylation stops splicing [36, 37], we looked for changes in genes associated with both processes. No mutations of the U snRNAs were found. A missense mutation was found in the Tb927.10.9660 open reading frame encoding a putative CRN/SYF3; this protein co-purified with the PRP19 complex, but did not co-sediment with the complex on a sucrose gradient [65]; its function is thus unclear. The Sm complex forms the core of spliceosomal snRNPs, and our two cell lines that had been selected in AN7973 both had 1–2 extra copies of the genes encoding four out of the seven components: Sm-B, Sm-E, Sm-F and Sm-G. The other three genes were, however, not amplified. After selection of proteins on an oxaborole affinity column, Jones et al. found enrichment of RBSR1 (Tb927.9.6870), a protein with an SR-domain that is potentially involved in splicing, and of the U2 splicing auxiliary factor (U2AF35, Tb927.10.3200) [22]. Our resistant lines had no changes in RBSR1 or U2AF35.
In P. falciparum and T. gondii, mutations in CPSF3 gave resistance to AN3661 [26, 27]. Moreover, in one of their AN2965 resistant lines, Jones et al. [22] observed a two-fold amplification of the gene encoding CPSF3 (Tb927.4.1340; also designated CPSF73); and AN2965 was a strong inhibitor of splicing (Fig 7A). We therefore compared the T. brucei sequence with those of Plasmodium and humans, concentrating on the residues that were mutated in AN3661-resistant Apicomplexa. H36 and Y408 of the P. falciparum sequence were mutated in AN3661-resistant lines but are retained as H and Y the human and trypanosome sequences (S8 Fig). The remaining important Plasmodium residues, however, are already different in trypanosomes. Y252 (mutated to C for AN3661 resistance) is N in T. brucei; T406 (mutated to I) is A; T409 (mutated to A) is C; and D470 (mutated to N) is already N in T. brucei (S8 Fig). From these changes alone one could predict that trypanosomes would be quite resistant to AN3661—as is indeed the case (S1 Fig). Although AN3661 did inhibit trypanosome mRNA processing and give a SAM/MTA effect, the concentrations used were 10–20 times higher than for AN7973 (S1 Table).
To test the role of CPSF3 we first attempted to make the equivalent of the Y408S mutation in trypanosomes by homologous gene replacement. Interestingly, although the transfections yielded several transgenic clones, none had the mutation. This, together with the fact that neither we nor Jones et al. [22] found the mutation after resistance selection, suggests that in the context of the trypanosome sequence, the Y408S equivalent (Y383S in the T. brucei sequence, S8 Fig) results in an unacceptable decrease in CPSF3 activity. If so, the result suggests either that the mutant CPSF3 has dominant-negative effects, or that CPSF3 is present in limiting amounts such that mutation of one gene copy results in haplo-insufficiency.
We next assessed how the differences between T. brucei and P. falciparum CPSF3 sequences would affect protein conformational dynamics. To do this we investigated a homology model of CPSF3 by elastic network normal mode analysis (S9A and S9B Fig), and found a relative rotational breathing motion of the domains (S9C Fig). Interestingly, in the model, many residues associated with AN3661 resistance, such as Y408 and N252 (P. falciparum numbering) were found to line the cleft between the breathing domains. Relative domain rotation might thus be able to enhance the accessibility of the binding site. Thus, mutations associated with AN3661 resistance might affect the conformational dynamics of the enzyme, the accessibility of the active site and the stability of the interdomain contacts.
As an alternative to mutation, we inducibly over-expressed RBSR1, U2AF35 and CPSF3 in bloodstream forms, as C-terminally myc-tagged versions. Expression of RBSR1-myc and U2AF35-myc (S10A Fig) did not affect the EC50 of AN7973. In contrast, expression of CPSF3-myc caused at least 3-fold increases in the EC50s of four tested benzoxaboroles, including AN7973 (Fig 10, S10B Fig, S8 Table). Statistically significant increases of 2-fold or more were also seen for several other benzoxaboroles (Fig 10, S10B Fig, S8 Table) and there was a moderate correlation with Y-structure inhibition (Fig 10B). This suggests that the effective intracellular concentration of the benzoxaboroles is reduced through binding to excess CPSF3. The modest level of resistance may be explained by the fact that CPSF3 normally functions as part of a complex: we do not know the extent to which CPSF3 can be accumulated independently, and its conformation might be influenced by protein-protein interactions.
Over-expression of CPSF3-myc had almost no effect on the EC50 of AN3661. Because of its low potency, AN3661 was always tested at relatively high (micromolar) concentrations, so this result is difficult to interpret. However poor binding of AN3661 to CPSF3 is expected, since the trypanosome protein has residues equivalent to those found in the AN3661 resistant version of the Plasmodium falciparum protein (S8 Fig).
To investigate whether AN7973 could interact with CPSF3, we performed induced fit docking to our comparative model of T. brucei CPSF3 (TbCPSF3). For AN3661, it was proposed that boron occupies the position of an mRNA substrate phosphate group and that the ring oxygen, as well as hydroxylations, of the benzoxaborole can interact with the two active site zinc ions [27]. We applied constraints accordingly when docking AN7973 in the TbCPSF3 active site. In our model, in addition to metal-oxygen interactions and hydrogen bonding contacts with D67, which stabilize the benzoxaborole binding mode, the pyrazole ring of AN7973 forms a hydrogen bond with the backbone amide nitrogen of A406 (S11 Fig). (All residue numbers mentioned correspond to those of the P. falciparum sequence, PfCPSF3, S8 Fig) The elongated tail of AN7973 lodges in a subpocket that is largely lined by hydrophobic residues such as F267, P284, F286, L294, A406 and F443. Notably, the same interaction pattern was observed for both the tetrahedral form of the compound with a formal negative charge on the boron atom and the trigonal planar neutral variant (S11A and S11B Fig, respectively). The benzoxaborole core of the planar variant however has a rotated orientation compared to the tetrahedral form, which leads to a displacement of the loop containing Y408. The results of the docking therefore suggested that inhibition of TbCPSF3 by AN7973 is feasible.
Benzoxaboroles are important drug candidates for both human and ruminant African trypanosomosis. Our results show that AN7973 inhibits mRNA processing in trypanosomes. Expression of additional CPSF3 increased the EC50 of AN7973, suggesting that AN7973 can bind to CPSF3. These results suggest that mRNA processing is an important target of AN7973, which might operate through CPSF3 inhibition. AN7973 also caused metabolite changes indicative of disturbed methylation, similar to those observed for acoziborole.
When this work was started, AN7973 was under consideration as a candidate for treatment of cattle trypanosomosis, but it was later found to be less effective against T. vivax. In the available T. vivax genome sequence there is a gap in the CPSF3 open reading frame. This gap spans N231 in the T. brucei sequence, which is Y in Apicomplexa and humans and mutated to C in resistant Apicomplexa (S8 Fig). All other CPSF3 residues that are implicated in AN3661 resistance in Apicomplexa are conserved between T. vivax, T. congolense and T. brucei (S8 Fig). Residues predicted to be in contact with AN7973 in our structural model of TbCPSF3 were also largely conserved in the T. vivax sequence. Only in the loop 378–385 (or 403–410 in the P. falciparum numbering) are two valine residues replaced by isoleucine, which might slightly restrain the space accessible to the compound in the active site (S8 Fig).
Our results suggested a mechanistic link between splicing inhibition and accumulation of specific methylated metabolites (Fig 7B). One hypothesis we had was that the effects on methylation intermediates could be caused by decreased requirements for cap methylation. However we found no evidence, either in vivo or in vitro, that SLRNA transcription, capping or methylation were affected by AN7973. Another hypothesis was that loss of mRNA production will lead to a selective reduction in the activities of enzymes that have relatively high mRNA and protein turnover rates, and that these enzymes are required for various aspects of methylation. This was partially supported by accumulation of methylated amino acids after Actinomycin D treatment.
Strangely, our results revealed no clear structure-function relationships for benzoxaborole effects on trypanosome viability, metabolites, or splicing. The diversity in structures of compounds that inhibited trypanosome mRNA processing, combined with molecular modelling, suggests to us that this might be an intrinsic property of the pharmacophore. Nevertheless, some trends with respect to molecules with the same scaffolds can be noted. For example the two compounds AN15368 and AN11736 are both L-valinate amide benzoxaboroles [21] and both showed very little Y structure inhibition (Fig 1, Fig 7). These two compounds also showed little change in EC50 with overexpression of CPSF3 (Fig 10) suggesting that the L-valinate amide benzoxaborole series may not act through inhibition of mRNA processing. AN7973 and AN4169, both of the carboxamide scaffold, also showed similar decreases in Y structure. An important caveat is that the uptakes and metabolisms of the various compounds are likely to differ. For example, after 2h incubation, AN3056 and the veterinary drug candidate AN11736 had less effect on processing than did AN7973: but their action might be delayed, since both are subject to activation within the parasites.
So far, selection of trypanosomes resistant to benzoxaboroles has met very limited success. Substantial resistance was obtained only for compounds which require intracellular metabolic activation, and the mutations responsible were in the activating enzymes. For benzoxaboroles that are probably not metabolised, such as AN7973 and AN2965, only very limited resistance could be obtained. A priori, one would expect that if these compounds have a single active site target, selection for resistance should be relatively straightforward. It is however possible that the necessary mutations are incompatible with function.
Inhibition of mRNA processing was the earliest effect that was seen after AN7973 treatment and is therefore likely to make a vital contribution to parasite killing. In addition, both docking studies and results from over-expression identify CPSF3 as a likely target. This raises the possibility that similar modes of action might be seen in oxaboroles under development against other Kinetoplastids, including Leishmania and Trypanosoma cruzi.
All in vivo mouse experiments were carried out in accordance with the strict regulations set out by the Swiss Federal Veterinary Office, under the ethical approval of the Canton of Basel City, under license number #2813.
The experimental protocols used for goat studies were approved by the ethics committee for animal experimentation by the Veterinary Faculty of the University of Las Palmas in Gran Canaria, Spain on July 21, 2012 with the reference number 240/030/0121-36/2012. The studies were conducted under the strict guidelines set out by the FELASA for the correct implementation of animal care and experimentation.
Tests on cattle were done in accordance with the principles of veterinary good clinical practice (http://www.vichsec.org/guidelines/pharmaceuticals/pharma-efficacy/good-clinical-practice.html). The ethical and animal welfare approval number was 00 l-2013/CE-CIRDES.
Testing of T. congolense in vitro (72 hrs) and T. vivax ex vivo (48 hrs) assays were done as described in [66]. Testing of T. congolense and T. vivax in vivo in mice was done as described in [21].
The proof of concept efficacy studies in goats using AN7973 were conducted as previously described in [67], but established and modified for T. congolense and T. vivax models of infection. The trials took place from January to May 2013 within the Veterinary Faculty of the University of Las Palmas and the Agricultural farm (Granja Agricola) of the Canarian Island Government in Arucas, Gran Canaria, Spain. In total, 45 female Canarian goats, weighing between 12–35 kg and no less than four months old, were purchased from a local dairy farmer and transported to the study site. The goats were placed in fly-proofed pens and allowed to acclimatise for two weeks, before being randomly selected and divided into test groups of four. Goats were experimentally infected intravenously from two highly parasitaemic donor goats, with 106 and 105 parasites per goat for T. congolense and T. vivax, respectively. AN7973 was administered intramuscularly accordingly, as either two injections of 10 mg/kg or as a single bolus dose of 10 mg/kg, on days 7 and 8 post-infection. Thereafter, the parasitaemia was monitored in the goats for up to 100 days post-treatment, after which any aparasitaemic and surviving goats were considered cured. Relapsed goats were removed immediately from the trial and humanely euthanized with an intravenous injection of sodium phenobarbital.
The efficacy of AN7973 against T. congolense and T. vivax in cattle was tested as described in [68]. The studies were conducted in fly-proof facilities and included negative (saline) controls; and the staff were blinded with regard to allocation of animals to treatment groups. Assessments were made for 100 days post treatment unless animals relapsed sooner.
Bloodstream-form T. brucei brucei 427 Lister strain were cultured in HMI-9 plus 10% foetal calf serum or CMM [69] plus 20% foetal calf serum at 37°C, 5% CO2. PCF T. b. brucei were cultured in SDM79 plus 10% foetal calf serum at 28°C. Compounds were dissolved at 20 mM in DMSO and aliquoted to avoid excessive freeze-thaw cycles.
EC50s were measured in two different ways. To obtain the EC50s in S1 Fig, compounds were serially diluted over 24 doubling dilutions in 100 μL culture medium in 96 well opaque plates from a starting concentration of 100 μM. Bloodstream form trypanosomes were added at a final density of 2x104/mL (100 μL) and incubated for 48 hours, while procyclic forms were added at a final density of 2x105/mL and incubated for 72 hours. After incubation, 20 μL of 0.49 mM Resazurin sodium salt in PBS was added to each well and plates were incubated for a further 24 hours. Plates were read on a BMG FLUOstar OPTIMA microplate reader (BMG Labtech GmbH, Germany) with λexcitation = 544 nm and λemission = 590 nm.
For the results in S1 Table, sheet 1, column E, bloodstream form trypanosomes were diluted to 4000/mL in the presence of compounds (diluted in water from DMSO) and incubated for 72h. 3-4h before the end of the incubation, Resazurin (Sigma) was added (final concentration 44 μM). Resazurin fluorescence was measured to assess the number of surviving viable cells [70, 71]. Each assay was performed with 3 technical and 3 biological replicates.
For the time to kill assay (S1 Fig, S2 Fig) bloodstream-form T. b. brucei were cultured in 24 well plates in triplicate. Cultures were seeded at 5 x 105/mL and compound was added at 5xEC50. Cells in each well were counted at 2, 4, 6, 8 and 24 hours using a haemocytometer.
For the all assays except the metabolomes shown in Fig 1 and S6 Table, compounds were used at 10x the 72-h EC50. To allow for variations between drug aliquots, EC50s were measured prior to every experimental series. The concentrations of compounds used in different experiments are listed in S1 Table and S6 Table.
For protein analysis, 2-3x106 cells were collected for each sample, resuspended in Laemmli buffer heated and subjected to SDS-PAGE gel electrophoresis. All assays of macromolecular biosynthesis and RNA processing were done at densities of less than 2 x 106/mL. Pulse-labeling was done as described in [72].
Total RNA was extracted from roughly 5x107 cells using peqGold TriFast (peqLab) following the manufacturer's instructions. The RNA was separated on formaldehyde gels and then blotted on Nytran membranes (GE Healthcare). Following crosslinking and methylene blue staining (SERVA), the northern blots were hybridized with the appropriate probes. For mRNA detection, the membranes were incubated with [α-32P]dCTP radioactively labelled DNA probes (Prime-IT RmT Random Primer Labelling Kit, Stratagene) overnight at 65°C. For spliced leader detection, a 39mer oligonucleotide complementary to the spliced leader was labelled with [γ-32P]ATP using T4 polynucleotide kinase (NEB) and incubated with the membrane overnight at 42°C. After washing the blot, it was exposed to autoradiography films and detection was performed with FLA-7000 (GE Healthcare). The images were processed with ImageJ.
Compound treatments were all done at cell densities of about 0.9x106 cells/mL. For each condition, 8-10x107 cells were used. Primer extension was done approximately as described in [47]; primers were: ACCCCACCTTCCAGATTC for SLRNA (KW01 or CZ6364) and TGGTTATTTCTCATTTAAGAGG (CZ6491) for U3 snRNA. Both primers and the ladder were radioactively 5'-end-labelled with [γ-32P]ATP. For extension, 10 μg of RNA was incubated for 5’ at 65° with 2 μL of dNTPs (10 mM) and roughly 200 000 counts per minute (cpm) of the corresponding primer. Afterwards, RNasin (Promega), SuperScript III Reverse Transcriptase (Thermo Fischer), DTT and buffer were added according to the manufacturers instructions. The mixture was incubated 60’ at 50°C and then inactivated 15’ at 70°C. The samples were run in 35 cm long 6% polyacrylamide gels, dried, and analysed by phosphorimaging. The images were analysed using Fuji / ImageJ.
In vitro transcription in permeabilised cells was done following the published procedure [48, 61] with minor modifications [73]. Briefly, cells (2.5 x 108/reaction) were permeabilised with lysolecithin for 1 min on ice, washed, then resuspended in 60 μL transcription buffer. 1 μL of either AN7973 or DMSO alone were added, and the reaction pre-incubated at 28°C for 2 min. After addition of 100 μL transcription cocktail containing 1 μL of either AN7973 or DMSO, the reaction was allowed to proceed for 10 min at 28°C. The permeabilised cells were pelleted (45 sec) and resuspended in 1 mL of TriFast. After the first phase separation, the aqueous fraction was re-extracted with phenol to remove residual protein; this was necessary to obtain good separation during gel electrophoresis. The final RNA volume was 20 μL. 10 μL of each reaction were separated on a 7% polyadcrylamide/urea gel and the products were detected by phosphorimaging with low molecular weight DNA markers (New England Biolabs).
For genomic DNA sequencing, libraries were prepared at the Cell Networks Deep Sequencing Core Facility (University of Heidelberg) and subjected to paired-end MiSeq (Illumina) at EMBL. The quality of sequencing was evaluated with FastQC and the reads were trimmed using Trimmomatic. The output was aligned to the T. b. brucei TREU927 genome (version 9.0) using bowtie2. Results are available in ArrayExpress under accession number E-MTAB-6307. The Picard option AddOrReplaceReadGroups was used to create a valid .bam file to then be piped into GATK for obtaining, as output, .vcf files containing SNP and indel information. SnpSift filtered the features of interest, excluding for example synonymous mutations and intergenic regions. Identified variations from all cell lines were pooled to look for mutations found in all strains compared to the wild type, and in addition, reads from each strain were processed separately to find all the mutated genes. The lists were then compared. The genes taken into consideration were identified with annotation and categories from the Clayton lab in-house annotation list. Based on these annotations, the genes were filtered in Excel, where highly repetitive genes were excluded from the analysis. At first, variant surface glycoproteins (VSG), expression site-associated genes (ESAG), receptor-type adenylate cyclase GRESAG, UDP-Gal or UDP-GlcNAc-dependent glycosyltransferases and pseudogenes were removed. In the case of non-homozygous mutations, the genes were further selected filtering out other repetitive genes such as leucine-rich repeat proteins (LRRP), various invariant surface glycoproteins (ISG), nucleoside transporters (TbNT) and retrotransposon hot spot (RHS) proteins. The list of gene IDs was compared with their translation level based on ribosome profiles [74, 75] and genes with values below 10 (non-translated) were excluded from the analysis.
For all assays at 5x EC50, bloodstream-form T. b. brucei were inoculated into medium at 1 x 106/mL (for a six or eight hour incubation) or 2 x 105/mL (for a 24 hour incubation) and compounds were added. Cells were incubated with compounds under normal growth conditions. At the desired time point, 1 x 108 cells were taken and cooled rapidly in a dry-ice ethanol bath to 4°C. Samples were centrifuged at 1250 g twice to remove all medium before 200 μL chloroform:methanol:water (1:2:1) were added. Extracts were shaken for one hour at max speed before cell debris was removed by centrifugation at 16,000 g. Metabolite extracts were stored at -80°C under argon gas.
Mass spectrometry–Metabolite samples were defrosted and run on a pHILIC column coupled to an Orbitrap mass spectrometer as previously described [76]. In batch 1 an Orbitrap Exactive (Thermo Scientific) was used with settings including mass range: 70–1400, a lock mass of 74.0964, capillary: 40V, Tube lens: 70V, Skimmer: 20V, Gate lens: 6.75V and C-trap RF: 700V. In batch 2 an Orbitrap QExactive (Thermo Scientific) was used with settings including mass range: 70–1050, lock masses of 74.0964 and 88.0757, S-lens: 25V, Skimmer: 15V, Gate lens: 5.88V and C-trap RF: 700V. For fragmentation analysis an MS2 isolation window of 4 m/z, an intensity threshold of 3.3e5 and dynamic exclusion of 10 seconds were used.
Metabolites were putatively annotated using IDEOM software [77] before verification of annotations using mass, retention time, isotope distribution and fragmentation pattern. Xcalibur (Thermo) was used to explore the raw data, MzCloud (mzcloud.org) was used to match fragments to database spectra. Metabolite analysis was done using four biological replicates per condition and cell line. Relative metabolite levels were based on raw peak height relative to the average raw peak height of untreated cells. Lists of metabolites were mapped on metabolic pathways using Pathos (http://motif.gla.ac.uk/Pathos/). Intersections between samples were found using bioVenn (http://www.cmbi.ru.nl/cdd/biovenn/). (http://www.cmbi.ru.nl/cdd/biovenn/). Metabolite identities are consistent with standards from the Metabolomics Standards Initiative and evidence for each identity is shown in S6 Table.
For over-expression of C-terminally myc-tagged proteins, the open reading frames encoding proteins of interest were amplified by PCR from genomic DNA, and cloned into pRPa-6xmyc [78]. After transfection, cells were selected and expression of myc-tagged protein was induced overnight with 100 ng/mL tetracycline.
The plasmid for creation of cells expressing YFP-DHH1 [59] was a kind gift from Susanne Krämer (University of Würzburg). It was transfected into bloodstream-form trypanosomes and two stable cell lines expressing the protein were selected. Trypanosomes (maximum density less than 1 million/mL in 10 mL) were treated for 30 min with 2 μg/mL Sinefungin, or for 1-2h with compounds at 10x EC50. After collection (5 min 1000g) and washing in PBS cells were resuspended in 20 μL PBS. 500 μL of 4% paraformaldehyde solution in PBS was added, cells were incubated without shaking for 18 min, washed 3x with PBS, then distributed onto poly-lysine coated chamber glass slides (all cells divided in two chambers) and left at 4°C over night. The PBS was then removed and cells were permeabilised using 0.2% Triton X-100 (w/v) in PBS, with shaking at room temperature for 20 min. After a further 3 washes, slides were incubate at room temperature with 200 ng/mL DAPI in PBS (15 min shaking), washed twice more, air-dried, embedded and covered for microscopy.
Slides were viewed and the images captured with Olympus IX81 microscope. The 100x oil objective was used. Digital imaging was done with ORCA-R2 digital CCD camera C10600 (Hamamatsu) and using the xcellence rt software. Bright field images were taken using differential interference contrast (DIC), Exposure time 30 ms, Lamp 4.0. Fluorescent images were made using DAPI and YFP filters. They were taken as Z-stacks, 30–40 images in a 8 μm thick layer, Exposure time 40 ms, Light intensity 100%, and afterwards they were deconvoluted (Numerical aperture 1.45, Wiener filter, Sub-Volume Overlap 20, Spherical Aberration Detection Accurate).
The fluorescence images were processed using ImageJ. For YFP and DAPI channels, layers containing signals ere selected, then the projection of maximum intensity of the deconvoluted stack was used. The images were saved in an 8-bit range then overlayed with DAPI in cyan and YFP as magenta. The colour balance was then adjusted with variable maxima for DAPI/cyan, but a set maximum of 80 for YFP/magenta.
After initial assessment after AN7973 and Sinefungin treatment, three independent experimental series, each including a negative control, were processed, and the images were read blinded. Cells were classified as having nuclear periphery granules if there were at least four strong granules on top of, or within one granule diameter of, the nucleus. Cells were classified as "strong" if nearly all granules were around the nucleus, and "possible" if there were several granules in other positions as well. For granule counts, only structures with at least four adjacent pixels at maximum intensity were considered.
|
10.1371/journal.pmed.1002072 | Glycemic Control and the Risk of Tuberculosis: A Cohort Study | Diabetes is a well-known risk factor for tuberculosis (TB) and is increasingly prevalent in low- and middle-income countries, where the burden of TB is high. Glycemic control has the potential to modify the risk of TB. However, there are few studies on the association between glycemic control and TB risk, and the results are inconsistent.
We assembled a cohort using 123,546 individuals who participated in a community-based health screening service in northern Taiwan from 5 March 2005 to 27 July 2008. Glycemic control was measured using fasting plasma glucose (FPG) at the time of screening. The cohort was followed up to 31 December 2012 for the occurrence of TB by cross-matching the screening database to the national health insurance database. Multiple imputation was used to handle missing information. During a median follow-up of 4.6 y, 327 cases of TB occurred. In the multivariable Cox regression model, diabetic patients with poor glycemic control (FPG > 130 mg/dl) had a significantly higher hazard of TB (adjusted hazard ratio [aHR] 2.21, 95% CI 1.63–2.99, p < 0.001) compared to those without diabetes. The hazard of TB in diabetic patients with good glycemic control (FPG ≤ 130 mg/dl) did not differ significantly from that in nondiabetic individuals (aHR 0.69, 95% CI 0.35–1.36, p = 0.281). In the linear dose-response analysis, the hazard of TB increased with FPG (aHR 1.06 per 10-mg/dl increase in FPG, 95% CI 1.03–1.08, p < 0.001). Assuming the observed association between glycemic control and TB was causal, an estimated 7.5% (95% CI 4.1%–11.5%) of incident TB in the study population could be attributed to poor glycemic control. Limitations of the study include one-time measurement of fasting glucose at baseline and voluntary participation in the health screening service.
Good glycemic control could potentially modify the risk of TB among diabetic patients and may contribute to the control of TB in settings where diabetes and TB are prevalent.
| Diabetes, a well-known risk factor for tuberculosis, is increasingly prevalent in countries with a high tuberculosis burden.
In order to curb the dual epidemic of diabetes and tuberculosis, there is an urgent need for evidence that clarifies whether glycemic control affects the risk of tuberculosis.
To date, few studies have investigated the association between glycemic control and the risk of tuberculosis disease.
Using a Taiwanese cohort of over 120,000 participants with five years of follow-up, we found that the risk of tuberculosis among individuals with diabetes depended on the level of fasting plasma glucose measured at the start of follow-up.
In those with poor glycemic control (fasting glucose > 130 mg/dl), the risk of developing tuberculosis was doubled compared to individuals without diabetes. On the other hand, the risk of tuberculosis in patients with good glycemic control (fasting glucose ≤ 130 mg/dl) did not differ significantly from that of individuals without diabetes.
There was a linear relationship between fasting plasma glucose at baseline and subsequent risk of tuberculosis.
Assuming that these findings imply a causal effect of glycemic control on tuberculosis, 7.5% of incident tuberculosis cases could be prevented if all diabetic patients in the study population achieved good glycemic control.
Diabetes control has the potential to complement current tuberculosis control efforts, above and beyond its impact on reducing the burden of non-communicable disease.
| In its post-2015 End TB Strategy, the World Health Organization considers diabetes mellitus (DM) an important risk factor and comorbidity to be addressed in several components of tuberculosis (TB) control [1]. Recent studies suggested that DM increased the risk of active TB and was associated with higher risks of TB treatment failure, relapse after treatment completion, and mortality [2,3]. It was also noted that the greater risk of TB in patients with diabetes varied substantially across studies [2]. Meanwhile, the prevalence of DM has been rising in most low- and middle-income countries [4]. The looming co-epidemic of DM and TB could therefore undermine TB control in these countries [5]. There is an urgent need for solutions and actions to reduce the impact of DM on TB and to prevent the colliding epidemics.
Despite the well-documented association between DM and TB risk, it remains unclear whether improving glycemic control in DM patients could modify this risk. Previous studies suggested that good glycemic control was associated with better clinical outcome in common infections and decreased the risk of infectious complications from surgery [6,7]. However, evidence on the association between glycemic control and TB risk has been limited and inconsistent. While some studies suggested that good glycemic control was associated with a lower risk of TB, others did not find such an association [8–11]. In a recent modeling study of 13 countries with high TB burden, model outcomes suggested that prevention of DM would accelerate the decline of TB incidence and mortality, averting millions of TB cases and TB deaths in the next two decades [12]. It follows that glycemic control in DM patients may also be an important strategy for global TB control. We hypothesized that adequate management of blood glucose would reduce the risk of TB among diabetic patients; therefore, we conducted a cohort study to investigate the association between glycemic control in DM patients and the risk of active TB disease.
We enrolled individuals participating in a community-based multiple screening service in New Taipei City from 5 March 2005 to 27 July 2008. The service provided free screening for chronic diseases and common cancers to adults ≥ 30 y old. The screening included a questionnaire about demographic and lifestyle information, a physical examination, and blood and urine tests. Of the 127,085 people who participated in the screening service, 124,455 provided written consent to be enrolled in the study. After excluding those with a previous history of TB and those with a diagnosis of TB within the first 28 d of follow-up (n = 909), 123,546 were included in the analysis. In order to obtain detailed information on DM and TB for each individual, we used patients’ unique national identification numbers to cross-match the screening service database to the national health insurance database and the vital registry. The participants were followed up until the occurrence of TB, death, or 31 December 2012, whichever came first.
DM status and glycemic control were defined using information from the screening service (fasting plasma glucose [FPG]) and the national health insurance database. DM was defined by the prescription of a hypoglycemic drug for ≥28 d within 2 y before the date of screening or FPG ≥ 126 mg/dl at screening [13]. The hypoglycemic agents included sulfonylureas, biguanides, alpha-glucosidase inhibitors, thiazolidinediones, meglitinides, and insulin. We divided DM status into three groups based on the recommendation of the American Diabetes Association: (i) no DM; (ii) DM with good glycemic control: FPG ≤ 130 mg/dl; and (iii) DM with poor glycemic control: FPG >130 mg/dl [14]. We also determined whether the diabetic patients had DM-related complications at baseline using the national health insurance database [15].
We identified incident TB disease from the national health insurance database. In Taiwan, TB care is provided for free, and the reimbursement is done through the national health insurance system, which has a coverage rate of over 99% nationwide [16]. We defined TB as ICD-9-CM code 010–018 in the patient’s medical record plus prescription of anti-TB treatment for ≥90 d (including inpatient and outpatient services). The 90-d cutoff was used because the turnaround time for mycobacterial culture examination might be longer than 2 mo. A previous validation study was conducted using confirmed cases in the National TB Registry as the gold standard. The case definition based on ICD-9 code and prescription record was found to have a sensitivity of 87% and a specificity of nearly 100% [17].
We collected information of other covariates that are known risk factors for TB. Information on demographic and lifestyle factors was obtained from the questionnaire of the health screening service. End-stage renal disease (ESRD) was defined by estimated glomerular filtration rate < 15 ml/min (calculated from the MDRD equation). We also identified malignancy (ICD-9-CM code 140–208), pneumoconiosis (ICD-9-CM code 500–503, 505), and use of systematic steroids (prescription of steroid for ≥30 d) in the previous 2 y before screening. Lastly, because diabetic patients may be more likely to attend clinics and therefore be exposed to TB patients, we used the national health insurance database to determine the frequency of outpatient visits in the year after screening as a proxy of health service utilization.
We computed the incidence rate of TB in all participants and by DM status. We used Kaplan-Meier curves to compare the time to incident TB among the different DM groups. The survival curves were adjusted for age by reweighting the data within each DM group using the age distribution (by 5-y span) of the study population [18,19]. A Cox proportional hazards regression model was used to estimate the adjusted hazard ratio (aHR) and corresponding 95% CI for diabetic patients with poor glycemic control and those with good glycemic control, using the nondiabetic population as the reference. We adjusted for other demographic and clinical risk factors for incident TB, including age, sex, body mass index (BMI), level of education, marital status, smoking status, alcohol use, betel nut use (as a proxy measurement of socioeconomic status) [20], ESRD, malignancy, pneumoconiosis, steroid use, and frequency of outpatient visits. Because BMI was strongly associated with both DM and TB, we adjusted for BMI categorically (<18.5, ≥18.5 to <25.0, ≥25.0 to < 30.0, ≥30.0) and continuously in two different models [21]. We examined the dose-response relationship between FPG and risk of TB both linearly and nonlinearly in the Cox regression model. The potential nonlinear relationship was investigated using penalized spline regression (with three degrees of freedom), and the test for nonlinearity was done using the likelihood ratio test [22].
We conducted subgroup analyses to explore whether the association between DM status and incident TB might be modified by the following factors: (i) age (<65, ≥65 y), (ii) sex, and (iii) BMI (<25, ≥25 kg/m2). To estimate the aHRs of DM status among different subgroups, we added cross-product terms to the multivariable Cox regression model, adjusting for all other covariates. We compared models with and without the cross-product terms using the likelihood ratio test to test for effect modification.
In all, 5.4% (6,643 out of 123,546) of participants had missing data for at least one of the covariates in the analysis (Table 1). A comparison of those with and without missing information showed similar basic characteristics in the two groups (S1 Table). Under the assumption of missing at random, we used multiple imputation to impute missing data using the chained equations approach, with five imputed datasets and 20 burn-in iterations [23]. For each covariate with missing information, we used all the other covariates in the analysis (Table 1) as the predictors to impute missing values. For continuous variables we set the lower and upper bounds of imputed values using the minimal and maximal values in the observed data. The distributions of observed and imputed values did not differ substantially for all imputed covariates (S2 Table). All regression analyses were conducted in each imputed dataset; results from all imputed datasets were combined using the standard rules from Rubin [24]. The only exception was the dose-response analysis of FPG and TB, where complete case analysis was used (because the nonlinear dose-response analysis cannot be conducted using multiple imputation). We used the procedures PROC MI and PROC MI ANALYZE in SAS 9.4 (SAS Institute) for multiple imputation.
Lastly, we estimated the population attributable fraction (PAF) of TB due to poor glycemic control using the following formula:
PAF=∑i=1nPiRRi−∑i=1nP′iRRi∑i=1nPiRRi
where Pi represents the current proportion of the population in the ith DM category (no DM, DM with good glycemic control, or DM with poor glycemic control), Pi′ represents the proportion of the population in the ith DM category in the alternative scenario (had all diabetic patients achieved good glycemic control), and RRi represents the aHR (if statistically significant) between DM status and active TB based on the present study [25]. We used 1,000 Monte Carlo simulations to obtain the mean and 95% uncertainty interval (UI) of the PAF.
All analyses were conducted using SAS software version 9.4 (SAS Institute) and R software version 3.1.2 (R Project). The original prospective analysis plan from the institutional review board submission is available (S1 and S2 Texts). The main analysis in the present report (glycemic control and hazard of active TB) was consistent with the prospective analysis plan. The dose-response analysis and the subgroup analyses were formulated at the data analysis stage.
This study was approved by the ethics committee of the Taiwan National Health Research Institutes (IRB No. EC1011004-E). Written consent was obtained from each participant during enrollment.
Of the 123,546 participants, 1,504 (1.2%) had unknown DM status because of missing FPG information. In the 122,042 participants with FPG information, 11,260 (9.2%) had DM at baseline, and 8,015 of those with DM (71.2%) had poor glycemic control (FPG > 130 mg/dl) (Table 1). At baseline, compared with nondiabetic individuals, those with DM were older and more likely to be male, had higher BMI, and had a lower level of education. Among diabetic patients, the difference in baseline characteristics between those with good and poor glycemic control was small (Table 1).
The 123,546 participants were followed up for a median of 4.6 y, and 327 cases of TB developed in 540,120 person-years. The overall incidence rate of TB was 60.5 (95% CI 54.0–67.1) per 100,000 person-years. Among those with DM information (n = 122,042), the incidence rate of TB was 54.2 (95% CI 47.7–60.8), 65.1 (95% CI 22.6–107.6), and 155.5 (95% CI 114.0–196.9) per 100,000 person-years in nondiabetic individuals, DM patients with good glycemic control, and DM patients with poor glycemic control, respectively. In the Kaplan-Meier plot, TB-free survival was significantly different by DM status (p-value from log-rank test for overall difference: 0.0019; Fig 1). Compared to DM patients with good glycemic control and individuals without DM, DM patients with poor glycemic control developed TB more quickly.
In the multivariable Cox regression analysis, DM was associated with a higher hazard of incident TB compared with nondiabetic individuals (aHR 1.70, 95% CI 1.27–2.27, p < 0.001) (Table 2). The hazard was higher among those with poor glycemic control (aHR 2.21, 95% CI 1.63–2.99, p < 0.001). The hazard of TB in those with good glycemic control did not differ significantly from that in nondiabetic individuals (aHR 0.69, 95% CI 0.35–1.36, p = 0.281). When we restricted the analysis to diabetic patients without DM-related complications, the association between glycemic control and TB risk remained unchanged (Table 2). Results from the complete case analysis were very similar to those from the main analysis using multiple imputation (S3 Table).
In the linear dose-response analysis, the hazard of TB increased with FPG (aHR 1.06 per 10-mg/dl increase in FPG, 95% CI 1.03–1.08, p < 0.001). In the penalized spline regression, the positive dose-response relationship between FPG and the hazard of TB persisted (Fig 2); the test for nonlinearity was not statistically significant (p = 0.081).
Across different subgroups of age, sex, and BMI level, poor glycemic control was associated with a higher hazard of TB and good glycemic control was not significantly associated with the hazard of TB (Table 3). There was a non-significant (p = 0.053) difference in the association between poor glycemic control and TB when grouped by age: for participants <65 y old, the aHR was 3.38 (95% CI 2.25–5.09, p < 0.001), while for those ≥65 y old, the aHR was 1.63 (95% CI 1.05–2.53, p = 0.028). We did not find evidence of effect modification by sex (p = 0.658) or BMI level (p = 0.167).
Assuming the observed association between glycemic control and risk of TB was causal, we estimated that 7.5% (95% UI 4.1%–11.5%) of all TB cases in our population would have been avoided if all diabetic patients had achieved good glycemic control.
In this cohort study, we found that people with DM had a 70% greater hazard of active TB compared to nondiabetic individuals. However, the higher TB risk was not uniform across all DM patients. The hazard of TB, compared to those without DM, was higher (over 2-fold) in patients with poor glycemic control (FPG > 130 mg/dl) but was not significantly different in those with good glycemic control (FPG ≤ 130 mg/dl). In the dose-response analysis, the hazard of TB increased with increasing levels of FPG. Assuming the observed association between glycemic control and TB was causal, we determined that 7.5% of incident TB in the study population could be attributed to poor glycemic control.
Previous observational studies have shown that the risk of TB is greater in patients with diabetes to varying degrees [2]. The present analysis suggests that the variation in DM-TB association might be partially explained by different levels of glycemic control in the study populations. Few studies have investigated the association between glycemic control and TB risk. In a cohort study of older individuals in Hong Kong, DM patients with hemoglobin A1c (HbA1c) ≥ 7% had a higher risk of developing active TB than individuals without DM (aHR 2.56), while the risk among patients with HbA1c < 7% was not elevated [9]. Another cohort study of DM patients in Chile reported that 24.2% of insulin-dependent DM patients developed active TB in 10 y, while the risk of TB for other DM patients was 4.8% [26]. Baker et al. used the number of DM-related complications as a proxy measurement for DM severity and found that the risk of TB increased with increasing DM severity [8]. On the other hand, in two population-based studies in Denmark and the UK, the level of HbA1c was not associated with the risk of TB [10,11]. We note that several factors including BMI, smoking status, and alcohol use were not adjusted for in the Denmark study. High BMI is associated with poor glycemic control and a lower risk of TB; therefore, the negative result in the Denmark study could be due to confounding by high BMI. In the UK study, glycemic control was generally good in diabetic patients, with nearly two-thirds of patients having HbA1c of < 7.5%. This was in contrast to our diabetic patients, in whom only one-third had good glycemic control. The lack of association between glycemic control and TB risk in the UK study might be explained by DM being well-controlled among diabetic patients in this population. In sum, the finding from the present study, together with previous research, suggests that good glycemic control could potentially modify the higher risk of TB among DM patients.
This study is an observational study, but the finding of a beneficial effect of glycemic control on TB was unlikely to be due to biases. First, the distribution of other major risk factors for TB was similar in the two DM groups (good glycemic control versus poor glycemic control; Table 1). We note, however, that we cannot rule out the possibility of confounding by other unmeasured covariates. Second, TB patients can have transient hyperglycemia before receiving anti-TB treatment [27]. Therefore, the apparently higher hazard among DM patients with poor glycemic control could be due to reverse causality. However, the long follow-up period (>4 y) and exclusion of TB cases that occurred within the first month of follow-up minimized the chance of reverse causality. In the Kaplan-Meier plot, the group of diabetic patients with poor glycemic control was separated from the other two groups during the whole follow-up period, and the separation gradually increased over time. Therefore, our results could not be explained by reverse causality. Third, people with long-term DM may be more likely to have poor glycemic control than those with new-onset DM. As a result, the observed lower hazard of TB in those with good glycemic control could be simply due to the early stage of DM instead of being the effect of glycemic control. However, when we restricted the analysis to those without any DM complications to adjust for the duration of DM, the result remained unchanged. Overall, the evidence from our analysis supports a probable causal effect of glycemic control on the occurrence of TB.
Although the exact mechanism underlying the association between DM and TB is yet to be clearly elucidated, previous laboratory studies have suggested that both the innate and adaptive immunity related to TB defense were impaired in DM patients [28]. A few studies further shed light on the possibility that improved glycemic control could restore immune function and reverse the risk of TB. Using serial whole blood chemiluminescence, MacRury et al. found that phagocytic function was below normal in non-insulin-dependent DM patients with poor glycemic control; phagocytic function was significantly elevated when glycemic control was improved [29]. In another study, impaired granulocyte adherence was noted in patients with poorly controlled DM. After 1–2 wk of antidiabetic treatment and lowering of fasting blood glucose, granulocyte adherence improved significantly [30]. Another study found that diabetic mice had lower expression of Th1-related cytokines in response to Mycobacterium tuberculosis infection, and insulin treatment significantly improved the synthesis of related cytokines [31]. Lastly, Gomez et al. found that the attachment and ingestion of M. tuberculosis in human monocytes was lower in diabetic than nondiabetic individuals. In multivariable analysis, poorly controlled DM (measured by HbA1c and plasma glucose level) was a significant predictor of lower interaction between monocytes and M. tuberculosis [32].
In our study, the point estimate of relative risk for DM patients with good glycemic control was lower than the null value of one (compared to nondiabetic individuals), although the confidence interval was very wide. A similar pattern was also observed in a previous study of the elderly population in Hong Kong (aHR 0.81 comparing DM patients with HbA1c < 7% to those without DM, 95% CI 0.44–1.48) [9]. Baker et al. suggested that residual confounding by BMI level might explain the lower risk in those with good glycemic control in their study [8]. In our analysis, however, BMI was adjusted for both continuously and categorically (Table 2 and S4 Table). Another possibility is the anti-TB effect of metformin. In a recent study, metformin was found to inhibit the intracellular growth of M. tuberculosis in a human monocytic cell line and to improve the treatment outcome of TB patients [33]. Since metformin is a commonly prescribed antidiabetic agent, it may be possible that the “metformin group” of patients with well-controlled diabetes was driving the trend towards a protective effect for patients with well-controlled diabetes relative to nondiabetic individuals. Further studies are required to confirm the efficacy of metformin against TB.
Our study has limitations. The information on glycemic control was based on a single FPG test at baseline, and this may not reflect the long-term status of individuals’ glycemic control during the study period. Previous large-scale studies showed a good correlation between FPG and HbA1c [34,35]. In addition, levels of FPG and HbA1c were both found to correlate well with the prevalence of diabetic retinopathy in several populations [34,36]. In our study, the single measurement of fasting glucose was still strongly predictive of subsequent development of TB. In case of any measurement error of glycemic control in our study, this error would likely be nondifferential with regard to the risk of TB (after adjusting for other major TB risk factors), and would mostly likely bias our results toward a less significant association between glycemic control and TB. In other words, the association between glycemic control and TB might have been even larger if we had obtained more complete information on long-term glycemic control over time. Furthermore, we do not have information on latent TB infection at baseline because tuberculin skin tests and interferon gamma release assays were not performed in the screening survey. Further studies are needed to better understand the role of DM in primary progressive TB versus reactivation disease.
The study population was voluntary participants of a community-based health screening service. It is possible that nonparticipants were at greater risk of poor glycemic control as well as greater risk of TB, causing selection bias. In addition, although major risk factors for TB were adjusted for in the analysis, we cannot rule out the possibility of unmeasured or residual confounding in this observational study. The definition of DM was based on prescription of antidiabetics and FPG. It was possible that nondiabetic patients with obesity or polycystic ovary syndrome were misclassified as DM patients because of metformin use. We conducted a sensitivity analysis excluding metformin from the list of antidiabetics in DM definition, and the results remained unchanged. Lastly, the diagnosis of TB was based on the national health insurance database. To explore the impact of outcome misclassification, we conducted a sensitivity analysis using different durations of anti-TB treatment (30 d and 60 d) to define TB, and the results were similar.
In some countries with low or intermediate burden of TB, non-foreign-born TB cases are increasingly concentrated in the elderly population as a result of reactivation from remote latent infection. In these settings, TB case detection and treatment will have limited impact on the incidence of TB disease. Preventive therapy can effectively reduce the risk of TB in those with latent TB infection, but potential drug toxicity limits its use in the elderly [37]. On the other hand, DM is a prevalent disease in the elderly and contributes substantially to TB burden, especially in populations with poorly controlled diabetes [38]. Management of DM provides an alternative solution to reduce TB in the elderly. In our study cohort, 70% of DM patients had suboptimal glycemic control (FPG > 130 mg/dl) despite the universal coverage of national health insurance. Consistent with our finding, the percentage of patients with poor glycemic control (defined as HbA1c ≥ 7.0%) was 68% and 66% in 2006 and 2011, respectively, in recent national surveys [39]. Further studies are needed to identify and evaluate effective strategies to improve glycemic control at the population level [40].
DM is a major risk factor for TB and will likely be an important driver of TB epidemiology in the upcoming decades [41]. In a modeling study, Pan et al. found that prevention of DM could avoid millions of TB cases and TB deaths in 13 high-burden countries over the next two decades [12]. Our study provides further evidence that, in addition to prevention of DM, improving glycemic control in DM patients may also benefit TB control. Echoing the new WHO End TB Strategy, we urge that more efforts be made to link non-communicable and communicable disease programs in order to leverage the overall impact on disease control and prevention. In practice, the comprehensive program for DM-TB management should include prevention of DM, early detection of DM followed by proper glycemic control, and bi-directional screening of DM and TB.
|
10.1371/journal.pntd.0001775 | Toxocariasis and Epilepsy: Systematic Review and Meta-Analysis | Human toxocariasis is a zoonotic infection caused by the larval stages of Toxocara canis (T. canis) and less frequently Toxocara cati (T. cati). A relationship between toxocariasis and epilepsy has been hypothesized. We conducted a systematic review and a meta-analysis of available data to evaluate the strength of association between epilepsy and Toxocara spp. seropositivity and to propose some guidelines for future surveys.
Electronic databases, the database from the Institute of Neuroepidemiology and Tropical Neurology of the University of Limoges (http://www-ient.unilim.fr/) and the reference lists of all relevant papers and books were screened up to October 2011.
We performed a systematic review of literature on toxocariasis (the exposure) and epilepsy (the outcome). Two authors independently assessed eligibility and study quality and extracted data. A common odds ratio (OR) was estimated using a random-effects meta-analysis model of aggregated published data.
Seven case-control studies met the inclusion criteria, for a total of 1867 participants (850 cases and 1017 controls). The percentage of seropositivity (presence of anti-Toxocara spp. antibodies) was higher among people with epilepsy (PWE) in all the included studies even if the association between epilepsy and Toxocara spp. seropositivity was statistically significant in only 4 studies, with crude ORs ranging 2.04–2.85. Another study bordered statistical significance, while in 2 of the included studies no significant association was found. A significant (p<0.001) common OR of 1.92 [95% confidence interval (CI) 1.50–2.44] was estimated. Similar results were found when meta-analysis was restricted to the studies considering an exclusively juvenile population and to surveys using Western Blot as confirmatory or diagnostic serological assay.
Our results support the existence of a positive association between Toxocara spp. seropositivity and epilepsy. Further studies, possibly including incident cases, should be performed to better investigate the relationship between toxocariasis and epilepsy.
| Human toxocariasis is an infection caused by the larval stage of the worms Toxocara canis and less frequently Toxocara cati, common parasites of domestic and peridomestic dogs and cats. It is a cosmopolitan infection, occurring whenever the man-soil-dog relationship is particularly close, especially in tropical countries, where the humid climate favours the survival of parasite eggs in the soil, and in rural settings, where the poor hygiene increases the probability of human infection. Epilepsy affects nowadays at least 65 million of people worldwide and is particularly common in tropical areas, probably because of the presence of cases caused by infectious diseases largely absent in industrialized countries. For several decades, researchers have investigated the possible association between toxocariasis and epilepsy. In this study we conducted a statistical analysis of all the data available on the relationship between these two conditions. The combined results of the 7 studies included indicate an association between the two diseases. Further studies are necessary to demonstrate a causal relationship (i.e. toxocariasis causes epilepsy). Considering that toxocariasis is a preventable and common disease, a better understanding of the relationship between toxocariasis and epilepsy may contribute to improving prevention of epilepsy worldwide.
| Human toxocariasis is a parasitic zoonosis caused by the larval stages of the ascarids Toxocara canis (T. canis), the common roundworm of dogs, and by the roundworm of cats, Toxocara cati (T. cati) [1]. The reported prevalence of soil contamination with Toxocara spp. eggs is variable between studies, going from a percentage of 6.6 to 87.1% [2]–[9]. Therefore toxocariasis is one of the most prevalent zoonotic helminth infections, occurring whenever the man–soil–dog relationship is particularly close. High seroprevalence rates of Toxocara spp. (presence of sera anti-Toxocara spp. antibodies) have been found in tropical countries, where the humid climate favours the survival of parasite eggs in the soil, and in rural settings, where the poor hygiene and the rare administration of anthelmintic treatments to dogs increases the probability of human infection [10]–[12]. Nevertheless, the reported seroprevalence in apparently healthy adults from urban areas of Western countries is of 2–5% [13], whit a wider range (2.4%–31.0% [14], [15]) when considering all the studies carried out in Europe, independently from age of participants and type of setting. Despite being the most prevalent human helminthic infection in some industrialized countries [16], toxocariasis remains relatively unknown to the public [17] and the true magnitude of the global burden of Toxocara spp.-associated human disease has still to be evaluated [18].
Humans are infected by the accidental ingestion of embryonated Toxocara spp. eggs present in contaminated soil or food, or by the ingestion of encapsulated larvae contained in the raw tissues of paratenic hosts, such as cows, sheep or chickens [1], [19]. The clinical manifestations of human toxocariasis vary from asymptomatic infection to severe organ injury, depending on the parasite load, the sites of larval migration and the host's inflammatory response [20]. Two severe clinical syndromes are classically recognised: visceral larva migrans (VLM), systemic disease caused by larval migration through major organs, and ocular larva migrans (OLM), in which the disease is limited to the eyes and the optic nerves. Two less severe syndromes have also been described: ‘covert toxocariasis’, seen mainly in children and characterized by fever, headache, behavioural and sleep disturbances, cough, anorexia, abdominal pain, hepatomegaly, nausea and vomiting, and ‘common toxocariasis’, seen predominantly in adults with weakness, pruritus, rash, difficult breathing and abdominal pain [20]. Clinical involvement of the central nervous system (CNS) in visceral larva migrans is thought to be rare, although in experimental animals the larvae frequently migrate to the brain [21]–[23]. The CNS migration may lead to a variety of neurological disorders such as meningo-encephalitis, myelitis, cerebral vasculitis, optic neuritis [23], [24] and probably cognitive [25] and behavioural [26] disorders.
Concerning epilepsy, early reports have suggested a high exposure rate to Toxocara spp. among people with epilepsy (PWE) [27], [28]. In particular, in 1966 Woodruff et al. [27] found that 7.5% of PWE had a positive skin reaction to an antigen prepared from adult T. canis, in contrast to 2.1% of apparently healthy persons. In addition, they noted a statistically significant association between contact with dogs and positive skin test to toxocaral antigen in PWE. Following these preliminary observations and prompted by the development of serodiagnostic tests with improved sensitivity and specificity, further studies have been carried out in different populations to investigate the possible association between Toxocara spp. seropositivity and epilepsy, suggesting that toxocariasis could play a role in the incidence of epilepsy in endemic areas [29]–[31].
Considering that toxocariasis is one of the most common helminthiasis worldwide and that it is a potentially preventable disease, a correct estimate of the association between toxocariasis and epilepsy is necessary.
We carried out a systematic literature revision and a meta-analysis to evaluate the possible association between human toxocariasis and epilepsy and to highlight some methodological points to be taken into account for the elaboration of future surveys.
A systematic search without past time or language restriction was conducted to identify published and unpublished articles dealing with the association between toxocariasis and epilepsy. The following online databases were independently examined by two researchers (GQ and BM): MEDLINE, IngentaConnect, ScienceDirect (Elsevier), Refdoc (ex ArticleScience), Scopus, Highwire. In addition, the database from the Institute of Neuroepidemiology and Tropical Neurology of the University of Limoges (IENT): Virtual Library on African Neurology, BVNA (http://www-ient.unilim.fr/), which contains more than 9000 references of medical dissertations, theses and articles dealing with tropical neurology and parasitology, was examined. In MEDLINE combined text words and Medical Subject Headings (MeSH) terminology were used. The following search key words and Boolean operators were entered: “toxocariasis” AND “epilepsy” AND “epidemiology”. The term “toxocarosis” as an alternative to “toxocariasis” was also considered. The literature search was adapted for the other databases. Titles and available abstracts were scanned for relevance, identifying papers requiring further consideration. Reference lists of all available reviews, primary studies and books found were screened manually. When necessary, corresponding authors of relevant studies were contacted. Experts in the field were also contacted to find out other eventual non-published studies. The systematic search was realized up to October 2011.
Considering epilepsy as the outcome and toxocariasis as the exposure, all the studies meeting the following eligibility criteria were included:
Studies including only acute symptomatic seizures or specific seizure patterns or epileptic syndromes were excluded.
Full copies of all reports identified by the electronic or hand searching were obtained and two reviewers (GQ and BM) independently assessed their eligibility and extracted data.
The following data were independently recorded in an ad hoc created collecting form: author, country, study design, study population (number, age group, gender, setting) and recruitment methods. For toxocariasis, specific information was recorded on methods used for diagnosis. Considering epilepsy, details on definition and assessment were extracted. Discrepancies between reviewers were rechecked and consensus was achieved by discussion.
For each survey, the crude odds ratio (OR) on the association between toxocariasis and epilepsy and the relative 95% confidence interval (CI) were recalculated. Furthermore, statistical power was calculated as a priori and a posteriori. A priori statistical powers were calculated following the hypothesis that the objective of the survey was to identify a minimum OR of 2 (i.e., Toxocara spp. exposure leads to twice more epilepsy) with one control per case, based on the number of PWE and the percentage of Toxocara spp. seropositivity in PWOE. The a posteriori statistical powers were calculated upon the results of the surveys. In both cases a 5% alpha risk was considered. Powers were calculated using Epi-Info 6.04 [32].
To estimate the association between toxocariasis and epilepsy we performed a meta-analysis applying a random effects model, assuming that the true effect size of exposure varies from one study to the other, and that the studies in our analysis represent a random sample of effects that could have been observed [33]. A common risk was estimated as a common OR from all the studies. The homogeneity was tested by the Cochran Q test of heterogeneity. In order to account for the different age groups considered, the analysis was then separately applied to the studies including an exclusively juvenile population [30], [31]. Furthermore, considering that Western Blot (WB) is as sensitive but more specific than enzyme-linked immunosorbent assay (ELISA) [34], we also conducted an analysis restricted to the studies using WB as diagnostic or confirmatory test [35]–[38].
The meta-analysis was performed using EasyMA, 2001 version [39]. The PRISMA (Preferred Reporting Items for Systematic reviews and Meta-Analyses) statement [40] was used as a guide in the reporting of this study.
A flowchart summary of the literature search is shown in Figure 1. A PRISMA flowchart is also shown (Figure S1). Electronic search produced 131 publications, among which 25 dealt with epilepsy and toxocariasis. The removal of duplicate citations and the screening of abstracts permitted to isolate 8 documents [27], [31], [35]–[38], [41], [42]. Two additional publications [28], [30] were found by hand searching (reference lists check). Full text review of the 10 documents permitted to exclude 3 of them for not fulfilling the inclusion criteria: one [27] was excluded because methods to assess epilepsy were not reported and toxocariasis infection was detected through a skin test; furthermore the included cases consisted of a highly selected group of severe patients with epilepsy. Another study [28] was excluded because toxocariasis infection was exclusively assessed in a sample of PWE without control group. The last study [42] was excluded because of the lack of reporting of aggregated data for each group.
Considering the 7 articles meeting the inclusion criteria, the materials and methods of the study reported by Nicoletti et al. (2007) [36] had been previously detailed in Nsengiyumva et al. [43], while the study population of Winkler et al. (2008) [38] has been better described in Winkler et al. (2009) [44]. The methodological aspects of these articles have been therefore assessed considering both the publications.
Seven case-control studies [30], [31], [35]–[38], [41] were included, providing a total subjects number of 1867 (850 PWE and 1017 PWOE). Two of them [30], [31] considered a population aged 1–17 years while one excluded children aged 10 years or younger [38]. The studies were carried out in 6 different countries (USA, Italy, Bolivia, Turkey, Burundi and Tanzania), both in rural [35], [36], [38] and urban [30], [31], [37], [41] settings. In the study by Akyol et al. [41] 10% of participants were from rural areas, but no significant relationship was found between residency and seropositivity rate. The general characteristics of the included studies are shown in table 1.
Three surveys had a matched case-control design [35]–[37] among them age was the only common matching criteria. Only one study was a population-based survey [35].
The epilepsy definition proposed by the International League Against Epilepsy (ILAE) in 1993 [45] was applied in 3 studies [35]–[37] while Glickman et al. [30] considered the definition proposed in 1972 by Alter [46], and Winkler et al. [38] defined epilepsy according to the World Health Organization (WHO) Neurosciences Research Protocol proposal [47]. In the work by Arpino et al. [31] a general definition of “positive seizure history” was considered as cases entry criteria. Considering seizures, 4 studies [35]–[37], [41] applied the classification of epilepsies and epileptic syndromes proposed by the ILAE in 1981 [48], while one [38] used an adjusted classification for rural African hospitals suggested in 2007 [49].
All PWE were prevalent cases and none of the studies clearly specified if active or lifetime epilepsy was considered, the second being more probable.
Controls were out- or in-patients attending the same hospital of cases [30] or people going to hospital for vaccination or haematological check [36], [37] or volunteers [41]. A negative history for seizures [31], [36]–[38], [41] and for both seizures and other neurological diseases [31], [36], [37] was considered for controls definition. In the population-based survey controls were selected from the same community, but different households, of cases [35], whereas another study selected controls from the same province of PWE excluding blood relationship [36]. In an attempt to determine the accuracy of the seizures classification EEG recordings were examined in some studies [31], [35]–[37].
A neurologist confirmed both cases and controls through anamnesis and complete neurological examination in 4 studies [35]–[38].
In order to obtain demographic data and information concerning factors possibly associated with Toxocara spp. exposure a questionnaire was administered to cases and control subjects in 5 studies [30], [31], [36], [37], [41]. Data were usually obtained by the patient's mother when the study population was infantile [30], [31]. The questionnaire version used was specified only in one study [36] and interviewers qualifications were stated only in 2 surveys [36], [37].
Presence of anti-Toxocara spp. antibodies in sera was assessed using antibodies-ELISA (Ab-ELISA, commercial or in-house kits) [30], [31], [41], or immunoblot [36], [37] or Ab-ELISA followed by WB confirmation [35], [38]. Laboratories performing the analysis were blind to the case-control status of sera samples in 3 studies [35]–[37].
The results of the included studies are shown in Table 2. Toxocara spp. seropositivity ranged from 6.5% to 50.8% in the control group and from 12.0% to 59.7% in PWE. Seroprevalence rate was higher among PWE than control subjects in all the 7 included surveys, even if the association between Toxocara spp. seropositivity and epilepsy was statistically significant in 4 of them [30], [31], [35], [37]. In one study the crude OR bordered on statistical significance, anyway, after adjustments on other variables according to a multivariate model using the conditional logistic regression, a stronger and significant association was found [36].
Significant crude ORs ranged between 2.04 and 2.85. A priori statistical power ranged 32.8–90.9% and a posteriori statistical power 8.0–89.6%.
A meta-analysis was at first performed on all the 7 studies included. Results are presented in figure 2. A significant (p<0.001) common OR of 1.92 (95%CI 1.50–2.44) was estimated. The test of heterogeneity was not significant (p = 0.545), indicating homogeneity of the studies included. When analysis was restricted to the 2 studies considering only a juvenile population [30], [31], as shown in figure 3, a common OR of 2.23 (95% CI 1.35–3.69; p = 0.002) was found. The test for heterogeneity was also not significant (p = 0.655). The meta-analysis was at last restricted to the 4 studies using WB test [35]–[38], as shown in figure 4, leading to an OR of 1.91 (95% CI 1.33–2.75, p<0.001) and a non significant test for heterogeneity (p = 0.430).
We performed a systematic literature revision and a meta-analysis of available data to evaluate the association between epilepsy and toxocariasis. To our knowledge this is the first meta-analysis on this argument. Based on our literature search, we analyzed data from 7 case-control studies carried on in rural or urban settings and in various countries worldwide. We are confident that our literature search is exhaustive as conducted on several electronic databases and also on a specific database containing literature on tropical neurology and parasitology including theses and memos unpublished in international or electronic databases.
Seroprevalence rate of anti-Toxocara spp. antibodies was higher among PWE than control subjects in all the 7 studies analysed [30], [31], [35]–[38], [41] even if only 4 showed a significant positive association between Toxocara spp. seropositivity and epilepsy [30], [31], [35], [37] and a fifth reached statistical significance only after adjustment for other variables [36]. In our meta-analysis we found evidence of positive association, with a common OR of 1.92 (95%CI 1.50–2.44) and a lack of heterogeneity between the studies.
Our result is noteworthy for coming from studies across different populations in disparate geographic locations and socioeconomic climates. This consistency of observations among different populations is in favor for a causal relationship between toxocariasis and epilepsy [50]. Anyway, various important points should be taken into account when interpreting our data.
First of all, the studies evaluated were retrospective case-controls ascertaining both Toxocara spp. seropositivity and epilepsy in a cross-sectional manner; thus, the inclusion of “prevalent” rather than “incident” cases does not permit to demonstrate a temporal relationship between the exposure (Toxocara spp.) and the outcome (epilepsy) and doesn't allow to exclude a possible “reverse causality”. It has been in fact hypothesized that the abnormal behavior patterns (e.g. pica and hyperactivity) and the elevated number of falls to the ground of PWE (especially children or mentally retarded) could predispose them to Toxocara spp. exposure rather than the contrary [30]. In particular, evidence of association has been reported between Toxocara spp. seropositivity and mental retardation [51], [52]. We underline anyway that the study by Nicoletti et al. (2002) [35] found no statistical difference in seroprevalence between PWE with or without mental retardation. On the other hand, a significant difference in the frequency of mental retardation between seropositive and seronegative subjects was found by Nicoletti et al. (2008) [37], but it lead to only a slight reduction of association after restriction of the analysis to the PWE without mental retardation.
Selection of cases and controls represents one of the most important pitfalls in case-control studies. In the studies evaluated, with the exception of the only population based survey [35], PWE and PWOE were generally enrolled from a hospital setting, and their source population was often not clearly defined. This constitutes a possible recruitment bias, especially in rural settings, where people receiving care are not representative of the general population. In particular, concerning controls, hospital controls might resemble cases more than population controls, biasing OR towards the null [53]. Furthermore, a volunteer bias could have affected the study by Akyol et al. [41], where the control group was composed by volunteers coming from an undefined source. However, the population based survey showed a positive association, similar to the results found by the positive hospital based studies, suggesting that the selection bias effect could be limited.
Considering cases and controls characteristics, PWE and PWOE should be comparable at least for age, because the prevalence of both epilepsy and anti-Toxocara spp. antibodies vary with age, and for geographical provenience and education, which are likely related to the level of exposure to Toxocara spp. On this point, the studies examined are often lacking of detailed descriptive data. We report as an example, the comment by Quet et al. [54] on the study by Akyol et al. [41], which noticed how the greater number of students observed in the control group could suggest a higher education in controls than cases. The educational level was in fact expressed as a binary variable (less or more than primary school) in this study, which could give an unreliable estimation of education; in such cases the number of school years might be a more precise measure. In order to account for the different age groups included, and considering that young age seems to contribute to Toxocara spp. exposure [1], we restricted our meta-analysis to the studies with an exclusively juvenile population and we obtained also in this case a significant positive association (OR 2.23, 95% CI 1.35–3.69, p = 0.002). In particular, in the study by Arpino et al. [31] the relationship was more remarkable in children under 5 years old. Based on these findings, it has been suggested that the parasite may act as a cofactor in determining the occurrence of seizures in children with a predisposing background [31]. Only a prospective cohort follow-up study could avoid these biases. However such a design, leaving subjects exposed to toxocariasis and without intervention, is ethically not acceptable.
A potential weakness of our study is the use of different and not always clears epilepsy definitions in the included articles. On this point, considering that the lag time between Toxocara spp. infection and epilepsy occurrence is not yet defined, we underline the importance of including lifetime and not only active epilepsy, as likely properly done in the studies examined. On the other hand, the lack of exhaustive descriptive data on the age of onset, on the probable etiology and on seizures classification didn't permit us to differentiate our analysis for these factors. The significant positive association found in some studies between Toxocara spp. seropositivity and partial epilepsy could in fact be biologically plausible, given the higher prevalence of idiopathic epilepsy among the generalized forms [35], [37]; while the lack of association between partial epilepsy and toxocariasis found by Nicoletti et al. (2007) [36] has been related by the authors to a lack of power. In the study by Akyol et al. [41], besides the lack of a precise definition of epilepsy, the authors reported a higher frequency of generalized epilepsies, as expected, because of the inclusion of only cryptogenic (in the abstract) or idiopathic (in the methods) epilepsies [54]. This could have affected the results, showing no statistically significant association between toxocariasis and epilepsy. A correct classification of seizures, possibly with the help of EEG recordings, is therefore an important element that should be taken into account in future studies to permit a correct interpretation of the results.
Regarding the diagnosis of toxocariasis, the major limitation in confronting different studies consists in the heterogeneity of techniques (Ab-ELISA or WB or both) used to detect sera anti-Toxocara spp. antibodies, mostly due to different cost and availability. Also when considering ELISA, different kits (commercial or in house) with different sera dilutions were utilized and a serum pre-adsorption with larval Ascaris extracts was carried on only in some studies. It would have been interesting evaluating and reporting the sensitivity and specificity of the used assays, which has never been done in the studies examined. Considering that the WB confirmation of positive results from the ELISA (especially where pre-absorption is not carried out) has been recommended [55], and given the higher specificity of WB [34], we restricted our meta-analysis to the studies applying WB, obtaining results similar to the global analysis (OR 1.91, 95%CI 1.33–2.75, p<0.001).
When interpreting these data, we are of course aware that other factors, such as Toxocara spp. excretory-secretory (TES) antigen preparations, parasite strains, and WB technical procedures, could have influenced the results obtained by different investigators. It should also be kept in mind that a single seropositivity has limited pathological significance and could probably represent past rather than recent infection. Furthermore, the presence of sera antibodies against Toxocara spp. does not provide evidence of either an active systemic infection or a CNS involvement. Diagnosis of neurotoxocariasis is in fact based on the history; blood tests, including differential blood cell count and determination of serum total IgE; CSF investigation, including the detection of anti-Toxocara spp. antibodies and neuroimaging [13].
The absence of significant results was associated with a lower power (type II error), making not really surprising the lack of statistical confirmation of the difference found. The statistical power of a study can be improved performing surveys in areas with high level of exposure assessed with the most sensitive assay or, when the number of cases is small, increasing the ratio of controls to cases up to 4/1 [53]. The low a posteriori power of the studies by Winkler et al. [38] (8.0%) and Akyol et al. [41] (11.3%) could be mostly accounted to the small sample size and in particular the lower number of controls than cases, highlighting one more time the central role of the elaboration of the control group.
In our paper we referred to toxocariasis etiological agent as Toxocara spp. and not only T. canis. TES in fact are not species-specific and the differentiation between T. canis and T. cati remains challenging. Considering the prominence historically given to T. canis, the role of T. cati in human toxocariasis could have been underestimated. Further work should be encouraged to differentiate the two parasites and to better address future prevention strategies [56].
The most frequent infectious agent involved in the differential diagnosis of subjects with late-onset epilepsy or inflammatory brain nodules is the larval stage of Taenia solium (T. solium), aetiological agent of neurocysticercosis (NCC). Concomitant T. solium and Toxocara spp. seropositivity is a possible event in areas endemic for both helminthes. Anyway, albeit there is yet no evidence on the mechanisms underlying toxocariasis-induced epilepsy, according to us toxocariasis should not be ruled out as an accidental association. First of all, the presence of anti-T. solium antibodies, as in the case of toxocariasis, could represent only a previous exposure without established infection. Furthermore, considering the studies included in our analysis, in the study by Nicoletti et al. (2002) [35] only 7 PWE over a total of 113 were positive to both T. solium and Toxocara spp. and in the study by Nicoletti et al. (2007) [36], finding a positive association between Toxocara spp. seropositivity and epilepsy, seropositivity for cysticercosis was considered as a variable in the multivarate analysis. Of course, the interpretation of serological results should always take into account the background seroprevalence of both Toxocara spp. and T. solium in the studied population and cysticercosis seropositivity should always be evaluated as a possible confounder when carrying on surveys on infectious agents and epilepsy.
In conclusion, a positive association between Toxocara spp.-seropositivity and epilepsy could be hypothesized; nevertheless, even the modestly strong association demonstrated in our meta-analysis does not necessarily prove causality (i.e., Toxocara spp. infestation caused the epilepsy). Further studies, considering incident rather than prevalent cases and with a population-based design, should be performed. An internationally accepted epilepsy definition and seizures classification should be applied and cases and controls should be comparable at least for age, sex, geographic provenience, education and socio-economic background. Pica, pet owning, mental retardation and other possible toxocariasis risk factors should be assessed trough a validated questionnaire administered by trained investigators and assessors and laboratory personnel should be blind to the status of participants. The improvement of techniques permitting to distinguish recent from past infections, such as antigen-ELISA (Ag-ELISA), should be encouraged in order to better investigate the time relationship between Toxocara spp. infection and epilepsy occurrence.
Assessing the link between toxocariasis and epilepsy is of interest as toxocariasis is a potentially preventable disease. Nowadays, despite the implementation of regular and intensive de-worming programs in western countries, the parasite still prevails, indicating that prevention is not easy in practice. Good hygiene practices should be encouraged and further strategies to prevent Toxocara spp. transmission should be identified and applied, permitting to experimentally investigate the causation hypothesis [50]. The existence of a causal relationship and the estimation of the impact of toxocariasis on the global burden of epilepsy may strongly contribute in encouraging further programs on toxocariasis prevention worldwide, in order to control both the Toxocara spp. transmission and the related epilepsy burden.
|
10.1371/journal.pcbi.1001024 | A Systems Model for Immune Cell Interactions Unravels the Mechanism of Inflammation in Human Skin | Inflammation is characterized by altered cytokine levels produced by cell populations in a highly interdependent manner. To elucidate the mechanism of an inflammatory reaction, we have developed a mathematical model for immune cell interactions via the specific, dose-dependent cytokine production rates of cell populations. The model describes the criteria required for normal and pathological immune system responses and suggests that alterations in the cytokine production rates can lead to various stable levels which manifest themselves in different disease phenotypes. The model predicts that pairs of interacting immune cell populations can maintain homeostatic and elevated extracellular cytokine concentration levels, enabling them to operate as an immune system switch. The concept described here is developed in the context of psoriasis, an immune-mediated disease, but it can also offer mechanistic insights into other inflammatory pathologies as it explains how interactions between immune cell populations can lead to disease phenotypes.
| A functional immune system requires complex interactions among diverse cell types, mediated by a variety of cytokines. These interactions include phenomena such as positive and negative feedback loops that can be experimentally characterized by dose-dependent cytokine production measurements. However, any experimental approach is not only limited with regard to the number of cell-cell interactions that can be studied at a given time, but also does not have the capacity to assess or predict the overall immune response which is the result of complex interdependent immune cell interactions. Therefore, experimental data need to be viewed from a theoretical perspective allowing the quantitative modeling of immune cell interactions. Here, we propose a strategy for a quantitative description of multiple interactions between immune cell populations based on their cytokine production profiles. The model predicts that the modified feedback loop interactions can result in the appearance of alternative steady-states causing the switch-like immune system effect that is experimentally observed in pathologic phenotypes. Overall, the quantitative description of immune cell interactions via cytokine signaling reported here offers new insights into understanding and predicting normal and pathological immune system responses.
| Inflammation is an organism's protective response to injury, pathogens or irritants and represents a complex multicomponent process that mobilizes immune cells to remove pathogens and restore tissue homeostasis. Healthy inflammatory reaction only lasts for a relatively short period of time, in contrast to pathological conditions where inflammation can persist over period of months or years. Chronic inflammation can be harmful and is attributed to the loss of balanced interactions between immune cells. Such interactions occur either via relatively small soluble proteins known as cytokines and chemokines, or through direct cellular interactions between ligands and their receptors expressed on the cellular surface [1]. Pathologies related to the immune system lead to a number of human diseases including psoriasis [2], arthritis [3], cancer [4], atherosclerosis [5], diabetes [6], inflammatory bowel disease [7], and asthma [8]. Even though each inflammation-mediated disease carries a set of unique features, a common trait between many inflammation-associated diseases is the chronic elevations of cytokine concentrations in the affected area.
Skin is a preferred system for studying inflammatory conditions, as tissue can be both easily observed and sampled. Due to its easy accessibility it can be viewed as the “window” to the human immune system. Skin is composed of mainly two layers containing different cell types: keratinocytes are the major cell type forming the outer epidermis, whereas fibroblasts are the major component of the underlying dermis. In addition, various immune cells such as dendritic cell, T cells, neutrophils or natural killer cells reside in the skin and increase in number under inflammatory conditions [9]–[11]. Perturbations in the local immune system are found to be essential factors mediating skin disease [2]. Psoriasis is a chronic inflammatory skin disorder in which keratinocytes proliferate at an unusually rapid rate. The disease affects about 0.6–4.8% of the population [12] and is characterized by red, scaly patches that reveal fine silvery scales. Psoriasis usually develops on the knees, elbows and scalp, but can appear anywhere on the body [13]–[14]. Psoriasis serves as a good model for studies of inflammatory mechanisms and it is an attractive disease for proof-of-principle studies of new anti-inflammatory therapeutic strategies [15]. A schematic view of the role of the immune system in normal and inflamed skin is provided in Figure 1.
A major histological feature of lesional psoriatic skin is the thickened epidermis which is due to hyperproliferation and abnormal differentiation of keratinocytes (Figure 2A and 2B). The increase in number of keratinocytes is about four-fold compared to normal skin [16]. The transition from normal to diseased skin has been shown to be dependent on immune cell infiltration into the dermis and epidermis (Figure 1) [2], [15]. Keratinocytes and immune cells interact via the release of soluble mediators such as cytokines and chemokines, as well as through cell-cell interactions mediated by surface-expressed ligands and receptors.
A widely held view is that psoriasis occurs as a result of unbalanced interactions between cells of the immune system, their mediators and keratinocytes [15]. Genetic studies have allowed the identification of a substantial number of loci harboring polymorphisms influencing the susceptibility to or protection from psoriasis. These studies are diverse and range from typing of serological variants of HLA-Cw6, to whole-genome linkage or association (GWAS) studies [17]. Although genetics studies validate the notion that key cytokine pathways are involved in the genetic susceptibility to psoriasis, they do not offer an explanation on how and why genetic variations in the cytokine mediated pathways lead to chronic inflammation. It is also unclear why in psoriasis some areas in skin are chronically inflamed, while others show no symptoms despite carrying the same disease-associated alleles. It is important to note that answering similar questions may be crucial in other inflammation-associated human conditions.
It has been suggested that chronic inflammation occurs as a result of a modified regulation in key immune cell populations via cytokine-mediated interactions. For example, T cells are reported to be regulated by dendritic cells via feedback control mechanisms [18], whereas Th17 cytokine mediated CCL20 expression in keratinocytes is implicated in psoriasis pathogenesis [19]. IL-21 has been shown to induce IL-17 through a self-amplifying loop [20]. IFN-γ and IL-17 secreted by activated CD4+ cells have been reported to up-regulate IL-6, IL-8, and CXCL10 production by benign prostatic hyperplasia cells [21], suggesting a positive feedback loop that amplifies inflammation in prostatic conditions. In another study, an example of the negative feedback loop in the NF-κβ-dependent cytokine pathway is reported to elevate the expression of proinflammatory cytokines [22]. Negative feedback control of the autoimmune response has been reported to occur through antigen-induced differentiation of IL-10-secreting Th1 cells [23]. Altogether, the above studies suggest that it is essential to investigate how immune cell populations switch from a healthy to a pathologic inflammatory response as a result of modified feedback interactions. Modified feedback loop interactions between immune cells in inflammation require the development of new computational strategies to describe how alterations in feedback loops relate to pathology.
A number of computational studies have offered insights into immune system signaling in the context of human disease. A model for a cell-cell interaction network has demonstrated that the loss of responsiveness in feedback signaling pathways is necessary and sufficient to induce leukemic transformation [24]. Immune system responses were evaluated for the tumor-immune system interactions by a mathematical model for melanoma invasion into healthy tissue [25]. It is reported that small metastatic lesions distal to the primary tumor mass can be held to a minimal size via the immune interaction with the larger primary tumor. A computational model has been used to determine the steady-state basal plasma glucose and insulin concentrations determined by their interaction in a feedback loop [26] and became one of the most well-recognized approaches for evaluating diabetes. Mathematical models developed to describe the dynamics of T cell homeostasis and proliferation were applied to provide insights into the CD4+ memory T cell depletion dynamics in HIV [27]. Other applications of translational systems biology in inflammation have been recently summarized in a comprehensive review [28].
These and other studies [29]–[34] have demonstrated that mathematical modeling can offer new insights into various aspects of inflammation by linking various experimental observations into an integrative model. However, the basic principles that distinguish healthy from pathologic inflammatory responses have not been elucidated or clearly explained yet. While it has been suggested that cytokine receptor polymorphisms can modify cytokine production by a small amount, there is currently no clear understanding of how such - seemingly insignificant - alterations can lead to disease. Experimental and computational studies need to lead to a framework that links genetic mutations to the (small) modifications of feedback loop interactions between immune cells which, in turn, may lead to pathology.
In order to address some of the outstanding questions and increase our understanding of how immune cell interactions contribute to normal or inflamed skin phenotypes, we developed a quantitative model that captures cytokine-dependent production profiles of cytokines in immune cell populations. The model represents the immune cell interactions as coupled cytokine concentration levels in human tissue by quantifying the underlying feedback loops. The approach allows the application of general concepts in dynamic systems modeling, such as stable homeostatic solution, feedback loops, bistability or oscillations, and thereby, uncovers the causes of chronic inflammation. Moreover, the methodology has the power to differentiate inflammatory disease phenotypes according to mechanisms of immune system imbalance. In this study we consider possible scenarios of cell population interactions and we show how even small changes in cytokine production rates by a single cell population can significantly affect systems properties due to altered feedback interactions and cause immune system-mediated pathology. The model also allows for discrimination between a healthy inflammatory response and chronic inflammation. Due to shared cytokine pathways between psoriasis and other chronic inflammatory diseases, the principles introduced in this study might be applicable to a wider range of immune system disorders.
Given the importance of cytokine-mediated interactions between immune cells, cytokine genes, gene products and their receptors have been subjected to genetic and immunological analysis. Cytokines form a group of candidate susceptibility genes in psoriasis [35]. For example, polymorphisms of the INF-γ and IL-10 genes were shown to be associated with different levels of cytokine production in patients with psoriasis [35]. Psoriasis is associated with over-expression of T-helper cell type 1 (Th1) cytokines, IFN-γ and TNF-α in the involved skin and relative underexpression of T-helper cell type 2 (Th2) cytokines, interleukin IL-4 and IL-10 [36]. Currently, the analysis of cytokine-mediated inflammatory conditions is performed on the bases of genetic association or case-control studies (GWAS) in combination with cytokine or expression production measurements. Frequently used causative indicators of disease occurrence are (i) disease-associated single nucleotide polymorphisms (SNPs) in cytokines and (ii) differentially expressed cytokine levels.
To evaluate the genetic association approaches and altered cytokine levels observed in psoriasis, we examined the degree of genetic association in polymorphisms located in the vicinity of key psoriasis cytokines. We re-analyzed the genetic association data obtained from GWAS for psoriasis [17]. In the Manhattan plot (Figure 2C), associations are highlighted corresponding to SNPs located in the genomic vicinity of a number of genes for key inflammatory cytokines crucial in psoriasis The Figure shows that none of the polymorphisms near the major cytokines IL-22, INF-γ, IL-1, IL-17A and IL-6 reached the significance association levels () determined by the GWAS [17]. Since all these cytokines are shown to participate in the mechanism that mediates psoriasis [15], this result suggests that genotyping experiments do not represent an infallible method for identifying key pathology-associated cytokines.
To further assess the predictive capabilities of genome-wide screens for marker identification in inflammatory conditions, the differences in the IL-10 and IL-22 production levels between psoriatic and healthy skin samples were compared using experimental data from the literature [37]. We found that IL-10 is significantly associated with psoriasis GWAS [17] whereas IL-22 is not (Figure 2C). A comparison of the IL-22 concentration in the normal and psoriatic skin shows a significant elevation of IL-22 levels in disease [37] (Figure 2D), despite of the lack of significant IL-22 SNP in GWAS [17] (Figure 2C). The IL-10 cytokine shows the opposite effect to IL-22, as the SNP observed in the vicinity of the IL-10 gene shows a clear genomic association with psoriasis [17] (Figure 2C). At the same time, the IL-10 production by lymphocytes does not change significantly between cases and controls [37] (Figure 2E). Therefore, IL-10 and IL-22 cytokines are examples to demonstrate that either presence or absence of a SNP in a cytokine does not always translate to modified cytokine levels in pathological tissue. More specifically, the IL-10 cytokine example illustrates the case where a significantly associated SNP found in the cytokine does not result in altered cytokine levels, while the IL-22 example shows difference between cases and controls production levels in the absence of any significance in the genome wide scan.
The above examples suggest that although GWAS and cytokine production/expression comparison allow identification of potential cytokine candidates, they may lead to conflicting conclusions and do not establish a specific cytokine function. Moreover, one can argue that even in situations when both genetic significance and cytokine production/expression differences between cases and controls are present, the mechanisms of molecular interaction between immune cell populations in normal and pathologic interactions cannot be ascertained. It is also unclear how statistically significant differences for cytokines in genotype or expression data of disease and control cases contribute to unbalanced interactions between the immune cell populations. Therefore, the need exists for the development of additional methodologies complementary to genome-wide association studies and expression level comparison that would provide further insights into how the immune system operates.
Genetic or expression level comparison studies are frequently complemented by cytokine concentration profiles, whereby the amounts of various cytokines produced by a specific cell population under normal and diseased conditions are measured by Luminex or Elisa assays. These techniques provide a closer insight into cellular interactions in disease, as individual SNPs or altered cytokine expression levels may not always translate into changes in cytokine production levels. In the previous section we showed that SNPs in cytokine genes may not always result in the modification of cytokine production profile.
In this section we demonstrate that up- or down-regulation of cytokine production levels in disease is due to the interactions between immune cells. Experimental measurements of cytokine production profiles in individual cell populations are usually performed in a physiological “cocktail” of other cytokines. Here we demonstrate that a random choice of the cytokine concentrations in such a physiological cocktail creates grounds for misconceiving the role of a particular cytokine in disease, as illustrated below.
Measurement of a particular cytokine concentration largely depends on the levels of other cytokines also present in the medium. We consider the dose-response curve for IL-17 production in bone marrow derived macrophages as a function of IL-23 concentration shown on Figure 3A, as adopted from [38]. Both IL-17 and IL-23 are major inflammatory cytokines, as identified by linkage analysis and functional characterization in a number of inflammatory conditions [13], [39]–[41]. The data show that IL-17 production has a complex dependence on extracellular IL-23 concentration. For example, the blue dotted line in Figure 3A indicates that for IL-23 concentration of 0.25 ng/ml, IL-17 provides a 120 pg/ml readout, while 10 ng/ml of IL-23 produces ∼180 pg/ml of IL-17. Therefore, variability in the IL-23 concentration within the physiological range is likely to cause significantly a different IL-17 production profile. The dotted green and red lines in Figure 3A indicate how the background concentration of IL-23 in the experimental medium can lead to either “upregulation” (Figure 3B) or “downregulation” effects in IL-17 production in disease even in the absence of any changes in bone marrow derived macrophage cytokine production properties (Figure 3C). This example illustrates that cytokine production profiles in immune cell populations cannot define the disease unambiguously and may lead to misinterpretation of cytokine production differences in control and disease samples (Figure 3).
It is essential to note that the overall cytokine production dependence in tissue combines both the cytokine production by a specific cell population as well as other cytokine-dependent effects, such as proliferation and apoptosis. Regulation through proliferation and apoptosis changes the number of cells in skin and therefore also modulates the dose-response profiles. For example, Figure 3D (adopted from [42]), shows the proliferation-apoptosis cycle of a T cell population with increasing IL-2 concentrations. Larger T cell pools produce greater amounts of cytokines and chemokines, therefore the total amount of cytokine production is by the cell numbers in Il-2 dependent manner.
Cytokine production in a cell population is complemented by a number of mechanisms that counterbalance cytokine production in tissues. Extracellular concentrations of cytokines are affected by diffusion, cleavage by metalloproteases and cytokine binding followed by uptake. The dose-dependence of cytokine B on cytokine A concentration represents a dose-dependent curve of homeostatic balance. It is mediated by immune cell populations and balanced by the cytokine removal mechanisms described above. According to the dose-response curve, any given extracellular concentration of cytokine A in tissue translates to a specific extracellular concentration of cytokine B, under conditions of equilibrium. However, it is also possible that additional cytokine A or B production by other cell populations can also occur in tissue, resulting in cytokine A and B concentrations that do not fit the line of homeostatic equilibrium for the immune cell population considered (Figure 4A). After such perturbation, the immune system returns to homeostasis, defined as the dose-dependent line of cytokine B production in a cytokine A-dependent manner and modulated by the cytokine removal mechanisms. As shown in Figure 4A, there is an infinite number of homeostatic cytokine A and B concentrations that the system can adopt as it returns to equilibrium.
Owing to the infinite combination of cytokine A and B concentrations in homeostasis mentioned above (Figure 4A), at least two interdependent cell populations need to be considered to establish the conditions required for a specific homeostatic equilibrium. One immune cell population produces cytokine B in a cytokine A-dependent manner as previously described (Figure 4B) and the other cell population produces cytokine A in a cytokine B dose-dependent manner, where the cytokine B is chosen to have an inhibitory effect to the second cell population (Figure 4C). Both dose-dependent cytokine production curves represent the lines of homeostatic equilibrium for two “opposite” cell populations. The intersection of the two dose-dependent cytokine production profiles represents the point of synergistic balance, where both cell populations reach a homeostatic equilibrium. Such mutual dependence of cytokine concentration via the immune cell populations creates the classical problem of two interdependent variables that has been extensively studied in life sciences, but insufficiently recognized in immunology to-date. Such system-level effects can be associated with the presence of numerous interdependent cytokine pairs, whereby the interdependence can arise from either direct cell-to-cell interactions or larger number of interacting cell populations. Therefore, understanding of the immune cell interactions is enhanced by studying the experimental data through a quantitative description of cytokine production by cell populations in a cytokine-dependent manner.
Physiologically relevant consideration of two cell populations jointly (Figure 4D), suggests that the intersection of the dose-dependent curves occurs at a specific point, as shown in Figure 4E. This intersection defines the cytokine A and B concentrations unambiguously, as this is the only point where both cell populations reach homeostasis in equilibrium. Therefore, homeostatic cytokine concentrations can be defined as the extracellular cytokine concentrations where the immune system remains in equilibrium in the absence of normal or pathologic inflammatory response. From a systems perspective, the inflammatory response can be defined as the system response to the temporally perturbed shift from equilibrium with the ensuing return to homeostasis.
In order to model the performance of the immune system under normal homeostatic conditions, we analyzed dynamic system responses shown in Figure 5. The phase diagram (Figure 5A) depicts two overlapping cytokine dose-response curves for two cell populations (red, blue curves) intersect at one point (violet circle). The dotted lines represent predicted homeostatic concentrations for cytokines A and B. The red and blue dose-response curves are defined as null clines or lines of equilibrium. Vector fields are also shown to represent the cytokine concentration dynamics at the non-equilibrium levels (Figure 5A).
It is noted that the cytokine dose-dependent relationships shown on Figure 4 are schematic and intended for illustrative purposes only, while Figure 5 describes the predictions of the mathematical model. According to the model, the dependence of cytokine A on cytokine B concentration (blue curve) represents a classical dose-dependent activation of one cytokine by another, while the reverse dependence of the cytokine B production as a function of cytokine A concentration (red line) reveals a significant nonlinearity. Mechanistically, such dependence can occur when the model parameters are set such that the cytokine production is nonlinearly related to the cytokine concentration-dependent uptake (please refer to the Materials and Methods section for the detailed description of the model and the underlying parameter values). At the same time, the highly nonlinear relationship between cytokine A and B concentrations (red curve) corresponds to the experimentally observed IL-17 production as a function of IL-23 concentration (Figure 3A) in bone-derived marrow fibroblasts [38]. The dose-dependent curve of IL-17 production as a function of IL-23 (Figure 3A) is “rotated” by 90° and superimposed on the dose-dependent curve for IL-23 production as a function of IL-17 concentration (Figure 5). While we believe that the proposed framework of immune cell interactions analysis is generic and applicable to various pairs of immune cell populations or pairs of cytokines, we note that cytokines IL-17 and IL-23 are good candidates to showcase the systems model presented in this manuscript.
The quantitative representation of immune cell interactions offers a number of mechanistic insights into the immune system responses, specifically in the activation dynamics in response to external application of cytokine A, applied at the state of homeostatic equilibrium (Figure 5B). Three cytokine dynamic profiles annotated as 1, 2 and 3 show the interconnected cell population responses to the temporal application of external cytokine A in increasing amplitude. In all three cases, both cytokine concentrations increase temporally and converge back into the same point of homeostasis. External perturbations of the highest amplitude that induced response 3 on Figure 5B applied to the system of two interacting cell populations are shown on Figure 5C. The temporal cytokine A and B dynamics in response to the small and temporal external perturbation by other immune cell populations (Figure 5C) or infection is presented on Figure 5D. This graph is a temporal projection of trajectory 3 from Figure 5B and clearly illustrates that a small external perturbation applied for a small duration induces cytokine A and B impulses of significantly higher amplitude and somewhat longer duration. The cytokine concentrations released into the extracellular space dramatically diminish in concentration as cytokine diffuses in all possible directions. Our model predicts that a normal immune system is very sensitive and capable of amplifying very small cytokine impulses followed by a return to the original level of homeostasis.
Inflammation-mediated skin conditions are characterized by chronically high cytokine concentrations maintained over extended periods of time. We employed our mathematical model to explore potential factors that can turn normal immune system responses into pathology.
To explore the ability of the model to predict pathologic immune responses, we varied parameter values (Table 1) of the governing equations in the model without changing the structure of equations used, to ensure that we simulate the same cell populations that could originally produce a normal immune response. The alteration of model parameter values reflects the influence of internal and environmental factors to the immune cell populations.
Figure 6A shows the nullclines that represent the dose-dependent cytokine production rates for two interacting cell populations. Since the underlying equations have not been modified and parameters have only been altered in a minor fashion, the shapes of the dose-dependent cytokine production profiles are similar to the ones predicted for a healthy immune system, shown on Figure 5A. However, slight modifications in the immune cell interaction parameters (that can be caused by genetic polymorphisms, environmental factors or a combination of the two) cause the nullclines to intersect at three different points (marked by violet circles and annotated as H1-H3): two intersections occur in the area of the low cytokine concentrations, and the third is observed at the region of significantly higher cytokine concentrations. Numerical simulations reveal that only two of the three solutions are stable (H1 and H3, shown as filled violet circles on Figure 6A), whereas the intermediate one is unstable (H2 indicated as a hollow violet circle). This suggests that two interacting immune cell populations can create more than one homeostatic level of cytokine concentration in extracellular space.
The described result implies that even minor alterations in cytokine production profiles (presumably initiated by a combination of genetic polymorphisms and environmental factors) can lead to pathological inflammation resulting from modification of feedback loop parameters. The model predicts that causative SNPs that contribute to the alteration of feedback loops via small modifications of cytokine profiles do not need to be identical across all disease phenotypes. The model relates SNPs to pathologic levels of cytokine concentration in tissue and predicts that both statistically significant and insignificant genetic polymorphisms from different immune cell populations can lead to the appearance of additional pathologic cytokine levels and describes how. This offers an explanation of why only some areas of skin can be inflamed in psoriasis while others exhibit symptomless phenotype, while all cells across the whole body carry the same genetic polymorphisms. The systems analysis indicates that genetic polymorphisms can operate in combination with external conditions and either lead to inflammation, or exhibit symptomless phenotype depending on the environmental stimulus. However, one can argue that feedback loop modifications between immune cell populations originates from genetic variants, which do not need to be the same in all disease states and can lead to the emergence of pathology with or without environmental factors.
The directed green lines on Figure 6A represent the vector field and show the dynamic cytokine trajectories that converge into one of the stable homeostatic solutions from any combination of cytokine concentrations. In this case, interacting cell populations can maintain two distinct homeostatic cytokine levels, one in the area with low and with high cytokine concentrations. The systems model predicts that under a certain combination of parameters, interacting cell populations are capable of operating as a switch that can shift between two distinct homeostatic levels of cytokine concentrations. The appearance of additional stable homeostatic solutions suggests that the immune system can remain in a state of elevated cytokine concentrations for a significant period of time. The existence of two stable solutions creates a different scenario than in the case where interactions between immune cells had only one single stable homeostatic solution. More specifically, in healthy immune system temporal elevation of cytokine concentrations are always followed by an imminent return to homeostatic concentrations. In this pathologic scenario, the alterations of cytokine concentrations can cause the immune system to return to either of two stable homeostatic levels; the state of low cytokine concentration or the pathologic one of high cytokine concentrations, where the immune system can remain for a significant duration. The possibility to switch between two stable cytokine concentration levels provides the trigger-like properties to the system of at least two immune cell populations.
Next, we analyzed the dynamics of cytokine alterations at the transition between the two stable homeostatic levels. Figure 6B shows the variations of cytokine as the system switches from the homeostatic point of low cytokine concentrations (H1) to the homeostasis point with high cytokine concentration (point H3, green trajectory 1) and back (green trajectory 2). Under the assumptions underlying the present model, the transition from H1 to H3 occurs upon external impulse of cytokine A (Figure 6C). The cytokine A and B alteration dynamics during the transition from H1 to H3 is shown on Figure 6D. The model predicts that the transition from H3 to H1 can be induced by application of external cytokine B (Figure 6E). The transition from chronically high cytokine concentrations to the low level is shown schematically on Figure 6F. The model predicts that cytokine B is capable of generating a significant spike before shifting to H1. The model predictions address the fundamental question of how lesional and perilesional skin phenotypes can simultaneously coexist in inflammatory condition affected patients. The trigger-like cytokine behavior emerging from the interactions between the cell populations can keep the skin either in the inflamed condition causing a lesion, or remain at the lower cytokine concentration steady-state level observed in perilesional skin samples.
Psoriasis is characterized by a variety of clinical phenotypes. After establishing the mechanism of chronic inflammation in the form of additional stable homeostatic level as described previously, we employed the systems model for immune cell interactions to elucidate whether it can uncover the causes of variety of clinical phenotypes observed in clinical practice. Similarly to the previous case, we tested combinations of parameters within physiological limits without changing the structure of the governing equations.
Under certain combination of parameters (Table 1), a stable solution H3 (Figure 6A) can become unstable (Figure 7A), and form a limit cycle that represents simultaneous oscillatory alterations of both cytokines. Stable oscillations of cytokine concentrations cause unbalanced proliferation and differentiation of keratinocytes, the main cell type constituting dermis and epidermis, and are thus pathologic for skin. At the same time, the oscillatory type of pathology is different from the cytokine-trigger mode described in the previous section. Trigger-like inflammation causes clearly defined areas of lesion, whereas oscillations are more likely to cause a phenotype with gradual transition between inflamed and non-inflamed areas of skin.
Variation of cytokine oscillation-driven pathology is shown on Figure 7B. The chosen combination of model parameters allows only one unstable solution H3 with the limit cycle in the area of high cytokine concentrations. The absence of stable homeostatic solutions leads to the most severe disease phenotypes, which are least susceptible to potential treatment.
In order to analyze the dynamic properties of immune cell interactions in relation to the type of pathology (Figure 7A) when cells either maintain the stable homeostasis or experience stable oscillations, we studied how the system responds to the applications of external cytokine concentrations. The present model predicts that the external cytokine A application can either switch the system from the homeostasis H1 to oscillatory mode around unstable solution H3 (trajectory 1 on Figure 7C) or generate an impulse and the system returns to homeostasis H1 (trajectory 2 on Figure 7C) as in the case of healthy immune reaction. The difference between healthy and pathologic responses is due to the amplitude of applied external perturbation of cytokine A (Figure 8). Small impulses shift the system out of homeostasis H1 into the oscillatory mode. The spike of higher amplitude leads to generation of a sizable response, before returning to homeostasis level H1. Our model predicts that small perturbations of either cytokine A or cytokine B is sufficient returning the system from oscillatory mode to the normal level of homeostasis (Figure 7D and Figure 9).
We propose a new systems biology model that captures crucial properties of immune cell interactions and predicts the conditions under which normal and pathological inflammatory responses are elicited. The model integrates individual characteristics of immune cell populations and allows the definition of homeostasis as specific cytokine concentrations estimated by the intersection of the immune cell population cytokine dose-response curves. The model predictions provide novel insights into the mechanism of elevated levels of inflammatory cytokines in disease [2], [15], [43]. While it is well known that (i) genetic variants change the susceptibility to disease [44] and (ii) the same disease phenotype can be elicited by different types of inflammation [15], the relationship between genetic variants and pathologic inflammation remains unclear. The present study reports a generic framework to explain why and how small alterations to cytokine production profiles (arising from genetic variants which can be different across cases and not always statistically significant) leads to the modification of feedback loop interactions between immune cells and the appearance of pathologic inflammatory levels.
This study suggests that cytokine concentrations can deviate from homeostatic levels even in the absence of any pathology, as long as such deviations are temporal and always return to homeostatic level in equilibrium. Normal immune response initiates temporal increase of key cytokines concentrations for a time span sufficient to execute the effector system and eliminate the cause of inflammatory reaction. Pathology occurs if the inflammatory response is not temporal and cytokine concentrations fail to return to the original levels. According to the model predictions, homeostatic cytokine concentrations can only be estimated from the interactions of interdependent cell populations. Homeostasis is therefore a systems effect and occurs at the crossing of the dose-dependent cytokine productions curves from at least two immune cell populations (Figures 4D and 4E). The analysis of model properties allows unravelling of mechanisms that cause stable chronic inflammation. According to the model, normal immune system can be described as a system with one stable homeostatic level defined by the cytokine feedback loop parameters of immune cell interactions. External perturbations applied to the healthy immune system induce a temporal cytokine concentration increase, followed by a return to the stable homeostasis (Figure 5).
Alterations in the feedback loop parameters [18]–[23], [45] can turn the immune system pathologic by inducing bistable behavior with discrete steady-states or loss of stability in homeostasis. The present study follows earlier modeling analyses of different types of inflammation [28]–[29], [32], [42], [46]–[53]. Similarly to previous studies, the framework reported here predicts that inflammatory response is a highly dynamic process that can be represented mathematically by incorporating experimentally derived feedback loop interactions between immune cell interactions. The presented model proposes new generic principles that can distinguish healthy and pathologic inflammation. Moreover, it offers a rational foundation to establish the relationship between causative genetic variants, alterations in the cytokine production profiles and modifications in the feedback loop interactions between immune cells, ultimately leading to the appearance of inflammatory pathology. The model also possesses a predictive capacity to distinguish between different types of inflammation that can arise from the same immune system. Overall, the application of systems modeling theory to simulate the immune cell regulation effects in psoriasis through altered properties of feedback loops can outline the key factors that distinguish normal immune system response from pathology.
The quantitative model for immune cell interactions in this study offers a mechanistic distinction between healthy inflammatory reaction and pathological inflammation. Internal and environmental factors can alter cellular interactions in the form of modified cytokine production curves. In order to investigate how such alterations can translate into various pathologies, the derived model was subjected to exhaustive evaluation of the underlying parameters of cytokine production and degradation rates without any modifications in the model structure. Such an assumption reflects the physiological situation where the interacting immune cell population pairs remain the same, but the parameters of the interactions can vary due to genetic mutations. The model predicts that the autoimmune mediated pathology occurs in those cases where the modified feedback interactions between immune cells lead to the appearance of additional levels of homeostasis (Figure 6). As a result, the immune system can start operating as a cytokine trigger and maintains either low or high cytokine concentrations levels. A different type of pathology can occur when the alterations of the cytokine-mediated feedback interactions between immune cells lead to the loss of stability of the homeostatic level. Variability in the interactions between immune cell populations can result in the appearance of oscillations (Figures 7, 8 and 9). The model therefore predicts that the same immune cell populations are not capable of mediating a normal immune reaction or operate as a biological trigger, instead, the immune system undergoes periodic temporal alterations. Stable oscillations of cytokine levels are also pathologic. The oscillations-based type of pathology is different from the trigger-type immune system pathology.
The healthy homeostatic and pathologic model predictions have been obtained through exhaustive screening of possible parameter values. The summary of representative sets of parameters chosen approximately in the middle range of the corresponding dynamic behavior is found in Table 1. While the listed parameter values may not be the only possible combinations of healthy and pathologic immune cell interactions for the described scenarios, they cover all possible types of dynamic behavior that the present model can achieve. One can choose different combinations of constants for the model so that it would oscillate or operate as a trigger, however there are no possible combinations of parameters where three or more stable homeostatic levels can exist, as it has been shown for example in multisite phosphorylation systems [54].
The combinations of parameter values are closely related to the model application on actual cytokines and, as noted earlier, two potential candidates for the proposed model are IL-17 and IL-23. Other cytokines that have been shown to be essential in skin inflammation include IL-22, oncostatin M, TNF-α, IL-1α [55], IL-6, IL-12, interferon-α and interferon-γ [15]. The difficulty of analyzing real cytokines rather than the immune cell interactions via hypothetical cytokines can be attributed to the fact that the majority of experimental investigations report static comparisons between experimental groups without considering either dose-dependent curves or dynamic information. While such comparisons are important, this model suggests that they may be insufficient for deeper understanding of mechanisms in inflammation. Further experimental investigations directed toward the dose-dependent cytokine production profiles would be required for estimation of the model parameters. It is essential to note that parameter values will be different in individual immune cell population pairs in a given tissue and that pathologic parameter alterations will depend on the combination of the causative genetic mutations found in specific cytokines.
Figure 10 summarizes the model-based description of normal and pathologic immune system performance in human skin. Under normal conditions, cytokine production mediated interactions between immune cells lead to one stable homeostatic level in tissue (Figure 10A). Combinations of the internal and external factors can change the interactions between immune cells, in such a way that additional stable or unstable homeostasis levels appear. Chronically increased cytokine concentrations are more likely to be observed in clearly defined inflamed lesions (Figure 10B). In those cases, the immune system is able to switch and remain at the elevated cytokine concentration state. Oscillating cytokine concentrations are likely to cause a different inflammation phenotype with diffused borders between inflamed and non-inflamed regions (Figure 10C). The model predicts that the immune system's ability to mediate either normal or pathologic inflammatory responses is a systemic effect which emerges from the imbalance of immune cell interactions, rather than an attributed feature of a favorite cell population or a genetic polymorphism.
According to the proposed model, pathology occurs as a result of one or a combination of SNPs in cytokine or any other genes with the net effect of altered type of homeostatic level via the modification of parameters in the feedback loop interactions between immune cells. The description of the homeostatic mechanism from the systems perspective explains why SNPs in some cytokines (e.g. in IL-22), can have very low statistical association with psoriasis, but can contribute to pathology in a number of cases (Figure 2). The proposed mechanistic description of inflammation suggests that different combinations of SNPs (some or all of which can have very low association with the disease) can cause similar cytokine production curve alterations.
The proposed quantitative model for immune system explains how normal and pathologic inflammatory immune reactions can be mediated by the same immune cell populations. Current research in immuno-genetics mainly focuses on the search of polymorphisms highlighting candidate genes responsible for pathological inflammation. This work proposes that the altered feedback loop parameters (potentially arising from genetic polymorphisms) in the interactions between immune cell populations participate in the maintenance of inflamed lesions. The system model predictions for the possible coexistence of multiple homeostatic levels explains how inflammatory disease affected individuals can simultaneously have both non-inflamed and inflamed areas of skin while carrying the same genotype with disease-associated SNPs. The proposed approach, therefore, offers a mechanistic explanation for why “causative” SNPs mediate inflammatory lesions at some regions of skin while they do not do so at others.
The systems model described in this work is relatively generic and applicable to analysis of a range of inflammatory conditions. The mathematical model allows the prediction of mechanisms in inflammatory disease and the formulation of requirements for therapeutic interventions. The model-guided screening of therapeutic agents can be performed on the bases of eliminating the second possible level of stable or unstable homeostasis or lowering existing cytokine concentrations.
The model describes different pathology phenotypes which are due to the appearance of additional stable or unstable levels of homeostasis, the loss of stability of the basal level of homeostasis or due to the shift of homeostasis to the levels of higher cytokine concentrations. The last case is probably the most frequent and “simple” scenario of inflammatory pathology that occurs when the cytokine production curves intersect at higher cytokine concentration levels. For example, different homeostatic concentrations of a cytokine A shown on Figure 11A occur as a result of altered cytokine production profiles by immune cell populations. According to the proposed methodology, the search for pharmaceutical interventions can be based on identifications of direct or indirect way to restore the original dose-response profile of immune cell population. The interdependence between cytokines via an immune cell population can be utilized by indirect target identification strategies for novel interventions, by using already available therapeutic agents. One possibility is the injections of a cytokine B know to reduce the levels of a different cytokine A (Figure 11B).
The proposed mathematical approach offers new exciting therapeutic opportunities for various inflammatory conditions. One interesting example where the ideas proposed in this study have already been utilized in a similar fashion is the type II diabetes. There are two ways of estimating insulin resistance in diabetes: the glucose clamp test [56] and the homeostasis model assessment [26]. The homeostasis model assessment system seeks values of resistance to the hypoglycemic effect of insulin and β-cell function from the measures of plasma insulin versus glucose, in comparison with a standard group of healthy young adults. The homeostasis model assessment approach takes into the consideration the interactions between glucose and insulin via the specialized cell populations (Figure 11C) and thereby increases diagnostic power in diabetes [26]. We propose that the model for quantitative inflammation assessment will offer advanced diagnostic tools for inflammatory conditions, as opposed to diagnostic methods based on readouts of a single biomarker (Figure 11D). The model suggests the necessary criterion for the properties of required treatment. A large number of currently available pharmaceutical agents offer a temporal relieve from the inflammation induced symptoms. In the context of systems representation of the disease, the drug action can be viewed as a temporal switch from higher to lower cytokine concentration steady state (Figure 12A). Any physiological alterations are likely to switch the system back into the level of pathological inflammation. The major criterion for the new treatments would require them to eliminate the second pathologic steady-state level (Figure 12B).
Systems modeling of inflammatory responses initiated by interdependent immune cell populations can offer new avenues for inflammatory disease-associated data interpretation. In a vast number of cases, the comparison between cytokine production or expression levels is performed statically, by calculating the medians between readouts obtained from cases and controls (Figure 13A). Such representation does not capture the regulatory alterations in cytokine expression or production during either normal or pathological events. As a result, there can be a significant variability in the experimental readouts. However, if data are viewed from a systems perspective, the possibility of dynamic alterations would explain the observed variability in both cases (Figure 13B) and controls (Figure 13C). The combination of the immune cell interactions in the form of a dynamic model with the measured cytokine production or expression levels in heath and disease can offer more explanation for the experimentally observed data points.
One of the major difficulties in research of inflammatory pathologies is the lack of unambiguous definition of disease. The systems model suggests that inflammatory skin disease is unlikely to be mediated by one gene or by a specific cell population. Instead, local inflammation of the immune system in the skin arises from systems-level effects emerging from the interactions between cell populations via cytokines, chemokines and cell surface expressed ligands. Interdependent cytokine production by cell populations creates a network of immune cells with a number of emergent properties such as integration of signals across the immune system, generation of distinct outputs depending on combinations of internal and external conditions. Of particular interest is the immune system ability to form discreet steady-states and switch between them. This study analyzes the effects arising from the interaction of two cell populations only. While the mathematical model covers the range of large number of possibilities, one needs to acknowledge that more sophisticated effects can arise from larger number of interacting cell populations via cytokines. Current study does not include a specific dose-dependent or time course data for cytokine dynamics. While inflammation-associated pathologies are likely to develop according the described principles, a specific subset of cytokines and immune cell populations is needed to be identified for each specific inflammatory condition.
The high quality genotypes for 438,670 markers of the 1359 psoriasis cases and 1400 controls from the genome-wide association scan performed by the Collaborative Association Study of Psoriasis [17], were used for association analysis. The dataset used for the analyses described in this manuscript were obtained from the database of Genotypes and Phenotypes (dbGaP) found at http://www.ncbi.nlm.nih.gov/gap through dbGaP accession number phs000019.v1.p1. Single marker case – control association analysis was performed by executing the –assoc option of the PLINK package (v1.06) developed by Shaun Purcell (http://pngu.mgh.harvard.edu/purcell/plink/) [57]. This option calculates the statistical significance as measured in odd ratios, P or χ2 values of the minor allele frequency differences between psoriatic cases and healthy controls.
Genome-wide association of each SNP is showed in a Manhattan plot as the −log10 (P) dependence on the genomic location using the coordinates of the NCBI Build 36.1 (March 2006). The association of the SNPs located within the 2 Mbp window centered at the selected inflammatory cytokines is shown in color for individual cytokines (Figure 2C).
Figure 4 provides a schematic framework of the two interacting cell populations. We investigate the interactions between immune cell population and the interaction-dependent properties of the immune system in homeostasis through a mathematical model that captures the extracellular cytokine concentrations. All possibilities of immune cell interactions are cdescribed in Supplementary Text S1. Given that cytokine production by immune cell populations can be represented as a function of another cytokine in a dose-dependent manner, inflammation can be defined quantitatively by considering cytokines as interdependent variables, where the specific inter-dependence of cytokines can be established experimentally through studying immune cell populations. The interdependence of cytokines A and B can be represented by a system of coupled ordinary differential equations:(1)The same principle can be applied to larger numbers of cytokines and chemokines produced by immune cell populations:(2)where n is the total number of considered cytokines.
In order to elucidate what distinguishes normal and pathological immune system performance, two cytokines interconnected via dose-dependent effects of corresponding cell populations are considered. Effects that occur in the multidimensional space of cell interactions via cytokines can be projected to two dimensions and we show below that alterations in an immune sub-system with two interacting cell immune cell populations have the potential to describe several different inflammatory phenotypes. We develop a systems model for cytokine, chemokine and surface ligand-mediated immune cell interactions that can unravel the mechanism of inflammation and provide mechanistic explanation for the inflammation in human skin. The model contains two cell populations interconnected via activatory and inhibitory cytokine production. The dose-dependent cytokine production is complemented by cytokine removal via diffusion, cleavage by metalloproteases and trapping mechanisms.
In the most general case, the speed of cytokine concentration dynamics in tissue can be represented as follows:(3)where is the cytokine concentration, is an elementary tissue volume, is a surface area of a cell that produces a cytokine, is the number of cells that produce a cytokine in volume , is the rate of cytokine production , is the rate of cytokine uptake via endocytosis, , is a surface area of cells that express the cytokine receptor, is the number of cells capable of endocytosis of the cytokine receptor in volume , is the capillary surface area in the volume V, is the capillary permeability to the cytokine, , is the cytokine concentration in blood, is the maximum cytokine degradation rate by proteases , is the concentration of proteases, is the rate constant , is the basal cytokine secretion rate by an immune cell population . is the Michaelis constant.
In order to develop the mathematical model capturing the interactions in immune cells via cytokines we defined a number of following immune cell subpopulation groups according to the classification shown on Figure 4: i) cells produce cytokine B in a dose-dependent manner from cytokine A (Figure 4B), ii) the production of cytokine A is inhibited by a cytokine B (Figure 4C). One can also consider a variety of other cases of the bell-shape or reverse bell-shape dependence on cytokine concentration or even more complex cases. We analyze the cytokine system properties under the framework of outlined assumptions and any specific cytokine-dependent cytokine production profiles can be developed as an extension of the model described below.
The rate of cytokine A production by a cell population when it interacts with another population that produces inhibitory cytokine B is given by:(4)where is a normalization coefficient, is the cytokine A concentration, is the dissociation constant for the cytokine A interaction with the cytokine A receptor, is the cytokine B concentration, is the dissociation constant for the cytokine B interaction with the cytokine B receptor.
The rate of the cytokine B production by a cell population when it interacts with another population that produces activatory cytokine A is given by:(5)where () is a normalization coefficient.
Cytokine production by a given cell populations can be modulated by several activatory or inhibitory cytokines. In this general case the cytokine production is given by:(6)where is a normalization coefficient, and are the numbers of activatory and inhibitory cytokines, respectively. is the concentration of the th activatory cytokine and is the concentration of the th inhibitory cytokine. is the dissociation constant for the cytokine interaction with the cytokine receptor. is the dissociation constant for the cytokine interaction with the cytokine receptor.
Cytokine production is complemented by mechanisms of cytokine elimination. Various routes of cytokine removal from extracellular space include cleavage by metalloproteases, diffusion, cytokine trapping, binding to the cytokine receptor and uptake. Cytokine removal by diffusion and cleavage by metalloproteases are nonspecific and do not play an active role in the regulation of the extracellular cytokine concentrations, whereas the cytokine binding to the receptor followed by either release or uptake can have significant implications on the cytokine concentration dynamics. Thus, we next develop governing equations for the cytokine-cytokine receptor interactions.
Cytokine binding to the receptor initiates intracellular signaling events. Under the conditions of dynamic equilibrium, in the absence of endocytosis, the number of cytokines bound to the soluble receptors would equal to the number of cytokines released. However, due to the cytokine-cytokine receptor complexes uptake certain amount of cytokine is internalized via endocytosis mechanism and degraded. The cytokine uptake decreases the cytokine concentration in the extracellular space in the cytokine concentration-dependent manner. The rate of cytokine uptake by a cell population is proportional to the number of receptors bound to the cytokine, multiplied by the total number of receptors on the cumulative cell surface:(7)where () is a normalization coefficient, is the concentration of cytokine A, is the dissociation constant for the cytokine interaction with the cytokine receptor. is the cell surface area, is the number of receptors expressed on a cell surface.
The total number of receptors can be divided into two fractions: receptors that are present on the surface and the subpopulation in the vesicles after the uptake event took place. In steady-state, the rate of receptor synthesis equals to the rate of receptor degradation by proteosomes; these rates are not considered in the present analysis. The conservation law applied to the two receptor populations at any given time point is given by:(8)where is the total number of receptors to a specific cytokine , is the number of receptors on the cell surface, is the number of internalized receptors.
The dynamics of the receptors present on the cell surface is given by:(9)where and are the coefficients that describe the rate of cytokine bound and cytokine free receptor internalization, respectively. reflects the rate of receptor recovery from proteosomes, is the dissociation constant for the cytokine interaction with the cytokine receptor, is the number of receptors on the cell surface, is the number of internalized receptors.
In the steady-state, the number of receptors on the cell surface as a function of extracellular cytokine concentration is given by:(10)The combination of equations (10) and (7) allows obtaining the rate of cytokine uptake as a function of cytokine concentration:(11)The full model for the immune cell interactions is given by:(12)
The relationship between the parameters in the normalized system of differential equations with the original description for the cytokine production and uptake rates is thus given by:The final system of differential equations for two interacting cell populations which was solved numerically to generate all the results presented in the paper is thus given by:(13)All parameter values used in the above equations are given in Table 1.
|
10.1371/journal.pntd.0001958 | Cruzipain Promotes Trypanosoma cruzi Adhesion to Rhodnius prolixus Midgut | Trypanosoma cruzi is the etiological agent of Chagas' disease. Cysteine peptidases are relevant to several aspects of the T. cruzi life cycle and are implicated in parasite-mammalian host relationships. However, little is known about the factors that contribute to the parasite-insect host interaction.
Here, we have investigated whether cruzipain could be involved in the interaction of T. cruzi with the invertebrate host. We analyzed the effect of treatment of T. cruzi epimastigotes with anti-cruzipain antibodies or with a panel of cysteine peptidase inhibitors (cystatin, antipain, E-64, leupeptin, iodocetamide or CA-074-OMe) on parasite adhesion to Rhodnius prolixus posterior midgut ex vivo. All treatments, with the exception of CA074-OMe, significantly decreased parasite adhesion to R. prolixus midgut. Cystatin presented a dose-dependent reduction on the adhesion. Comparison of the adhesion rate among several T. cruzi isolates revealed that the G isolate, which naturally possesses low levels of active cruzipain, adhered to a lesser extent in comparison to Dm28c, Y and CL Brener isolates. Transgenic epimastigotes overexpressing an endogenous cruzipain inhibitor (pCHAG), chagasin, and that have reduced levels of active cruzipain adhered to the insect gut 73% less than the wild-type parasites. The adhesion of pCHAG parasites was partially restored by the addition of exogenous cruzipain. In vivo colonization experiments revealed low levels of pCHAG parasites in comparison to wild-type. Parasites isolated after passage in the insect presented a drastic enhancement in the expression of surface cruzipain.
These data highlight, for the first time, that cruzipain contributes to the interaction of T. cruzi with the insect host.
| Chagas' disease, a neglected tropical disease caused by Trypanosoma cruzi, is transmitted to vertebrate hosts by hematophagous insects. Cruzipain is a lysosomal cysteine peptidase, which plays an important role in parasite infectivity, intracellular growth and differentiation, and is abundantly expressed on the surface of epimastigotes. Since these forms face the insect vector environment during the life cycle, it is conceivable that cruzipain may participate in the interaction process with the invertebrate host. Here, we showed that adhesion of T. cruzi to the insect midgut cells was inhibited by the blockage of cruzipain function. Cysteine peptidase inhibitors, in a dose-dependent manner, and anti-cruzipain antibodies were able to reduce the binding of epimastigote forms to the Rhodnius prolixus midgut. Similarly, T. cruzi transfectants that overexpress chagasin, the endogenous cruzipain inhibitor, displayed low levels of adhesion. Accordingly, the supplementation of exogenous cruzipain partially restored the adherence of the transfected line. Additionally, the ability of the chagasin overexpressing transfectants to colonize the insect in vivo was drastically reduced, and the levels of cruzipain expression by wild-type parasites were enhanced after in vivo passage in R. prolixus. Collectively, our results strongly suggest that cruzipain is required for successful colonization of R. prolixus by T. cruzi.
| Chagas' disease remains one of the most important neglected diseases of Latin America, and it has become a world health problem due to migration of infected people from endemic countries [1]. Every year, it is estimated that 14,000 people die as consequence of the infection in Latin America [2]. The etiological agent Trypanosoma cruzi is transmitted in nature to vertebrate hosts through hematophagous insects from the Reduviidae family. During their development within insects, the parasites undergo profound morphological changes, modulating surface molecules to enable interactions with specific insect tissues that are essential for their survival, development and successful transmission to the vertebrate host. T. cruzi-insect vector interaction begins when the insect feeds on blood of an infected vertebrate host. Once ingested, most of the bloodstream trypomastigotes differentiate into a mammalian non-infective epimastigote forms. In the posterior midgut, they divide repeatedly by binary fission and adhere to perimicrovillar membranes (PMM) of the intestinal cells. In the rectum, where the highest parasite population occurs, a proportion of epimastigotes attach to the rectal cuticle by hydrophobic interactions and transforms into non-replicative metacyclic trypomastigotes (metacyclogenesis), which are eliminated with the feces and urine during blood feeding, infecting another mammalian host [3]. In this scenario, the PMMs act as an adhesion site, becoming essential to the establishment of the parasite in the insect vector. In addition, T. cruzi-PMM interaction appears to be necessary for metacyclogenesis, but there is a general lack of information regarding which parasite molecules are implicated in this process [3], [4].
Cruzipain, a member of the papain superfamily, is a cysteine peptidase of T. cruzi that is an important virulence factor of this parasite, which is involved in several crucial steps in the interaction with mammalian cells, such as in the host cell invasion, and parasite survival, differentiation and multiplication within host cell [4]–[12]. The involvement of cruzipain in the metacyclogenesis process has been indirectly demonstrated by several approaches [4], [11], [12]. The participation of cruzipain in host cell invasion by trypomastigotes is mediated through at least two distinct pathways [8], [9]. One pathway involves the triggering of the B2 type of bradykinin receptor (B2R), whereas the other pathway is independent of the kinin receptors [8], [9]. More recently, it was uncovered that cruzipain also participates in the mobilization of endothelin receptors during the invasion of smooth muscle [13]. Also, cruzipain can cleave at the hinge of all human IgG subclasses, which might be relevant to parasite escape from the adaptive immune response [14]. The drug candidate, N-methyl-piperazine-Phe-homoPhe-vinyl sulphone phenyl (K777), a potent cruzipain inhibitor, is in late preclinical trials for Chagas' disease chemotherapy. This drug rescued mice from a lethal infection of T. cruzi, promoting parasitological cure in most of them, even in an immunodeficient mouse model [15].
Cruzipain is expressed at variable levels in all developmental forms and strains of the parasite, being abundantly detected in epimastigote forms. This enzyme is found in the endosomal-lysosomal system of epimastigote (especially in reservosomes), amastigote and trypomastigote forms, and is profusely detected on the surface of epimastigotes and amastigotes [5]. Some isoforms are associated to the plasma membrane of epimastigotes, presumably through a glycosylphosphatidylinositol (GPI) anchor [16]. There is a common notion that cruzipain, as the major lysosomal peptidase of T. cruzi, may play a prominent role in nutrition of the parasite, at least in the gut of the hematophagous insect vector, however, up to now, no experimental evidence supports such concept. In T. cruzi, a potent endogenous cruzipain inhibitor, chagasin, forms tight binding complexes with the enzyme in vivo, regulating its activity. Parasites lines transfected with an episomal plasmid containing the chagasin gene express four-fold more chagasin than wild-type parasites and exhibit 70–80% reduction in the overall cysteine peptidase activity [10]. These transfectants provide an interesting tool to asses cruzipain function. For instance, it was shown that those lines have reduced capacity to differentiate into trypomastigotes, as well as reduced infectivity to mammalian cells in vitro [10].
In the present study, we sought to investigate whether T. cruzi cruzipain might be involved in the interaction of epimastigotes with R. prolixus midgut. For this purpose, we analyzed the effects of anti-cruzipain antibodies, as well as, of a panel of cysteine peptidase inhibitors on the parasite adhesion to R. prolixus posterior midgut ex vivo. We also compared the adhesion rate to the insects among chagasin transfectants and wild-type parasites, both ex vivo or in vivo. Our findings point to a prominent role for cruzipain in the interaction of T. cruzi with the invertebrate host.
Trypanosoma cruzi, Dm28c (COLPROT 010), G (COLPROT 216), Y (COLPROT 106) and CL Brener (COLPROT 005) isolates, obtained from the Coleção de Protozoários da Fundação Oswaldo Cruz (COLPROT-FIOCRUZ), were used in this work. The transgenic lines were obtained using the episomal pTEX shuttle vector containing the chagasin-encoding gene (pCHAG). A parasite line harboring the empty vector (pTEX) was also used in parallel for control [10]. The epimastigote forms of T. cruzi were grown in 3.7% brain heart infusion medium (BHI), containing 0.002% hemin, supplemented with 10% heat-inactivated fetal bovine serum (FBS), at 28°C for 4 days, to reach late-log growth phase. The transgenic parasites were maintained in BHI, supplemented with 800 µg/mL geneticin. For all experiments, epimastigotes were harvested by centrifugation (1500× g for 5 min at 25°C), washed three times in 0.15 M NaCl, 0.01 M phosphate-buffer pH 7.2 (PBS) and immediately used.
Rhodnius prolixus were reared and maintained as previously described [17]. Briefly, fifth-instars larvae were randomly chosen, starved for 30 days after the last ecdysis and then allowed to feed on defibrinated rabbit blood through a membrane feeder. Ten days after feeding, the insects were dissected; the posterior midguts removed, longitudinally sectioned and washed three times in PBS to expose their luminal surfaces [18]. After the washing, the tissue fragments were processed as described below. The insects were obtained from the insectary of the Laboratório de Bioquímica e Fisiologia de Insetos, Instituto Oswaldo Cruz, FIOCRUZ.
For the interaction assays, the tissue fragments were placed into Eppendorf microtubes and then, incubated with the parasites (2.0×107 in 100 µL of PBS) for 15 min at room temperature, under gentle shaking. Only one dissected midgut was added to parasites per treatment. Wild-type Dm28c parasites were pre-treated or not for 1 h with a panel of different cysteine peptidase inhibitors: iodoacetamide, leupeptin, antipain, CA-074-OMe [L-3-trans-(propylcarbamoyl)oxirane-2-carbonyl]-L-isoleucyl-L-proline methyl ester or E-64 [(trans-(epoxy-succinyl)-L-leucylamino-(4-guanidino)butane] at 10 µM, or chicken egg white cystatin at 1 µg/mL. For the dose-dependent assay, we used cystatin concentrations ranging from 0.1 to 10 µg/µL. Wild-type parasites were treated with anti-cruzipain antibodies at 1∶1000 or 1∶2500 dilution or with rabbit pre-immune serum (1∶1000) for 1 h. The viability of the parasites throughout the experiment was assessed by mobility and trypan blue dye exclusion. After each treatment, parasites were washed three times with PBS prior to the interaction assays. Alternatively, the adhesion rate of several T. cruzi isolates was compared: wild-type parasites, pTEX and pCHAG (Dm28c), as well as G, Y and CL Brener. After incubation with the parasites, the explanted midguts were spread onto glass slides and the numbers of attached parasites per 100 randomly chosen epithelial cells in 10 different fields of each midgut explanted were quantified by counting under the light microscope [18]. Results are shown as the mean ± standard error of two experiments performed in quadruplicate.
In this set of experiments, tissue fragments were pre-treated for 15 min at room temperature with 1.5 µg/µL of exogenous active cruzipain obtained from T. cruzi Dm28c epimastigotes as previously described [19] or heat-inactivated cruzipain. After this treatment, the midguts were gently washed in PBS and the interaction with Dm28c epimastigotes was performed as described before. Alternatively, pCHAG parasites were incubated for 15 min at room temperature with R. prolixus dissected midguts in the absence or presence of increasing concentrations of exogenous active cruzipain ranging from 1.875 to 7.5 µg/µL. The interaction process was carried out as described above.
After a starvation period of 30 days, fifth-instars larvae were fed through a membrane feeder on defibrinated rabbit blood containing 9×106 T. cruzi cells/mL (wild-type, pTEX or pCHAG). Twenty days after infection, the entire posterior midgut of 4 insects or the entire rectum of 8 insects were obtained, pooled and gently homogenized in 1 mL of PBS. Then, two aliquots from the homogenate were used to quantify in a hemocytometer chamber the total number of live flagellates. The same tissue preparations were also used for parasite quantification through real-time PCR assays, as described below.
A pool of 4 midguts or a pool of 8 recta were resuspended in 100 µL of 10 mM Tris-HCl, 1 mM ethylenediaminetetraacetic acid (EDTA) buffer, pH 8.0 (TE buffer) containing 100 µg/mL proteinase K and incubated for 2 h at 56°C. Then, total DNA was extracted using Wizard Genomic DNA Purification Kit (Promega) according to the manufacturer's instructions, with slight modifications. Briefly, after the treatment with lysis buffer, the samples were centrifuged at 2,000× g for 5 min, the collected supernatant was incubated at 56°C for 2 min to proceed the following steps. At the final step, the DNA was eluted with 200 µL of ultrapure water, and incubated at 25°C for 10 min. The purity (A260/280 nm ratio) and the concentration of DNA were estimated by spectrophotometry using a NanoDrop (Thermo Scientific). After that, absolute quantification of T. cruzi in each sample was performed through real-time quantitative PCR in a thermocycler ABI Prism 7500 Fast Sequence Detection System (Applied Biosystems, Foster City, CA, USA). The quantification was performed in a final volume of 20 µL containing: 2 µL DNA, 10 µL 2× Power SYBR Green master mix (Applied Biosystems, CA), 0.3 µM primers for the T. cruzi satellite DNA region [20] or 0.1 µM primers for the R. prolixus 12S ribosomal RNA gene. Primers used for T. cruzi and insect DNA sequences were, respectively: Cruzi 1 (Forward) 5′-ASTCGGCTGATCGTTTTCGA-3′, Cruzi 2 (Reverse) 5′-AATTCCTCCAAGCAGCGGATA-3′, P2b (Forward) 5′-AAAGAATTTGGCGGTAATTTAGTCT-3′ and P6 (Reverse) 5′-GCTGCACCTTGACCTGACATT-3′. The PCR conditions were: 50°C for 2 min, 95°C for 10 min followed by 40 cycles at 95°C for 30 sec and 58°C for 1 min. Parasites were quantified using the absolute quantification method, and samples were normalized to the R. prolixus 12S ribosomal RNA gene. The standard curves were prepared from parasite DNA serially 10-fold diluted in TE buffer.
For parasite re-isolation, a pool of 8 recta from infected insects was resuspended in 1 mL of PBS and centrifuged at 300× g for 1 min. The supernatant was discharged and the pellet was incubated with 500 µL of PBS for 30 min at 28°C. The supernatant was collected and centrifuged again at 1500× g for 5 min. The pellet contained T. cruzi cells and, in minor quantity, small fragments of the insect rectum. Finally, re-isolated cells from the rectum (at least 106 cells) were fixed in 0.1% paraformaldehyde in PBS (pH 7.2) for 30 min at 26°C, followed by extensive washing in the same buffer. The fixed cells maintained their morphological integrity, as verified by optical microscopic observation. After this step, the cells were incubated for 1 h at room temperature with a 1∶1000 dilution of the anti-cruzipain antibody. Cells were then incubated for an additional hour with a 1∶100 dilution of fluorescein isothiocyanate (FITC)-labeled goat anti-rabbit IgG. The cells were then washed 3 times in PBS and the parasite associated fluorescence was quantified in a flow cytometer (FACSCalibur, BD Bioscience, USA) equipped with a 15 mW argon laser emitting at 488 nm. Non-treated parasite cells from culture or from insect rectum and cells treated with the secondary antibody alone were run in parallel for control. Each experimental population was then mapped by using a two-parameter histogram of forward-angle light scatter versus side scatter. The mapped population (n = 10,000) was then analyzed for log green fluorescence by using a single parameter histogram [21].
All ex vivo experiments were repeated two times in quadruplicate. All in vivo infection assays, including qPCR assays, were performed as three independent experiments in triplicate. The data was analyzed statistically by means of Student's t test, or the analysis of variance between groups was performed by means of ANOVA test using EPI–INFO 6.04 (Database and Statistics Program for Public Health) computer software. P values of 0.05 or less were considered statistically significant.
In order to evaluate whether cysteine peptidase inhibitors influence the adhesion of T. cruzi to dissected R. prolixus posterior midgut, we performed experiments in which each cysteine peptidase inhibitors were incubated with epimastigote forms, followed by their exposure to dissected R. prolixus posterior midgut. The parasites maintained their viability under this condition, as judged by their mobility and trypan blue dye exclusion. After this time, untreated T. cruzi epimastigotes were allowed to bind to dissected R. prolixus posterior midgut, revealing many parasites adhered to the insect epithelial cells mainly by their flagella. Our results showed that iodoacetamide, leupeptin, antipain, E-64 and cystatin significantly reduced, on average 70%, the adhesion of T. cruzi to R. prolixus posterior midgut in relation to untreated parasites (Fig. 1). Considering the presence of a significant activity of a 30-kDa cathepsin B like-cysteine peptidase in T. cruzi extracts [22], we analyzed the effect of CA-074-OMe, a specific inhibitor of cathepsin B. Our results revealed that this inhibitor showed no significant change in the interaction process (Fig. 1).
Cystatin is a high affinity tight binding inhibitor of cathepsin L cysteine peptidases, such as cruzipain. Therefore, we evaluated the effect of increasing concentrations of cystatin on the proportion of parasite adhesion to midgut. In cystatin doses ranging from 0.1 to 10 µg/mL, the adhesion of epimastigotes diminished from 73% to 15% in relation to the control (Fig. 2). Supporting the hypothesis that parasite cysteine peptidases play a role in the adhesion to the insect, the pre-treatment of parasites with anti-cruzipain antibodies considerably reduced the interaction process, in relation to the control (Fig. 3). The antibody concentrations used did not promote parasite agglutination (data not shown). Parasites treated with the pre-immune serum adhered to the midguts at a rate similar to that of the control (Fig. 3).
Transgenic parasite lines overexpressing chagasin present a four-fold increase in cysteine peptidase inhibitory activity and reduced levels of active cruzipain, posing as a tool to address the role of this parasite peptidase in the interaction with the insect vector [10]. Chagasin overexpressing line (pCHAG) displayed low rates (73% lower than control) of adhesion to insect dissected midguts ex vivo. Parasites carrying empty vector (pTEX) were used as controls and did not show any significant alteration in the adhesion rate in relation to wild-type parasites (Fig. 4). In order to assess if the diminished capacity of pCHAG to adhere was related to reduced cruzipain activity, we added exogenous cruzipain to the interaction media. In this condition, the adhesion of pCHAG raised systematically from 27% (no supplementation) up to 60% in relation to the control, as function of exogenous cruzipain concentration supplemented to the assay (Fig. 5).
It is known that distinct T. cruzi isolates have natural differences in the stoichiometric balance of cruzipain:chagasin [10], which impact on the overall cysteine peptidase activity of the parasite. In this sense, the adhesion rate to R. prolixus dissected midguts was compared, revealing a correlation with the overall cysteine peptidase activity. Our results demonstrated that the G isolate, which presents ten times less cruzipain activity than Dm28c [10], presented the lower capability to adhere to the insect luminal midgut surface of R. prolixus in relation to the other isolates (Fig. 6).
In order to assess if cruzipain could act as a direct ligand for possible receptors in the insect epithelial midgut cells, as previously described for the gp63 (a metallopeptidase) from a lower trypanosomatid in the adhesion to an insect host model [23], we tested the effect of the pre-treatment of dissected midguts with cruzipain molecules. Our results showed that the pre-incubation of either active or heat-inactivated cruzipain did not promote any significant alteration in the interaction rate (data not shown).
In order to compare the infection levels in vivo of epimastigote forms of T. cruzi (wild-type, pTEX and pCHAG), insects were fed with defibrinated rabbit blood containing parasites. Twenty days after blood feeding, the insect midguts and recta were screened for parasites by direct microscopic counting. Parasites were only detected in the rectum. As expected, chagasin-transfectants (pCHAG) displayed low rates of colonization in comparison to both wild-type and pTEX parasites (Fig. 7A). Although the possible participation of cruzipain on the colonization process was demonstrated through this approach, the absence of parasites in the midgut was unexpected, and led us to develop, for the first time, a methodology using quantitative real-time PCR (qPCR), targeting T. cruzi satellite-DNA, to quantify with higher sensibility and accuracy T. cruzi infection in R. prolixus midgut and rectum. The samples were normalized to the R. prolixus 12S ribosomal RNA gene. qPCR assays revealed that control parasites were detected both in the midgut and rectum, being more abundant in the latter, while pCHAG parasites were detected at considerably lower levels (Fig. 7B,C).
In order to assess the levels of cruzipain expression of T. cruzi after passage in R. prolixus, parasites were re-isolated after the colonization and the levels of anti-cruzipain binding to the parasite surface was compared through flow cytometry with cells obtained from cultivation in axenic medium (BHI). Our data revealed that, after colonization of the insect host rectum, parasites demonstrated a significant increase in the surface cruzipain expression (Fig. 8).
Pathogenic protozoa express large amounts and varied patterns of intracellular and/or extracellular peptidases that are involved in specific and extremely necessary functions in the parasite life-cycle, either directly through its catalytic properties or indirectly by regulating other proteins [24]. Cysteine peptidases from the papain superfamily of kinetoplastid parasites are considered as key factors for survival and interaction with the vertebrate host. Due to the importance of peptidases in physiological processes, they have emerged as promising targets for antiparasitic drugs. Therefore, T. cruzi cruzipain has been extensively investigated as a target for Chagas' disease chemotherapy [15], [25]. Although T. cruzi cruzipain is expressed abundantly on the surface of epimastigote forms, found in the insect vector, its role in parasite interaction with the insect has never been addressed before. Our research group has been studying some peptidases believed to be essential in this part of the life cycle of trypanosomatids, like gp63 and cruzipain [21], [23], [26]–[28]. Cruzipain homologues have been described in insect and plant trypanosomatids, namely Blastocrithidia culicis (recently reassigned as Strigomonas culicis [29]) and Phytomonas serpens, respectively [21], [27], [28]. In the latter, it was shown that cruzipain-like proteins are located at P. serpens cell surface, and are implicated on the adhesion to the salivary glands of Oncopeltus fasciatus, a phytophagous insect employed as an experimental model [21], [28].
The present study investigated the relevance of T. cruzi cruzipain in the interaction process with R. prolixus midgut. Our results showed that the five cysteine peptidase inhibitors used (cystatin, antipain, E-64, leupeptin and iodocetamide) significantly decreased the adhesion of T. cruzi to R. prolixus posterior midgut. T. cruzi possesses two major cysteine peptidases, the cathepsin L cruzipain [15], [25], and a 30-kDa cathepsin B like-peptidase [22]. We showed that a specific cathepsin B inhibitor, CA-074-OMe, promoted no significant alteration in the adhesion rate, which is in accordance to its intracellular localization [22]. These findings suggest that a cathepsin L may be the molecule responsible for the reduced parasite binding to the insect midgut.
In plants and mammals, endogenous inhibitors of the cystatin superfamily are regulators of cysteine peptidase activity of enzymes from the papain superfamily with high affinity for cathepsin L. Chagasin, identified in T. cruzi, is a tight-binding high affinity reversible inhibitor of papain-like cysteine peptidases. In T. cruzi, chagasin interacts with cruzipain, regulating the activity of this enzyme [10], [30]. Herein, we showed that parasites treated with cystatin displayed reduced adhesion to the luminal surface midgut of the R. prolixus in a dose-dependent manner. Therefore, despite of the doubtful selectivity of some inhibitors, the results obtained using cystatin strongly suggest that papain-like peptidases are required for efficient interaction between the parasite and the insect midgut. Moreover, the blockade of cruzipain by antibodies also led to a significant reduction in the capacity of adhesion to the midgut of the insects in relation to untreated parasites. This effect may be caused by steric intervention, where the antibodies prevent the access of insect gut molecules to specific sites in cruzipain, which could act as a recognition molecule. It has been shown in several monoxenic trypanosomatids that the metallopeptidase gp63 participates in parasite attachment to the insect midgut through a proteolytic-independent mechanism, which involves the recognition and binding to a yet unidentified insect midgut receptor [23], [26]. Our results indicate that the pre-incubation of the R. prolixus midgut with cruzipain did not influence the interaction rate. This could be an indication that cruzipain does not act as an adhesion molecule, suggesting that it could act cleaving off either surface molecules from the parasite or the insect epithelial cells, thus exposing hidden relevant molecules, which may facilitate parasite access to binding sites, promoting adhesion and colonization. For instance, our group has previously shown that P. serpens cruzipain-like molecules are able to degrade O. fasciatus salivary gland proteins [31].
However, the lack of inhibition by the pre-treatment of the insect midgut with the purified enzyme (either active or inactive) cannot rule out the participation of cruzipain as a ligand/adhesion molecule. The attachment of microorganisms to a biological surface is a complex process involving specific interactions between adhesins and complex receptors on host tissues. It should be kept in mind that surface molecules do not exist in their isolated form in cellular systems. Experimental models describing structural and functional aspects of proteins have historically used purified molecules, mutants lacking genes coding for the enzymes, and specific protein-binding probes, including antibodies, peptides and inhibitors. These classic approaches have traditionally focused on isolated molecules for structural and/or functional testing [32]. However, these approaches do not take into account the molecular associations at the cell surface, for instance, protein-protein, lipid-protein, glycan-protein and all possible inter-associations. The study of isolated molecules is insufficient to fully elucidate the functional impact of the complex structures that can be formed and are upon influence of the microenvironment of the insect midgut.
It is well known that T. cruzi presents remarkable genetic diversity between isolates [33]. Distinct expression profile and activity either of the enzyme (cruzipain) or the inhibitor (chagasin) could contribute to the biologic heterogeneity found between different isolates of the T. cruzi. Tissue-culture trypomastigotes (TCT) from the G isolate are less infective to mammalian cells and presents reduced activity of cruzipain [9], [10]. In this sense, we sought to compare the ability of several T. cruzi isolates as well as chagasin transfectants to adhere to R. prolixus. The interaction rate of the chagasin transfectants to the insect midgut was considerably lower in relation to wild-type parasites. The addition of exogenous active cruzipain partially restored in a dose-dependent manner the adherence of chagasin overexpressing parasites to the insect midgut, which further supports that this reduction in the adhesion rate was linked to the reduced levels of cruzipain and not to other phenotypic alterations induced by the overexpression of chagasin. Also, T. cruzi G isolate adhered to the insect gut to a lesser extent in comparison to Dm28c, Y and CL Brener isolates, indicating a biological deficiency of the G isolate, possibly linked to the lower activity of cruzipain. In epimastigotes, chagasin and cruzipain co-localize in at least two compartments of the secretory pathway of epimastigotes: the Golgi complex and the reservosome. It was also shown that cruzipain-chagasin complexes are formed in living parasites. Therefore, it is conceivable that chagasin associates with cruzipain molecules in the Golgi complex before sorting to reservosomes, flagellar pocket and plasma membrane [10], [30], suggesting that the surface cruzipain is inactive/inaccessible in the chagasin transfectants. Also, it is still unknown if epimastigote surface cruzipain is proteolytically active or not [25]. Interestingly, it has been shown in other microorganisms that classical cytosolic enzymes are on the surface, present activity and act as an adhesin, even if these enzymes do not possess the classical N-terminal sequence that predicts surface location [34]. Accordingly, our research group has shown that T. cruzi calpain molecules, which are typical cytosolic proteins, are present on T. cruzi surface and are involved in the metacyclogensis, interaction with mammalian cells, parasite proliferation and adhesion to the insect vector [35]–[37]. Ultimately, it illustrates that the interaction process involves a pool of molecules both on the microorganism and the host.
The data from the ex vivo assays are very suggestive of the involvement of cruzipain in the interaction with the invertebrate host. This hypothesis was further supported by in vivo colonization experiments, which revealed parasites only in the rectum by direct microscopic counts. The traditional parasite quantification method, through direct microscopic observation, is exhaustive, subjected to errors, and with reduced sensitivity. This led us to develop, for the first time, a methodology using quantitative real-time PCR (qPCR) using SYBR-green targeting T. cruzi satellite-DNA to quantify T. cruzi infection in R. prolixus. qPCR assays performed with chagasin transfectant or wild-type parasites revealed that the ability of the former to colonize in vivo was drastically reduced, being detected both in the midgut and rectum at considerably lower levels than wild-type parasites, which were more abundant in the rectum. The rectal portion is considered a site of stress, which induces the parasite metacyclogenesis in the vector. It is worth mentioning that cruzipain is abundantly expressed on the surface of epimastigote forms, while in tissue culture-derived trypomastigotes, the surface labeling is either absent or faint [5]. This fact together with the hypothesis that cruzipain might be involved in the attachment to the insect midgut could help to explain why T. cruzi is released after metacyclogenesis. However, tissue-derived and insect-derived trypomastigotes may present distinct surface properties, and this should be further investigated.
It is well known that long periods of in vitro culture reduce the expression of parasite virulence factors. In this sense, we showed through flow cytometry analysis a considerable increase in the levels of T. cruzi surface cruzipain after colonization of R. prolixus rectum, in comparison to cells cultivated in BHI medium. Previous studies indicated that an attenuated strain of Leishmania major produced low amounts and low enzymatic activity of gp63. After serial passages of these parasites through either Phlebotomus duboscqi or through mice, the recovery of the proteolytical activity was seen in a similar level of that presented in a virulent strain of L. major [38]. Reduced levels of gp63 are frequent in Leishmania promastigotes that undergo long-termed maintenance in vitro [38]. In addition, our research group reported previously the enhancement in the expression of gp63-like molecules in Herpetomonas samuelpessoai after colonization of an insect host model, Aedes aegypti [39]. Attenuated T. cruzi strains display reduced content of active cruzipain compared to virulent strains [40], which suggests a strong correlation between the virulence/attenuation of long-term T. cruzi cultures and the activity of cruzipain.
Although cruzipain is defined as the major cysteine peptidase detected in T. cruzi epimastigotes, this peptidase is a member of a large multigene family composed of polymorphic genes, whose expression are stage regulated in the parasite [25]. In epimastigotes, the majority of cruzipain RNA encodes highly similar isoforms, while in trypomastigotes and amastigotes, the expression of more divergent cruzipain genes can be detected [41]. The majority of the biochemical studies on cruzipain were performed using the natural enzyme purified from epimastigotes. The major isoform isolated from epimastigotes has been referred as cruzipain 1 (n-cruzipain 1) [42]. Cruzipain 2 is preferentially expressed by trypomastigotes and amastigotes. Although the isoforms show distinct substrate preferences, which would implicate on unique functions [42], both n-cruzipain 1 and n-cruzipain 2 may participate cooperatively in relevant biological processes such as host cell signaling and invasion by T. cruzi [8], [9], as well as in the interaction with the insect host.
Altogether, these findings establish that cruzipain is one of the molecules involved in the interaction between T. cruzi and its invertebrate host. Indeed, our results demonstrated that this enzyme is involved in the successful adhesion to the epithelial cells of insect vector both ex vivo and in vivo, although the exact molecular mechanism should be further explored. Collectively, our work adds new insights, never assessed before, about the relevance of cruzipain in the infection of the insect vector, R. prolixus.
|
10.1371/journal.pcbi.0040043 | Accurate Structural Correlations from Maximum Likelihood Superpositions | The cores of globular proteins are densely packed, resulting in complicated networks of structural interactions. These interactions in turn give rise to dynamic structural correlations over a wide range of time scales. Accurate analysis of these complex correlations is crucial for understanding biomolecular mechanisms and for relating structure to function. Here we report a highly accurate technique for inferring the major modes of structural correlation in macromolecules using likelihood-based statistical analysis of sets of structures. This method is generally applicable to any ensemble of related molecules, including families of nuclear magnetic resonance (NMR) models, different crystal forms of a protein, and structural alignments of homologous proteins, as well as molecular dynamics trajectories. Dominant modes of structural correlation are determined using principal components analysis (PCA) of the maximum likelihood estimate of the correlation matrix. The correlations we identify are inherently independent of the statistical uncertainty and dynamic heterogeneity associated with the structural coordinates. We additionally present an easily interpretable method (“PCA plots”) for displaying these positional correlations by color-coding them onto a macromolecular structure. Maximum likelihood PCA of structural superpositions, and the structural PCA plots that illustrate the results, will facilitate the accurate determination of dynamic structural correlations analyzed in diverse fields of structural biology.
| Biological macromolecules comprise extensive networks of interconnected atoms. These complex coupled networks result in correlated structural dynamics, where atoms and residues move and evolve together as concerted conformational changes. The availability of a wealth of macromolecular structures necessitates the use of robust strategies for analyzing the correlated modes of motion found in molecular ensembles. Current strategies use a combination of least-squares superpositions and statistical analysis of the structural covariance matrix. However, the least-squares treatment implicitly requires that atoms are uncorrelated and that each atom has the same positional uncertainty, two assumptions which are violated in structural ensembles. For example, the atoms in the proteins are connected by chemical bonds, covalent and non-covalent, resulting in strong correlations. Furthermore, different atoms have different variances, because some atoms are known with less precision or have greater mobility. Using maximum likelihood (ML) analysis, we have developed a technique that is markedly more accurate than the classical least-squares approach by accounting for both correlations and heterogeneous variances. The improved ability to accurately analyze the major modes of dynamic structural correlations will benefit a diverse range of biological disciplines, including nuclear magnetic resonance (NMR) spectroscopy, crystallography, molecular dynamics, and molecular evolution.
| Biological macromolecules, like proteins and catalytic RNAs, are dynamic structures. Each of the atoms in a macromolecule is coupled with other atoms via covalent bonds and various non-covalent interactions. This large and complex network of interconnections produces correlated structural dynamics, in which a perturbation or movement of one structural element covaries with the positional displacement of other elements. Thus, over a given time frame, macromolecules exist as an ensemble of correlated substates which span a large configurational space. Relevant time scales for dynamic structural change can range from picoseconds in molecular dynamics studies, to milliseconds for large structural movements, to millennia in evolutionary analyses of conformational perturbations due to amino acid substitutions. Understanding the correlated dynamics of such systems is essential for mapping structure to function. However, structural biologists currently have few tools for analyzing the correlations found in an ensemble of structures.
Previous work characterizing the structural correlations in macromolecules has been limited to analysis of molecular dynamics (MD) simulations. Two general methods have been used to extract major modes of functionally relevant motions: normal mode analysis [1] and principal components analysis (PCA) of atomic covariance matrices [2,3]. Studies using these methods have largely shown that protein motions are dominated by only a few major distinct modes of correlated movement. Normal mode analysis assumes that dynamics are harmonic. In contrast, PCA does not make this assumption, and it has been found to be useful for finding major modes when the dynamics are highly anharmonic, which is more biologically realistic since proteins have multiple energetic minima [1].
In standard practice, PCA of an MD trajectory first involves removal of arbitrary rotational and translational effects by conventional least-squares superpositioning [4–8]. From this least-squares superposition one then calculates a covariance matrix, which is subsequently used as input for eigendecomposition in PCA (also see [9]). However, the use of least squares is problematic in both theory and practice. As a statistical technique, least squares relies on two strong physical assumptions: that all atoms have the same variability, and that each atom is uncorrelated with the others. When these assumptions do not hold, least squares can give very misleading results [10]. In biomolecular applications, individual atoms in a superposition do not have equal variances, as some regions superposition closely while others show more conformational heterogeneity. Similarly, the atoms in macromolecular structures are strongly correlated by physical coupling via chemical bonds. Thus, both of the assumptions of least squares are violated in real biological data. In fact, performing PCA of a least-squares superposition is logically contradictory; the least-squares method assumes that no correlations exist, yet PCA is then performed on the least-squares derived covariance matrix to analyze those “nonexistent” correlations.
We use a maximum likelihood (ML) method that overcomes the drawbacks of conventional least-squares superpositioning methods [11–13]. Unlike least squares, ML superpositioning is valid in the presence of heterogeneous variances and correlations, thereby providing more accurate superpositions [12,13] and corresponding covariance (and correlation) matrices. Rather than performing separate superpositioning and covariance matrix calculation steps, our ML superpositioning method simultaneously determines the optimal superposition and the optimal covariance matrix. We show that, as expected, PCA of our ML superposition provides markedly more accurate structural correlations than those extracted from least-squares superpositions. Furthermore, we show that use of the correlation matrix, rather than the covariance matrix, automatically corrects for biases that may be introduced due to experimental uncertainty in atomic positions or due to large differences in the magnitude of dynamic motion. We provide examples of the generality of the method by applying it to alternate crystal forms of the same protein, nuclear magnetic resonance (NMR) ensembles, and distant homologs with differing amino acid sequences.
We performed two simulation analyses to confirm the ability of our ML method to accurately determine the structural correlations found in sets of conformationally similar molecules. Two sets of conformationally perturbed protein structures were generated randomly by assuming a Gaussian distribution with known mean and known covariance matrices (and, hence, based on known correlation matrices; see Figure 1A and 1E). In this case, the covariance matrix is a mathematical description of the positional variation and correlations among the atoms in an ensemble of molecular structures (for more background regarding covariance and correlation matrices, see Methods). Two different covariance matrices were used: one with a range of variances, yet no correlations, and another with the same range of variances plus strong correlations (the corresponding “true” correlation matrices are plotted in Figure 1A and 1E). The correlation structure and the range of variances are typical of NMR solution structures found in the PDB database (see Methods). We then randomly translated and rotated each of the perturbed structures. Both least-squares and ML superpositions were performed independently on these two sets of simulated protein structures to obtain estimates of the true covariance/correlation matrix that was used to generate the structures (Figure 1B–1D and 1F–1H).
We found that, when calculated from an ML superposition, both the covariance matrix and the corresponding correlation matrix are considerably more accurate than those calculated from least-squares superpositions (Figure 1). When compared to the true (known) correlation matrix, the least-squares correlation matrix is highly biased, showing an artifactual pattern of correlation (Figure 1B and 1F). As shown in Figure 1C and 1G, the least-squares correlation matrix remains artifactually biased even when the majority of highly variable atoms are excluded from the analysis, as often done in common practice (“truncated least squares,” where disordered regions are subjectively removed from the analysis with intent to obtain lower RMSDs). Interestingly, the least-squares procedure imparts a highly similar, artifactual correlation structure regardless of the true correlations (compare Figure 1B and 1C, with no true correlations, to Figure 1F and 1G, in which the structures had true strong correlations). In contrast, the ML-based correlation matrix reliably recapitulates the true complex patterns of correlation (Figure 1D and 1H).
To extract major modes of structural correlation from a superposition, we use the statistical method of principal components analysis (PCA; see Methods). PCA produces multiple principal components, each of which represents the predominant modes of structural correlation within the superposition. Generally, only the first few principal components (that is, those with the largest eigenvalues) are of practical interest, since they usually account for the majority of correlations in the data. As shown below, when significant covariation exists in a family of structures, PCA based on a least-squares superposition will yield erroneous principal components, resulting in artifactual modes of correlation.
As with the correlation matrices, we found that the principal components determined from an ML superposition are likewise more accurate than principal components from a least-squares superposition (Figure 2). In these images, the largest (or first) principal component has been plotted in color on a single representative structure from the superposition. We refer to these types of graphs as “PCA plots.” Red regions are correlated with each other, meaning that these regions tend to “move together” on average within the set of structures. Similarly, blue regions are also correlated with each other. However, the red regions are anti-correlated with the blue regions, meaning that red and blue regions tend to “move” in opposition to each other. White regions represent atoms whose positions are completely uncorrelated.
In the first analysis, the PCA plots shown in Figure 2A–2D were calculated from simulated structures that had no bona fide correlations among their atoms (using the correlation structure plotted in Figure 1A). Nevertheless, the largest principal components from the least-squares superpositions indicate a substantial, yet completely artifactual, mode of correlation, even when only the well-ordered residues are included in the superposition (compare the true first principal component in Figure 2A with Figure 2B and 2C). In contrast, the first principal component from the ML superposition faithfully shows very little correlation, as indicated by the lack of colored patterns (Figure 2D). PCA of the ML superposition also avoids the need for a subjective judgment on which residues to remove from the analysis.
In the second, complementary analysis, protein structures were simulated which had strong correlations, using the correlation matrix plotted in Figure 1E. As before, the first principal component from the least-squares superposition indicates a large, artifactual mode of correlation, which is still present even when the highly variable residues are excluded (Figure 2F and 2G). PCA of the ML superposition, however, accurately estimates the true correlation (Figure 2H).
Results from our ML method differ most from the conventional least-squares method when there is a wide range of variances among the atoms (that is, when some regions of the structures are well-superpositioned and other regions are highly disordered) and when correlations are strong. As the variances for the atoms become more uniform, and as the correlations approach zero, our method converges on the conventional least-squares method. Even so, the poor performance of the least-squares PCA method persists despite the removal of the majority of the most highly variable residues (residues 1–5 at the N-terminus; see Figure 2C and 2G). Thus, with the improved accuracy of ML superpositions, PCA can be used reliably to find the major modes of positional variation and dynamical correlation within a family of structures.
The method presented here for identifying major modes of structural correlation is general, and in principle it can be used to analyze any structural superposition, including independent solutions of the same protein, different homologous proteins, or a series of MD conformations. As one example, Figure 3 shows the second principal component from an ML superposition of a series of 10 crystal structures structures of the 70S ribosomal subunit from Haloarcula marismortui, including nine structures of the subunit bound to different antibiotics [14–16]. Remarkably, the majority of the correlation is localized to the active site of ribosome, the subunit interface, and the active site cleft, which binds the actively transcribed mRNA, tRNAs, translation factors, and the nascent polypeptide. The regions of strong correlated positional displacement also roughly correspond to regions of high RNA sequence conservation (see, for example, Figure 5 of [14]). Thus, this PCA plot suggests that conformational perturbations of the ribosome during binding by various antibiotics are accompanied by correlated changes in distant yet functionally important regions.
Our method can also be used to analyze the correlated conformational changes that have occurred during the evolution of protein homologs. The ML superposition and first principal component for a set of homologous telomere end-binding protein OB-fold domains are shown in Figure 4. The PCA plot indicates a clear correlation between the two upper loops in blue and also within the red β-barrel, a fact that is otherwise difficult to ascertain from inspection of the structural alignment alone. The two blue loops are known to be critical for recognition of the proteins' single-stranded DNA ligand [17,18]. Thus, this PCA analysis implies that these loops (and also the β-barrel) have co-evolved in terms of conformation during the divergence of these domains from a common ancestor [19–21].
The correlations found in PCA plots are also useful for analyzing ensembles of solution structures of macromolecules solved by NMR spectroscopy. For instance, Figure 5A and 5C shows the largest principal mode of correlation from solution structures of ubiquitin solved by dynamic ensemble refinement, which takes into account the dynamic heterogeneity of a protein as measured by NMR relaxation experiments in addition to NOE distance constraint data [22]. Two independent NMR refinements of the ubiquitin structure are shown to give a sense of the reproducibility of our ML PCA method [22,23]. Two key residues in the core of the protein, Val5 and Ile30, pack against each other and are highly anti-correlated, indicating that during the “fluid-like” dynamic motion of the protein's interior these residues move in opposition to each other. Val5 and Ile30 are both members of a small set of core residues that have been implicated in forming a folding nucleus in ubiquitin [24]. Furthermore, these residues are notable for being some of the most highly conserved among ubiquitin homologs [25], for exhibiting the slowest rates of hydrogen exchange in the protein [26], and for decreasing the thermodynamic stability of the protein when mutated [27]. Together with these experimental results, ML PCA suggests that strongly correlated residues in ubiquitin are important for its folding and stability.
Our method is reminiscent of previous work that has used PCA of covariance matrices to extract major modes of functionally relevant motions from MD trajectories [2–8]. However, the interpretation of PCA of a covariance matrix is problematic, as that method results in modes of covariation that are a convolution of both the correlation and the variance of the atoms (see Equation 3 in Methods). In structural superpositions, two very different factors contribute to the conformational variance: (1) random experimental uncertainties and (2) dynamic motion or conformational heterogeneity. Because we use the correlation matrix, rather than the covariance matrix, our method cleanly separates pure correlations from the variance, and thus the resulting principal components can be interpreted as bona fide modes of correlation.
For instance, often the variances in a covariance matrix are composed of stochastic contributions that can be physically irrelevant or uninteresting. In NMR ensembles, the variance of each atom reflects not only the dynamics of that atom but also the number of experimental constraints for the position of that atom. Highly uncertain regions of a structure can therefore dominate the largest principal component from a covariance matrix, thereby artifactually inflating the importance of these imprecise regions. An example is shown in Figure 5B, where the disordered C-terminal tail of ubiquitin has a large variance largely due to experimental imprecision (from a paucity of NOE distance constraints), resulting in its unilateral contribution to the largest principal component of the covariance matrix. PCA of a correlation matrix, on the other hand, circumvents this problem by down-weighting uncertain regions in proportion to their variances (see Equation 2 and compare Figure 5A and 5C with Figure 5B and 5D).
Furthermore, in an MD trajectory, a highly mobile loop with little correlated movement with other parts of the structure can nevertheless dominate the first mode of covariation. As a result, the largest principal components from the covariance matrix will primarily represent large magnitude motions with little or no real correlated movement. Covariance matrix PCA is useful, then, for analyzing major modes of motion when coordinate precision is high. However, covariance PCA is generally uninformative about true conformational correlation.
In sum, correlation matrix PCA produces modes of pure correlation that are independent of the uncertainties in atomic positions, since the variance components have been normalized away (Equation 2). Our ML method thus provides correlations that are unlikely to be artifacts of experimental imprecision or of the magnitude of dynamic motions in localized regions of the structure.
Our maximum likelihood method provides principal components that accurately describe the modes of coordinated motions and correlations found in an ensemble of structures. By using correlation matrices rather than covariance matrices, the modes of correlation that are found are largely free of artifacts that can result from experimental imprecision and the magnitude of dynamic motion. Taken together, various experimental results suggest that highly correlated residues from PCA plots are likely to be functionally significant. Thus, maximum likelihood PCA of structural superpositions, and the structural PCA plots that illustrate the results, should prove to be of wide utility in analyzing and comparing macromolecules in diverse fields of structural biology.
A covariance matrix is a mathematical description of the variation and covariation among members of a dataset. In the case of macromolecular structures, the covariance matrix describes the positional variation and correlations among the atoms observed in properly superpositioned family of structures. For example, given a protein K amino acids in length, here we consider the K × K covariance matrix representing the covariation of each of the K α-carbons with each of the others. If the orientations of the structures are known with certainty, then the diagonal elements σi,i of the covariance matrix Σ are simply the variances for each of the atoms. Each off-diagonal element σi,j≠i is the covariance of the ith atom with the jth atom. The elements σi,j of the covariance matrix Σ can be defined as:
where
denotes the arithmetic average of yi over all i, and the xi here are 3-vectors representing the 3-D x, y, and z coordinates of each atom.
The correlation matrix C, on the other hand, is a simple function of the covariance matrix that has been normalized by the variances, leaving only pure correlations. Each element ci,j of the correlation matrix C is given by:
Unlike a covariance matrix, the diagonal elements of a correlation matrix all equal 1, and the non-diagonal elements range from −1 to 1 (corresponding to perfect negative correlation and positive correlation, respectively). Clearly, the accuracy of both the covariance and correlation matrices directly depends on the accuracy of the superposition. Note that, if the covariance matrix is known, then the correlation matrix is also necessarily known. However, the transform is not symmetric, as the correlation matrix does not contain all the information needed to reconstruct the covariance matrix; the variances are also required:
Major modes of structural correlation within a given structural dataset were found using the statistical method of principal components analysis (PCA). To perform PCA, the correlation (or covariance) matrix is diagonalized by spectral decomposition. The resulting eigenvectors are ranked according to their corresponding eigenvalues, largest to smallest. The eigenvector with the largest eigenvalue corresponds to the first principal component, which summarizes the major mode of correlation (or covariance) in the data. The second principal component corresponds to the second largest mode of correlation, and so on. Unless otherwise indicated, all examples reported here used PCA of the correlation matrix, although our program THESEUS will also perform PCA on the covariance matrix if desired (see Implementation).
A statistical likelihood model for superpositioning structures. A detailed treatment of the following likelihood analysis can be found elsewhere [12,13]. We present here a simplified account of the ML method and its rationale, focusing on simultaneous estimation of the covariance matrix in the macromolecular structural superpositioning problem. In the following, we specifically consider the superpositioning problem per se, as opposed to the structural alignment problem. We assume that the one-to-one correspondence between atoms or residues (i.e., the alignment) is known.
Consider superpositioning N different structures (Xi, i = 1…N), each with K corresponding atoms. Each structure is mathematically represented as a K × 3 matrix of K rows of atoms. We assume a statistical perturbation model in which each macromolecular structure Xi is drawn from a matrix normal (Gaussian) probability distribution [28,29]. Each structure Xi to be superpositioned is considered as an arbitrarily rotated and translated Gaussian perturbation of a mean structure M:
where ti is a 1 × 3 translational row vector, 1K denotes the K × 1 column vector of ones, and Ri is an orthogonal 3 × 3 rotation matrix. The entries of the K × 3 matrix Ei are filled with normal random errors, each with mean zero, i.e., Ei ∝ NK,3(0, Σ, I3). The K × K covariance matrix Σ describes the (spherical) variance of each atom and the covariances among the atoms.
The likelihood equation for matrix Gaussian superpositioning. In the superposition problem with arbitrary translations, the covariance matrix Σ is poorly identified and singular unless it is parametrically constrained. Thus, to render the covariance matrix estimable, we assume that its eigenvalues are hierarchically distributed according to an inverse gamma probability density. An inverse gamma distribution is physically reasonable, as extremely small or large variances are relatively unlikely. The full joint log-likelihood for a structural superposition is then the sum of the log-likelihood for the eigenvalues of the atomic covariance matrix and the log-likelihood for the multivariate matrix normal density [30,31] corresponding to the statistical model given by Equation 4. The full superposition log-likelihood
is thus
where |U| denotes the determinant of a matrix U,
denotes a squared Frobenius Mahalanobis matrix norm, and α and γ are the scale and shape parameters, respectively, of an inverse gamma distribution for the K eigenvalues (λj) of the atomic covariance matrix Σ:
ML superposition solutions. In the following, we briefly give the ML solutions for each of the unknown parameters of the superposition log-likelihood equation from above.
Each observed structure must be translated to its row-weighted centroid:
where
t̂" i is the ML estimate of the translation:
The optimal rotations are calculated using a singular value decomposition (SVD). Let the SVD of an arbitrary matrix D be UΛV'. Then, the ML rotations
R̂" i are estimated by
The mean structure is estimated as the arithmetic average of the optimally translated and rotated structures:
Finally, the ML estimate of the atomic covariance matrix
is given by:
where the unconstrained ML estimate of the covariance matrix
is:
Because the estimate of the covariance matrix Σ is a function of the other unknown parameters, the ML solutions given above must be solved simultaneously by numerical methods [12,13]. We use an iterative algorithm based on the Expectation-Maximization (EM) method [32,33]. The algorithm assumes that the alignment (the one-to-one correspondence among atoms/residues in the structures) is known a priori, and it aims to determine the ML superposition given that alignment. In brief:
1. Initialize: Set
= I. Randomly choose one of the observed structures for use as the mean structure
M̂" .
2. Translate: Translate (i.e., center) each according to Equation 7.
3. Rotate: Calculate each rotation
R̂" i (Equation 8), and rotate each translated structure by setting
.
4. Estimate the mean: Recalculate the average structure
M̂" (Equation 9). Return to step 3 and loop until convergence.
5. Estimate the inverse gamma distributed eigenvalues: Estimate
(Equation 11) and find its sample eigenvalues. Estimate the inverse gamma parameters by iteratively fitting them to the eigenvalues of the ML estimate of the covariance matrix, treating the zero eigenvalues (or the smallest variance) as missing data in an expectation-maximization algorithm.
6. Estimate the atomic covariance matrix: Modify
according to Equation 10. Return to step 2 and loop until convergence.
7. PCA: Perform a principal components analysis on the correlation matrix (or corresponding covariance matrix).
If all variances are assumed to be equal and all covariances are assumed to be zero (i.e., Σ ∝ I), then this algorithm corresponds to the classical least-squares algorithm for the simultaneous superpositioning of multiple structures [34–37]. The algorithm presented above (like that of Theobald and Wuttke [13]) is similar to that given in Theobald and Wuttke [12], with three exceptions. First, the algorithm of [12] is much more general, e.g., it is applicable to data in an arbitrary number of dimensions. Here we assume D = 3 for 3-D, spatial data. Second, here no scaling factors are necessary (i.e., βi = 1 for all structures), since molecules are inherently in the same scale, as bond lengths are fixed by the laws of physics. Third, we further assume that the variance about each atom is spherical (i.e., Ξ = I), an assumption that greatly simplifies the calculations.
The algorithm described above for calculating ML superpositions and performing PCA of the estimated covariance matrix is implemented in the command-line UNIX program THESEUS [12,13]. THESEUS operates in two different modes: (1) a mode for superpositioning structures with identical sequences and (2) an “alignment mode,” which superpositions homologous structures with different residues given a known alignment (for instance, as determined from a sequence alignment program or from a structure-based alignment program). THESEUS does not perform structure-based sequence alignments, which is a distinct bioinformatic problem [38]. As with all superposition methods, THESEUS requires an a priori one-to-one mapping among the atoms/residues (i.e., it requires a known alignment). With NMR models or different crystal structures of identical proteins, the one-to-one mapping is trivial. When superpositioning different molecules with different sequences, however, a sequence alignment must be provided as a guide. THESEUS accepts sequence alignments in standard CLUSTAL and A2M (FASTA) formats.
In addition to the ML superposition for a set of structures, THESEUS will calculate the principal components of either the covariance or correlation matrix. For input, THESEUS takes a set of standard PDB formatted structure coordinate files (http://www.wwpdb.org/docs.html [39,40]). PCA analysis is requested with the “-Pn” command line option, where “n” is substituted with the number of principal components desired (usually three are sufficient). PCA of the correlation matrix is performed by default; the “-C” option specifies that the covariance matrix should be used. Each principal component is written into the temperature factor field of two output files: (1) a PDB formatted file of the optimal ML superposition (each structure is represented as a different MODEL) and (2) a PDB formatted file of the estimate of the mean structure. Principal components can then be visualized as PCA plots (described in Results/Discussion) with any visualization software, such as PyMOL [41], RasMol [42], or MolScript [43], that can color the structures by values in the temperature factor field.
Two artificial datasets of protein coordinates were prepared as described previously [12]. Briefly, for each set, 300 protein structures were generated randomly, assuming a matrix Gaussian error distribution with a known mean protein structure and known atomic covariance matrix. The α-carbon atoms from model 1 of PDB:ID 2sdf (the human cytokine stromal cell-derived factor-1 protein [44]) were used as the mean protein structure (67 atoms/landmarks, squared radius of gyration = 152 Å2). The 67 × 67 atomic covariance matrices were based on values calculated from the superposition given in 2sdf, with variances ranging from 0.0452 to 79.2 Å and correlations from 0 to 0.99. Thus, in this simulation, the variances range over 3.2 orders of magnitude, a value that is typical for NMR solution structure ensembles (of 3,150 single-domain NMR families in the PDB database, the average range for the variance is 2.9 ± 1.1 (SD) orders of magnitude). The first simulated set of structures used a diagonal covariance matrix in which all covariances were set to zero. The second simulated set of structures used the full covariance matrix. Hence, both sets were generated with the same variances, differing only in their correlation structure. After generating the perturbed protein structures, each was then randomly translated and rotated.
Our ML superposition procedure was then performed on these simulated data sets, providing estimates of the atomic covariance matrix, along with estimates of the coordinates of the mean structure and of the original “true” superposition before translations and rotations had been applied. Default THESEUS parameters were used (version 1.2.6), except that the full covariance and correlation matrices were estimated with the “-c” command line option. For comparison, conventional least-squares superpositions were also calculated for the same dataset. The corresponding sample covariance and correlation matrices were calculated based on these least-squares superpositions. In order to show the effect of discarding a subset of highly variable (“disordered”) regions, separate least-squares analyses were performed using all atoms and also excluding residues 1–5 from the N-terminus, the atoms with the highest variance (referred to as “truncated least squares”).
Images of rendered macromolecules in Figures 2, 4, and 5 were made with POVScript+ [43,45] and Raster3D [46]. Figure 3 was made with PyMOL [41]. |
10.1371/journal.ppat.1006022 | CD8+ T Cells Induce Fatal Brainstem Pathology during Cerebral Malaria via Luminal Antigen-Specific Engagement of Brain Vasculature | Cerebral malaria (CM) is a severe complication of Plasmodium falciparum infection that results in thousands of deaths each year, mostly in African children. The in vivo mechanisms underlying this fatal condition are not entirely understood. Using the animal model of experimental cerebral malaria (ECM), we sought mechanistic insights into the pathogenesis of CM. Fatal disease was associated with alterations in tight junction proteins, vascular breakdown in the meninges / parenchyma, edema, and ultimately neuronal cell death in the brainstem, which is consistent with cerebral herniation as a cause of death. At the peak of ECM, we revealed using intravital two-photon microscopy that myelomonocytic cells and parasite-specific CD8+ T cells associated primarily with the luminal surface of CNS blood vessels. Myelomonocytic cells participated in the removal of parasitized red blood cells (pRBCs) from cerebral blood vessels, but were not required for the disease. Interestingly, the majority of disease-inducing parasite-specific CD8+ T cells interacted with the lumen of brain vascular endothelial cells (ECs), where they were observed surveying, dividing, and arresting in a cognate peptide-MHC I dependent manner. These activities were critically dependent on IFN-γ, which was responsible for activating cerebrovascular ECs to upregulate adhesion and antigen-presenting molecules. Importantly, parasite-specific CD8+ T cell interactions with cerebral vessels were impaired in chimeric mice rendered unable to present EC antigens on MHC I, and these mice were in turn resistant to fatal brainstem pathology. Moreover, anti-adhesion molecule (LFA-1 / VLA-4) therapy prevented fatal disease by rapidly displacing luminal CD8+ T cells from cerebrovascular ECs without affecting extravascular T cells. These in vivo data demonstrate that parasite-specific CD8+ T cell-induced fatal vascular breakdown and subsequent neuronal death during ECM is associated with luminal, antigen-dependent interactions with cerebrovasculature.
| Cerebral malaria (CM) is a severe and potentially fatal complication of malaria in humans that results in swelling and bleeding within the brain. The mechanisms that cause this fatal disease in humans are not completely understood. We studied an animal model known as experimental cerebral malaria to learn more about the factors that drive this disease process. Using a technique referred to as intravital microscopy, we captured movies of immune cells operating in the living brain as the disease developed. At the peak of disease, we observed evidence of immune cells interacting with and aggregating along blood vessels throughout the brain. These interactions were directly associated vascular leakage. This caused the brain to swell, which gave rise to an unsustainable pressure that ultimately killed neurons responsible for heart and lung function. The fatal swelling was induced by immune cells (referred to as T cells) interacting with bits of parasite presented by blood vessels in the brain. Removal of this parasite presentation protected the mice from fatal disease. We also evaluated a straightforward therapy that involved intravenous administration of antibodies that interfered with T cell sticking to blood vessels. Our movies revealed that this therapeutic approach rapidly displaced T cells from the blood vessels in the brain and prevented fatal disease.
| Malaria, a disease caused by protozoan parasites of the genus Plasmodium, is a leading cause of morbidity and mortality in the developing world. Of the 627,000 annual deaths due to malaria, the vast majority are caused by Plasmodium falciparum infections [1]. Human cerebral malaria (HCM) is one of several clinical manifestations of severe P. falciparum infection and is diagnosed by coma and parasitemia in the absence of meningitis, hyperglycemia, and postictal state [2]. HCM is fatal in 15–30% of affected individuals [3, 4], while an additional 10% of survivors suffer long-term neurological sequelae such as ataxia, hemiplegia, and cognitive impairment [5]. Yet, the underlying cause of HCM remains unknown.
Several characteristic pathologies are observed in the brains of patients suffering from HCM including vascular hemorrhage [6], breakdown of the blood brain barrier (BBB) [7, 8], and edema [9, 10]. At the cellular and molecular level, HCM is associated with an increase in systemic pro-inflammatory cytokines [11, 12], endothelial cell (EC) activation [13], and sequestration of parasite-infected red blood cells (iRBCs) [2] and leukocytes [8] within the brain vasculature. These conditions are hypothesized to contribute to the observed BBB disruption and cerebral edema as well as ischemia throughout the CNS [14]. However, interpretations of these HCM data are limited by the fact that most information about CNS pathology and the cellular response to P. falciparum is derived from post-mortem analyses. Although real-time in vivo imaging techniques such as MRI [9] and ophthalmoscopy [15] have been used in patients suffering from HCM, they lack the resolution needed to observe cellular dynamics in the CNS and have been used mostly to improve the fidelity of CM diagnoses. Examination of cellular dynamics in animal model systems is therefore needed to uncover mechanistic insights into HCM.
Infection of mice with Plasmodium berghei ANKA (PbA) induces a neurological disease called experimental cerebral malaria (ECM) that mirrors many of the pathological features observed in HCM. These include increased pro-inflammatory cytokines, vascular pathology, disruption of the BBB, and cerebral edema [16–19]. ECM in mice is a widely used model of HCM and provides a valuable tool for elucidating the mechanisms involved in CM pathogenesis and identifying cellular and molecular targets for adjunctive therapy [20]. Leukocytes have been shown to accumulate in the brains of mice throughout the course of ECM [21, 22]. In addition, mice that are genetically deficient in peripheral leukocyte chemokine receptors such as CCR5 and CXCR3 (or, their ligands) are resistant to ECM [23–26]. Many individual immune cell populations including neutrophils [27, 28], macrophages/monocytes [18, 29], NK cells [30], and CD4+ T cells [31, 32] have been implicated in the pathogenesis of this disease. However, other studies have shown that neither antibody-mediated nor genetic depletion of these cells affected the accumulation of iRBCs in the CNS [33] or the ability of mice to develop ECM [21, 34–36]. Therefore, the contribution of these immune cell subsets to ECM is still a matter of debate.
In contrast, the role of CD8+ T cells in ECM is unequivocal. Numerous studies have demonstrated that CD8+ T cell depletion [21, 31, 33, 34] or ablation of effector functions [22, 37] completely abrogates this disease. Furthermore, parasite-specific CD8+ T cells can mediate ECM in the absence of bystander T cells [38]. Despite the critical role played by CD8+ T cells during ECM, little is known about the dynamics, kinetics, anatomical localization, and function of these cells in vivo. One recent intravital study demonstrated that perivascular CD8+ T cell arrest is a signature of the disease [39], but it is unclear how these cells cause neurological symptoms or why mice succumb to ECM.
Using several techniques, including intravital imaging, we set out in this study to conduct an unbiased examination of how innate and adaptive immune cells contribute to cerebrovascular perturbations during ECM and to identify the cause of this fatal disease. We also sought in vivo insights into how cerebrovascular ECs respond to ECM and whether immune interactions with these cells could be therapeutically manipulated to ameliorate disease pathogenesis. Our data demonstrate that CD8+ T cells drive fatal vascular leakage during ECM, and they accomplish this by interacting in an antigen-dependent manner with cerebrovascular ECs. This results in profound BBB dysfunction and secondary death of neurons, most notably in the brainstem, which likely gives rise to autonomic dysfunction and death. ECM can be prevented by eliminating antigen presentation in cerebrovascular endothelial cells or by displacing parasite-specific CD8+ T cells from CNS blood vessels using an anti-adhesion molecule therapy.
Human cerebral malaria is associated with several hallmark pathologies in the brain parenchyma including vascular hemorrhaging, breakdown of the BBB, and cerebral edema. Similarly, we observed that mice infected with PbA were highly parasitemic, moribund, and showed evidence of profound BBB breakdown and edema at day 6 p.i. (S1A–S1E Fig). We also observed evidence of significant vascular hemorrhaging in the brain parenchyma (S1F Fig) that was almost always associated with iRBCs (S1G Fig). Interestingly, evidence of vascular pathology was also found in the meninges, which has not been reported previously (Fig 1A).
To gain insights into the immunopathogenesis of ECM, we conducted a temporal analysis of the immune subsets that arrive in the CNS as PbA-infected mice develop neurological symptoms. We used a clinical scoring system of 0 (moribund) to 20 (asymptomatic) developed by Carroll et al. [40] to assess disease severity (S1A Fig). We separately evaluated the meninges and brains of naïve, d5 p.i. (parasitemic but asymptomatic), and d6 p.i. (highly parasitemic and symptomatic) mice. This study revealed a significant increase in CD8+ T cells and Ly6Chi inflammatory macrophages/monocytes in the brains and meninges as mice developed symptoms on d6 p.i. (Fig 1B and 1C). To monitor PbA-specific CD8+ T cells, we generated pentamers consisting of H-2Db loaded with an immunodominant PbA epitope (SQLLNAKYL) described by Howland et al. [41]. Interestingly, PbA-specific CD8+ T cells increased in frequency and number in the spleen, meninges, and brain as mice progressed from asymptomatic (d5 p.i.) to symptomatic (d6 p.i.) (Fig 1D and 1E). Thus, neurological symptoms during ECM are associated with the migration of Ly6Chi monocytes and parasite-specific CD8+ T cells to the CNS.
Because myeloid cells are recruited to the CNS during ECM, we examined the dynamics and anatomical distribution of these cells using intravital two-photon microscopy (TPM) through a thinned skull window as described [42, 43]. We monitored the dynamics of myelomonocytic cells (monocytes and neutrophils) using lysozyme M-GFP (LysMgfp/+) reporter mice [44]. Consistent with our flow cytometric data Fig 1B and 1C we observed a significant increase in the number of myelomonocytic cells in symptomatic mice at d6 p.i. compared to d5 p.i. and uninfected animals. Interestingly, the myelomonocytic cells were contained almost entirely within the vasculature and were found in both the brain and periphery (ear) indicating systemic inflammation (Fig 2A) (S1–S3 Movies). These cells were also occasionally associated with vascular leakage in the brain (S2 Movie).
To determine the anatomical relationship between myelomonocytic cells and iRBCs during the development of ECM, we infected LysMgfp/+ mice with recombinant PbA expressing mCherry and OVA (PbA-OVA-mCherry) [45]. In symptomatic mice at d6 p.i., we observed fluorescent iRBCs adherent to cerebral vasculature (Fig 2B; S3 Movie), similar to what is found in human CM patients. We also visualized myelomonocytic cells actively phagocytosing these iRBCs while patrolling the vascular lumen (Fig 2B; S3 Movie). In fact, we found that on average nearly 60% of the fluorescent iRBCs were associated with LysM-GFP signal at any one time. In addition to the vascular lumen, some fluorescent iRBCs localized to the perivascular spaces of intact vessels, suggesting extravasation (S4 Movie). These parasites were rapidly acquired by perivascular macrophages visualized in CX3CR1gfp/+ mice (S4 Movie). These data indicate iRBCs were positioned on the luminal and abluminal surface of cerebral blood vessels during the development of ECM.
The contribution of innate immune cells to the pathogenesis of ECM is still a matter of contention. Previous studies have either implicated monocytes and neutrophils in the pathogenesis of ECM [18, 27–29] or shown them to be irrelevant [21, 33, 35, 36]. Our flow cytometric (Fig 1) and imaging (Fig 2) data suggested that these cells might be involved in disease pathogenesis. To determine if myelomonocytic cells contribute to BBB breakdown or mortality during ECM, we depleted neutrophils with anti-Ly6G antibodies in CCR2-/- mice (S2A Fig, which are deficient in circulating monocytes [46, 47]. This resulted in an 85% reduction of the total circulating myelomonocytic compartment. When compared to PbA-infected controls, mice lacking monocytes and neutrophils showed no preservation of BBB integrity or reduced mortality rate during ECM (Fig 2C–2E). In addition, depletion of myelomonocytic cells did not improve clinical scores or alter parasitemia levels (S2B and S2C Fig). Based on these results, myelomonocytic cells do not appear to play a significant role in the pathogenesis of ECM.
Having ruled out myelomonocytic cells as the cause of fatal pathology during ECM, we focused on the adaptive immune response, with a specific emphasis on T cells given that mice lacking B cells are still susceptible to ECM [34]. We initiated this line investigation by conducting a series a T cell depletion experiments. On d4 p.i. PbA-infected mice were administered antibodies specific for CD8+ or CD4+ T cells (S3A Fig) and then monitored for development of ECM relative to untreated control mice (Fig 3A). Although depletion of CD4+ T cells had no effect on survival, CD8+ T cell depleted mice were completely resistant to ECM (Fig 3A) despite having levels of parasitemia comparable to control mice (S3B Fig). Depletion of CD8+ T cells also prevented the vascular hemorrhaging (S3C Fig), BBB breakdown (S3D and S3E Fig), and edema (S3F Fig) normally associated with ECM.
To determine the anatomical localization of activated PbA-specific CD8+ T cells within the brain at the peak of ECM, naïve B6 mice were seeded i.v. with 104 naïve mCerulean+ OT-I T cell receptor transgenic CD8+ T cells and then infected with PbA-OVA-mCherry. Symptomatic mice were imaged 6 days later by TPM. These intravital imaging studies revealed that nearly all PbA-specific CD8+ T cells in symptomatic mice were arrested on or slowly crawling along the luminal and extravascular surfaces of cerebral blood vessels and were often associated with significant vascular breakdown. (Fig 3B; S5–S7 Movies). Interestingly, this activity was specific to the brain, as PbA- specific CD8+ T cells were not observed arresting along the vasculature in a peripheral tissue (ear) within the same mouse (Fig 3B; S5 Movie). Although PbA-specific CD8 T cells in the CNS were observed within perivascular spaces and the parenchyma, the majority appeared to be interacting with the luminal surface of blood vessels. To quantify the percentage of luminal vs. extravascular PbA-specific CD8 T cells, we created volumetric masks corresponding to the Evans blue signal in each blood vessel (Fig 3C, S8 Movie). After applying the mask, all visible PbA-specific CD8+ T cells, including those located on the surface blood vessels, were counted as extravascular, whereas cells obscured by the mask were considered luminal (Fig 3C, S8 Movie). We found that the vast majority of PbA-specific CD8 T cells were associated with the luminal surface of CNS blood vessels (Fig 3D). Collectively, these data indicate that CD8+ T cells arrest along the cerebral vasculature during ECM and are responsible for the vascular pathology.
Because PbA-specific CD8+ T cells were intimately associated with cerebral vasculature, we postulated that cytotoxic lymphocyte-mediated killing of vascular ECs might serve as the cause of BBB breakdown and death during ECM. To simultaneously assess cell death and vascular leakage in vivo, we injected naïve and symptomatic mice at d6 p.i. intravenously with propidium iodide (PI) (to label dead cells) and Evans blue (to assess vascular leakage). This assay revealed a striking pattern of pathology in all mice succumbing to ECM (Fig 4A). Whereas evidence of cell death was observed in multiple brain regions (e.g. olfactory bulb, cortex, cerebellum, brainstem, choroid plexus), the brainstem and olfactory bulb pathology were particularly severe (Fig 4A). Both brain regions showed evidence of profound vascular leakage and cell death. The brainstem pathology is consistent with cerebral herniation [48] and would likely give rise to autonomic dysfunction. To determine if the PI+ cells were in fact vascular ECs, we co-stained sagittal brain sections with anti-CD31 antibodies and performed quantitative analyses (Fig 4B and 4C). Although EC death was observed in multiple brain regions during ECM, only a small fraction of the total ECs was PI+ (Fig 4B and 4C). Thus, it is unlikely that EC death is the cause of fatal disease in mice with ECM.
We found a small number of PI+ ECs in the brain during ECM, but the vast majority of the dead cells were unknown. Based on cellular morphology and anatomical location, we hypothesized that at least some of the PI+ cells were neurons. Interestingly, co-staining with anti-NeuN antibodies revealed that nearly all of the PI+ cells in the brainstem were neurons—a pattern of cell death that was unique to this brain region (Fig 4D and 4E). Because the brainstem controls vital functions such as the cardiovascular and respiratory systems, it is likely that mice succumb to ECM due to the widespread neuronal death observed in this brain region.
We observed evidence of profound vascular leakage in several brain regions, including the brainstem; however, previous studies have suggested that this is not due to reduced expression of endothelial tight junction proteins [18]. We hypothesized that this negative result might be explained by a failure to directly compare tight junction (TJ) protein expression in leaking versus intact cerebral vasculature during the development of ECM. To test this hypothesis, we injected uninfected or symptomatic mice at d6 p.i. with Evans blue to locate areas of vascular leakage within the brain (Fig 5A). Next, we stained sagittal brain sections with antibodies against CD31 and claudin-5 to identify ECs and TJs, respectively (Fig 5B). Thick sections were used in order to generate volumetric 3D masks of individual blood vessels in the frontal cortex, cerebellum, and brainstem (Fig 5B). This method provides a more accurate representation of claudin-5 staining over an entire blood vessel than would be obtained by performing 2D analyses on thin sections. When we quantified the intensity of claudin-5 staining on 3D reconstructed blood vessels in various regions of the brain, we consistently found reduced expression in areas of Evans blue+ vascular leakage within symptomatic mice when compared to the same areas in uninfected mice (Fig 5C). Furthermore, we found that claudin-5 levels in brain regions of symptomatic mice where there was no vascular leakage were comparable to the levels observed on naïve blood vessels (Fig 5C). We were unable to find areas without vascular leakage in the brainstem due to the extensive amount of pathology in this brain region. Thus, by comparing to leaking to intact cerebral blood vessels, we uncovered that vascular leakage is indeed associated with reduced TJ expression.
We observed the PbA-specific CD8+ T cells interact heavily with cerebral vasculature during ECM, but the vast majority of ECs survive these engagements. To better understand the mechanisms guiding these interactions, we conducted flow cytometric analyses of cerebral ECs during the development of ECM and compared them to ECs extracted from a peripheral tissue (ear) (Fig 6A–6C). By gating on live, CD45-CD31+ cells (S4A Fig), we noted that adhesion (ICAM-1, VCAM-1) and antigen-presenting (I-Ab, Db, Kb) molecules were all significantly increased on ECs from symptomatic mice at d6 p.i. (Fig 6A and 6B). Elevated expression of these molecules appeared as early as d4 p.i. and usually increased further as the disease progressed (Fig 6B). ECs extracted from meningeal blood vessels were similarly activated (S4B Fig). Interestingly, this EC activation phenotype was unique to the CNS, as ECs extracted from a peripheral site (ear) showed a significantly reduced expression level of adhesion and antigen presentation molecules (Fig 6C). This is consistent with our intravital imaging data showing slow crawling and arrest of PbA-specific CD8+ T cells along cerebral, but not ear vasculature (Fig 3B; S5 Movie). This finding also suggests that the increase in myelomonocytic cells observed in the ear vasculature of PbA-infected mice did not induce EC activation.
We next sought insights into the mechanism underlying cerebral EC activation during ECM. Previous studies have shown that IFNγ-deficient mice are protected from ECM [34, 49, 50]. We confirmed these findings (S4C–S4E Fig) and hypothesized that IFNγ produced by lymphocytes recruited to the CNS might be responsible for cerebral endothelial cell activation during ECM (Fig 6B). A link has been established between IFNγ and ICAM-1 expression in the brain during ECM, but the cell type(s) affected by the absence of IFNγ was not determined [51, 52]. To address the role of IFNγ in EC activation during ECM, we infected wild type and IFNγ-/- mice with PbA and examined EC phenotype on day 6. ECs from IFNγ-/- mice had significantly reduced levels of adhesion and antigen presentation molecules (Fig 6D). In fact, with the exception of ICAM-1 and H-2Db, the loss of IFNγ reduced the activation state of brain ECs to uninfected levels. These data correlated with a severe reduction in PbA-specific CD8 T cell accumulation in the brain, despite normal peripheral expansion and migration (S4F Fig). Furthermore, PbA-infected IFNγ-/- mice lacked the severe vascular breakdown and brainstem neuronal cell death observed in wild type mice (Fig 6E and 6F). These data indicate that IFNγ is responsible for cerebral EC activation during ECM and that IFNγ-deficiency likely protects mice in part by keeping ECs in a naïve state.
Previous studies using CBA/J mice have demonstrated that antibody blockade of LFA-1 during ECM is highly efficacious at preventing disease [17, 53, 54]. However, we have found that this treatment is ineffective in PbA-infected C57BL/6 mice. Activated CD8+ T cells express multiple adhesion molecules, including LFA-1 and VLA-4, which are ligands for ICAM-1 and VCAM-1, respectively. Given that brain ECs upregulate both ICAM-1 and VCAM-1 during ECM (Fig 6B), we hypothesized that administration of a combination of anti-LFA-1 and anti-VLA-4 antibodies (anti-LFA-1/VLA-4) could be used to therapeutically displace PbA-specific CD8+ T cells from cerebral vasculature and prevent fatal disease. We administered this therapy on day 5.5 post-infection to avoid interfering with T cell priming. Mice at this time point were parasitemic (S5B Fig) and had symptoms associated with ECM (S5C Fig). Importantly, treatment with anti-LFA-1/VLA-4 completely reversed these symptoms and prevented death in PbA-infected mice (S5A and S5C Fig). This treatment had no effect on PbA-specific CD8+ T cell expansion (S5D and S5E Fig).
To assess the impact of anti-LFA-1/VLA-4 therapy on PbA-specific CD8+ T cell dynamics in the brain, we performed a series intravital imaging studies. At day 6 following infection with PbA-OVA, we imaged the dynamics of mCerulean+ OT-I T cells in the brain for 30 min by TPM. This was followed by intravenous administration of anti-LFA-1/VLA-4 therapy and an additional 30 min of imaging in the same anatomical location. Before antibody treatment, PbA-specific CD8+ T cells were observed slowly crawling along and arresting on cerebral blood vessels (Fig 7A; S9 Movie). In contrast, anti-adhesion antibody treatment resulted in an immediate displacement of PbA-specific CD8+ T cells from the vasculature (Fig 7A; S9 Movie). A significant reduction in the frequency of PbA-specific CD8+ T cells associated with brain vasculature was observed for the entire viewing period after anti-adhesion therapy (Fig 7B), which was not seen in mice treated with istoype control antibodies (Fig 7C). Interestingly, the frequency of PbA-specific CD8+ T cells in the parenchyma and perivascular spaces was not affected by blocking adhesion molecules (Fig 7D), suggesting that intravascular CD8+ T cell interactions are the ones responsible for fatal pathology during ECM. Furthermore, disruption of luminal CD8 T cell interactions with brain ECs via adhesion molecule blockade also prevented the death of brainstem neurons observed in isotype control treated mice (Fig 7E and 7F). In concert, these data show that blocking adhesion to ICAM-1 and VCAM-1, which are highly expressed on brain ECs during ECM, prevents PbA-specific CD8+ T cells from arresting along cerebral vasculature and rescues mice from fatal pathology and disease.
During the development of ECM, we routinely observed PbA-specific CD8+ T cells dividing following arrest on the luminal surface of cerebral vasculature (Fig 8A; S10 Movie). Because cognate peptide-MHC I interactions can advance the cell cycle program of effector CD8+ T cells [55], we hypothesized that PbA-specific CD8+ T cells interact with the brain vasculature in an antigen dependent manner. To address this hypothesis, we first compared the dynamics of PbA-specific vs. bystander CD8+ T cells of an irrelevant specificity in cerebral blood vessels. Mice were seeded with mCerulean+ OT-I T cells and then infected with PbA-OVA. When these mice became symptomatic on day 6 post-infection, we intravenously injected yellow fluorescent protein (YFP)+ DbGP33-41 CD8+ T cells (YFP+ P14) purified from the spleens of a separate group of mice infected 8 days earlier with lymphocytic choriomeningitis virus (LCMV). YFP+ P14 cells were used as activated bystander CD8+ T cells because they are specific to the LCMV glycoprotein (GP), not PbA. TPM imaging and subsequent analysis of these two CD8+ T cell populations revealed that PbA-specific CD8+ T cells moved at a significantly slower speed (Fig 8B) and spent more time arrested along cerebral vasculature (Fig 8C and S6A Fig) than the bystander CD8+ T cells. These results demonstrate that antigen-specificity dictates the interaction between CD8+ T cells and brain microvasculature during ECM.
To further demonstrate the specificity of PbA-specific CD8+ T cell interactions, we used TPM to evaluate the dynamics of these cells following injection of an anti-peptide MHC I blocking antibody. PbA-OVA infected mice seeded mCerulean+ OT-I cells were imaged by TPM on day 6 post-infection. Midway through the imaging session, mice were injected i.v. with anti-Kb-SIINFEKL (the peptide MHC complex recognized by OT-I cells) or isotype control antibodies. Injection of anti-Kb-SIINFEKL, but not isotype, control antibodies significantly elevated the velocity PbA-specific CD8+ T cells (Fig 8D), further supporting that the interactions with cerebral ECs are antigen-specific.
Next, we set out to determine the functional importance of PbA-specific CD8+ T cell interactions with cerebral ECs during the development of ECM. This was accomplished by generating bone marrow (BM) chimeras in which MHC I deficient hosts (Kb-/-Db-/- mice) were reconstituted with wild type bone marrow (Fig 8E). These mice were incapable of presenting MHC I peptides on ECs and other stromal cells while maintaining normal hematopoietic presentation. Irradiated wild type mice receiving wild type bone marrow served as a control for this experiment. Interestingly, the Kb-/-Db-/- chimeras were nearly all resistant to fatal ECM, whereas the wild type controls succumbed to disease as expected (Fig 8F and S6C Fig). Protection from ECM was observed in Kb-/-Db-/- mice despite normal parasitemia levels (S6B Fig) and generation of an equal, if not greater, PbA-specific CD8+ T cell response relative to the wild type controls (Fig 8G and 8H). To evaluate how MHC I deficiency affected PbA-specific CD8 T cell interactions with the brain vasculature during ECM, we seeded wild type and Kb-/-Db-/- BM chimeras with mCerulean+ OT-I cells and then used TPM to monitor their intravascular motility 6 days following infection with PbA-Ova. Quantification of PbA-specific CD8+ T cells in the cerebral vasculature of the Kb-/-Db-/- BM chimeras revealed significantly increased velocities (Fig 8I) and reduced arrest within the lumen of brain blood vessels (Fig 8J and S6D Fig) relative to the wild type controls. Because there are no other radio-resistant cells within the vascular lumen, these data suggest that PbA-specific CD8+ T cells engage cerebral ECs in an antigen-dependent manner during ECM.
Because PbA-infected Kb-/-Db-/- BM chimeras were resistant to ECM, we set out to determine whether they were also free from brainstem pathology normally associated with this disease (Fig 4A, 4D and 4E). Brains from wild type BM chimeras injected with PI and Evans blue at the peak of disease revealed extensive vascular leakage and cell death in the brainstem (Fig 9A). However, Kb-/-Db-/- BM chimeras were free of pathology at this same time point (Fig 9A). Nearly all of the dead cells in the brainstems of wild type BM chimeras were neurons, whereas brainstem neurons in Kb-/-Db-/- BM chimeras were unaffected (Fig 9B and 9C). These results suggest that antigen presentation by brain ECs, which fosters increased interactions with PbA-specific CD8 T cells, leads to severe brainstem pathology during ECM.
The pathological mechanisms underlying HCM are not entirely understood. Because ECM shares many of the pathological features of HCM, we set out to uncover novel mechanistic insights into the immunopathogenesis of this disorder. We made several important observations that significantly advance our understanding of cerebral malaria. Studies have shown that CM in humans and rodents is associated with BBB breakdown, edema, and hemorrhaging. During the peak of ECM, we noted that the meninges in addition to the brain parenchyma show evidence of profound vascular pathology, and this was associated a reduction in tight junction protein expression. Importantly, death from ECM is linked to marked vascular leakage and neuronal cell death in the brainstem, which is consistent with the edema and subsequent cerebral herniation recently observed in children with HCM [9]. Mechanistically, we uncovered that this fatal disease is caused by the activities of parasite-specific CD8+ T cells operating along cerebral blood vessels. As the disease developed, cerebrovascular ECs were highly activated by IFNγ, which promoted induction of cell adhesion and antigen presenting molecules. This in turn facilitated cognate peptide-MHC I dependent engagement by parasite-specific CD8+ T cells primarily on the luminal surfaces of cerebral blood vessels. The pathological significance of these interactions was demonstrated in mice rendered genetically deficient in their ability to present antigen in MHC I on ECs. These mice had reduced cerebrovascular engagement by CD8+ T cells and were resistant to fatal disease. Lastly, therapeutic administration of antibodies specific for VLA-4 and LFA-1 rapidly displaced CD8+ T cells from cerebral blood vessels and promoted survival, thus providing a simple yet effective means to treat this disease.
One of the most interesting findings in our study is the pathology observed in mice with severe ECM. By simultaneously injecting Evans blue and propidium iodide, we were able to evaluate the relationship between vascular leakage and cell death in mice succumbing to ECM. While vascular leakage was notable through the brain and meninges, the most striking areas of pathology were the olfactory bulb and brainstem. Vascular pathology was previously reported in the olfactory bulb of mice with ECM and linked to a decline in their sense of smell [56]. The severe pathology observed in the brainstem, however, is more relevant from the perspective of survival. Profound vascular leakage and neuronal cell death was seen in all mice succumbing to ECM and was associated with reduced expression of the tight junction protein, claudin-5. Sudden neuronal depolarization and death in this brain region would cause cardiorespiratory failure as observed in other neurological disorders [57, 58]. Importantly, a recent magnetic resonance imaging (MRI) study revealed evidence of severe brain swelling in 84% of children with HCM [9]. Brainstem herniation was also demonstrated in fatal instances of this disease. Our ECM results are consistent with brainstem herniation and pathology being the cause of death, which is supported by a MRI study showing significant displacement of the cerebellum and brainstem in mice with fatal ECM [48]. Therefore, increased intracranial pressure leading to cerebral herniation is the likely cause of death in rodents and children with CM.
To gain insights into the mechanisms that give rise to fatal edema during CM, we examined the activities of innate and adaptive immune cells. Our flow cytometric and two-photon imaging data revealed that ECM is associated with recruitment of innate and adaptive immune cells to blood vessels in the meninges and brain parenchyma. A strong consensus exists in the literature among several studies showing that CD8+ T cells play an essential role in ECM pathogenesis [21, 22, 31, 33, 34, 37, 38], and our data support this conclusion. There is, however, some debate regarding the role of innate immune cells, such as monocytes and neutrophils, in this disease. As mice developed ECM, our intravital time lapses revealed myelomonocytic cells migrating along cerebral blood vessels and acquiring adherent iRBCs. These cells were associated on occasion with vascular leakage; however, depletion had no impact on BBB breakdown, edema, or survival. This is different from the significant vascular disruption induced by synchronously extravasating myelomonocytic cells during fatal viral meningitis [42]. Some studies have implicated myelomonocytic cells in the pathogenesis of ECM [18, 27–29], whereas others have not [21, 33, 35, 36]. The difference in the outcome of these studies could be linked to many different variables, including strain of mice, depletion strategy, inadvertent blockade of T cells, and genetic variation in the strain of Plasmodium used. Regardless of the explanation, our findings suggest that while myelomonocytic cells may contribute to disease, they are not a dominant participant like CD8+ T cells.
At the peak of ECM, we observed parasite-specific CD8+ T cells slowly rolling, arresting, and dividing within the lumen of CNS vasculature. These interactions were highly associated with breakdown of the BBB and the flow of vascular contents into the meninges and parenchyma. Several lines of investigation have suggested that CTL-mediated killing of cerebrovascular ECs might be the cause of CNS vascular breakdown during ECM. Mice deficient in CTL effector pathways such as perforin and granzyme B are resistant to ECM [22, 37], apoptotic ECs have been identified in the retina [59] and brain [60, 61] during disease development, and CD8+ T cells isolated from ECM mice kill parasite-loaded ECs in vitro [62]. However, using a sensitive in vivo approach to quantify cell death, we were unable to demonstrate evidence of widespread EC death in the brains of highly symptomatic mice, but were able to show that Evans blue leakage was associated with reduced expression of tight junction proteins. This is consistent with other studies showing minimal EC death during ECM [17, 39]. Therefore, it is unlikely that CD8 T cells mediate ECM by directly killing brain ECs. Alterations in tight junction protein expression are the most likely explanation for vascular leakage in this model, although the low percentage of EC death we observed could certainly contribute to vascular leakage and surrounding brain pathology.
From the standpoint of vascular pathogenesis, we favor a mechanism whereby CD8+ T cells induce reversible alterations in EC tight junctions (such as claudin-5) that cause cerebrovascular leakage during ECM. A previous study demonstrated that CD8+ T cells can traverse the BBB following recognition of cognate peptide MHC I complexes on the lumen of cerebral ECs [63]. It was discovered more recently that CD8+ T cells can actually use granzyme B in a nontraditional manner to cleave vascular basement membrane [64]. This mechanism allows CD8+ T cells to extravasate across vasculature. Considering that we and others [39] have found parasite-specific CD8+ T cells along CNS vasculature during ECM, it is possible that the cumulative vascular breaks associated with CD8+ T cell extravasation contribute to global breakdown of the BBB. This event would be mitigated in granzyme B knockout mice, which are resistant to ECM [37].
CD8+ T cells could also use IFNγ to remodel the CNS vasculature. Studies using in vitro cultured ECs have shown that IFNγ causes cytoskeletal rearrangement and decreased barrier integrity [65, 66]. During HSV-2 infection, T cell-derived IFNγ has been shown to open the BBB, facilitating antibody access to the CNS [67]. Therefore, PbA-specific CD8+ T cells engaging CNS ECs in an antigen-specific manner could induce barrier openings through IFNγ release. In addition, IFNγ can promote cross presentation of antigen by cerebral ECs [62] as well as production of CXCL9 and 10 [68], chemokines known to attract CD8+ T cells. Thus, it is conceivable that engagement of peptide MHC I complexes by CD8+ T cells on cerebral ECs results in IFNγ release that opens tight junctions and recruits additional parasite-specific T cells. This type of amplification loop localizing primarily to cerebral vasculature would result in a rapid deterioration of the CNS barrier system. Consistent with this theory, deletion of IFNγ significantly reduced accumulation of parasite-specific CD8+ T cells along cerebral vasculature and completely eliminated fatal brainstem pathology.
The precise localization of parasite-specific CD8+ T cells to cerebrovascular ECs led us to consider whether this pattern was unique to the CNS. To address this question, we examined the vasculature of the ear skin as a representative peripheral location. Interestingly, our intravital imaging studies revealed that parasite-specific CD8+ T cells localized along brain but not ear vasculature. Moreover, vascular ECs extracted from the ear showed a significantly reduced expression of adhesion (ICAM-1 and VCAM-1) and antigen presenting (MHC I, MHC II) molecules when compared to cerebral ECs. This is an intriguing observation given that malaria is a systemic disease characterized by circulating iRBCs and inflammatory cytokines. Thus, vascular ECs throughout the body should have access to the same inflammatory mediators. The specific changes in cerebral ECs are best explained by parasite-specific T cell engagement following local antigen presentation. We observed upregulation of MHC I and II molecules on cerebral ECs, which would allow direct engagement by CD8+ and CD4+ T cells, respectively. These two T cell subsets are known to secrete IFNγ upon engagement and amplify the activities of one another during ECM [32]. Although we and others have shown that activation of the CNS endothelium occurs during ECM (Fig 6C) [69], we have provided direct evidence that this is driven by IFNγ(Fig 6D). Because antigen recognition occurs on cerebral ECs, it is likely that T cell-derived IFNγ plays a major role in the activation of the CNS vasculature. However, it remains to be determined why antigen presentation and T cell engagement occurs primarily on cerebral blood vessels and not vasculature residing in a peripheral site such as the ear. Features of the infecting parasite are also likely to contribute to severe disease. In children, one particular var gene product expressed on the iRBC surface to endothelial protein C receptor (EPCR) is associated with CM [70, 71], and brain autopsies of Malawian children who died from CM showed loss of EPCR at sites of sequestered iRBCs [72]. In mice, ECM-inducing PbA differs by only 18 non-synonymous mutations in open reading frames from the NK65 parasite that does not cause ECM, suggesting that these parasite genes may contribute to the unique brain pathology during ECM.
The localization of parasite-specific CD8+ T cells along cerebral vessels [39] does not alone prove that these cells are engaged in cognate peptide MHC I interactions or that these interactions are important for disease development. To demonstrate the specificity of these interactions, we conducted several lines of investigation. We demonstrated previously that CTL division could be advanced by peptide MHC I interactions in the virally infected meninges [55]. Interestingly, we observed a similar pattern of parasite-specific CD8+ T cell division in the lumen of cerebral blood vessels during ECM, suggesting antigen-specific interactions. The specificity of the vascular CD8+ T cell interactions was proven by injecting anti-Kb-SIINFEKL antibodies intravenously during ECM. This increased the velocity of luminal parasite-specific CD8+ T cells, indicating that the interactions with ECs were in fact antigen specific. This conclusion was further supported by the preferential arrest of parasite-specific CD8+ T cells on cerebral vasculature when compared to bystander CD8+ T cells of an irrelevant specificity. Lastly, the functional importance of the interactions was demonstrated by removing MHC I from the stromal compartment through the generation of bone marrow chimeras. Expression of MHC I on the hematopoietic system but not stromal cells such as vascular endothelium significantly reduced parasite-specific CD8+ T cell arrest on cerebral ECs and promoted survival. In vitro studies have shown that human and murine brain ECs have the ability to internalize and cross-present Plasmodium antigen to CD8+ T cells [41, 62]. In fact, ex vivo cultured brain ECs from mice with ECM have the capacity to stimulate CD8+ T cells in an antigen-specific manner. These data collectively demonstrate that CD8+ T cells mediate fatal vascular pathology during ECM via antigen-dependent interactions with cerebrovascular ECs.
While murine and human CM share similar pathologies, the role of CD8+ T cells in HCM remains unknown. Less attention is given to CD8+ T cells during HCM because they are difficult to find in human post-mortem brain samples [73]. It should be noted, however, that we and others [21] have also found it difficult to identify intravascular CD8+ T cells in post-mortem brain samples from mice with ECM. An inability to observe an abundance of these cells histologically does not negate their involvement in HCM. It is unfortunately not possible to examine parasite-specific CD8+ T cells intravitally during HCM (as we have done in mice); however, genetic studies offer some clues regarding their involvement in this disease. For example, resistant vs. susceptibility to HCM has been linked to specific human leukocyte antigen class I alleles [74, 75]. In addition, the CD8 T cell chemoattractant, CXCL10, is a strong biomarker for HCM [76–78], and a genetic polymorphism that elevates CXCL10 levels was linked to an increased incidence of HCM [79]. Importantly, CXCL10 blockade or deficiency is partially protective against development of ECM [25, 26]. It remains to be determined whether CD8+ T cells are definitively involved in the pathogenesis of HCM, but further studies are warranted given the similarities between HCM and ECM.
Without knowing the specificity of disease-inducing parasite-specific T cells, it would be difficult to therapeutically target these cells during CM. Thus, we surmised that a non-specific displacement of T cells from the cerebral vasculature would provide a more effective means to thwart this disease. We demonstrated that late blockade with antibodies against LFA-1 and VLA-4 completely prevented development of ECM, which is consistent with other studies showing the effectiveness of adhesion molecule blockade in this model [17, 53, 54]. However, the mechanism by which this therapy prevented disease was unclear. Using intravital imaging, we observed that anti-adhesion molecule therapy completely disengages CD8+ T cells from the CNS vasculature at the peak of disease without affecting the number of extravascular CD8+ T cells. These data indicate that this disease can be treated by interfering with luminal interactions between CD8+ T cells and cerebrovascular ECs. Natalizumab (anti-VLA-4) and Efaluzimab (anti-LFA-1) are both FDA approved drugs used to treat human inflammatory diseases [80–82], opening the possibility of an expedited intervention in patients with HCM. It would be important to begin treatment in patients with edema prior to the development of cerebral herniation, which would result in irreversible brainstem pathology.
In conclusion, we have provided the first in vivo evidence detailing how CD8+ T cells cause ECM. Parasite-specific CD8+ T cells engage CNS ECs in an antigen-dependent manner, which leads to profound vascular breakdown (associated with alterations in tight junction protein expression), edema, loss of brainstem neurons, and subsequent death. Despite CTL engagement, we observed little evidence of EC death during ECM. We therefore hypothesize that CD8+ T cells use a noncytopathic mechanism to disrupt cerebrovascular EC tight junctions. This signifies that the disease is reversible up to the point when severe edema gives rise to cerebral herniation and death of brainstem neurons (an irreversible event). The reversibility of the disease is demonstrated by the effectiveness of late anti-adhesion molecule blockade. Because many patients with CM likely arrive in the hospital with active BBB breakdown, effective therapies need to target parasite acquisition by ECs and subsequent CD8+ T cell engagement. In addition to anti-malarial drugs and supportive care, consideration should also be given to therapeutics that interfere with T cell function / metabolism [83, 84] or that temporarily displace these cells from cerebral vasculature (e.g. anti-VLA-4 / LFA-1). A reduction of parasite burden in combination with a transient disruption of T cell activity should give the BBB sufficient time to repair and for neurological disease to subside.
C57BL/6J (B6), B6.129S7-Ifntm1Ts/J (IFNγ-/-), B6.129S4-Ccr2tm1Ifc/J (CCR2-/-) B6.129P-Cx3cr1tm1Litt/J (CX3CR1gfp/gfp)[85], B6.SJL-Ptprca Pepcb/BoyJ (Ly5.1), and BL/6-Tg(TcraTcrb)1100Mjb/J (OT-I) mice were purchased from The Jackson Laboratory (Bar Harbor, ME). CX3CR1gfp/gfp, OT-I, and Ly5.1 were then bred and maintained under specific pathogen free conditions at the National Institute of Health (NIH). B6 DbGP33–41 TCR-tg (P14) [86], B6 LysMgfp/+ [44], B6 H-2Kb-/-Db-/-, B6 actin-mCerulean, and B6 actin-YFP were also bred and maintained at the NIH. CX3CR1gfp/+ mice were generated by crossing B6 mice with CX3CR1gfp/gfp mice. Actin-mCerulean and actin-YFP mice were made as described below. YFP+ P14 and mCerulean+ OT-I mice were derived from the following F1 crosses: actin-YFP x P14 and actin-mCerulean x OT-I, respectively. All mice bred in house were confirmed to be on a pure C57BL/6J background by SNP analysis (Charles River).
All transgenic mice were generated by the National Institute of Mental Health (NIMH) Transgenic Core Facility. Transgenic mice expressing monomeric cerulean (mCerulean) under the chicken β-actin promoter were generated as described [87] by first PCR amplifying the entire 717 bp mCerulean coding region from the pCMV-mCerulean vector (Addgene) using primers (fwd: 5’ATATATGAATTCGCCACCATGGTGAGCAAGGGCGAGG3’; rev: 5'ATATATCTC GAGTTACTTGTACAGCTCGTCCATG3') containing EcoRI and XhoI restriction sites. This cDNA was cloned into the same sites following removal of the GFP sequence from the pCAG-GFP vector (Addgene). The resultant plasmid was digested with Sal I / HindIII, and a 2925 bp fragment containing the CMV early enhancer/chicken β-actin (CAG) promoter, mCerulean, and PolyA sequence was prepared for microinjection into the pronuclei of fertilized mouse eggs. Mice expressing the Venus variant of yellow fluorescent protein (YFP) under the CAG promoter were generated in a similar manner. The entire 717 bp coding sequence for YFP was PCR amplified using the aforementioned primers and cloned into pCAG-GFP following removal of GFP with EcoRI and XhoI. A 2925 bp fragment consisting of the CAG promoter, YFP, and a PolyA sequence cut with Sal I / HindIII and prepared for microinjection. To generate all transgenic mice, linearized constructs were injected into C57BL/6J eggs. Following selection of transgene positive founder lines, all mice were backcrossed one additional generation onto C57BL/6J background before intercrossing.
PbA was maintained and used as previously reported [83]. PbA.OVA::mCherryhsp70 (PbA-OVA-mCherry) parasites were kindly provided by C. Janse and S. Khan [45]. PbTg (PbA-OVA-GFP) parasites were kindly provided by W. Heath [38]. All mice were infected intraperitoneally (i.p.) with 106 parasitized red blood cells (pRBCs). Parasitemia was determined by Giemsa-stained thin blood smear or by flow cytometry as described [88]. Mice were monitored daily for neurological symptoms of ECM using a quantitative scale as described [40]. For survival studies, mice were euthanized upon reaching a score of 0. For LCMV infections, mice were injected with 105 plaque forming units of LCMV Armstrong 53b strain i.p.
Mice were i.v. injected with 20mg Evans blue (Sigma) per kg body weight at the indicated time points p.i. After 4 hours, brains were extracted following saline perfusion and then processed as described previously [42].
Brain water content was measured as previously described [19]. Weights of brains removed at indicated time points were compared to the dry weight after overnight incubation at 80°C.
Mice were anesthetized with chloral hydrate and perfused with 4% paraformaldehyde in PBS for H&E analysis, 10% neutral buffered formalin (NBF) for meningeal whole mounts, or 2% NBF in PBS for all other immunohistochemical stains. For meningeal whole mounts, skull caps were removed and incubated overnight in 10% NFB. Following a brief wash, meninges (including dura, arachnoid, and partial pia layers) were carefully removed from the bone with fine-tipped forceps and placed in PBS + 2% fetal bovine serum (FACS buffer) at 4°C and stained with primary antibody overnight. After a brief wash, the meninges were stained with secondary antibodies for 1 hr at room temperature, followed by another brief wash and then a DAPI stain for 5 min. The meninges were placed in mounting medium on a glass slide, spread out and flattened with forceps, and cover-slipped. Brain tissues were extracted following perfusion and incubated either 2 hours for TJ stains or overnight for all other stains in the same solution with which they were perfused. After a brief wash in PBS, brain tissues were incubated in 30% sucrose in PBS for 24 hours. H&E staining was performed by HistoServe, Inc. For all other stains, tissues were frozen in Tissue-Tek optimal cutting media (Thermo Fisher Scientific). A Leica CM1850 cryostat was used to cut 40 μm thick tissue sections for TJ staining or 20 μm sections for all other stains. Primary stains were performed overnight at 4°C, and secondary and tertiary stains were performed for 1 hr at room temperature with 3 x 5 min washes after each stain. DAPI (Sigma) was added for 5 min at RT to label cell nuclei. Following staining, 1 drop of FluorSave Reagent (Calbiochem) was added to each section before addition of a coverslip. The following primary antibodies were used: anti-Claudin-5 Alexa Fluor 488 (4C3C2) (Thermo Fisher Scientific) (1:100), polyclonal anti-laminin (Abcam) (1:200), anti-NeuN (A60) (EMD Millipore) (1:250), polyclonal anti-PECAM-1 (CD31) (EMD Millipore) (1:40), and anti-TER119 (Biolegend) (1:200). All secondary antibodies and staining reagents used for immunohistochemistry were purchased from Jackson ImmunoResearch including anti-rat Alexa Fluor 647, anti-goat Rhodamine Red-X, biotin anti-goat, biotin anti-rabbit, streptavidin Alexa Fluor 488, and streptavidin Rhodamine Red-X. All secondary antibodies were used at a concentration of 1:500 with the exception of streptavidin conjugates, which were used at a concentration of 1:1000. H&E images were acquired using a Nikon Eclipse Ci microscope with 4x/0.2 NA and 40x/0.75 objectives. Fluorescent images were acquired using an Olympus FV1200 laser scanning confocal microscope equipped with 405, 458, 488, 515, 559, and 635 laser lines, 4 side window PMTs for simultaneous 4 channel acquisition, and 4x/0.16 NA, 10x/0.4 NA, 20x/0.75, 40x/0.95, and chromatic aberration corrected 60x/1.4 NA objectives. For cell death analyses, 10–12 0.4 mm2 field images were collected from the brain sections of each mouse. Each field was chosen at random within one of four brain regions: olfactory bulb, cerebrum, cerebellum, and brainstem. Images were analyzed using Imaris 7.6.4 software. Only cells with PI staining throughout the nucleus were counted as dead. To determine co-localization of PI and NeuN, PI+ nuclei were first identified and labeled using the spots function with Imaris software. Each spot was then masked in the NeuN channel to reveal all the NeuN+PI+ co-staining. For tight junction analysis, areas of vascular hemorrhaging were first identified within brain sections of symptomatic mice at d6 p.i. by Evans blue staining. We focused specifically on the frontal cortex, brainstem, and cerebellum. Within those regions, xyz images (each 808,992 μm3) were captured by confocal microscopy from each brain region per mouse. Three to four z-stacks were captured from areas with and without Evans blue leakage. Images from the corresponding brain regions of uninfected mice were also collected. Within each group, we then quantified the intensity of claudin-5 expression on 30–32 blood vessels as described previously [89]. Briefly, using Imaris 7.6.4 software, contours were generated around individual blood vessels based on CD31 staining and used to create a 3D surface. The claudin-5 intensity per unit area of an individual blood vessel within this surface was then calculated as follows: (total # voxels x claudin-5 MFI) / total surface area of vessel.
Mice were injected with 200μg Evans blue dye i.v. for one hour and 50μg propidium iodide (Invitrogen) i.v. for 30 minutes. Mice were anesthetized, perfused, and brain tissue was processed as described above.
Anesthetized mice received an intracardiac perfusion with PBS to remove all blood leukocytes other non-adhered cells. Single-cell spleen suspensions were prepared by mechanical disruption through a 100 mm mesh barrier followed by RBC lysis with Ack lysis buffer (0.15M NH4Cl, 10mM KHCO3, 0.1mM EDTA). Leukocytes were isolated from the meninges, using forceps to gently separate them from the underside of skull cap (exactly as performed for immunohistochemistry whole mounts) followed by enzymatic digestion in 2mg/ml collagenase D (Roche) + 50 mg/ml DNase (Roche) in RPMI for 30min at 37°C. Leukocytes were isolated from the brain as described [90]. EC isolation was adapted from [91]. Briefly, whole meninges and minced brains from perfused mice were enzymatically digested in 0.8mg/ml Dispase II + 0.2mg/ml Collagenase P + 0.1mg/ml DNase (all Roche) in RPMI at 37°C for 15 min with gentle shaking followed by 15 min of mechanical comminution at 37°C. Following digestion, supernatants were isolated and washed. Meningeal ECs were ready for staining but brain EC preps were resuspended in a 40% Percoll (GE Healthcare) gradient in HBSS and centrifuged to remove excess myelin.
Surface staining and Fc blocking was performed as described [90]. Dead cells were excluded from the analysis by using the LIVE/DEAD fixable Blue Cell Staining kit (Invitrogen). The following antibodies and reagents from BioLegend were used: CD4 PE (RM4-5), CD8a APC (53–6.7), CD8β.2 FITC (53–5.8), CD11b Brilliant Violet 605 (M1/70), CD31 PE/Cy7 (390), CD45.1 AlexaFluor 647 (A20), CD45.2 Alexa Fluor 700 (104), CD45.2 FITC (104), CD54 Alexa Fluor 488 (YN1/1.7.4), CD106 biotin (429), Gr-1 PE (RB6-8CJ), H-2Db Alexa Fluor 647 (KH95), H-2Kb PerCP/Cy5.5 (AF6-88.5), IAb/IEb Pacific Blue (M5/114.15.2), Ly6C APC/Cy7 (HK1.4), Ly6G PE (1A8), Streptavidin Brilliant Violet 605, Thy1.2 Alexa Fluor 700 (30-H12), and corresponding isotype controls. To identify PbA-specific CD8+ T cells, single cell suspensions were first stained with H-2Db-SQLLNAKYL+ pentamers (ProImmune) in PBS + 10% BSA for 10 min at room temperature before surface staining as described above. PbA-specific CD8+ T cells were also identified by IFNγ production following in vitro stimulation with SQLLNAKYL peptide. Briefly, 2 x 106 splenocytes from the denoted mice were incubated with 1μg/ml SQLLNAKYL peptide (AnaSpec), 100U/ml IL-2 (NIH), and 10μg/ml BFA (Sigma) in RPMI complete media at 37°C for five hours. To detect intracellular IFNγ production, single cell suspensions were surface stained as described above, treated with cytofix/cytoperm (BD), and then stained intracellularly with anti-IFNγPE/Cy7 (XMG1.2) (BioLegend). Samples were acquired using an LSRII flow cytometer (BD), and data were analyzed using FlowJo software version 9.7.2 (Tree Star).
All antibodies used for cell depletion and blocking assays were purchased from BioXcell. Mice were depleted of neutrophils by injecting 500 μg of anti-Ly6G (clone 1A8) i.p. on days -1 and 3 p.i. Mice were depleted of CD4 or CD8+ T cells by i.p. injection of 500 μg (clone GK1.5) or 200 μg anti-CD8 (clone YTS 169.4), respectively, on d4 p.i. Cell depletion efficiency was calculated in the blood using the following formula: 100 –((frequency of targeted cell population of a mouse / the average frequency of the targeted cell population of the 4–5 untreated mice) x 100). For adhesion molecule blocking assays mice were treated with a combination of 500 μg anti-LFA-1 (clone M17/4) and 500 μg anti-VLA-4 (clone PS/2) i.v. at the indicated time points. For cognate peptide-MHC blocking assays, mice were treated with 680 μg anti-Kb-SIINFEKL (clone 25-D1.16) or mIgG1 isotype control (clone MOPC-21) antibodies i.v. at the indicated time points.
Mice were seeded i.v. with 104 mCerulean+ OT-I or YFP+ P14 CD8+ T cells purified from the splenocytes of naïve transgenic mice using a CD8 negative selection kit (Stem Cell Technologies). To compare activated PbA-specific and non-specific CD8+ T cell responses in the same mice, activated YFP+ P14s were first purified from the splenocytes of previously seeded mice using a CD8 positive selection kit (Stem Cell Technologies), 8 days following infection with LCMV. Symptomatic PbA-infected mice previously seeded with mCer OT-1 cells were then i.v. injected with 5x106 activated YFP-P14s i.v.
Mice were anesthetized and thin skull windows were prepared as previously described [42, 43]. Mice were injected with 5 μl Qdot655 (BD) or 50 μg Evans blue where indicated to visualize blood vessels. 4D datasets were acquired using an SP5 two-photon microscope (Leica) equipped with two Mai Tai HP DeepSee lasers (SpectraPhysics), an 8,000-Hz resonant scanner, a 20×/1.0 NA objective, an NDD4 detector array, and a custom-environment chamber. Simultaneous excitation and detection of multiple fluorophores was achieved using custom dichroic mirrors (Semrock) and by tuning one laser to 905 nm and the other to 990 nm. All imaging studies focused on the meninges and superficial neocortex. Imaging data were processed with Imaris 7.6.4 software. A surface was created for each CNS blood vessel and PbA-specific CD8+ T cells located within that surface (vessel lumen) were quantified, whereas T cells residing outside the surface (perivascular or parenchymal) were quantified separately. Mean track velocities (μm/min), arrest coefficients (proportion of time a cell spent arrested <2 μm/min), and arrest duration (the total amount of time a cell spent arrested <2 μm/min) were calculated for all individual luminal PbA-specific CD8+ T cell tracks using Imaris 7.6.4 and T Cell Analyzer software (TCA 1.7.0; Strathclyde Institute of Pharmacy and Biomedical Sciences). All time lapses used for these analyses were at least 40 minutes in length.
BM was harvested from femurs and tibias of Ly5.1 mice and 5x106 BM cells were i.v. injected into recipient mice following a lethal irradiation dose of 900 RAD. Mice received antibiotics in drinking water for 4 weeks following irradiation and were allowed 8 weeks to fully reconstitute bone marrow and donor peripheral cells.
Statistical analyses for data were performed using a Student’s t test (two groups) or ANOVA (more than two groups) in Prism 6 (GraphPad Software). Groups were considered statistically different at a p value of <0.05. All data are displayed as the mean ± SD.
This study was carried out in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The protocol was approved by the NINDS Animal Care and Use Committee (Protocol Number: 1295–14).
|
10.1371/journal.pcbi.1002646 | The Landscape of the Prion Protein's Structural Response to Mutation Revealed by Principal Component Analysis of Multiple NMR Ensembles | Prion Proteins (PrP) are among a small number of proteins for which large numbers of NMR ensembles have been resolved for sequence mutants and diverse species. Here, we perform a comprehensive principle components analysis (PCA) on the tertiary structures of PrP globular proteins to discern PrP subdomains that exhibit conformational change in response to point mutations and clade-specific evolutionary sequence mutation trends. This is to our knowledge the first such large-scale analysis of multiple NMR ensembles of protein structures, and the first study of its kind for PrPs. We conducted PCA on human (n = 11), mouse (n = 14), and wildtype (n = 21) sets of PrP globular structures, from which we identified five conformationally variable subdomains within PrP. PCA shows that different non-local patterns and rankings of variable subdomains arise for different pathogenic mutants. These subdomains may thus be key areas for initiating PrP conversion during disease. Furthermore, we have observed the conformational clustering of divergent TSE-non-susceptible species pairs; these non-phylogenetic clusterings indicate structural solutions towards TSE resistance that do not necessarily coincide with evolutionary divergence. We discuss the novelty of our approach and the importance of PrP subdomains in structural conversion during disease.
| Prion Proteins (PrP) cause a variety of incurable TSE diseases, and are among a small number of proteins for which large numbers of NMR ensembles have been resolved for sequence mutants and diverse species. Here, we perform a comprehensive PCA study to assess conformational variation and discern the landscape of the PrP structural response to sequence mutation. This is to our knowledge the first large-scale analysis of multiple NMR ensembles for a specific protein, and the first study to perform a multivariate PCA on the native globular structures of PrP. We conducted exhaustive PCA on three PrP subsets: human and mouse subsets that include structures of sequence mutants, and the set of wild-type PrP (16 PrP species). PCA shows that different non-local patterns of variable subdomains arise for different pathogenic mutants. These subdomains may thus be key areas for initiating PrP conversion during disease. Furthermore, we observed that some evolutionarily divergent species that are non-susceptible to TSEs have surprising structural similarities in their PrPs. We discuss the novelty of our approach with respect to prions, and the advantage of this analysis as a fast, reliable starting point to identify interesting domains that may warrant further experimental and computational analysis.
| The extraordinary conformational change witnessed between the normal, non-pathological prion protein, PrPC, and its virulent pathological form, PrPSC, in which the latter acquires substantial β-sheet content, is a significant contributor to the role this protein plays as an agent of many incurable Transmission Spongiform Encephalopathies (TSEs). Such diseases, including human Creutzfeldt-Jakob Disease (CJD) and Bovine Spongiform Encephalopathy (BSE), are caused by the misfolding and subsequent aggregation of PrPSC to produce amyloid fibrils, highly ordered and distinct β-sheet-rich molecular aggregates [1], [2]. The PrP protein is a 208 residue protein (residues 23–230, hPrP numbering) composed of a largely disordered N-terminal tail (23–124) and a C-terminal globular domain (125–231), in addition to two signal peptides (1–23, 232–253) [3], [4]. The globular domain contains three α-helices (H1,H2,H3) and two anti-parallel β-sheets (S1,S2). Globular domains of multiple PrP species have been resolved to develop an understanding of PrP structures in relation to TSE-susceptibility, and discern subdomains of the protein that are involved in the PrP conversion process [4], [5], [6], [7], [8], [9]. The S2-H2 loop and H2-H3 regions, for example, demonstrate structural plasticity in pathogenic PrP and are proposed to be involved in the conversion process, making them candidate sites for transmissibility studies and potential target sites for drug design [10], [11], [12], [13], [14], [15]. The prion protein is one of few proteins with a large number of pathogenic mutants, and the increasing availability of these structures in the protein databank (PDB) provides ample material for a multivariate analysis of structural plasticity of PrP domains.
Principal Component Analysis (PCA) [16] is a dimensionality reduction technique that can be used to analyze protein structures by reducing variation observed within 3D atomic coordinates of the protein structures. PCA has been used on several protein families to analyze key regions of interest, including ligand-binding sites and cavities [17], [18], receptor sites [19], catalytic subunits [20], as well as large-scale analysis of whole proteins [21]. Most interesting is the recent application of PCA towards modeling protein flexibility computationally, and characterizing structural variation of protein domains [22], [23]. Identifying structural plasticity within protein domains is especially advantageous for proteins involved in conformational diseases, such as amyloid-forming proteins.
In this work, we perform an exhaustive PCA analysis on the tertiary structures of PrP globular proteins to discern PrP subdomains that exhibit conformational plasticity in response to pathogenic point mutations and clade-specific evolutionary sequence mutation trends; these subdomains may thus be key areas for initiating the conversion of PrPC to PrPSC. To our knowledge, this is the first PCA study on native globular structures of PrP, using NMR ensembles, and without relying on structures generated from protein dynamics methods. We focus our analysis on three subsets of PrP, human and mouse PrP subsets that include structures of sequence mutants, and the set of wild-type PrP globular proteins (representing 16 PrP species). From this analysis, we identify five conformationally variable subdomains of PrP whose relative importance changes for different pathogenic mutations and species groupings. Also, PCA indicates that PrPs exhibit a marked non-phylogenetic clustering, with some notable divergent pairs of species that are non-susceptible to TSEs. We discuss the implications of these results for the conformational basis of TSEs.
PCA was conducted on the NMR ensembles of 11 human wildtype, variant and mutant prion proteins (230 models in total), to examine major conformational changes between the structures and map them onto a lower (mostly 2-dimensional) space. The resulting eigenvalue contribution of PCA shows that 65% of the total mean-square displacement of atom positional fluctuations was captured in the first three components (Figure 1C).
Plotting hPrP structures onto the two most significant principal components (PC1 and PC2) characterizes conformational relationships between the hPrP structures that are reflective of human prion TSEs. Four major conformational clusters have been observed, of which the largest cluster (encircled in the black oval in Figure 1A) corresponds to PDB structures of WT proteins, as well as hPrP artificial variant structures [PDBs 1E1G, 1E1P, 1E1U, 1H0L] that maintain a similar structure to WT PrPs (mPrP, shPrP) [24], [25]. For each of the remaining three clusters, each cluster is composed of the models of the NMR ensemble representing the PDB structure of each of the human TSE diseases of GSS (red oval) [PDB 2KUN] [26], FFI (blue oval) [PDB 2K1D], and CJD (green oval) [PDB 1FO7] [27] (Figure 1A, 1B). These four clusters, as observed by projection of the hPrP structures onto PC1 and PC2 (Figure 1A), as well as PC1 and PC3 (Figure 1B), indicate that these principal component projections facilitate the discrimination of key, pathogenic mutant structures that reflect PrP diseases. Interestingly, such projections also highlight variation between models within an NMR ensemble, as is clearly demonstrated for the structure 2K1D (encircled in blue in Figure 1A), whereby an additional hierarchical cluster is introduced for some models (model numbers 8, 14, 16, 20, encircled in a dashed brown oval in Figure 1A) which cluster further away from the 2K1D ensemble along PC1 (Figure 1A, Figure 1B). This contrasts with other NMR ensembles whose models remain tightly clustered together along the PCs, such as 1FO7 (encircled in green in Figure 1A).
The contribution of each residue in hPrP to each of the first three PCs is displayed, whereby the height of each bar indicates the maximum atomic displacement of each residue for a given PC, and regions of increased displacement highlight structurally variable subdomains in the hPrP structures (Figure 2A, 2C–2E). The mutant structure ensembles are separable on the conformer plots (Figure 1) because of distinct patterns of variable subdomains observed in the residue contribution plot (Figure 2A). The variable subdomains captured by PC1 include the S2-H2 loop and the C-terminal end of H3. PC2, which contributes to the large separation between the FFI and GSS clusters on the conformer plot (Figure 1A), is characterized by concerted structural variability of the H2-H3 loop, the N-terminus of the globular domain, and S1. The remaining variations captured by PC3 include the S1-H1 loop, and increased displacement of the S2-H2 loop region witnessed in PC1. In total, 5 variable subdomains have been identified: the N-terminal region of the globular domain and S1, the S1-H1 loop, the S2-H2 loop, the H2-H3 loop, and the C terminus of H3 (Figure 2B). Strikingly, these subdomains of structural variation are not localized to the variant or mutation spots of the protein, which reflects on the nonlocal changes in the protein that are induced by these highly localized substitutions (Figure 2B).
For comparison, we also performed a PCA analysis just on the structural variation observed in the WT PrPs (totaling 4 NMR ensembles), while excluding NMR ensembles of mutant and variant PrP structures. The resultant residue contribution plot indicates that all five subdomains of concerted variation contribute to PC1 of the WT dataset (Figure 3A), implying that they share equal degrees of importance in representing variance between the structures (PC1 captured 30% of the variance of the dataset) (Figure 3B). Intriguingly, displacement of the H2-H3 loop and the C terminus of H3 are not readily observed in PC2, but are observed in PC3. The lack of additional clustering between the NMR ensembles in PC3, except for the dispersion of models within each NMR ensemble, suggests that these subdomains might play a greater role in discerning conformational changes between models of the NMR ensembles (not shown). Conversely, the N terminus and the S1-H1 loop are readily observed in PC1 and PC2, but not in PC3, showing that these regions play a greater role in separation of the NMR ensembles, instead of inter-model variation.
To check which subdomains vary in a mutant-specific way, we performed three separate analyses, each analysis consisting of the set of WT and variant hPrP structures (encircled by the black oval in Figure 4A) and an NMR ensemble from each of the CJD, FFI, and GSS mutant structures (Figure 4B–D). The resultant conformer plots indicate that the pathogenic mutant structures are successfully separated from the WT and non-pathogenic hPrP structures (Figure 4B–D). Comparison of residue contribution to each PC indicates that the C-terminus of H3, as well as the S2-H2 loop, differentiate the mutant structures for all analyses, as both subdomains appear in PC1 (Figure 4B–D). This observation is reinforced by comparison to the residue contribution plot of the WT, variant, and mutant hPrP structures (Figure 4A). The remaining subdomains representing the N terminus, S1-H1 loop and H1, and the H2-H3 loop display different levels of importance that are reflective on each of the mutant structures. For example, the H2-H3 loop is strong contributor to conformational separation of the CJD mutant structure, as it appears in PC1 in the residue contribution plot (Figure 4B), compared to the FFI mutant where it appears in PC3 (Figure 4C). Similarly, the S1-H1 loop and N terminus of H1 exhibit greater importance in differentiating the GSS mutant structure (Figure 4D), as they appear in a later PC for the FFI and CJD structures (Figure 4B–C). To ascertain our observations, we calculated the residue difference profile between each of the datasets in (Figure 4B–D) with hPrP WT and variant dataset (black oval in Figure 4A) for PC1 (Figure S1). The resultant plots (Figure S1) indicates the residue contribution that is specific to each of the hPrP mutant structures, from which we confirm our observations that the S2-H2 loop exhibits the greatest conformational perturbation for all three mutant structures, and that the H2-H3 loop is clearly important for structural differentiation of the CJD mutant (Figure S1).
In aggregate, these PCA analyses succeed in delineating and ranking structural subdomains in terms of their relative importance for different pathogenic mutants.
We conducted PCA analysis on a set of 14 wildtype, variant, and mutant mouse PrPs NMR ensembles (280 models in total) to examine structural differences between mPrP structures and compare these changes to hPrP. Aside from WT mPrP [PDBs 1XYX, 2L1H, 2L39], 9 PrPs contain mutations in the S2-H2 region (between residues 166–175), and 2 PrP structures [PDBs 2KFM, 2L1K] contain mutations at the C-terminus of H3 (Y255A and Y226A). PCA analysis of mPrP including 2KFM and 2L1K reveals a prominent concerted variation of the C-terminus of H3 that far exceeds any other atomic displacement in the protein, for all three PCs (Figure S2). One might argue that 2KFM and 2L1K, as the only two structures with conformational differences in H3, are “conformational outliers” that contribute to the displacement of the H3 region in all PCs and overshadow structural differences of the H2-H3 loop. To test this hypothesis, we re-ran the analysis without 2KFM and 2L1K, such that the mPrP dataset consisted only of the WT and variant structures and those with mutations in the S2-H2 loop. Contrary to our expectations, the observed pattern of atomic displacements indicates that the H3 subdomain, in addition to the N terminus of the proteins, remains responsible for conformational variation.
PCA was conducted on NMR ensembles of 16 species of WT PrP (21 PDB ensembles corresponding to 420 models in total) (Figure 5). Among the species studied, 8 species (mouse, bovine, human, hamster, cat, pig, elk, bank vole) are known to develop TSEs, and 7 species (dog, horse, rabbit, chick, turtle, frog, and wallaby) are “TSE-non-susceptible”, taken collectively here to refer to PrP species that are experimentally proven to be resistant to TSEs or for which TSEs remain undetected. In our analysis, sheep is the only species which has been considered in both categories, as sheep with the H168 polymorphism [PDB 1XYU] are TSE-susceptible, but those with the R168 variant [PDB 1Y2S] are highly resistant to disease [28]. PCA successfully clusters many of the TSE-non-susceptible species from TSE-susceptible ones, as indicated by the conformer plots (Figure 5A–C). PC1 separates chicken (chPrP) and turtle (tPrP) from the rest of the species, such that they form their own subgroup (Figure 5A). This is to be expected since they are divergent species evolutionarily. Detailed analysis of residue contribution in this PC indicates that the H2-H3 loop undergoes a significant displacement relative to the rest of the protein (Figure 5G–H). However, unexpectedly from an evolutionary point of view, PC2 also contributes to the clustering of the two TSE-non-susceptible species, frog and rabbit (Figure 5A–B) (when n = 3 in hierarchical clustering). Residue contribution to PC2 characterizes the concerted maximum displacement of the S2-H2 loop and the H1 helix (Figure 5G–H). With the exception of an additional clustering for pig that is introduced in PC3 (Figure 5B–C), analysis of the residue contribution to PC3 does not introduce any newer subdomains than those identified in PC1 or PC2. Thus, the first two PCs are sufficient in describing the range of structural differences between PrP species.
As the H2-H3 loop is longest in chPrP compared to other PrP species [5], [15], we wished to assess whether the concerted displacement of the H2-H3 loop in PC1 is the biased result of major conformational differences in chPrP. To this end, we performed a PCA analysis on all the WT PrPs without chPrP (Figure 5D–F). Despite the removal of chPrP, the dominant feature described by PC1 remains the displacement of the H2-H3 loop, followed by the displacement of the S2-H2 loop and H1 in PC2 (not shown). Similarly, no additional regions of displacement are witnessed in PC3. With respect to conformational clustering, removal of chPrP has decreased the amount of variation observed in the first 3 PCs (46% compared to 51% with chPrP). Conformational clusters of the dataset without chPrP indicate that the turtle, frog, rabbit, and cat species cluster further away from the TSE-susceptible species (Figure 5D–F), and the clustering of the NMR ensemble for pig PrP is also observed in PC3 (Figure 5E–F). However, an additional clustering of the sheep resistant R168 polymorphism (PDB 1Y2S) is observed at PC3, while the TSE-susceptible sheep polymorphism H168 (PDB 1XYU) remains closely clustered with the TSE-susceptible PrPs (Figure 5E–F). In summary, we demonstrate that our PCA analysis detects major “structural signatures” for PrPs of different evolutionary groups, and highlight PrP subdomains that are worthwhile to explore in TSE-transmissibility studies.
PCA analyses of the entire WT dataset (Figure 5A–C) raises the following question: does the structural variation in these analyses reflect upon species evolutionary relationships, and is there discernible clustering that reflects TSE susceptibility and non-susceptibility/resistance? Analysis of WT PrP reveals that distantly-related, non-mammalian species (frog, chicken, and turtle) form separate clusters from the mammalian cluster in the conformer plot (Figure 5A). To discern the behavior of PrP subdomains in the evolutionary and structural separation of a large subset of closely-related species, we ran PCA on a set of 13 mammalian TSE-non-susceptible and TSE-susceptible PrP NMR ensembles. The resultant conformer plots (Figure 6A–C) show that rabbit and pig PrP structures quickly separate from the remaining PrPs. Analysis of residue contribution to the PCs indicates a different pattern of “subdomain importance” that differentiates between the mammalian PrPs (Figure 7A), compared to the complete WT species set that includes non-mammalian PrPs (Figure 5G). The residue contribution plot of the mammalian PrPs (Figure 7A) indicates that the C-terminus of the H3, as opposed to the H2-H3 loop, exhibits the largest atomic displacement in PC1, while the remaining four subdomains appear in PC2 and PC3.
We compared subdomain displacement of the mammalian dataset (n = 13 total species) (Figure 7A) to subsets of TSE-non-susceptible mammals (n = 5 species, including Sheep R168 variant) (Figure 7B) and TSE-susceptible mammals (n = 9 species, including Sheep H168 variant) (Figure 7C). With respect to the combined set of mammalian and non-mammalian TSE-non-susceptible PrP structures (presented in Figure S3, part B), removal of the non-mammalian PrPs from that set shifts subdomain importance from the H2-H3 loop (Figure S3, part B) to the C-terminus of H3 in the TSE-non-susceptible mammalian dataset (Figure 7B), such that the pattern of conformational variation and subdomain importance is similar to the total WT mammalian dataset (Figure 7A). Notably however, H1 and its flanking loops still exhibit strong displacement at PC2 in both TSE-non-susceptible residue contribution plots (Figure 7B, Figure S3), which suggests that for all TSE-non-susceptible species, including or excluding non-mammals (Figure 7B, Figure S3), H1 represents a large percentage of conformational variation within that dataset.
It is interesting to note that PCA analysis of mammalian PrPs (n = 13), and TSE-non-susceptible mammals (n = 5), indicates that TSE-non-susceptible mammals (ex: horse, wallaby, rabbit) exhibit a “structural differentiation”, such that they cluster at the periphery of the conformational space away from TSE-susceptible mammals (Figure 6A, 6D–F). This indicates different structural solutions towards resistance that don't necessarily coincide with evolutionary divergence. This is clearly demonstrated by examination of a PC-based cluster dendrogram of all of the 16 PrP NMR ensembles (420 models) under study and of a neighbor-joining tree for the PrP sequences of the 16 species (Figure S4); horse and wallaby, for example, are closely clustered together in the PC-based dendrogram, even though they are evolutionarily divergent species.
Five subdomains displaying structural plasticity in PrP have been identified in NMR ensembles of hPrP, mPrP, and WT datasets (Figure 8). The pattern of concerted displacement of these subdomains for all three PCs, for each of the datasets, is summarized (Table 1).
We have conducted exhaustive PCA analyses on a large set of PrP globular structures, as well as several subsets representing particular species of interest (human and mouse), or groupings which hold biological significance (TSE susceptibility or non-susceptibility); from these analyses we identified five conformationally variable subdomains in PrP undergoing varying levels of correlated movements in all datasets, and which are thought to be significant for the PrP conformational conversion process that underlies prion disease. We have demonstrated the benefits of exploring prion protein conformational variation using PCA, and the importance of the identified subdomains towards understanding the PrP conformational conversion process.
One obvious concern with the PCA analysis is that increased structural plasticity in the loop regions and protein termini would bias selection toward these regions, and outweigh identification of other regions in ordered, structured subdomains of the protein. However, for several of our PCA runs, structural variation within the protein datasets does not directly result from increased displacement in protein termini (the WT PrP set is an obvious example). In datasets where termini play a significant role in conformational differentiation of the structures, this variation is supported by weakened NMR definition in the protein (for example, hPrP and its variants vary in length and definition of residues 220–228 of H3 [24]). Additionally, our analysis identified structural variation within regions with repetitive secondary structures (ex: S1 and H1). Finally, for all PrP datasets we considered, structural plasticity of the loop regions has only been identified for selected portions of the loops, not the entire loop. For example, we only identify the latter half of the S1-H1 loop as conformationally variable in the hPrP and WT datasets, but the first half of the loop (residues 134–138, hPrP numbering) is relatively invariable.
To our knowledge, the presented work is the first study to perform a multivariate PCA on the native globular structures of PrP. Generally, few publications on prion structural biology have utilized multivariate analysis to comprehend the structural complexity of this protein and model protein flexibility computationally, with the exception of a couple that have conducted PCA of MD simulations to determine protein flexibility [10], [15]. Strikingly, some of the structurally variable subdomains we have identified (e.g., the S2-H2 loop) are “complementary” to the ‘domains of collective movement’ (rigid domains) identified by these studies [10], [15]. Much of the computational analysis on PrP structures, however, involves the use of molecular dynamic simulations [13], [29], [30], [31], or longer dynamic simulations such as normal mode analysis (NMA) [32]. Such methods, as in the case with molecular dynamics, are continuously challenged by their computational expense, involvement of complex force fields, size of the query protein, and long time spans required to run the simulations [23], [33]. Comparatively, our PCA analysis on native PrP, without the reliance on any structures generated by long- or short-term dynamics studies, succeeds in identifying key regions that may be involved in the conversion process and which have been previously highlighted in MD and NMA studies [10], [14], [15], [29], [31], [32]. Accordingly, PCA is advantageous in rapid identification of important subdomains in PrP while saving computational time and effort, and may be used as starting point to identify key subdomains that can be further analyzed over longer time scales using protein dynamics.
This study is the first large-scale analysis of multiple NMR ensembles for a specific protein, and it poses unique challenges for principal component data analysis and interpretation. While static X-ray structures only provide a snapshot of potential motions of proteins, ensemble analysis of multiple X-ray structures may provide insight into the conformational changes of proteins and elucidate structural mechanisms of biological activity. The abundance of X-ray models for several protein families in the PDB facilitated PCA analysis of these proteins [34], [35], [36], and development of computer tools for systematic multivariate analysis of X-ray ensembles is gaining increasing importance [37], [38]. In the case of the PrP family, however, few X-ray structures of PrP exist in the PDB (<40% of all deposited PrP structures in the PDB), and even fewer structures represent globular PrP (as opposed to peptide segments, for example). For this analysis, we could only identify 11 relevant crystal structures, as opposed to the 41 NMR structures we have selected. Use of a reduced sample size based on X-ray structures severely limits the number of PCA analyses that could be performed on PrP subgroups and produces inaccurate estimates of collective motions in PrP. Structural analyses with multiple NMR ensembles, while increasing the sample size multi-fold, poses a considerable analytical challenge as two sources of structural variation need to be considered: variation of models within an ensemble, and variation between ensembles. As variation between ensembles is expected, and sought for by PCA, eliminating variation within the ensemble remains an issue. To reduce the effect of inter-model variation, we have opted to use entire NMR ensembles, as random selection of any model may inadvertently introduce biases if the selected model is a structural outlier within the ensemble. Additionally, where selection of ensemble representatives was warranted, we used OLDERADO [39] to select for models representing the largest central core of the NMR ensemble, i.e., the “average” of the ensemble. Accordingly, PCA on the NMR ensembles allowed for identification of structural differences between NMR ensembles, but also successfully outlined inter-model differences within the ensembles.
Our PCA analysis has indicated that different subdomains are variable in different pathogenic mutants of PrP structures. Our PCA analysis has succeeded in providing a ranking for these subdomains that correlates with pathogenicity. In hPrP for example, by comparing displacements in residue contribution plots of the combined hPrP dataset and the mutant hPrP subset, we have demonstrated that the S2-H2 loop (residues 165–175) and the C-terminus of H3 (residues 220–228) are the first subdomains to differentiate pathogenic and nonpathogenic PrP structures. The S2-H2 loop is one of the most affected regions of PrP in terms of structure and flexibility, and may influence stability of PrP during PrPC→PrPSC conversion [29]. Mutant hPrPs exhibit weakened hydrophobic intramolecular interactions between this loop and the H3 helix, compared to native hPrP [29]. Weakened interactions between Y169-F175-Y218 have been reported for the E200K and Q212P mutants, as well as M166-Y225 π-stacking interactions [29]. The mutual orientation of aromatic residues in S2-H2 loop is affected by increased solvent exposure of Y169 in mutant PrP, yielding higher flexibility and greater solvent exposure of these hydrophobic residues compared to the observed stabilized aromatic interactions of Y163-Y169-F175 in the native hPrP [26], [29], [40]. As weakened hydrophobic interactions of the S2-H2 loop also weaken the interactions with H3 helix, it is not a surprise that the C-terminus of H3 (residues 220–228) is equally important in differentiating wildtype from mutant PrPs. The C-terminus of H3 is observed to gain flexibility as a result of a breakdown in salt bridges between the H2 and H3 helices [13], [14], [29]. Interestingly, our conformer plot of all hPrP structures succeeds in separating the E200K, Q212P, and other pathogenic mutants displaying similar behavior (ex: D178N) from the remaining hPrP, reflecting on the specificity of the PCA in differentiating the structures by the plasticity of S2-H2 loop. This is particularly intriguing, as the S2-H2 loop (residues 166–170, hPrP numbering) and the C-terminal of H3 (residues 215–230) form a solvent-accessible disease-linked epitope for monoclonal antibody, and may serve as a recognition area for “protein X” involved in the conversion process [41]. Additionally, the S2-H2 loop has been observed to exhibit varying levels of flexibility within TSE-susceptible species, and is rigid in TSE-non-susceptible species, making it a prime candidate for PrP transmissibility studies [10], [15].
PCA of WT PrP structures has summarized areas that change concertedly over evolution, e.g. the H2-H3 loop. This was a particularly interesting result, as the H2-H3 loop is longer for chicken (the most outlying protein structure) than in any other species, and compared to other TSE-non-susceptible PrPs, is a flexible subdomain within that protein [5]. Generally, the structural variation observed does not correlate phylogenetically with organismal speciation. Intriguingly, the two most ‘non-phylogenetic’ clusterings are for TSE-non-susceptible species, rabbit (a placental mammal) clustering with frog, and horse (a placental mammal) clustering with wallaby (a marsupial). This is evidence for evolutionary ‘re-visiting’ of different structural solutions to TSE resistance, in different evolutionary lineages. PCA profiles clearly show that different PrP subdomains vary amongst the TSE-susceptible and TSE-non-susceptible mammalian subsets. Also, the NMR ensembles for TSE-non-susceptible mammalian PrP structures tend to be peripheral on the PCA conformer plots, and overall, show a greater structural diversity, suggesting that TSE susceptibility may be linked to a greater degree of PrP structural similarity between infecting and receiving species/organisms.
To conclude, we performed an exhaustive analysis of PrP globular structures to identify subdomains of conformational change, as these subdomains of structural plasticity may contribute to PrP conversion and misfolding, and ultimately, to TSEs. Our PCA analysis succeeds in ranking these subdomains of as a function of species variation and disease-susceptibility. This is the first study to perform a multivariate PCA analysis on the native structures of the globular PrP, and one of very few studies to conduct PCA on NMR ensembles to detect biologically significant conformational variability in proteins and protein families. Our identified subdomains within PrP for all datasets studied compare favorably against those identified in computationally-intensive dynamic simulations and experimental data, suggesting that PCA analysis of the native structures can be used as a fast, reliable starting point to identify regions of interest that may warrant further analysis by computational and experimental methods.
We collated all known PrP structures in the RCSB Protein Data Bank [42], by searching for all proteins within the ‘Prion-like’ family and superfamily of SCOP [43], proteins which match the architecture of the Major Prion Protein as specified in CATH [44] (Mainly alpha, orthogonal bundle, 1.10.790), as well as searches based on PFAM [45] Hidden Markov Models (HMMs) representing the Prion-like protein Doppel [PF11466], Prion/Doppel alpha-helical domain [PF00377], and the major prion protein bPrP-N terminal [PF11587]. These searches yielded a total of 112 prion PDB structures, from which only PrP globular domains were selected. The list of PrP globular domains was further refined to exclude dimers (ex: [PDB 3O79]), domain-swapped structures (ex: [PDB 1I4M]), and pdb models representing the average minimized structure of an NMR ensemble (ex: [1E1J, 1E1S, 1E1W, 1FKC, 1HJM, 1QLX, 1QM0, 1QM2] in human PrP, [1AG2] in mouse PrP, and [1DWY], [1DX0] in bovine PrP). A total of 41PDB structures, all of which are NMR-derived, were selected for analysis.
The analysis was performed on three separate cohorts of PrP globular proteins: (i) all human PrP (hPrP), (ii) all mouse PrP (mPrP), and (iii) all wildtype (WT) PrP, representing 16 species of PrP.
For each of the datasets studied, an analysis was performed all models of the PDB NMR Ensembles, as well as the subset of representative models for each ensemble, identified using EBI OLDERADO [39].
For each dataset being studied, a multiple sequence alignment of all structures, based on ATOM residues, was generated using EBI MUSCLE [46]. This alignment and the corresponding structures were used as input in the Bio3D [38] package within the R statistical program [47]. Iterated rounds of structural superposition of PrP structures by Cα atoms, ignoring gap/insertion regions and missing residues, was performed to identify invariant core residues of PrP with a 1°A core cutoff. The structurally invariant core was used as a reference frame for structural alignment of the PrP NMR models, and Cartesian coordinates of the aligned Cα atoms were used as input for principal component analysis (PCA).
PCA maps high-dimensional data into fewer dimensions by a linear transformation [16], and has been employed in several studies to provide insight into the nature of conformational changes within proteins and protein families. In this study, PCA finds axes along which the high-dimensional ensemble of PrP protein structures can be best separated. The input is a coordinate matrix, X, composed of N by P dimensions, where N represents the number of structures and P represents three times the number of residues [23], [36], and each row of the matrix corresponds to the Cα coordinates of each structure. PCA is based on diagonalization of the covariance matrix, C, with elements Cij built from X as follows:
where
i,j = all pairs of 3N Cartesian coordinates
< > = average over N atoms under consideration
Principal components (orthogonal eigenvectors) describe axes of maximal variance of the distribution of structures, and eigenvalues provide the percentage of variance (total mean square displacement) of atom positional fluctuations captured along each PC. Projecting PrP structures onto the conformational subspace defined by the largest PCs produces a low-dimension “conformer plot” which allows for the identification of dominant conformational changes and the characterization of inter-conformer relationships [38]. Additionally, the relative displacement of each residue described by a given PC can be represented in a “residue contribution” plot. Collectively, both plots allow for the identification of “conformationally variable subdomains” that are responsible for conformational clustering of the PrP structures, and which contribute to the structural variation observed in the datasets. These subdomains represent the largest segments of structural plasticity within the prion protein, making them candidate sites in the PrP conversion process.
Variation within models of an NMR ensemble poses a challenge for PCA analysis: how does the selection of a particular model influence the structural variation of a dataset? To test the extent to which inter-model variation within an NMR ensemble influences identification of variable PrP subdomains, we conducted PCA analyses on randomly selected NMR models within the hPrP and mPrP datasets. Using the total hPrP (11 PDBs) and mPrP (14 PDBs) datasets listed above, an NMR model was selected at random from each of the NMR ensembles within that set, creating a subset of ‘representative’ NMR models for all the structures. The process was repeated 50 times and PCA was performed on each of the selected subsets. These random PCA runs on NMR models (Figure S5, S6) succeed in identifying the same variable subdomains as those identified using ensembles, for hPrP (Figure S5), and for mPrP (Figure S6).
Molecular figures have been rendered using PyMOL [48] and VMD [49].
|
10.1371/journal.pntd.0007721 | Impact evaluation of Zika epidemic on congenital anomalies registration in Brazil: An interrupted time series analysis | This study aimed to assess the impact of the Zika epidemic on the registration of birth defects in Brazil. We used an interrupted time series analysis design to identify changes in the trends in the registration of congenital anomalies. We obtained monthly data from Brazilian Live Birth Information System and used two outcome definitions: 1) rate of congenital malformation of the brain and eye (likely to be affected by Zika and its complications) 2) rate of congenital malformation not related to the brain or eye unlikely to be causally affected by Zika. The period between maternal infection with Zika and diagnosis of congenital abnormality attributable to the infection is around six months. We therefore used September 2015 as the interruption point in the time series, six months following March 2015 when cases of Zika started to increase. For the purposes of this analysis, we considered the period from January 2010 to September 2015 to be “pre-Zika event,” and the period from just after September 2015 to December 2017 to be “post-Zika event.” We found that immediately after the interruption point, there was a great increase in the notification rate of congenital anomalies of 14.9/10,000 live births in the brain and eye group and of 5.2/10,000 live births in the group not related with brain or eye malformations. This increase in reporting was in all regions of the country (except in the South) and especially in the Northeast. In the period “post-Zika event”, unlike the brain and eye group which showed a monthly decrease, the group without brain or eye malformations showed a slow but significant increase (relative to the pre-Zika trend) of 0.2/10,000 live births. These findings suggest an overall improvement in the registration of birth malformations, including malformations that were not attributed to Zika, during and after the Zika epidemic.
| Zika can be characterized as one of the most significant emerging arboviruses. The Zika epidemic in Brazil and the subsequent increase in the number of serious brain anomalies, such as microcephaly, has demonstrated the importance of analysing the impact of Zika infection on the rate of congenital anomalies in an affected population. From the analysis of the monthly data on the Live Birth Information System, the authors found that immediately after the Zika event there was a large increase in the notification rate of congenital anomalies reported as a complication of which infection (malformations of brain and eye) and also an increase in the rate of congenital malformations not related with Zika. This growth was seen throughout the country as a whole and in all regions (except in the South), especially in the Northeast where the infection rates were the highest. In the period post-Zika event, the group not related with brain or eye malformation there was an increase in the monthly notification rate while in the brain and eye group there was a decrease in the monthly notification rate. These findings suggest an overall growing awareness of health professionals to improve the registration of birth malformations trigged by the Zika epidemic.
| Zika is a vector-borne disease that has become an important concern for the global health agenda after the World Health Organization (WHO) declared a public health emergency of international concern (PHEIC) when it was associated with an epidemic of severe microcephaly cases[1–3].
Ever since the causal relationship between prenatal Zika virus (ZIKV) infection and microcephaly (among other serious brain anomalies) has been established [4]. One important research question is the impact of Zika infection on the rate of congenital anomalies in an affected population. However, this question has proved to be difficult to answer in Brazil and maybe in other Zika affected countries. In Brazil, the main source of information on birth anomalies is the Live Birth Information System (SINASC) that is well known for underreporting cases of congenital anomalies[5–9].
Surveillance data on congenital anomalies is an attractive source of information due to its universality (it covers more than 98% of live births in Brazil, for example) [10,11]. However, the surveillance system that relies on passive case finding strategies, such as SINASC, may be more susceptible to underreporting[12–14]. The rate of under estimation observed in SINASC varies from 36% to 47% in general, but in categories such as microcephaly, this rate be as high as 75% [7].
In 2015, a series of events, triggered by the Zika epidemic, had great potential to change the practices that impact the registration of congenital anomalies. These events were: the alarming growth in suspected cases of a rare condition (microcephaly); measures adopted to strengthen surveillance systems of congenital anomalies in regions where Zika cases had been reported; and finally the massive media coverage on the birth of babies with small heads[15]. Despite the well-known rise in the rate of congenital anomalies related with Zika complications during the epidemic, little has been described on whether some changes in the patterns of recording congenital anomalies not related to Zika have occurred over time. Therefore, this study aims to assess the impact of Zika epidemic on the registration of congenital anomalies in infants in Brazil.
We used an interrupted time series analysis (ITSA) design to identify changes in the trends of the registration of congenital anomalies in the country and its regions before and during the Zika epidemic from 2010 to 2017.
In this study we used the data from Brazilian live birth information system (SINASC). This system is updated with the registration of a live birth. The system uses a legal document, created in 1990 and compulsorily used throughout the country. The forms are pre-numbered and in three copies are identified by colours (white—the form kept by the local health council that digitizes the information and sends it to the Brazilian Information System headquarters; yellow—kept by the local registry office that generates a birth certificate; pink—kept with the health records of the pregnant women or the neonate in the facility). Mothers do not have to consent to the registration. Data available on this system are collected in a standard form which is completed by the health professional who assisted the delivery, mostly physicians as more than 98% of deliveries take place in hospital. SINASC includes information on the newborn (sex; birth weight, presence of abnormality), the mother (name, place of residence, age, marital status, education) and the pregnancy (length of gestation, type of delivery). Congenital anomalies observed at birth must be described using the International Classification of Diseases (ICD-10) 10th revision. In case of doubt about the condition, paediatricians or neonatologists should be contacted. If none of these professionals were available in the institution the SINASC headquarters must be contacted. These data have a high completeness rate, with missing data not exceeding 10% of most variables[16]. An evaluation of the birth registration system in Brazil found that 98% of Brazilian live births are registered in the system, although some difference are found within regions. However, it should be noted that all regions reach percentages of coverage over 90%[17].
We downloaded the SINASC files on live birth information registered in the period January 2010 to December 2017. We selected seven variables: (i) maternal age, (ii) maternal place of residence, (iii) presence of congenital malformation in the newborn (iv) malformation diagnosis according to International Classification of Diseases, 10th Revision (ICD-10), (vi) newborn date of birth (vii) newborn sex.
We divided our outcome in two categories:1) the rate of congenital malformation of the brain and eye coded as Q00-Q04 and Q10-Q15 in (ICD-10) per 10.000; 2) the rate of congenital malformation not related to the brain or eye, coded as Q05-Q07, Q16-Q18, Q20-Q28, Q30-Q34, Q35-Q37, Q38-Q45, Q50-Q56, Q60-Q64, Q65-Q79, Q80-Q89, Q90-Q99 in ICD-10 per 10,000. We separated the outcomes in these two categories because after the identification of Zika related abnormalities, the first group of ICD-10 codes were potentially related to ZIKV and its complications[18].
The event analysed in this study was the Zika epidemic in Brazil. The design of this study is an interrupted time series because the “event analysed” is expected to “interrupt” the level and/or trend of the outcome variable after its introduction. However, because we are analysing events occurring at birth, we expected a delay in the outcome after maternal exposure during pregnancy. Therefore we considered that: at the beginning of the Zika epidemic, cases were not compulsorily notifiable (the Brazilian surveillance system was not able to record all the disease cases systematically), and there was a delay of about six months between exposure (maternal infection) and outcome (live birth with congenital abnormalities). We therefore used the following approach to select the interruption point in the time series. Firstly we searched in the literature, and considered published studies that estimated the rise of Zika infections cases or exanthematous illness related to Zika. The rise in cases was reported to have started in March 2015 [1, 19]. We implemented a delay of six months from the month when the number of cases started to increase and therefore used September 2015 as the interruption point in the time series. For the purposes of this analysis, we considered the period from January 2010 to September 2015 to be “pre-Zika event,” and the period from just after September 2015 to December 2017 to be “post-Zika event.”
To summarize the characteristics of congenital anomalies according to newborn, sex, maternal age and ICD-10 diagnosis categories, we used descriptive statistics. To assess the impact of Zika epidemic on registry of congenital anomalies, we used an Interrupted Time-series Analysis (ITSA) for a single group. ITSA model for a single group (i.e. a single time series) assumes the following form:
yt=β0+β1Tt+β2Xt+β3XtTt+ϵt,
Where yt is the number of cases of malformation divided by the number of births multiplied by 100000 (rate) in each month; Tt is the time since the start of the study; Xt is a dummy variable that was 0 if the birth date was before Sept 2015 or 1 otherwise; XtTt is the interaction between the time and the dummy variable[20]. We use this model to estimate four parameters: (i) β0 that is the rate of malformation at T0 (“Pre-zika starting level”), (ii) β1 the mean increase in the malformation rate monthly before Sept 2015 (“Pre Zika event”), (iii) β2 is a change in the slope after Sept 2015 (immediately change) and (iv) β3 the mean increase in malformation rate after Sept 2015 (“Post Zika event”). Furthermore, for each β estimated in our model a t-test is performed to check the parameter values is equal to 0. We assumed that any time-varying unmeasured confounder is relatively slowly changing so that it would be distinguishable from the sharp jump of the event indicator (Zika epidemic).
We performed the ITSA for each of the groups defined in the data preparation section. We did our calculations using Stata SE version 15.
We obtained ethical approval from the Federal University of Bahia research ethics committee, Salvador, Brazil (CAAE: 70745617.2.0000.5030). All the data analysed were anonymized.
A total of 141,969 (0.6% of 23,359,499 live births) congenital abnormalities were registered in SINASC from 2010 to 2017. In Brazil, the starting rate of overall congenital malformations was estimated at 77.2/10,000 live births varying from 56/10,000 live births in the North to 89.3/10,000 live births in Southeast. The distribution of congenital anomalies by ICD-10 group varied over the years, mainly in the number of malformations of the nervous system that peaked in 2016; rates of malformations of the eye, ear, face and neck and malformations of the circulatory system increased over the years from 6.5% in 2010 to 9.2% and 7.2% in 2010 to 11.1% in 2017 respectively. Reporting of other congenital malformations has slowly decreased over time from 6.8% in 2010 to 5.3% in 2017. Maternal age and newborn sex distributions remained similar over the years, although the proportion of babies with congenital anomalies who were born to women over the age of 35 increased from 13.4% to 16.9%. (see Table 1).
Changes in the rates of reporting congenital malformation rates of the brain and eye are shown in Fig 1 and Table 2. In Brazil, the starting level of congenital malformation rate of the brain and eye was estimated at 8.2/10,000 live births. It was decreasing slowly monthly in the baseline, but it was not significant. Immediately after the interruption point (Sept 2015), the notification rate rose significantly, by 14.9/10,000 live births (CI 95% 6.7–23.2) per month, followed by a significant decrease in the monthly trend (relative to the pre-Zika trend) of 0.6/10,000 (CI 95% -1.1–0.2).
The North, Northeast, Midwest and Southeast regions showed similar patterns of change. Immediately after the interruption point, the notification rate rose significantly, followed by a significant decrease in the monthly trend (relative to the pre-Zika trend).The most dramatic change occurred in the Northeast region, where the notification rate of brain and eye anomalies immediately post the Zika event went up by 37.1/10,000 live births (95% CI 12.9–61.3) per month, over four times higher than observed in other regions. In the South region, where the circulation of Zika was low, there were no significant changes neither immediately nor over time post the Zika event.
As shown in Fig 2 and Table 2, the starting rate of congenital malformations, not coded as brain or eye related, were estimated at 69.6/10,000 live births, and this rate decreased every month prior to Zika by 0.03/10,000 live births (CI 95%-0.08–0.01). However, this was not significant. Right after the interruption point (September 2015), the notification rate of non brain or eye congenital anomalies increased significantly to 5.2/10,000 live births (CI 95% 2.3–8.1), three times lower than observed in the brain and eye group. A slow but significant increase of 0.2/10,000 (CI 95% 0.1–0.4) was observed (relative to the pre-Zika trend) in the monthly notification rates of no brain or eye anomalies.
The North, Northeast, and Southeast regions showed similar patterns of change. An increase in the notification rate of no brain or eye congenital anomalies was observed either right after the interruption point and monthly, however, the effect immediately after the Zika event was not significant in Southeast and the effect over time was not significant in Northeast. In the Midwest and South region, there was no significant change neither immediately nor over time post the Zika event.
We finally conducted a post hoc sensitivity analyses to investigate the earliest month where we got a positive result as an interruption point, and found that March was the first point that the series had broken, it would add to the hypothesis that these findings are Zika-driven.
This study showed that immediately post the Zika event in Brazil, there was a considerable increase in the notification rate of congenital anomalies, mainly malformations of the brain and eye that were reported as a complication of the infection. This growth was reported in all region of the country except in the South, especially in the Northeast, where the majority of Zika cases were concentrated1. When the frequency of Zika cases, and consequently risk of maternal infection decreased, the malformations related to its complications also went down significantly, as expected. However, the increased observed (compared with the pre-Zika trends) in the rate of congenital malformations not related to the brain or eye remained at the same level suggesting an overall improvement in the registration of birth malformations. A natural conjecture arises, that such an increase in the registration of cases was in part due to surveillance actions and overall growing awareness of health professionals at the time of the Zika epidemic.
The live birth information system is an attractive source of information on congenital anomalies. Before the circulation of Zika in Brazil the prevalence of congenital anomalies recorded in SINASC was less than 1%, however, it was estimated that the prevalence of congenital anomalies among live births in Brazil was about 2%- 3%[21]. The variation observed across Brazilian regions in the reporting of congenital anomalies rate is possibly due to the heterogeneity of the quality of the notification system, and higher rate of sub registration occurring in the poorest regions of the country. The level of underreporting can also vary by diagnosing groups. A high rate of underreported anomalies has been observed for hydrops, microcephaly, cleft palate, congenital heart disease and Down syndrome[7]. The reported findings suggest that, in part, the increase observed in this study was the result of an active search for cases. Therefore, after the Zika epidemic the live birth information system began to reflect closer to expected levels of notification of birth abnormalities as the reporting system improved.
There are many causes associated with the under-registration of congenital anomalies in the live birth information system, such as uncertainty and delayed diagnosis, deficient knowledge on how to correctly complete the form, and a lack of standardized case definition[22,23]. During the Zika epidemic, the broad press coverage of the malformations resulting from the virus, especially microcephaly, had the effect of changing health care practices and the way cases were recorded. This drew attention to clinical pictures which previously may have been overlooked or incorrectly reported[15]. Improving the quality of medical records of births can lead to a better understanding of the characteristics of children with congenital anomalies, the prevalence of the different types of congenital anomalies and the distribution of these across the country. This can provide crucial information for decision making processes by policy makers.
The great repercussions of the Zika and Congenital Zika Syndrome epidemic may also result in an improvement in prenatal care, either by alerting health professionals to the importance of protocols of care and by making the pregnant women more aware of the importance of pre-natal care and about measures to protect themselves against potentially dangerous infections such as Zika.
In the face of Congenital Zika Syndrome as a result of the Zika epidemic, the overall national emergency response was essential to identify gaps and take steps to strengthen the structure and correct distortions in the registration systems to produce more reliable surveillance systems capable of detecting and notifying cases of birth anomalies. However, after the drop in the number of Zika cases and its complications, there should be concern that some of the operational capacity structured during the epidemic may be dismantled, together with the extra funding and health care resources[24].
Our findings have several limitations. First, there is a lack of knowledge on the spectrum of Congenital Zika Syndrome, therefore in part the excess of cases registered in the no brain or eye ICD-10 group could be explained by unknown Zika complications. Although studies to better understand the spectrum of outcomes associated with maternal ZIKV infection have been developed, and knowledge about the syndrome is improving, the full spectrum of CZS has yet to be defined[25,26]. Secondly, while we have over two years of post-Zika data, it is possible that some effects have not yet become evident. Finally, our study uses routine data that were not specifically created to answer this research question. However, we use the data in high-level aggregate analysis and only use final, rather than provisional data, which are regarded as complete. Despite these limitations this study has provided evidence of improvements in the live birth notification system in registering congenital anomalies triggered by the Zika epidemic.
Congenital anomalies surveillance should be a priority on the public health agenda and CZS has highlighted its importance. Birth defect registration needs to be improved in all developing countries especially now that Zika is also circulating in Africa[27] and Asia[28,29], where birth defect surveillance systems may be even worse than in Brazil.
|
10.1371/journal.pcbi.1004650 | Protein Connectivity in Chemotaxis Receptor Complexes | The chemotaxis sensory system allows bacteria such as Escherichia coli to swim towards nutrients and away from repellents. The underlying pathway is remarkably sensitive in detecting chemical gradients over a wide range of ambient concentrations. Interactions among receptors, which are predominantly clustered at the cell poles, are crucial to this sensitivity. Although it has been suggested that the kinase CheA and the adapter protein CheW are integral for receptor connectivity, the exact coupling mechanism remains unclear. Here, we present a statistical-mechanics approach to model the receptor linkage mechanism itself, building on nanodisc and electron cryotomography experiments. Specifically, we investigate how the sensing behavior of mixed receptor clusters is affected by variations in the expression levels of CheA and CheW at a constant receptor density in the membrane. Our model compares favorably with dose-response curves from in vivo Förster resonance energy transfer (FRET) measurements, demonstrating that the receptor-methylation level has only minor effects on receptor cooperativity. Importantly, our model provides an explanation for the non-intuitive conclusion that the receptor cooperativity decreases with increasing levels of CheA, a core signaling protein associated with the receptors, whereas the receptor cooperativity increases with increasing levels of CheW, a key adapter protein. Finally, we propose an evolutionary advantage as explanation for the recently suggested CheW-only linker structures.
| Receptor clusters of the bacterial chemotaxis sensory system act as antennae to amplify tiny changes in concentrations in the chemical environment of the cell, ultimately steering the cell towards nutrients and away from toxins. Despite bacterial chemotaxis being the most widely studied sensory pathway, the exact architecture of the receptor clusters remains speculative, with understanding suffering from a number of paradoxical observations. To address these issues with respect to the protein arrangement in the linkers connecting receptors, we present a statistical-mechanics model that combines insights from electron cryotomography on the linker architecture with results from fluorescence imaging of signaling in living cells. Although the signaling data for different expression levels of key molecular components in the linkers seems contradictory at first, our model reconciles these predictions with structural and biochemical data. Finally, we provide an evolutionary explanation for the observation that some of the incorporated linkers do not seem to transmit signals from the receptors.
| Escherichia coli cells are able to sense changes in the chemical environment, allowing the bacteria to move towards higher concentrations of attractants and lower concentrations of repellents. The chemotaxis system is remarkable for its high sensitivity, wide dynamic range, and precise adaptation while only involving a small number of molecular components [1–3]. Despite the importance of receptor clustering in accounting for these signaling properties [4–7], there are still unresolved issues with the clusters, in particular with respect to the nature of the coupling mechanism between receptors [8]. It has been proposed that receptors assemble into larger arrays via the connection of the kinase CheA and the adapter protein CheW [9, 10], with potentially complementary effects of membrane-mediated interactions [11]. Unexpectedly, in vivo Förster resonance energy transfer (FRET) shows that increasing the expression level of CheA of engineered non-adapting receptors decreases the cooperativity among receptors. In contrast, expressing more CheW increases the cooperativity, albeit in different ranges of expression levels [12]. This raises the question of how these different observations can be reconciled.
In E. coli, there are four types of methyl-accepting chemoreceptors: the high-abundance Tar and Tsr receptors that sense serine and aspartate, respectively, and the low-abundance Trg and Tap receptors [13, 14]. In addition, Aer is a chemoreceptor-like sensor of redox potential [15]. The chemoreceptors form homodimers, which assemble into trimers of dimers (TDs) [16, 17]. On a larger scale, these TDs cluster at cell poles [18–20]. CheW and CheA, which interact with the cytoplasmic domain of the receptors [21], are involved in the stabilization of these clusters [22], which in turn consist of smaller complexes (signaling teams) [6, 23, 24]. Signal transduction is triggered by ligand-receptor binding, which leads to a conformational change in the cytoplasmic domains of the receptors [25–27]. The removal of attractant (or addition of repellent) activates autophosphorylation of the kinase CheA, which is associated with the receptors via the adapter protein CheW (Fig 1A). The phosphoryl group is then transferred to the response regulator protein CheY, which diffuses through the cytoplasm. CheY-P binds to the flagellar motors to induce clockwise rotation and tumbling of the cell. In contrast, addition of attractant (or removal of repellent) inhibits autophosphorylation of CheA. CheY-P dephosphorylation by phosphatase CheZ leads to counterclockwise rotation and straight swimming [1].
To avoid saturation of the sensory system, adaptation is implemented via covalent receptor modification. This is achieved through changing the receptor-methylation level by the activities of the methyltransferase CheR and the methylesterase CheB, which antagonistically add and remove, respectively, methyl groups at four or five, depending on the receptor, specific glutamate residues on each receptor monomer [33], respectively. Methylation by CheR increases the activity of CheA, i.e., its autophosphorylation rate, thus counteracting the effect of attractant binding. In contrast, CheB activation by phosphorylation by CheA-P decreases CheA activity [12]. Through genetic engineering, the glutamate residues (E) can be replaced by one to four glutamine residues (Q) to mimic increasing receptor-methylation levels in the absence of CheR and CheB [2]. The E. coli chemotaxis pathway is exceptionally well characterized and is thus amenable to modeling at a high quantitative level.
To explain the receptor cooperativity, which generates the high sensitivity of the system, the mechanism of receptor-receptor coupling has attracted much interest [8, 34–37]. Electron cryotomography (EC) images of the TDs in quick-frozen cells led to the idea that TDs form densely packed hexagonal ‘honeycomb’ arrays (Fig 1B) [28, 32, 38]. These and other in vitro experiments using nanodiscs and nanoscale plugs to imitate cellular membranes suggest that –CheW–CheA2–CheW– is the structural core unit linking two TDs (see Fig 1C for a simplified depiction) [9]. An approach to indirectly study the cooperative behavior of the specific receptors inside the cells is to monitor the signaling activity of CheY-P/CheZ pairs via FRET, with the FRET signal being proportional to the overall CheA activity [39]. An increase in the concentration of CheW was observed to enhance the cooperativity of the FRET response mechanism, whereas, unexpectedly, an increase in CheA concentration led to the opposite effect [12]. It is well known that multimeric protein complexes can be inhibited by high concentrations of one of their components, similar to the prozone phenomenon in precipitin tests [40]. However, it is unclear how the FRET results relate to other experimental observations, including the proposed linker and lattice structures.
Here, we use statistical-mechanics modeling within the framework of the Monod-Wyman-Changeux (MWC) model [41] for cooperative receptor complexes to unify the assumed linker and lattice structures with the seemingly contradictory FRET results. By implementing the linker structure we initially fit our model of receptor complexes of up to four TDs to FRET data obtained with cells that express only the Tar receptor in different non-adapting modification states. Next, we apply our model to Tar–Tsr–Tap and Tsr–only cells in the non-adapting QEQE modification state, which mimics half-methylated receptors. As a result we recover the experimentally observed decrease in cooperativity of the response to serine with increasing CheA concentration, whereas increasing CheW yields the observed enhanced cooperativity. Note, other higher order effects of protein overexpression, such as membrane invaginations or interference of CheA/ CheW with clustering, are not included. Our results surmise that the observed opposing trends in cooperativity are based on a critical combination of the correct linker architecture and a constant average complex size.
At the heart of our approach lies the MWC model [5, 6, 12]. Chemoreceptors are regarded as two-state systems being either active (on) or inactive (off), with conformation-dependent dissociation constants K D on and K D off for a specific ligand. As the attractant affinity of inactive receptors is higher than for active receptors (K D on ≫ K D off), the state ratio tips towards inactive receptors with increasing ligand concentration c. In contrast, receptor modification m favors the active state in the absence of ligands represented by an energy offset Δϵ(m). The resulting single-dimer free energies in the active and inactive states are given by
f on = Δ ϵ ( m ) - ln ( 1 + c K D on ) + μ f off = - ln ( 1 + c K D off ) + μ , (1)
with μ the chemical potential of the receptors in the membrane. All energies are expressed in units of the thermal energy, kB T. In our approach, we allow for an ensemble of different complexes with varying complex size x (i.e. number of connected TDs) and partially developed linkers as rest groups R (Fig 1D). All receptors within a complex are assumed to share the same conformational state because of tight coupling. For simplicity, we consider the –CheW–CheA2–CheW– linker structure [9], which we incorporate by assigning energies μW and μA2 for each CheW and CheA2 molecule integrated in a specific receptor-complex type (see Discussion section for an alternative linker structure). These energies are of the forms
μ W = ln ( ( K D W · K D A ) 1 / 2 [ W ] ) μ A 2 = ln ( K D A [ A ] ) , (2)
where [W] and [A] indicate monomer concentrations and K D W and K D A are dissociation constants for CheW–receptor and CheW–CheA2 binding, respectively. In particular [W] and [A] are expressed as fractional changes i and j of wild-type expression levels [W]0 and [A]0, respectively:
[ W ] ( i ) = i · [ W ] 0 [ A ] ( j ) = j · [ A ] 0 . (3)
The TD is assumed to be the smallest receptor unit [9, 42], and the maximal number of connected TDs is restricted to four, in line with observed Hill coefficients from FRET [12, 23]. (Including larger complex sizes does not alter the model predictions, but increases the computational complexity significantly; see Materials and Methods.) Each dimer can maximally bind to one molecule of CheW, whereas CheA is assumed to not interact with receptor dimers directly. In order to restrain the combinatorial complexity partially developed linkers are only considered in a symmetric manner, i.e. all rest groups are assumed to be identical in a complex. Furthermore, we attribute an attractant energy J to each linker within an active complex, a treatment in line with the previously proposed enhanced coupling among active receptor dimers [24], albeit independent of receptor-modification level.
The resulting free energies for a complex of size x and rest group R are given by (cf. Fig 1D)
F on ( x , R ) = 3 x f on + ( x - 1 ) ( μ A 2 + 2 μ W + J ) + R ( μ A 2 , μ W ) F off ( x , R ) = 3 x f off + ( x - 1 ) ( μ A 2 + 2 μ W ) + R ( μ A 2 , μ W ) , (4)
with 3x receptor dimers per complex of size x and x − 1 linkers. Such a complex has x + 2 rest groups with R(μA2, μW) given by
R 1 = 0 R 2 ( μ W ) = ( x + 2 ) μ W R 3 ( μ A 2 , μ W ) = ( x + 2 ) ( μ W + μ A 2 ) , (5)
for Eq (1) no rest group, Eq (2) a CheW and Eq (3) a CheW and a CheA dimer, respectively. The probability PS for a certain complex type S(x, R) and its probability P S on of being active follow from standard combinatorial reasoning and the partition function Z Z ≡ 1 + ∑ S ( e - F on ( S ) + e - F off ( S ) ) (6) P S = e - F on ( S ) + e - F off ( S ) Z (7) P S on = ( 1 + e F on ( S ) - F off ( S ) ) - 1 , (8)
where the number 1 in the partition function Z reflects the possibility of an empty membrane site.
Assuming the FRET signal to report the number nA2(S) of CheA2 dimers within an active complex, we define the receptor activity as
A = ∑ S P ( S , on ) · n A 2 ( S ) = ∑ S P S · P S on · n A 2 ( S ) . (9)
In contrast, the classical MWC model for coupled receptors describes the response of a single complex of N TDs to a change in ligand concentration. Without incorporating the receptor coupling explicitly, the corresponding activity A ˇ reads [23]
A ˇ = ( 1 + exp [ N { Δ ϵ ( m ) + log ( 1 + c / K D off 1 + c / K D on ) } ] ) - 1 . (10)
In the past, the Hill coefficient nH and complex size N have broadly been treated as equivalent to quantify the cooperative behavior of receptor complexes, and in [23], an increase in N with receptor-modification level was equated with an increase in receptor cooperativity. However, both quantities are not necessarily the same as approximating Eq (10) by a Hill function with nH = N requires c ⪡ K D on [6]. We found that, in the classical MWC model, the response of differently modified Tar receptors to MeAsp, a non-metabolizable analog of aspartate, can also be described with a fixed N for all modification levels. This treatment results in a similar quality of fit when relating the reduced number of parameters to the new χ2 goodness-of-fit value (see S1 Fig). As our model incorporates an ensemble of complexes of varying sizes, the finding of a constant complex size N in the classical MWC model is naturally generalized by a constant average complex size 〈N〉 with respect to ligand concentration and receptor-modification state. The average complex size, which we term receptor density ρ, is given by
ρ = ∑ S 3 · x · P S = 3 ⟨ x ⟩ ≡ constant , (11)
with x being the number of dimers of a given complex type S. The chemical potential μ in Eq (1) is adjusted throughout the simulation to fulfill this condition, reflecting anticipated regulation of the receptor-expression level by the cell. Biologically, a constant receptor density can be achieved by random receptor insertion into a growing membrane at constant rate [35]. Since wild-type cells express and insert receptors in the QEQE modification state [2], we do not expect a modification-dependent insertion rate. Although allowing for a modification-dependent ρ would increase the quality of fit because of an increased number of fitting parameters, our minimal model with constant ρ can describe the data very well.
In order to test our model, we firstly applied it to FRET data of Tar-only receptors in different non-adapting receptor-modification states from Ref. [23] i.e. Tar{QEQE}, Tar{QEQQ} and Tar{QQQQ}. The dose-response curves of the chemoreceptors match closely the statistical-mechanics model with fixed receptor density, and hence fixed average complex size (Fig 2A). Fig 2B displays the fitted receptor density ρ next to the Hill coefficients nH of the experimental curves (see Materials and Methods) and the complex size N of the classical MWC model, taken from [23]. Although the classical MWC model predicts a rise in complex size with modification level [23], including its implementation based on a dynamic Ising model [24], this is not true for the Hill coefficients (see also S1 Fig). This finding shows that receptor modification is not the main determinant of receptor cooperativity.
In our model, the chemical potential μ can be regarded as the cost function for the cell to provide a constant complex size in the membrane. By definition, the chemical potential μ ≡ ∂F/∂N reflects the amount of energy required for adding a particle to a system with free energy F. Although the value of the parameter μ, introduced to ensure constant receptor density ρ, is gained by solving a highly nonlinear equation, its behavior with respect to ligand concentration is very homogeneous and characterized by two regimes, as shown in Fig 2C. While this cost is approximately constant for c < cH, with cH being the half-maximum concentration obtained from Hill fits, the cost necessary to maintain a constant density increases rapidly for ligand concentrations beyond cH. In this second regime, the curves for all modification levels m are of the form f(c) = f0 + ln c, which is the functional description of an ideal chemical potential. Although the slope in the second regime is the same for all values of m, the different offsets f0(m) reflect the modification-dependent energy Δϵ(m). Note, if we were instead to keep μ constant (and not ρ), then bumps would appear in the dose-response curves as a result of the receptor density increasing with ligand concentration (see S2 Fig).
In summary, our model is capable of quantitatively describing dose-response curves from in vivo FRET, in particular the receptor-receptor cooperativity. Although in spirit similar to other recent statistical-mechanics models, most noticeably by Hansen et al. [24] and Lan et al. [43], only our model addresses the protein connectivity in receptor complexes.
While the receptor density ρ is assumed to be constant on a short time scale, the rate of receptor expression and insertion into the membrane can be regulated by the cell on a longer time scale. Hence, as a further test of our statistical-mechanics model, we investigated how a change in receptor density ρ affects CheA activity at wild-type expression levels for CheA and CheW. Fig 3A shows modeled dose-response curves for different ρ values of 1.5 ⋅ ρ0, ρ0 and 0.5 ⋅ ρ0 with ρ0 = 7.5 the wild-type receptor density and otherwise using the same parameter set as in Fig 2. An increase in receptor density is directly associated with an enhanced signal amplitude because more CheA molecules are incorporated into the complexes. Fig 3B reflects the associated trend in cooperativity by comparing density ρ and Hill coefficient nH. In qualitative agreement with experimental observations [12] and in line with previous modeling [6], larger complex sizes lead to higher sensitivities and hence steeper dose-response curves given a certain receptor-modification state. Since the expression level of receptors (and other chemotaxis proteins) is highest under nutrient-poor conditions, the resulting increase in receptor density and cooperativity leads to enhanced sensitivity when it is most crucial for cell survival [44].
To gain insight into the role of CheA and CheW in forming receptor complexes, we varied the expression levels [A] and [W] to study the effect on receptor activity. According to the experimental observations in [12], we set the CheW concentrations to 0.7, 0.1 and 0.01 and the CheA concentrations to 8, 0.3 and 0.25 times the wild-type values [W]0 and [A]0, respectively. This allowed us to make the comparison with experimental dose-response curves from FRET of Tsr–only cells (for varying CheW) and Tar–Tsr–Tap cells (for varying CheA), both in the non-adapting QEQE modification state. To keep the overall number of parameters small, the data for changes in [A] and [W] was fitted with the same parameter set (Δϵ, ρ, K D,Tsr on, K D,Tsr off, J, μ W 0 and μ A 2 0). Multiplication of the calculated activities with scaling parameters sA and sW, respectively allows for comparison with the FRET signal amplitudes. Subsequently, a Hill function was fitted to the model curves and the model Hill parameters were compared with the experimental values. Note that our minimal model does not account for alternative forms of signaling disruption upon over- or underexpression of CheA/CheW, such as zipper-like invaginations of the cell membrane [45] or interference with trimer formation [16].
Fig 4A and 4B show the model data next to the experimentally determined Hill curves for variations in [W]. Enhanced CheW expression results in raised activity amplitudes and Hill coefficients (Fig 4C and 4D). Although the nH values from the model change significantly with expression level [W] at a 95% confidence level, which is in qualitative agreement with the experimental data, especially with respect to the highest CheW expression level, the change in nH is less pronounced for the model than the experimental data. The positive correlation between kinase activity and amount of available CheW becomes evident in the distribution of complex species at half-maximum concentration (Fig 4E). Whereas low levels of [W] favor independent, single TDs, larger complexes are more likely to form for larger [W]. As the probability for an empty membrane site also increases, the receptor density remains constant.
Changing [A] in our model has the opposite effect on the Hill coefficient as changing [W]. This result is in line with experimental data (Fig 5A, 5B and 5C). The activity amplitude reflecting the amount of active CheA molecules benefits from higher CheA levels, as one would expect (Fig 5D). In contrast, Hill coefficients are higher for smaller [A], recovering the naively unexpected experimental observations (Fig 5C). Looking at the distribution of complexes at half-maximum ligand concentration (Fig 5E), we note that although high CheA concentrations favor rest groups including CheA, complex sizes of 3 and 4 TDs are more likely at lower concentrations of CheA.
The opposing trends in nH concerning variations in [A] and [W] are a direct result of the linker stoichiometry and fixed average complex size. For complexes with rest groups, the ratio of CheW molecules per TD is independent of the complex size (Fig 6A). However, for species without rest groups, this ratio increases with the number of coupled TDs. As a result, an enhancement in [W] yields larger complexes that directly incorporate more CheA molecules. Furthermore, empty sites ensure a constant receptor density even when expression levels of CheW and CheA are extremely low. In this case, the receptor density still remains constant as empty sites can be occupied by individual TDs. This requires a dilute membrane, i.e., a receptor density not much larger than 〈ρ〉 = 3〈x〉 = 9 (see Fig 2B).
In contrast, the corresponding ratio of CheA dimers per TD is highest for single TDs with full rest groups and decreases with increasing complex size (Fig 6B). The CheA molecules within the rest groups contribute to the FRET amplitude but not to the receptor cooperativity. An accompanying rise in the number of occupied membrane sites ensures a constant receptor density.
Our model qualitatively reproduces the experimental results obtained when the expression levels of CheW and CheA were changed. However, there are quantitative differences, especially with respect to the change in cooperativity as a function of the expression level of CheW. This change is less pronounced in the model than in the experiment. Recent findings from electron cryotomography may shed light on the reasons for these discrepancies. Although both studies stressed the importance of one dimeric CheA and two CheWs as the minimal unit needed for kinase activation, Briegel et al. [30] and Liu et al. [31] proposed additional CheW-only linkers, underlining the role of CheW in the cooperative behavior of TDs. Such structures could explain how increased levels of CheW contribute to the cooperativity of TDs. In order to quantify this effect, we allowed for additional CheW-only linkers in our model (Fig 7). The dimeric appearance of CheW in the linker is accounted for by a new parameter μW2; we keep the previously introduced rest groups for simplicity.
Fig 8 shows the results for varying expression levels of CheW and CheA. The dose-response curves of the new model exhibit the same trends in Hill coefficient and amplitude for variation in [W] (Fig 8A) and [A] (Fig 8B) as before, in agreement with experimental results (see also S3 Fig). However, the difference in behavior is manifested in the comparison panels below. The previously obtained minor changes in receptor cooperativity as a function of [W] are now much more pronounced (Fig 8C), although the modeled Hill coefficients for [A] variation are larger than the experimental ones (Fig 8D). The excess CheW leads to formation of CheW-only linkers and hence larger complex sizes when the amount of available CheA is held constant.
In order to make predictions beyond the data used to fit the model, we created surface plots of amplitudes and Hill coefficients covering several orders of magnitude for expression levels of CheW and CheA (Fig 9). The receptor activity and hence amplitude increases monotonically with the level of CheA, whereas the increase in amplitude with respect to the level of CheW is only pronounced in a subspace around the experimental data (Fig 9A). In the case of high CheA levels, CheW-only linkers exclude CheA from signaling. This also occurs at the wild-type CheA level, although the extent of the effect strongly relies on model parameters. The surface plot showing the Hill coefficients as a function of the expression levels of CheW and CheA has a saddle-like form (Fig 9B). Although the right flank is consistent with the FRET data at high levels of CheA (small Hill coefficients), the Hill coefficient also decreases at very low levels of CheA as the receptor activity diminishes. To test to what extent the model predictions depend on the actual values of parameters μ W 0, μ A 2 0 and μ W 2 0, we varied these parameters and found that the general shape of the surface plot was preserved. Taken together, these observations suggest the need for regulation of both CheW and CheA by the cell to balance signaling amplitude and sensitivity. Indeed, as CheW and CheA are required in comparable amounts [13], both are expressed from the same operon [46].
Receptor coupling plays a key role in the remarkable sensing and signaling properties of bacterial chemotaxis. These networks can explain the high sensitivity, wide dynamic range and precise adaptation. In this work we present a statistical-mechanics model of different complex sizes, modeling for the first time a molecular linker architecture consistent with (i) FRET dose-response curves, (ii) cryotomography data and (iii) nanodisc experiments. The linker –CheW–CheA2–CheW– proposed by Li and Hazelbauer [9] is incorporated by assigning expression level-dependent energies μW and μA2 respectively for each CheW and CheA2 molecule within a complex as part of a fully or partially developed linker. A coupling energy J < 0 attributed to linkers between active TDs indicates that the coupling between active trimers is stronger than between inactive trimers, in agreement with previous modeling [24]. Although the actual distribution of complex sizes is influenced by expression levels [W] and [A], a readily adapted chemical potential μ ensures a fixed average complex size ρ with respect to ligand concentration c.
Our model was first applied to describe the dose-response of Tar receptors in different modification states to MeAsp, a non-metabolizable analog of aspartate. We mainly considered a constant, modification-independent ρ, a constraint that not only reduces the number of parameters but also calls into question that the complex size increases with receptor-modification level [23]. In our work we discovered the discrepancies between the number of connected TDs N and the curves’ Hill coefficients nH within the classical MWC model. An increase in N is not directly associated with an increase in nH. In our statistical-mechanics model, the approximately constant nH is explained by a constant average complex size across all receptor-modification levels. Indeed, experiments show that both the level of expression of receptors and the insertion of newly synthesized receptors into the inner membrane by the Sec-machinery are highly regulated [47, 48].
Hansen et al. [24] previously presented a dynamic-signaling-team approach to describe the data obtained with Tar-only cells in which the allosteric coupling among trimers is represented by a modification-dependent trimer-trimer interaction energy J ^ ( m ) without modeling the actual protein connectivity. Limited conformational spread and hence a finite complex size is achieved by using a long-range repulsion energy U between all trimers within a complex. In contrast, our model is simpler while providing valuable insights. Neither μW and μA2 nor J in our ensemble model depend on the modification state of the receptor, and μ ensures constant average complex size without introducing a repulsive term. Furthermore, the chemical potential μ(c) provides insights into the energetic cost of insertion of receptors into the membrane and its dependence on ligand concentration c, albeit based on an equilibrium mechanism.
For constant J and ρ, we conclude that receptor modification mainly governs the ‘turn off’-ligand concentration, whereas its influence on receptor clustering is limited. This finding is supported by Briegel et al. [49], who found that the receptor array order and the spacing of receptors in different modification states were indistinguishable. This is in stark contrast to Hansen et al. [24], who predict a strong increase in average complex size with increasing receptor-modification level. High-resolution imaging of equilibrated receptors in artificial membranes by electron or total internal reflection fluorescence (TIRF) microscopy may allow direct determination of receptor-complex distributions and their dependence on receptor-modification level and ligand concentration. Using photoactivated localization microscopy (PALM) [35] or quantitative immunoblotting [13], such an investigation could also be performed on intact cells.
Although CheA and CheW have long been known to mediate receptor interactions [12, 22], an increase in the expression level of CheA leads to a reduction in receptor cooperativity [12]. Varying expression levels of CheA and CheW in our model produced results in agreement with experimental data of Sourjik and Berg [12], thereby supporting the linker architecture we employed. The striking observation that increased CheA levels lead to higher kinase activities but lower cooperativity is based on the fact that the number of CheA dimers per TD is highest for single trimers with almost fully developed linker rest groups (Fig 6B). Hence, overexpression of CheA, a bridging molecule at the center of the linker, promotes smaller complex sizes. CheA molecules within the rest groups do not contribute to TD coupling and curve steepness, but nevertheless add to the activity of the FRET signal.
In contrast to what is observed with CheA, raising the level of CheW leads to larger complex sizes and an increased number of empty membrane sites. Again, this behavior becomes comprehensible when the number of CheW molecules per TD (Fig 6A) is taken into account. While this ratio is constant for complexes with rest groups, it increases with complex size in the absence of partially developed linkers. Larger complexes directly incorporate more CheA to enhance cooperativity as well as the amplitudes of FRET signals observed both in the model and experimentally. In light of our model the experimental observations are produced by a combination of constant receptor density and (partial) linkers. Although partial linkers play a crucial role in the mechanism of our model, their inclusion might appear arbitrary at first. Interestingly, Briegel et al. [30] recently observed a range of assembly intermediates and partial receptor hexagons forming when [W] and [A] were varied. Our surface plots of amplitudes and Hill coefficients also make testable predictions for wide-ranging CheA and CheW expression levels (Fig 9). Is there any evidence to suggest that ρ remains constant when CheA and CheW expression levels change? First, CheA and CheW binding to the receptors occurs after insertion of the receptors into the membrane. Second, increasing the expression of a protein, e.g., of CheW, should remove ribosomes from translating receptor mRNA [50, 51]. Although expected to be a minor perturbation, this may lead to a reduced receptor density and hence cooperativity. However, the opposite trend is observed in FRET experiments [12].
Although our assumed linear linker structure –CheW–CheA2–CheW– matches observed stoichiometries [9, 13], electron cryotomography images suggest that reality is more complicated [30, 52]. Modeling of the electron density and spin-labeling studies suggest that CheW and the P5 domain of CheA form alternating CheW/CheA rings connecting the trimers, with P5 occupying positions approximately equivalent to CheW (see Fig 10). This arrangement is consistent with the strong structural homology between P5 and CheW. However, to describe the FRET data obtained with cells with overexpressed CheA and CheW [12], our model predicts that CheA2 has the role of a bridging molecule and connects trimers via a CheW associated with each trimer. Indeed, an alternative linker with direct receptor-CheA binding and hence symmetric roles of CheA and CheW upon clustering does not match the FRET data (see panel D in S4 Fig). This view is supported by binding assays, which show that CheW binds much firmer to receptor trimers than CheA to trimers (see Fig. 5A,B in [9] and also discussion in [29]).
Although our model qualitatively reproduces the experimental FRET data, the change in cooperativity with variation in [W] is less pronounced in the simulation than in experiments. Recent findings based on electron cryotomography offer a possible explanation for this shortcoming. Briegel et al. [30] and Liu et al. [31] stress the importance of the implemented core unit stoichiometry, but they propose a second type of linker that only involves CheW, with P5/CheW interactions replaced by CheW/CheW interactions [31]. To investigate the consequences of these findings for signaling behavior, we allowed for an additional –CheW–CheW2–CheW– linker in our model. The simulated dose-response curves show a greatly enhanced change of cooperativity with variation in [W] (Figs 8C and 11A). The generally increased Hill coefficients, and hence sensitivity, may reveal an evolutionary advantage that is not apparent in the tomography images but is detected by FRET. However, whereas CheW-only linkers fit the FRET observations, their incorporation into complexes needs to be tightly regulated. Moreover, in addition to excluding CheA from signaling (Fig 11B), high levels of CheW were also claimed to disrupt receptor clustering [53]. Taken together these observations suggest that an optimal level of CheW is required for cooperative signaling by receptors (Fig 11C).
In conclusion, our work integrates functional (FRET) and structural (nanodisc and electron cryotomography) data, explains the paradoxes that increased levels of CheA lead to less cooperativity, and provides a functional role for CheW-only linkers. Our proposed linker –CheW–CheA2–CheW– is consistent both with the data from experiments with nanodiscs [9] and with images from electron cryotomography [29–31], if the P5 domain of CheA binds more weakly to the receptor than does CheW. We predict that the observed tetrameric CheW linker, if incorporated at an optimal level, increases the cooperativity while keeping the receptor activity at a sufficiently high level. An increased understanding of the protein connectivity in receptor clusters may aid not only in describing the fundamental biology of receptor signaling, including the role of cytoplasmic receptor clusters in Rhodobacter sphaeroides and Vibrio cholerae [52], but may also contribute to the design of novel biosensors [54].
Keeping ρ constant requires nonlinear optimization of μ at every ligand concentration. For performance reasons we therefore chose to implement the model in C# and used a custom-written toolbox to connect to MATLAB 2014a for parameter optimization and plotting. The value for μ is determined based on Brent’s method for root-finding [55]. Fitting of model parameters employs Global Search from MATLAB Global Optimization Toolbox. Multiple start points are generated using scatter-search options (5000 trial points). For the different start points square deviations from experimental data are minimized using the function fmincon with interior point optimization. Note while the number of molecular species in the model increases linearly with the maximal complex size, the computational time is determined by the root finding. The latter becomes considerably harder with additional exponentials of increasing arguments in Eqs (6), (7) and (11).
In order to quantify the cooperative behavior of the complexes, Hill functions A(c) Eq (12) with amplitude A0, half-maximum concentration cH and Hill coefficient nH are fitted to the model evaluated at 50 logarithmically spaced concentrations between c = 0.001mM and c = 1mM. The Hill coefficients in the comparative plot Fig 2B result from direct fitting to the experimental data.
A ( c ) = A 0 1 + ( c c H ) n H (12)
Though parameter confidence intervals can be calculated based on robust regression and the resulting covariance matrix, especially for highly nonlinear models as ours their validity is questionable given the underlying linear theory [56]. We therefore decided against including confidence intervals except for the fitted Hill curves.
We note that for all simulations with variations in expression of CheA and CheW the Hill amplitudes match quantitatively much better their experimental counterparts than do the Hill coefficients. This observation is partly owed to the fitting routine. With logarithmically spaced concentrations, a difference in amplitude between model and experimental curve directly impacts the corresponding χ2 goodness-of-fit value. In contrast, a small variation in the Hill coefficient only influences the slope of the curve within a relatively narrow range of ligand concentrations and hence is less reflected in the optimization function value.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.