In Silico Molecular Comparisons of C. elegans and Mammalian Pharmacology Identify Distinct Targets That Regulate Feeding

This paper takes advantage of similarities between the C. elegans and human pharmacopeia to identify and validate pharmacological targets that regulate C. elegans feeding rates.


Introduction
Before the molecular biology era, pharmacological targets were typically classified by the effects of organic molecules on whole tissues [1]. Many pathways were first recognized based on phenotypic responsiveness to compounds without knowledge of underlying molecular mechanisms. Examples include the inference of the aand b-adrenergic pathways in the 1940s [2], the inference of the H2 histaminergic receptor [3] and of the m, and k-opioid receptors in the 1970s [4], and the proposal of the 5-HT3 serotonergic receptor in the mid-1980s [5]. Although these targets were eventually characterized by molecular biology, the tissue and organism approach had the advantage that the compounds emerging from it were active on a physiologically intact tissue or organismal circuit, and directly linked functional perturbation of targets to biological effects.
Phenotypic compound screens return to this classical approach to capture some of the same advantages for the discovery of molecules with systemic activity. Such screens have generally relied on high content microscopy assays in cell-based systems [6][7][8]. However, certain biological processes such as physiology and behavior are the result of integrated organism-wide processes that only manifest themselves in intact multicellular organisms. For example, as a physiological process, feeding behavior is the outcome of integration of extrinsic and intrinsic cues of food availability and energy demand and thus is best understood when studied in whole organisms. Elucidation of the neural circuits that determine feeding is a fundamental challenge in the neuroscience of energy homeostasis [9]. Small molecules that alter feeding behavior can serve as useful reagents for investigating these circuits and provide exquisite temporal control in ways not easily achieved through genetic manipulations.
Given their small size and ease of manipulation, C. elegans have been used in pharmacology-based phenotypic screens [10][11][12][13][14]. These animals are also well suited for study of molecular and neural circuits that underlie food intake behavior. C. elegans feeds using peristaltic contractions of a muscular pharynx to aspirate microbes into the lumen of the intestine [15]. This pharyngeal pumping rate directly correlates with the transport of nutrients into the intestinal lumen [16,17]. C. elegans' central nervous system integrates signals from external cues such as food availability, food quality, and internal nutritional status to regulate feeding behavior [17][18][19][20]. Multiple pathways in the nervous system that are dependent on serotonin, glutamate, and neuropeptide release regulate the pharyngeal pumping rate in C. elegans [21][22][23][24][25][26][27]. Thus, C. elegans' feeding behavior is subject to regulation by some of the same physiological parameters and molecular components as those in mammals.
Feeding behavior is a relatively challenging read-out for a screen-scale phenotypic effort. Therefore, we focused on surrogate phenotypes that could potentially enrich for identification of feeding regulatory compounds. From a high-content microscopy assay, we previously discovered 84 compounds, which increased or decreased Nile Red in C. elegans [11]. Nile Red is a vital dye that has been broadly used for detecting fat levels in numerous experimental systems [28][29][30][31][32]. However, its use as a read-out of fat content in C. elegans has been challenged [33][34][35][36]. Nevertheless, for a subset of compounds emerging from the initial Nile Red screen, effects on lipid content were further verified by other vital dyes, biochemical methods, molecular read-outs of fat content, as well as efficacy in mammalian cell-based models of adipogenesis [11]. Therefore, we hypothesized that compounds emerging from the Nile Red screen would be enriched for those that also alter feeding behavior.
For all but one of the compounds, no biological targets were previously known. To predict targets for these compounds, we used chemoinformatic inference based on ligand patterns against mammalian receptors. For a subset of the predictions, we tested the compounds against the predicted mammalian targets in vitro and subsequently tested orthologs of these targets in C. elegans by chemical-genetic epistasis. We then combined the compounds with the mutants to map epistasis relationships for feeding behavior, identifying four signaling pathways previously unassociated with C. elegans feeding regulation. Together, our findings reveal that chemical screens in C. elegans lead to molecules with activity toward known human targets and highlight the utility of C. elegans for unambiguous assessment of compound-target-phenotype relationships in the context of an intact organism.

C. elegans Screen Actives Are Similar to Ligands for Multiple Mammalian Targets
Computational chemoinformatics methods have been used to query annotated ligand-target interactions to identify novel targets for known drugs [37] and, recently, targets for ligands identified by a screen in zebrafish [38]. We used the Similarity Ensemble Approach (SEA) [39] to interrogate the ChEMBL database for targets, represented by their ligand sets, resembling compounds of the 84 identified in a C. elegans phenotypic screen [11]. ChEMBL annotates more than 8,000,000 ligand-target interactions for 2,456 targets; mostly, mammalian. SEA uses extended connectivity fingerprints to measure chemical similarity, quantified as Tanimoto coefficients (T c ), between a query molecule and each target ligand. The ensemble of all pairwise T c 's for a target's ligand set to the query are summed and compared to an expectation of random association. Using statistical machinery similar to BLAST, an expectation value (E-value) quantifies the possibility that observed structural similarities could occur by chance. Because SEA relies on reported ligand-target interactions, it cannot predict associations for chemically novel molecules or for targets for which no ligands have been identified. Nevertheless, SEA provides a rapid and systematic approach for discovering the pharmacological relevance of C. elegans actives vis-à-vis mammalian targets with known ligands.
At an E-value threshold of less than 10 25 , 79 of the 84 active compounds were associated with at least one target in ChEMBL. Most compounds were predicted to have two or more targets, with 572 distinct targets predicted overall ( Figure 1A). Target predictions often spanned two or more classes per compound as has been observed in GPCR pharmacological relationships [40]. We first considered the possibility that the profile of predicted targets simply recapitulates the underlying distribution of target classes in ChEMBL. This was not the case. Our ChEMBL-derived dataset is composed of 35% enzymes, 15% ion channels, 30% membrane receptors, and 20% transporters, transcription factors, and other proteins. Conversely, more than 60% of the SEApredicted targets for the phenotypic actives were enzymes, whereas both membrane receptors and ion channels were comparatively underrepresented ( Figure 1B). This deviation may reflect the targets with known ligands that are relevant to metabolic phenotypes or peculiarities of the C. elegans model, a question that we will address below.

Author Summary
Many beneficial pharmacological interventions were first discovered by observing the effects of perturbation of intact biological systems by small organic molecules without a priori knowledge of their targets. This forward pharmacological approach has the advantage of directly identifying new pharmacological agents that are active on complex biological processes. However, because of experimental feasibility, systematic application of this approach is generally limited to small animals such as the roundworm C. elegans and zebrafish, raising the question of whether use of these animals could identify compounds that act on ortholgous mammalian targets. A significant challenge in addressing this question is the determination of the molecular identities of the compounds' targets responsible for the desired phenotypic outcomes. Here we describe a computational approach for target identification based on structural similarities of newly identified compounds to known ligand interactions with mostly mammalian targets. For several of the compounds emerging from a C. elegans phenotypic screen, we predict and confirm mammalian targets using in vitro binding assays. Using genetic and pharmacological assays, we then demonstrate that a subset of these compounds alter C. elegans feeding rates through the C. elegans counterparts of the predicted mammalian targets.

Testing SEA Predictions of Mammalian Targets in Vitro
To test the SEA predictions, we assayed selected molecules in vitro against their putative mammalian targets. Given the many potential ligand-target interactions (79 ligands on 572 targets, 1,024 overall pairs), we only tested compound-target interactions for which established assays were readily accessible. We tested 16 different compounds against 20 targets (21 ligand-target pairs) ranging from G-protein coupled receptors (GPCRs) and nuclear hormone receptors (NHRs) to kinases and phosphatases ( Figure 2, Table S1). Nine compounds had significant activity at a 10 mM assay concentration against nine predicted targets, including the dopamine, tachykinin, oxytocin, and metabotropic glutamate receptors, the Flt-3 receptor tyrosine kinase, PI3-kinases, and the NHRs peroxisome proliferator activator receptor-gamma (PPARc) and the androgen receptor ( Figure 2). Significant activity was not observed for 12 target-ligand predictions in vitro (Table S1). A final ligand-target prediction did not show activity on the human receptor in vitro, but was confirmed by chemical-genetic epistasis in C. elegans (below). This 43% in vitro hit rate resembles that observed for chemoinformatic linkage of human drugs to new [37] and to adverse drug reaction targets [41].
The potencies of the hits ranged from 8.6 nM to 9.8 mM in full concentration response analysis ( Figure 2, Figures S1, S2, S3). As expected, there was no correlation between E-value and potency, since target affinity is not considered when calculating ligand similarities. In several cases, SEA successfully predicted the overall target family but not the exact isoform. For instance, H6 and B16 were predicted to antagonize tachykinin receptor-3 (Tkr-3) and metabotropic glutamate receptor 5 (mGluR-5), respectively, but instead antagonized Tkr-1 and mGluR-8 ( Figure 2). Indeed, compared to other mGluR antagonists, B16 appears to be uniquely specific for the mGluR8 isoform (Table S2).

Identification of C. elegans Feeding-Regulatory Compounds
To search for feeding regulatory compounds whose SEApredicted targets could be confirmed by direct testing against mammalian targets in vitro, we focused on the compounds listed in Figure 2. We found that B16, H6, F15, and D20 each increased the pharyngeal pumping rate ( Figure 3A). Our standard assay for pharyngeal pumping rate is a real-time assay over a short time interval (10 s), in which contractions of the posterior bulb of the pharynx are manually counted. The rapid assay is particularly amenable to measuring pumping rate in large numbers of animals and experimental conditions. However, to alleviate concerns that the short duration of the assay may capture an unrepresentative aspect of the pumping rate, we compared its results with those obtained over a longer time course (60 s) using time-lapse microscopy ( Figure S4, Movies S1, S2, S3, S4, S5). The long time course study revealed that for young, egg-laying gravid adults foraging on E. coli OP-50 lawns, the C. elegans pharynx contracts almost continuously with occasional brief pauses of less than a second, similar to that seen during short-term measurements. Importantly, the results of the long-term video measurements confirmed our short-term manual protocol that B16, H6, F15, and D20 all increase pharyngeal pumping in a range of 7%-12%. The percent increase in pharyngeal pumping rate caused by these compound treatments is within the physiological range C. elegans exhibits in response to food following a fast, or that reported with animals treated with serotonin [17,27], one of the best characterized modulators of C. elegans pharyngeal pumping. While reduced pharyngeal pumping could reflect deleterious effects on animal health such as a general disruption of neuromuscular junctions, increased pumping is less likely to be due to such nonspecific effects.
To determine whether the feeding increasing effects of B16, H6, F15, and D20 were dependent on specific developmental stages, we evaluated their effects when administered at different developmental exposure times ( Figure 3B). C. elegans treated with compounds beginning at the first larval stage (L1) assayed as L4 animals exhibited similar percentage increases in feeding in response to B16, H6, F15, and D20 as animals that were allowed an extra day under treatment then assayed as adults. L4 animals do have a basal pumping rate that is lower than adults, and this is reflected in different absolute pump counts for the two stages ( Figure 3B). Developmental exposure to the compounds was not required to elevate the pumping rate, since day 1 gravid adults raised without compound exposure still exhibited elevated pumping once exposed to these compounds. The feeding elevating effects of B16, D20, and F15 were notable within 1 h of exposure but that of H6 required ,16 h ( Figure 3B).

Pharmacological Similarity Predictions Identify Feeding Regulatory Genes in C. elegans
To ascertain whether the feeding increasing compounds mediated their effects through the targets identified in Figure 2, we used genetic epistasis to test target engagement in vivo. We first examined compound D20, which was predicted and shown in vitro to act on the human Flt-3 receptor tyrosine kinase, a member of the PDGF-b receptor superfamily that is involved in the early stages of hematopoiesis and is active in certain cancers [42]. While C. elegans do not have hematopoiesis, their genome encodes dozens of tyrosine kinase domains with sequence similarity to the human Flt-3 receptor with no one sequence being an obvious candidate (Table S3). To determine whether D20 interacted with any of these kinase domains to regulate feeding, we measured the pharyngeal pumping rates of RNAi-treated populations, treated with either D20 or DMSO as a vehicle control. We reasoned that if D20 exerts its effects through a receptor tyrosine kinase in C. elegans, inactivation of such a receptor should mimic the effects of D20 on feeding and, importantly, render pharyngeal pumping insensitive to further modulation. In contrast, combined pharmacological and genetic perturbations that act through independent pathways will exhibit additive or synergistic effects on feeding when combined.
For each of 26 receptor tyrosine kinases with significant BLAST similarity to the Flt-3 receptor, we examined the effects of their gene knockdowns on pharyngeal pumping with or without D20 treatment ( Figure 4A). Among the 26 receptor tyrosine kinases, RNAi exposure of only one resulted in an interaction insensitive to D20 treatment-that of the VEGF-related receptor encoded by ver-3 ( Figure 4A). C. elegans subjected to ver-3 RNAi exhibit an elevated pharyngeal pumping phenotype relative to vector controltreated animals, thus mimicking the effects of D20 treatment ( Figure 4A). While some of the other kinase knockdowns elevated the feeding rate, they all remained sensitive to further modulation by D20. For example, while F09A5.2 RNAi and frk-1 RNAi each caused increased pumping relative to RNAi vector control, D20 treatment further increased pumping rates of these RNAi-treated animals ( Figure 4A). Thus, the resistance of ver-3 mutants to further enhancement of feeding by D20 was not simply due to an upper physiological limit on the pumping rate. Because pharyngeal pumping is regulated by the nervous system and RNAi is sometimes ineffective at gene knockdown, particularly in the C. elegans nervous system [43], it is possible that some of the intended gene products could not be sufficiently knocked down by our RNAi strategy. However, using RNAi to knock down gene expression products of each of ver-3, ver-2, egl-15, and vab-1, all of which have reported nervous system expressions (Table S3), led to effects on pharyngeal pumping that were similar to those obtained when we examined mutants in each of these genes ( Figure 4C). Thus, the combination of in vitro binding assays and patterns of phenotypic interactions in vivo strongly supported the notion that D20 mediates its feeding increasing effects in a ver-3-dependent mechanism.
To further compare the pharmacological parallels between Flt-3 inhibitors and pharyngeal pumping, we tested a known Flt-3 receptor inhibitor (59-fluoroindirubinoxime: 5-flurox, Figure 4B) for its effects on pharyngeal pumping. D20 induces a dosedependent increase in the pharyngeal pumping rate with an EC 50 of 600 nM ( Figure 4B). Similarly, 5-flurox mimicked the dosedependent effect of D20 with an EC 50 of 40 nM ( Figure 4B). The increased in vivo efficacy correlates with the measured in vitro activities of the Flt-3 receptor for D20 (K i = 165 nM) and 5-flurox (IC 50 = 15 nM) [44]. Finally, joint D20 and 5-flurox treatment did not increase pharyngeal pumping rate beyond that seen by individual compound treatments ( Figure S5A). These combined results suggest that a target pharmacologically similar to the Flt-3 receptor exists in C. elegans to regulate pharyngeal pumping.
To verify the RNAi results and further ascertain that D20 and 5-flurox elicit similar effects on C. elegans feeding behavior, we focused on ver-3 and its closest family members, ver-1, ver-2, ver-4, and egl-15. Similar to the RNAi results, mutants in either ver-2 or ver-3 exhibited elevated rates of pharyngeal pumping, while those of ver-1 and ver-4 resembled wild-type, and egl-15 mutants had reduced rates of pumping ( Figure 4C). As with D20, the elevated feeding rates of ver-3 mutants were insensitive to further increase with 5-flurox treatment ( Figure 4C). In contrast, treatment of egl-15, ver-1, and ver-4 receptor mutants with either D20 or 5-flurox led to similar increases in pumping rate despite different basal pumping rates of these mutants. Treatment of the ver-2 mutants, whose basal pumping rate was increased relative to WT, returned to untreated, WT rates with either compound, recapitulating the effect observed when D20 treatment was combined with ver-2(RNAi) ( Figure 4A). The reason for this antagonistic relationship is unclear to us but may reflect a dependence of D20's feeding elevated phenotype on intact ver-2 signaling, perhaps due to a compensatory mechanism between the different receptors. These observations support the notion that the tyrosine kinase receptor VER-3 is pharmacologically orthologous to the human Flt-3 receptor and responsible for D20's feeding phenotype.
SEA predictions may prove informative for finding mechanistic targets in C. elegans even if they fail to modulate the predicted mammalian targets. This appears to be the case for K9, an analog of D20 also predicted to inhibit Flt-3 kinase (Table S1, Figure 4C). K9-treated C. elegans resembled D20-treated animals in the dosedependent increase in feeding rate observed ( Figure 4D). In addition, K9 exhibited the same genetic interactions as D20-and 5-flurox-treated animals: wild-type, egl-15, ver-1, and ver-4 mutants all exhibited elevated pharyngeal pumping ( Figure 4C); ver-3 mutants were insensitive to treatment; and the fast pumping ver-2 mutant reverted to the wild-type rate upon K9 treatment. Therefore, the similarity of the pharmacological response between The effects of the compounds on the pharyngeal pumping rate when exposed for differing developmental periods was evaluated for C. elegans exposed to each 10 uM of each compound during different times: L1 to L4 (2 d at 20uC), L1 to gravid adult (3 d at 20uC, and naïve day 1 gravid adults exposed to B16, F15, and D20 for 1 h, H6 for 16 h). The pharyngeal pumping rate of 10-13 animals per condition was quantified. Error bars represent the standard deviation. *p,0.01: ANOVA, Dunnett's multiple comparisons test. In (B) gravid adults exposed to H6 for 16 h was compared to DMSO 16 h (t test: two tailed *p,0.01). doi:10.1371/journal.pbio.1001712.g003 Pharyngeal pumping rates of wild-type C. elegans treated with serial 3-fold dilutions of 5-flurox or D20. (C) The pharyngeal pumping rates of wildtype, egl-15(n484), ver-1(ok1738), ver-2(ok897), ver-3(ok891), and ver-4(ok1079) mutant animals cultured on 0.1% DMSO, 10 mM D20, 10 mM K9, or 2 mM 5-flurox. Error bars represent the standard deviation. Significance levels: **p,0.001 were determined by one-way ANOVA using Bonferroni's multiple the C. elegans ver-3 and the human Flt-3 receptors, while substantial, is clearly not identical.
Based on the SEA and in vitro binding data, we next examined the possibility that an oxytocin receptor-like system may underlie the feeding regulatory effects of F15 (Figure 2). The oxytocin receptor system modulates mammalian feeding [45][46][47], but the existence of a C. elegans oxytocin receptor ortholog with a role in feeding regulation has not been defined. Both F15 and a comparison test. (D) C. elegans were treated with serial 3-fold dilutions of K9 and the pharyngeal pumping rate was quantified at each dose. Error bars represent the s.e.m. In (A-D) 10-20 animals were evaluated per condition. doi:10.1371/journal.pbio.1001712.g004 structurally independent oxytocin receptor antagonist L-371257 exhibited dose-dependent increases in C. elegans pharyngeal pumping with in vivo EC 50 's of 400 and 10 nM, respectively ( Figure 5A). Like D20 and 5-flurox, their in vivo C. elegans efficacies parallel their in vitro affinities toward the human receptor (Oxtr K i 's for F15, 1.6 mM and for L-371257, 4.6 nM) [48]. Furthermore, simultaneous treatment with F15 and L-371257 did not further elevate pumping (Figure S5B), suggesting the existence of a target with pharmacological similarity to the human oxytocin receptor that regulates feeding.
BLAST comparisons suggest multiple C. elegans sequences with similarity to the human oxytocin receptor. These include an oxytocin/vasopressin-like receptor NTR-1 (T07D10.2) that modulates male mating behavior [49] and associative learning [50], a tachykinin receptor-like protein TKR-1, and three related receptors (GNRR-1, -2, -3) (Table S3). We examined the pharyngeal pumping rates in ntr-1 and the gnrr-1, -2, and -3 mutants ( Figure 5B) and animals treated with tkr-1 RNAi ( Figure 5C). Both ntr-1 and gnrr-2 mutants exhibited wild-type feeding rates and were sensitive to F15-and L-371257-induced pharyngeal pumping increases. Both gnrr-1 and -3 mutants ( Figure 5B) as well as tkr-1(RNAi) animals ( Figure 5C) exhibited elevated pumping in the absence of compound treatment. However, only the pharyngeal pumping rate of the gnrr-1 mutants was resistant to further increase by F15 and oxytocin-antagonizing L-371257 treatments ( Figure 5B,C). Turning to the tachykinin system, each of wild-type, ntr-1, gnrr-1, -2, and -3 mutants responded with increased feeding on treatment with SB222200, a high-affinity human tachykinin receptor antagonist ( Figure 5C). In contrast to their F15 sensitivity, tkr-1(RNAi)-treated animals were indeed insensitive to the pharyngeal pumping rate increases elicited by either SB222200 or H6 treatment ( Figure 5C), a compound identified in our C. elegans phenotypic screen and shown in vitro to act on a human tachykinin receptor (Figure 2). Wild-type H6-dosed animals were insensitive to SB222200-induced feeding increase ( Figure S5C) consistent with their in vitro activities as tachykinin receptor antagonists. Together these results indicate that a tachykinin-like and an oxytocin-like receptor pathway function in parallel to regulate pharyngeal pumping in C. elegans. Despite relative similarities in sequence and feeding phenotype, gnrr-1 and tkr-1 mutants were differentiated by their antagonists and pharmacologically linked to the human oxytocin and tachykinin receptors, based on responsiveness to F15/L-371257 and H6/SB222200, respectively.
Finally we tested whether B16, the mammalian mGluR-8 inhibitor, regulates the pharyngeal pumping rate through any of the C. elegans metabotropic glutamate receptors. The C. elegans genome encodes three orthologs of human mGluR-8: mgl-1, -2, and -3 [51]. These receptors are expressed in the C. elegans nervous system and regulate diverse aspects of C. elegans behavior and physiology, but have not previously been implicated in feeding [52]. We found that mgl-2 but not mgl-1 mutant animals exhibited an elevated pharyngeal pumping rate in the absence of B16 treatment ( Figure 5D). The elevated feeding of mgl-2 mutants resembled that of WT animals treated with B16, and that of animals treated with MMPIP [53], a human mGluR-7 allosteric antagonist (IC 50 's, 26-220 nM) structurally distinct from B16 ( Figure 5D). Combined treatment of wild-type animals with B16 and MMPIP resulted in no further increase over either alone ( Figure S5D). The pharyngeal pumping rate of mgl-2 mutant animals, unlike both WT and mgl-1 mutant animals, was unaffected by treatment with either B16 or MMPIP, consistent with the notion that both agents mediate their feeding phenotype though inhibition of MGL-2. Similar to WT and mgl-1 mutants, mgl-2 mutant animals further elevated pumping when treated with the Oxtr/GNRR-1 antagonist F15 ( Figure 5D). The additive effect of F15 on the mgl-2 mutants' pharyngeal pumping rate distinguishes its underlying biological mechanism from that of the mGluR/mgl-2 antagonists.

The Regulatory Relationships of the Newly Identified Feeding Pathways
To further examine the in vivo specificity of the compoundinduced feeding phenotypes and determine whether each target regulates feeding independently of one another, we examined the feeding phenotypes of all possible binary combinations of compounds with the high pumping mutants identified by this study (Table S4). The resulting 54-interaction matrix ( Figure 6A) classifies interactions between compounds and mutants based on the sensitivity of a mutant to a compound's effect on pharyngeal pumping. Considering that relative to wild-type animals on vehicle control each of the gene knockdowns and compound treatments individually were sufficient to cause feeding increases, two patterns of interactions were expected: interactions where the feeding increasing effects are additive, representing likely parallel mechanisms of actions, and those that are nonadditive, suggesting a single regulatory pathway. Pharyngeal pumping rates that exceeded those observed in the vehicle-treated mutant controls are classified as additive. For example, both H6 and SB222200 increase the pharyngeal pumping rates of mgl-2, ver-3, gnrr-1, ver-2, and gnrr-3 mutants, a series of additive interactions, but not that of animals treated with tkr-1 RNAi, a genetic inactivation in their common target.
We also noted 10 interactions in which elevated pumping rates of specific mutants were lowered upon compound treatment. For instance, while each of D20, K9, and 5-flurox treatments alone, as does inactivation of ver-2, elevate feeding rates in wild-type and in many mutant backgrounds, treatment of ver-2 mutants with each of these compounds lowers feeding ( Figure 4C, Figure 6A). Similar antagonistic patterns were also seen, for example, when ver-3 mutants were treated with the Oxtr antagonists F15 and L-371257. While the precise reasons for antagonistic interactions are not known, one likely possibility is that distinct signaling pathways normally act in compensatory manners that are revealed by the simultaneous inhibition of both of these pathways.
Strikingly, we noted that in all cases where compounds share a common target, for example F15, L-371257, and the oxytocin receptor, both compounds exhibited identical interactions, unique to each compound pair across the mutant series. This includes additive, nonadditive, and antagonist interactions. This result strongly supports the hypothesis that the compounds share a common in vivo target identified by this study that drives the feeding phenotype. Beyond specificity, this matrix also indicates a higher level pathway organization of these mutants ( Figure 6A). Both antagonists of the mgl-2 pathway (B16 and MMPIP) and VER-3 inhibitors (D20, K9, and 5-flurox) interact nonadditively and reciprocally with ver-3 and mgl-2 mutants, indicating that these gene products may act in a common pathway. However, the additive effect of B16/MMPIP versus the antagonistic phenotypic effect of D20/K9/5-flurox activity on gnrr-1 mutants is consistent with their targets being distinct entities. In addition, GNRR-1 antagonists (F15 and L-371257) interact nonadditively with ver-2 mutants and antagonistically with ver-3, recapitulating the antagonistic interaction of VER-3 inhibitors with ver-2 mutants ( Figure 4C, Figure 6A). This indicates that a second feeding regulatory pathway combines GNRR-1 and VER-2 signaling.
To evaluate whether the chemical-genetic epistasis interactions observed were the result of pharmacological peculiarities or could be confirmed by standard genetic epistasis tests, we tested genetic interactions in the implied mgl-2/ver-3 and ver-2/gnrr-1 pathways. The feeding rate of ver-3 mutants was insensitive to effects of mgl-2(RNAi), but was reduced by ver-2(RNAi) ( Figure 5B). Mgl-2(RNAi) interacted additively with gnrr-1 mutants to increase the pumping rate, and ver-3(RNAi) antagonized the feeding increasing effects of the gnrr-1 mutation ( Figure 5B). Ver-2(RNAi) interacted nonadditively with the gnrr-1 mutants and was antagonistic to mgl-2 mutants ( Figure 5B). Similar interactions between gnrr-1 mutants and ver-3(RNAi), ver-2(RNAi), and mgl-2(RNAi) were observed by measuring pumping over longer intervals by time lapse microscopy ( Figure S6). These results indicate that the chemical-genetic interactions observed in this study accurately predict the interactions of loss-of-function perturbation combinations on the pharyngeal pumping rate. As in chemical-genetic interactions, examination of mutant combinations also indicated that mgl-2/ver-3 and ver-2/gnrr-1 function in parallel pathways but with significant crosstalk in regulating the pharyngeal pumping rate.

Discussion
Whole organism phenotypic screens retain key advantages of classical pharmacological approaches, such as the discovery of compounds that are biologically active and that alter physiologically intact, integrated circuits without predisposed conceptions as to which circuits should be targeted. To prove biologically informative, this forward pharmacological approach requires the determination of in vivo molecular targets as well as the mode of action by which the phenotype is modulated [54]. Five key observations emerge from this study. First, by chemoinformatic inference, targets may be rapidly prioritized for experimental testing on isolated receptors in vitro. Whereas this method did not always succeed, the confirmation of the predicted targets was high enough, at 43%, to be practical. Second, as the identified targets are overwhelmingly mammalian, the ease of phenotypic screening strategies in C. elegans can be linked to identification of humanrelevant targets. Third, in a model system such as C. elegans, the relationships of orthologous targets to in vivo phenotypes can be parsed by applying the rationale of genetic epistasis analysis. This is critical for unambiguous in vivo establishment of mechanisms of action, as in vitro activities of even highly characterized compounds are only suggestive of the in vivo efficacy targets. Fourth, despite significant differences in primary sequence identity of the targets, the in vivo efficacy in C. elegans can reflect the in vitro activity against human targets. Finally, the chemical-genetic interactions described in this study illuminate four previously uncharacterized, parallel molecular targets that regulate food intake. Together, these findings demonstrate an experimental and computational path from phenotypic screens in C. elegans to the discovery of humanrelevant targets and elucidation of mechanisms of actions of newly identified compounds in C. elegans.
Most drugs interact with multiple targets in vivo [41], which can confound the assignment of phenotypic effects to particular targets. This can be especially true of compounds emerging from screening campaigns prior to any efforts aimed at optimizing the potential specificities of compounds. An advantage of a pharmacological approach in C. elegans is that genetic perturbations can test target engagement in vivo. Identifying a chemical-genetic epistatic interaction does not imply that a given compound has absolute specificity for a particular target in a biological system. However, the epistasis interaction does confirm that the compound induces a particular phenotype specifically through its interaction with the pathway defined by the genetic perturbation. Thus, in a manner similar to classical double mutant analysis, chemical and genetic interactions are combined to interrogate the pharmacological relevance of hypothesized target interactions to specific phenotypes.
Our findings suggest that there is a substantial pharmacological intersection between mammals and C. elegans. These results indicate that a C. elegans phenotypic screen can lead to identification of compounds that are sufficiently similar to the mammalian pharmacopeia to allow for prediction and confirmation of their interactions with mammalian targets. In turn, evolutionary conservation of these targets, in both sequence and ligand recognition, makes it possible to accurately predict the C. elegans target whose perturbation results in the phenotype. It could have easily been the case that the compounds emerging from a C. elegans screen are so diverse as to belie prediction of targets, and that the C. elegans phenotypes could be irrelevant to the human target space, or could reflect new targets not previously seen. As such, it may be astonishing that this approach worked at all. In fact, 79 of 84 active compounds could be chemoinformatically linked to human targets, suggesting that even a diversity library retains substantial and, for our purposes, highly useful biases towards previously ''liganded'' targets. Moreover, whereas these in vitro mammalian targets need have no relevance for C. elegans in vivo pharmacology, for compounds with confirmed activity in vitro, orthologous targets were indeed found to mediate their C. elegans phenotypes.
Whether the specific compounds identified from the C. elegans screen act on mammalian feeding and regulatory systems remains to be determined. However, there are already some hints that functionally related circuits modulate feeding behavior in both mammals and C. elegans. For instance, several independent studies in chickens, mice, and rats indicate that administration of oxytocin reduces food intake [55]. Oxytocin appears to be a target of satiety signals since a lipid-related signal, oleoylethanolamide, positively requires intact hypothalamic oxytocin signaling to mediate its anorexigenic effects [45]. Conversely, hyperphagia associated with a high-fat diet requires synaptotagmin-4-mediated suppression of oxytocin vesicle exocytosis [46]. Furthermore, a key function of neurons of the hypothalamus that stimulate feeding involves the inhibition of a separate population of oxytocin neurons [47]. Thus, analogous to C. elegans gnrr-1, signaling through the oxytocin receptor is a negative regulator of food intake in mice. Similarly, central administration of substance P, a TKR ligand, inhibits feeding in chicks [56], consistent with the negative feeding regulatory role of the TKR-1 in C. elegans. While the Flt-3 receptor has no known role in the mammalian nervous system, closely related growth factor receptors, such as the platelet-derived growth factor-b (PDGF-b) receptor, are expressed in the hypothalamus, and administration of PDGF-b depresses food intake and anti-PDGF-b antibodies elevate food intake in rats [57]. This resembles our observation that loss-of-function in a C. elegans receptor tyrosine kinase (ver-3) through either mutation or pharmacological inhibition elevates pharyngeal pumping. These targets may thus have ancient evolutionary origins in the regulation of feeding behavior.
Key weaknesses of our approach merit discussion. First, whereas it is comforting that we can predict targets for most biologically active synthetic compounds even in a ''diversity'' library, this also reflects the restricted chemical-target space in which the field is working. A library composed of genuinely novel chemotypes might be more likely to illuminate unprecedented targets, a widely desired goal of the field. On the other hand, any such truly diverse library risks missing that small part of chemical space that is relevant to terrestrial biology, the bias towards which is, after all, a pragmatic advantage of the current libraries [58]. Second, even within this restricted ligand-target space, the chemoinformatic linkage was far from perfect, and about half of the tested compounds remain unlinked to predicted targets. While the inferential computational approach cannot replace experiment, it is a rapid, comprehensive, and quantitative assessment of biologically active small molecules with unknown mechanism. These predictions generate testable hypotheses with regards to in vivo function. Third, while for feeding, the in vitro and in vivo data provided a compelling case for mechanisms of actions of F15, H6, B16, and D20, it remains to be determined whether these compounds also act on as-of-yet undetermined molecular pathways to alter other biological processes in C. elegans. Finally, while the link between compounds, targets, and C. elegans phenotypes now seems strong for several of the active compounds, linkage between C. elegans and mammalian in vivo pharmacology remains to be drawn.
In summary, whereas chemoinformatic linkage retains important liabilities, its success rate here and in earlier studies [37][38][39]41] is high enough to be pragmatic for target hypothesis testing. Similarly, despite critical differences between C. elegans and mammals, some of them target-based, some biology-based, the targets and ligand networks for a substantial number of small molecules are conserved enough to allow target and phenotypic association across phyla. We envision that such chemical-genetic epistasis maps could be extended to saturation mapping of the pharmacological target networks underlying feeding regulation and other processes in C. elegans. The amenability of C. elegans to genetic manipulation and pharmacological screening may find broad utility as a means to identify new small molecules with interesting phenotypes and human-relevant targets.
For the target panel, we first used ChEMBL_7 (released November 11, 2010) and later moved to an updated 2,482 molecular target panel derived from ChEMBL_11 (released August 9, 2011). We filtered reference ligands by molecular weight (#1,000 Da) and by reported affinity (#10 mM), and then subjected all ligand structures to cleaning, standardization, and deduplication as before [39].

In Vitro Compound Assays against Mammalian Targets
Tachykinin, ghrelin, and calcium-sensing receptor activity were measured using a cell-based Ca flux assay by Multispan, Inc. (Hayward, California). Cholescystokinin A and B receptors were assayed for effects on cAMP production, and nicotinic acid receptor activity was measured in a cell-based assay of forskolinstimulated cAMP production by Multispan, Inc. Calcium-sensing receptor, PP2A phosphatase activity assays were performed by CEREP (Celle l'Evescault, France) using PP2A from human erythrocytes. Phospholipase C from Bacillus cereus was assayed by CEREP, Inc. using glycero-phosphatidyl ethanolamine as a substrate, monitoring diacyl glycerol production. Radioligand binding assays for PPAR-a, c, d and the androgen receptor were performed by CEREP, Inc. CARNA Biosciences (Kobe, Japan) performed in vitro kinase activity assays using recombinant catalytic domains and measured phosphorylation of an Src-derived peptide for Flt-3 and phosphatidyl inositol for PI3KCA. Radioligand binding assays for the oxytocin receptor, D4 dopamine receptor, cannabinoid CB2 receptor, and angiotensin type I and II receptors were performed as described previously, as were cell-based activity assays for mGluR 1a, 2, 4, 5, 6, and 8 [60][61][62].

RNA Interference
Synchronized first larval stage C. elegans were cultured for 3 d at 20uC on a lawn of HT115 E. coli induced with IPTG to express double-stranded RNAi as described [64].

Pharyngeal Pumping Assay
Compounds at 1,0006 stock concentrations (0.1-10 mM) in DMSO or an equal volume of DMSO were diluted in a suspension of E. coli OP-50 (100 ml of a 36concentrated overnight culture in LB broth) and absorbed as a single drop onto 3.5 cm plates containing 2.5 ml of NGM agar forming a well-defined lawn of bacteria. For developmental exposures, 20-30 synchronized first larval stage C. elegans derived from alkaline hypochlorite treatment of gravid adults were applied to the preseeded plates and cultured until assay at mid-L4 stage (2 d, 20uC) or as day 1 gravid adults (3 d, 20uC). For naïve adult exposures, synchronized first larval stage animals were cultured on NGM-agar plates seeded with E. coli OP-50 for 3 d at 20uC, then transferred to assay plates, and pumping was assayed 1 to 16 h later. Comparisons between vehicle-treated and compound-treated animals were always performed between animals at the same developmental stage and same exposure time. Pharyngeal pumping was counted by live observation at 115-2006 magnification using a stereo microscope and recorded at 10 s intervals using a manually controlled digital cell counter. Alternatively, C. elegans pumping was recorded for longer intervals (30-60 s) using bright-field time-lapse microscopy at 1206 magnification with a 20 images per second acquisition rate. The resulting movies were analyzed manually at a playback rate of 10 images per second.

Statistics
Significance was determined by one-way ANOVA applying a Dunnett's posttest when comparing multiple treatments to a single control and a Bonferroni posttest when comparing the more than two treatments against one another. For pairwise nonrepeated measured comparisons, a student's t test was used. To express data as a percentage of control, the pumping rates of compound-treated animals were divided by the mean pumping rate of the DMSOtreated wild-type animals measured in the same experiment, unless indicated otherwise. Error bars on the control samples indicate the variation around that mean, which is utilized in all statistical calculations. Table S1 Target predictions that could not be confirmed in vitro.   Figure 6A.