Identification of Novel Allosteric Modulators of Metabotropic Glutamate

May 27, 2019 - (37) In brief, the test compound was applied to cells expressing rat mGlu5 in .... length independent from size and composition of the ...
0 downloads 0 Views 1MB Size
Download PDF
Research Article Cite This: ACS Chem. Neurosci. XXXX, XXX, XXX−XXX

pubs.acs.org/chemneuro

Identification of Novel Allosteric Modulators of Metabotropic Glutamate Receptor Subtype 5 Acting at Site Distinct from 2‑Methyl-6-(phenylethynyl)-pyridine Binding Mariusz Butkiewicz,†,§ Alice L. Rodriguez,‡,⊥ Shane E. Rainey,‡,⊥ Joshua Wieting,‡,⊥ Vincent B. Luscombe,‡,⊥ Shaun R. Stauffer,†,‡,⊥ Craig W. Lindsley,†,‡,⊥ P. Jeffrey Conn,‡,⊥ and Jens Meiler*,†,‡,§,∥ †

Department of Chemistry, ‡Department of Pharmacology, §Center for Structural Biology, ∥Institute of Chemical Biology, and Center for Neuroscience Drug Discovery, Vanderbilt University, Nashville, Tennessee 37232, United States

Downloaded via GUILFORD COLG on July 18, 2019 at 12:59:44 (UTC). See https://pubs.acs.org/sharingguidelines for options on how to legitimately share published articles.

⊥

S Supporting Information *

ABSTRACT: As part of the G-protein coupled receptor (GPCR) family, metabotropic glutamate (mGlu) receptors play an important role as drug targets of cognitive diseases. Selective allosteric modulators of mGlu subtype 5 (mGlu5) have the potential to alleviate symptoms of numerous central nervous system disorders such as schizophrenia in a more targeted fashion. Multiple mGlu5 positive allosteric modulators (PAMs), such as 1-(3-fluorophenyl)-N-((3-fluorophenyl)-methylideneamino)methanimine (DFB), 3-cyano-N-(1,3-diphenyl-1H-pyrazol-5-yl)-benzamide (CDPPB), and 4-nitro-N-(1,3-diphenyl-1Hpyrazol-5-yl)-benzamide (VU-29), exert their actions by binding to a defined allosteric site on mGlu5 located in the seventransmembrane domain (7TM) and shared by mGlu5 negative allosteric modulator (NAM) 2-methyl-6-(phenylethynyl)pyridine (MPEP). Actions of the PAM N-{4-chloro-2-[(1,3-dioxo-1,3-dihydro-2H-isoindol-2-yl)methyl]phenyl}-2-hydroxybenzamide (CPPHA) are mediated by a distinct allosteric site in the 7TM domain different from the MPEP binding site. Experimental evidence confirms these findings through mutagenesis experiments involving residues F585 (TM1) and A809 (TM7). In an effort to investigate mGlu5 PAM selectivity for this alternative allosteric site distinct from MPEP binding, we employed in silico quantitative structure−activity relationship (QSAR) modeling. Subsequent ligand-based virtual screening prioritized a set of 63 candidate compounds predicted from a library of over 4 million commercially available compounds to bind exclusively to this novel site. Experimental validation verified the biological activity for seven of 63 selected candidates. Further, medicinal chemistry optimizations based on these molecules revealed compound VU6003586 with an experimentally validated potency of 174 nM. Radioligand binding experiments showed only partial inhibition at very high concentrations, most likely indicative of binding at a non-MPEP site. Selective positive allosteric modulators for mGlu5 have the potential for tremendous impact concerning devastating neurological disorders such as schizophrenia and Huntington’s disease. These identified and validated novel selective compounds can serve as starting points for more specifically tailored lead and probe molecules and thus help the development of potential therapeutic agents with reduced adverse effects. KEYWORDS: virtual screening, machine learning, quantitative structure−activity relationship, high-throughput screening, cheminformatics, PubChem, BioChemistryLibrary, G protein-coupled receptor, GPCR, metabotropic glutamate receptor, mGlu, 1-(3-fluorophenyl)-N-((3-fluorophenyl)-methylideneamino)-methanimine, 3-cyano-N-(1,3-diphenyl-1H-pyrazol-5-yl)-benzamide, 4-nitro-N-(1,3-diphenyl-1H-pyrazol-5-yl)-benzamide, 2-methyl-6-(phenylethynyl)-pyridine, MPEP, positive allosteric modulator, PAM, N-{4-chloro-2-[(1,3-dioxo-1,3-dihydro-2H-isoindol-2-yl)methyl]phenyl}-2-hydroxybenzamide, CPPHA, QSAR

1. INTRODUCTION

Received: May 7, 2018 Accepted: May 27, 2019 Published: May 27, 2019

Metabotropic glutamate (mGlu) receptors are part of the G protein-coupled receptor (GPCR) family also known as seven© XXXX American Chemical Society

A

DOI: 10.1021/acschemneuro.8b00227 ACS Chem. Neurosci. XXXX, XXX, XXX−XXX

Research Article

ACS Chemical Neuroscience

O’Brien et al.10 In a later study, Chen et al.12 characterized a novel mGlu5 PAM labeled 4-nitro-N-(1,3-diphenyl-1H-pyrazol-5-yl)-benzamide (VU-29), which was shown to act at a site overlapping with the binding site of MPEP. The same study demonstrated that the structurally distinct PAM CPPHA potentiates the mGlu5 response by an alternative mechanism to the one of VU-29. Additionally, a neutral ligand at the MPEP allosteric site termed 5-methyl-2-(phenylethynyl)-pyridine (5MPEP) blocks the VU-29- and CPPHA-induced potentiation of mGlu5 responses. Remarkably, an increase in 5MPEP concentration invokes a parallel rightward shift in the concentration−response curve of VU-29. In contrast, CPPHA potentiation is inhibited by 5MPEP noncompetitively.13,21 Furthermore, this evidence of an alternative binding site is supported by mutagenesis experiments where a mutation (A809V) decreases binding of ligands to the MPEP site. The mutation abolishes binding of VU-29 but does not influence the response to CPPHA. Likewise, a second mutation (F585I) eliminates the response to binding CPPHA but does not change binding of VU-29. Both results suggest that CPPHA does not bind at the MPEP site but rather acts at a distinct novel allosteric site on mGlu5 as a potentiator to the receptor. Additional studies have identified other mGlu5 PAMs that interact noncompetitively with the MPEP site including VU035712114 and (N-(4-chloro-2-((4-fluoro-1,3-dioxoisoindolin-2-yl)-methyl)-phenyl)-picolinamide (NCFP).26 In the case of VU0357121, while similar in structure to CPPHA, radioligand binding and mutagenesis studies suggest the compound and its analogs bind to a non-MPEP and nonCPPHA site. In studies by Noetzel et al.,26 the CPPHA analog NCFP was found to likely bind to the previously mentioned CPPHA site on mGlu5. It exhibits stronger mGlu5 subtype selectivity in comparison to CPPHA, thereby increasing its suitability for studies of mGlu5 influence in the CNS. NCFP was shown to potentiate multiple responses to mGlu5 in a manner similar to CPPHA; however, it was not able to potentiate LTD and LTP in the hippocampus, demonstrating the differential effects that mGlu5 PAMs can have on responses in the CNS. While the functional selectivity or stimulus bias demonstrated by compounds such as these may complicate drug discovery, it is possible these differential effects could result in pharmaceutical intervention tailored to individual needs. 1.3. Drug Discovery Guided by in Silico Virtual Screening Can Prioritize Modulation-Specific mGlu Potentiators. Previous quantitative structure−activity relationship (QSAR) studies were successful in identifying lead scaffolds serving as starting points for successful probe development campaigns. A research study by Mueller et al.27 identified novel mGlu5 PAMs based on a high-throughput screen (HTS) of a diverse library of 144 475 substances, which revealed 1382 compounds as PAMs. A subsequently trained QSAR model with a theoretical enrichment ratio of up to 38 for an independent data set was applied to screen a database of approximately 450 000 commercially available drug-like compounds. A set of 824 compounds was acquired for testing based on the highest predicted potency values. Experimental validation confirmed 28.2% (232/824) of these compounds with various activities at mGlu5. These results represent an enrichment factor of 23 for pure potentiation of the mGlu5 glutamate response and 30 for overall mGlu5 modulation activity when compared with those of the original mGlu5 experimental screening data (0.94% hit rate).

transmembrane (7TM) domain receptors. As the primary excitatory neurotransmitter in the mammalian nervous system, glutamate acts as an endogenous ligand to mGlu receptors binding orthosterically at a large extracellular globular domain. mGlu activation is tied to a downstream effector mechanism through guanine nucleotide binding proteins. The mGlu family consists of eight subtypes classified into three major groups.1,2 Group I includes mGlu1 and mGlu5, which couple primarily to Gq/11 and mediate IP3/Ca2+ signal transduction,3 whereas group II (mGlu2 and mGlu3) and group III (mGlu4, mGlu6, mGlu7, and mGlu8) mGlu receptors couple primarily to inhibition of adenylyl cyclase.4 Through the diverse physiological roles, heterogeneous distribution, and wide diversity of mGlu receptor subtypes, the possibility arises for new therapeutic agents. These next generation agents have the capacity to selectively interact with a mGlu receptor subtype associated with a single or limited number of central nervous system (CNS) functions. Such targeted therapeutics could have incisive importance in the development of novel treatment strategies for a range of neurological disorders.5−7 1.1. Modulators of mGlu5 Have Promising Potential for Treatment of Schizophrenia and Other Diseases Associated with Cognitive Impairment. Modulators of mGlu5 are of particular interest due to an increasing body of studies providing evidence that they play key roles in potential novel treatment strategies for schizophrenia, Alzheimer’s disease, and other disorders linked to cognitive impairment.8,9 Current mGlu5 positive allosteric modulators (PAMs) have been developed based on multiple scaffolds.10−19 Wellcharacterized structural classes of mGlu5 PAMs include 1-(3fluorophenyl)-N-((3-fluorophenyl)-methylideneamino)-methanimine (DFB), 3-cyano-N-(1,3-diphenyl-1H-pyrazol-5-yl)benzamide (CDPPB), and N-{4-chloro-2-[(1,3-dioxo-1,3dihydro-2H-isoindol-2-yl)methyl]phenyl}-2-hydroxybenzamide (CPPHA).20 Numerous challenges encountered during the development of mGlu5 PAMs have been documented. For example, slight structural changes introduced during chemical optimization often induce a large spectrum of pharmacological responses (DFB series), making prediction of activity difficult. Also, as seen with the CPPHA series, a “flat” structure−activity relationship (SAR) is often confronted, meaning slight structural changes result in complete loss of activity. Despite these difficulties, efforts have led to the development of positive modulators able to allosterically potentiate mGlu5mediated electrophysiological responses in the CNS.10,13,20,21 Furthermore, antipsychotic-like effects of these compounds were confirmed by in vivo studies based on animal models.11,22 More recently, a distinct mGlu5 PAM chemotype was presented by Addex pharmaceuticals (ADX47273).23 The compound showed in vivo efficacy in preclinical behavioral models comparable to other known antipsychotics.24 In general, these mGlu5 PAMs improved cognitive function in animals in which object recognition was impaired24 and demonstrated improved behavioral flexibility.25 These exciting discoveries provide a strong emphasis for mGlu5 PAMs having potential application as novel antipsychotic agents and cognition-enhancing therapeutics. 1.2. Allosteric Modulation of mGlu5 through Multiple Distinct Binding Sites. The primary binding site for many mGlu5 PAMs correlates with the same site for the negative allosteric modulator (NAM) 2-methyl-6-(phenylethynyl)-pyridine (MPEP) located in TM 3, 6, and 7.8 Evidence for a possible distinct allosteric binding site was provided by B

DOI: 10.1021/acschemneuro.8b00227 ACS Chem. Neurosci. XXXX, XXX, XXX−XXX

Research Article

ACS Chemical Neuroscience

Figure 1. Clustering of 130 active and 145 inactive compounds in the CPPHA-like series. (A) Structural similarity of 17 identified clusters. (B) Scaffold representation of each cluster. (C) Cluster composition with respect to actives (green) and inactives (red).

involved in treatment of a range of CNS disorders. Research efforts are focused around the development of allosteric modulators due to receptor subtype selectivity and activity reliance upon receptor activation. However, the current understanding of mechanisms that involve mGlu receptors is limited when mediated in vivo effects are considered. In addition, awareness is increasing that distinct allosteric modulators can induce differential effects on mGlu5-mediated psychological responses in the CNS, referred to as functional selectivity or stimulus bias.30−32 Albeit challenging to characterize, promising evidence has demonstrated the existence of a novel mGlu5 binding site distinct from the MPEP binding site. The identification of novel allosteric potentiators binding to this alternative site will help elucidate the pharmacological significance of binding interactions and bias effects of differential binding sites. Computational techniques like virtual screening are imperative for prioritization and identification of novel selective potentiators. The present study identified 63 compounds through virtual screening and rigorous filtering, which were purchased and experimentally validated (Supplementary Table S2). Approximately 11% (7/63 compounds) exhibited mGlu5 activity upon initial screening. Six compounds were PAMs. This represents a virtual screening hit rate increased by 2 orders of magnitude compared to an initial experimental hit rate of 0.09% = 130 CPPHA-like active molecules/(1382 PAMs + 144 475 inactive molecules). Binding studies suggest the compounds act via a

A subsequent study applied the same approach to investigate the activation of metabotropic glutamate receptor subtype 4 (mGlu4), which has been shown to be efficacious in rodent models of Parkinson’s disease.28 A high-throughput screen identified 434 positive allosteric modulators of mGlu4 out of a set of approximately 155 000 compounds. A QSAR model has a theoretical enrichment of 15-fold when selecting the top 2% of compounds of an independent test data set. The same external commercial database of approximately 450 000 druglike compounds was screened as in the previous study. From these, 1100 predicted active small molecules were tested experimentally using two distinct assays of mGlu4 activity. This experiment yielded 67 positive allosteric modulators of mGlu4 that were confirmed in both experimental systems. Compared to the initial hit rate of 0.3% in the primary screen, this resulted in a 22-fold enrichment. Noeske et al.29 applied a self-organizing map to discern antagonists for mGlu1 and mGlu5. This approach employed topological pharmacophore descriptors which were applied to a compiled library of 338 compounds of noncompetitive antagonists and an external library of 5376 molecules of known drugs and lead candidates for different drug targets. The resulting self-organizing maps were able to differentiate between separate localized distributions for the two mGlu receptor targets. 1.4. Significance. Preclinical studies have implied the role of mGlu receptors as possible candidates for drug targets C

DOI: 10.1021/acschemneuro.8b00227 ACS Chem. Neurosci. XXXX, XXX, XXX−XXX

Research Article

ACS Chemical Neuroscience

used in the studies of refs 28 and 29 (Figure 2b,c). Given this QSAR model, it was possible to identify compounds that bind to a distinct site, different from the known MPEP binding site. Cross-validation was used to ensure the QSAR model was not overfitted. As a quality metric, we computed the receiver operating characteristics (ROC) curve and positive predictive value (PPV = TP/(TP + FP)), shown in Figure 3. The

site distinct from the MPEP binding pocket. These newly identified molecules have the potential to serve as lead compounds in future drug discovery campaigns and can subsequently help to establish specifically tailored therapeutics with reduced side effects.

2. RESULTS With the identification of a putative allosteric mGlu5 binding site different from the MPEP binding site, a series of CPPHA analogs was screened for mGlu5 PAM activity.14,15,33 A total of 275 compounds were screened, and 130 compounds were found to be active (see Table S1). This data set is comprised of CPPHA-like compounds. Structural similarity between the different involved scaffolds is shown in Figure 1. The dendrogram shows the structural similarity of each cluster scaffold with a representative structure. Additionally, for each cluster, the composition of active and inactive compounds is indicated relative to the overall total number of compounds. 2.1. Development of QSAR Model To Prioritize Specific mGlu5 PAMs for Alternative Site Distinct from MPEP Binding. In a previous HTS campaign for mGlu5 modulators,27 a set of 1382 PAMs and 343 NAMs were identified to engage mGlu5 allosterically. Based on these data, we developed two specific QSAR models to predict allosteric mGlu5 PAMs28 and NAMs.29 In this study, a third QSAR model was developed specifically to identify compounds that bind to the putative additional binding site evident through the noncompetitive binding character of CPPHA-like compounds. The data set configuration for all involved QSAR models is presented in Figure 2. To hone in on the attributes of the CPPHA-like compounds, a QSAR model was trained considering only the 130 confirmed CPPHA-like PAMs as active. The 145 CPPHA-like inactive molecules as well as all compounds screened in the HTS campaign were labeled as inactive (Figure 2a). This setup contrasts the QSAR models

Figure 3. ROC curve for the specific mGlu5 PAM model predicting based on CPPHA-like compounds. The ideal PPV curve is given in gray, the actual PPV curve is shown in black, and the true positive rate (TPR) is shown in red. Plotting the TPR (y-axis) against the false positive rate (x-axis) constitutes the ROC curve. The x-axis is on a logarithmic scale.

selective mGlu5 PAM model for CPPHA-like binders achieved a PPV of 73% when considering a false positive rate (FPR) of 0.1%. To put this into perspective, when allowing a FPR of 0.1%, the model successfully predicted 73% of the active compounds in an independent data set not used for training the model. More information on data set composition and regarding the QSAR models for general mGlu5 PAM and NAM prediction can be found in Supplementary Section S3. While retraining the QSAR models, the quality measure results improved, observing PPV rates of 51% (PAM) and 19% (NAM) indicating an improvement compared to the previously published results when applying a similar FPR cutoff. 2.2. Virtual Screening Predicts Site Specific mGlu5 PAMs among Commercially Available Compounds. The compound library eMolecules34 was chosen for virtual screening with the aforementioned QSAR models. It contains over 4 million compounds covering several large vendor compound libraries. To query for selective hits of the mGlu5 PAM CPPHA binding site, the three QSAR models were applied simultaneously to virtually screen for activity of each respective mGlu5 target. An mGlu5 PAM CPPHA site hit was defined by molecules classified as inactive by the general mGlu5 PAM model, inactive by the general mGlu5 NAM model, and active by the selective mGlu5 CPPHA-like PAM model. Through this query scheme, 4719 compounds were predicted to be mGlu5 CPPHA-like PAMs.

Figure 2. Schematics of data set composition of the three QSAR models. (green, labeled active; red, labeled inactive). D

DOI: 10.1021/acschemneuro.8b00227 ACS Chem. Neurosci. XXXX, XXX, XXX−XXX

Research Article

ACS Chemical Neuroscience 2.3. Medicinal Chemistry Filters and Clustering Analysis Narrow down Selective Hit Compounds. All 4719 hits were assessed further by the rapid elimination of swill (REOS) filter35 and pan-assay interference compounds (PAINS) filter36 to remove non-drug-like compounds and frequent false positives. The resulting set of compounds contained 353 molecules, which were further filtered by allowing for not more than one halogen atom. The final set contained 138 molecules, 109 with one halogen atom and 29 with no halogen atoms. A subsequent hierarchical cluster analysis narrowed this set of molecules further down to 103 candidate clusters. The candidate compound with the highest predicted mGlu5 CPPHA-like PAM activity was selected for experimental verification. Taking pricing and availability into consideration, 63 of 103 identified candidate molecules were acquired. To further investigate the chemical similarity between the 63 candidate compounds and the 1382 PAMs from the HTS campaign, an analysis of the similarity between every compound pair was conducted (see Figure 4). The distribution of the highest

similarity of each compound pair clearly indicates a scaffold similarity peak at approximately 40%. The majority of compound pairs exhibit a scaffold similarity below 50%, which is indicative of high scaffold diversity between the 63 candidate compounds and known PAMs. 2.4. Experimental Validation of Identified Compounds by Virtual HTS. The 63 compounds were evaluated for mGlu5 activity by utilizing a cell-based functional assay that measures receptor-induced mobilization of intracellular calcium. The assay is described in detail by Rodriguez et al.37 In brief, the test compound was applied to cells expressing rat mGlu5 in a 10-point concentration−response curve (CRC) format. After an incubation period, cells were stimulated with a submaximal concentration of the mGlu5 agonist glutamate, and the calcium response was measured, followed by stimulation with a near maximal concentration of glutamate and measurement of the calcium response. In this manner both potentiator and antagonist activity can be assessed simultaneously. Six compounds demonstrated an increased mGlu5 response to glutamate while one compound inhibited the response to glutamate. The seven compounds that displayed activity were retested in 10-point CRC format in triplicate. Of the seven compounds, three showed significant PAM activity upon repeat (Figure 5 and Table 1). PAM activity of these three compounds reached a maximal glutamate response of 40−50% with potency values above 10 μM as shown in Figure 5A. Competition binding studies using [3H]methoxyPEPy, a radioligand known to bind to the MPEP binding site, were performed to determine if the novel modulators also interacted at that site or at an alternative binding site such as the CPPHA site. Confirmed PAMs were tested in triplicate using membranes harvested from the same mGlu5 cell line used for the aforementioned calcium assay. A concentration range up to 300 μM was tested (Figure 5B). Compounds VU0603830 and VU0603841 showed no radioligand displacement up to the highest concentration tested, indicating that they do not bind to the MPEP site. Compound VU0603805 exhibited weak binding at concentrations of 100 and 300 μM. PAM activity was observed at 10-fold lower concentrations, 10 and 30 μM, indicating a significant disconnect between potentiator activity and binding, suggesting these compounds interact via an alternate or overlapping binding site. While promising as non-MPEP site PAMs, the potency values of these compounds are not adequate for additional characterization. Compound VU0603805 was chosen to be investigated further through SAR optimization. A synthesized compound series around the VU0603805 compound revealed

Figure 4. This distribution plot shows the similarity of the ordered compounds compared to known PAMs.

Figure 5. (A) Three compounds exhibit concentration-dependent PAM activity in calcium assay. (B) mGlu5 PAMs do not compete for binding to the MPEP site in [3H]methoxyPEPy binding assay at concentrations PAM activity is present in calcium assay. E

DOI: 10.1021/acschemneuro.8b00227 ACS Chem. Neurosci. XXXX, XXX, XXX−XXX

Research Article

ACS Chemical Neuroscience Table 1. List of Relevant and Experimentally Validated Compounds in This Studya

a Potency, maximum glutamate response, and affinity for the compounds of interest. Values are an average of n ≥ 3 independent experiments. N.B. denotes no binding up to highest concentration tested (300 μM).

Figure 6. (A) VU6000591 displays significant PAM activity in calcium assay. (B) VU6000591 exhibits partial inhibition in [3H]methoxyPEPy binding assay. (C) VU6003586 shows a significant increase in PAM potency compared to VU6000591 in calcium assay. (D) VU6003586 exhibits partial inhibition in [3H]methoxyPEPy binding assay.

[3H]methoxyPEPy at concentrations greater than 10 μM (see Figure 6D and Table 1). This demonstrates a difference between the functional potency (EC50 = 174 nM) and the binding affinity for the MPEP site (Ki > 10 μM) of greater than 50-fold, which is consistent with a noncompetitive mechanism of action at an alternative binding site.

a new promising candidate compound, VU6000591, which demonstrated enhanced potency and efficacy (2.4 μM, 86% max). When evaluated in radioligand binding assays, VU6000591 only partially inhibited [3H]methoxyPEPy at concentrations greater than 10 μM. This weak partial inhibition suggests a noncompetitive interaction with the MPEP site (see Figure 6A,B). VU6000591 was optimized further to provide compound VU6003586 with significantly enhanced potency (174 nM, 85% max) (see Figure 6C and Table 1). This potency range is better suited for a more detailed investigation of the binding site. Subsequent radioligand binding experiments demonstrated that in a manner similar to VU6000591, VU6003586 only partially inhibited

3. DISCUSSION Functional selectivity for receptors of the mGlu GPCR family is driven in part by ligand binding behavior at the receptors TM domain. With the discovery of a potential alternative allosteric mGlu5 binding site in addition to the already known MPEP binding site, new possibilities to investigate more F

DOI: 10.1021/acschemneuro.8b00227 ACS Chem. Neurosci. XXXX, XXX, XXX−XXX

Research Article

ACS Chemical Neuroscience

results. Though VU6000591 showed increased potency compared to previous compounds, a compound would need to have a potency roughly in the range of 300−500 nM to allow for a full exploration of its pharmacological effects. Further SAR optimization employing VU6000591 as a starting structure led to candidate compound VU6003586, showing an improved potency in comparison. Subsequent radioligand binding experiments with VU6003586 showed that the compound exhibits partial inhibition of [3H]methoxyPEPy at concentrations well above those that demonstrate functional activity (see Figure 6D). This evidence suggests that the potentiator activity of VU6003586 is most likely the result of interaction at a nonMPEP binding site, but other possibilities, such as high cooperativity, cannot be ruled out. While additional studies would be required to assess which specific binding site these compounds interact with, these data are significant in that they provide additional support that mGlu5 PAMs can act through multiple binding sites and present novel chemotypes from which to begin chemical optimization. Given the limited data set used to build the site-specific QSAR model, it is impressive that novel mGlu5 PAMs were discovered, distinct in structure from known PAMs and focused on a pocket distinct from the MPEP binding site. The identified structures, specifically VU6003586, show significant divergence from the reference CPPHA-like structures indicative of scaffold hopping through the involved QSAR modeling. This approach demonstrates that the identification of target-specific binders is possible and is a valuable method in identifying novel selective lead compounds with reduced adverse effects.

selective allosteric modulators for the mGlu5 receptor are revealed. The goal of this study is to elucidate the chemical space of ligands interacting with this distinct binding site. Thus, a series of CPPHA-like compounds was developed to investigate determinants of the alternative binding site further. The set of compounds contained 275 molecules comprised of 130 active and 145 inactive compounds that show structural similarity to CPPHA. In an effort to investigate the structural distinctiveness of these compounds, a similarity analysis identified 17 clusters. Each cluster center is shown with its main scaffold and a percentage of involved active and inactive compounds (see Figure 1). A clear separation between clusters of solely active or inactive compounds was only given in clusters 3, 11, and 14 and 4, 7, 10, 15, and 16, respectively. This clustering scheme suggests a partially intertwined structural landscape for active and inactive molecules. Therefore, a pure structural separation is not possible. On this basis, a QSAR model was trained to identify selective allosteric mGlu5 binders for the distinct non-MPEP binding site (see Figure 3). The extreme imbalanced character of the training data set leads to a restricted chemical space coverage of the active compounds (see Figure 2); at the same time, it allows for the model to be trained to predict sitespecific binders. In addition, two general prediction models for mGlu5 PAMs27 and NAMs38 from previous work were retrained, observing improved PPV rates of 51% (PAM) and 19% (NAM) (see Supplementary Section S3). This improvement is due to regularization features applied to the ANN algorithm such as neuronal dropout,39 a technique that allows training the QSAR model with increased predictive generalization. These three QSAR models were combined to virtually screen the eMolecules compound library containing over 4 million molecules. A set of 4719 molecules was predicted to be active for specific interaction with the non-MPEP site, predicted inactive by the general mGlu5 PAM QSAR model, and predicted inactive by the mGlu5 NAM model. Medicinal chemistry filters, REOS and PAINS, reduced this set further to only 103 compounds. From this pool of compounds, 63 molecules were chosen by predicted activity, structural diversity, and acquisition cost. A total of seven compounds were identified as exhibiting activity at mGlu5 in a functional assay measuring calcium mobilization. Six of these compounds were categorized as PAMs and one as an antagonist. The three most potent compounds reached a glutamate response of 40−50% of the maximal signal with potencies above 10 μM. This relatively weak signal strength is an expected result given the primary goal of this study was to find new chemotypes through scaffold-hopping that exhibit at minimum a weak response but interact selectively with the mGlu5 site distinct from the MPEP binding site. Competition binding assays measuring the displacement of [3H]methoxyPEPy provided evidence that five of the identified seven compounds indeed interact with a different binding site distinct from the MPEP site and suggested that two compounds, VU0603805 and VU0603849, interact via an alternate or overlapping site. However, these compounds exhibit only weak potentiator activity and thus were considered as starting structures for further SAR optimization. As a result, candidate compound VU6000591 arose from the VU0603805 series. Compound optimization attempts for both other series, VU0603830 and VU0603841, yielded negative

4. CONCLUSIONS In summary, the mGlu5 receptor offers great potential as a possible candidate for drug targets involved in treatment of a range of CNS disorders. An alternative binding site distinct from the already known MPEP binding site was previously identified through mutagenesis and binding experiments involving residues F585 (TM1) and A809 (TM7) in the 7TM helices region. Herein, a QSAR model was employed to query the commercially available compound library “eMolecules” for specific binders to the mGlu5 binding site distinct from MPEP binding. A resulting set of 4719 candidate molecules was identified. After the application of medicinal chemistry filters, a subset of 63 molecules was chosen for compound acquisition. Subsequent calcium mobilization experiments revealed a concentration-dependent modulation of the biological response to mGlu5 in vitro for seven out of the ordered 63 compounds. Further experimental validation through displacement of the radioligand [3H]methoxyPEPy confirmed five out of seven compounds failed to bind the MPEP site, indicating binding to an alternative mGlu5 allosteric pocket. This corresponds to an approximated 8% (5/63 compounds) hit rate compared to the initial 0.09%. Further medicinal chemistry optimizations revealed compound VU6003586 with an experimentally validated potency of 174 nM. Radioligand binding experiments with [3H]methoxyPEPy showed only partial inhibition at very high concentrations, which is most likely indicative of VU6003586 binding at a non-MPEP site. These are very exciting results demonstrating that QSAR modeling can identify novel chemotypes through scaffold hopping given only limited data sets. The confirmed G

DOI: 10.1021/acschemneuro.8b00227 ACS Chem. Neurosci. XXXX, XXX, XXX−XXX

Research Article

ACS Chemical Neuroscience

coated 384-well plates in 20 μL of assay medium (DMEM containing 10% dialyzed FBS, 20 mM HEPES, and 1 mM sodium pyruvate) at a density of 20000 cells/well. The cells were grown overnight at 37 °C in the presence of 5% CO2. The next day, medium was removed, and the cells incubated with 20 μL of 2 μM Fluo-4 AM prepared as a 2.3 mM stock in DMSO, mixed in a 1:1 ratio with 10% (w/v) pluronic acid F-127, and diluted in assay buffer (Hank’s balanced salt solution, 20 mM HEPES, and 2.5 mM probenecid) for 45 min at 37 °C. Dye was removed, 20 μL of assay buffer was added, and the plate was incubated for 10 min at room temperature. Compounds were serially diluted 1:3 in DMSO into 10 point concentration−response curves and transferred to daughter plates using the Echo acoustic plate reformatter (Labcyte, Sunnyvale, CA) followed by further dilution into assay buffer to a 2× stock using a Thermo Fisher Combi (Thermo Fisher, Waltham, MA). Ca2+ flux was measured using the Functional Drug Screening System (FDSS7000, Hamamatsu, Japan). After establishment of a fluorescence baseline for about 3 s, the test compounds were added to the cells, and the response in cells was measured; 2.3 min later, an EC20 concentration of the mGlu5 receptor agonist glutamate was added to the cells, and the response of the cells was measured for 1.9 min. An EC80 concentration of agonist was added, and readings were taken for an additional 1.7 min. Data were collected at 1 Hz. Concentration−response curves were generated using a four point logistical equation with XLfit curve fitting software for Excel (IDBS, Guildford, U.K.) or GraphPad Prism (GraphPad Software, Inc., La Jolla, CA). 5.4. Radioligand Binding Assay Determines Interaction of Ligands with MPEP Site. The allosteric antagonist MPEP analog 3 [ H]methoxyPEPy was used to evaluate the ability of test compounds to interact with the MPEP site on mGlu5. Membranes were prepared from rat mGlu5 HEK 293 cells as previously described.21 Compounds were serially diluted in DMSO then added to assay buffer (50 mM Tris/0.9% NaCl, pH 7.4) to reach a 5× stock, and 50 μL of test compound was added to each well of a 96 deep-well assay plate. Aliquots (150 μL) of membranes diluted in assay buffer (20 μg/well) were added to each well. [3H]MethoxyPEPy (50 μL, 2 nM final concentration) was added, and the reaction was incubated at room temperature for 1 h with shaking. After the incubation period, the membrane-bound ligand was separated from free ligand by filtration through glass-fiber 96 well filter plates (Unifilter-96, GF/B, PerkinElmer Life and Analytical Sciences, Boston, MA). The contents of each well were transferred simultaneously to the filter plate and washed 3−4 times with assay buffer using a cell harvester (Brandel Cell Harvester, Brandel Inc., Gaithersburg, MD). Scintillation fluid (40 μL) was added to each well, and the membrane-bound radioactivity was determined by scintillation counting (TopCount, PerkinElmer Life and Analytical Sciences). Nonspecific binding was estimated using 5 μM MPEP. Concentration−response curves were generated using a four parameter logistical equation in GraphPad Prism (GraphPad Software, Inc., La Jolla, CA). 5.5. Medicinal Chemistry Experimental Details. All reactions were carried out employing standard chemical techniques under inert atmosphere. Solvents used for extraction, washing, and chromatography were HPLC grade. All reagents were purchased from commercial sources and were used without further purification. Analytical HPLC was performed on an Agilent 1200 LCMS with UV detection at 215 and 254 nm along with ELSD detection and electrospray ionization, with all final compounds showing >95% purity and a parent mass ion consistent with the desired structure. All NMR spectra were recorded on a 400 MHz Brüker AV-400 instrument. 1H chemical shifts are reported as δ values in ppm relative to the residual solvent peak (CDCl3 = 7.26). Data are reported as follows: chemical shift, multiplicity (br = broad, s = singlet, d = doublet, t = triplet, q = quartet, m = multiplet), coupling constant (Hz), and integration. 13C chemical shifts are reported as δ values in ppm relative to the residual solvent peak (CDCl3 = 77.16). Low resolution mass spectra were obtained on an Agilent 1200 LCMS with electrospray ionization, with a gradient of 5−95% 0.1% TFA water in MeCN over a run time of 90 s. High resolution mass spectra were recorded on a Waters QToFAPI-US plus Acquity system with electrospray ionization. Automated

compounds target a pocket distinct from the well-characterized known MPEP binding site, implying a separate binding site. The binding affinity can be refined through SAR optimization. As shown in this research virtual screening enables the prioritization of site-specific novel modulators, and thus can contribute to the reduction of adverse effects of drug candidates. The here identified molecules have great potential to serve as novel lead or probe compounds in subsequent drug discovery campaigns.

5. METHODS 5.1. Machine Learning Applied to Develop QSAR Models. A large fraction of proteins targeted by therapeutics have no suitable structural model for structure-based virtual screening. Quantitative structure−activity relationship (QSAR) modeling is an area of computational research that builds in silico models to predict the biological activity for small organic molecules with respect to a protein target.40 The models seek to quantitatively correlate complex nonlinear relationships between chemical and physical properties of a molecule with its biological response for a specific biological target. The application of QSAR models have impact on a variety of fields like drug discovery41 and toxicity prediction.42 Knowledge acquired from compounds already biologically tested has impact on the design of novel molecules. Molecular descriptors play a fundamental role in encoding chemical information of a molecule.43 Chemical properties and structural information of a molecule are encoded in a numerical representation or descriptor. A list of all applied molecular descriptors is available in the Supporting Information of ref 44. Each encoding function is independent of translation or rotation of the molecule. They can be readily applied to conformational ensembles and yield a numerical description of constant length independent from size and composition of the compound. Machine learning algorithms, such as artificial neural networks (ANNs), have shown exciting potential in estimating biological target data through approximation of highly nonlinear relationships between molecules and biological targets.45 The algorithms learn to recognize complex patterns and make intelligent decisions based on an established compound library. Imposing such acquired sets of patterns obtained by a learning process, the algorithms are able to recognize not yet tested molecules and categorize them toward a given outcome. ANNs consist of an interconnected group of artificial neurons and process information adaptively using a restricted set of transformations. ANNs have been used for several years in chemistry and biochemistry to describe QSARs.46−48 Their critical advantage compared to linear methods such as multiple linear regression lies in the flexibility of the models. ANN models can adapt to complex interrelations and are capable of detecting even small signals at high noise levels. Here, ANNs were trained by means of 5-fold cross-validation and evaluated by prediction performance metrics49 on an independent data set, such as ROC Curve, Enrichment, Positive predicted value (PPV), and accuracy. 5.2. Clustering of Identified Hits from Virtual Screening. To assess the similarity of predicted compounds, a hierarchical clustering approach based on average linkage as cluster distance measure was chosen. The similarity calculation among all compound pairs was based on common occurring small molecule fragments. The fragment library was established by determining the largest common substructure between pairs of molecule fragments keeping ring systems intact. Each fragment had a minimum of four atoms. Pairwise compound similarities were calculated based on molecule fragment overlap using the Tanimoto coefficient50 between molecular fragment sets. All atoms were considered equivalent by element type. Bonds were compared by order (single, double, triple), ring membership, and aromaticity. 5.3. Calcium Mobilization Assay Assesses Ligands for mGlu5 Modulation Response. HEK 293 cells expressing rat mGlu5 were plated in black-walled, clear-bottomed, poly(D-lysine)H

DOI: 10.1021/acschemneuro.8b00227 ACS Chem. Neurosci. XXXX, XXX, XXX−XXX

Research Article

ACS Chemical Neuroscience purification of library compounds was performed on a Gilson 215 preparative LC system. More experimental details can be found in Supplementary Section S2 and synthesis information for compound VU6003586 in Supplementary Section S4. 5.6. Implementation. Developed in the Meiler laboratory, the BioChemistryLibrary (BCL) is an object-oriented C++ programming library hosting an array of applications to simulate, interact, and modify biological molecules, such as small organic molecules and proteins. Specifically, the BCL::ChemInfo software suite was applied in this study providing implementations of machine learning algorithms, molecular descriptors, and a cross-validation scheme for in silico virtual screening. The BCL is free of charge for the academic community and can be downloaded from www.meilerlab.org.

■

(2) Schoepp, D. D., Jane, D. E., and Monn, J. A. (1999) Pharmacological agents acting at subtypes of metabotropic glutamate receptors. Neuropharmacology 38 (10), 1431−1476. (3) Abe, T., Sugihara, H., Nawa, H., Shigemoto, R., Mizuno, N., and Nakanishi, S. (1992) Molecular characterization of a novel metabotropic glutamate receptor mGluR5 coupled to inositol phosphate/Ca2+ signal transduction. J. Biol. Chem. 267 (19), 13361−13368. (4) Nickols, H. H., and Conn, P. J. (2014) Development of allosteric modulators of GPCRs for treatment of CNS disorders. Neurobiol. Dis. 61, 55−71. (5) Marino, M. J., and Conn, P. (2002) Modulation of the Basal Ganglia by Metabotropic Glutamate Receptors: Potential for Novel Therapeutics. Curr. Drug Targets: CNS Neurol. Disord. 1 (3), 239− 250. (6) Chavez-Noriega, L. E., Schaffhauser, H., and Campbell, U. C. (2002) Metabotropic Glutamate Receptors: Potential Drug Targets for the Treatment of Schizophrenia. Curr. Drug Targets: CNS Neurol. Disord. 1 (3), 261−281. (7) Wiśniewski, K., and Car, H. (2002) (S)-3,5-DHPG: A Review. CNS Drug Rev. 8 (1), 101−116. (8) Gregory, K. J., Dong, E. N., Meiler, J., and Conn, P. J. (2011) Allosteric modulation of metabotropic glutamate receptors: Structural insights and therapeutic potential. Neuropharmacology 60 (1), 66−81. (9) Vinson, P. N., and Conn, P. J. (2012) Metabotropic glutamate receptors as therapeutic targets for schizophrenia. Neuropharmacology 62 (3), 1461−1472. (10) O’Brien, J. A., et al. (2004) A Novel Selective Allosteric Modulator Potentiates the Activity of Native Metabotropic Glutamate Receptor Subtype 5 in Rat Forebrain. J. Pharmacol. Exp. Ther. 309 (2), 568−577. (11) Kinney, G. G., et al. (2004) A novel selective positive allosteric modulator of metabotropic glutamate receptor subtype 5 has in vivo activity and antipsychotic-like effects in rat behavioral models. J. Pharmacol. Exp. Ther. 313 (1), 199. (12) Chen, Y., Goudet, C., Pin, J.-P., and Conn, P. J. (2007) N-{4Chloro-2-[(1, 3-dioxo-1, 3-dihydro-2H-isoindol-2-yl) methyl] phenyl}-2-hydroxybenzamide (CPPHA) acts through a novel site as a positive allosteric modulator of group 1 metabotropic glutamate receptors. Mol. Pharmacol. 73 (3), 909−918. (13) Chen, Y., et al. (2007) Interaction of novel positive allosteric modulators of metabotropic glutamate receptor 5 with the negative allosteric antagonist site is required for potentiation of receptor responses. Mol. Pharmacol. 71 (5), 1389−1398. (14) Hammond, A. S., et al. (2010) Discovery of a Novel Chemical Class of mGlu5 Allosteric Ligands with Distinct Modes of Pharmacology. ACS Chem. Neurosci. 1 (10), 702−716. (15) Rodriguez, A. L., et al. (2010) Identification of a glycine sulfonamide based non-MPEP site positive allosteric potentiator (PAM) of mGlu5, in Probe Reports from the NIH Molecular Libraries Program, National Center for Biotechnology Information (US), Bethesda, MD. (16) Noetzel, M. J., et al. (2012) Functional Impact of Allosteric Agonist Activity of Selective Positive Allosteric Modulators of Metabotropic Glutamate Receptor Subtype 5 in Regulating Central Nervous System Function. Mol. Pharmacol. 81 (2), 120−133. (17) Packiarajan, M., et al. (2012) Azetidinyl oxadiazoles as potent mGluR5 positive allosteric modulators. Bioorg. Med. Chem. Lett. 22 (20), 6469−6474. (18) Packiarajan, M., et al. (2012) N-Aryl pyrrolidinonyl oxadiazoles as potent mGluR5 positive allosteric modulators. Bioorg. Med. Chem. Lett. 22 (17), 5658−5662. (19) Lamb, J. P., et al. (2011) Discovery of molecular switches within the ADX-47273 mGlu5 PAM scaffold that modulate modes of pharmacology to afford potent mGlu5 NAMs, PAMs and partial antagonists. Bioorg. Med. Chem. Lett. 21 (9), 2711−2714. (20) Niswender, C. M., Jones, C. K., and Conn, P. J. (2005) New Therapeutic Frontiers for Metabotropic Glutamate Receptors. Curr. Top. Med. Chem. 5 (9), 847−857.

ASSOCIATED CONTENT

S Supporting Information *

The Supporting Information is available free of charge on the ACS Publications website at DOI: 10.1021/acschemneuro.8b00227.

■

List of abbreviations used in the manuscript and their corresponding explanations, NMR peak information on experimental determination of key compounds for this study, description of the data set composition used to create the QSAR model and its applied quality metrics, step by step synthesis information for the key compound of this manuscript (VU6003586), chemical structures and IUPAC names of 63 compounds identified through virtual screening with the developed QSAR model, experimentally evaluated, and a subset considered as starting structures for subsequent SAR optimization (PDF)

AUTHOR INFORMATION

Corresponding Author

*E-mail: [email protected]. Tel: +1-615-936-5662. Fax: +1-615-936-2211. ORCID

Mariusz Butkiewicz: 0000-0003-2582-9336 Craig W. Lindsley: 0000-0003-0168-1445 Author Contributions

M.B. performed QSAR modeling and wrote the manuscript. A.L.R. conducted experimental work and wrote the experimental sections of the manuscript. S.E.R., J.W., and V.B.L. helped with conducting the experimental work. S.R.S. provided compound information for QSAR modeling. C.W.L., P.J.C., and J.M. designed the project and guided the writing process of the manuscript. All coauthors approved the manuscript. Notes

The authors declare no competing financial interest.

■

ACKNOWLEDGMENTS We thank the Advanced Computing Center for Research and Education (ACCRE) at Vanderbilt University for use of their cluster. Work in the Meiler laboratory is supported through NIH (R01 MH090192, R01 DK097376) and NSF (CHE 1305874).

■

REFERENCES

(1) Conn, P. J., and Pin, J. P. (1997) Pharmacology and functions of metabotropic glutamate receptors. Annu. Rev. Pharmacol. Toxicol. 37, 205−37. I

DOI: 10.1021/acschemneuro.8b00227 ACS Chem. Neurosci. XXXX, XXX, XXX−XXX

Research Article

ACS Chemical Neuroscience (21) Rodriguez, A. L., Nong, Y., Sekaran, N. K., Alagille, D., Tamagnan, G. D., and Conn, P. J. (2005) A close structural analog of 2-methyl-6-(phenylethynyl)-pyridine acts as a neutral allosteric site ligand on metabotropic glutamate receptor subtype 5 and blocks the effects of multiple allosteric modulators. Mol. Pharmacol. 68 (6), 1793−1802. (22) Lindsley, C. W., et al. (2004) Discovery of Positive Allosteric Modulators for the Metabotropic Glutamate Receptor Subtype 5 from a Series of N-(1,3-Diphenyl-1H- pyrazol-5-yl)benzamides That Potentiate Receptor Function in Vivo. J. Med. Chem. 47 (24), 5825−5828. (23) Liu, F., et al. (2008) ADX47273 [S-(4-Fluoro-phenyl)-{3-[3(4-fluoro-phenyl)-[1,2,4]-oxadiazol-5-yl]-piperidin-1-yl}-methanone]: A Novel Metabotropic Glutamate Receptor 5-Selective Positive Allosteric Modulator with Preclinical Antipsychotic-Like and Procognitive Activities. J. Pharmacol. Exp. Ther. 327 (3), 827−839. (24) Epping-Jordan, M. P., et al. (2005) In vivo characterization of mGluR5 positive allosteric modulators as novel treatments for schizophrenia and cognitive dysfunction. Neuropharmacology 49, 243−243. (25) Darrah, J. M., Stefani, M. R., and Moghaddam, B. (2008) Interaction of N-methyl-D-aspartate and group 5 metabotropic glutamate receptors on behavioral flexibility using a novel operant set-shift paradigm. Behav. Pharmacol. 19 (3), 225−234. (26) Noetzel, M. J., et al. (2013) A Novel Metabotropic Glutamate Receptor 5 Positive Allosteric Modulator Acts at a Unique Site and Confers Stimulus Bias to mGlu5 Signaling. Mol. Pharmacol. 83 (4), 835−847. (27) Mueller, R., et al. (2010) Identification of Metabotropic Glutamate Receptor Subtype 5 Potentiators Using Virtual HighThroughput Screening. ACS Chem. Neurosci. 1 (4), 288−305. (28) Mueller, R., et al. (2012) Iterative experimental and virtual high-throughput screening identifies metabotropic glutamate receptor subtype 4 positive allosteric modulators. J. Mol. Model. 18 (9), 4437− 4446. (29) Noeske, T., Sasse, B. C., Stark, H., Parsons, C. G., Weil, T., and Schneider, G. (2006) Predicting Compound Selectivity by SelfOrganizing Maps: Cross-Activities of Metabotropic Glutamate Receptor Antagonists. ChemMedChem 1 (10), 1066−1068. (30) Sheffler, D. J., and Conn, P. J. (2008) Allosteric Potentiators of Metabotropic Glutamate Receptor Subtype 1a Differentially Modulate Independent Signaling Pathways in Baby Hamster Kidney Cells. Neuropharmacology 55 (4), 419−427. (31) Niswender, C. M., and Conn, P. J. (2010) Metabotropic Glutamate Receptors: Physiology, Pharmacology, and Disease. Annu. Rev. Pharmacol. Toxicol. 50, 295−322. (32) Gregory, K. J., et al. (2012) Investigating Metabotropic Glutamate Receptor 5 Allosteric Modulator Cooperativity, Affinity, and Agonism: Enriching Structure-Function Studies and StructureActivity Relationships. Mol. Pharmacol. 82 (5), 860−875. (33) Rodriguez, A. L., et al. (2012) Discovery and SAR of a novel series of non-MPEP site mGlu 5 PAMs based on an aryl glycine sulfonamide scaffold. Bioorg. Med. Chem. Lett. 22 (24), 7388−7392. (34) eMolecules (May 2013). [Online]. Available at https://www. emolecules.com/. (35) Walters, W. P., Stahl, M. T., and Murcko, M. A. (1998) Virtual screeningan overview. Drug Discovery Today 3 (4), 160−178. (36) Baell, J. B., and Holloway, G. A. (2010) New Substructure Filters for Removal of Pan Assay Interference Compounds (PAINS) from Screening Libraries and for Their Exclusion in Bioassays. J. Med. Chem. 53 (7), 2719−2740. (37) Rodriguez, A. L., et al. (2010) Discovery of novel allosteric modulators of metabotropic glutamate receptor subtype 5 reveals chemical and functional diversity and in vivo activity in rat behavioral models of anxiolytic and antipsychotic activity. Mol. Pharmacol. 78 (6), 1105−1123. (38) Mueller, R., et al. (2012) Discovery of 2-(2-Benzoxazoyl amino)-4-Aryl-5-Cyanopyrimidine as Negative Allosteric Modulators (NAMs) of Metabotropic Glutamate Receptor 5 (mGlu5): From an

Artificial Neural Network Virtual Screen to an In Vivo Tool Compound. ChemMedChem 7 (3), 406−414. (39) Krizhevsky, A., Sutskever, I., and Hinton, G. E. (2012) ImageNet Classification with Deep Convolutional Neural Networks, in Advances in Neural Information Processing Systems 25 (Pereira, F., Burges, C. J. C., Bottou, L., and Weinberger, K. Q., Eds.), pp 1097− 1105, Curran Associates, Inc. (40) Wang, L., Ma, C., Wipf, P., Liu, H., Su, W., and Xie, X.-Q. (2013) TargetHunter: An In Silico Target Identification Tool for Predicting Therapeutic Potential of Small Organic Molecules Based on Chemogenomic Database. AAPS J. 15 (2), 395−406. (41) Kapetanovic, I. M. (2008) Computer-aided drug discovery and development (CADDD): in silico-chemico-biological approach. Chem.-Biol. Interact. 171 (2), 165−176. (42) Cronin, M. T. (2004) Predicting chemical toxicity and fate. CRC press. (43) Todeschini, R., and Consonni, V. (2009) Molecular Descriptors for Chemoinformatics, Vol. 41, John Wiley & Sons. (44) Butkiewicz, M., et al. (2013) Benchmarking ligand-based virtual high-throughput screening with the PubChem database. Molecules 18 (1), 735−756. (45) Almeida, J. S. (2002) Predictive non-linear modeling of complex data by artificial neural networks. Curr. Opin. Biotechnol. 13 (1), 72−76. (46) So, S.-S., and Karplus, M. (1996) Evolutionary optimization in quantitative structure-activity relationship: an application of genetic neural networks. J. Med. Chem. 39 (7), 1521−1530. (47) Simmons, K., et al. (2008) Comparative Study of MachineLearning and Chemometric Tools for Analysis of In-Vivo HighThroughput Screening Data. J. Chem. Inf. Model. 48 (8), 1663−1668. (48) Butkiewicz, M., Mueller, R., Selic, D., Dawson, E., and Meiler, J. (2009) Application of Machine Learning Approaches on Quantitative Structure Activity Relationships, Presented at the IEEE Symposium on Computational Intelligence in Bioinformatics and Computational Biology, Nashville, pp 255−262. (49) Florkowski, C. M. (2008) Sensitivity, specificity, receiveroperating characteristic (ROC) curves and likelihood ratios: communicating the performance of diagnostic tests. Clin. Biochem. Rev. 29 (Suppl 1), S83. (50) Martin, E. J., Blaney, J. M., Siani, M. A., Spellmeyer, D. C., Wong, A. K., and Moos, W. H. (1995) Measuring Diversity: Experimental Design of Combinatorial Libraries for Drug Discovery. J. Med. Chem. 38 (9), 1431−1436.

J

DOI: 10.1021/acschemneuro.8b00227 ACS Chem. Neurosci. XXXX, XXX, XXX−XXX