Predictive Models to Identify Small Molecule Activators and Inhibitors of Opioid Receptors

Abstract

Opioid receptors (OPRs) are the main targets for the treatment of pain and related disorders. The opiate compounds that activate these receptors are effective analgesics but their use leads to adverse effects and they often are highly addictive drugs of abuse. There is an urgent need for alternative chemicals that are analgesic and reducing/avoiding the unwanted effects in order to relieve the public health crisis of opioid addiction. Here, we aim to develop computational models to predict the OPR activity of small molecule compounds based on chemical structures and apply these models to identify novel OPR active compounds. We used four different machine learning algorithms to build models based on quantitative high throughput screening (qHTS) datasets of three OPRs in both agonist and antagonist modes. The best performing models were applied to virtually screen a large collection of compounds. The model predicted active compounds which were experimentally validated using the same qHTS assays that generated the training data. Random forest was the best classifier with the highest performance metrics and the mu OPR (OPRM)-agonist model achieved the best performance with AUC-ROC (0.88) and MCC (0.7) values. The model predicted actives resulted in hit rates ranging from 2.3% (delta OPR-agonist) to 15.8% (OPRM-agonist) after experimental confirmation. Comparing to the original assay hit rate, all models enriched hit rate by ≥ 2-fold. Our approach produced robust OPR prediction models that can be applied to prioritize compounds identified from large libraries for further experimental validation. The models identified several novel potent compounds as activators/inhibitors of OPRs that were confirmed experimentally. The potent hits were further investigated using molecular docking to find the interactions of the novel ligands in the active site of the corresponding OPR.

Graphical Abstract

Introduction

Opioid receptors (OPRs) belong to the superfamily of G protein-coupled receptors (GPCR), consisting of 3 main types: mu (OPRM), kappa (OPRK), and delta (OPRD). These receptors are important for expressing pain transmission and modulation pathways, and are largely distributed in the central nervous system, while to a less extent in the periphery including gastrointestinal tract, heart and immune system etc. ¹. OPRs are activated both endogenously and exogenously. The endogenous ligands include endorphins, dynorphins, and enkephalins for OPRM, OPRK, and OPRD respectively ². Whereas the exogenous opioid drugs include codeine, fentanyl, hydrocodone, methadone, morphine, oxycodone, buprenorphine, naloxone, naltrexone etc. with varying effect on different opioid receptor types ³. Most of these drugs that are administered as analgesics have side effects leading to drug addiction and abuse ^{4, 5}. In recent years there is a significant increase in death rate due to the drug overdose. According to the Centers for Disease Control and Prevention (CDC), more than 67,000 drug overdose related deaths occurred in the United States in 2018 and opioid-involved overdose accounted for ~70% of the total drug overdose deaths ⁶. Mainly the compounds targeting OPRM are known to produce several side effects that could be fatal. Search for analgesics with fewer side effects and/or for compounds targeting OPRK and OPRD has emerged as an alternative to produce safer drugs ⁷.

The vast amount of data generated from high-throughput screens are commonly used as training data for developing quantitative structure-activity relationship (QSAR) models to predict the activity of novel chemicals on biological targets using machine learning techniques ^{8, 9}. In order to build predictive models for the identification of novel activators/inhibitors of OPRs, we screened a collection of ~3000 approved drugs against three OPRs, OPRD, OPRK and OPRM, in a quantitative high throughput screening (qHTS) ¹⁰ format in both agonist and antagonist modes. qHTS generates a concentration-response for every compound in the primary screen producing high quality data that are ideal for training machine learning models. Several research works were published in the past using HTS data to build predictive models for various endpoints using machine learning algorithms and produced robust models ^{11, 12}. In this study, we developed predictive models to identify activators/inhibitors of OPRs using the qHTS assay data as training datasets. Six QSAR models were developed, which were trained with the experimental qHTS datasets of agonist/antagonist modes of OPRM, OPRK, and OPRD. The models with good performance were applied to virtually screen our large in-house collections of 49,018 compounds to identify potential new OPR actives. The model predicted active compounds were confirmed experimentally. The potent actives were further evaluated by docking them to the crystal structures of the respective OPRs to study their interactions. Several independent research groups have identified novel hits through docking, such as discovery of active molecules against mu ¹³, and kappa ¹⁴ OPRs with new scaffolds that are unrelated to the known opioids. Through our current study, we identified several potent compounds with novel structures that activate or inhibit OPRs. Our models can be applied to make predictions on large chemical libraries, which lack experimental data, to prioritize the rapidly increasing drug-like new compounds for further testing.

Materials and Methods

qHTS cAMP assay in OPR cells

The CHO-K1 cells that express full-length human recombinant μ-, κ-, and δ-OPRs (HMOR, HKOR, and HDOR respectively) were purchased from Multispan, Inc. (Hayward, CA). The cells were cultured in DMEM/F12, 10% FBS, 100U/ml penicillin-100μg/ml streptomycin, and 10μg/mL puromycin (HMOR and HKOR) or 10μg/mL puromycin + 250μg/mL hygromycin (HDOR). The cell culture was maintained at 37°C, 5% CO₂, and 99% humidity. The cells were plated at 1,000/well in 3μL of the culture medium without the antibiotic marker in a 1,536-well white solid-bottom plates (Greiner Bio-One North America, NC) using Multidrop combi dispenser (Thermo Fisher Scientific Inc., Waltham, MA). The assay plates were incubated at 37°C for 18hr for cell adhesion to the plates, then 23nL of the positive control and test compounds were transferred to each well of the assay plates using Pintool station (Wako, San Diego, CA). The agonist positive controls used were DAMGO (Abcam, Cambridge, MA) for HMOR, and Dynorphin B peptide (Abcam) for HKOR and HDOR. The antagonist positive controls used were naloxone for HMOR and HKOR (Sigma-Aldrich, St. Louis, MO), and naltrindole (Sigma-Aldrich) for HDOR. Compound transfer was followed by the addition of 1μL of 0.5mM IBMX (3-Isobutyl-1-methylxanthine, Sigma-Aldrich) to each well of the assay plates using a Flying Reagent Dispenser (FRD, Aurora Discovery, Carlsbad, CA). Whereas for antagonist mode, 1μL of a mixture of 0.5mM IBMX and agonist positive control (2nM DAMGO for HMOR, 0.6nM or 2nM dynorphin B for HKOR or HDOR respectively) was added to each well of the assay plates using an FRD. The assay plates were incubated at 37°C for 20min and followed by the addition of 1μL of 1.0μM NKH477 (Sigma-Aldrich) to each well of the assay plates using an FRD. The assay plates were incubated at 37°C for 30min. Then the detection reagents were added at 2.5μL of cAMP-Cryptate (cAMP-Gi kit, Cisbio US, Inc., Bedford, MA), followed by 2.5μL of Anti-cAMP-d2 (cAMP-Gi kit, Cisbio) to each well of the assay plates using an FRD. After 1hr incubation at room temperature, the fluorescence intensity was quantified using Envision plate reader (PerkinElmer, Waltham, MA) at excitation 340nm and emissions at 665 and 620 nm. Data were expressed as ratio of 665nm/620nm.

qHTS data analysis

For primary data analysis, raw plate reads for each titration point were first normalized relative to positive control (agonist mode: 100%, antagonist mode: 0%) and DMSO only wells (agonist mode: 0%, antagonist mode: −100%). Percent activity is then calculated as: % Activity = ((Vtest compound−V_DMSO)/(Vpositive control−V_DMSO)) × 100, where Vtest compound are the values of compound wells, Vpositive control is the median value of the positive control wells, and V_DMSO is the median value of DMSO-only wells, and then corrected by applying a pattern correction algorithm using compound-free control plates (DMSO plates). Concentration-response titration points for each compound were fitted to the Hill equation and concentrations of half-maximal activity (AC₅₀) and maximal response (efficacy) values were calculated ¹⁵. Compounds were designated as class 1–4 according to the type of concentration–response curve observed. Class 4 compounds were considered inactive. Compounds with class 1.1, 1.2, 2.1 curves or 2.2 curves with >40% efficacy in the agonist mode assays or >50% efficacy in the antagonist mode assays, and inactive or >6-fold less potent in the wild type counter screen were considered active. All other classes of compounds were considered inconclusive and excluded from modeling ¹⁶.

Compound library and datasets for modeling

The training set used in our study consists of 2805 compounds from the NCATS Pharmaceutical Collection (NPC) of approved and investigational drugs ^{17, 18}. The prediction set used in our study consists of 49,018 compounds from both Sytravon (a library of retired pharma screening collection containing a diversity of novel small molecules with an emphasis on medicinal chemistry-tractable scaffolds) and NPACT (NCATS Pharmacologically Active Chemical Toolbox; a library of annotated compounds that inform on novel phenotypes, cellular processes, and biological pathways) collections. The qHTS data obtained from in vitro assays of HMOR, HKOR, and HDOR cells screened against NPC were used to train and test models. In vitro qHTS assay data against a particular target, consisting of both active and inactive compounds, were randomly split into two sets, roughly two-thirds (1888) for model training and testing (cross-validation) and one-third (917) for external validation. Three different binary fingerprints: Molecular ACCess System (MACCS) ¹⁹, PubChem substructure fingerprints (ftp://ftp.ncbi.nlm.nih.gov/pubchem/specifications/pubchem_fingerprints.txt), and Extended-Connectivity Fingerprints (ECFP) ²⁰ were used to represent the compound structures, which are in 166, 881, and 1024-bit length, respectively. MACCS and PubChem fingerprints were generated from a workflow-based cheminformatics tool, KNIME-CDK ²¹ and ECFP from Dragon 7 software.

Supervised machine learning algorithms

Data processing was performed using R 3.5.3. QSAR models were developed using four machine learning algorithms: Random Forests (RF), Support Vector Machines (SVM), Neural Networks (NN), and eXtreme Gradient Boosting (XGBoost) ^22–25 to classify the compounds based on their chemical structures for a given target. The packages used in R were randomForest, kernlab, nnet, and xgboost for implementing RF, SVM, NN, and XGBoost methods respectively. The models were trained and tested using a 5-fold cross validation. The cross-validation process was repeated 20 times with different random splits. The key parameters chosen for RF were a default value of 500 for number of trees, and randomly selected variables for each split was set to the square root of the number of predictors. SVM algorithm implemented was a Guassian Radial Basis kernel method for classification with cost of constraints violation set to 100. A feed-forward NN with a single hidden layer of unit size 4 and decay of 0.1 were set as parameters. The boosting parameters for XGBoost were set to 0.05 (control the learning rate), 2 (maximum depth of trees) and the objective specified for the learning task was a logistic regression for binary classification, with 200 boosting iterations. For making predictions on the validation and prediction datasets, the Caret package was used for model fitting on the training set using RF algorithm and the hyper-parameters were selected based on the optimal model with the largest “AUC-ROC” metric ²⁶.

Class rebalancing and feature selection

Random under- and over-sampling methods were employed to the training set. Under-sampling was implemented by random selection of majority class at each iteration to balance with the active class compounds and over-sampling was implemented using Rose package in R ²⁷. Features that have zero variance were eliminated and the significant features that show bivariate relationships were identified via a chi-square test ²⁸. Only those features that had p-value <0.05 were used in our current study ²⁹.

Evaluation of model performances

The 5-fold cross validation performance of the training set was evaluated by computing area under the ROC (receiver operating characteristic) curve (AUC-ROC) values and were computed using “ROCR” package in R. The models generated using the training set were validated on the hold-out test set. The predictions fall into the four categories: true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN). The following measures were used to evaluate the model performances in addition to AUC-ROC ^30–32:

$$\text{Sensitivity}=\frac{TP}{TP+FN}$$

$$\text{Specificity}=\frac{TN}{FP+TN}$$

$$\text{Balanced}\text{Accuracy}(\text{BA})=\frac{Sensitivity+Specificity}{2}$$

$$\text{Matthews}\text{Correlation}\text{Coefficient}(\text{MCC})=\frac{(TPXTN)-(FPXFN)}{\sqrt{(TP+FP)(TP+FN)(TN+FP)(TN+FN)}}$$

$$\text{Positive}\text{Predictive}\text{Value}(\text{PPV})=\frac{TP}{TP+FP}$$

The boxplots and ROC curves were generated using “ggplot2” and “ggpubr” packages in R. Compounds represented by PubChem fingerprints (categorical data) were clustered using k-modes method, an extension to k-means algorithm especially developed for categorical datasets ³³. The clustering was performed in R using klaR package by a simple-matching distance method. The structural similarity between two compounds in our study was measured by the Tanimoto score ^{34, 35}.

Molecular docking

The three dimensional structures of OPRs bound to a ligand molecule were retrieved from Protein Data Bank (PDB) with the following PDB codes: 5C1M (OPRM-BU72 agonist), 4DKL (OPRM-βFNA antagonist), 6B73 (OPRK-MP1104 agonist), 4DJH (OPRK-JDTic antagonist), 6PT3 (OPRD-DPI287 agonist), and 4EJ4 (OPRD-naltrindole antagonist). Molecular docking was performed using Autodock Vina, an open source docking program ³⁶. Using AutoDock Tools, the ligands and the protein molecules were prepared (deletion of heteroatoms and addition of polar hydrogen atoms, and Kollman charges) and were saved in pdbqt format. The active site amino acid residues had been indicated in the published crystal structures of each protein-ligand complexes like OPRM-BU72 ³⁷, OPRM-βFNA ³⁸, OPRK-MP1104 ³⁹, OPRK-JDTic ⁴⁰, OPRD-DPI287 ⁴¹, and OPRD-naltrindole ⁴². Based on these binding sites of the receptors, a grid box was defined ⁴³. From the correctly aligned grid box that covers the entire active site of the receptor’s binding pocket, the coordinates were saved. The center coordinates are different for each target, but the size of the grid box for each target is the same (size_x: 40, size_y: 40, and size_z: 40). The potent active (AC₅₀ ≤ 2.5μM) compounds from each set were docked into the active site of the respective OPRs and the binding affinity values were reported in Table 2. The structures of the protein-ligand docked complexes were visualized using PyMOL tool.

Table 2.

List of novel active compounds (AC₅₀ ≤ 2.5 μM).

ID	Name	AC50 (μM)	Binding Affinity (kcal/mol)	Target	Structure
NCGC00356417-05	LLY-507	0.13	−8.5	OPRM-agonist
NCGC00386484-01	LY 426965	0.18	−9.5	OPRM-agonist
NCGC00378719-01	VUF-2274	0.41	−10.1	OPRM-agonist
NCGC00118549-01		0.46	−9.6	OPRM-agonist
NCGC00114741-01		0.52	−10.1	OPRM-agonist
NCGC00370807-05	FIPI	0.58	−11.4	OPRM-agonist
NCGC00485045-01	N-Methylspiperone	0.65	−9.4	OPRM-agonist
NCGC00247751-02		0.65	−10.4	OPRM-agonist
NCGC00123899-01		0.92	−10.2	OPRM-agonist
NCGC00107633-01		0.92	−10.0	OPRM-agonist
NCGC00114743-01		1.64	−10.0	OPRM-agonist
NCGC00378879-01	NP-118809	1.83	−10.4	OPRM-agonist
NCGC00118605-01		1.83	−8.4	OPRM-agonist
NCGC00141762-01		2.06	−8.2	OPRM-agonist
NCGC00346481-02	Ridaforolimus	0.13	−9.9	OPRK-agonist
NCGC00135974-01		1.16	−9.7	OPRK-agonist
NCGC00117293-01		2.50	−9.0	OPRK-agonist
NCGC00139128-01		2.06	−7.4	OPRD-agonist
NCGC00136158-01		2.31	−7.9	OPRD-agonist
NCGC00485585-01	Adenosine 3’,5’-cyclothiophosphate	0.06 0.07 0.04	−6.7 −7.3 −6.8	OPRM-antagonist OPRK-antagonist OPRD-antagonist
NCGC00114968-02		0.65	−8.4	OPRM-antagonist
NCGC00485288-01	Tocladesine	1.03 0.92 0.46	−6.8 −7.6 −7.4	OPRM-antagonist OPRK-antagonist OPRD-antagonist
NCGC00106344-01		2.06	−9.8	OPRK-antagonist
NCGC00386726-01	IN-1130	2.06	−10.4	OPRK-antagonist

Results

Assay performances and activity distribution

The compounds from the NPC library were screened at 7 concentrations in both agonist and antagonist modes against three cells-based assays (OPRM, OPRK, and OPRD) to identify activators/inhibitors of OPRs. All the assays performed well with coefficient of variances (CV) <7.0%, and signal/background (S/B) ratios ≥2.0 for OPRM-agonist and OPRK-antagonist screening, and Z’ factors >0.40 for OPRM- & OPRD-agonists and OPRK-antagonist screening. The rest of the assays had lower S/B and Z’ factors, but the overall performances were compensated by the good CV values. The compound activity distributions in terms of active, inactive, or inconclusive for each OPR assay are shown in Figure 1 ¹⁶. The assays produced the highest active (hit) rates were OPRM-agonist (15%) and OPRK-antagonist (12%), and the OPRD-agonist (3%) assay yielded the lowest hit rate.

Figure 1.

The activity class distribution of compounds in the qHTS assay data.

Model training and cross-validation

Four machine learning algorithms: RF, SVM, NN, and XGBoost were applied to develop QSAR models for the 6 OPR targets (agonist or antagonist of OPRM, OPRK, and OPRD) using cell-based cAMP assay data in OPR cell lines. These models were trained using 3 different structural fingerprints for compounds representation: ECFP, MACCS, and PubChem. All 6 datasets were highly imbalanced as the hit (active) rates were low (Figure 1). An under-sampling technique was applied to balance the active/inactive classes by random selection from the majority class (inactive compounds) to maintain a ratio of 1:2 (active: inactive). A 5-fold cross-validation with 20 iterations was implemented on the training set and iterating ensures that the majority class was sufficiently sampled to cover all inactive compounds. The model performances from the 20 iterations of 5-fold cross-validation were reported as mean ± standard deviation of a total of 100 AUC-ROC values (Supplementary Table S1), and the summary of distributions are shown as box plots in Figure 2. The RF classifier generated high AUC-ROC values, and the highest was 0.87 ± 0.03 for OPRM-agonist with PubChem fingerprints. The next best AUC-ROC values were 0.81 ± 0.07 and 0.82 ± 0.04 for OPRK-agonist with PubChem fingerprints and OPRK-antagonist with ECFP fingerprints, respectively. The next best classifier was SVM, which yielded an AUC-ROC value of 0.85 ± 0.04 for OPRM-agonist.

Figure 2.

Box plots for the AUC-ROC values showing distributions of 4 machine learning algorithms for different fingerprint types.

To further evaluate the robustness of our models, Y-randomization was performed for all 6 datasets by randomly shuffling the dependent variable (class). The RF method and ECFP were used to build models based on the Y-randomized training sets, as this combination generated the highest AUC-ROC values (Supplementary Table S1). The performance results from 20 iterations of 5-fold cross validations on the randomized datasets are provided in Supplementary Table S2. The models built with the randomized sets showed predictive performances close to a random classifier with AUC-ROC values close to 0.5, demonstrating the robustness of our original models.

External validation and predictions

Dealing with imbalanced classes, an over-sampling strategy was applied by random duplication of the minority class (active compounds) to 30%. Both ECFP and PubChem fingerprints were used as predictors to train the 6 models using two-thirds of the NPC compounds by the RF algorithm, and the models were applied to make predictions on the remaining one-third of the compounds that served as the external validation set. The predictive performances of all 6 models on the external validation set in terms of balanced accuracy, AUC-ROC and MCC were calculated (Table 1) and the ROC curves were plotted (Figure S1). The ROC curves for agonist (Figure S1.A) and antagonist modes (Figure S1.B) for each fingerprint type showed slight variations at different thresholds of true positive and false positive rates. PubChem fingerprints produced slightly better performance metrics when compared to ECFP in terms of AUC-ROC and MCC values (Table 1). The AUC-ROC values were >0.76 for all the models (Table 1), except for the OPRM- and OPRD-antagonist models built with ECFP. We then tried to see if consensus modeling could help improve the model performances. Individual models built with the RF, SVM, NN, and XGBoost methods were used for consensus modeling. The AUC-ROC values from predicting the validation data for the individual OPRM-antagonist models were 0.72, 0.71, 0.67, and 0.68, respectively, and 0.73, 0.74, 0.67, and 0.71, respectively, for the OPRD-antagonist models. To build the consensus models using the weighted average method, higher weights were assigned to more accurate individual models. A weight of 0.4 was assigned to the RF and SVM models, and 0.1 was assigned to the NN and XGBoost models. Consensus modeling improved the predictive performances of OPRM- and OPRD-antagonists slightly to AUC-ROC values of 0.73 and 0.75, respectively, but still not better than the best individual models built with PubChem fingerprints. Similarly, consensus models based on the majority vote were not able to outperform the best individual models. Thus, the RF classification method that produced the best models was used for further analyses.

Table 1.

Performance measures of balanced accuracy (BA), AUC-ROC and MCC from the evaluation on the validation set with 6 models.

	PubChem fingerprints			ECFP fingerprints
	BA	AUC	MCC	BA	AUC	MCC
OPRM-agonist	0.73	0.88	0.7	0.67	0.84	0.59
OPRK-agonist	0.62	0.76	0.33	0.61	0.76	0.31
OPRD-agonist	0.58	0.8	0.17	0.52	0.78	0.07
OPRM-antagonist	0.61	0.76	0.31	0.58	0.72	0.26
OPRK-antagonist	0.62	0.79	0.39	0.61	0.78	0.39
OPRD-antagonist	0.62	0.76	0.32	0.58	0.73	0.27

The training and external validation sets were combined, and only PubChem fingerprints were used, to build the final models, which were applied to make predictions on the large prediction set of 49,018 compounds that have no experimental data.

Compound selection and experimental validation

Compounds with predicted probability of 0.5 or higher were classified as active (Figure S2). The whole prediction set of compounds were clustered using the k-modes algorithm resulting in 2450 clusters, with cluster sizes ranging from one to 112 compounds. The compounds from each dataset were ranked based on the probability score and the cluster size. For the current study, to reduce false positives from our predictions, we implemented a precise prediction probability cutoff for selection of compounds from each dataset (OPRM-agonist > 0.7, OPRK-antagonist > 0.6, and others > 0.5). From each cluster, the compounds with the highest probability scores were selected if the cluster size was < 10, and the top two were selected if the cluster size was > 10. Based on the in-house chemical availability, 2816 compounds, which fit exactly into two screening plates, were selected for experimental validation.

These predicted active compounds were tested in the same qHTS assays that generated the corresponding training data. Experimentally confirmed compounds were counted as true positive (TP) and false positive (FP) otherwise. The confusion matrices describing the performance of the models based on the experimental validation results are given in Supplementary Table S3. The performances were assessed by PPV measures, shown as histograms in Figure 3A. In our current study, the highest PPVs were obtained for OPRM-agonist (0.31), OPRK-agonist (0.30), and – antagonist (0.24), whereas OPRD-agonist and -antagonist obtained the lowest PPV of 0.18 and 0.15 respectively. To assess the applicability domain (AD) of the models, the Tanimoto similarity score was calculated between each predicted active compound and all active compounds in the training set for every model. The Tanimoto score (Tmax) between the predicted actives and the most similar compound to it in the training set was recorded. The predicted actives with Tmax>0.8 (fall within the model AD) were selected to re-evaluate the PPV for each model. The histograms representing the PPV in comparison with the active hit rate from the original training set is shown with and without AD consideration in Figures 3A and 3B, respectively. From the initial analysis (Figure 3A), only two models enriched the active rate by ≥ 2-fold, which are OPRK- and OPRD-agonists, with 5- and 3.5-fold enrichment, respectively. When an AD was defined with a similarity of Tmax>0.8, all 6 models enriched the % active by ≥ 2-fold and the enrichment by the OPRK- and OPRD-agonist models even increased up to 6- and 7-fold, respectively (Figure 3B).

Figure 3.

Comparison of initial (A) and final (B: Tanimoto score consideration) data analysis of PPV with % active hit rate from the original training set.

Docking analysis

To further evaluate the binding potential of the active compounds confirmed in the follow-up experimental study, all the novel potent active (AC₅₀ ≤ 2.5μM) compounds were docked into the binding pocket of the respective active targets. All the ligands yielded binding affinities ranging from −6.7 to −11.4 kcal/mol (Table 2), and these values are comparable to the binding affinities of the known OPR ligands with their corresponding receptors, namely OPRM-BU72 agonist (−10.3 kcal/mol), OPRM-βFNA antagonist (−9.2 kcal/mol), OPRK-MP1104 agonist (−10.4 kcal/mol), OPRK-JDTic antagonist (−9.9 kcal/mol), OPRD-DPI287 agonist (−9.6 kcal/mol), and OPRD-naltrindole antagonist (−10.5 kcal/mol). From the experimental validation, the OPRM-agonist model produced the largest number of potent compounds and the docking interactions of these OPRM-agonist positive compounds showed the highest binding affinities (majority of them are ≤ −9.5 kcal/mol) compared to other OPR models. The most potent compound for each target in complex with the protein is shown in Figure 4A–F, and the active site amino acid residues are indicated as a single letter code followed by their positional number. All the potent compounds are well embedded in the binding pocket. The interactions with the agonist targets are: the amide nitrogen atom of LLY-507 forms a hydrogen bond with Tyr148 residue of OPRM, ridaforolimus forms three hydrogen bonds with Gln115, His304 and Tyr313 residues of OPRK, and NCGC00139128 forms two hydrogen bonds with Asp108 and His301 residues of OPRD. For antagonist targets, the most potent ligand for all three targets is adenosine 3’,5’-cyclothiophosphate, which forms hydrogen bonds with Asp147 residue of OPRM, with Thr111, Gln115, and Asp138 residues of OPRK, and with Lys108 residue of OPRD. The novel ligands were shown to have interactions with the amino acid residues present in the binding pockets of published crystal structures of the opioid targets ^37–42.

Figure 4.

Docked poses of the most potent active compounds. The target receptors are shown as grey helices, the active site amino acid residues are represented as lines in grey, the potent compounds shown as sticks with carbons colored in cyan, and the interactions are shown as black dashed lines. Docked poses in top row are for opioid receptor agonists (A-OPRM; B-OPRK; C-OPRD) and the bottom row is for antagonists (D-OPRM; E-OPRK; F-OPRD)

Discussion

The aim of this study was to develop QSAR models to predict a compound’s agonistic and/or antagonistic effect on OPRM, OPRK, and OPRD targets. A total of 6 QSAR models were developed based on in vitro cell-based assay data. The active compounds constituted only a small percentage in all six qHTS datasets (Figure 1). The HTS hit rates from a diverse compound library are typically ≤ 1%, unless there are some exceptions where the compounds selected are for multi-targets, for which the hit rates can go beyond or up to 10% ⁴⁴. Two approaches were used in our study to balance the classes: random under-sampling and over-sampling. For evaluating the model performances on a cross-validation of training set, an under-sampling strategy was applied due to its computational inexpensiveness, and for the rest of the analysis over-sampling was applied. The two strategies showed comparable performances in terms of AUC-ROC (Table 1 and Supplementary Table S1). The classifiers that were generated without employing any class-balancing strategy were more biased towards the majority class (inactive compounds), for example, a true-positive rate of 0.01 was observed for the OPRM-agonist model though it had the highest percentage of the minority class (active compounds). Based on the performance metrics given in Table 1 and Supplementary Table S1, the RF algorithm was adopted as the method of classification, and PubChem fingerprints were chosen to develop the final predictive models.

The experimental validation of the predictions generated from the RF classifier, resulted in higher rate of false positives (Supplementary Table S3). Even though the QSAR models developed in our study were trained on a structurally diverse set of compounds, the compounds in the NPC are all drugs whereas most of the compounds in the prediction set are novel synthetic molecules that may fall out of the model’s AD ⁴⁵. Predictions made outside of a model’s AD are often unreliable ⁴⁶. We found this to be true in this study as well. When the compounds that fell outside of the model AD were excluded based on a structure similarity cutoff, the models showed significant improvement in performance on the experimental validation set (Figure 3B) ⁴⁷.

Herein, the OPRM-agonist and OPRK-antagonist models predicted more active compounds than other models, and the OPRM-agonist model predictions yielded the largest number of novel potent active compounds that were experimentally validated. Most opioids in use for pain treatment are OPRM agonists, and with some activity exerting on OPRK as well ⁴⁸. The mu OPR knockout mice has demonstrated that they are the sole receptors in mediating morphine’s analgesic and addictive properties ⁴⁹. From our study, 24 novel potent compounds were identified to have effect on the OPRs with the majority active against OPRM (Table 2), these compounds could be developed into new therapies to combat the opioid crisis. Only a few of these compounds have previously reported targets, for example, LLY-507 is a potent and selective inhibitor of protein-lysine methyltransferase SMYD2 ⁵⁰; LY 426965 is an aryl piperazine compound that acts as a serotonin_1A (5-hydroxytryptamine_1A) antagonist ⁵¹; VUF 2274 is a human cytomegalovirus encoded US28 (a GPCR) inhibitor ⁵²; FIPI (a halopemide derivative) is a potent phospholipase D inhibitor ⁵³; NP-118809 (39-1B4) is a potent N-type calcium channel blocker ⁵⁴; and ridaforolimus (MK 8669) is a selective inhibitor of the mammalian target of rapamycin (mTOR) and has an anti-tumor activity ⁵⁵. It is a macrolide known to form a complex with the intracellular receptor FK506-binding protein (FKBP12), and interfere with the mTOR activity ^{56, 57}. The linkage of OPRK with the mTOR system is not well understood, but OPRK-mediated mTOR signaling was shown in the mouse brain as the activation of the mTOR pathway occurs in neurons expressing OPRK ⁵⁸.

The OPR antagonist compounds compete with the agonists and block the receptor thus reverse the agonistic effects, so they are used in the clinic for partial/complete reversal of opioid toxicity, and to relieve opioid-related adverse effects etc. ⁵⁹. The most commonly used antagonists for reversing the opioid toxicity are naloxone, naltrexone (both compounds inhibit all types of OPRs), and naltrindole (OPRD specific) ³. For our initial assay optimization, naloxone was used as a positive control compound, and its IC₅₀ was 1.2nM and 0.22μM for OPRM and OPRK, respectively. Our models also identified novel compounds that exhibited similar potent inhibitory effect on OPRs. Two compounds that are analogs of cyclic adenosine monophosphate (cAMP) showed antagonistic effects on mu, kappa, and delta OPRs, and they are adenosine 3’,5’-cyclothiophosphate and 8-chloro cAMP (tocladesine, an anticancer drug). The first compound was the most potent against all three receptors (IC₅₀ in range of 40–70nM) (Table 2). Another novel compound, NCGC00114968, that contains the 8-hydroxyquinoline (8HQ) moiety, showed potent inhibition against OPRM (IC₅₀ = 0.65μM). 8HQ derivatives have been used as fungicides and a few of its derivatives with the piperazine ring (like in NCGC00114968) were reported to exert antineurodegenerative effect ⁶⁰. Two other novel compounds, which were shown to have a specific inhibitory effect on OPRK, both with IC₅₀ of 2.06μM, are NCGC00106344, a quinazolineacetamide derivative with no known target reported yet and NCGC00386726 (IN-1130), a well-studied drug for its potency in inhibiting the TGF-beta signaling pathway ⁶¹.

We performed docking to gain insights into the interactions of the novel potent (AC₅₀ ≤ 2.5μM) compounds that were validated experimentally, with the active sites of the respective crystal structures of OPRs. The experimentally validated, most potent compound for each OPR was shown to have interactions with residues in the active site (Figure 4). The search for safer drugs to alleviate pain without exerting severe adverse effects was mostly in silico driven, and these approaches have provided important information regarding the structural determinants that are responsible for binding affinity and selectivity of newly identified ligands ⁶². Also through pharmacophore-based modeling, novel antagonists for the mu OPR were identified and evaluated in in vitro ⁶³ and in vivo ⁶⁴ for its significant inhibition of morphine-induced antinociception. Docking approaches to predict the binding affinities of fentanyl derivatives to the mu-OPR have been developed recently ^{65, 66}. Thousands of fentanyl analogues were identified, and a strong correlation was found between the docking scores and experimental binding affinities. These approaches are exploited when in vitro data are not available and may facilitate temporary scheduling of those substances that pose risks to the public. Other studies included the design and synthesis of analogues of known OPR agonists and antagonists, which were evaluated in in vitro pharmacological assays. Such efforts involved modifying the 6^th position of the morphinan that plays a key role in the mu OPR activity ⁶⁷, and replacing the hydroxyl groups with other groups in JDTic to examine their effect on mu, kappa, and delta OPRs ⁶⁸. These target structure-based virtual screening approaches often have limited capacity in identifying novel chemical scaffolds, whereas models developed based on assay data may discover compounds with more diverse structures. Our study presented the first predictive models built on in vitro assay data, which were generated from a large, diverse set of known drugs against OPRs. These models could be applied to virtually screen large compound libraries to identify novel OPR active compounds.

In summary, we developed models based on qHTS data for the prediction of compound activity on three different OPRs. The models identified a number of novel compounds, which were validated experimentally. The potent active compounds were shown to have interactions within the receptor’s binding pocket via molecular docking. All models were able to enrich active hit rate by ≥ 2-fold. These models have the potential to be used for larger collections to predict the compound’s effect on OPRs.

Data and Software Availability

Data and source codes are available for download at:

https://tripod.nih.gov/tox/oprpred/codes.zip

https://tripod.nih.gov/tox/oprpred/data.zip