Skip to main content
NCBI home page
RSC Advances logoLink to RSC Advances
. 2026 Aug 19;16(41):49012–49023. doi: 10.1039/d6ra03044k

µORScreen: a lightweight consensus modeling framework for µ-opioid receptor ligand prediction and virtual screening

Keyu Chen a,†, Juan Huang b,†, Jiangcheng Xu d, Dengwen Bai c, Tao Bi b, Zengjin Liu b, Qin Sun b,✉, Yong Dai a,✉
PMCID: PMC13488679  PMID: 42622020

Abstract

Opioid use disorder (OUD) remains a major public health challenge, and the human µ-opioid receptor (µOR) is a central target in opioid pharmacology. Here, we report a reproducible ligand-based workflow for µOR antagonist classification named µORScreen which integrates rigorous split design, systematic model benchmarking, interpretation, virtual screening and lightweight local deployment. A curated set of 982 human µOR ligands was partitioned under three complementary strategies (similarity-based, scaffold-based, and random-based), each with a held-out test set and five train/validation folds. On the test evaluation, LightGBM (ECFP4 with RDKit 2D descriptors) generalized best (AUROC 0.714), closely followed by TabPFN (0.705) and Random Forest (0.696). The three top-ranked models were combined into a consensus classifier that prioritized unanimously predicted compounds as high-confidence antagonist-like candidates. Applied to GPCRdb, ZINC, REINVENT, and OUROBOROS, µORScreen revealed pronounced source dependence, with the strongest enrichment of antagonist-like candidates in GPCRdb. On an independent set of 17 non-overlapping, literature-derived ligands (10 antagonists, 7 non-antagonists), the consensus achieved a balanced accuracy of 0.68. SHAP analysis attributed predictions to a concentrated subset of fingerprint features, and the workflow was deployed as a web server supporting SMILES-based prediction and RF-based SHAP analysis. µORScreen thus provides a computationally efficient, openly accessible framework for early-stage µOR ligand prioritization and external-library triage.

µORScreen integrates leakage-aware benchmarking and a LightGBM–TabPFN–RF consensus to prioritize µ-opioid receptor antagonist-like compounds, with SHAP interpretation and an openly accessible deployment workflow.graphic file with name d6ra03044k-ga.webp

1. Introduction

The ongoing opioid crisis remains a major global public health challenge, driven by widespread opioid misuse, dependence, and overdose-related mortality.1 Consequently, effective treatment of opioid use disorder (OUD), particularly during withdrawal management and relapse prevention, remains an urgent unmet clinical need.2 Pharmacological modulation of the human µ-opioid receptor (µOR) is central to current therapeutic strategies because µOR mediates not only the analgesic and rewarding effects of opioids, but also the neuroadaptive processes that underlie tolerance, dependence, and withdrawal.3 Clinically used agents such as methadone, buprenorphine, and naltrexone further underscore the importance of µOR-centered pharmacology in OUD management.

At the molecular level, µOR ligands exhibit substantial chemical and pharmacological diversity, and even subtle structural changes can lead to markedly different functional outcomes.4 Efficient identification and prioritization of µOR-related compounds are therefore important both for advancing mechanistic understanding of opioid pharmacology and for supporting early-stage ligand discovery. However, experimental screening of large chemical libraries remains time-consuming and resource-intensive, creating a strong need for computational approaches that enable scalable and efficient ligand prioritization.5,6

Several recent studies have explored the computational prioritization of µOR ligands. Oh et al.7 curated the [35S]GTPγS-derived intrinsic-activity dataset used in this study and established ExtraTrees and message-passing neural network baselines for agonist–antagonist classification. Using the same dataset, Ugurlu8 introduced µOR-ligand, a target-aware framework that augments ligand descriptors and fingerprints with a large space of receptor–ligand interaction descriptors and performs view-based hybrid feature selection over stacked ensembles, demonstrating that target-conditioned features carry complementary signal for functional classification. In a distinct problem setting, Sakamuru et al.9 developed random-forest models over quantitative high-throughput screening data for three opioid receptors in both agonist and antagonist modes, and – uniquely among these studies – experimentally confirmed model-prioritised compounds by qHTS. These studies demonstrate that µOR functional activity can be predicted accurately and that receptor-conditioned features can provide additional predictive value. However, the generalizability of efficient ligand-based models under stringent, similarity-separated evaluation remains insufficiently explored.

Here, we present µORScreen (Fig. 1), a lightweight consensus framework for µOR ligand classification and virtual screening, emphasizing practicality, reproducibility, and deployability. First, µORScreen benchmarks diverse machine-learning (ML) methods – including Random Forest (RF),10 Support Vector Machine (SVM),11 XGBoost,12 LightGBM,13 Chemprop,14,15 Logistic Regression (LR), k-Nearest Neighbors (KNN) and TabPFN16 – under three data-splitting strategies: random, scaffold-based, and similarity-based splitting using DataSAIL17 to minimize information leakage. Comparing performance across these strategies provides a direct assessment of how data leakage and structural overlap between training and test sets influence model evaluation. Second, rather than deploying the entire model panel, µORScreen combines the three best-performing models—LightGBM, TabPFN, and RF, ranked in that order—into a computationally efficient consensus predictor. These models provide complementary signals from gradient-boosted trees, a tabular foundation model, and bagged decision-tree ensembles, respectively. Third, a unanimity-based decision rule prioritizes compounds consistently classified across all three models, enabling the selection of high-confidence antagonist candidates and the conservative exclusion of unanimously predicted negatives. Finally, µORScreen provides a web interface for SMILES prediction, facilitating its practical application.

Fig. 1. Overview of µORScreen for µ-opioid receptor ligand classification and virtual screening. A curated benchmark dataset of opioid-related molecules was partitioned into training, validation and independent test sets using a similarity-, scaffold-, random-based splitting strategies. Eight ML models (RF, SVM, XGBoost, LightGBM, TabPFN, LR, KNN and Chemprop) were benchmarked, and the three best-performing models were integrated into a consensus predictor. RF prediction results were further interpreted using SHAP. µORScreen was subsequently evaluated against an independent literature-derived dataset and applied to the virtual screening of GPCRdb, ZINC, REINVENT, and OUROBOROS libraries. Compounds unanimously predicted as antagonists were prioritized as high-confidence candidates. Finally, µORScreen was deployed as a web server for user-submitted molecule prediction.

Fig. 1

Open in a new tab

2. Methods

2.1. Dataset source and curation

A curated dataset of 982 human µ-opioid receptor (µOR) ligands was used for supervised learning. Molecular records were obtained from the dataset reported by Oh and colleagues,7 who manually curated literature reports of ligand efficacy based on [35S]GTPγS functional assays at human µOR. Compounds were assigned as agonists or antagonists according to the original pharmacological annotations whenever such annotations were available. When explicit efficacy annotations were unavailable, naltrexone (Emax = 14%) served as the reference threshold for label assignment. The resulting dataset comprised 755 agonists and 227 antagonists after removing one duplicate, with each compound represented by a valid SMILES string and a binary activity label.

2.2. Data splitting strategies

To assess the effects of chemical overlap and information leakage, we compared random, Bemis–Murcko scaffold,18 and DataSAIL similarity-based splitting. Random splitting served as the conventional reference, scaffold splitting reduced overlap between core structures, and DataSAIL imposed stricter chemical separation between the development and test sets. As shown in Fig. 2, the mean maximum Morgan Tanimoto similarity to the development set decreased from 0.789 for random splitting to 0.689 for scaffold splitting and 0.486 for DataSAIL. Thus, scaffold splitting reduced—but did not eliminate—chemical overlap, whereas DataSAIL provided the most stringent assessment of generalization.

Fig. 2. Chemical similarity between the held-out test and model-development sets under different splitting strategies. The distribution shows the maximum Morgan fingerprint-based Tanimoto similarity of each test molecule to the development set. Random splitting produced the greatest chemical overlap (n = 98; mean = 0.789), followed by scaffold splitting (n = 98; mean = 0.689). DataSAIL yielded the lowest similarity (n = 94; mean = 0.486), indicating more effective separation of related compounds and a lower risk of information leakage.

Fig. 2

Open in a new tab

Models were first benchmarked by five-fold cross-validation within the development set, with validation and held-out-fold (CV-test) performance recorded for each fold. The selected models were then retrained on the combined training and validation data and evaluated on an independent test set. Comparing results across the three splitting strategies enabled us to quantify the extent to which chemical similarity and information leakage influenced model performance.

2.3. Molecular representation

Several representations were evaluated: 1024-bit Morgan fingerprints, 2048-bit Morgan fingerprints, 200-dimensional normalized RDKit19 descriptors (rdkit2d_normalized_200, adopted from KPGT20), and the concatenation of these descriptors with each Morgan fingerprint. The two combined representations therefore comprised 1224 and 2248 features, respectively. Morgan fingerprints were generated using RDKit and encode atom-centered chemical environments as fixed-length binary vectors, whereas the normalized RDKit descriptors capture complementary physicochemical and structural properties.

Chemprop was evaluated separately using three input configurations: its native molecular-graph representation alone, the graph augmented with Morgan fingerprints, and the graph augmented with both Morgan fingerprints and RDKit descriptors. These representations enabled systematic comparison of graph-based, fingerprint-based, descriptor-based, and multimodal molecular encodings.

2.4. Benchmarked model families

We benchmarked eight classifiers spanning complementary modelling paradigms: LR, KNN, SVM, RF, XGBoost, LightGBM, Chemprop and TabPFN. LR, KNN, SVM and the three tree-based models were evaluated with each of the five ligand representations described in Section 2.3.

Chemprop, a directed message-passing neural network (D-MPNN),21 was evaluated using molecular graphs alone, graphs with Morgan fingerprints, and graphs with both Morgan fingerprints and RDKit descriptors. TabPFN, a transformer-based foundation model for tabular prediction,16,22–24 was evaluated with all five fixed-length representations. All model-representation combinations were assessed using random, scaffold and DataSAIL similarity-based splits.

2.5. Model training and selection

All model-representation combinations were evaluated using a unified pipeline. Within each splitting strategy, hyperparameters and molecular representations were selected by mean validation AUROC. AUPRC, accuracy, F1 and MCC were retained as secondary metrics to assess performance under class imbalance but were not used for primary model selection. The independent test set remained strictly isolated throughout optimization and was evaluated only after the final configuration for each model family had been fixed.

The selected configurations were retrained on the combined training and validation data before independent testing. Model families were then ranked using the predefined validation-based procedure, and LightGBM, TabPFN and RF were retained for consensus screening. After benchmark evaluation, these configurations were retrained on all 982 benchmark compounds for external validation and library screening. All other models were retained as comparative baselines.

2.6. Performance metrics

This study adopts widely used binary classification metrics to evaluate the performance of various machine learning models. These metrics include the area under the receiver operating characteristic curve (AUROC), the area under the precision–recall curve (AUPRC), accuracy (ACC), and F1 score. For binary classification with true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN), accuracy was computed as:2.6.

Precision and recall were defined as:2.6.and the F1 score was calculated as:2.6.

AUROC summarizes classifier performance across all thresholds by plotting the true-positive rate (TPR) against the false-positive rate (FPR), where:2.6.

AUPRC was included because it is particularly informative under class imbalance. For the independent external validation analysis, we additionally report specificity and balanced accuracy. Specificity was defined as:2.6.and balanced accuracy was computed as:2.6.

Among these metrics, AUROC was used as the primary model-ranking metric.

2.7. Consensus screening framework

To improve prediction robustness and reduce model-specific bias, we implemented a consensus-based inference framework. Although µORScreen benchmarks the full baseline panel, only the top-ranked models were retained for downstream screening. Each molecule received per-model probabilities and binary predictions from the selected models. Compounds predicted as positive by all selected models were designated as high-confidence antagonist-like candidates, whereas compounds predicted as negative by all selected models were designated as high-confidence predicted negatives or screening exclusions. All remaining compounds were treated as intermediate-confidence cases. In addition, the mean predicted probability across the top three models was used as a composite score for library-level distribution analysis. This framework provides a practical and risk-aware strategy for large-scale ligand prioritization.

2.8. Model interpretation and SAR analysis

To enhance model interpretability, we conducted global and compound-level SAR analyses of the well-performing RF model. SHAP analysis identified the fingerprint bits that contributed most strongly to predictions, while representative compounds were examined to map antagonist- and agonist-associated contributions onto specific substructures. Together, these analyses improve model transparency and provide testable hypotheses for medicinal chemistry.

2.9. External virtual screening

To evaluate the practical utility and source dependence of µORScreen, we applied the trained models and the top-3 consensus framework to four external libraries with distinct chemical origins: a GPCR-focused set from GPCRdb, a scaffold-matched subset from ZINC, and generative libraries from REINVENT and OUROBOROS.

GPCRdb25 was included as a biologically relevant reference source because it is a curated resource that integrates GPCR-centered sequence, structure, ligand, and bioactivity information, including ligand annotations aggregated from multiple pharmacological databases. For this study, we collected opioid-related molecules associated with µOR-relevant pharmacology from GPCRdb. After initial collection and preprocessing, this set contained 11 312 molecules before final de-duplication.

ZINC26 was used as a broader purchasable screening background. It is a public database of commercially available compounds developed to support virtual screening and is also a major chemical source used in the pretraining of modern molecular language models such as MolFormer.27 Bemis–Murcko decomposition of the 982 reference compounds yielded 502 unique scaffolds: 129 exclusive to antagonists, 347 exclusive to non-antagonists and 26 shared between classes. ZINC was queried against each scaffold without class-specific filtering, yielding 14 555 exact scaffold matches. Canonicalization, validity filtering, removal of benchmark overlaps and within-source deduplication reduced this set to 7041 compounds. This scaffold-constrained subset therefore represents a chemically focused screening background rather than an unrestricted sample of ZINC.

REINVENT28 and OUROBOROS29 were included as generative sources to test whether model-guided molecular generation could provide additional candidate space beyond database-derived compounds. REINVENT is a reinforcement-learning-based de novo molecular design platform that optimizes generated molecules against a user-defined multi-component scoring function. OUROBOROS, as described in its public release, combines learned molecular representations with molecular generation for efficient exploration of chemical space. Using these two generators, we produced 58 056 and 19 562 molecules, respectively, under QED-guided generation settings. All four libraries then underwent the same structural standardization and de-duplication procedure before screening.

All molecules from the four external libraries were processed using the same RDKit-based fingerprint pipeline as the labeled training data. Predictions were generated in batch mode, and the top three performers identified under DataSAIL-validation/shared-test selection procedure for the final consensus screen. For each compound, we calculated the mean predicted probability across the top three models as a consensus score and assigned consensus labels according to unanimity among binary model outputs. For each library, we summarized the proportions of consensus-positive and consensus-negative compounds and compared score distributions using the empirical cumulative distribution function (ECDF) of the top-3 mean probability.

2.10. Independent external validation

To further assess out-of-distribution generalization beyond scaffold-aware splitting and large-library screening, we constructed a small independent external validation set comprising 17 literature-derived human µ-opioid receptor (µOR) ligands, including 10 antagonists and 7 non-antagonists. Molecules were collected from recent or authoritative sources, standardized to canonical SMILES, and filtered to retain only valid structures. To ensure strict independence from model development, all compounds were explicitly de-duplicated against the full curated benchmark dataset, including the training, validation, and held-out test partitions, resulting in zero overlap and no invalid structures.

The external set was evaluated using the three-model consensus comprising LightGBM, TabPFN and RF. For each model, the best-performing configuration identified by cross-validation was retrained on all 982 benchmark compounds and subsequently used for both external validation and library screening.30 Consensus labels were assigned only when all three models agreed.

2.11. Web server implementation

To support practical use and result sharing, µORScreen is provided as a set of deployment scripts implementing the final LightGBM–TabPFN–RF consensus (Fig. 3). Each model uses the representation and parameters selected during cross-validation, ensuring consistency with the offline screening pipeline. Compounds are prioritized as antagonists only when all three models agree. RF, although not the strongest individual model, is retained to provide efficient SHAP explanations.

Fig. 3. µORScreen web interface for consensus prediction and local interpretation. Users submit a SMILES string to obtain predictions from LightGBM, TabPFN and RF, together with a unanimity-based consensus label. The example shown is unanimously classified as antagonist-like, with the individual model scores displayed separately. The lower panel presents the RF SHAP explanation, highlighting fingerprint-derived substructures that increase or decrease the antagonist prediction score.

Fig. 3

Open in a new tab

The scripts support single-SMILES prediction and batch screening via CSV input. In single-compound mode (Fig. 3), they report the canonical SMILES, individual model scores and labels, and the consensus prediction; RF feature contributions are mapped to molecular substructures to provide a local explanation. Batch files may contain up to 1000 molecules in a smiles or SMILES column. Original metadata are preserved, prediction results and parsing status are appended, and invalid SMILES are retained and flagged. Overall, the web server translates the offline µORScreen workflow into a practical, reproducible, and user-friendly tool for rapid µ-opioid receptor ligand screening.

3. Results and discussion

3.1. Data-splitting strategy dominates benchmark performance

We benchmarked eight classifiers under random, scaffold and DataSAIL similarity-based splits. Performance was consistently high under random and scaffold splitting, with held-out AUROCs of 0.93–0.97 and 0.93–0.97, respectively (Fig. 4A). By contrast, DataSAIL substantially reduced performance across all models: held-out AUROC ranged from 0.59 to 0.78, and AUPRC from 0.35 to 0.61 (Fig. 4A and B). This decline indicates that random and scaffold splits retain chemical similarity between the development and test sets, leading to optimistic estimates of generalization.

Fig. 4. Benchmark performance across models and data-splitting strategies. Performance of SVM, RF, XGBoost, LightGBM, TabPFN, Chemprop, KNN and LR under scaffold, random and DataSAIL similarity-based splitting. Light and dark bars denote validation and held-out test performance, respectively. Models were evaluated using the metrics shown in panels (A–D): (A) AUROC; (B) AUPRC; (C) accuracy; and (D) F1 score. The marked performance reduction under DataSAIL indicates that reduced chemical overlap between training and test sets provides a more challenging assessment of model generalization.

Fig. 4

Open in a new tab

The effect was particularly pronounced for minority-class prediction. Under DataSAIL, held-out accuracy remained relatively stable at 0.77–0.78, whereas F1 scores fell to 0.02–0.47 (Fig. 4C and D), highlighting the limited informativeness of accuracy for this imbalanced dataset. Differences among model families were smaller than those induced by the splitting strategy, suggesting that chemical separation, rather than model complexity, was the principal determinant of benchmark performance. The top-performing models exhibited broadly comparable predictive capability; however, we provide a flexible framework that enables users to rapidly substitute and integrate their own models, facilitating efficient ensemble construction and combined prediction workflows. LightGBM, TabPFN and RF were retained for consensus screening according to the predefined model-selection procedure. Full results are provided in Tables S1 and S2.

The top three models represent distinct learning paradigms—tree-based ensemble learning (RF), a tabular foundation model (TabPFN), and gradient-boosted decision trees (LightGBM)—suggesting that they capture partially complementary predictive signals. This complementarity motivated their integration into the downstream consensus framework.

Although the held-out test set was scaffold-separated from the model-development pool, it remained chemically similar to the training and validation compounds, which may partly explain the relatively strong performance observed across multiple baseline models. Overall, these results establish a robust benchmark for µORScreen and provide the foundation for the subsequent consensus-screening strategy.

3.2. RF-based SHAP interpretation

RF was selected for post hoc interpretation because it supports efficient and robust SHAP analysis,31,32 rather than because it was the best-performing classifier.31,32 Antagonists were defined as the positive class; positive SHAP values therefore shift predictions towards antagonist-like activity, whereas negative values favour the non-antagonist class.

Global analysis showed that RF predictions were dominated by a limited subset of fingerprint bits (Fig. 5), indicating concentrated rather than diffuse structural contributions. Mapping high-impact bits to their corresponding atomic environments identified substructures associated with agonist–antagonist discrimination and provided chemically interpretable hypotheses for subsequent analysis.

Fig. 5. Global RF SHAP analysis and representative compound predictions. (A) Representative compounds showing molecular structure, experimental antagonist annotation, prediction outcome and model score. RF-01 is a boundary false positive; RF-02–RF-04 are correctly classified boundary compounds; RF-05–RF-08 and RF-09–RF-12 are high-confidence non-antagonist and antagonist predictions, respectively. (B) SHAP summary plot for the leading fingerprint bits. Positive SHAP values shift predictions towards the antagonist class (label 1), whereas negative values favour the non-antagonist class. (C) Global feature importance ranked by mean absolute SHAP value. (D) Class-conditional occurrence frequencies of selected high-impact bits. (E) Representative atomic environments associated with these bits. Together, the panels summarize the global decision logic of RF and its behaviour across representative prediction-confidence regimes.

Fig. 5

Open in a new tab

Analysis of representative compounds contextualized the global SHAP results. RF-01 was a false positive near the decision boundary (score, 0.501), whereas RF-02–RF-04 were correctly classified antagonists with moderate scores (0.559–0.638). By contrast, RF-05–RF-08 were confidently classified as non-antagonists, and RF-09–RF-12 as antagonists. These examples illustrate robust predictions at the extremes of the score distribution and greater uncertainty near the classification threshold.

Global SHAP analysis identified a limited set of fingerprint bits with substantial influence on RF predictions. Their class-conditional frequencies and mapped chemical environments revealed structural patterns associated with antagonist-like or non-antagonist predictions. These attributions explain the model's decision process but should not be interpreted as causal pharmacological determinants. More generally, 2D fingerprints capture local substructures but do not explicitly represent molecular conformation, receptor–ligand interactions or assay context, limiting their ability to resolve functional activity in unfamiliar chemical space.

3.3. The screening results

To evaluate source-dependent screening behavior, we applied the three-model consensus framework to external libraries from GPCRdb, ZINC, REINVENT, and OUROBOROS. Following RDKit canonicalization, invalid-SMILES removal, exclusion of compounds overlapping the training corpus, and within-source de-duplication, the final screening sets contained 11 312 GPCRdb compounds, 7041 ZINC compounds, 58 056 REINVENT compounds, and 19 562 OUROBOROS compounds.

GPCRdb produced the highest consensus-positive fraction, with 674 compounds prioritized by all three models (5.96%). The corresponding results were 174 compounds for REINVENT (0.30%), 28 for OUROBOROS (0.14%), and 9 for ZINC (0.13%). Thus, the source ordering was GPCRdb, REINVENT, OUROBOROS, and ZINC. These differences reflect source-dependent enrichment in chemical structures favored by the current uncalibrated, antagonist-oriented screening ensemble and should not be interpreted as experimentally confirmed antagonist rates.

Source dependence was also evident in the distribution of three-model mean antagonist scores (Fig. 6). GPCRdb was markedly right-shifted and exhibited the broadest upper-score tail, consistent with its higher unanimity-based hit rate. By contrast, REINVENT was concentrated predominantly at low scores. ZINC and OUROBOROS showed intermediate distributions, although neither approached the population-level score enrichment observed for GPCRdb. Sparse upper tails were nevertheless present in the generative libraries, indicating that they may contribute a small subset of candidates for further prioritization.

Fig. 6. Source-dependent distributions of three-model mean antagonist scores. Empirical cumulative distribution functions of the mean uncalibrated scores from LightGBM, TabPFN and RF for compounds from GPCRdb, ZINC, REINVENT and OUROBOROS. At a given score threshold, a lower cumulative fraction indicates a larger proportion of compounds exceeding that threshold. GPCRdb shows the strongest rightward shift and the broadest high-score tail, whereas the other libraries are concentrated primarily in the low-to-intermediate score range.

Fig. 6

Open in a new tab

The low predicted hit rate for ZINC should be interpreted in the context of library construction. The screened subset was generated using stringent scaffold-matching criteria, which constrained its accessible chemical space relative to the other libraries. Its limited high-score tail therefore reflects the selected subset and should not be generalized to ZINC as a whole.

Together, these results identify GPCRdb as the most enriched source of predicted antagonist-like compounds, whereas REINVENT and OUROBOROS provide lower-yield but chemically distinct candidate pools. Because these scores are uncalibrated model outputs, experimental validation remains necessary to establish pharmacological activity.

3.4. Validation on a literature-derived µOR ligand set

The literature-derived validation set comprised 20 µOR ligands, including 10 antagonists and 10 non-antagonists. Three non-antagonists—PZM21 (EXT-N05), cebranopadol (EXT-N08) and endomorphin-1 (EXT-N10)—were also present in the benchmark dataset, leaving 17 compounds with no exact molecular overlap. Given the small sample size, performance was summarized using accuracy only. The LightGBM–TabPFN–RF consensus correctly classified 14 of 20 compounds (balanced accuracy, 0.70). After excluding the three overlapping molecules, 11 of 17 compounds were correctly classified (balanced accuracy, 0.679).

Methocinnamox,33 naloxonazine,34 axelopran,35 CTOP36 and CTAP37 received unanimous antagonist predictions. By contrast, LY255582,38 6β-naltrexol39 and bevenopran40 were missed by the unanimity rule, whereas JDTic was the only false positive. Compound-level annotations, overlap status and individual model predictions are provided in Table S3.

These results indicate that µORScreen retains predictive utility on literature-derived compounds, although the limited sample size and partial overlap with the benchmark dataset preclude definitive conclusions about external generalization.

4. Conclusion

We developed µORScreen, a lightweight and reproducible framework for µOR ligand classification and virtual screening. Benchmarking eight classifiers across random, scaffold and DataSAIL similarity-based splits showed that data partitioning had a greater effect on apparent performance than model family, highlighting the risk of information leakage in chemically related datasets. LightGBM, TabPFN and RF were integrated into a unanimity-based consensus, with RF additionally supporting SHAP interpretation.

Screening of GPCRdb, ZINC, REINVENT and OUROBOROS revealed marked source-dependent differences in predicted antagonist-like enrichment. GPCRdb produced the largest high-scoring candidate pool, whereas the generative libraries contributed smaller prioritized subsets. Deployment scripts support single-SMILES prediction, batch screening and RF-based local explanations. Evaluation on a small literature-derived set provided preliminary evidence of transferability but was limited by sample size and partial overlap with the benchmark dataset. Further work should therefore emphasize prospective experimental validation, probability calibration, broader chemical coverage and extension to richer pharmacological endpoints.

Author contributions

Keyu Chen: writing – original draft, visualization, validation, methodology, investigation, formal analysis, data curation. Juan Huang: investigation, formal analysis, data curation. Jiangcheng Xu: investigation, formal analysis, data curation. Dengwen Bai: investigation, formal analysis, data curation. Tao Bi: methodology, project administration, formal analysis. Zengjin Liu: methodology, investigation. Qin Sun: writing – review & editing, supervision, methodology, investigation. Yong Dai: writing – review & editing, supervision, project administration, funding acquisition.

Conflicts of interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Supplementary Material

RA-016-D6RA03044K-s001
RA-016-D6RA03044K-s002
RA-016-D6RA03044K-s003

Acknowledgments

This research was funded by the Sichuan Science and Technology Program (No. 2024YFFK0123); Sichuan Provincial Administration of Traditional Chinese Medicine (No. 25ZD1IZX027); The Research Project on the Integration of Traditional Chinese and Western Medicine Conducted at Southwest Medical University (No. 2025ZXYZX01) and Science and Technology Strategic Cooperation Project of Luzhou Municipal People's Government-Southwest Medical University (No. 2025LZXNYDYG02). The funders had no role in study design, data collection, and analysis, decision to publish, or preparation of the manuscript.

Data availability

All data, source code, analysis and visualization scripts, tutorial materials, and web server deployment files generated or used in this study are freely available at the GitHub repository: https://github.com/MolAstra/opioid.

Supplementary information (SI): additional benchmarking and validation results. See DOI: https://doi.org/10.1039/d6ra03044k.

References

  1. The Lancet Public Health Opioid overdose crisis: time for radical rethink. The Lancet Public Health. 2022;7(3):e195. doi: 10.1016/S2468-2667(22)00043-3. [DOI] [PubMed] [Google Scholar]
  2. Blanco C. Wiley T. R. A. Lloyd J. J. Lopez M. F. Volkow N. D. America's Opioid Crisis: The Need for an Integrated Public Health Approach. Transl. Psychiatry. 2020;10(1):167. doi: 10.1038/s41398-020-0847-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
  3. Herman T., Cascella M. and Muzio M. R., Mu Receptors, StatPearls, 2024 [PubMed] [Google Scholar]
  4. Valentino R. J. Volkow N. D. Untangling the Complexity of Opioid Receptor Function. Neuropsychopharmacology. 2018;43(13):2514–2520. doi: 10.1038/s41386-018-0225-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Moshawih S. Bu Z. H. Goh H. P. Kifli N. Lee L. H. Goh K. W. Ming L. C. Consensus Holistic Virtual Screening for Drug Discovery: A Novel Machine Learning Model Approach. J. Cheminformatics. 2024;16(1):62. doi: 10.1186/s13321-024-00855-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
  6. Noor F. Junaid M. Almalki A. H. Almaghrabi M. Ghazanfar S. Tahir ul Qamar M. Deep Learning Pipeline for Accelerating Virtual Screening in Drug Discovery. Sci. Rep. 2024;14(1):28321. doi: 10.1038/s41598-024-79799-w. [DOI] [PMC free article] [PubMed] [Google Scholar]
  7. Oh M. Shen M. Liu R. Stavitskaya L. Shen J. Machine Learned Classification of Ligand Intrinsic Activities at Human µ-Opioid Receptor. ACS Chem. Neurosci. 2024;15(15):2842–2852. doi: 10.1021/acschemneuro.4c00212. [DOI] [PMC free article] [PubMed] [Google Scholar]
  8. Ugurlu S. Y. µOR-Ligand: Target-Aware View-Based Hybrid Feature Selection for µ-Opioid Receptor Ligand Functional Classification. J. Comput. Aided Mol. Des. 2025;39(2):103. doi: 10.1007/s10822-025-00686-1. [DOI] [PubMed] [Google Scholar]
  9. Sakamuru S. Zhao J. Xia M. Hong H. Simeonov A. Vaisman I. Huang R. Predictive Models to Identify Small Molecule Activators and Inhibitors of Opioid Receptors. J. Chem. Inf. Model. 2021;61(6):2675–2685. doi: 10.1021/acs.jcim.1c00439. [DOI] [PMC free article] [PubMed] [Google Scholar]
  10. Svetnik V. Liaw A. Tong C. Culberson J. C. Sheridan R. P. Feuston B. P. Random Forest: A Classification and Regression Tool for Comound Classification and QSAR Modeling. J. Chem. Inf. Comput. Sci. 2003;43(6):1947–1958. doi: 10.1021/ci034160g. [DOI] [PubMed] [Google Scholar]
  11. Cristianini N. and Shawe-Taylor J., An Introduction to Support Vector Machines and Other Kernel-Based Learning Methods, 2010 [Google Scholar]
  12. Chen T. and Guestrin C., XGBoost: A Scalable Tree Boosting System, in Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, ACM: San Francisco California USA, 2016, pp. 785–794, 10.1145/2939672.2939785 [DOI] [Google Scholar]
  13. Ke G. L., Meng Q., Finley T., Wang T. F., Chen W., Ma W. D., Ye Q. W. and Liu T. Y., LightGBM: A Highly Efficient Gradient Boosting Decision Tree, in In Proceedings of the 31st Annual Conference on Neural Information Processing Systems, Long Beach, CA, USA, 2017 [Google Scholar]
  14. Heid E. Greenman K. P. Chung Y. Li S.-C. Graff D. E. Vermeire F. H. Wu H. Green W. H. McGill C. J. Chemprop: A Machine Learning Package for Chemical Property Prediction. J. Chem. Inf. Model. 2024;64(1):9–17. doi: 10.1021/acs.jcim.3c01250. [DOI] [PMC free article] [PubMed] [Google Scholar]
  15. Graff D. E. Morgan N. K. Burns J. W. Doner A. C. Li B. Li S.-C. Manu J. Menon A. Pang H.-W. Wu H. Zalte A. S. Zheng J. W. Coley C. W. Green W. H. Greenman K. P. Chemprop v2: An Efficient, Modular Machine Learning Package for Chemical Property Prediction. J. Chem. Inf. Model. 2026;66(1):28–33. doi: 10.1021/acs.jcim.5c02332. [DOI] [PubMed] [Google Scholar]
  16. Hollmann N. Müller S. Purucker L. Krishnakumar A. Körfer M. Hoo S. B. Schirrmeister R. T. Hutter F. Accurate Predictions on Small Data with a Tabular Foundation Model. Nature. 2025;637(8045):319–326. doi: 10.1038/s41586-024-08328-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
  17. Joeres R. Blumenthal D. B. Kalinina O. V. Data Splitting to Avoid Information Leakage with DataSAIL. Nat. Commun. 2025;16(1):3337. doi: 10.1038/s41467-025-58606-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
  18. van Tilborg D. Alenicheva A. Grisoni F. Exposing the Limitations of Molecular Machine Learning with Activity Cliffs. J. Chem. Inf. Model. 2022;62(23):5938–5951. doi: 10.1021/acs.jcim.2c01073. [DOI] [PMC free article] [PubMed] [Google Scholar]
  19. Landrum G., Tosco P., Kelley B., Rodriguez R., Cosgrove D., Vianello R., Gedeck P., Jones G., Kawashima E. and Nealschneider D., et al., Rdkit/Rdkit: 2025_03_1 (Q1 2025) Release, Zenodo, 2025 [Google Scholar]
  20. Li H. Zhang R. Min Y. Ma D. Zhao D. Zeng J. A. Knowledge-Guided Pre-Training Framework for Improving Molecular Representation Learning. Nat. Commun. 2023;14(1):7568. doi: 10.1038/s41467-023-43214-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
  21. Liao H.-C. Lin Y.-H. Peng C.-H. Li Y.-P. Directed Message Passing Neural Networks for Accurate Prediction of Polymer–Solvent Interaction Parameters. ACS Eng. Au. 2025;5(5):530–539. doi: 10.1021/acsengineeringau.5c00027. [DOI] [Google Scholar]
  22. Vaswani A., Shazeer N., Parmar N., Uszkoreit J., Jones L., Gomez A. N., Kaiser Ł. and Polosukhin I., Attention Is All You Need, in Advances in Neural Information Processing Systems, ed. I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan and R. Garnett, Curran Associates, Inc., 2017, vol. 30 [Google Scholar]
  23. Chen W. Tian Y. Liao N. Deng Y. Jiang D. Cao D. TabPFN Opens New Avenues for Small-Data Tabular Learning in Drug Discovery. J. Chem. Inf. Model. 2026;66(7):3525–3539. doi: 10.1021/acs.jcim.5c02823. [DOI] [PubMed] [Google Scholar]
  24. Chupakhin V. DiBella J. Descriptor-First Approach for ADMET Prediction in the PolarisHub Antiviral Challenge. J. Chem. Inf. Model. 2026;66(1):406–412. doi: 10.1021/acs.jcim.5c02094. [DOI] [PubMed] [Google Scholar]
  25. Pándy-Szekeres G. Munk C. Tsonkov T. M. Mordalski S. Harpsøe K. Hauser A. S. Bojarski A. J. Gloriam D. E. GPCRdb in 2018: Adding GPCR Structure Models and Ligands. Nucleic Acids Res. 2018;46(D1):D440–D446. doi: 10.1093/nar/gkx1109. [DOI] [PMC free article] [PubMed] [Google Scholar]
  26. Jin H. Merz Jr K. M. Modeling Zinc Complexes Using Neural Networks. J. Chem. Inf. Model. 2024;64(8):3140–3148. doi: 10.1021/acs.jcim.4c00095. [DOI] [PMC free article] [PubMed] [Google Scholar]
  27. Ross J. Belgodere B. Chenthamarakshan V. Padhi I. Mroueh Y. Das P. Large-Scale Chemical Language Representations Capture Molecular Structure and Properties. Nat. Mach. Intell. 2022;4(12):1256–1264. doi: 10.1038/s42256-022-00580-7. [DOI] [Google Scholar]
  28. Loeffler H. H. He J. Tibo A. Janet J. P. Voronov A. Mervin L. H. Engkvist O. Reinvent 4: Modern AI–Driven Generative Molecule Design. J. Cheminf. 2024;16(1):20. doi: 10.1186/s13321-024-00812-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
  29. Wang L. Wu Y. Luo H. Liang M. Zhou Y. Chen C. Liu C. Zhang J. Zhang Y. Learned Conformational Space and Pharmacophore Into Molecular Foundational Model. Adv. Sci. 2026;13(17):e13556. doi: 10.1002/advs.202513556. [DOI] [PMC free article] [PubMed] [Google Scholar]
  30. Zhan X. Liu T. Yu C. Huang Y.-A. You Z. Siu S. W. I. MAARDTI: A Multi-Perspective Attention Aggregation Model for the Prediction of Drug–Target Interactions. Digit. Discov. 2025;4(10):2994–3007. doi: 10.1039/D5DD00311C. [DOI] [Google Scholar]
  31. Mosca E., Szigeti F., Tragianni S., Gallagher D. and Groh G., SHAP-Based Explanation Methods: A Review for NLP Interpretability, in Proceedings of the 29th International Conference on Computational Linguistics, 2022, pp. 4593–4603 [Google Scholar]
  32. Lundberg S. M. Erion G. Chen H. DeGrave A. Prutkin J. M. Nair B. Katz R. Himmelfarb J. Bansal N. Lee S.-I. From local explanations to global understanding with explainable AI for trees. Nature Machine Intelligence. 2020;2:56–67. doi: 10.1038/s42256-019-0138-9. https://dx.doi.org/10.1038/s42256-019-0318-9 [DOI] [PMC free article] [PubMed] [Google Scholar]
  33. Broadbear J. H. Sumpter T. L. Burke T. F. Husbands S. M. Lewis J. W. Woods J. H. Traynor J. R. Methocinnamox Is a Potent, Long-Lasting, and Selective Antagonist of Morphine-Mediated Antinociception in the Mouse: Comparison with Clocinnamox, β-Funaltrexamine, and β-Chlornaltrexamine11This Work Was Supported by United States Public Health Service Grant DA-00254. J. Pharmacol. Exp. Ther. 2000;294(3):933–940. doi: 10.1016/S0022-3565(24)39155-4. [DOI] [PubMed] [Google Scholar]
  34. Ling G. S. F. Simantov R. Clark J. A. Pasternak G. W. Naloxonazine Actions in Vivo. Eur. J. Pharmacol. 1986;129(1):33–38. doi: 10.1016/0014-2999(86)90333-X. [DOI] [PubMed] [Google Scholar]
  35. Long D. D. Armstrong S. R. Beattie D. T. Campbell C. B. Church T. J. Colson P.-J. Dalziel S. M. Jacobsen J. R. Jiang L. Obedencio G. P. Rapta M. Saito D. Stergiades I. Tsuruda P. R. Van Dyke P. M. Vickery R. G. Discovery of Axelopran (TD-1211): A Peripherally Restricted µ-Opioid Receptor Antagonist. ACS Med. Chem. Lett. 2019;10(12):1641–1647. doi: 10.1021/acsmedchemlett.9b00406. [DOI] [PMC free article] [PubMed] [Google Scholar]
  36. Gulya K. Kriván M. Nyolczas N. Sarnyai Z. Kovács G. L. Central Effects of the Potent and Highly Selective µ Opioid Antagonist (CTOP) in Mice. Eur. J. Pharmacol. 1988;150(3):355–360. doi: 10.1016/0014-2999(88)90018-0. [DOI] [PubMed] [Google Scholar]
  37. Steinmiller C. L. Young A. M. Pharmacological Selectivity of CTAP in a Warm Water Tail-Withdrawal Antinociception Assay in Rats. Psychopharmacol. 2008;195(4):497–507. doi: 10.1007/s00213-007-0898-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
  38. Sahr A. E. Sindelar D. K. Alexander-Chacko J. T. Eastwood B. J. Mitch C. H. Statnick M. A. Activation of Mesolimbic Dopamine Neurons during Novel and Daily Limited Access to Palatable Food Is Blocked by the Opioid Antagonist LY255582. Am. J. Physiol.: Regul., Integr. Comp. Physiol. 2008;295(2):R463–R471. doi: 10.1152/ajpregu.00390.2007. [DOI] [PubMed] [Google Scholar]
  39. Raehal K. M. Lowery J. J. Bhamidipati C. M. Paolino R. M. Blair J. R. Wang D. Sadée W. Bilsky E. J. In Vivo Characterization of 6β-Naltrexol, an Opioid Ligand with Less Inverse Agonist Activity Compared with Naltrexone and Naloxone in Opioid-Dependent Mice. J. Pharmacol. Exp. Ther. 2005;313(3):1150–1162. doi: 10.1124/jpet.104.082966. [DOI] [PubMed] [Google Scholar]
  40. Luthra P. Burr N. E. Brenner D. M. Ford A. C. Efficacy of Pharmacological Therapies for the Treatment of Opioid-Induced Constipation: Systematic Review and Network Meta-Analysis. Gut. 2019;68(3):434–444. doi: 10.1136/gutjnl-2018-316001. [DOI] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

RA-016-D6RA03044K-s001
RA-016-D6RA03044K-s001.pdf (126.6KB, pdf)
RA-016-D6RA03044K-s002
RA-016-D6RA03044K-s002.xlsx (20KB, xlsx)
RA-016-D6RA03044K-s003
RA-016-D6RA03044K-s003.xls (27.5KB, xls)

Data Availability Statement

All data, source code, analysis and visualization scripts, tutorial materials, and web server deployment files generated or used in this study are freely available at the GitHub repository: https://github.com/MolAstra/opioid.

Supplementary information (SI): additional benchmarking and validation results. See DOI: https://doi.org/10.1039/d6ra03044k.

Articles from RSC Advances are provided here courtesy of Royal Society of Chemistry

RESOURCES