Machine Learning and Structure-Based Virtual Screening for Reliability-Aware G Protein-Coupled Receptor Ligand Prioritization

Background: Predictive performance is necessary but insufficient for G protein-coupled receptor (GPCR) ligand prioritization, where heterogeneous bioactivity data, analog-series overlap, receptor-state dependence, and polypharmacology can weaken high-scoring predictions. Methods: We developed an auditable ligand- and structure-based workflow across five human GPCRs: DRD2, CB2, ADORA2A, OPRM1, and CCR5. ChEMBL Ki, IC50, and EC50 data were evaluated using scaffold-disjoint validation, external validation, Venn–Abers calibration, conformal prediction, applicability-domain analysis, explainability, and receptor-state-aware docking. Results: Deployed classifiers achieved held-out ROC-AUC values of 0.876–0.929. Under stricter temporal validation, models developed exclusively on pre-cutoff data achieved ROC-AUC values of 0.617–0.894 on the same post-cutoff scaffold-held-out compounds, indicating optimism in the original temporal stress test arising from post-cutoff analog information. De-overlapped BindingDB validation was informative for ADORA2A (n = 134) and OPRM1 (n = 49), whereas smaller DRD2, CB2, and CCR5 strata were descriptive only. Venn–Abers provided validity-aware probability intervals, although calibration metrics did not improve uniformly. Classification error increased outside the applicability domain for all targets, while regression conformal coverage decreased. Morgan fingerprints dominated predictive discrimination, whereas physicochemical descriptors recurred more frequently among top SHAP features across targets. DrugBank screening yielded 92 high-confidence, dataset-novel predictions, with structure-based filtering reducing prioritization to three evidence-classified compounds across three targets. Conclusions: Overall, the study presents a reliability-aware framework for evidence-qualified GPCR ligand prioritization rather than an experimentally confirmed candidate-discovery campaign. Prospective biochemical and functional validation, including potency and selectivity assessment where appropriate, is required before candidate-level conclusions can be established.

Authors

Institutions

Publication Details

Journal
Pharmaceuticals
Published
2026-11-17
DOI
https://doi.org/10.3390/ph19111829
Primary Topic
Computational Drug Discovery Methods
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Machine Learning and Structure-Based Virtual Screening for Reliability-Aware G Protein-Coupled Receptor Ligand Prioritization

Hezekiel Mathambo Kumalo, Sifiso S. Makhathini, Sphamandla E. Mtambo, Cornelius Cano Ssemakalu
2 links
Pharmaceuticals
Computational Drug Discovery Methods
article

Machine Learning and Structure-Based Virtual Screening for Reliability-Aware G Protein-Coupled Receptor Ligand Prioritization

Hezekiel Mathambo Kumalo, Sifiso S. Makhathini, Sphamandla E. Mtambo, Cornelius Cano Ssemakalu
article en

Abstract

Background: Predictive performance is necessary but insufficient for G protein-coupled receptor (GPCR) ligand prioritization, where heterogeneous bioactivity data, analog-series overlap, receptor-state dependence, and polypharmacology can weaken high-scoring predictions. Methods: We developed an auditable ligand- and structure-based workflow across five human GPCRs: DRD2, CB2, ADORA2A, OPRM1, and CCR5. ChEMBL Ki, IC50, and EC50 data were evaluated using scaffold-disjoint validation, external validation, Venn–Abers calibration, conformal prediction, applicability-domain analysis, explainability, and receptor-state-aware docking. Results: Deployed classifiers achieved held-out ROC-AUC values of 0.876–0.929. Under stricter temporal validation, models developed exclusively on pre-cutoff data achieved ROC-AUC values of 0.617–0.894 on the same post-cutoff scaffold-held-out compounds, indicating optimism in the original temporal stress test arising from post-cutoff analog information. De-overlapped BindingDB validation was informative for ADORA2A (n = 134) and OPRM1 (n = 49), whereas smaller DRD2, CB2, and CCR5 strata were descriptive only. Venn–Abers provided validity-aware probability intervals, although calibration metrics did not improve uniformly. Classification error increased outside the applicability domain for all targets, while regression conformal coverage decreased. Morgan fingerprints dominated predictive discrimination, whereas physicochemical descriptors recurred more frequently among top SHAP features across targets. DrugBank screening yielded 92 high-confidence, dataset-novel predictions, with structure-based filtering reducing prioritization to three evidence-classified compounds across three targets. Conclusions: Overall, the study presents a reliability-aware framework for evidence-qualified GPCR ligand prioritization rather than an experimentally confirmed candidate-discovery campaign. Prospective biochemical and functional validation, including potency and selectivity assessment where appropriate, is required before candidate-level conclusions can be established.

PharmaceuticalsVol. 19(11)
Vaal University of Technology (ZA), University of KwaZulu-Natal (ZA)
Good health and well-being
Computational Drug Discovery Methods
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon