<p class="MsoNormal" style="margin-bottom: 12.0pt; text-align: left; mso-line-height-alt: 14.0pt; layout-grid-mode: char; mso-layout-grid-align: none; border: none; mso-padding-alt: 31.0pt 31.0pt 31.0pt 31.0pt; mso-border-shadow: yes;" align="left">MultiEndpointTox: A Chemoinformatics Platform for Multidimensional Drug Toxicity Profiling Using Interpretable Machine Learning, Multi-Task Learning, and Integrated Risk Scoring

Drug-induced toxicity remains a principal driver of attrition in pharmaceutical development, yet conventional screening paradigms typically address individual toxicity endpoints in isolation. Here, we introduce MultiEndpointTox, a chemoinformatics platform that simultaneously predicts seven critical drug toxicity endpoints—hERG cardiotoxicity, hepatotoxicity (DILI), nephrotoxicity (DIKI), Ames mutagenicity, skin sensitization, cytotoxicity, and reproductive toxicity (exploratory)—from molecular structures using curated datasets totaling over 18,000 compounds. The platform employs optimized classical machine learning models with systematic benchmarking of 2D topological descriptors (2240 features), enhanced multi-conformer 3D descriptors (1975 features from 5-conformer ensembles incorporating AUTOCORR3D, RDF, WHIM, and pharmacophore fingerprints), and hybrid representations. Under the tested conditions, 2D descriptors achieved the highest classification performance (AUC-ROC 0.859 ± 0.02), while enhanced 3D descriptors substantially narrowed the previously reported gap (AUC-ROC 0.833 ± 0.03 versus 0.69–0.73 for basic 14-feature 3D). Scaffold-based splitting provided rigorous generalization assessment, with an average performance reduction of approximately 8%. A multi-task learning framework via stacked generalization demonstrated cross-endpoint information sharing improves performance for 5 of 6 endpoints (average +2.1% AUC). The platform integrates leverage-based applicability domain assessment (31–100% coverage), SHAP-based feature importance analysis, and a confidence-weighted multi-endpoint risk scoring system validated on known drugs (AUC = 0.83, p = 4.06 × 10−14, Cliff’s δ = 0.66), with sensitivity analysis confirming robustness across five weight configurations (AUC range 0.72–0.98). External validation on independent benchmark datasets revealed the challenge of cross-dataset domain shift in computational toxicology. MultiEndpointTox is deployed as a production-ready REST API and publicly available at https://github.com/sharhabileltahir/MultiEndpointTox.