MultiEndpointTox 2.1.2: Multi-Endpoint Toxicity Prediction Platform for Drug Safety Assessment with OpenTox 2.0-Aligned Workflows

Abstract

MultiEndpointTox 2.1.2 (MET2.0) is an open-source, machine-learning-powered REST API that predicts seven drug toxicity endpoints from a single SMILES input: hERG cardiotoxicity — regression (R² = 0.621 hold-out / R² = 0.378 scaffold-split) Drug-induced liver injury (DILI) — AUC = 0.791 (scaffold-split), calibrated threshold = 0.717 Nephrotoxicity — AUC = 0.868 (scaffold-split) Ames mutagenicity — AUC = 0.852 (scaffold-split) Skin sensitization — AUC = 0.800 (scaffold-split) Cytotoxicity — AUC = 0.916 (scaffold-split) Reproductive toxicity — AUC = 0.929 (hold-out) / AUC = 0.588 (scaffold-split); excluded from integrated risk score due to limited training data (n = 127) All classification metrics above are reported under scaffold-based GroupKFold cross-validation (Bemis–Murcko generic scaffolds, k = 5), providing realistic out-of-distribution generalization estimates. A confidence-weighted integrated risk score (AUC = 0.81; Mann–Whitney p = 4.06 × 10⁻¹⁴) aggregates endpoint-level predictions using applicability-domain (AD) confidence as weights. Every prediction includes SHAP feature attribution, structural alert detection, AD assessment via leverage, and per-endpoint risk tiers. Architecture Feature pipeline: 2,048-bit ECFP4 (Morgan, radius = 2) + 167 MACCS keys + 25 RDKit physicochemical descriptors → 2,240 raw features → 500 selected via VarianceThreshold (> 0.01) + Pearson correlation filtering (|r| < 0.95). Per-endpoint models are individually optimized from Random Forest, XGBoost, LightGBM, and SVM/SVR using Optuna TPE (100 trials per algorithm–endpoint combination). Post-hoc probability calibration (Youden's J threshold) is applied to all classification endpoints. Deployed via FastAPI + Uvicorn in a non-root Docker container with OpenAPI 3.0 documentation at /docs. Statistical validation (v2.1.0+) DeLong's test confirmed statistically significant AUC superiority over ADMETlab 3.0 on a shared 284-compound benchmark: hepatotoxicity (MET2.0 AUC = 0.904 vs ADMETlab3 AUC = 0.682, p = 0.016), hERG (0.937 vs 0.749, p = 0.010), and nephrotoxicity (0.965 vs 0.704, p < 0.0001). Bootstrap 95% confidence intervals computed with n = 2,000 resamples. In an FDA-withdrawn drug case study (n = 30), MET2.0 correctly flagged 11 of 20 withdrawn drugs for their specific withdrawal-reason endpoint (sensitivity = 0.55), whereas ProTox-3.0, ADMETlab 3.0, and pkCSM flagged none. Additional capabilities SHAP-based feature attribution (TreeExplainer / KernelExplainer) Structural alert detection for known toxicophore patterns SyGMa-based metabolite screening (Phase I/II biotransformations) Pharmacogenomic Sensitivity Index (PSI) across CYP2D6, CYP3A4, CYP2C9 phenotypes AutoDock Vina molecular docking against six protein targets (hERG, CYP3A4, CYP2D6, CYP2C9, ERα, AR) Three-dimensional molecular descriptors FAIR-compliant REST API aligned with OpenTox 2.0 principles v2.1.2 audit fixes (2026-04-15) Integrated risk score now uses confidence-weighted aggregation (AD confidence × endpoint risk score) with reproductive toxicity explicitly excluded SMOTE data leakage in Optuna hyperparameter optimization corrected — moved to per-trial ImbPipeline inside CV folds FastAPI router refactored for maintainability 13 new unit tests for integrated risk score logic XGBoost deprecated parameter removed Scientific validity scorecard included Author name corrected to full legal name across all metadata files Performance metrics reconciled with scaffold-split values throughout Known limitations hERG scaffold gap: R² drops 39% from hold-out (0.621) to scaffold-split (0.378); extrapolation to structurally novel scaffolds is unreliable. Reproductive toxicity underpowered: n = 127 training compounds; scaffold-split AUC = 0.588 (near-random). Excluded from integrated risk score. Hepatotoxicity class imbalance: 86% positive training labels; specificity remains 34.5% even after threshold calibration (MCC = 0.330). AD is leverage-based only; no consensus AD implemented yet. Not for regulatory use: Predictions do not meet ICH/FDA/EMA requirements as standalone evidence.