MultiEndpointTox 2.1.0: Regulatory-Grade Multi-Endpoint Toxicity Prediction API

Abstract

MultiEndpointTox (MET2.0) is an open-source machine-learning REST API for predicting 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=0.847, calibrated threshold=0.717 • Nephrotoxicity — AUC=0.823 • Ames mutagenicity — AUC=0.891 • Skin sensitization — AUC=0.876 • Cytotoxicity — AUC=0.834 • Reproductive/developmental toxicity — AUC=0.723 (excluded from integrated risk score; scaffold-split AUC=0.588, n=127 underpowered) Every prediction includes SHAP feature attribution, structural alert detection, applicability domain (AD) assessment via leverage, and an integrated risk score across six endpoints. ─── What is new in v2.1.0 ─────────────────────────────── • 35+ automated tests across four modules (API, predictions, interpretability, edge cases) • FAIR compliance documentation (docs/FAIR_compliance.md) • Reproducibility guide — scaffold-split algorithm, calibration procedure, feature pipeline (docs/reproducibility.md) • Model card in Mitchell et al. (2019) format (docs/model_card.md) • Pinned requirements.txt and requirements-dev.txt • Three new Pydantic response models: ApplicabilityDomainResult, SHAPAttribution, IntegratedRiskScore • Machine-readable reproductive toxicity exclusion reason embedded in /predict/integrated schema • GitHub Actions CI: mypy, ruff, black, isort, safety, and Docker smoke-test workflows • DOI-stamped /metadata endpoint (10.5281/zenodo.19478366) ─── Architecture ──────────────────────────────────────── Feature pipeline: 2,048-bit ECFP4 + 167 MACCS keys + 25 RDKit descriptors → 2,240 raw features → 500 selected (VarianceThreshold + SelectKBest χ²). Ensemble: XGBoost + LightGBM + calibrated logistic regression. Deployed via FastAPI + Uvicorn in a non-root Docker container with OpenAPI 3.0 docs at /docs. ─── Known Limitations ─────────────────────────────────── 1. hERG scaffold gap: 39% R² drop (hold-out 0.621 → scaffold 0.378); extrapolation to novel scaffolds unreliable. 2. Reproductive toxicity dataset underpowered (n=127); excluded from integrated risk score until ≥1,000 compounds. 3. Applicability domain is leverage-based only; no consensus AD implemented yet. 4. No prospective validation against ClinicalTrials.gov clinical outcomes.