All model cards
hERG Cardiotoxicity
regression · XGBoost (regression)
Purpose
Predicts hERG (human Ether-à-go-go-Related Gene) potassium channel inhibition — a proxy for drug-induced QT prolongation and cardiac arrhythmia risk.
Prediction Output
pIC50 (regression) — higher values indicate stronger predicted hERG inhibition.
Input Requirements
A single valid SMILES string, standardized (canonicalized) server-side before featurization.
Training Methodology
Algorithm: XGBoost (regression) · Optuna, TPE sampler, 100 trials, random_state=42
Feature representation: 2,048-bit Morgan ECFP4 (radius=2) + 167 MACCS keys + 25 RDKit physicochemical descriptors → 2,240 raw features → variance threshold (>0.01) → Pearson correlation filter (|r|<0.95) → top 500 features → StandardScaler (fit on the training fold only).
- Source: ChEMBL v34, target CHEMBL240. N = 10,661 compounds.
- Activity values (nM) converted to pIC50 = −log₁₀(IC50 / 1e9).
- Duplicates removed by canonical SMILES; median pIC50 kept per compound.
Validation Methodology
- Primary: scaffold-split GroupKFold (k=5), Murcko generic scaffolds, random_state=42.
- Comparison: random hold-out split (same data, no scaffold grouping).
Scaffold-split R² = 0.378 (primary, reported). Random-split R² = 0.621 (comparison only — see Known Limitations).
Known Limitations
- 39% R² drop from random-split (0.621) to scaffold-split (0.378) — a documented generalization gap. Treat predictions for structurally novel scaffolds with caution.
- Stereochemistry is removed during SMILES standardization (ADR-002) — enantiomers/diastereomers receive identical predictions.
- A D-MPNN (Chemprop) replacement is roadmapped, conditional on achieving ΔR² ≥ 0.10 (ADR-003) — not yet implemented.
Applicability Domain
Leverage-based applicability domain: h* = 3×(p+1)/n (Williams plot). in_domain=true → confidence 1.0; in_domain=false → confidence 0.0, and the prediction should be treated as exploratory only. Supplementary checks: Tanimoto k-NN distance > 0.6 or descriptor values outside [min−3σ, max+3σ].
Interpretability Support
SHAP (SHapley Additive exPlanations) feature attribution is available via /predict/interpret and /predict/integrated, returning per-feature contribution values and a global/local breakdown by feature category (physicochemical / structural fingerprints / substructure keys).
Research-Use Disclaimer
Research use only. Not validated for regulatory submissions or clinical decision-making. Predictions must not be the sole basis for safety decisions without independent experimental validation.