All model cards

Hepatotoxicity

classification · RandomForest (classification)

Purpose

Predicts drug-induced liver injury (DILI) / hepatotoxicity potential.

Prediction Output

Binary classification (hepatotoxic / non-hepatotoxic) + probability.

Input Requirements

A single valid SMILES string, standardized server-side before featurization.

Training Methodology

Algorithm: RandomForest (classification) · 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).

  • Sources: ChEMBL v34 + DILIrank (Chen et al., 2016). N = 1,511 compounds.
  • Train/test: 1,057 / 303 (scaffold split). Class balance ~60% hepatotoxic before SMOTE.
  • SMOTE oversampling applied inside each training fold only — never on the test fold.

Validation Methodology

  • Scaffold-split GroupKFold (k=5).
  • Decision threshold calibrated to 0.717 by maximizing the Youden index on the scaffold-split validation set, constrained to specificity ≥ 20%.

Scaffold-split AUC = 0.791. Sensitivity 0.848, specificity 0.238, MCC 0.131 (N test = 303).

Known Limitations

  • Specificity is deliberately low (23.8%) — the calibrated threshold favors catching true hepatotoxins (sensitivity) over avoiding false alarms, since missed hepatotoxins are considered more dangerous in early-stage safety screening.
  • Applicability domain coverage on the training set: 78.9%.

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.