Scientific Credibility

Validation Hub

Internal and external validation, kept strictly separate, with R² (hERG's regression metric) never combined with the six classification endpoints' AUC in the same chart.

1. Internal Validation — Scaffold Split

Primary methodology: scaffold-split GroupKFold (k=5), Murcko generic scaffolds — the harder, more realistic estimate (see §2 below for why it matters).

Classification Endpoints (AUC)

hERG Cardiotoxicity — regression, reported separately

hERG is a pIC50 regression model — R², not AUC, and never plotted alongside the classification endpoints above. Scaffold-split R² = 0.38.

Integrated Risk Score

The aggregated score itself has its own reported validation: AUC = 0.81 on a withdrawn/safe compound set.

Reproductive toxicity is excluded from every comparison above — n=127, scaffold AUC=0.588 (near-random). It is not compared against the other six endpoints anywhere on this site; see its model card for the full reasoning.

2. Why Scaffold Split, Not Random Split

Random splits let a model exploit scaffold-level similarity between train and test, inflating performance. These are the two endpoints where the backend documents both numbers.

3. External Validation

Per the model card's TRIPOD-AI checklist (item T14): partial. Reported as status, not invented results.

DatasetPurposeStatus
ClinToxExternal classification benchmark (FDA-approved vs. withdrawn-for-toxicity)In progress
Tox21External multi-assay toxicity benchmarkIn progress
Temporal splitTime-split validation (train on older compounds, test on newer)Methodology implemented
Train/test overlap checkScaffold & exact-structure leakage audit between splitsImplemented

4. Case Studies

Live predictions on 6 real, well-characterized drugs — not cached screenshots.

Explore case studies