Toxicity predictors struggle to beat simple baselines on approved drugs
- Toxicity predictors, from fingerprint QSAR to foundation-model encoders and dedicated tools, were tested on approved drugs flagged by black-box warnings or market withdrawal.
- Learned representations gave no gain over plain fingerprints; most dedicated tools, liver-injury ones included, hovered near chance and trailed a chemistry-free approval-year baseline.
- Preprint, not yet peer reviewed; the authors say exposure and metabolism data are needed.
💡 Points Based on the abstract
If an approval-year baseline with zero chemistry beats a liver-injury model, the labels are mostly encoding era and regulatory history, so the split deserves suspicion before the models do. Even so, near-chance DILI tools on approved drugs are not something to use as a gate. Bigger encoders adding nothing over fingerprints is worth remembering before any GPU spend. There is no prospective test, though.
Toxicity Prediction Tools Do Not Generalize to Novel Chemistry ↗