Toxicity Prediction Tools Do Not Generalize to Novel Chemistry

On a curated panel of 569 approved drugs, no toxicity prediction model—across 35 readouts—generalizes to novel chemistry, and only one toxicity-trained tool performs above chance.

• Why it matters: The field assumes molecular structure alone contains enough signal to forecast adverse effects in human patients, but this assumption is rarely tested directly against real-world drug labels.
• What they did: The study benchmarks fingerprint-based QSAR, fine-tuned foundation encoders, and toxicity-targeted tools using labels defined by black-box warnings or market withdrawals, plus comparisons to an approval-year-only predictor.
• The result: No learned representation beats a plain fingerprint, DILI-specific tools do not reliably exceed chance, and any external successes track largely with approval year—implying structure-only inputs are limiting.

AI summary of the abstract · may contain errors · not medical advice · read the paper.