Drug-safety models fail on new chemistry, so labs are rethinking what counts as evidence

The benchmark results are uncomfortable

Drug discovery leans heavily on computational models that predict whether a molecule will be toxic before anyone synthesizes it or tests it in cells. A preprint benchmarking 35 toxicity models, spanning quantitative structure-activity models, foundation encoders, and specialized tools, against a curated panel of 569 approved drugs, reports that this reliance is shakier than the field's leaderboards suggest. No learned representation outperformed a plain fingerprint baseline, and approval year alone was a better predictor of toxicity than the specialized tools. Tools trained specifically for drug-induced liver injury did not reliably beat chance on the curated panel. m:…4768

The reported pattern is stark. No learned representation outperformed a plain fingerprint for toxicity prediction, and approval year alone was a better predictor of toxicity than the specialized toxicity tools. Tools trained specifically for drug-induced liver injury did not reliably beat chance on the curated panel. Those three results are the preprint's reported comparisons on the curated panel, and they are the full basis for this section's reading of the benchmark. m:…4768

The same pattern inside ADMET prediction

Two more preprints find the same fragility in ADMET prediction, the related task of forecasting absorption, distribution, metabolism, excretion, and toxicity properties that determine whether a candidate survives early screening. A benchmark comparing active-learning acquisition strategies and two uncertainty estimation methods across regression endpoints found that active-learning benefits were endpoint-dependent, with solubility showing the most consistent gains. Deep Ensembles achieved lower error than MC Dropout in 45 of 56 comparisons, yet that predictive edge did not consistently translate into better acquisition efficiency. The authors attribute the failures to weak alignment between uncertainty estimates and actual errors, and to reduced coverage of chemical types in the batches the models selected. m:…1055

A second benchmark, evaluating graph neural networks, pretrained models, and fingerprint baselines on six Therapeutics Data Commons endpoints under a strictly controlled, leakage-free protocol with matched tuning budgets, found that no single model family dominated across all endpoints. The graph networks' advantage on Caco-2 permeability came from auxiliary physicochemical descriptors rather than from the graph representation itself, and Morgan fingerprints were competitive on hERG and superior for clearance endpoints. The authors also show that evaluation habits inflate reported advantages: reporting single best-trial runs, or choosing optimizers based on test-set performance, significantly exaggerated model superiority, while hyperparameter optimization did not consistently beat random search at a 50-trial budget. m:…0288

What replaces the model leaderboard

The fourth preprint offers a different kind of evidence altogether: a wet-lab measurement instead of a better leaderboard score. Its high-throughput screening platform uses human induced pluripotent stem cells, with four distinct models covering the three germ layers, to identify teratogenic hazards early in drug discovery, as a scalable and automated alternative to traditional animal-based testing. Across a reference set of 146 compounds, the platform achieved 83% overall accuracy and identified 88% of known teratogens, using an automated high-content imaging workflow that supports dose-response testing at scale. m:…1109

Taken together, the four preprints suggest a reordering of priorities for early-stage screening. This suggests model rankings built on loose evaluation protocols are unreliable guides: reporting single best-trial runs or selecting optimizers based on test-set performance significantly inflates model advantage, and no single model family dominated across all six Therapeutics Data Commons endpoints under fair evaluation. This suggests the more useful evidence is measurement that tracks its own errors and results that hold under controlled evaluation — weak uncertainty-error alignment and reduced chemotype coverage were identified as drivers of active-learning failure, while a human stem-cell platform identified teratogens with 83% overall accuracy and 88% sensitivity across a reference set of 146 compounds. Those are reported results on the tested panels, not a guarantee for novel chemistry beyond them. m:…0288 m:…1055 m:…1109