Approximately 90% of drug candidates entering clinical trials fail, primarily due to inadequate pharmacokinetics and unacceptable toxicity. The 2025 OpenADMET blind challenge had 65+ teams submit predictions before experimental results were revealed. Deep learning beat classical methods for ADME prediction; classical methods remained competitive for potency.
The AI-PBPK Platform
Wang et al. at Macau University published an AI-PBPK platform predicting eight molecular properties from structure alone and feeding them into a physiologically-based PK model. Validated against 677 human PK datasets, most AUC predictions fell within 2-3x fold error of experimental data, acceptable for early-stage decision-making.
Why the Blind Format Matters
The design of the OpenADMET challenge is the reason its result carries weight. Teams committed predictions before experimental values were revealed, which removes the possibility of tuning a model against the answer. That distinction separates this result from most published benchmark claims in computational biology, where evaluation happens after the ground truth is available and contamination is difficult to rule out.
How much that matters is visible in fields that lack the blind format. In transcriptomics, seven independent benchmarks found single-cell foundation models losing to simple linear baselines, and no evaluated model publishes a cell-level training manifest, so contamination has to be inferred from platform inclusion rules rather than measured directly. One benchmark team caught a train-test leak in its own run only during pre-submission audit: a misconfigured filter silently matched zero rows, leaving every held-out test case in the training pool and inflating the score from 0.402 to 0.949. A blind challenge makes that class of error structurally impossible, which is why the ADMET result should be read as stronger evidence than a comparable number from a retrospective benchmark.
Where AI ADMET Fails
Idiosyncratic toxicity reactions depend on individual immune variability that cannot be predicted from molecular structure. Potency prediction still favors classical methods for many targets.
Related coverage: Generative AI for Small Molecule Drug Discovery | AlphaFold 3 in Drug Discovery: Where It Works and Where It Fails | Single-Cell Foundation Models Keep Losing to Linear Baselines
Primary sources: Fischer Y et al., J Chem Inf Model 2025;65(24); Wang W et al., Clin Pharmacol Ther 2025;118(4).