The results of the openADMET competition suggest that better data may be more important than bigger AI models when predicting how the body will metabolize drug candidates. Inductive Bio took the top prize, but a 28-way statistical tie among entrants revealed that data quality — not model size — drove the best predictions.
The PXR problem
The competition focused on predicting whether a drug candidate would activate the pregnane X receptor, or PXR. When activated, PXR ramps up production of an enzyme that breaks down roughly 50% of all marketed drugs, causing candidates to exit the body too fast or creating dangerous drug-drug interactions. Most drug developers only discover this problem late in development, forcing costly do-overs.
Lessons for drug development
The results echo findings from other AI domains: adding more computational power and parameters offers diminishing returns compared with curating high-quality, diverse training data. For biotech companies racing to apply AI to drug discovery, the competition underscores that data infrastructure and experimental design remain as critical as model sophistication.
