The OpenADMET drug metabolism AI competition produced a surprise result: bigger models are not always better. Inductive Bio took the top spot, but a 28-way statistical dead heat among very different approaches suggests that pharmaceutical companies may need to rethink how they evaluate AI for drug development.
OpenADMET is a blind competition that challenges AI teams to predict how drug compounds are absorbed, metabolized, and excreted by the human body — factors that determine whether a promising molecule becomes a viable medicine. Poor ADMET (absorption, distribution, metabolism, excretion, toxicity) profiles account for roughly 40% of clinical trial failures.
The close finish between large foundation models and smaller, task-specific architectures carries a practical implication for biotech and pharma R&D leaders: throwing more compute at drug metabolism prediction does not guarantee better results. Domain-specific training data and thoughtful model design may matter more than raw parameter count.
The results arrive as pharmaceutical companies and CDMOs race to deploy AI across their pipelines. The OpenADMET outcome suggests that the winning strategy may not be to license the largest available model but to carefully match model architecture to the specific prediction task.
