Inductive Bio wins drug metabolism AI contest in 28-way tie

An artificial intelligence competition focused on predicting drug metabolism has produced an unexpected result: the winning model from Inductive Bio shared its victory with 27 other approaches in what amounted to a statistical tie, suggesting that bigger AI models may not hold the advantage many assume in pharmaceutical development.

The OpenADMET competition challenged participants to predict whether drug candidates would activate the pregnane X receptor (PXR), a biological sensor that triggers production of enzymes capable of breaking down approximately 50% of all marketed drugs. When PXR activates in response to a drug molecule, it can cause the medication to exit the body too quickly or create dangerous drug-drug interactions.

Why it matters

Pharmaceutical companies typically discover PXR activation problems late in development, forcing costly returns to early-stage research. An AI system that reliably predicts this interaction could save years of work and millions in development costs. But the 28-way tie suggests the industry's path forward may depend less on sophisticated model architectures and more on access to high-quality training data—a finding with significant implications for how drug developers should invest their AI resources.

The PXR prediction challenge

The pregnane X receptor represents a critical bottleneck in drug development. This protein acts as the body's defense mechanism against foreign molecules, ramping up production of cytochrome P450 enzymes that metabolize and eliminate drugs. Most development programs only identify PXR activation issues during late-stage testing, after substantial time and capital investment.

Predicting PXR interactions earlier could help developers either modify drug candidates to avoid triggering the receptor or account for faster metabolism in dosing strategies. The competition aimed to determine whether current AI approaches could deliver that predictive capability reliably.

Data quality over model complexity

While Inductive Bio technically won the competition, the statistical tie across 28 different approaches points to a broader conclusion about AI in drug discovery. The results suggest that the quality and characteristics of training data may matter more than the sophistication of the underlying model architecture.

This finding contrasts with the trajectory of AI competitions in other domains, where increasingly large and complex models have dominated leaderboards. In pharmaceutical applications, where training data remains limited compared to fields like natural language processing or computer vision, data curation and selection may represent the true competitive advantage.

Beyond protein structure prediction

The competition represents a shift in pharmaceutical AI ambitions. Following AlphaFold's breakthrough in protein structure prediction—which earned a Nobel Prize after its 2020 CASP competition victory—the industry has moved toward more complex challenges that directly impact drug development timelines and success rates.

Predicting drug metabolism, toxicity, and efficacy requires AI systems to model dynamic biological processes rather than static structures. The OpenADMET results suggest these problems may require different approaches than those that succeeded in protein folding.

The details were first reported by STAT News.