A new AI drug discovery competition suggests that high-quality data may be more valuable than simply building larger AI models. As reported by STAT, the OpenADMET PXR Blind Challenge tested whether AI systems could predict how experimental drug compounds activate the pregnane X receptor (PXR)—a key protein involved in drug metabolism. Correctly predicting PXR activation is critical because it can determine whether a drug is broken down too quickly, interacts with other medicines, or fails during development. The competition attracted more than 350 participants from pharmaceutical companies, biotech firms, universities, and AI organizations.
The winning entry came from Inductive Bio, whose Beacon AI model achieved the highest performance. However, one of the competition's most striking findings was that 28 different submissions were statistically tied, despite using different AI architectures and techniques. Researchers said this indicates that the field may be approaching the limits of what current datasets can reveal. Rather than simply scaling up model size, future improvements are likely to come from collecting richer, more diverse, and higher-quality biological data.
The article explains that ADMET—Absorption, Distribution, Metabolism, Excretion, and Toxicity—remains one of the biggest challenges in drug development. Many promising drug candidates fail in late-stage clinical development because of unexpected metabolism or safety issues, costing companies years of research and billions of dollars. AI models that accurately predict these properties early in the discovery process can help scientists eliminate weak candidates sooner, reduce laboratory testing, shorten development timelines, and improve the chances of bringing safer medicines to patients.
The article concludes that the future of AI-driven drug discovery may depend less on building ever-larger foundation models and more on developing better datasets, rigorous benchmarking, and real-world validation. Blind competitions such as OpenADMET, where models are evaluated on previously unseen compounds, provide a more realistic measure of AI performance than conventional benchmarks. As pharmaceutical companies increasingly integrate AI into research, the results suggest that data quality, scientific expertise, and robust evaluation methods will be just as important as advances in AI algorithms themselves.
