A team of scientists from the University of Florida College of Pharmacy placed second in the OpenADMET PXR Blind Challenge, a global competition focused on advancing computational and artificial intelligence models to accelerate drug discovery and development. UF’s AI model outperformed more than 350 submissions and ranked No. 1 among all entries developed exclusively using publicly available data.
The OpenADMET competition is a series of open, community-driven artificial intelligence and computational challenges designed to benchmark and improve the predictive models using realistic drug discovery datasets. These include datasets focused on ADMET properties — absorption, distribution, metabolism, excretion and toxicity — which affect how medicines act in the body.
The PXR Blind Challenge focused on predicting whether drug candidates induce the pregnane X receptor, or PXR, a key regulator of drug-metabolizing enzymes and transporters. PXR detects foreign compounds and activates biological pathways that promote their metabolism and elimination, thereby negatively impacting drug efficacy and safety. Because unwanted PXR induction is often identified late in the drug development process, it can lead to costly delays and setbacks. Developing AI models that can accurately predict PXR activation earlier in drug discovery could help researchers identify safer drug candidates and streamline the development of new medicines.
“In the challenge, participants were asked to predict the PXR induction activities of previously unseen compounds before experimental measurements were disclosed,” said Yanjun Li, Ph.D., an assistant professor of medicinal chemistry in the UF College of Pharmacy and a member of the UF AI-Driven Drug Discovery team that participated in the competition. “The prospective blind evaluation provides a rigorous opportunity to benchmark AI and computational models under conditions that closely resemble real-world drug discovery, offering a more realistic assessment of their ability to generalize to novel chemical space.”
The OpenADMET PXR Blind Challenge was built around a large experimental dataset of more than 11,000 compounds measured using a high-fidelity assay to determine whether the compounds activated PXR.
The UF AI-Driven Drug Discovery team developed a novel AI-native strategy that combines agentic AI with human scientific expertise. Using only publicly available data, the team used multiple advanced AI and machine learning approaches to analyze the chemical structures of compounds and predict their potential to induce PXR. Human experts remained closely involved throughout the process, evaluating the AI-generated hypotheses, refining the modeling choices, validating results and guiding final decisions. This combination of AI-driven exploration and expert oversight created an efficient and rigorous approach to addressing real-world drug discovery challenges.
“Our success in the challenge was the result of a combination of strong multidisciplinary collaborations, advanced and complementary modeling strategies and rigorous scientific validation,” Li said. “Our team brought together expertise in artificial intelligence, computational chemistry, medicinal chemistry, structural biology and drug discovery, which allowed us to evaluate the problem from both computation and experimental perspectives. This breadth was especially important for the PXR Blind Challenge because prospective activity prediction requires not only machine learning capabilities, but also a deep understanding of molecular structure, assay biology, chemical data quality and the practical constraints of real-world drug discovery.”
The OpenADMET PXR Blind Challenge gave the UF AI-Driven Drug Discovery team an opportunity to put its AI modeling approaches to the test against some of the world’s leading computational drug discovery experts. The competition attracted teams from academia, biotechnology and pharmaceutical companies and global AI organizations, creating a highly competitive environment for advancing new approaches to drug discovery.
UF’s model performed only marginally below and was statistically indistinguishable from the top-ranked model, developed by a private company that attributed its performance to training on one of the industry’s largest and most diverse proprietary datasets.
“Our results demonstrate that highly competitive prospective performance can be achieved using exclusively publicly available data,” Li said. “This is encouraging for academic laboratories, smaller biotechnology companies and open-science initiatives that lack access to large propriety datasets. It highlights the value of openly releasing challenge data, evaluation protocols, models and methodology reports so the broader community can learn from both successful and unsuccessful approaches.”
Members of the UF AI-Driven Drug Discovery team included Li; Chenglong Li, Ph.D., chair of medicinal chemistry and the Nicholas Bodor Professor in Drug Discovery; Gustavo Seabra, Ph.D., research associate professor; and Jinhang Wei, a graduate student in the Department of Medicinal Chemistry.
This work is part of the team’s broader AI-driven drug discovery program, supported by the National Institutes of Health, the UF Health Cancer Institute and the UF-NVIDIA Artificial Intelligence and Complex Computational Research Award, to develop advanced AI models for drug discovery and development.