Assessing the state of the art in ADMET prediction: Lessons from the OpenADMET-ExpansionRx community blind challenge

Computational blind challenges are essential for advancing drug discovery, especially in research areas with limited public data, such as absorption, distribution, metabolism, excretion, toxicity (ADMET) prediction. In this paper, we present the results of our latest community blind challenge, organized in collaboration with Expansion Therapeutics (ExpansionRx). Participants were tasked with predicting nine distinct ADMET endpoints routinely evaluated during lead optimization of compounds in drug discovery programs. Following unprecedented engagement from the global community, this challenge provided valuable insights into the state of the art in machine learning for ADMET modeling. Evaluating submissions against a realistic, challenging dataset from active ExpansionRx drug discovery programs revealed that message-passing neural networks and chemical foundation models dominated among the top performers. Furthermore, the results highlighted the power of proprietary industry datasets and a critical need to optimize predictive workflows for different endpoints independently via multitasking and task-affinity grouping. We also demonstrate how hosting our open-source blind challenge infrastructure on a flexible platform like Hugging Face was pivotal for driving community engagement and visibility. Ultimately, blind challenges supported by robust infrastructure, widespread dissemination, and rigorous post-challenge analysis are fundamental to maximizing the utility of public data releases. We anticipate that these community benchmarks will continue to drive the advancement of ADMET models, enabling their deployment in both early- and late-stage drug discovery.