Collaborative Research: CDS&E: Adaptive Uncertainty-Driven AI for Drug Discovery: A Conformal Prediction-Simulation Loop Approach

Yihang Wang of Case Western Reserve University, and collaborator Lu Cheng of the University of Illinois Chicago, are supported to develop artificial-intelligence (AI) methods that quantify their own uncertainty and refine themselves through molecular simulation for more reliable computational drug discovery. The discovery of new medicines depends on accurately predicting how candidate molecules bind to the proteins that drive disease, a process known as molecular docking. AI can now perform such predictions across millions of compounds, greatly accelerating the early stages of drug discovery. These models, however, frequently produce incorrect predictions, especially when applied to molecules that differ from those encountered during training. The wrong predictions generate large numbers of false leads that consume valuable laboratory time and resources. To address this challenge, PIs Wang and Cheng, together with their research groups, will develop computational frameworks that report how confident these models are in each prediction and use the predicted confidence level to automatically guide the initiation of detailed physics-based simulations to acquire additional evidence and improve themselves. This work will help determine when an AI model’s prediction can be trusted, when additional computation is needed, and how to use limited computing resources more effectively. By advancing reliable AI for chemistry, the project will support faster and more trustworthy molecular discovery, with potential impact in areas such as drug design and materials discovery. The project will also train students in chemistry, AI, and statistics; develop open-source software and educational materials; and broaden participation in computational chemistry through mentoring and research opportunities. In this collaborative project, Wang and Cheng will establish the theoretical and practical foundations for reliable, uncertainty-aware AI in ligand pose prediction by integrating conformal prediction with physics-based molecular simulation. Building on graph-neural-network (GNN)-based calibrated conformalized quantile regression, the team will develop chemically and structurally adaptive methods that construct atom-level uncertainty intervals for predicted ligand binding poses. These intervals will remain statistically rigorous and chemically meaningful across diverse binding-site environments and novel targets. The uncertainty estimates will guide a closed-loop AI simulation workflow. In regions of high uncertainty, a guided diffusion model will generate diverse initial configurations, which will seed enhanced-sampling simulations to produce new structural ensemble data. These data will enable iterative retraining to reduce false positives and improve generalization. The framework will be applied to virtual screening of ultra-large chemical libraries, where calibrated uncertainty will guide compound prioritization, limit error propagation, and support the discovery of new chemical scaffolds and mechanistic insights into molecular recognition. By unifying machine learning, statistical uncertainty quantification, and computational biophysics, the project will advance self-correcting molecular modeling and contribute to NSF priority areas in AI, with applications in biotechnology and translational science. The algorithms, open-source software, and web-accessible tools developed through this project will be made available to support applications across scientific fields. The project will also integrate results into coursework, support community workshops, and provide research and outreach opportunities for trainees. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.