Automating Decision-Making in Drug Discovery with First-Principles Graph Machine Learning
Drug discovery is slow, costly, and prone to failures, partly because of human decision-making processes whose (sub)optimality cannot be quantitatively assessed, let alone improved. Computational methods are here to help, but physics-driven (structure-based) models typically rely on force fields that are either fast or accurate, but never both; data-driven (ligand-based) models usually employ machine learning backbones borrowed from other fields, exhibiting challenges when facing small, noisy, and graph-structured data ubiquitous in drug discovery.
To make drug discovery autonomous by 2050, my research program celebrates the synergy between physics-based modeling and machine learning. We develop deep learning models to improve force fields and sampling methods, and conversely, en-code physicochemical inductive biases to guide the design of machine learning archi-tectures of, by, and for chemists. Along the journey, such innovations provide new in-sights into key biomolecular processes, such as protein-ligand binding. Finally, these understandings are unified by a foundation model and a Bayesian active learning module to iteratively design the next experiment, leading us one step closer to life-saving therapeutics.
Hosted by Professor Sagar Khare