From protein structures to effective drugs via machine learning and molecular simulation

Project Summary Recent years have seen dramatic breakthroughs in methods for determining structures of proteins and other drug targets experimentally (e.g., cryo-EM) and predicting them computationally (e.g., AlphaFold). These advances tremendously increase the potential of structure-based drug discovery, but going from an accurate target structure to an effective drug remains a difficult problem. To do so, once must design molecules that bind tightly and selectively to the target, have appropriate (drug-like) physiochemical properties, and exert a desired effect on the target’s function. This project will make advances in several areas critical to enabling the efficient structure-based design of effective, safe drugs: 1. Creation of machine learning methods to accurately predict a chemical compound’s binding affinity, drawing simultaneously on the target’s structure and experimentally determined affinities of other compounds that do or do not bind the target. 2. Development of machine learning methods to design drug-like ligands that bind tightly and selectively, exploiting both generative artificial intelligence and iteration between computation and experiments. 3. Determination of the mechanisms by which drugs can selectively stimulate or selectively block arrestin signaling at diverse GPCRs, achieving desired therapeutic effects without dangerous side effects. This research builds on recent achievements of the Principal Investigator, Ron Dror, in two areas: (a) machine learning methods for structural biology, and (b) elucidation of the structural basis of GPCR signaling. It continues his record of collaborating with experimentalists to do innovative computational work with a strong impact on experimental biochemistry, structural biology, pharmacology, and drug design.