Title: Accurate and fast virtual drug screening with atomic embeddings
Abstract: Virtual drug screening searches databases of billions of candidate compounds for the few that may bind a target protein. Physics-based methods such as docking show limited efficacy in predicting binding affinity; just as importantly, they cost seconds to minutes per compound, so they cannot be run across a library that size.
I will describe a retrieval approach that changes the shape of the computation. We train a model to place every atom (in the protein and in every candidate drug) into a shared high-dimensional space, where atoms that interact favorably are placed near one another. Binding affinity then reduces to a sum over nearby atom pairs in this high dimensional embedding space and screening a billion-compound library becomes a local search over embeddings that we computed once. Our model, Simpatico, achieves state-of-the-art enrichment, meaning it ranks known binders near the top of the list more reliably than competing methods, and it scans millions of compounds per minute on a single L40S GPU. It trains in a few hours on orders of magnitude less data than competitor models.
Simpatico and competing tools see each protein as a single static structure. The problem: proteins move, and that movement plays a role in (and is impacted by) binding interactions. I will discuss our efforts to gather massive new data sets of protein dynamics data to train the next generation of Deep Learning methods for virtual drug screening that accounts for dynamics.
Bio: Travis Wheeler is a Professor in the University of Arizona Department of Pharmacy Practice and Science. He earned his bachelors in Evolutionary Biology from the University of Arizona, then his PhD in Computer Science from UArizona (2009). He spent 5 years as a postdoc in the research group of Sean Eddy at HHMI Janelia Research Campus, then joined the Computer Science faculty at the University of Montana in 2014, where he remained until his recent move back to Arizona in 2022. Dr. Wheeler leads a large group whose research can be broadly described as “algorithms and machine learning approaches for computational biology”, primarily focused on applications to genomics, drug discovery, and animal behavior classification.