Speaker: Prof. Xuhui Huang

Host: Prof. JC Gumbart

Title: Physics-Informed Machine Learning for Protein Dynamics: From RNA Polymerase Inhibition to Targeted Protein Degradation

Abstract:

 Protein dynamics are fundamental to protein function and encode complex biomolecular mechanisms. In this talk, I will explore how incorporating dynamics memory (i.e., non-Markovian effects) into machine learning models can greatly improve both the efficiency and accuracy of predicting long-time dynamics in complex biomolecules. Specifically, I will introduce MEMnets, a deep learning framework for identifying the slow collective variables (CVs) of protein dynamics. Unlike conventional deep learning models such as VAMPnets, which assume Markovian dynamics, MEMnets builds on our own integrative generalized master equation (IGME) theory with a novel loss function that minimizes the time integration of memory kernels. We demonstrate that MEMnets-derived CVs elucidate the molecular mechanism underlying the loading gate opening of bacterial RNA polymerase (RNAP), revealing a transiently open cryptic pocket that binds the antibiotic Myx. In addition, we apply these methods to Targeted Protein Degradation (TPD)—an emerging therapeutic strategy that eliminates, rather than inhibits, disease-causing proteins. TPD agents such as PROTACs and molecular glues work by stabilizing weak or transient protein–protein interactions (PPIs) between the target protein and an E3 ubiquitin ligase, thereby marking the target for proteasomal degradation. These metastable PPIs are central to TPD activity but remain notoriously difficult to predict. Using our approach, we predict metastable PPIs between E3 ligases (VHL) and target proteins (KRAS or RIP1 kinase). Remarkably, one of our predicted complexes closely matches an experimentally determined co-crystal structure. Building on this, we use these metastable PPIs to steer AlphaFold3’s generative process for virtual screening of TPD agents. Finally, I will introduce TS-DART, another deep-learning method that automatically identifies transition states (TS) across multiple free energy barriers in biomolecular systems. Inspired by trustworthy AI, TS-DART detects TS as out-of-distribution data in a hyperspherical latent space, providing a robust and automated way to characterize rare events in biomolecular dynamics.