Chemistry Seminar with Dr. Pratyush Tiwary from the University of Maryland at 4:00pm

Abstract: Structure prediction tools using Generative Artificial Intelligence (AI) have made great progress in predicting, with the click of a button, the most stable structure given the sequence of generic proteins/RNA, and even an ensemble of structures with associated dynamics. It would appear as if the days of molecular dynamics (MD), development of statistical mechanics and enhanced sampling methods are thus over. In this talk, I will demonstrate why this is not the case, and why MD and especially enhanced sampling will continue to play a fundamental role in making AI methods predictive and trustable for biophysics. I will show how it is possible that the outputs from current AI based structure predictions might sometimes be just memorization or worse, hallucination. I will then show how integrating Generative AI methods with enhanced MD and other statistical mechanics concepts can be a powerful approach, which while not lightning fast, is far more predictive than Generative AI or MD either on their own. The examples I show will likely include kinases and RNA, where we will obtain thermodynamic and dynamic observables such as drug residence times, conformational populations, effect of mutations, and melting curves starting solely from chemical identity and associated force-fields.