CAREER: Learning to learn - Artificial Intelligence Augmented Chemistry for Molecular Simulations and Beyond

点击翻译按钮获取中文摘要

Dr. Pratyush Tiwary of University of Maryland, College Park, is supported by an award from the Chemical Theory, Models and Computational Methods program in the Division of Chemistry to develop simulation algorithms at the interface of statistical mechanics and artificial intelligence (AI) for the study of rare events. The synergistic use of statistical mechanics and AI enables the automatic, human bias-free modeling of very slow processes in chemistry and biochemistry that unfold across many time and length scales. The tools Dr. Tiwary and his team are developing are to be incorporated into efficient open-source computational platforms for widespread use by the scientific community. One of many applications he is pursuing is the quantification of time spent by small molecules inside biological hosts, a property fundamental to the chemistry of life processes, yet very hard to calculate through experiments or simulations. The results of Tiwary’s modeling will be compared against experimental investigations by his partners at Stony Brook University and the National Cancer Institute. Under this award, Dr. Tiwary will also be developing platforms to introduce coding and AI to high school/college students and educators in physical sciences through workshops and online tutorials, providing the workforce of the next generation with transferable skills for today's job markets. These efforts are being carried out through collaborations with Prince George’s Community College, Bowie State University and through virtual means with other partners across the country.Pratyush Tiwary’s research seeks to develop the next generation of ultra-long timescale molecular dynamics (MD) simulation methods by integrating AI with statistical mechanics through a “learning to learn” framework. This framework uses AI to learn the reaction coordinate (RC) characterizing a generic molecular system, interpreting it as a past-future information bottleneck. The knowledge of the RC is used through biased sampling methods to systematically sample more of the configuration space and thereby generate more relevant data to train AI. Furthermore, the use of statistical mechanics helps AI in different ways, by (i) making “black box” AI techniques more transparent, and (ii) dealing with the problem of poor training data from which AI can produce misleading results. The iteration between AI and MD continues till the RC converges, leading to estimates of thermodynamic constants, rates and rate-limiting steps in one shot. These methods will be applied to the study of protein-ligand and riboswitch-ligand interactions, both of which are central to life processes, yet poorly understood. These long timescale all-atom simulation methods are designed to unravel the complexity and richness that arises from the interplay between different degrees of freedom in these and other generic molecular systems. Finally, this research program also strives to demonstrate how mixing statistical mechanics and AI can lead to interpretable and trustworthy use of AI, thereby increasing confidence in large-scale deployments of AI across chemistry.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.