ARTIFICIAL INTELLIGENCE FOR ACCELERATING AND UNDERSTANDING MOLECULAR SIMULATIONS

Abstract

Computational techniques such as molecular dynamics (MD) simulations offer detailed spatiotemporal insights into biomolecular systems, playing a critical role in uncovering mechanisms and informing therapeutic design. However, a major limitation of MD is its high computational cost, which makes it challenging to study many biophysically relevant processes.In this dissertation, I address this challenge by integrating artificial intelligence (AI) with statistical physics to develop enhanced sampling methods that significantly accelerate MD simulations. These physics-driven, representation learning approaches enable efficient implementation of molecular dynamics in systems that would otherwise be computationally intractable. Moving beyond traditional model systems, I demonstrate the practical utility of these methods in drug discovery, particularly in characterizing the interactions between small molecules and RNA, an emerging class of therapeutic targets. Furthermore, such AI-driven approaches typically operate in data-sparse regimes and it is essential to establish the robustness of the trained models before deploying them. For this purpose, I design an algorithm to validate general-purpose AI models, particularly in the context of MD. Overall, this dissertation demonstrates how the principled integration of AI with physics-based computational methods enables more efficient and insightful molecular simulations.