BRIDGING ARTIFICIAL INTELLIGENCE AND THERMODYNAMICS FOR ADVANCING CANCER IMMUNOTHERAPY AND DRUG DISCOVERY

Abstract

Protein function is often discussed from the perspective of a single representative structure. This static view has led to the development of experimental structure determination techniques, including X-ray crystallography, NMR spectroscopy, and cryo-EM, and has also shaped the development of the highly accurate structure-prediction methods such as AlphaFold2 (AF2). However, proteins are dynamic molecules, and their biological functions are governed by conformational ensembles composed of interconverting metastable states. This ensemble perspective is especially important in drug discovery and immunotherapy, where ligand binding, receptor engagement, and downstream signaling can depend on those rare or transient conformations that are not apparent from a single structure. Therefore, a key challenge is to develop computational methods that can start from sequence, generate relevant conformational diversity, assign thermodynamic weights to the generated states, and connect these ensembles to biological function. To address this, in this thesis, I develop and apply computational approaches that bridge Artificial Intelligence(AI)-based structure prediction with physics-based sampling and reweighting. First, I investigated how stochasticity can be introduced into AF2 through reduced multiple sequence alignment(MSA) approach, showing that AF2-generated ensembles can provide useful structural hypotheses but do not directly encode thermodynamic or Boltzmann weights. To this end, I developed AlphaFold2-RAVE (AF2-RAVE), a framework that combines AF2-derived structural diversity with AI-augmented enhanced sampling molecular dynamics to obtain thermodynamically ranked conformational ensembles. I apply AF2-RAVE to systems involving rotameric metastability, missense mutations, G-protein-coupled receptor motions, kinase DFG-loop and activation-loop conformational changes, and calcium-bound S100 homodimers. Together, these applications demonstrate that the post hoc integration of physics-based methods with generative-AI methods provides a framework that leverages the strengths of both approaches. Specifically, I show that AF2-RAVE type methods can use the conformational selection hypothesis to generate druggable metastable states for structure-based drug discovery in a computationally efficient manner. I then extend the ensemble perspective to immunotherapy. I apply AF2-RAVE-based ideas to evaluate T-cell receptor - peptideMHC complexes generated by cofolding models. I also extend this approach to model the synaptic cleft formed by CAR T cells and develop a theory for CAR T-cell responsiveness based on the canonical T-cell immunological synapse. Finally, I develop a protocol that fits AI-generated structural ensembles to the Langevin Equation for Protein Dynamics (LE4PD), allowing residue-level relaxation timescales to be inferred from AF2 and other generative AI methods. Overall, this thesis develops a "sequence--ensemble--function" framework for biomolecular modeling. By combining generative AI, molecular dynamics, enhanced sampling, and coarse-grained dynamical models, the work presented here moves beyond static structure prediction toward a more complete description of protein function. This framework led to the development and application of several methods, including reduced-MSA approach for AF2, AF2-RAVE and its variants, CAR-BOOST, and an LE4PD-based approach for assigning dynamical information to AI-generated structural ensembles.