Computational approaches for studying enzyme function and protein-ligand interactions: from atomistic simulations to deep learning

DNA and RNA polymerases are essential enzymes involved in diverse biological processes, making a detailed understanding of their molecular mechanisms vital for elucidating disease progression, including cancer. While experimental techniques like X-ray crystallography and steady-state kinetics have provided significant insights, computational methods offer atomic-level precision and dynamic understanding of these processes, often with advantages in cost and detail. This thesis explores the application of molecular dynamics simulations, free energy calculations, and convolutional neural networks to investigate polymerase function and inhibition. First, we investigate how mutations distant from the active site influence nucleotide selection in DNA polymerase β (Pol β). Through molecular dynamics simulations and free energy calculations, we demonstrate that the R283A mutation, located 14.8 Å from the catalytic site, reduces fidelity by disrupting interactions that stabilize correct base pairing geometries. Structural alignment reveals this mechanism is conserved across the polymerase X family. Second, we systematically benchmark alchemical free energy calculations for ribonucleotide discrimination in Pol β and RB69 polymerase, demonstrating that specialized nucleotide force fields like SHAW substantially outperform general-purpose parametrizations (GAFF2), though significant errors persist for challenging nucleotide analogs. These limitations motivate the development of Neural Network Potential/Molecular Mechanics hybrid methods, for which we have curated a specialized training dataset. Finally, we develop a convolutional neural network for fragment-based drug design. Through chemistry-informed data curation and systematic optimization, our model achieves reasonable accuracy across 30 fragment classes while demonstrating robust generalization to novel protein families. Validation on clinically relevant targets reveals both current capabilities and limitations that guide future development. Together, these investigations advance our mechanistic understanding of polymerase fidelity and establish more reliable computational methods for predicting ligand binding and guiding inhibitor design.