Graduation Semester and Year
Spring 2026
Language
English
Document Type
Dissertation
Degree Name
Doctor of Philosophy in Chemistry
Department
Chemistry and Biochemistry
First Advisor
Kwangho Nam
Abstract
Molecular dynamics (MD) simulation approaches are crucial for understanding enzyme function at the atomic level. In principle, the most accurate description requires quantum mechanics (QM); however, QM calculations are prohibitively expensive for large systems such as enzymes, which necessitates reliance on classical molecular mechanics (MM) descriptions.
Chapter 1 provides an overview of computational methods, beginning with MM approaches and progressing to multiscale QM/MM frameworks and machine-learning potentials (MLPs). In addition, it introduces the two enzyme systems (adenylate kinase and dihydrofolate reductase) that are studied in Chapters 2 ~ 5.
Chapter 2 combines classical MD with structural biology to reveal key mechanistic insights into how Mg2+ drives a crucial reorganization of the adenylate kinase active site into a reaction-ready arrangement. However, because MM descriptions cannot capture reactive processes, simulations of the chemical step must rely on hybrid approaches such as QM/MM, which explicitly treat electronic rearrangements while keeping the computational cost manageable.
Chapter 3 introduces a new QM/MM interfacer module in the CHARMM-GUI web-platform that enables QM/MM simulations through a user-friendly, step-by-step setup workflow. By streamlining system preparation and input generation, this module lowers the practical barriers to performing QM/MM studies of enzyme reactions.
Even with QM/MM, obtaining reaction thermodynamics typically requires long simulations and repeated evaluations of quantum energies and forces, making such calculations computationally demanding. Chapters 4 and 5 tackle this bottleneck using MLPs. Chapter 4 demonstrates that MLPs can reproduce reaction free energies with sub-kcal/mol accuracy and remain reliable across mutations, ranging from point mutations to cross-species variations, as shown for the hydride transfer reaction catalyzed by multiple variants of dihydrofolate reductase. Building on these results, Chapter 5 presents an optimized, high-throughput implementation in CHARMM program that accelerates simulations by approximately 500-fold compared with conventional ab initio (or density functional theory) QM/MM methods, along with a Python software platform and workflow for training QM/MM-specific MLP models.
Overall, this thesis progresses from structural insight to accessible reaction simulations and ultimately to substantial computational acceleration, serving as a steppingstone toward next-generation enzyme reaction simulation strategies.
Keywords
Bio-molecular Simulations, Enzyme Catalysis, Machine Learning Potentials
Disciplines
Biophysics | Computational Chemistry | Physical Chemistry
License
This work is licensed under a Creative Commons Attribution-NonCommercial-No Derivative Works 4.0 International License.
Recommended Citation
Arattu Thodika, Abdul Raafik, "Accelerated and Accurate Enzyme Reaction Free Energy Simulations with Machine Learning Potentials" (2026). Chemistry & Biochemistry Dissertations. 297.
https://mavmatrix.uta.edu/chemistry_dissertations/297