Computational Investigation of GNMT-Catalyzed Methyl Transfer Reaction: Integrating MD, QM, and ML Approaches

The glycine N-methyltransferase (GNMT) reaction was examined using an integrated workflow combining molecular dynamics (MD), quantum mechanical (QM) cluster calculations, and machine learning (ML) analysis. Instead of relying on a single crystal-like conformation, multiple MD simulations were used to sample diverse reactant (SAM + glycine bound to GNMT) and product (SAH + sarcosine bound to GNMT) state geometries for QM cluster modeling. Across more than 150 QM-cluster models constructed by the Residue Interaction Network ResidUe Selector (RINRUS) from selected MD frames with and without explicit waters, the computed activation and reaction free energies span broad ranges (7 to 25 kcal mol^-1 and -36 to +3 kcal mol^-1), demonstrating a strong dependence on the initial MD conformation. Product-state consistently yields lower reaction barriers, while explicit water introduces only small shifts in energetics and preserves the relative ordering among frames. The two-coordinate potential energy surface (PES) offers only limited insight and cannot fully account for the observed energetic variability. These QM-cluster models were further analyzed using machine-learning methods to identify structural descriptors that correlate with the observed energy variations and provide insight into their structural origin. ML models trained on multiple feature representations show that the donor-methyl-acceptor distances are the most informative and yield the strongest predictive accuracy, while higher dimensional solvent- or residue-based features contribute comparatively little. Overall, the results highlight the importance of conformational sampling for reliable QM-cluster energetics and point toward more expressive structure-to-property representations for analyzing enzymatic reactions.