BPS2026 – Dynamics informed machine learning: Physics-guided feature importance identifies key residues for enzyme design

Artificial enzyme design aims to create catalysts capable of performing reactions not found in nature. Current workflows typically begin by computationally engineering an enzyme structure featuring an optimized active site, followed by directed evolution to improve catalytic activity. However, approaches that focus only on static structures fail to capture conformational dynamics, often yielding enzymes with catalytic efficiencies many orders of magnitude lower than natural counterparts. To address this shortcoming, we developed a machine-learning workflow based on random forest classifiers that identifies residues whose dynamic behavior correlates with increased catalytic efficiency. To evaluate our workflow, we select kemp eliminase, a well-established de novo enzyme design that has been extensively characterized through directed evolution. Protein dynamics are introduced through enhanced sampling trajectories driven by collective variables derived from our recently developed frequency selective anharmonic (FRESEAN) mode analysis, which extracts low-frequency vibrational modes responsible for slow conformational changes. For each kemp eliminase variant, the model ingests a contact distance matrix at each trajectory timestep and predicts which evolutionary variant the snapshot belongs to. We then apply a novel physics-based feature importance metric that quantifies how the dynamics of individual residues contribute to the model’s classification decisions. To prioritize candidate residues for mutagenesis, we highlight amino acids whose motions are most perturbed by directed evolution and demonstrate how additional mutations can reshape the conformational flexibility of kemp eliminase, offering a basis for identifying mutational hotspots that can improve catalytic performance. In future work, we aim to integrate our computational predictions with experimental mutagenesis to validate predicted hotspots and extend our approach to other artificial enzymes beyond kemp eliminase.