Modeling Solvated Phosphoryl Transfer using Machine Learning Potentials and Transfer Learning

Phosphoryl transfer plays a critical role in biochemistry, both in protein regulation and as a key energy source, yet its mechanistic details remain elusive. The complexity arises from multiple possible reaction pathways, environmental influences like electrostatics and protein interactions, and the challenge of bond-breaking, which requires highly scaling computational methods that account for electronic correlation. To address these challenges, we employ machine learning potentials trained on quantum mechanical datasets using a transfer learning approach. Initial models trained on lower-level electronic structure methods are refined using a minimal but strategically selected high-level dataset, chosen through active learning. Combined with enhanced sampling techniques, this approach enables accurate modeling of phosphoryl transfer across diverse contexts, providing insight into how environmental factors shape reaction pathways.