Learning Potential Energy Surfaces of Hydrogen Atom Transfer Reactions in Peptides

Abstract

Hydrogen atom transfer (HAT) reactions are essential in many biological processes, such as radical migration in damaged proteins, but their mechanistic pathways remain incompletely understood. Simulating HAT processes is challenging due to the conflicting requirements of quantum chemical accuracy and biologically relevant time and length scales; thus, neither classical force fields nor DFT-based molecular dynamics simulations are applicable. Machine-learned potentials offer an alternative, with the ability to learn potential energy surfaces (PESs) that capture reactions and transitions with near-quantum accuracy. However, training such models to generalize across diverse HAT configurations—especially at radical positions in proteins—requires tailored data generation strategies and careful model selection. In this work, we systematically generate HAT reaction configurations in peptides to build large datasets using semiempirical methods as well as DFT. We benchmark three atomistic machine-learned potential architectures, SchNet, Allegro, and MACE, on their ability to learn HAT potential energy surfaces and indirectly predict reaction barriers through direct energy predictions. MACE consistently outperforms the other models in energy, force, and reaction barrier prediction accuracy, achieving a mean absolute error of 1.13 kcal/mol on DFT barrier predictions. Short molecular dynamics simulations indicate that the learned potential is numerically stable at finite temperature and can sustain reactive HAT sampling under moderate biasing, serving as a feasibility check for downstream simulation workflows. We show that the trained MACE potential generalizes well beyond our training data by performing out-of-distribution evaluations and analysis of HAT barriers in collagen I snapshots. This level of accuracy can enable integration of machine-learning-based barrier predictions into large-scale simulation workflows to compute reaction rates from predicted barriers, advancing the mechanistic understanding of HAT and radical migration in peptides. We analyze scaling laws, model transferability, and cost-performance trade-offs, and outline strategies for improvement through the combination of machine-learned potentials with transition state search algorithms and active learning. The presented approach is generalizable to other biomolecular systems, offering a method toward quantum-accurate barrier predictions of chemical reactivity in complex biological environments.