Abstract
Molecular simulations have become indispensable in biological research. Their accuracy continues to improve, but directly modelling biochemical reactions - central to all life processes - remains computationally challenging. Here, we present a biomolecular reaction emulator that models reactions across conformational ensembles using kinetic Monte Carlo. Our method, KIMMDY, is capable of handling dynamic, large-scale systems with successive, competing reactions, even on the second timescale or slower. It leverages graph neural networks for large-scale prediction of reaction rates, while also being capable of using simpler physics-based or heuristic models. We validate our approach against experimental data and showcase its power and versatility through a series of applications, including radical reactions, nucleophilic substitutions, and photodimerization. Example systems span proteins and DNA. KIMMDY aids the understanding of biochemical reaction cascades in complex systems, helps to re-interpret experimental data, and can inspire future wet-lab experiments.