Accelerated Sampling of Protein Dynamics Using BioEmu-Augmented Molecular Simulation.

Generative models, such as BioEmu, can produce diverse conformational ensembles of proteins; however, determining which predicted states are physiologically relevant and how they are populated remains a major challenge. In particular, it is unclear which conformations correspond to functional states and whether intermediate states and transitions between them can be reliably identified. Here, we investigate these questions using the serine-threonine kinases BRAF and CDK2. We examine how the disease-associated BRAF mutation V600E alters the populations of metastable conformational states relative to the wild type. To quantify conformational populations, we developed a workflow that combines conformational ensembles generated by BioEmu, with short molecular dynamics simulations and Markov State Model to estimate Boltzmann-weighted state populations. In addition, we integrate BioEmu ensembles with experimental cryo-electron microscopy data to construct all-atom conformational ensembles of biomolecules. Compared to the AlphaFold2 reduced multiple sequence alignment approach (rMSA-AF2), BioEmu-seeded molecular simulations more effectively sample functionally important metastable states and capture mutation-induced population shifts in serine-threonine kinases. However, it fails to capture conformational heterogeneity in multiple systems, including glycine transporter 1 (GlyT1) and plasmepsin-II (PlmII). In systems where side-chain conformational heterogeneity governs dynamics, such as cryptic pocket opening in PlmII or transitions between metastable states in GlyT1, molecular simulations initiated from BioEmu-generated ensembles do not reproduce the experimentally observed conformational variability. Together, this work introduces a framework for integrating generative protein ensemble prediction models with statistical mechanical reweighting to recover Boltzmann-weighted conformational ensembles at scale while highlighting important limitations that require system-specific evaluation when interpreting AI-generated protein conformational landscapes.