Scalable emulation of protein equilibrium ensembles with generative deep learning

TL;DR

BioEmu is a methodology that uses generative deep learning to rapidly sample protein structural equilibrium ensembles, potentially replacing molecular dynamics simulations.

Problem

Understanding protein function requires dynamic structural ensembles and thermodynamic properties, but experiments and molecular dynamics simulations suffer from low throughput and long timescales.

Approach

A generative model is trained using novel methods on protein structure data, over 200 milliseconds of molecular dynamics simulation, and experimental stability data to directly generate equilibrium ensembles.

Results & Contribution

BioEmu samples functional conformational changes including cryptic pockets, protein region unfolding, and large-scale domain rearrangements, validated with relative free energy errors around 1 kcal/mol. It provides mechanistic insights such as causes of fold destabilization by mutations and can efficiently generate experimentally testable hypotheses.