Scalable emulation of protein equilibrium ensembles with generative deep learning

Area of Science:

• Computational Biology
• Structural Biology
• Biophysics

Background:

• Predicting protein structure changes is crucial but challenging.
• Existing methods struggle to capture functional dynamics at scale.

Purpose of the Study:

• Introduce BioEmu, a deep learning system to emulate protein equilibrium ensembles.
• Enable scalable prediction of functionally relevant protein structure changes.

Main Methods:

• BioEmu integrates molecular dynamics (MD) simulations, static structures, and experimental stability data.
• New training algorithms enable rapid generation of thousands of protein structures per hour on a GPU.
• Jointly models structural ensembles and thermodynamic properties.

Main Results:

• BioEmu captures diverse functional motions like cryptic pocket formation and domain rearrangements.
• Predicts relative free energies with 1 kcal/mol accuracy compared to MD and experimental data.
• Achieves millisecond-scale simulation capabilities.

Conclusions:

• BioEmu provides mechanistic insights into protein function by modeling ensembles and thermodynamics.
• Amortizes costs of MD and experimental data generation.
• Offers a scalable approach for protein function understanding and design.