Data-Efficient Generation of Protein Conformational Ensembles with Backbone-to-Side-Chain Transformers

Area of Science:

• Computational Biology
• Biophysics
• Machine Learning

Background:

• Data-driven generative models show promise for sampling high-dimensional data.
• Current models struggle with the limited data for biomolecular conformational ensembles.
• Existing methods lack the ability to propose rare, yet physically relevant, conformations.

Purpose of the Study:

• To develop a novel generative model for sampling biomolecular conformational ensembles.
• To enable the discovery of rare conformations not present in existing datasets.
• To facilitate scalable physical simulations using generative neural networks.

Main Methods:

• A modular strategy using backmapping from a fixed protein backbone.
• Integration of statistical models of side-chain conformations (rotamer libraries).
• Utilizing the transformer architecture for atomistic accuracy and conformational diversity.

Main Results:

• The model maintains side-chain conformational diversity.
• Side-chain fluctuations are coupled using global protein conformation information.
• Achieved atomistic accuracy in sampling conformations.

Conclusions:

• The proposed strategy enables rapid data acquisition for generative models.
• This approach is a crucial step towards scalable physical simulations.
• The backmapping method enhances the ability of generative models to explore biomolecular conformations.