FlowPacker: protein side-chain packing with torsional flow matching

Area of Science:

• Computational Biology
• Structural Biology
• Machine Learning

Background:

• Accurate prediction of protein side-chain conformations is crucial for understanding protein folding and interactions.
• It also facilitates de novo protein design, a key area in biotechnology.

Purpose of the Study:

• To develop a fast and performant model for predicting protein side-chain conformations.
• To improve upon existing state-of-the-art methods in terms of accuracy and runtime.

Main Methods:

• Application of torsional flow matching and equivariant graph attention.
• Development of the FlowPacker model, conditioned on protein sequence and backbone.

Main Results:

• FlowPacker outperforms previous state-of-the-art baselines across most metrics.
• Achieved improved runtime compared to existing methods.
• Demonstrated utility in inpainting missing side-chain coordinates and handling multimeric targets.
• Showcased strong performance on antibody-antigen complex datasets.

Conclusions:

• FlowPacker represents a significant advancement in predicting protein side-chain conformations.
• The model's speed and performance offer new possibilities for protein structure prediction and design.
• Its versatility extends to applications like structure inpainting and analysis of complex biological targets.