An end-to-end deep learning method for protein side-chain packing and inverse folding

Area of Science:

• Computational biology
• Structural bioinformatics
• Machine learning in biochemistry

Background:

• Protein side-chain packing (PSCP) is crucial for protein structure prediction, refinement, and design.
• Existing PSCP methods often face limitations in speed and accuracy.
• There is a need for more efficient and precise computational tools for PSCP.

Purpose of the Study:

• To introduce AttnPacker, a deep learning method for direct prediction of protein side-chain coordinates.
• To enhance the computational efficiency and accuracy of PSCP.
• To enable simultaneous prediction of side-chain conformations and protein design.

Main Methods:

• AttnPacker utilizes deep learning to directly predict protein side-chain coordinates from backbone atom positions.
• The method incorporates 3D backbone geometry to simultaneously compute all side-chain coordinates.
• It avoids reliance on discrete rotamer libraries and expensive conformational search or sampling.

Main Results:

• AttnPacker demonstrates over 100× increased computational efficiency compared to DLPacker and RosettaPacker.
• The method produces physically realistic side-chain conformations, reducing steric clashes.
• AttnPacker improves root-mean-square deviation (rmsd) and dihedral accuracy over state-of-the-art methods.
• It enables sequence and side-chain codesign, yielding designs with subnative Rosetta energy.

Conclusions:

• AttnPacker offers a significant advancement in protein side-chain packing prediction.
• The method provides a computationally efficient and accurate solution for PSCP.
• AttnPacker has potential applications in protein structure prediction, refinement, and advanced protein design.