Robust deep learning-based protein sequence design using ProteinMPNN.

Area of Science:

• Computational Biology
• Protein Engineering
• Deep Learning

Background:

• Traditional de novo protein design relies on physically based methods like Rosetta.
• Deep learning has transformed protein structure prediction but not yet sequence design.

Purpose of the Study:

• Introduce ProteinMPNN, a deep learning-based protein sequence design method.
• Demonstrate its superior performance compared to existing methods.
• Showcase its versatility across diverse protein design challenges.

Main Methods:

• Developed ProteinMPNN, a deep learning model for protein sequence design.
• Evaluated performance on native protein backbones, comparing sequence recovery with Rosetta.
• Applied the method to various complex protein structures, including monomers, oligomers, nanoparticles, and binding proteins.

Main Results:

• ProteinMPNN achieved 52.4% sequence recovery on native backbones, significantly higher than Rosetta's 32.9%.
• The method enables coupled sequence design across single and multiple chains.
• Successfully rescued previously failed designs and validated new designs using X-ray crystallography, cryo-electron microscopy, and functional studies.

Conclusions:

• ProteinMPNN offers a powerful and accurate deep learning approach for de novo protein sequence design.
• Its ability to handle complex, multi-chain designs broadens the scope of protein engineering.
• Experimental validation confirms the method's high utility and accuracy for diverse protein design applications.