Sidechain conditioning and modeling for full-atom protein sequence design with FAMPNN

Area of Science:

• Computational Biology
• Protein Engineering
• Deep Learning

Background:

• Current deep learning methods for protein sequence design often neglect explicit modeling of sidechain conformation.
• Sidechain conformation is critical for protein structure, stability, and function.
• Existing models infer sidechain interactions indirectly from backbone geometry and sequence labels.

Purpose of the Study:

• To develop a deep learning method that explicitly models both protein sequence identity and sidechain conformation.
• To improve the accuracy and applicability of computational protein design.

Main Methods:

• Introduction of FAMPNN (Full-Atom MPNN), a novel graph neural network architecture.
• Jointly learning discrete amino acid identity and continuous sidechain conformation using a combined categorical cross-entropy and diffusion loss objective.
• Utilizing full-atom representations for each residue during sequence generation.

Main Results:

• FAMPNN demonstrates improved sequence recovery compared to existing methods.
• The method achieves state-of-the-art performance in sidechain packing.
• Benefits of full-atom modeling extend to accurate zero-shot prediction of experimental binding and stability.

Conclusions:

• Explicitly modeling sidechain conformation alongside sequence identity is a synergistic approach that enhances protein design.
• FAMPNN offers a more comprehensive and accurate method for computational protein sequence design.
• The developed method has practical implications for predicting protein biophysical properties.