An all-atom protein generative model.

Area of Science:

• Biochemistry and Structural Biology
• Computational Biology and Bioinformatics
• Artificial Intelligence in Molecular Design

Background:

• Proteins function via chemical interactions, primarily through sidechains, necessitating accurate modeling for protein design.
• Existing generative models face challenges in managing the continuous and discrete properties of protein structure and sequence.
• All-atom modeling is crucial for capturing the detailed chemical behavior of proteins.

Approach:

• Introduced Protpardelle, an all-atom diffusion model for protein structure generation.
• Implemented a novel 'superposition' scheme to handle discrete sidechain states within a continuous diffusion process.
• Integrated Protpardelle with sequence design methods for joint structure-sequence co-design.

Key Points:

• Protpardelle generates high-quality, diverse, and novel protein structures and sequences.
• The model accurately reproduces the chemical features and behavior of natural protein sidechains.
• Generated proteins demonstrate good quality metrics (quality, diversity, novelty).

Conclusions:

• Protpardelle enables all-atom protein design, including the scaffolding of functional motifs.
• The model offers a backbone- and rotamer-free approach to protein engineering.
• This work advances generative modeling for de novo protein design and functional site engineering.