De novo Design of All-atom Biomolecular Interactions with RFdiffusion3

Area of Science:

• Computational biology
• Structural biology
• Artificial intelligence in drug discovery

Background:

• Deep learning advances protein design but often overlooks interactions with other biomolecules.
• Existing methods primarily generate protein backbones, limiting their application.

Purpose of the Study:

• Introduce RFdiffusion3 (RFD3), a diffusion model for generating protein structures.
• Enable protein design within the context of ligands, nucleic acids, and other non-protein atoms.
• Improve efficiency and effectiveness of protein design with complex constraints.

Main Methods:

• Developed RFdiffusion3 (RFD3), a diffusion model explicitly modeling all polymer atoms.
• Utilized atom-level constraints for conditioning the model on complex molecular interactions.
• Evaluated RFD3 performance on various in silico benchmarks.

Main Results:

• RFD3 demonstrates superior performance on in silico benchmarks compared to prior methods.
• Achieved comparable results with one-tenth the computational cost of existing approaches.
• Successfully designed and experimentally validated DNA binding proteins and cysteine hydrolases.

Conclusions:

• RFD3 enables simpler and more effective protein design with complex atom-level constraints.
• The model's ability to incorporate non-protein atoms expands attainable protein functions.
• RFD3 significantly advances the field of AI-driven protein engineering.