Generalized biomolecular modeling and design with RoseTTAFold All-Atom.

Area of Science:

• Computational biology
• Structural biology
• Biochemistry

Background:

• Deep learning has advanced protein structure prediction but is limited to protein-only systems.
• Modeling complex biological assemblies requires integrating diverse chemical entities.

Purpose of the Study:

• To develop a deep learning framework capable of modeling and designing complex biological assemblies beyond proteins.
• To create a versatile tool for designing proteins that interact with non-protein components.

Main Methods:

• Introduced RoseTTAFold All-Atom (RFAA) using a hybrid residue- and atomic-level representation.
• Developed RFdiffusion All-Atom (RFdiffusionAA) by fine-tuning on denoising tasks for targeted design.
• Utilized crystallography and binding assays for experimental validation.

Main Results:

• RFAA successfully models assemblies of proteins, nucleic acids, small molecules, metals, and covalent modifications.
• RFdiffusionAA designed de novo proteins that bind specific small molecules.
• Validated designs include proteins binding digoxigenin, heme, and bilin.

Conclusions:

• RoseTTAFold All-Atom and RFdiffusion All-Atom represent a significant advancement in modeling and designing complex molecular assemblies.
• This approach opens new avenues for engineering proteins with desired functions and interactions.