Atom-level enzyme active site scaffolding using RFdiffusion2.

Area of Science:

• Biochemistry
• Computational Biology
• Protein Engineering

Background:

• Traditional enzyme design relies on idealized functional group arrangements and complex protein structure generation.
• Existing AI methods for enzyme design often require predefined residue positions and limited flexibility.

Purpose of the Study:

• To introduce RoseTTAFold diffusion 2 (RFdiffusion2), a deep generative model for de novo enzyme design.
• To overcome limitations of current AI-based enzyme design by enabling direct generation from functional group geometries.

Main Methods:

• Developed RFdiffusion2, a deep generative model for enzyme design.
• Utilized functional group geometries as direct input, bypassing residue order specification and inverse rotamer generation.
• Benchmarked RFdiffusion2 against existing methods on a diverse set of active sites.

Main Results:

• RFdiffusion2 successfully generated scaffolds for all 41 active sites in a benchmark set, significantly outperforming previous methods (16 active sites).
• Designed novel enzymes for three distinct catalytic mechanisms.
• Identified active enzyme candidates with fewer than 96 experimental tests per mechanism.

Conclusions:

• RFdiffusion2 demonstrates a powerful new approach for de novo enzyme creation.
• Atomic-level generative modeling offers a direct pathway from reaction mechanisms to functional enzyme designs.
• This method significantly enhances the efficiency and scope of computational enzyme engineering.