RFdiffusion2 is Now Available On GitHub

The long awaited inference code for RFdiffusion2 is now public! Developed by the Baker Lab, RFDiffusion2 builds on the capabilities of RFdiffusion AA to model more complex enzyme structures than was previously possible with the original RFdiffusion method. 

What’s New?

In the first RFdiffusion release, enzyme design was restricted: the geometry of the active site had to be represented at the residue level, losing important information about the placements of the residue side chains. In RFdiffusion2, users can

• Describe their enzyme active site either using side chain or backbone atomic coordinates to obtain a more diverse set of structures
• Use stochastic centering to specify their active site by defining the center of mass of the target enzyme
• Control the depth of each reactant/cofactor in the protein structure 
• Specify partial ligands – useful when the atoms comprising the ligand are known, but not the transition state structure

Benchmarking

To test this new method, a new enzyme benchmark was created: Atomic Motif Enzyme (AME). 

RFdiffusion2 found solutions to all 41 active sites included in the benchmarking set, outperforming RFdiffusion, which was only able to find 16. In particular, the original RFdiffusion failed to generate backbones for enzymes with ‘residue islands’ – contiguous segments of catalytic residues given in the input.

In Vitro Results

Beyond benchmarking, the authors tested the capabilities of RFdiffusion2 for designing enzymes with in vitro catalytic activity. While not as active as natural enzymes, RFdiffusion2 successfully designed functional enzymes for four different catalytic reactions.

Try It Out

We are excited to bring this new tool to the Commons. If you have any questions, comments, or concerns open an Issue in the RFdiffusion2 repository or reach out to the Rosetta Commons team via our Contact Form.