What it is
Designing an enzyme has meant fixing catalytic residue positions in advance and building protein backbones inward from side-chain placements, which sharply limits what can be attempted. RFdiffusion2 removes both constraints, generating protein scaffolds directly from the geometry of catalytic functional groups without specifying residue order or performing inverse rotamer generation. On a diverse benchmark it produced scaffolds for all 41 active sites, against 16 for previous methods. For three distinct catalytic mechanisms, active enzyme candidates were found after testing fewer than 96 sequences each.
Why it matters
Placing catalytic residues correctly is the central bottleneck in de novo enzyme design, and prior tools cleared 16 of the 41 benchmark sites while RFdiffusion2 clears all 41. Designing straight from reaction geometry lets a chemist specify a mechanism and get a candidate structure, rather than reverse-engineering a backbone from guessed residue positions. Recovering active enzymes after screening fewer than 96 sequences per mechanism means the hit rate is high enough for ordinary wet-lab throughput.
Underlined numbers link to their source. Every metric and quoted figure is listed under Sources and data below.
Filed underprotein design, enzymes, generative models, de novo proteins, biocatalysis
Watch
A short explainer of this result.