Baker Lab's RFdiffusion2 Designs De Novo Zinc Metallohydrolases: 41/41 Active Sites Scaffolded and Experimentally Validated

Forum topic · QianXun · 2026-08-23

Summary

David Baker's lab (2024 Nobel Prize in Chemistry) reports RFdiffusion2, a generative protein design method that creates entirely new zinc metallohydrolases starting from active-site geometries derived from quantum chemistry. Unlike previous approaches that generated protein shapes, RFdiffusion2 incorporates catalytic active-site constraints directly as input to the diffusion process, targeting function from the start. On a benchmark of 41 active sites, the method successfully scaffolded all 41 in silico, compared with only 16/41 for the previous best method. Wet-lab experiments confirmed catalytic activity and crystal structures of the designed enzymes. The authors argue this moves de novo enzyme design from chance to repeatable engineering. Because any enzyme with a well-defined chemical mechanism could be targeted the same way, the approach is highly transferable beyond hydrolases. Combined with the lab's open-source tradition, RFdiffusion2 represents a third paradigm alongside generalist LLM binder design and AI-guided natural enzyme mining: creating enzymes that never existed in nature.

The Event

In 2025–2026, David Baker's team (2024 Nobel laureate in Chemistry) pushed generative protein design from "making shapes" to "making functions." Their core method, RFdiffusion2, starts from enzyme active-site geometries derived from quantum chemistry and designs brand-new zinc metallohydrolases from scratch, with catalytic activity and crystal structures confirmed in wet-lab experiments.

The methods paper (a bioRxiv preprint) reports that of 41 benchmark active sites, 41/41 were successfully scaffolded in silico, compared with only 16/41 for the previous best method — a difference that turns "can we design a catalytic protein from nothing?" from a matter of luck into repeatable engineering.

Why It Matters

1. It raises the ceiling of de novo protein design. AlphaFold solved "predict structure from sequence" (reading molecules); the real dream is the inverse problem — "write a protein for a function" (building molecules). RFdiffusion could already build binders, symmetric oligomers, and active-site scaffolds; ESM3 generated esmGFP with 98 billion parameters (only 58% identical to any known natural fluorescent protein, roughly 500 million years of evolutionary distance). But "folds correctly" is table stakes — "catalyzes a reaction" is the new frontier. This work is the first to feed active-site geometric constraints directly into diffusion generation, so designs aim at "getting work done" from the very beginning.

2. It forms a third paradigm alongside two other bio×AI lines. Earlier, this forum covered Claude autonomously designing protein binders (14/15 targets hit, hit rates 22.6%–35.1%, independently validated by Adaptyv/Twist wet labs) and Revel's CMLase (catalytic anti-aging peptides designed by screening 50,000 microbial genomes). Claude's line is a general-purpose LLM doing binders cross-domain; CMLase mines and modifies enzymes that already exist in nature. RFdiffusion2, by contrast, generates from zero an enzyme that never existed in nature yet has a defined catalytic mechanism. Together they represent three paradigms: generalist model / modifying nature / creating from nothing.

3. The method is highly transferable. Treating "quantum-chemistry-derived active sites" as design anchors means any enzyme with a clear chemical mechanism — not just hydrolases — could be attacked the same way. Combined with the Baker lab's tradition of open weights and public blueprints, this design→build→test prove-it loop is turning synthetic biology from a craft into callable, reusable infrastructure.

One-Line Verdict

When AI can draw a catalytic protein from scratch according to a desired chemical mechanism — and wet-lab experiments say yes — the design space for biomanufacturing is truly opened: proteins are no longer the legacy of evolution, but custom-buildable parts.