In January 2026, a team from the Institute for Protein Design at the University of Washington, led by Nobel laureate David Baker, published RFdiffusion2 in Nature Methods, demonstrating a new AI model that designs functional enzymes directly from atomic-level descriptions of desired active sites. The paper marks a significant advance in de novo enzyme design, showing that AI can create proteins with catalytic activities approaching those found in natural enzymes.
Designing from Functional Group Geometries
RFdiffusion2 overcomes a key limitation of previous protein design approaches. Earlier methods required specifying the identity and order of amino acid residues at the active site, severely constraining the design space. RFdiffusion2 instead takes as input only the geometry of functional groups, the precise spatial arrangement of atoms needed for catalysis, and generates complete protein scaffolds around them. This allows the model to explore a vastly larger space of possible protein architectures.
Benchmark Performance
The model successfully generated scaffolds for all 41 active sites in a diverse benchmark, compared to just 16 using previous methods. This dramatic improvement in coverage means that RFdiffusion2 can tackle a much wider range of enzymatic reactions and target chemistries than was previously possible with computational design tools.
Experimental Validation
In laboratory testing, the results were striking. From an initial set of 96 computationally designed enzymes, the most active achieved a catalytic efficiency (kcat/KM) of 16,000 M-1 s-1. A second round of 96 designs yielded three additional highly active enzymes, with catalytic efficiencies reaching up to 53,000 M-1 s-1. These activities approach those of some natural enzymes, a threshold that computational enzyme design has long struggled to reach.
Using quantum chemistry calculations from naturally occurring zinc metallohydrolases to define key active site atoms, RFdiffusion2 designed proteins with scaffolds correctly positioning atoms to break ester linkages. The experimentally synthesized designs achieved enzymatic activities close to those of their natural counterparts.
RFdiffusion3: The Next Generation
Alongside the RFdiffusion2 publication, the Baker lab also released RFdiffusion3, an even more advanced model that treats individual atoms as the fundamental design units rather than amino acid residues. RFdiffusion3 is ten times faster than RFdiffusion2 despite using a more computationally intensive all-atom approach, and handles a broader range of non-protein molecules and catalytic sites. RFdiffusion3 has been made freely available as open-source software.
Implications for Drug Discovery
The ability to design functional enzymes from scratch has direct implications for drug discovery, particularly in the development of enzyme replacement therapies, prodrug-activating enzymes, and biocatalysts for pharmaceutical manufacturing. The open-source availability of both RFdiffusion2 and RFdiffusion3 means these tools are immediately accessible to researchers worldwide, accelerating the application of AI-driven protein design across academia and industry.
