Use RFdiffusion2 Online
Commercially Available RFdiffusion2 No-Code Web Server
RFdiffusion2: Atom-Level Active Site Scaffolding for De Novo Enzyme Design
RFdiffusion2 is a generative model for scaffolding atomic-level enzyme active sites, described in a 2025 bioRxiv preprint. It extends RoseTTAFold All-Atom diffusion (RFdiffusionAA) so that a designer can specify catalytic residues at the level of individual atoms, without having to decide in advance which sequence positions or side-chain rotamers those residues will occupy. The model is trained with flow matching and builds a protein backbone around the functional site.
Designing enzymes from scratch means placing a few catalytic groups, often called a theozyme, in exactly the right geometry and then building a stable protein around them. Earlier backbone-generation tools generally needed the residue indices and rotamers of the motif to be fixed up front, which forces the designer to guess answers the model could work out itself. RFdiffusion2 is built to remove that constraint.
Fact | Detail |
|---|---|
Task | Scaffold atomic active sites (theozymes) into new protein backbones |
Builds on | RoseTTAFold All-Atom diffusion (RFdiffusionAA) |
Training approach | Flow matching with atomized residues and unindexed motifs |
Benchmark | Atomic Motif Enzyme (AME) benchmark of 41 active sites |
Training compute | 17 days on 24 A100 GPUs |
How RFdiffusion2 Works: Unindexed Motifs and Atomized Residues
Atomized residues: catalytic residues are given as atoms rather than as fixed residue types at fixed positions.
Unindexed motifs: the model infers the rotamers and sequence indices of the motif during generation.
Flow-matching training: atoms follow Gaussian paths in 3D, and residue frames use Riemannian flow matching in SE(3).
Stochastic centering: training structures are randomly centered, and an origin pseudo-atom sets the approximate center of mass of the scaffold.
Ligand conditioning: per-atom relative solvent accessibility labels can be given for ligands, and partial ligands are supported, with the model sampling the rest of the conformer.
In the reported pipeline, RFdiffusion2 generates backbones, LigandMPNN designs sequences, and an all-atom structure predictor checks the result (Chai-1 was used for the benchmark).
RFdiffusion2 Results: Enzyme Design Benchmark and Experimental Validation
On the AME benchmark, 100 structures were sampled per active site and 8 sequences were designed per structure. A case counted as solved if at least one sequence gave a catalytic-residue heavy-atom RMSD below 1.5 Å in the predicted structure with no ligand clashes.
Method | Active sites solved (of 41) |
|---|---|
RFdiffusion2 | 41 |
RFdiffusion | 16 |
RFdiffusion2 did better than RFdiffusion in 40 of 41 cases and handled motifs split across up to seven residue islands. Letting the model infer rotamers and indices gave the best success rate.
Experimental Enzyme Activity
Retroaldolase: 4 active variants out of 96 tested.
Cysteine hydrolase: up to 48 designs screened, with a best kcat/KM of 248 ± 34 per M per s.
Zinc hydrolase: functional enzymes were found for two substrates, and a design with a general base gave 11 functional enzymes from 96 tested, with a best kcat/KM of 53,000 ± 5,000 per M per s.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with RFdiffusion2 on Tamarind Bio
De novo enzyme design: build new protein scaffolds around a theozyme for a reaction of interest.
Hydrolase design: the authors report functional zinc and cysteine hydrolases.
Multi-island motifs: scaffold active sites whose residues are spread across several segments.
Ligand-aware design: condition on a ligand, including partial ligands, when designing a binding pocket.
How to Use RFdiffusion2 on Tamarind Bio
Log in and open the tool: sign in at tamarind.bio and open RFdiffusion2.
Define the active site: provide the catalytic residue atoms (the theozyme) and the ligand structure, as the tool exposes them.
Set motif options: choose which motif atoms are unindexed, if exposed.
Set design size: specify the scaffold length range, if exposed.
Run the job: submit the run and wait for backbones to be generated.
Download outputs: retrieve the generated backbones.
Design sequences: in the paper's pipeline, sequences come from LigandMPNN.
Validate: check designs with an all-atom structure predictor and filter on motif RMSD and clashes.
Things to Keep in Mind
Designed enzymes in the paper are less active than native enzymes.
The chosen theozymes may not capture every interaction needed for high activity.
The authors suggest sequence co-design outside the active site and newer architectures as future improvements.
Source: RFdiffusion2 preprint, bioRxiv, 2025, DOI 10.1101/2025.04.09.648075 (https://www.biorxiv.org/content/10.1101/2025.04.09.648075v2). The authors state that inference and training code will be made freely available.