De novo Design of All-atom Biomolecular Interactions with RFdiffusion3
Jasper ButcheriD, Rohith KrishnaiD, Raktim MitraiD, Rafael I. BrentiD, Yanjing LiiD, Nathaniel CorleyiD, Paul T. KimiD, Jonathan FunkiD, Simon V. MathisiD, Saman SalikeiD, Aiko MuraishiiD, Helen E. EisenachiD, Tuscan Rock ThompsoniD, Jie CheniD, Yuliya PolitanskaiD, Enisha SehgaliD, Brian CoventryiD, Odin ZhangiD, Bo QiangiD, Kieran DidiiD, Maxwell W. KazmaniD, Frank DiMaioiD, David A. BakeriD
Abstract
Deep learning has accelerated protein design, but most existing methods are restricted to generating protein backbone coordinates and often neglect interactions with other biomolecules. We present RFdiffusion3 (RFD3), a diffusion model that generates protein structures in the context of ligands, nucleic acids and other non-protein constellations of atoms. Because all polymer atoms are modeled explicitly, conditioning the model on complex sets of atom-level constraints for enzyme design and other challenges is both simpler and more effective than previous approaches. RFD3 achieves improved performance compared to prior approaches on a range of in silico benchmarks with one tenth the computational cost. Finally, we demonstrate the broad applicability of RFD3 by designing and experimentally characterizing DNA binding proteins and cysteine hydrolases. The ability to rapidly generate protein structures guided by complex sets of atom-level constraints in the context of arbitrary non-protein atoms should further expand the range of functions attainable through protein design.
02 /Specification & Classification
Click any tag to browse matching papers- Citations
- 74
- Organism
- Assay
- Research Focus
- Approach
In-depth summary
Claude Opus 5.5 · xhigh effort · 2 Nov 2026 · from the full text of version 2 (19 Nov 2025) · CC BY 4.0
An all-atom diffusion model using a 14-atom residue representation and sparse U-Net architecture accelerates biomolecular design 10-fold, enabling experimentally validated DNA-binding proteins and multi-turnover cysteine hydrolases.
Existing protein design diffusion models primarily operate on amino acid backbones or condition residue-level diffusion on a small subset of side-chain atoms, limiting specific side-chain interactions with nucleic acids and small molecules. RFdiffusion3 diffuses all backbone and side-chain atoms simultaneously alongside non-protein biomolecules using a lightweight conditioning architecture, improving in silico benchmark performance at one-tenth the compute cost of RFdiffusion2.
Key results
ShownRFdiffusion3 achieves high unconditional backbone designability and approximately 10-fold faster inference than RFdiffusion2 across typical protein lengths.
Using a 2-layer Pairformer (168M trainable parameters vs ~350M for AlphaFold3) and sparse attention without triangle updates, 98% of unconditional designs of length 100-200 have at least 1 of 8 ProteinMPNN sequences predicted by AlphaFold3 within 1.5 Å RMSD, yielding 41 clusters out of 96 generations (length 100-250, TM-score 0.5). · Abstract; 2 RFdiffusion3 architecture and training; 3 In silico results; Figure legends
ShownIn silico benchmarks across protein-protein, protein-DNA, protein-ligand, and enzyme scaffolding tasks show higher success rates and structural diversity than prior RFdiffusion variants.
Across 5 protein targets (400 designs each), RFdiffusion3 beat RFdiffusion1 on 4/5 targets with 8.2 vs 1.4 unique successful clusters (TM-score 0.6); beat RFdiffusionAA on all 4 ligand targets; beat RFdiffusion2 on 37 of 41 (90%) Atomic Motif Enzyme active sites (15% vs 4% passing for >4 residue islands, n = 12); and reached 8.67% monomeric and 6.67% dimeric DNA-binder pass rates (<5 Å DNA-aligned RMSD). · 3.1 Design of protein-binding proteins; 3.2 Design of protein-nucleic acid interactions; 3.3 Designing small molecule binders; 3.4 Enzyme design
ShownAtom-level hydrogen-bond conditioning and classifier-free guidance increase the fraction of targeted polar interactions with small molecules and DNA bases.
Specifying hydrogen-bond donor or acceptor atoms increased targeted small-molecule hydrogen-bonding interactions from 26.67% to 32.67%, rising to 36.67% with classifier-free guidance, and similarly improved DNA-base hydrogen-bonding rates (11% vs 11.3% vs 12.5%). · 2 RFdiffusion3 architecture and training; Figure legends
ShownProspective wet-lab characterization validates a de novo micromolar DNA-binding protein and 35 multi-turnover cysteine hydrolases designed with RFdiffusion3.
Among 5 tested DNA-binding protein designs against target sequence CGAGAACATAGTCG, 1 bound in yeast surface display flow cytometry with EC50 = 5.89 ± 2.15 µM (3 replicates); screening 190 esterase designs scaffolding a Cys-His-Asp triad and Gln against 4-methylumbelliferyl phenyl acetate yielded 35 multi-turnover enzymes, with the top design reaching Kcat/Km = 3557. · Abstract; 4 Experiments; Figure legends
Methods, data and reproducibility
RFdiffusion3 represents each residue with 4 backbone and 10 side-chain atoms (placing unused virtual atoms on C-beta or C-alpha) and denoises all-atom coordinates using a transformer U-Net with sparse local attention, cross-attention pooling, and a 2-layer Pairformer. Designs are sequence-fitted with ProteinMPNN or LigandMPNN, filtered with AlphaFold3 or Chai-1, and tested in vitro.
- Biomolecular complexes from the Protein Data Bank (PDB) deposited through December 2024 and AlphaFold2 distillation structures · multiple species · n = all PDB complexes through December 2024 plus Hsu et al. AF2 distillation set · PDB IDs: 2GY5, 7RTE, 7N5U, 7M5W, 1CTT
- In silico benchmark targets for protein-protein, DNA, small-molecule, and Atomic Motif Enzyme (AME) design · n = 5 PPI targets, 3 DNA sequences, 4 small molecules, 41 AME active sites
- Experimental validation of de novo DNA-binding proteins via yeast surface display flow cytometry · Saccharomyces cerevisiae (yeast surface display) · n = 5 synthetic designs tested across 3 replicates
- Experimental screening of de novo cysteine hydrolase designs on 4-methylumbelliferyl phenyl acetate · Escherichia coli (purified C6 kinetics) and IVTT screen · n = 190 designs screened (35 multi-turnover)
- Methods & tools
- AtomWorks, ProteinMPNN, LigandMPNN, AlphaFold3, Chai-1, RFdiffusion, RFdiffusion2, RFdiffusionAA, Rosetta, FoldSeek, RDKit, 3DNA-Analyze
- Software & environment
- Built on the AtomWorks framework; software versions and container specifications are not stated in the main text.
- Experimental conditions
- Inference runtimes benchmarked on NVIDIA A6000 GPUs; model contains 168M trainable parameters and achieves ~10-fold speedup over RFdiffusion2.
- Reproducibility gaps
- Training dataset curation scripts and accession numbers for the experimentally validated designs are not included in the main text (PDB ID for Ulp-1 is left blank in S9).
Limitations (3)
- Unconditional sequences decoded directly from RFdiffusion3 coordinates show a bias toward alanine compared with native PDB proteins, requiring sequence redesign with ProteinMPNN or LigandMPNN for optimal structure consistency. · stated by the authors
- Increasing the diffusion inference step scale improves designability at the expense of structural diversity. · stated by the authors
- Experimental validation of DNA-binding protein design tested only five candidates against a single target DNA sequence, yielding one micromolar binder without specificity profiling against off-target DNA sequences. · noted in this summary
Who should read:Read this if you design proteins that bind small molecules, nucleic acids, or transition-state motifs and need atom-level control over hydrogen bonding and active-site geometry.
Reproducibility ledger
Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.
Sign in to file a reproducibility report.
Related papers
- Broadly applicable and accurate protein design by integrating structure prediction networks and diffusion generative models
Joseph L. Watson, David Juergens, … David A. Baker·NaturePeer-reviewed·Jul 11, 2023
- Designing protein/non-protein binding interactions using a full-atom diffusion model
Kale Kundert, George McDonald Church·bioRxiv·Feb 4, 2026
- Atom level enzyme active site scaffolding using RFdiffusion2
Woody Ahern, Jason Yim, … David A. Baker·Nature MethodsPeer-reviewed·Dec 3, 2025
- Protein Hunter: exploiting structure hallucination within diffusion for protein design
Yehlin Cho, Griffin Rangel, … Sergey Ovchinnikov·bioRxiv·Oct 10, 2025
- Constrained Diffusion for Protein Design with Hard Structural Constraints
Jacob K. Christopher, Austin Seamann, … Ferdinando Fioretto·bioRxiv·Oct 15, 2025
- RFOptimization: Guiding Design Optimization with All-Atom Structure Prediction
Odin Zhang, Jiaqi Wang, … David A. Baker·bioRxiv·Sep 9, 2026