Broadly applicable and accurate protein design by integrating structure prediction networks and diffusion generative models
Joseph L. WatsoniD, David JuergensiD, Nathaniel R. BennettiD, Brian L. TrippeiD, Jason YimiD, Helen E. EisenachiD, Woody AherniD, Andrew J. BorstiD, Robert J. RagotteiD, Lukas F. MillesiD, Basile I. M. WickyiD, Nikita HanikeliD, Samuel J. PellockiD, Alexis CourbetiD, William SheffleriD, Jue WangiD, Preetham VenkateshiD, Isaac SappingtoniD, Susana Vázquez TorresiD, Anna LaukoiD, Valentin De BortoliiD, Émile Mathieu, Sergey OvchinnikoviD, Regina BarzilayiD, Tommi JaakkolaiD, , David A. BakeriD
Abstract
There has been considerable recent progress in designing new proteins using deep learning methods1-9. Despite this progress, a general deep learning framework for protein design that enables solution of a wide range of design challenges, including de novo binder design and design of higher order symmetric architectures, has yet to be described. Diffusion models10,11 have had considerable success in image and language generative modeling but limited success when applied to protein modeling, likely due to the complexity of protein backbone geometry and sequence-structure relationships. Here we show that by fine tuning the RoseTTAFold structure prediction network on protein structure denoising tasks, we obtain a generative model of protein backbones that achieves outstanding performance on unconditional and topology-constrained protein monomer design, protein binder design, symmetric oligomer design, enzyme active site scaffolding, and symmetric motif scaffolding for therapeutic and metal-binding protein design. We demonstrate the power and generality of the method, called RoseTTAFold Diffusion (RFdiffusion), by experimentally characterizing the structures and functions of hundreds of new designs. In a manner analogous to networks which produce images from user-specified inputs, RFdiffusion enables the design of diverse, complex, functional proteins from simple molecular specifications.
02 /Specification & Classification
Click any tag to browse matching papers- Citations
- 2,373
- FWCI
- 210.01
- Field percentile
- top 1%
- Assay
- Research Focus
- Approach
In-depth summary
Claude Sonnet 4.6 · max effort · 8 Nov 2026 · from the full text of version 1 (9 Dec 2022) · CC BY 4.0 · version 2 was posted after this summary
Fine-tuning RoseTTAFold as a diffusion denoiser gives RFdiffusion, a general protein backbone generator; a designed influenza haemagglutinin binder matched its model by cryo-EM.
Earlier protein diffusion models rarely gave sequences that fold as intended in silico and were untested experimentally, Hallucination fails for higher-order symmetries, and prior Rosetta binder design needed thousands of screened designs. RFdiffusion tackles these tasks in one framework.
Key results
ShownDe novo binders were found for all five targets with fewer than 100 designs tested each. A cryo-EM structure of an influenza haemagglutinin (HA) binder matched its design model at atomic accuracy.
95 designs per target tested by BLI at 10 μM; overall success 19%, roughly two orders of magnitude above the prior Rosetta method on the same targets. HA and IL-7Rα binders roughly 30 nM; HA_20 KD 28 nM; 2.9 Å cryo-EM map, 0.63 Å to design, binder alone 0.6 Å. · Design of protein-binding proteins; Figure legends
ShownOn a 25-problem motif-scaffolding benchmark, RFdiffusion solved more problems than Hallucination or RFjoint Inpainting without hyperparameter tuning. It also scaffolded the p53 helix into MDM2 binders far tighter than the peptide.
Solved 23 of 25 problems vs 15 (Hallucination) and 19 (RFjoint Inpainting); higher success fraction in 19 of the 23. MDM2 binders of 0.5 and 0.7 nM vs 600 nM for p53 peptide; 55 of 96 designs showed detectable binding at 10 μM. · Functional-motif scaffolding; Figure legends
ShownRFdiffusion produced symmetric oligomers across cyclic, dihedral and icosahedral groups. SEC and electron microscopy often supported the intended oligomeric state and overall shape; of 608 selected designs, at least 87 matched within the 95% CI by SEC.
608 designs selected; at least 87 had SEC oligomeric states within the 95% CI and 126 within the 99% CI. 38 D2, seven D3 and three D4 designs had expected molecular weights. One of forty-eight icosahedra (HE0902, 15 nm) formed homogeneous particles matching the design by nsEM. · Design of higher-order oligomers; Figure legends
ShownCombining symmetric design with motif scaffolding gave C4 assemblies that bind Ni2+ via the scaffolded histidine. Mutating that histidine to alanine eliminated or reduced binding.
44 designs expressed, 37 with SEC consistent with the intended oligomer; 36 tested by isothermal titration calorimetry, 18 bound Ni2+ (low nanomolar to low micromolar). H52A abolished or notably reduced binding in 17 out of 17 expressed mutants. nsEM of four designs showed fourfold symmetry. · Symmetric functional-motif scaffolding; Figure legends
ShownRFdiffusion outperformed RF Hallucination at unconditional monomer generation and ran faster. Of 18 designs of 200 and 300 amino acids tested, 9 expressed, and these were folded and thermostable.
Two-proportion z-test, n = 400 designs per condition, z = 9.5, P = 1.6 × 10^−21. A 100-residue protein takes as little as 11 s on an NVIDIA RTX A4000 vs around 8.5 min for RF Hallucination. 9/18 designs of 200 and 300 amino acids expressed, folded at 95 °C. · Unconditional protein monomer generation; Figure legends
Methods, data and reproducibility
RoseTTAFold is fine-tuned as a denoiser on PDB structures noised over up to 200 steps (Gaussian Cα noise, Brownian rotations, MSE loss, self-conditioning). ProteinMPNN designs sequences and AlphaFold2 validates designs in silico.
- PDB structures sampled and noised to train the denoising network
- In silico motif-scaffolding benchmark of problems from six publications; 100 designs generated per problem · n = 25 problems
- Experimentally tested unconditional monomers (200 and 300 amino acids) with circular dichroism · n = 18 designs tested, 9 expressed
- Symmetric oligomer designs screened by SEC and nsEM · n = 608 designs selected
- C4 Ni2+-binding designs expressed in E. coli, tested by SEC and isothermal titration calorimetry · Escherichia coli (expression host) · n = 44 designs expressed; 36 tested by ITC
- Binder designs to HA, IL-7Rα, PD-L1, InsR and TrkA, screened by BLI · n = 95 designs per target
- Cryo-EM of HA_20 binder in complex with influenza A/USA:Iowa/1943 H1N1 HA · Influenza A virus · PDB 8SK7; EMDB-40557
- Electron microscopy data for D4 oligomer HE0537 · EMDB-40602
- Methods & tools
- RoseTTAFold (RF), RFdiffusion, ProteinMPNN, AlphaFold2, ESMFold, RF Hallucination, RFjoint Inpainting, Biolayer interferometry, Isothermal titration calorimetry, Negative stain EM, Cryo-EM, Circular dichroism
- Data & accessions
- https://figshare.com/s/439fdd59488215753bc3https://doi.org/10.2210/pdb8SK7/pdbhttp://www.ebi.ac.uk/pdbe/entry/emdb/40557
- Assay conditions & compute
- A 100-residue design takes as little as 11 s on an NVIDIA RTX A4000 GPU, vs around 8.5 min for RF Hallucination. Training hardware is not stated in the provided text.
- Reproducibility gaps
- Training details are in the Supplementary Methods, not the provided text. A Google Colab notebook is available through GitHub, but software versions and containers are not stated.
Limitations (5)
- RFdiffusion cannot currently model bound small molecules explicitly. The substrate is only approximated implicitly with an external potential, which acts on backbone Cα atoms and ignores sidechains. · stated by the authors
- Binder specificity was only checked by competition BLI for six IL-7Rα binders; proteome-wide specificity needs further work. Success on polar target sites and sites without native partners remains untested. · stated by the authors
- On hard motif-scaffolding problems where all methods have low success, RFdiffusion can still struggle. Enzyme active-site in silico success required task-specific fine-tuning. · stated by the authors
- Enzyme active-site scaffolds were validated only in silico with AlphaFold2, with no experimental activity or structure data reported. In silico success may not predict catalysis. · noted in this summary
- High-resolution structural confirmation covers only the HA_20 binder. Oligomers were characterized by SEC and low-resolution nsEM or cryo-EM 2D classes, and at least 87 of 608 selected designs matched the design oligomeric state by SEC. · noted in this summary
Who should read:Read this if you design proteins computationally and want benchmarks and experimental hit rates for diffusion-based monomer, symmetric, motif-scaffolding and binder design.
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
- Joint Generation of Protein Sequence and Structure with RoseTTAFold Sequence Space Diffusion
S. L. Lisanza, J. M. Gershon, … D. Baker·bioRxiv·May 8, 2023
- De novo Design of All-atom Biomolecular Interactions with RFdiffusion3
Jasper Butcher, Rohith Krishna, … David A. Baker·bioRxiv·Sep 18, 2025
- Proteus: exploring protein structure generation for enhanced designability and efficiency
Chentong Wang, Yannan Qu, … Longxing Cao·bioRxiv·Feb 10, 2024
- Protein generation with evolutionary diffusion: sequence is all you need
S. Alamdari, N. Thakkar, … K. K. Yang·bioRxiv·Sep 11, 2023
- De novo protein design by deep network hallucination
I. Anishchenko, T. M. Chidyausiku, … D. Baker·bioRxiv·Jul 22, 2020
- Improving de novo Protein Binder Design with Deep Learning
N. Bennett, B. Coventry, … D. Baker·bioRxiv·Jun 15, 2022