Designing antibodies with RFdiffusion • Baker Lab

Today we are sharing a significant update in the use of AI to generate antibodies. In a new preprint, we introduce a version of RFdiffusion fine-tuned to design human-like antibodies. We are also making this software free to use for both non-profit and for-profit research, including drug development.

Why antibodies?

Antibodies are proteins in the immune system that recognize target molecules and trigger protective responses in the body. By redirecting natural antibodies to new targets, drug makers can control certain biological activities linked to disease. This has yielded treatments for cancer, autoimmunity, and infection. However, creating new antibodies using traditional drug development strategies is often challenging, slow, and expensive.

RFdiffusion for antibody design

Building on previous breakthroughs in AI-driven protein design, we trained a new version of RFdiffusion specialized in building antibody loops—the intricate, flexible regions responsible for antibody binding. This model produces new antibody blueprints unlike any seen during training that bind user-specified targets. In an initial preprint posted last March, only short but functional antibody fragments called nanobodies could be made. Now, RFdiffusion has been trained to also generate more complete and human-like antibodies called single chain variable fragments (scFvs).

This project was led by Nate Bennett, Joe Watson, Rob Ragotte, Andrew Borst, DéJenaé See, Connor Weidle, Riti Biswas, and Yutong Yu.

“RFdiffusion was already great at designing binding proteins with rigid parts, but it struggled with flexible loops. By extending the model to the challenge of antibody loop design, brand new functional antibodies can now be developed purely on the computer,” said Nate Bennett, a recent postdoc in the lab.

Experimental validation

The team used the fine-tuned models to make antibodies against several targets relevant to disease, including influenza hemagglutinin and a potent toxin produced by the bacteria Clostridium difficile (C. diff).

Lab tests confirmed these proteins bound their intended targets with moderate affinity, providing a foundation for further optimization. Electron microscopy on five such antibodies revealed that four interacted with their binding partner exactly as intended, verifying RFdiffusion as a design tool and pinpointing areas for improvement. Collaborators in the Liu Lab at the University of California, Irvine then used their OrthoRep system to dramatically improve the binding strength of several AI-designed antibodies.

“Building useful antibodies on a computer has been a holy grail in science. This goal is now shifting from impossible to routine,” said postdoc Rob Ragotte.

“In this study, we characterized our computer-generated antibodies and gained insights that further improved the design process. This coupling of simulation and real-world experiments is key to advancing protein design,” added Andrew Borst, head of the Electron Microscopy R&D Core at the Institute for Protein Design.

**Figure 5: Cryo-EM structural characterization of two TcdB-binding scFvs**. A) Labeled cryo-EM 2D class averages of a designed scFv, scFv6, bound to TcdB. B) A 3.6 Å cryo-EM 3D reconstruction of the complex shows scFv6 bound to TcdB along the Frizzled-7 epitope. C) Cryo-EM structure of scFv6 bound to TcdB. D) The cryo-EM structure of scFv6 in complex with TcdB closes matches the design model. E) Cryo-EM reveals the accurate design of scFv6 using RFdiffusion (RMSD to the RFdiffusion design of the scFV is 0.9 Å). F) Superposition of each of the six designed scFv6 CDR loop predicted structures as compared to the built cryo-EM structure (RMSD values: CDRH1 = 0.4 Å; CDRH2 = 0.3 Å; CDRH3 = 0.7 Å; CDRL1 = 0.2 Å; CDRL2 = 1.1 Å; CDRL3 = 0.2 Å). G) Comparison of predicted CDRH3 rotamers compared to the built 3.6 Å cryo-EM structure. H) Labeled cryo-EM 2D class averages of the designed scFv, scFv5, bound to full-length TcdB. I) A 6.1 Å cryo-EM 3D reconstruction of the complex shows the scFv5 bound to the target epitope as predicted. J) Due to the modest resolution, a fragment of TcdB was first docked into the cryo-EM density map, and the full design model—including both the TcdB fragment and the designed scFv—was then aligned to the pre-fitted TcdB fragment. This approach demonstrates that the predicted design closely matches the experimentally determined complex in structure, epitope targeting, and overall conformation. (Yellow = TcdB; purple = variable heavy chain fragment (cryo-EM); pink = variable light chain fragment (cryo-EM); gray = computational design prediction)

Free and open source

We are making our latest AI tools for antibody design free for all scientists to use under an MIT License. Code for running the software can be downloaded on GitHub.

The biotechnology company Xaira Therapeutics has exclusively licensed the RFantibody training code from the University of Washington. David Baker, Nate Bennett, Joe Watson, Phil Leung, and Buwei Huang are among the company’s co-founders.

“We are really excited to put this technology in the hands of researchers around the world and to see what they can accomplish with it,” said Joe Watson, a recent postdoc in the lab.