Atomically accurate de novo design of antibodies with RFdiffusion
Nathaniel R. Bennett, Joseph L. Watson, Robert J. Ragotte, Andrew J. Borst, DéJenaé L. See, Connor E. Weidle et al. · 2025 · 305 citationsRead the paper
Abstract Despite the central role of antibodies in modern medicine, no method currently exists to design novel, epitope-specific antibodies entirely in silico. Instead, antibody discovery currently relies on immunization, random library screening or the isolation of antibodies directly from patients 1 . Here we demonstrate that combining computational protein design using a fine-tuned RFdiffusion 2 network with yeast display screening enables the de novo generation of antibody variable heavy chains (VHHs), single-chain variable fragments (scFvs) and full antibodies that bind to user-specified epitopes with atomic-level precision. We experimentally characterize VHH binders to four disease-relevant epitopes. Cryo-electron microscopy confirms the binding pose of designed VHHs targeting influenza haemagglutinin and Clostridium difficile toxin B (TcdB). A high-resolution structure of the influenza-targeting VHH confirms atomic accuracy of the designed complementarity-determining regions (CDRs). Although initial computational designs exhibit modest affinity (tens to hundreds of nanomolar K d ), affinity maturation using OrthoRep 3 enables production of single-digit nanomolar binders that maintain the intended epitope selectivity. We further demonstrate the de novo design of scFvs to TcdB and a PHOX2B peptide–MHC complex by combining designed heavy-chain and light-chain CDRs. Cryo-electron microscopy confirms the binding pose for two distinct TcdB scFvs, with high-resolution data for one design verifying the atomically accurate design of the conformations of all six CDR loops. Our approach establishes a framework for the computational design, screening and characterization of fully de novo antibodies with atomic-level precision in both structure and epitope targeting.
What the full text shows
Working prototype- De novo design of VHHs targeting disease-relevant epitopes with atomic-level precision. — RMSD = 1.45 Å for VHH backbone vs. design (influenza haemagglutinin) (Cryo-EM validation of VHH_flu_01 binding to influenza haemagglutinin stem epitope)
- De novo design of scFvs targeting the Frizzled epitope of TcdB with near-atomic precision across all six CDRs. — Backbone RMSDs: CDRH1=0.4Å, CDRH2=0.3Å, CDRH3=0.7Å, CDRL1=0.2Å, CDRL2=1.1Å, CDRL3=0.2Å (Cryo-EM structure of scFv6 bound to TcdB)
- Initial de novo VHH binders exhibit modest affinity, which can be improved via affinity maturation. — Initial Kd: tens to hundreds of nM; matured Kd: single-digit nM or sub-nanomolar (Using OrthoRep for continuous hypermutation in yeast display)
- Successful design of full-length IgG1 antibodies from de novo scFv designs. — Kd = 68 nM (Conversion of scFv6 (Kd=72 nM) to full-length IgG1 targeting TcdB Frizzled epitope)
- VHH design success rates vary by target. — 0% to 2% success rate (Screening of 9,000 designs per target via yeast display or 95 via SPR)
- Setup
- VHHs designed using h-NbBcII10FGLA framework; scFvs created via combinatorial pairing of heavy and light chains from structurally similar designs. Targets included TcdB, influenza haemagglutinin, RSV, SARS-CoV-2 RBD, and PHOX2B-MHC.
- How
- Fine-tuned RFdiffusion to generate antibody CDR loops and rigid-body docking based on framework and epitope conditioning, followed by ProteinMPNN for sequence design. Experimental validation used yeast surface display, SPR, and cryo-EM, with OrthoRep for affinity maturation.
- Limitations
- Low initial design success rates (0-2%) requiring high-throughput screening. · Designed sequences are less 'human' (lower OASis score) than therapeutic antibodies, potentially increasing immunogenicity. · Initial affinities are modest (tens to hundreds of nM) before maturation.
Summarized by Gemma 4 from the open-access full text · 9 Oct 2026. Read it