March 18, 2024 | Nathaniel R. Bennett†1,2,3, Joseph L. Watson*†1,2, Robert J. Ragotte†1,2, Andrew J. Borst†1,2, Déjenaé L. See†1,2,4, Connor Weidle#1,2, Riti Biswas1,2,3, Ellen L. Shrock1,2, Philip J. Y. Leung1,2,3, Buwei Huang1,2,4, Inna Goreshnik1,2,5, Russell Aull6,7, Kenneth D. Carr2, Benedikt Singer1,2, Cameron Criswell1,2, Dionne Vafeados2, Mariana Garcia Sanchez2, Ho Min Kim8,9, Susana Vázquez Torres1,2,10, Sidney Chan2, David Baker*1,2,5
The study demonstrates the capability of a fine-tuned RFdiffusion network to design de novo antibody variable heavy chains (VHHs) that bind to user-specified epitopes. The network is trained on a fine-tuned version of AlphaFold2 and RosettaTFAfold2, which are adapted for antibody structure design and prediction. The designed VHHs are experimentally validated against four disease-relevant epitopes, and the cryo-EM structure of a designed VHH bound to influenza hemagglutinin is nearly identical to the design model. The results highlight the potential of computational methods to revolutionize antibody discovery and development by enabling rational design of antibodies targeting specific epitopes with high affinity and specificity.The study demonstrates the capability of a fine-tuned RFdiffusion network to design de novo antibody variable heavy chains (VHHs) that bind to user-specified epitopes. The network is trained on a fine-tuned version of AlphaFold2 and RosettaTFAfold2, which are adapted for antibody structure design and prediction. The designed VHHs are experimentally validated against four disease-relevant epitopes, and the cryo-EM structure of a designed VHH bound to influenza hemagglutinin is nearly identical to the design model. The results highlight the potential of computational methods to revolutionize antibody discovery and development by enabling rational design of antibodies targeting specific epitopes with high affinity and specificity.