Antibody Fv structure predictor that turns AntiBERTy language-model embeddings into backbone coordinates with graph networks in seconds.
No providers recorded yet. Browse all providers
An antibody binds its target through six hypervariable loops, and five of them fall into a small set of canonical shapes that can be predicted from similar known structures. The sixth, the heavy-chain CDR H3 loop, is created by gene recombination and mutation, varies widely in length and sequence, and sits at the interface between the heavy and light chains. It is the part of the structure that matters most for binding and the part that template-based methods predict worst. General structure predictors such as AlphaFold-Multimer handle it better but need a multiple sequence alignment and minutes per antibody, which rules out folding the millions of sequences a repertoire experiment produces.
IgFold, developed by Jeffrey Ruffolo, Lee-Shin Chu, Sai Pooja Mahajan and Jeffrey Gray in the Gray lab at Johns Hopkins University, removes the alignment step. It reads the antibody sequence with AntiBERTy, a language model pretrained on 558 million natural antibody sequences, and uses the embeddings and attention maps as a stand-in for evolutionary information. Graph networks then place every residue's backbone directly in 3D space. The method appeared as a bioRxiv preprint in April 2022 and was published in Nature Communications in April 2023.
The structure module has four graph transformer layers interleaved with triangle multiplicative edge updates from AlphaFold, followed by two invariant point attention (IPA) layers that inject template coordinates and three IPA layers that assemble residue frames into the final structure. IgFold itself has 1.6M trainable parameters on top of the roughly 26M-parameter AntiBERTy encoder. Because only about 4,275 experimental antibody structures were available in SAbDab, training was augmented with AlphaFold-predicted structures for 16,141 paired and 22,132 unpaired OAS sequences, sampled so each source supplied about a third of examples. Four models were trained from different random seeds for about 110 hours each on one A100 GPU. At inference all four run and the prediction with the lowest 90th-percentile predicted error is kept.
On a benchmark of 197 paired antibodies and 71 nanobodies deposited after July 2021, IgFold reached a mean CDR H3 backbone RMSD of 3.27 Å, compared with 3.56 Å for AlphaFold-Multimer, 3.54 Å for ABlooper and 3.57 Å for DeepAb, with framework RMSDs under 0.5 Å. For nanobody CDR3 loops it scored 4.25 Å against 4.00 Å for AlphaFold2. Full-atom prediction averaged 23 seconds per paired antibody, against about 7 minutes for the ColabFold implementation of AlphaFold-Multimer.
IgFold is a working tool for antibody engineers who need structures for many candidates: scoring designed or humanized variants, supplying models for docking, or filtering sequences whose predicted loops carry high error before committing to synthesis. Its template option fits lead optimization, where a crystal structure of the parent is known and only mutated loops need predicting. Immunologists can use it to add structure to repertoire sequencing data. The authors folded 1.4 million paired antibodies this way, 104,994 non-redundant sequences from OAS and 1.3 million paired human antibodies from four donors, and released both sets for download.
IgFold showed that a pretrained antibody language model can substitute for multiple sequence alignments in antibody structure prediction, trading a small accuracy loss on the hardest nanobody loops for a speedup of more than an order of magnitude. Later predictors such as FlowAb reuse its 197-antibody benchmark split, ImmuFold and others report against it as a baseline, and it serves as a fast folding step inside design pipelines including AffinityFlow and NanoDesigner. Its limits are the ones its authors report: it predicts only variable domains, not antibody-antigen complexes; its predicted error tends to underestimate true CDR H3 error; and long, structured nanobody CDR3 loops remain better handled by AlphaFold. Code and weights are free for non-commercial use under the JHU Academic Software License, with commercial licensing through Johns Hopkins Technology Ventures.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.