IgFold, developed by Ruffolo et al., is one of the state-of-the-art methods for antibody (Ab) structure prediction. IgFold is an end-to-end approach based on two key components: the protein language model AntiBERTy trained on 558 million sequences of naturally occurring Abs, and graph neural networks for Ab backbone structure prediction. In addition to being able to predict structures with comparable or better quality than existing methods, IgFold also boasts of being particularly fast in making predictions.
The accuracy and speed of IgFold facilitate the investigation of challenges that were unreachable by other methods. The authors utilized IgFold to predict the structures of 1.4 million Ab sequences. This feat enables 500-fold structural insights of Abs than is feasible with only experimental structural determination.
Abs are central to the adaptive immune system-based body's response to and neutralization of antigens (Ags). Although there exists massive sequencing data on natural Abs, sequence data by itself does not allow sufficient insights into Ab investigations and therapeutic Ab development. Accurate structural modeling of Ab, especially the complementarity-determining regions (CDRs), would facilitate the understanding of Ag binding and rational design of Ag-specific Abs.
The prediction of protein monomers is becoming a routinely successful task using methods like AlphaFold3 (AF3) and RoseTTAFold. Multimeric structure prediction has also seen huge progress especially with AF2-Multimer and potentially with AF3. However, accurate Ab structure prediction remains elusive.
Compared to AF2-Multimer, IgFold demonstrates similar accuracy but faster predictions. Unlike many other Ab structure prediction methods, IgFold enables utilization of template structures, allows nanobody modeling, and provides per-residue accuracy measures. Structures predicted by IgFold are further fed to Rosetta for molecular geometry refinement and filling of missing atoms (side chains). However, based on benchmark tests, the backbones of the structures before and after refinement do not deviate by more than 0.5 Å.
References
Paper: Fast, accurate antibody structure prediction from deep learning on massive set of natural antibodies
Code: https://github.com/Graylab/IgFold
Check our LinkedIn post on tFold: tFold: Fast and Accurate Prediction of Structures of Antibodies and Antibody-Antigen Complexes
