Antibodies are among the most powerful tools in modern medicine, but their binding regions rarely exist as a single fixed shape. Instead, they can sample multiple conformations that influence how they recognize targets and how they behave as therapeutic candidates. Capturing this structural heterogeneity is an important challenge for antibody design.
In a study published in Protein & Cell, Yang Cao’s team at Sichuan University reports AbCVista, a deep learning framework designed to predict not only accurate antibody structures, but also the conformational ensembles antibodies can adopt in solution. The work focuses on CDR-H3, a highly variable loop that often dominates antigen recognition and remains one of the most difficult antibody regions to model.
“A single structure is like one photograph of an antibody,” said a member of the research team. “But one photograph may miss other important poses. What matters for recognition and developability is often the set of conformations an antibody can sample. With AbCVista, we wanted to capture that fuller structural picture.”
AbCVista combines antibody-specific sequence search and alignment with customized multiple sequence alignment sampling and an optimized OpenFold modeling stage. For ensemble generation, the team developed a cross-chain clustering strategy that learns heavy- and light-chain patterns separately, then reconstructs biologically coherent paired conformational states. This helps avoid a common artifact in multi-chain modeling, where sequences cluster by chain type rather than meaningful antibody pairing.
In benchmark tests against leading methods including AlphaFold3, AlphaFold2, ImmuneBuilder, DeepAb, IgFold, ColabFold and AF-Cluster, AbCVista showed strong performance in both static structure prediction and ensemble generation. On recently released SAbDab antibody structures, it achieved a CDR-H3 RMSD of 2.117 Å for paired antibodies, outperforming AlphaFold3 and IgFold. In a multi-conformation benchmark containing 167 experimentally observed CDR-H3 states, AbCVista recovered more experimental conformations than competing approaches, covering 49.1% at a 2.0 Å threshold. The same ensemble-based view also improved prediction of aggregation-related behavior in a dataset of 137 antibody drugs, where hydrophobic surface features derived from AbCVista ensembles performed better than features from single static structures.
DOI：10.1093/procel/pwag040