Summary
A bioRxiv preprint introduces LangAAI, a model that uses antibody and antigen sequences to model antibody hypervariable regions without requiring 3D structures. The authors report that changing the antigen input to a non-binding decoy particularly affects sequence recovery in the CDR3 loop.
Researchers at the University of California, Riverside, have introduced LangAAI, a masked language model that uses antibody and antigen sequences to model antibody hypervariable regions. The bioRxiv preprint, posted on October 24, 2026, reports that the model can incorporate antigen context without requiring a 3D structure, a docking pose or an annotated binding site.
Why antigen context matters
Antibodies recognise targets through regions called complementarity-determining regions, or CDRs. These parts of an antibody vary and help determine which antigen it can bind. The CDR3 loop is particularly variable, making it an important region for antibody design.
A model that considers an antibody sequence without its intended target may miss information relevant to selecting or changing these regions. LangAAI is designed to condition on both the antibody’s and antigen’s primary sequences—their strings of amino acids—rather than on a structural model of how the two molecules fit together. As a masked language model, it works with sequence context to model amino-acid patterns.
A decoy test probes whether the model uses the antigen
To test whether antigen input influences its predictions, the researchers replaced the matching antigen with a non-binding decoy. They report that antigen-agnostic models were largely insensitive to this substitution, while LangAAI showed a marked reduction in sequence recovery, concentrated in the CDR3 loop. This result indicates that the model’s outputs respond to which antigen sequence is supplied.
Across benchmarks, the authors report higher amino-acid recovery than existing antibody language models. They also report performance matching structure-based methods under those methods’ respective evaluation protocols. In a separate comparison, LangAAI ranked experimentally measured antibody variants by affinity comparably to inverse-folding models that use backbone coordinates.
Potential use and evidence limits
The approach could help prioritise antibody variants for targets whose antigen–antibody complexes have not been structurally solved, because it requires sequence information rather than a solved complex. The reported role is computational prioritisation: experimental measurements would still be needed to establish how a particular candidate performs.
These findings are from a bioRxiv preprint. The abstract reports comparative benchmark outcomes but gives no numerical scores, so the size of the reported performance differences cannot be assessed from the supplied text. It also notes that comparisons with structure-based methods used their respective protocols, rather than one shared evaluation setup.