Genki Masuda 増田 元希

Research

• Antigen-Prompted Antibody Sequence Generation and Immune Response Evaluation

  Using antigens as prompts for antibody language models to generate sequences, and evaluating antigen-specific immune responses with the generated sequences.

  Antibody language models can not only identify existing sequences but also generate new ones. I am working on generating sequences from antigen information supplied as a prompt, and on evaluating how well those sequences reflect the response to the target antigen — with antibody design and immune response analysis as the longer-term aim.

  This work has been presented as oral talks at the IPSJ SIG Bioinformatics meetings (SIGBIO): evaluating immune responses using antigen-prompted generated sequences (84th meeting, March 2026), and developing a sequence-conditioned antibody generative model that acquires antigen specificity (85th meeting, July 2026).

  • Antibody Design
  • Sequence Generation
  • Language Model

• Identifying Antigen-Specific Sequences from BCR Repertoires — LM-QASAS

  Developing LM-QASAS, a reference-free framework that identifies antigen-specific sequences from B cell receptor (BCR) repertoires without relying on existing antibody databases.

  The B-cell receptor (BCR) repertoire records the immunological events a person has been through. But existing methods for finding antigen-specific sequences in it need to reference a known antibody database — which rules them out for emerging pathogens, where no such reference exists.

  LM-QASAS removes that dependency. Within the semantic space of an antibody language model (AbLang2), it looks for clusters of sequences whose local clonal density rises transiently after immune stimulation, and treats them as antigen-responsive candidates. In SARS-CoV-2 mRNA vaccinees, those candidates overlapped a known neutralizing-antibody database (CoV-AbDab) far more often than a random baseline, and using them as a pseudo-reference database reconstructed immune-response dynamics without any external reference. Sensitivity was limited in an influenza vaccine cohort, which pins down where the method works best: strong, synchronized clonal expansion such as that induced by mRNA vaccination.

  • BCR Repertoire
  • Antibody Language Model
  • Reference-free
  • Paper (medRxiv)

CV