IMMUNOLOGY2026™ Conference Recordings For Attendees-AI and decoding the TCR repertoire

Video Transcription

Video Summary

The speaker discusses the challenge of predicting what antigen a T cell receptor (TCR) recognizes from sequence alone. They distinguish between a practical problem—identifying targets for known TCRs—and the “zero-shot” holy grail of predicting specificity for any TCR or peptide-MHC pair. Most existing machine-learning models fail in true unseen testing, often learning artifacts of the data rather than biology.  <br /><br />However, structural models such as Hermes and TCRdoc show promising partial success, especially when trained on real binding or activation data. The speaker presents new results from SARS-CoV-2 vaccine responses, where clustering TCRs and using HLA context enabled a correct peptide prediction for one cluster.  <br /><br />They also describe a new strategy to generate curated negative data by pairing known antigen-specific chains with many donor-derived partners. This reveals that single-chain data alone is usually insufficient to infer specificity. Overall, progress is real, but the zero-shot decoding problem remains unsolved.

Meta Tag

Date April 17, 2026 4:30 PM - 4:50 PM

Room 156

Session Resource Highlight: Artificial Intelligence (AI) in Research

Speaker Paul Thomas

Track Computational and Systems Immunology (COMP)

Year 2026

Keywords

T cell receptor

antigen specificity

zero-shot prediction

structural models

SARS-CoV-2 vaccine responses

April 17, 2026 4:30 PM - 4:50 PM

156

Resource Highlight: Artificial Intelligence (AI) in Research

Paul Thomas

Computational and Systems Immunology (COMP)

2026