Unravelling T Cell Recognition Specificity: One Epitope at a Time

In this webinar, we explore a landmark study from the Gfeller Lab at the University of Lausanne that takes a fresh look at how T cells recognise their targets — one epitope at a time. T cell receptors (TCRs) sit at the heart of our immune system's ability to detect infected and cancerous cells. But understanding how they recognise specific targets has long been hampered by the sheer complexity of the TCR sequence space and the flexibility of the TCR-epitope interface. This talk unpacks a new framework that cuts through that complexity. The study introduces TCR Specificity Profiles (TSPs) — an interpretable, probabilistic visualisation tool that compares epitope-specific TCRs against large baseline repertoires to reveal the key molecular signatures driving recognition. Using TSPs, the researchers show that much of TCR specificity is actually encoded in V and J gene usage — not just the CDR3 loop sequences that have traditionally received the most attention. The talk covers how TSPs can: Predict cross-reactivity between epitopes, including clinically relevant cases like the MAGE-A3/TITIN example that caused lethal toxicity in cancer trials Track how specificity evolves as peptide sequences, binding modes, or MHC alleles change Interpret and stress-test machine learning TCR-epitope prediction tools Leverage AlphaFold3 to approximate specificity profiles even without experimental structural data The team also introduces TEMPO, a fast and interpretable TCR-epitope interaction predictor that matches or outperforms state-of-the-art deep learning tools — while scoring a million TCRs in under a minute on a standard CPU. Whether you work in immunology, computational biology, or cancer immunotherapy, this talk offers both conceptual clarity and practical tools for understanding and predicting T cell recognition. 🔗 Explore the TCR Motif Atlas: https://tcrmotifatlas.unil.ch 🔗 TEMPO tool: https://github.com/GfellerLab/TEMPO 🔗 MixTCRcross: https://github.com/GfellerLab/MixTCRc...