A huge thank you to all collaborators, including Yan Liu, Giancarlo Croce, Dana Moreno, Daniel Tadros, Julien Racle, Anne-Christine Thierry, Petra Baumgartner, Alexandra Michel, Vincent Zoete and Alexandre Harari

The associated TCR-Epitope Motif-based interaction Predictor (TEMPO) can be accessed at github.com/GfellerLab/T... and can score a million TCRs in one minute on a standard CPU.

The TSP for hundreds of epitopes can be explored at the TCR Motif Atlas (tcrmotifatlas.unil.ch/home), including many sequence, length and MHC restriction variants of an immunodominant epitope.

Most of this work is based on experimentally identified TCRs, but we also we demonstrate that TSPs can be approximated with AlphaFold3 by scoring baseline TCR repertoires and focusing on the best scoring TCRs, thereby paving the way for in silico sorting of epitope-specific TCRs.

This does not mean that CDR3 sequences are not important for epitope recognition. Instead it reflects that a large fraction of CDR3 residues (including many directly contacting the epitope in X-ray structures) are determined by V/J usage.

In particular, our results demonstrate that much of the specificity in TCR-epitope recognition is encoded in V/J usage, with many epitopes showing up to ten-fold enrichment in specific V or J genes.

We show that TSPs unravel key determinants of TCR-epitope recognition specificity, predict cross-reactivity and reveal how TCR specificity evolves with epitope sequence, binding mode and MHC restriction.

Here we introduce a fully interpretable framework, called TCR Specificity Profiles or TSPs, to model TCR-epitope recognition.

Recent developments in machine learning and structure-based approaches (e.g., AF3) are revolutionizing TCR-epitope interaction predictions. However, interpreting such predictions to understand how specificity is encoded in TCR sequences remains challenging.

Beautiful work and very useful resource!