Predicting the cognate epitope target of a T-cell receptor (TCR) with machine learning has been the subject of many academic studies. It is also something we heavily work on at ImmuneWatch.

Figure 3 from the manuscript.

However several questions on the advantages and disadvantages of these different approaches remain unresolved, as most methods have only been evaluated within the context of their initial publications and data sets. In an effort to make these different approaches more comparable, our CTO Pieter Meysman (also affiliated to the University of Antwerp) joined forces with other academics during an Immune Repertoire workshop (ImmRep22) TCR-specificity workshop.

They motivate the reason for this workshop in the manuscript as follows:
“[Because epitope recognition [by T cells] is crucial for pathogen defense, vaccine response, tumor control and autoimmune diseases and since TCR specificity helps understanding the function of a T cell, it is essential to learn to decipher it.”

Some of the results include:

But mainly:
There is a large need for a true independent benchmark on the myriad of methods within the field. As already seen in other prediction fields such as the CASP competition (where AlphaFold 2 showed to be a game-changer), such an independent benchmark could really make a difference.

The biggest challenge however, is generating a good unpublished dataset. As highlighted in the study, it would ideally involve paired alpha-beta TCR sequence data, and would therefore likely be derived from a single cell sequencing experiment. In addition, the use of oligo-tagged multimers
would enable both identification of those TCRs that are specific for an epitope, along with those that are likely not.
Read the manuscript here: