Abstract
Accurate TCR-epitope interaction predictions have the potential to unlock the use of TCR repertoires for diagnostics, TCR discovery, and cross-reactivity predictions. In this issue of Cell Genomics, Drost et al.1 developed a streamlined framework for benchmarking such predictions. Careful understanding of the strengths and limitations of existing approaches will be instrumental to improve them and expand the scope of TCR-epitope recognition predictions.
Accurate TCR-epitope interaction predictions have the potential to unlock the use of TCR repertoires for diagnostics, TCR discovery, and cross-reactivity predictions. In this issue of Cell Genomics, Drost et al. developed a streamlined framework for benchmarking such predictions. Careful understanding of the strengths and limitations of existing approaches will be instrumental to improve them and expand the scope of TCR-epitope recognition predictions.
Main text
The prediction of T cell receptor (TCR)-epitope interactions remains one of the great challenges in immunology, with applications spanning vaccine design, TCR discovery for TCR-T cell therapy, and cross-reactivity predictions. The development of dedicated databases,2,3 combined with the availability of powerful and user-friendly machine-learning packages in many different programming languages, has made training predictors of TCR-epitope recognition on publicly available data accessible to most scientists with some programming skills. Rigorous benchmarking of these predictive tools is therefore of critical importance.
In their paper,1 Drost et al. leveraged two recent datasets—neither of which has been used in the training of existing methods—to systematically evaluate the predictive power of 21 TCR-epitope interaction predictors. The first dataset comprises hundreds of TCRs interacting with 14 epitopes restricted to five distinct major histocompatibility complexes (MHCs). These epitopes include both widely studied epitopes and epitopes with very little to no known TCRs (i.e., the so-called “unseen” epitopes). The second dataset consists of single-amino-acid variants of well-studied epitopes for which TCR recognition was assessed using TCRs recognizing the parent epitopes. This benchmarking effort is accompanied by an integrated interface (ePytope-TCR) that simplifies the use of the evaluated tools on user-provided data and is compatible with multiple data formats. This will facilitate exploration of success and limitations of current methods in different contexts.
Several important conclusions emerge from this study. First, the results confirm that quite accurate predictions can be achieved for a subset of epitopes, which consist mainly of immunodominant viral epitopes and for which abundant training data are available. Identifying these epitopes is essential for guiding the practical application of TCR-epitope interaction prediction tools and avoiding common pitfalls. For instance, the annotation of TCR repertoires to detect enrichment in TCRs recognizing epitopes from a specific pathogen may be feasible for epitopes with very accurate predictions but is likely not worth undertaking for epitopes with low prediction accuracy.
Second, the results demonstrate that the predictive power of current machine-learning tools remains very limited for unseen epitopes. This observation applies even to single-amino-acid variants of epitopes with abundant training data. This finding, which aligns with results from other recent benchmarking efforts,4 is particularly significant in light of the numerous publications claiming broad generalizability of machine-learning TCR-epitope interaction predictions across epitopes, MHCs, and even species. While this may initially seem discouraging, a clear understanding of the limitations of today’s algorithms is essential for designing the next generation of experiments and tools addressing these limitations.
An additional important observation is that, as of today, machine-learning predictors likely implicitly treat epitopes as categorical features. This is supported by the fact that pan-epitope (referred to as “general” by Drost et al.) tools did not show improved performance compared with epitope-specific (referred to as “categorical”) tools in this benchmark. This observation could inform the design of future machine-learning TCR-epitope prediction algorithms where one has to balance the complexity of the architecture and the risk of overfitting.
Some remaining questions have not been directly addressed in this work. These include, for instance, the choice of negatives and especially how much existing tools using TCRs binding to other epitopes as negatives are at risk of “learning the negatives.” Another important question is the extent to which false positives in existing databases impact prediction accuracy.5 We also suggest that future benchmarking efforts should include structure-based approaches, which may offer predictive power even for unseen epitopes, as these methods do not depend on the availability of extensive training data.6 The robust ePytope-TCR benchmarking framework developed in this study will help address some of these questions. This will be useful for guiding the design of both future experimental strategies and improved algorithms to address the current limitations of TCR-epitope recognition predictions.
Declaration of interests
The author declares no competing interests.
References
- 1.Drost F., Chernysheva A., Albahah M., Kocher K., Schober K., Schubert B. Benchmarking of T cell receptor-epitope predictors with ePytope-TCR. Cell Genom. 2025;5 doi: 10.1016/j.xgen.2025.100946. [DOI] [PubMed] [Google Scholar]
- 2.Goncharov M., Bagaev D., Shcherbinin D., Zvyagin I., Bolotin D., Thomas P.G., Minervina A.A., Pogorelyy M.V., Ladell K., McLaren J.E., et al. VDJdb in the pandemic era: a compendium of T cell receptors specific for SARS-CoV-2. Nat. Methods. 2022;19:1017–1019. doi: 10.1038/s41592-022-01578-0. [DOI] [PubMed] [Google Scholar]
- 3.Vita R., Blazeska N., Marrama D., Duesing S., Bennett J., Greenbaum J., De Almeida Mendes M., Mahita J., Wheeler D.K., et al. IEDB Curation Team Members The Immune Epitope Database (IEDB): 2024 update. Nucleic Acids Res. 2025;53:D436–D443. doi: 10.1093/nar/gkae1092. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Nielsen M., Eugster A., Jensen M.F., Goel M., Tiffeau-Mayer A., Pelissier A., Valkiers S., Martínez M.R., Meynard-Piganeeau B., Greiff V., et al. Lessons learned from the IMMREP23 TCR-epitope prediction challenge. ImmunoInformatics. 2024;16 doi: 10.1016/j.immuno.2024.100045. [DOI] [Google Scholar]
- 5.Messemaker M., Kwee B.P.Y., Moravec Ž., Álvarez-Salmoral D., Urbanus J., De Paauw S., Geerligs J., Voogd R., Morris B., Guislain A., et al. A functionally validated TCR-pMHC database for TCR specificity model development. bioRxiv. 2025 doi: 10.1101/2025.04.28.651095. Preprint at. [DOI] [Google Scholar]
- 6.Bradley P. Structure-based prediction of T cell receptor:peptide-MHC interactions. eLife. 2023;12 doi: 10.7554/eLife.82813. [DOI] [PMC free article] [PubMed] [Google Scholar]