Beyond Better Models: Barriers and a Path Forward in Generalisable TCR Antigen Specificity Prediction

T cells are at the centre of adaptive immunity, protecting the organism against infection and cancer. T cell receptors (TCRs), expressed on the surface of T cells, enable the recognition of target cells through interactions with peptides presented by major histocompatibility complex (pMHC) molecules. Reliable, large-scale mapping of TCRs to their cognate pMHCs remains a holy grail of modern systems immunology, with far-reaching implications for cancer immunotherapy, vaccine development, and the understanding and treatment of autoimmune diseases. While exhaustive experimental mapping remains intractable due to the enormous combinatorial sequence space of TCRs and pMHCs, computational approaches offer a promising alternative. Consequently, we have witnessed increasingly sophisticated computational models of TCR antigen specificity, from small neural networks to deep generative models and large protein language models. However, several independent benchmarks have shown that these models, despite their size and sophistication, largely fail to generalise to TCRs or pMHCs not present in the data available during model development. Here, we investigate possible reasons for this and suggest a path forward to generalisable TCR:pMHC prediction. First, our analysis of the largest public repositories of TCR:pMHC data reveals significant biases towards a small number of antigens and MHC molecules, sparse sampling of TCR sequence space per pMHC, as well as inconsistencies and inaccuracies in data collation. Second, we show the importance of carefully setting up model evaluation to measure discrimination based on antigen specificity instead of confounding factors present in the evaluation data. Finally, we highlight the need for and propose how best to match the evaluation metric with the specific downstream application of the predictions. Our investigation suggests that we, as a community, need more representative and better curated data, and a robust set of metrics and benchmarks to guide future model development towards generalisable TCR:pMHC specificity prediction.