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

T cells play a critical role in adaptive immunity by safeguarding the body against infections and cancer. T cell receptors (TCRs), located on the surface of T cells, facilitate the identification of target cells through interactions with peptides displayed by major histocompatibility complex (pMHC) molecules. Constructing a comprehensive and reliable map of TCRs to their corresponding pMHCs is a major goal in systems immunology, with significant impacts on cancer immunotherapy, vaccine development, and the management of autoimmune diseases.

Due to the vast combinations of TCR and pMHC sequences, exhaustive experimental mapping is currently unfeasible. As a result, computational models have emerged as a promising approach to tackle this challenge. Over time, these models have become increasingly advanced, ranging from small neural networks to deep generative models and large protein language models.

Despite their complexity, however, several independent evaluations have demonstrated that these models often struggle to generalize to TCRs or pMHCs not included in the training data. This study aims to explore the reasons behind this limitation and propose a roadmap for developing generalizable TCR:pMHC prediction models. By examining the largest public repositories of TCR:pMHC data, the researchers identified several biases, including a limited focus on specific antigens and MHC molecules, sparse sampling of TCR sequences per pMHC, and inconsistencies and inaccuracies in data compilation.

To address these issues, the study emphasizes the importance of designing model evaluation methods that prioritize antigen specificity over other potentially confounding factors present in the evaluation data. Furthermore, the researchers argue that the choice of evaluation metric should be tailored to the intended downstream application of the predictions.

Ultimately, the findings suggest that the immunology community must strive for more representative and better curated datasets, along with a robust set of metrics and benchmarks, to facilitate the development of generalizable TCR:pMHC prediction models.