HELP-TCR: Harmonized Explainable Language Processing Toolkit for T Cell Antigen Receptor Repertoires

Functional characterization of T cell antigen receptor (TCR) repertoires is critical for understanding adaptive immune responses across diverse contexts, including infectious diseases, cancer, autoimmune conditions, and allergic disorders. Detailed analysis of TCR repertoires can reveal disease-specific signatures and support biomarker discovery and the development of immunotherapies as well as vaccines. Current computational approaches often prioritize global repertoire metrics or employ deep learning models that, while powerful, offer limited interpretability. Here, we present HELP-TCR, a novel machine learning framework based on natural language processing that integrates low-dimensional, explainable feature extraction with robust sample classification performance. HELP-TCR aims to classify the per-sample TCR repertoires via a nonparametric probabilistic approach. TCR repertoires are represented by modeling within sample position-specific distributions of single amino acids and amino acid pairs, transforming sequences into multidimensional tensor structures. To increase reproducibility, a consensus grouping method is proposed to merge the features with highly similar position-wise distributions. A modified ResNet-18 deep learning architecture, adapted to process these tensors, enables accurate sample classification, while post hoc analysis based on saliency map highlights the most informative features contributing to model predictions. Using a dataset of bootstrapped TCR sequences, HELP-TCR achieved an area under the curve (AUC) of 0.96, outperforming existing methods including DeepTCR (AUC 0.76) and TCR-BERT embeddings, which exhibited limited class separability. Beyond performance, HELP-TCR enables identification of position-specific amino acid motifs (pairs) associated with sample TCR repertoire classification decisions, offering biologically interpretable insights into TCR repertoire differences. By emphasizing model interpretability alongside predictive accuracy, HELP-TCR provides a versatile platform for functional TCR repertoire analysis with potential applications in immunotherapy development, vaccine design, and immune monitoring.