Abstract 4196: Predicting self-antigen recognition of TCRs derived from tumor infiltrating CD8+ T cells via contrastive learning and T cell phenotyping

Abstract T cell receptor (TCR)-based immunotherapies leverage the ability of a TCR to bind with high specificity and selectivity to peptides derived from intracellular targets, including those overexpressed on tumors. Consequently, TCRs identified from tumor-infiltrating lymphocytes (TILs) that respond to antigens in the tumor microenvironment represent attractive substrates for TCR immunotherapy development. However, de-orphaning TCRs is exceedingly challenging given the diversity of TCR and target sequence space. Although multiplexed peptide-MHC (pMHC) staining assays have improved throughput of direct, ex vivo measurement of TCR-pMHC binding, they are limited by TIL sample quality, T cell infiltration, TCR repertoire diversity, and limited pMHC library size. Machine-learning based techniques may overcome these limitations by training models to predict TCR-pMHC interactions and then inferring TCR-pMHC binding on large databases collected in other contexts. Here, we applied a contrastive model of TCR-pMHC binding prediction (1) to 382 tumor samples from 9 indications containing 490,000 T cells. We found that 5,244/490,000 (∼1.1%) of TCRs from this dataset were predicted to recognize one of the 4,438 pMHCs in our model training set. Of these, 1,949 were predicted to bind common viral peptides, an observation indicative of bystander recruitment and infiltration. An additional 2,549 TCRs were predicted to bind to Class I onco-antigens, including QLLALLPSL PRAME, YLEPGPVTA GP100 and GLYDGMEHLI MAGEA10. The model was able to capture TIL TCRs with diverse complementarity-determining regions (CDRs) as well as highly homologous sequences. TCRs predicted to bind self-antigens were phenotyped using Repertoire’s DECODE™ platform (2), showing that cancer-specific TCRs tended to have a more cytotoxic and effector like phenotype compared to a memory-like signature on viral specific TCRs (e.g. TCRs specific to cytomegalovirus, Epstein-Barr virus, and influenza). These results demonstrate a way to computationally de-orphan TCR specificities and phenotypically characterize T cells in a high-throughput manner to further our understanding of the TCR landscape in cancer patients. Importantly, this strategy could enable the discovery of TCR sequences suitable for therapeutic development using existing datasets. Future work should focus on increasing the diversity of validated TCR-pMHC specificities and in-vitro validation of predicted TCR-pMHC interactions (as well as reported model performance metrics).1. Abel, John, et al. "REPTRA: Mapping Immune T Cell Receptor Activity from Full Sequences with a Debiased Contrastive Loss." bioRxiv (2025): 2025-10.2. Francis, Joshua M., et al. "Allelic variation in class I HLA determines CD8+ T cell repertoire shape and cross-reactive memory responses to SARS-CoV-2." Science immunology 7.67 (2021): eabk3070. Citation Format: Brinda Vijaykumar, Qiaomu Tian, Jack Prazich, Preet Joshi, Neel Patel, John Abel, Anthony Coyle, Daniel Pregibon, . Predicting self-antigen recognition of TCRs derived from tumor infiltrating CD8+ T cells via contrastive learning and T cell phenotyping abstract. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 4196.