Revealing pan-cancer clonal niches of T cells from single-cell RNA sequencing using contrastive learning
TL;DR
CloneTrast enables scalable and interpretable inference of clonal organization from single-cell transcriptomes and uses paired scRNA/TCR-seq during training to learn clonal-functional states from gene expression, while requiring only scRNA-seq at inference.
Abstract
T-cell clonal organization is central to anti-tumor immunity, yet its functional characterization typically requires paired single-cell RNA and TCR sequencing (scRNA/TCR-seq). Existing multimodal methods integrate transcriptome and TCR information but are limited when TCR data are unavailable. Here we present CloneTrast, a supervised contrastive-learning framework that uses paired scRNA/TCR-seq during training to learn clonal-functional states from gene expression, while requiring only scRNA-seq at inference. Using paired scRNA/TCR-seq profiles from 2.6 million human T cells across 16 cancer types, CloneTrast identified distinct T-cell clonal niches not explained solely by TCR sequence similarity or conventional cell-state annotations. Among these, an NK-like T-cell niche enriched for tumor-blood shared TCRs was consistently associated with poor immunotherapy response at the pan-cancer level. Analysis of antigen-specific T-cell clones revealed that tumor-resident and blood-borne antigen-specific T cells were preferentially associated with distinct clonal niches. Overall, CloneTrast enables scalable and interpretable inference of clonal organization from single-cell transcriptomes.