T-cell clonal organization without TCR sequencing: CloneTrast infers clonal niches from transcriptomes alone
Synopsis
CloneTrast uses clonal identities from paired single-cell RNA and TCR sequencing as training labels, so the model learns T-cell clonal-functional states from gene expression alone and needs no TCR data at inference. Training used 26 datasets across 14 cancer types with more than 1.5 million T cells, and external testing used 11 datasets with more than 1 million T cells, including two cancer types unseen in training. On both training and test sets, patient clonal purity improved significantly over a conventional expression embedding, and one NK-like niche enriched for tumor-blood shared TCRs was associated with poor immunotherapy response across studies.
Interpretation
CloneTrast makes T-cell clonal organization inferable from gene expression alone, without requiring TCR sequences at inference. Earlier multimodal methods require paired gene expression and TCR information for each cell at inference, which limits their use on data with low TCR capture or on large public resources holding only transcriptomes. On the training set, CloneTrast improved the patient clonal purity score over a conventional expression embedding computed from the same genes (p = 2.24·10⁻⁵⁴, 321 patients, paired two-sided Wilcoxon signed-rank test).
The model uses supervised contrastive learning to pull cells of the same clone together and push cells of different clones apart in expression space. TCR sequences serve only to define clonal labels during training and are not model inputs; only the encoder embedding is retained at inference. Training used 26 datasets across 14 cancer types with more than 1.5 million T cells, and external testing used 11 datasets with more than 1 million T cells, including two cancer types unseen in training.
On the independent test set, the pretrained model yields clonal structure without retraining and recapitulates the main niche-composition patterns seen in training. The test data were entirely external to training and hyperparameter tuning, and the model aligns new data to a fixed gene space. Test-set patient clonal purity also improved significantly over the expression embedding (p = 2.47·10⁻⁵⁵, 327 patients), and predicted clone size correlated with observed clone size at Spearman r = 0.459.
Perspective
The model was trained and evaluated on tumor and blood samples from human cancer patients; the external test set comprises 11 studies, 488 samples and more than 1 million T cells, including two cancer types absent from training. The authors state that generalizability to normal tissues, lymphoid tissues, autoimmune disease, infection or non-human systems remains to be tested.
The authors note that the identified niches depend on the learned representation and on the clustering resolution, with different resolutions revealing broader or finer levels of organization; the model does not directly infer TCR sequence, clonotype identity or antigen specificity, so antigen-related conclusions need independent validation by paired TCR sequencing or antigen-binding assays. Clinical associations were found across public datasets heterogeneous in cancer type, treatment regimen, response definition and sampling time, are described as hypothesis-generating, and some individual studies showed contradicting trends.