scXpand

Paired single-cell RNA and T-cell receptor sequencing is the standard route to identifying expanded T-cell clones and linking them to the transcriptional states that drive anti-tumor immunity. But scTCR-seq is absent from most published single-cell datasets, and where present, TCR capture varies unevenly between studies: sample size, tumor location, and sequencing quality all change how many productive TCRs are recovered. A T cell whose TCR sequence is shared with no other cell — a "singleton" — may genuinely belong to no clone, or may be a member of a large clone the assay failed to observe, confounding attempts to correlate phenotype with clonality across studies.

scXpand, developed by Ofir Shorer, Ron Amit, and Keren Yizhak at Technion – Israel Institute of Technology and posted as a bioRxiv preprint in September 2025, predicts clonal expansion status for each individual T cell from its transcriptome alone. It is a supervised, task-specific framework rather than a general-purpose embedding model like scGPT or SCimilarity: it ships fixed checkpoints trained once on a pan-cancer corpus of paired scRNA/TCR-seq data, then applies them to new datasets carrying no TCR information at all.

#Key Features

• TCR-free expansion calls: Assigns each T cell a probability of belonging to an expanded clone from unique molecular identifier counts alone, extending clonality analysis to datasets never paired with TCR sequencing.
• Pan-cancer training corpus: Built from 38 publicly available datasets spanning 14 cancer types and both tumor and blood samples, quality-controlled independently on the RNA and TCR sides.
• Five model families in one interface: A multi-task count autoencoder, an MLP, LightGBM, a linear SVM, and logistic regression are each trained and tuned, trading interpretability against the ability to capture gene-gene interactions.
• SHAP-based interpretation: Per-gene Shapley values recover known expansion biology and expose context-dependent gene-gene interactions invisible to linear models.
• Packaged and memory-efficient: Distributed on PyPI with streaming from disk during training, so millions of cells can be processed without loading the full matrix into memory.

#Technical Details

The training database comprises more than 2.6 million T cells passing quality control: 26 datasets covering 963 samples and over 1.5 million cells for training and tuning, and 12 external datasets covering 482 samples and over 1 million cells held out for testing. Input features are UMI counts over the 11,950 genes shared across all training datasets, keyed by Ensembl ID; MAGIC imputation recovers dropouts of CD8A/B, CD4, and FOXP3 for subtype assignment. Expansion labels are derived from CDR3 sequence identity within each sample, and splits are stratified but patient-exclusive. The autoencoder is a forked, masked design with a zero-inflated negative binomial reconstruction loss, an auxiliary head predicting tissue of origin (tumor or blood), and a second head performing the main expansion classification; only the latter is used at inference.

On the held-out validation set all five models discriminate comparably, with ROC AUC between 0.916 and 0.926 and precision-recall AUC between 0.742 and 0.785, but they diverge at a fixed threshold: the autoencoder, MLP, and LightGBM models hold false positive rates of 0.109-0.163 against 0.258-0.282 for the linear models. The autoencoder achieved the highest Matthews correlation coefficient at a naive 0.5 threshold and was carried forward. Across the external test datasets it reached AUCs of 0.886 on non-small-cell lung cancer tumors, 0.937 on melanoma blood, 0.877 on triple-negative breast cancer tumors, 0.961 on glioma blood, and 0.847-0.963 on the remaining eight. On a pancreatic cohort — a cancer type absent from training — it reached 0.868, 0.969, and 0.935 on tumor, blood, and adjacent normal samples.

#Applications

The immediate use is retrospective: immuno-oncology groups can apply the pretrained checkpoints to the large body of existing tumor scRNA-seq data that lacks TCR sequencing and recover a clonality axis that was previously unavailable. Applied in the preprint to a breast cancer dataset without TCR data, scXpand found expanded CD8+ T cells significantly more abundant in one immune environment than another, and a 12-gene prognostic signature separated the two groups as well. The hyperparameter optimization and training pipeline are exposed, so the same infrastructure can be repurposed for other binary single-cell classification tasks at atlas scale.

#Impact

scXpand is the first framework to predict pan-cancer T-cell clonal expansion from expression alone, and it is released openly: code under MIT, the tuned checkpoints on Figshare under CC BY 4.0, and tutorials with an API reference on Read the Docs. Its limitations are stated plainly by the authors. The model recovers a transcriptional state associated with expansion, not clonal identity or TCR sequence, so it cannot substitute for scTCR-seq where repertoire structure is the question. Training covered 14 cancer types — excluding ovarian, gastric, cervical, and hematologic malignancies — and only tumor and blood tissue. Applicability outside human cancer is untested, and the work remains a preprint awaiting peer review.