scTransMIL bridges patient-level disease states and single-cell transcriptomics for cancer screening and heterogeneity inference

Abstract

Single-cell sequencing offers profound insights into tumor complexity. However, reliably linking individual cellular profiles to overall sample-level phenotypes remains computationally challenging, particularly when cell-specific annotations are absent. This disconnect limits the technology’s utility in studying underlying tumor mechanisms. Here we show how scTransMIL, an artificial intelligence framework, bridges this gap. By conceptualizing a biological sample as a collection of individual cells, scTransMIL employs a transformer-based multi-instance learning approach to directly connect single-cell transcriptomic profiles with overarching sample-level labels. We demonstrate that scTransMIL accurately predicts sample-level cancer phenotypes, including the tissue-of-origin for metastatic tumors. At cellular resolution, it robustly identifies tumor-associated cell populations and maps biological progression trajectories using only minimal sample-level annotations. Furthermore, the model’s attention mechanisms facilitate full-transcriptome biomarker discovery. By systematically integrating molecular and cellular scales with broader phenotypic states, our approach provides a computational tool to dissect tumor heterogeneity and advance fundamental cancer biology. Reliably linking patient-level cancer phenotypes to single-cell transcriptomes re-mains challenging due to limited cell-level labels. Here, the authors develop scTransMIL, a transformer-based multi-instance learning framework that identifies cancer states, cancer subtypes, and potential cancer biomarkers with good performance across datasets.