Profiling Cell-state Fingerprints Based on Deep Learning Model with Meta-programs of Pan-cancer

Cell states within cancer have garnered significant attention, yet the mechanisms through which malignant cells assert dominance in pan-cancer commonalities remain elusive. In this study, we employed label-free multiplexed single-cell RNA sequencing (scRNA-seq) to analyze cell states in 159,372 cells across 245 cell lines spanning 14 tissue types, integrating both public and proprietary datasets. We identified 21 meta-programs (MPs) representing shared characteristics across pan-cancer landscapes, encompassing 16 biological processes. Subsequently, we developed a deep learning model StateNet to generate cell-state fingerprints for delineating the individuality of each cell line based on these MPs. Leveraging StateNet, we pinpointed ACAT2 as a potential mediator bridging hypoxia and the lipid metabolism pathway, and we also showcased that epithelial-mesenchymal transition programs are vital for classifying cell lines through perturbation experiments. StateNet not only elucidates the overarching manifold structure of scRNA-seq data but also furnishes cell-state fingerprints of cell clusters, unveiling prognosis-related programs and distinguishing between patients with varying survival outcomes. Utilizing these prognosis-related programs on 3210 cancer samples, we constructed Cox models and identified risk-associated programs and genes responsible for different cancer types. StateNet thus emerges as a novel and efficient tool for cancer profiling, unraveling the shared commonalities and distinct individualities of pan-cancer cells across expansive datasets.