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

Area of Science:

• Cancer Biology
• Genomics
• Computational Biology

Background:

• Understanding shared cell states in cancer is crucial for developing effective treatments.
• Mechanisms of malignant cell dominance in pan-cancer commonalities are not well understood.

Purpose of the Study:

• To analyze cell states across diverse cancer types using single-cell RNA sequencing.
• To develop a deep learning model (StateNet) for cancer cell profiling and identifying shared characteristics.
• To uncover prognosis-related programs and risk-associated genes in cancer.

Main Methods:

• Label-free multiplexed single-cell RNA sequencing (scRNA-seq) on 159,372 cells from 245 cell lines across 14 tissue types.
• Integration of public and proprietary scRNA-seq datasets.
• Development and application of the deep learning model StateNet for cell-state fingerprinting and analysis.
• Perturbation experiments to validate the role of epithelial-mesenchymal transition programs.
• Construction of Cox models using prognosis-related programs on 3210 cancer samples.

Main Results:

• Identification of 21 meta-programs (MPs) representing shared characteristics and 16 biological processes across pan-cancer landscapes.
• StateNet generated cell-state fingerprints, delineating cell line individuality.
• Pinpointed ACAT2 as a potential mediator between hypoxia and lipid metabolism.
• Epithelial-mesenchymal transition programs were found vital for cell line classification.
• Unveiled prognosis-related programs and identified risk-associated genes linked to patient survival outcomes.

Conclusions:

• StateNet is a novel tool for cancer profiling, elucidating shared commonalities and individualities of pan-cancer cells.
• The model effectively analyzes scRNA-seq data, reveals prognosis-related programs, and distinguishes patient survival.
• Findings provide insights into cancer mechanisms and potential therapeutic targets.