Cross-modal mapping of cancer stem-like cell plasticity using deep learning

Area of Science:

• Computational biology
• Cancer research
• Genomics

Background:

• Cancer stem-like cells (CSCs) drive tumor heterogeneity, resistance, and progression.
• Single-cell RNA sequencing (scRNA-seq) offers insights into tumor hierarchies.
• Robust tools are needed to profile CSCs across single-cell and bulk transcriptomic data.

Purpose of the Study:

• To develop a unified, machine learning-based framework (ACSCeND) for CSC state classification and tissue deconvolution.
• To enable high-resolution profiling of CSC states from transcriptomic data.
• To integrate single-cell precision with bulk-level applicability for CSC analysis.

Main Methods:

• Developed ACSCeND, a framework with a supervised classifier for CSC state assignment (pluripotent-like, multipotent-like, unipotent-like) and an attention-guided autoencoder for CSC subtype deconvolution from bulk RNA sequencing data.
• Trained the classifier on curated scRNA-seq datasets.
• Validated performance against existing tissue deconvolution tools using synthetic and real-world samples.

Main Results:

• ACSCeND demonstrated superior accuracy compared to existing tissue deconvolution tools.
• Analysis of over 25,000 tumor profiles revealed that CSC abundance correlates with poor disease-free survival and reduced immunotherapy efficacy.
• Identified distinct CSC-state-specific molecular programs, providing insights into CSC-driven heterogeneity and tumor evolution.
• The model successfully recapitulated developmental hierarchies in noncancerous tissues.

Conclusions:

• ACSCeND provides a robust and interpretable approach for profiling CSC dynamics.
• CSC state serves as a clinically meaningful, pan-cancer biomarker.
• The framework facilitates the development of stemness-informed therapies.