Artificial intelligence approaches for tumor phenotype stratification from single-cell transcriptomic data

Area of Science:

• Computational biology
• Genomics
• Cancer research

Background:

• Single-cell RNA-sequencing (scRNA-seq) reveals tumor heterogeneity but lacks clinical risk assessment for cell subpopulations.
• Intra-tumoral complexity and limited clinical data hinder the evaluation of individual cell subtype aggressiveness.

Purpose of the Study:

• Introduce SCellBOW, a novel computational framework for scRNA-seq analysis.
• Enhance identification and visualization of single-cell subpopulations.
• Enable risk stratification of tumor cell subpopulations based on aggressiveness.

Main Methods:

• SCellBOW utilizes Natural Language Processing-inspired document embedding techniques.
• Framework performance was validated against existing methods using diverse scRNA-seq datasets.
• Risk assessment is achieved by simulating subpopulation impact on disease prognosis.

Main Results:

• SCellBOW accurately represents phenotypically divergent cell types.
• Identified a novel, aggressive AR-/NElow malignant subpopulation in metastatic prostate cancer.
• Demonstrated SCellBOW's capability to stratify cell clusters by aggressiveness.

Conclusions:

• SCellBOW offers effective identification and visualization of single-cell subpopulations.
• The framework facilitates risk stratification, aiding in the development of tailored cancer therapies.
• Highlights the clinical relevance of identifying specific tumor subpopulations and their prognostic impact.