CEP-IP: An explainable framework for cell subpopulation identification in single-cell transcriptomics.

Area of Science:

• Computational biology and bioinformatics
• Genomics and transcriptomics
• Machine learning in biology

Background:

• Single-cell RNA sequencing (scRNA-seq) analysis often lacks explainable methods for identifying cell subpopulations with specific gene expression relationships.
• Identifying cells with strong pairwise monotonic gene-module relationships is crucial for understanding cellular heterogeneity.

Purpose of the Study:

• Introduce CEP-IP, a novel explainable machine learning framework to identify cell subpopulations based on gene-module relationships.
• Address the gap in explainable approaches for scRNA-seq data analysis.

Main Methods:

• Utilized prostate cancer scRNA-seq data with TRPM4 as the gene of interest (GOI) and ribosomal genes (Ribo) as co-expressed genes.
• Employed Spearman-Kendall dual-filter (DFG) to identify co-expressed genes and generalized additive modeling to quantify gene-module relationships (deviance explained, DE).
• Applied cell explanatory power (CEP) classification and inflection point (IP) analysis to stratify cells into distinct subpopulations (pre-IP and post-IP TREP cells) for pathway analysis, validated in independent datasets.

Main Results:

• CEP-IP identified four distinct cell subpopulations in each prostate cancer patient, outperforming alternative gene set modules.
• Pre-IP TREP cells showed enrichment in immune-related processes, while post-IP TREP cells were enriched for ribosomal, translation, and cell adhesion pathways.
• Validation in middle temporal gyrus and glioblastoma multiforme datasets confirmed CEP-IP's ability to identify biologically distinct subpopulations and revealed continuous cell trajectories in 3D space.

Conclusions:

• CEP-IP successfully identifies biologically distinct cell subpopulations in multiple independent scRNA-seq datasets.
• The framework demonstrates generalizability for analyzing pairwise gene-module relationships in single-cell transcriptomics.