Reusability report: Leveraging supervised learning to uncover phenotype-relevant biology from single-cell RNA...

Area of Science:

• Computational biology
• Single-cell transcriptomics
• Immunogenomics

Background:

• Single-cell RNA sequencing (scRNA-seq) reveals cellular heterogeneity.
• Associating cell subsets with specific phenotypes remains a challenge.
• PENCIL is a supervised learning framework designed to identify phenotype-relevant cells.

Purpose of the Study:

• To evaluate the reproducibility and transferability of the PENCIL framework.
• To identify and address limitations in the original PENCIL version.
• To enhance PENCIL's utility for identifying phenotype-relevant cell subsets.

Main Methods:

• Comprehensive evaluation of PENCIL across 12 scRNA-seq datasets.
• Identification and correction of PENCIL's sensitivity to input perturbation.
• Integration of gene set variation analysis with PENCIL for enhanced cell subset identification.

Main Results:

• PENCIL's reproducibility was enhanced through identified corrections.
• A cytotoxic T cell immunotherapy response signature (CyTIR) was developed by boosting PENCIL.
• CyTIR accurately predicts immune checkpoint blockade response in skin cancer (AUC >0.75, accuracy >0.71) across datasets.

Conclusions:

• The enhanced PENCIL framework demonstrates improved reproducibility and utility.
• PENCIL has significant potential for identifying phenotype-relevant cell subsets in various biomedical applications.
• The developed CyTIR signature offers a promising biomarker for predicting immunotherapy response.