Pan-cancer gene set discovery via scRNA-seq for optimal deep learning based downstream tasks.

Area of Science:

• Genomics
• Computational Biology
• Cancer Research

Background:

• Machine learning advances cancer research using transcriptomics.
• High dimensionality of RNA sequencing (RNA-seq) data challenges pan-cancer studies.
• Single-cell RNA sequencing (scRNA-seq) offers higher resolution data.

Purpose of the Study:

• To test if scRNA-seq derived gene sets outperform bulk RNA-seq gene sets in pan-cancer downstream tasks.
• To develop and validate a robust feature selection method for cancer genomics.
• To identify novel pan-cancer biomarkers.

Main Methods:

• Analysis of scRNA-seq data from 181 tumor biopsies across 13 cancer types.
• High-dimensional weighted gene co-expression network analysis (hdWGCNA) for gene set identification.
• XGBoost for feature selection, refined gene sets applied to TCGA pan-cancer data, and evaluated using deep learning models (MLPs, GNNs).

Main Results:

• The XGBoost-refined hdWGCNA gene set showed superior performance in most pan-cancer tasks.
• Tasks included tumor mutation burden assessment, microsatellite instability classification, mutation prediction, cancer subtyping, and grading.
• Genes like DPM1, BAD, and FKBP4 were identified as significant pan-cancer biomarkers, with DPM1 showing consistent importance.

Conclusions:

• Integrating scRNA-seq data with advanced computational methods provides a powerful approach for feature selection in cancer genomics.
• This strategy significantly improves predictive accuracy in pan-cancer downstream tasks.
• The identified gene sets and biomarkers offer promising avenues for advancing cancer research and precision medicine.