Fig. 3
From: Pan-cancer gene set discovery via scRNA-seq for optimal deep learning based downstream tasks
Comparative performance analysis of gene sets across downstream tasks. Bar plots showing the performance of nine different gene selection methods evaluated across downstream tasks using the TCGA pan-cancer RNA-seq dataset. Each bar represents the mean AUROC across 5-fold cross-validation, with error bars indicating 95% confidence intervals obtained from bootstrapping. Dashed horizontal lines mark the average AUROC across all methods within each task category for visual reference. a. Tumor mutation burden (TMB) assessment performances across four cancer types: CRC, LUAD, LUSC, and SKCM. b. Microsatellite instability (MSI) classification for CRC and STAD. c. Mutation prediction (MUT) performance for six cancer gene combinations: BRCA-TP53, LUAD-EGFR, LUAD-KRAS, LUAD-TP53, PAAD-KRAS, and STAD-TP53. All models were trained using a multi-layer perceptron (MLP) architecture with identical hyperparameters to ensure fair comparison. Higher AUROC values indicate stronger discriminative performance of the corresponding gene set.
