Fig. 3

Experimental results and downstream analyses of machine learning baselines applied to MLOmics datasets. (a) PREC bar plots for each baseline method across all datasets. Overall, machine learning-based methods (Subtype-GAN, DCAP, MAUI, XOmiVAE, CustOmics, and DeepCC) outperformed traditional statistical methods (SVM, XGBoost, RF, and LR). (b) SIL heatmaps for each baseline method across all datasets. Methods employing deep generative neural network architectures (Subtype-GAN, DCAP, MAUI, XOmiVAE, and MCluster-VAEs) generally outperformed other methods (SNF, NEMO, CIMLR, iClusterBayes, and moCluster). (c) Box plots for each baseline method across three imputation datasets. Matrix decomposition methods (SVD, Spectral) outperformed deep learning-based methods (GAIN, GNN). (d,e) Schematic illustrations of downstream analysis results based on the clustering outcomes of XOmiVAE applied to specific cancer patient clustering datasets. In the survival analysis plot, survival curves in different colors correspond to distinct clustering groups. In the volcano plot and KEGG pathway analysis, red and blue indicate downregulated and upregulated genes between patient groups, respectively. In the patient group clustering plot, different colors represent samples belonging to different clusters. In the simulated gene knockout analysis, red and green dots indicate sample clustering before and after expression knockout, respectively.