Table 1.
Summary of number of clusters.
Fig 1.
Kaplan-Meier survival curves for cancer subgroups identified using CA-CAE across 15 multi-omics cancer datasets.
Each panel represents one cancer type, with survival-related subgroups identified through K-means clustering.
Table 2.
Summary of P-values for different models (Log-rank test results).
Table 3.
C-index (Concordance index) comparison among different prognostic models.
Fig 2.
Comparison of model performance in terms of -log10(P-values) and C-index for five models across 15 cancer types.
(A) Distribution of -log10(P-values) for CA-CAE, ProgCAE, DeepProg, PCA and NMF. (B) Distribution of C-index values for the same five models.
Table 4.
Summary of 95% confidence intervals for the C-index across 15 cancer types.
Fig 3.
Ablation study of CA-CAE components on LUAD.
Kaplan–Meier survival curves of LUAD patients illustrating the impact of different modules on model performance:(A) The full CA-CAE model achieved the highest discriminative power. (B) Removing the channel attention mechanism (CA-CAE_noAtt) led to decreased stratification performance. (C) Excluding the Cox-based prognostic filtering (CA-CAE_noCox) further reduced discrimination.
Fig 4.
Kaplan-Meier survival curves for SVM model predictions in ACC, BLCA, and LUAD datasets.
Fig 5.
Comparison of single-omics and multi-omics approaches in predicting survival outcomes for LUAD patients.
(A) Hazard ratio analysis of significant mRNA features identified through univariate Cox regression in single-omics analysis. (B) Kaplan-Meier survival curve based on mRNA single-omics features for LUAD patients. (C) Kaplan-Meier survival curve based on integrated multi-omics features (mRNA and DNA methylation) for LUAD patients.
Table 5.
Summary of top 20 ranked genes across 3 omics datasets.
Fig 6.
Functional and pathway enrichment analyses of survival-related genes identified from the LUAD dataset.
(A) Gene Ontology (GO) enrichment analysis results. (B) KEGG pathway enrichment analysis results.
Table 6.
Associations between clinical features and cancer subtypes (Chi-square test results).
Table 7.
Associations between CA-CAE subtypes and WGD status across cancers.
Table 8.
Comparison of CA-CAE and NMF clustering consistency across representative cancer types.
Fig 7.
Kaplan–Meier survival analysis for GBM and LUAD patients from the CPTAC cohort using the CA-CAE model.
(A) Cancer: GBM (CPTAC). (B) Cancer: LUAD (CPTAC).
Table 9.
Summary of datasets.
Fig 8.
Overview of CA-CAE: The model includes feature normalization, dimensionality reduction, feature selection, and survival analysis for three types of omics data (DNA methylation, mRNA-seq, and miRNA-seq).
Each omics dataset is modeled with a convolutional autoencoder (CAE) combined with an attention mechanism to improve flexibility and scalability for heterogeneous data types.