An urgent need in bladder cancer. Bladder cancer is the ninth most commonly diagnosed cancer worldwide and among the most sex-disparate, with men facing four times the incidence and death risk of women. The five-year survival rate at stage IV is only about 15%, making early and accurate risk stratification essential for directing patients toward appropriate therapies.
Previous molecular subtyping of bladder cancer identified luminal and basal subtypes characterized by distinct gene markers, and TCGA catalogued four subtypes based on RNA sequencing alone. However, none of these approaches combined multi-layered molecular data with survival information during the subtype discovery step itself.
The gap this study fills. Prior single-omics prognostic models and multi-omics subtype identification studies worked in isolation. This paper presents DeepProg, a pipeline that simultaneously integrates mRNA, miRNA, and DNA methylation data through deep learning to discover survival-linked molecular subtypes and then classify any new patient sample.
Three-omics input from TCGA. The study used 402 bladder cancer samples from The Cancer Genome Atlas with mRNA expression, miRNA expression, and DNA methylation data. Each omic layer was independently normalized using rank normalization followed by Pearson correlation distance between samples, producing sample-by-sample correlation matrices that remove platform-specific technical variation.
A denoising autoencoder with 100 hidden nodes was trained separately for each omic layer. The autoencoder learns a compressed, nonlinear representation of each omic by reconstructing the input from a corrupted version, reducing overfitting through 50% dropout. The compressed representations capture biologically meaningful patterns that simple linear dimensionality reduction would miss.
Survival-linked feature selection. From the 100 features produced by each autoencoder, those with a statistically significant association with patient survival (log-rank p-value below 0.01 in univariate Cox proportional hazards models) were retained. Significant features from all three omics were stacked into a single integrated matrix used for unsupervised clustering into survival subtypes via Gaussian mixture modeling.
The optimal number of subtypes was determined objectively by silhouette and Calinski-Harabasz scores, which both favored two subtypes. This two-subtype solution was then used to train Support Vector Machine classifiers, one per omic and one combining all three, enabling prediction of the survival subtype of any new patient sample that shares at least one omic with the training data.
Highly significant survival differences. The two inferred subtypes showed clear Kaplan-Meier curve separation in the full TCGA cohort with a log-rank p-value of 1.8e-05. Across 20 random 66/33 training and test splits, the geometric mean p-values were 8e-4 for training folds and 0.04 for test folds, demonstrating that the subtype separation is robust and not an artifact of a particular data split.
The model also successfully classified a held-out final test set with a p-value of 0.03, and an independent external validation cohort of 62 muscle-invasive bladder cancer samples with only miRNA data yielded a significant survival difference with p-value of 0.02. The ability to classify samples using a single omic layer when multi-omics data are unavailable is an important practical advantage.
mRNA carries the most prognostic signal. Comparing the contribution of each omic layer, mRNA produced an average of 14 survival-linked features out of 100 encoded dimensions, far ahead of miRNA with 6 and methylation with 2. Individual omic classifiers showed corresponding average p-values of 0.001 for mRNA, 0.002 for miRNA, and 0.004 for methylation, confirming transcriptomics as the dominant survival signal.
Known aggressive-subtype markers confirmed. Differential expression analysis of the high-risk subtype (S1) versus the low-risk subtype (S2) identified 801 upregulated and 228 downregulated genes. Among the most notable upregulated genes were KRT6 and KRT14, which are established markers of the aggressive basal bladder cancer subtype, validating that DeepProg's unsupervised discovery recovers biologically meaningful groupings.
miRNA analysis found 18 upregulated and 10 downregulated microRNAs in the aggressive subtype. The downregulation of miR-194, a known tumor suppressor in bladder cancer, and the upregulation of miR-133b, a marker associated with high risk and shorter progression-free survival, were consistent with prior literature.
Methylation and pathway activation. Differential methylation identified 918 hypermethylated and 46 hypomethylated genes in the aggressive subtype, including hypermethylation of KRT13, associated with high-grade non-invasive bladder cancer, and FHIT and LAMC2, previously linked to poor survival. Pathway enrichment analysis revealed the PI3K-Akt signaling pathway, cytokine-cytokine receptor interaction, and ECM-receptor interaction as the most activated pathways in the high-risk subtype, all previously implicated in bladder cancer progression.
Flexible one-omic-at-a-time classification. A practical challenge in clinical translation is that external datasets rarely have all three omic layers measured. DeepProg addresses this by training an SVM classifier for each individual omic using the top 10 features most discriminative between subtypes, as well as a combined 30-feature multi-omics SVM, allowing classification regardless of which omics are available in a new sample.
New samples are normalized by computing their Pearson correlation distance to all training samples using the shared omic features, producing a normalized vector that can be input directly into the appropriate pre-trained classifier. This design means the only preprocessing required for a new sample is the correlation distance computation, making the pipeline practical for real-world deployment where standardized omics pipelines may differ across centers.
Probability output for risk grading. Beyond subtype assignment, each classifier outputs the probability of a sample belonging to the low-survival subgroup. This continuous risk score, rather than a binary label, can be used to fit Cox proportional hazards models and provides a graded measure of patient risk suitable for clinical decision support.
Tumor purity as a confounding factor. Bladder cancer samples in TCGA are reported to belong to one of the least pure cancer types, meaning that sequencing data reflect a mixture of tumor and normal cells. This tumor impurity adds noise to multi-omics measurements and makes survival analysis more challenging. The authors expect performance to improve as larger cohorts of higher-purity samples become available.
The modular architecture of DeepProg, where each omic has its own autoencoder, allows straightforward extension to additional data types such as clinical imaging, proteomics, or clinical variables. This extensibility positions the pipeline as a general framework for multi-modal cancer risk stratification beyond bladder cancer.
A broadly applicable template. The authors demonstrated DeepProg on bladder cancer as a proof of concept, but the pipeline's flexibility means it can be applied to any cancer type with multi-omics data. The requirement for only a subset of shared features between training and test samples makes it more realistic for adoption in settings where complete multi-omics profiling is not yet standard of care.