The Problem Hepatocellular carcinoma (HCC) is the most common type of liver cancer and accounts for 70-90% of cases. With a 5-year survival rate below 32% and extremely limited treatment options, better tools for predicting which patients will fare poorly are urgently needed to guide therapy choices.
The Approach This study is the first to apply deep learning to integrate three types of molecular data simultaneously - RNA-seq gene expression, miRNA-seq, and DNA methylation - to identify HCC patient subgroups with markedly different survival outcomes.
Why Multi-Omics? Each data type captures a different dimension of tumor biology. Gene expression shows what proteins are being made; miRNAs regulate gene expression post-transcriptionally; DNA methylation reflects epigenetic silencing. Integrating all three provides a richer, more predictive picture than any single data type alone.
Key Result The deep learning model identified two robust survival subgroups from 360 TCGA HCC patients and validated them across five independent external cohorts from different countries and platforms, with concordance indices ranging from 0.67 to 0.82.
Autoencoder Design An autoencoder is an unsupervised neural network that compresses high-dimensional input data into a lower-dimensional representation and then reconstructs it. The network used here had three hidden layers with 500, 100, and 500 nodes. The 100 nodes in the bottleneck layer produced compressed features capturing the most essential patterns across all three omics types.
Feature Selection with Survival After compression, each of the 100 bottleneck features was tested with univariate Cox proportional hazards models. Only the 37 features significantly associated with survival (log-rank p less than 0.05) were retained. K-means clustering on these features identified the optimal two survival subgroups.
Supervised Classification The cluster labels from K-means were then used to train a Support Vector Machine (SVM) classifier. For the TCGA data, the top 100 mRNA, 50 methylation, and 50 miRNA features selected by ANOVA were used. This classifier was then applied to external cohorts after appropriate normalization.
Validation Strategy The model was trained on all 360 TCGA samples for label generation, then tested using 60/40 cross-validation splits. Five external cohorts spanning RNA-seq, mRNA microarray, miRNA array, and DNA methylation platforms were used for independent validation, providing a stringent test of generalizability.
Significant Survival Difference The two subtypes identified in TCGA data showed dramatically different survival curves (log-rank p = 7.13 x 10^-6). Subtype S1 is more aggressive (higher risk) and S2 is more moderate (lower risk), with the difference persisting across all cross-validation folds.
TP53 Mutations in Aggressive Subtype TP53 inactivation mutations were significantly more frequent in the aggressive S1 subtype compared to S2 in both TCGA (Fisher's test p = 0.042) and the LIRI-JP cohort (p = 0.024). TP53 is a well-known tumor suppressor whose loss is associated with poor HCC prognosis.
Stemness Markers Elevated In the S1 aggressive subtype, key cancer stemness markers were significantly upregulated: EPCAM (p = 5.7 x 10^-6), KRT19 (p = 6.7 x 10^-15), and the tumor survival marker BIRC5 (p = 1.2 x 10^-13). These markers identify a progenitor-like, more dangerous tumor phenotype.
Pathway Differences Pathway analysis revealed that S1 tumors have activated Wnt signaling and PI3K-Akt signaling pathways - known oncogenic drivers - while S2 tumors show activated metabolic pathways (drug metabolism, amino acid and fatty acid metabolism), suggesting a less aggressive, more differentiated tumor state.
Cross-Cohort Performance The model was validated in five independent cohorts: LIRI-JP (n=230, C-index=0.75), NCI (n=221, C-index=0.67), Chinese (n=166, C-index=0.69), E-TABM-36 (n=40, C-index=0.77), and Hawaiian (n=27, C-index=0.82). The average C-index across external cohorts was 0.74, indicating good predictive accuracy.
Platform Diversity These five cohorts used different data platforms - RNA-seq, mRNA microarray, miRNA array, and DNA methylation arrays - demonstrating that the classifier generalizes across technological platforms, not just patient populations.
Multi-Omics Advantage When comparing three-omics to single-omics models in cross-validation, the multi-omics approach consistently outperformed any single data type, confirming that data integration provides meaningful benefit over using gene expression, miRNA, or methylation alone.
Deep Learning vs. Alternatives Principal Component Analysis (PCA) with the same number of features completely failed to find significant survival subgroups (p = 0.14 vs. 6.0 x 10^-9 for deep learning). Direct Cox-PH feature selection also gave less robust results across confirmation cohorts, underscoring the value of the autoencoder approach.
Clinical Variables Not Enough When clinical features alone (stage, grade, race, gender, age, and risk factor) were used for survival prediction, performance was much poorer (C-index 0.45-0.70) than the molecular model. Adding clinical data to the molecular model did not significantly improve performance, suggesting the molecular data already captures much of what clinical variables would add.
Risk Factor Subgroups The model performs well across all major HCC risk factor subgroups: HBV infected (C-index 0.74, p = 0.04), alcohol consumers (C-index 0.79, p = 0.005), and other groups (C-index 0.77, p = 0.0035). Only the HCV group was non-significant, likely due to small size (n=31).
Tumor Grade and Stage Of all clinical variables tested, only tumor grade (p = 0.0004) and stage (p = 0.002) were significantly associated with the molecular survival subtypes, consistent with the subtypes representing genuine biological aggressiveness rather than just staging artifacts.
Implications for Therapy The two molecular subtypes with their distinct pathway activations could guide treatment selection. S1 tumors with Wnt and Akt pathway activation might benefit from targeted inhibitors of these pathways, while S2 tumors may respond differently to systemic treatments.
Training Data Limitations The model was trained on TCGA data, which may have inflated C-statistics for the TCGA test set since labels were generated using all TCGA samples. External datasets provide more unbiased performance estimates and are the true test of generalizability.
Tumor Purity TCGA liver samples, while better than many cancer types, are not pure tumor cells. Heterogeneous tumor content from diverse etiological backgrounds (HBV, HCV, alcohol) adds noise that likely limits performance ceiling on TCGA training data.
Single-Omics Prediction The model can predict from a single omics type when multi-omics data is unavailable, but multi-omics integration provides the highest accuracy. Developing methods to make high-quality predictions from the most readily available data type (mRNA expression) would improve clinical utility.
Prospective Validation The authors plan to collaborate with clinicians to validate the model in prospective cohorts. Truly prospective validation with standardized data collection protocols is essential before this approach can guide real treatment decisions.