Abstract
Objectives: Cancer, characterized by its profound complexity and heterogeneity, arises from a multitude of molecular disruptions. The pursuit of identifying distinct cancer subtypes is driven by the need to stratify patients into clinically coherent subgroups, each exhibiting unique prognostic outcomes. The integration of multi-omics datasets enhances the precision of subtyping and advances precision medicine. Methods: Considering the high-dimensional nature inherent to various multi-omics data types, we introduce an innovative deep learning framework, DAE-MKL, which integrates denoising autoencoders with multi-kernel learning for identifying cancer subtypes. Leveraging the capabilities of DAE, we extract non-linearly transformed features that retain pertinent information while mitigating noise and redundancy. These refined data representations are then funneled into the MKL framework, thereby enhancing the accuracy of subtype identification. We applied the DAE-MKL framework to both simulated studies and empirical datasets derived from two distinct cancer types, low-grade glioma (LGG, n = 86) and kidney renal clear cell carcinoma (KIRC, n = 285), thereby validating its utility and feasibility. Results: In simulations, DAE-MKL achieved superior performance with NMI gains up to 0.78 compared to other state-of-the-art methods. For real datasets, DAE-MKL identified three LGG subtypes and three KIRC subtypes, showing significant survival differences (KIRC log-rank p = 3.33 × 10(-8), LGG log-rank p = 3.99 × 10(-8)). Additionally, we explored potential cancer-related biomarkers. Conclusions: The DAE-MKL effectively identifies molecular subtypes, reduces data dimensionality, and improves prognostic stratification in multi-omics cancer datasets, providing an effective tool for precision oncology.