Adaptive deep shared latent representation enables novel multi-omics cancer subtype classification

Abstract

Variations in outcomes among cancer patients are significant even when they have the same type of tumor. Identifying and classifying molecular subtypes of cancer offers a valuable opportunity to enhance prognosis and tailor treatment plans for individuals. Recent efforts have been made to generate extensive multidimensional genomic data to achieve this potential. However, existing algorithms still face challenges in integrating and analyzing such intricate datasets. In this study, we present Adaptive Deep Shared Latent Representation (ADSLR), a novel approach for cancer subtyping that utilizes shared latent representation to reveal distinct molecular subtypes in cancer. It incorporates a cycle autoencoder with a nonnegative matrix factorization layer, capturing consistent signals of nonlinear features at various omics levels. This enables the generation of adaptable representations for shared latent representation across multiple omics levels. We apply ADSLR to multi-omics data obtained from eight different cancer types in the “The Cancer Genome Atlas” dataset, demonstrating significant improvements in the identification of biologically meaningful cancer subtypes. These identified subtypes exhibit noteworthy variations in patient survival rates across seven out of the eight cancer types. Our analysis uncovers integrated patterns involving mRNA expression, miRNA expression, DNA methylation, and protein across multiple cancers while showcasing ADSLR’s versatility for integrating various other omics types.