MoVAE: Multi-Omics Variational Auto-Encoder for Cancer Subtype Detection

With the advent of advanced genomic data extraction methods, numerous studies have been utilized these data to identify cancer subtypes. Given the complexity of cancer subtyping and the limitations of single-omics data, multi-omics approaches have emerged as a promising solution. Despite the challenges of multi-omics data integration, these methods have shown acceptable results. Various strategies, including matrix decomposition, kernel-based fusion, and deep learning, have been proposed for data fusion and subtyping. As a deep model, we present the Multi-omics Variational Auto Encoder (MoVAE), an innovative strategy for integrating multi-omics data. MoVAE facilitates adaptive fusion by extracting unique and synergistic features from high-dimensional multi-omics data using an AggreEncoder network. Unlike previous methods, MoVAE effectively handles the integration of heterogeneous data, accommodates missing omics, and manages high-dimensionality. As a variational autoencoder, MoVAE encodes each multi-omics input into a low-dimensional latent space while extracting the omics-dependency and complementary information of multi-omics data according to a prior joint distribution. Using the mined knowledge in the cancer subtype detection model, its prediction efficiency would be improved. By decoding the latent vectors, MoVAE reconstructs the input multi-omics data as accurately as possible. To demonstrate MoVAE's effectiveness in identifying biologically meaningful and clinically relevant cancer subtypes, we conducted several experiments and compared the results with state-of-the-art VAE-based methods.