SGEGCAE: A Sparse Gating Enhanced Graph Convolutional Autoencoder for Multi-omics Data Integration and Classification

Integration of multi-omics data is essential for obtaining comprehensive insights into molecular mechanisms of complex diseases. While several methods have been proposed for analyzing multi-omics data in various applications, challenges persist in effectively handling heterogeneous and rich multi-omics data. In this paper, a Sparse Gating Enhanced Graph Convolutional AutoEncoder, named SGEGCAE, is proposed for multi-omics data integration and classification. Specifically, an enhanced graph convolutional autoencoder is developed, which integrates a basic autoencoder with a sparse gating strategy, aiming to combine attribute information with topological structure information of the graph for obtaining more comprehensive feature representations. To address the inherent variability and fluctuations in different omics data quality among samples, true class probability is introduced into the SGEGCAE to acquire reliable classification confidence. Furthermore, a tensor fusion network is designed to explore both inter-omics and intra-omics relationships in the label space to achieve ultimately multi-omics integration and classification. Extensive biomedical classification experiments are carried out on four datasets. In these experiments, the superior performance of the SGEGCAE is clearly validated compared to some state-of-the-art integrative analysis methods, demonstrating that the SGEGCAE might serve as an alternative method for multi-omics data integration and classification. The code and datasets for the SGEGCAE are available online at https://github.com/CDMBlab/SGEGCAE .