ScMOGAE: A Graph Convolutional Autoencoder-Based Multi-omics Data Integration Framework for Single-Cell Clustering

The integration of single-cell multi-omics data is a significant step forward in our understanding of the complex biological systems at the cellular level. This approach allows for the simultaneous analysis of various molecular layers, and provides insights into the heterogeneity and clustering of cells. However, the fusion of single-cell multi-omics data poses several challenges on how to effectively represent joint distributions due to their high dimensionality, sparsity and dropout events. In this study, we propose a deep graph neural network framework for single-cell multi-omics data fusion(scMOGAE), which integrates scRNA-seq data and scATAC-seq data. Specifically, scMOGAE first estimates cell-cell similarity for each modality and then employs graph convolutional autoencoders to learn their latent embedded representations, respectively. Subsequently, scMOGAE aligns and weights adaptively to obtain joint representations of these two modalities for cell clustering. Furthermore, by incorporating cross-modality prediction in the training process, scMOGAE is capable of imputing missing data. Extensive experiments on multiple datasets demonstrate that scMOGAE achieves superior performance for single-cell clustering.