Deep information fusion based on a transformer graph encoder for single-cell multi-omics clustering

Abstract Single-cell multi-omics clustering has emerged as a critical technology for deciphering cellular heterogeneity and functional diversity, enabling the simultaneous measurement of multiple omics layers within individual cells. Nevertheless, the inherent characteristics of single-cell multi-omics data, such as high noise, sparsity, and heterogeneity, continue to pose significant challenges to achieving accurate clustering analyses. Consequently, the effective integration of multi-omics data to enhance clustering performance remains a critical focus in current research. To overcome these challenges, we propose scTGIC, a clustering method based on a transformer graph autoencoder (TGAE) for deep information fusion. The TGAE integrates a multihead attention mechanism with local structural similarity, fusing the normalized adjacency matrix with the attention matrix to directly model multi-hop relationships and higher-order topological features, optimizing inter-node topology and overcoming the limitations of traditional graph convolutional neural networks (GCNs) in capturing global patterns. Furthermore, we introduce structural information in the information fusion mechanism, which combines a collaborative supervised clustering strategy and a dual-level redundant reduction mechanism. The experimental results demonstrate that the scTGIC exhibits strong competitiveness across five single-cell multi-omics datasets, providing more robust and reliable clustering results.