scADGH: scRNA-seq clustering utilizing on attention-based DAE and hybrid similarity GAE

The high dimensionality of single-cell sequencing (scRNA seq) data and the noise generated during the preparation process pose challenges to single-cell classification. Existing data denoising methods uniformly process gene expression values without distinguishing between important and unimportant features. Considering the potential feature relationships between cells, traditional similarity graph construction methods usually choose a single similarity measure, which cannot capture both local and global information simultaneously, resulting in insufficient utilization of information. Based on this, this paper proposes a single cell clustering method (scADGH) using attention based denoising autoencoder (DAE) and mixed similarity graph based graph autoencoder (GAE). The algorithm first uses attention based DAE to denoise the data, helping the model identify key features and reducing the impact of noise on the KNN algorithm. Secondly, a similarity graph is constructed using normalized Euclidean distance and Pearson correlation coefficient to comprehensively reflect the correlation of the data and better capture the potential features between cells. Then, a graph convolutional autoencoder is used to preserve the topological structure information between cells and reduce the dimensionality. Finally, k-means is used to cluster the reduced information. The analysis of 8 scRNA-seq datasets shows that the scADGH method outperforms other methods and exhibits excellent clustering performance.