Graph-Attentional Deep Sparse Subspace Clustering for Single-Cell Transcriptomics.

Abstract

Accurate identification of cell types constitutes a critical step in downstream analysis of single-cell sequencing data. However, the inherent high noise levels and high dimensionality characteristics pose significant challenges for clustering tasks. To address these issues, we propose method scGADSSC (Graph-Attentional Deep Sparse Subspace Clustering for Single-Cell Transcriptomics), an innovative end-to-end framework that achieves joint optimization and mutual enhancement of graph attention learning and subspace self-representation. It consists of two core collaborative components: (1) A Denoising Autoencoder for explicit modeling and noise reduction of raw expression data; (2) A Graph Attention Autoencoder to learn a discriminative self-expression matrix, which is subsequently used to construct a similarity matrix for spectral clustering-based cell type classification. Comprehensive evaluations across 15 biological datasets demonstrate that scGADSSC outperforms ten state-of-the-art single-cell clustering methods on most test datasets, achieving superior clustering performance.