Structure-guided integrative soft deep clustering analysis of scRNA-seq and scATAC-seq data

Introduction Current single-cell clustering methods often rely on hard clustering assignments, which fail to capture the dynamic and transitional states of cells during development. This study introduces the Structure-Guided Soft Deep Clustering (sgSDC) framework to address this limitation by integrating multimodal data and enabling probabilistic cluster assignments. Methods The sgSDC model combines scRNA-seq and scATAC-seq data using a structure-guided fusion module with global attention. It employs contrastive learning to align modality-specific representations with a consensus representation and introduces a novel soft clustering loss that allows cells to belong to multiple clusters with varying probabilities. Results Evaluations on four benchmark datasets demonstrate that sgSDC outperforms eight state-of-the-art methods in Accuracy (ACC), Normalized Mutual Information (NMI), and Adjusted Rand Index (ARI), achieving significant improvements-up to 52.62% in ARI on one dataset. Discussion The results validate the effectiveness of structure-guided contrastive learning and soft clustering in capturing cellular heterogeneity. sgSDC provides a robust tool for analyzing complex single-cell data, with potential applications in developmental biology and tumor microenvironment research.