KAN-Guided Dynamic Graph Improves Single-Cell RNA-seq Clustering.

Single-cell RNA sequencing (scRNA-seq) clustering is vital for identifying distinct cell types, but the data presents significant challenges due to its high dimensionality, sparsity, dropout events, and technical noise. Existing methods often fall short, with masked autoencoders primarily focusing on expression recovery and graph clustering relying on static k-Nearest Neighbor graphs without feedback loops for optimization. This research introduces scKDGM, a novel KAN-guided dynamic graph masked learning framework specifically tailored for scRNA-seq clustering. scKDGM incorporates several innovative components: Graph-Aware Distribution Preserving Gene Masking (GDP-Mask) to perturb cell identity, a KAN-based TAKGCN encoder for learning masked-view representations, and mask-guided expression recovery to construct a dynamic graph. Crucially, it uses cross-view contrastive learning to transfer recovery signals into topology updates, ensuring the graph evolves with improved expression. A Zero-Inflated Negative Binomial (ZINB) loss function is also employed to accurately model data overdispersion and zero inflation. Extensive experiments across 12 real scRNA-seq datasets demonstrate that scKDGM significantly outperforms 10 baseline methods in average Normalized Mutual Information (NMI) and Adjusted Rand Index (ARI), indicating superior performance in robust cell type identification.