scKDGM: New AI method beats 10 baselines in single-cell clustering
A KAN-guided dynamic graph model improves cell type identification across 12 datasets.
Single-cell RNA sequencing (scRNA-seq) is critical for identifying cell types, but technical noise, sparsity, and dropout effects make robust clustering difficult. Existing masked autoencoders recover expression but don't feed it back into graph construction, while graph clustering methods often rely on fixed KNN graphs. To bridge this gap, a team of researchers (Jun Tang et al.) introduced scKDGM, a KAN-guided dynamic graph masked learning framework.
scKDGM introduces several innovations: a graph-aware distribution preserving gene masking (GDP-Mask) that perturbs cell identity; a KAN-based TAKGCN encoder to learn masked-view representations; a mask-guided expression recovery module that dynamically updates the cell graph; and cross-view contrastive learning that transfers recovery signals into topology improvements. Additionally, a ZINB loss models overdispersion and zero inflation common in scRNA-seq data. Tested on 12 real scRNA-seq datasets against 10 baselines, scKDGM achieved top performance in average NMI and ARI, demonstrating its effectiveness for cell type clustering.
- scKDGM combines KAN-guided encoding with dynamic graph masking to improve cell graph construction.
- It introduces GDP-Mask for gene perturbation and cross-view contrastive learning to connect expression recovery with graph topology.
- Outperforms 10 baseline methods on 12 real datasets in average NMI (Normalized Mutual Information) and ARI (Adjusted Rand Index).
Why It Matters
More accurate cell type clustering from scRNA-seq data accelerates biological discovery and disease research.