GCAN: A data feature learning method for scRNA-seq data.

INTRODUCTION: Single-cell RNA sequencing (scRNA-seq) data exhibit extreme sparsity, technical noise, and complex nonlinear structures that obstruct accurate biological interpretation. Current methods inadequately model intercellular relationships and suffer from feature selection instability. PURPOSE: (1) Automatically identify biologically relevant features in noisy scRNA-seq data,(2) Model multi-scale cellular relationships for robust clustering,(3) Provide interpretable representations for mechanistic insights. METHODOLOGY: Adaptive HVG selection: Improved Random Forest to mitigate dropout artifacts. Hybrid graph autoencoder: Fusion of Graph Attention Networks (local interactions) and Graph Convolutional Networks (global neighborhoods). Biologically informed optimization: MMD regularization + Spearman-correlation feature filtering. RESULTS: Evaluated across 17 scRNA-seq datasets: Showed better clustering performance than CellVGAE in our experiments (Silhouette Coefficient and Davies-Bouldin Index),Selection of highly variable genes mitigated technical noise while enhancing biological signal retention. CONCLUSION: GCAN establishes a new paradigm for scRNA-seq analysis by unifying adaptive feature selection with context-aware relational modeling. Its architecture implements adaptive screening of biologically relevant features and enables accurate cell typing.