Abstract <p>Single-cell RNA sequencing (scRNA-seq) data typically contains substantial missing values, which often leads to the loss of critical gene signal information and severely limits downstream analyses. Deep learning-based imputation methods generally outperform shallow approaches in handling scRNA-seq data; however, most methods fail to consider the intrinsic relationships among genes, where gene expression is frequently regulated by other genes. Furthermore, existing methods remain inadequate in capturing complex intercellular spatial relationships and long-range dependencies. Therefore, we propose MGCGImpute, a deep learning model for scRNA-seq data imputation that integrates the multi-scale relational modeling capabilities of mobile graph convolutional networks with the distributional learning advantages of generative adversarial networks, achieving precise data reconstruction through learning enhanced spatial relational patterns and global data distribution characteristics. Experimental validation on multiple representative real-world scRNA-seq datasets using multidimensional evaluation metrics and 5-fold cross-validation demonstrates that MGCGImpute achieves superior performance in both imputation accuracy and biological plausibility.</p> Graphical Abstract <p></p> <p>MGCGImpute is a generative adversarial network framework enhanced with moving graph convolution (MGC), specifically designed for imputing missing events in single-cell RNA sequencing data. The input layer receives the original gene expression matrix with a mask. The generator consists of a multi-layer graph attention network and an MGC module, which are used to learn the topological structure between cells and generate enhanced feature outputs. The discriminator distinguishesbetween generated data and real data, improving imputation accuracy through adversarial training.</p>