scGACL: a generative adversarial network with multi-scale contrastive learning for accurate single-cell RNA sequencing imputation

Abstract Single-cell RNA sequencing is a powerful technology for investigating cell-to-cell heterogeneity, yet its application is often hindered by dropout events, making accurate imputation essential for downstream analyses. Existing imputation methods, however, frequently suffer from the over-smoothing problem, which results in the loss of cell-to-cell heterogeneity in the imputed outcomes and affects downstream analyses. To overcome this limitation, we propose scGACL, a generative adversarial network (GAN) integrated with multi-scale contrastive learning. The GAN architecture facilitates the distribution of the imputed data to approximate that of the real data. To fundamentally address over-smoothing, the model incorporates a multi-scale contrastive learning mechanism: cell-level contrastive learning preserves fine-grained cell-to-cell heterogeneity, while cell-type-level contrastive learning maintains macroscopic biological variation across different cellular groups. These mechanisms function synergistically to ensure accurate imputation and effectively address the over-smoothing challenge. Comprehensive evaluations across diverse simulated and real-world datasets confirm that scGACL consistently outperforms existing methods in accurately recovering gene expression and improving downstream analyses such as cell clustering, gene differential expression analysis, and cell trajectory inference.