Pengwei Hu, Pengcheng Song, Bingjie Dai, Chunshen Long, Hanshuang Li, Yongchun Zuo, Yongqiang Xing
Single-cell RNA sequencing (scRNA-seq) enables the investigation of alternative splicing (AS) at cellular resolution. However, the analysis of AS in scRNA-seq data is constrained by sparse splice-junction coverage, a consequence of low sequencing depth per cell. This limitation is particularly pronounced in 3'-biased, droplet-based protocols. To overcome this, we developed scASprofiler, a tailored deep convolutional generative network designed to decipher AS with single-cell resolution. scASprofiler performs missing-value imputation of junction read counts by leveraging cells generated by a mask-aware variational autoencoder-generative adversarial network (VAE-GAN), reducing oversmoothing of imputed values and preserving biologically meaningful heterogeneity. Across benchmarks, scASprofiler enhances delineation of cell populations and recovery of splicing signals. When applied to datasets generated using plate- and droplet-based platforms, scASprofiler uncovers cryptic AS events and reveals cell-type-specific AS patterns that complement and extend insights derived from gene expression. Together, our study establishes scASprofiler as a robust and versatile tool for dissecting AS landscapes from scRNA-seq data.