Learning association for single-cell transcriptomics by integrating profiling of gene expression and alternative polyadenylation

AbstractSingle-cell RNA-sequencing (scRNA-seq) has enabled transcriptome-wide profiling of gene expressions in individual cells. A myriad of computational methods have been proposed to learn cell-cell similarities and/or cluster cells, however, high variability and dropout rate inherent in scRNA-seq confounds reliable quantification of cell-cell associations based on the gene expression profile alone. Lately bioinformatics studies have emerged to capture key transcriptome information on alternative polyadenylation (APA) from standard scRNA-seq and revealed APA dynamics among cell types, suggesting the possibility of discerning cell identities with the APA profile. Complementary information at both layers of APA isoforms and genes creates great potential to develop cost-efficient approaches to dissect cell types based on multiple modalities derived from existing scRNA-seq data without changing experimental technologies. We proposed a toolkit called scLAPA forlearning association forsingle-cell transcriptomics by combing single-cell profiling of gene expression andalternativepolyadenylation derived from the same scRNA-seq data. We compared scLAPA with seven similarity metrics and five clustering methods using diverse scRNA-seq datasets. Comparative results showed that scLAPA is more effective and robust for learning cell-cell similarities and clustering cell types than competing methods. Moreover, with scLAPA we found two hidden subpopulations of peripheral blood mononuclear cells that were undetectable using the gene expression data alone. As a comprehensive toolkit, scLAPA provides a unique strategy to learn cell-cell associations, improve cell type clustering and discover novel cell types by augmentation of gene expression profiles with polyadenylation information, which can be incorporated in most existing scRNA-seq pipelines. scLAPA is available athttps://github.com/BMILAB/scLAPA.