Regulatory network-based imputation of dropouts in single-cell RNA sequencing data

Single-cell RNA sequencing (scRNA-seq) methods are typically unable to quantify the expression levels of all genes in a cell, creating a need for the computational prediction of missing values ('dropout imputation'). Most existing dropout imputation methods are limited in the sense that they exclusively use the scRNA-seq dataset at hand and do not exploit external gene-gene relationship information. Further, it is unknown if all genes equally benefit from imputation or which imputation method works best for a given gene. Here, we show that a transcriptional regulatory network learned from external, independent gene expression data improves dropout imputation. Using a variety of human scRNA-seq datasets we demonstrate that our network-based approach outperforms published state-of-the-art methods. The network-based approach performs particularly well for lowly expressed genes, including cell-type-specific transcriptional regulators. Further, the cell-to-cell variation of 11.3% to 48.8% of the genes could not be adequately imputed by any of the methods that we tested. In those cases gene expression levels were best predicted by the mean expression across all cells, i.e. assuming no measurable expression variation between cells. These findings suggest that different imputation methods are optimal for different genes. We thus implemented an R-package called ADImpute (available via Bioconductor ) that automatically determines the best imputation method for each gene in a dataset. Our work represents a paradigm shift by demonstrating that there is no single best imputation method. Instead, we propose that imputation should maximally exploit external information and be adapted to gene-specific features, such as expression level and expression variation across cells. Author summarySingle-cell RNA-sequencing (scRNA-seq) allows for gene expression to be quantified in individual cells and thus plays a critical role in revealing differences between cells within tissues and characterizing them in healthy and pathological conditions. Because scRNA-seq captures the RNA content of individual cells, lowly expressed genes, for which few RNA molecules are present in the cell, are easily missed. These events are called 'dropouts' and considerably hinder analysis of the resulting data. In this work, we propose to make use of gene-gene relationships, learnt from external and more complete datasets, to estimate the true expression of genes that could not be quantified in a given cell. We show that this approach generally outperforms previously published methods, but also that different genes are better estimated with different methods. To allow the community to use our proposed method and combine it with existing ones, we created the R package ADImpute, available through Bioconductor.