Abstract
SummaryDespite the advances in single-cell transcriptomics, the reconstruction of gene regulatory networks remains challenging. Both the large amount of zero counts in experimental data and the lack of a consensus preprocessing pipeline for single-cell RNA sequencing (scRNA-seq) data make it hard to infer networks. Imputation can be applied in order to enhance gene-gene correlations and facilitate downstream analysis. However, it is unclear what consequences imputation methods have on the reconstruction of gene regulatory networks. To study this, we evaluate the differences on the performance and structure of reconstructed networks before and after imputation in single-cell data. We observe an inflation of gene-gene correlations that affects the predicted network structures and may decrease the performance of network reconstruction in general. However, within the modest limits of achievable results, we also make a recommendation as to an advisable combination of algorithms while warning against the indiscriminate use of imputation before network reconstruction in general.
Highlights
Information on the seven cell types was derived from five experimental scRNA-seq datasets: human embryonic stem cell,[24] human hepatocyte,[25] mouse embryonic stem cell,[26] mouse dendritic cell,[27] and mouse hematopoietic stem cell.[28]
These were further separated into the following subtypes: erythrocyte, granulomonocyte, and lymphocyte
We preselected the datasets according to significantly varying transcription factors (TFs) and the most highly variable genes across pseudotime
Summary
GRN reconstruction from single-cell RNA-seq (scRNA-seq) data draw on a plethora of statistical approaches.[1,2,3,4,5,6] Pratapa et al.[4] provide an extensive benchmark study evaluating the performance of various methods. While normalization attempts to correct for different read depths between cells,[11,12] imputation attempts to recover gene counts by predicting missing data and eventually smoothen gene expression values.[13,14,15,16,17,18,19,20] In some tools, a prior normalization step is not required but is integrated within the imputation method.[15,18] Hou et al.[21] extensively evaluated the impact of imputation on clustering, differential expression analysis, and pseudotime inference and invoked cautious interpretations of the results
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
