Fig. 5

Zero values make up a large portion of the scRNA-seq expression counts. It’s expected that a significant portion of these values corresponds to true zero expression or is zero since there is not enough information to impute any other value credibly. The bar plots show the proportion of zero count values that are replaced by a value of 0.5 or higher due to imputation. Many imputation approaches modify a significant fraction of zero values without being correlated with improved downstream analysis performance. In other words, these approaches introduce more bias, which negatively impacts the scRNA-seq expression data analysis. ccImpute outperforms the competitors in downstream analysis while modifying much fewer values. scImpute, DCA, and DeepImpute only work with raw unnormalized datasets and cannot impute the Usoskin dataset. Further, scImpute and DrImpute timed out on the larger datasets