I-Impute: a self-consistent method to impute single cell RNA sequencing data

Single-cell RNA-sequencing (scRNA-seq) is essential for the study of cell-specific transcriptome landscapes. The scRNA-seq techniques capture merely a small fraction of the gene due to \"dropout\" events. When analyzing with scRNA-seq data, the dropout events receive intensive attentions. Imputation tools are proposed to estimate the values of the dropout events and de-noise the data. To evaluate the imputation tools, researchers have developed different clustering criteria by incorporating the ground-truth cell subgroup labels. There lack measurements without cell subgroup knowledge. A reliable imputation tool should follow the \"self-consistency\" principle; that is, the tool reports the results only if it finds no further errors or dropouts from the data. Here, we propose \"self-consistency\" as an explicit evaluation criterion; also, we propose I-Impute, a \"self-consistent\" method, to impute scRNA-seq data. I-Impute lever-ages continuous similarities and dropout probabilities and refines the data iteratively to make the final output self-consistent. On the in silico data sets, I-Impute exhibited the highest Pearson correlations for different dropout rates consistently compared with the state-of-art methods SAVER and scImpute. On the datasets of 90.87%, 70.98% and 56.65% zero rates, I-Impute exhibited the correlations as 0.78, 0.90, and 0.94, respectively, between ground truth entries and predicted values, while SAVER exhibited the correlations as 0.58, 0.79 and 0.88, respectively and scImpute exhibited correlations as 0.65, 0.86, and 0.93, respectively. Furthermore, we collected three wetlab datasets, mouse bladder cells dataset, embryonic stem cells dataset, and aortic leukocyte cells dataset, to evaluate the tools. I-Impute exhibited feasible cell subpopulation discovery efficacy on all the three datasets. It achieves the highest clustering accuracy compared with SAVER and scImpute; that is, I-Impute displayed the adjusted Rand indices of the three datasets as 0.61, 0.7, 0.52, which improved the indices of SAVER by 0.01 to 0.17, and improved the indices of scImpute by 0.19 to 0.4. Also, I-impute promoted normalized mutual information of the three datasets by 0.01 to 0.09 comparing with SAVER, and by 0.15 to 0.34 comparing with scImpute. I-Impute exhibits robust imputation ability and follows the \"self-consistency\" principle. It offers perspicacity to uncover the underlying cell subtypes in real scRNA-Seq data. Source code of I-Impute can be accessed at https://github.com/xikanfeng2/I-Impute.