EnsembleKQC: An Unsupervised Ensemble Learning Method for Quality Control of Single Cell RNA-seq Sequencing Data

摘要

Single cell RNA sequencing (scRNA-seq) provides a view of high-resolution to reveal the cellular heterogenicity. A series of analysis, such as cell-type identification, differential expression analysis, regulatory relationship detection, could uncover unprecedented biological findings. Prior to these downstream analysis, it's crucial to remove low-quality cells because they are technical noises which weaken true biological signal and mislead downstream analysis. Existing methods either require setting threshold manually or require true labels for supervised training, which is not appropriate in many cases. We present an unsupervised ensemble learning method, which could automatically identify low-quality cells from single cell RNA-seq sequencing data. This method integrates weak classifiers base on five selected features from housekeeping genes, reads mapping rate and detected genes. To avoid setting thresholds of classifiers manually, it enumerates threshold values within a reasonable range and chooses the most suitable threshold values based on a scoring function. In experiments, it exhibits high and steady accuracy on multiple datasets.