Automated quality control and cell identification of droplet-based single-cell data using dropkick

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Droplet-based single-cell RNA sequencing (scRNA-seq) generates large datasets but faces challenges in distinguishing true cells from technical noise.
• High technical noise, such as batch-specific ambient RNA, confounds accurate cell identification and complicates analysis of datasets with varying library sizes.

Purpose of the Study:

• To present dropkick, a fully automated software tool designed for quality control and filtering of scRNA-seq data.
• To effectively exclude ambient RNA barcodes and recover genuine cells that may be near the quality threshold.
• To provide a robust and reproducible method for cell identification in scRNA-seq analysis.

Main Methods:

• dropkick utilizes a weakly supervised machine learning approach to learn gene-based representations of real cells and ambient noise.
• The tool automatically determines dataset-specific training labels using predictive global heuristics.
• It calculates a cell probability score for each barcode to facilitate filtering.

Main Results:

• dropkick demonstrated superior performance in excluding empty droplets and noisy barcodes compared to conventional thresholding and EmptyDrops.
• The software showed greater recovery of rare cell types, particularly in datasets with high ambient RNA background.
• dropkick's model proved robust to dataset-specific variations, establishing a more reliable multidimensional boundary for cell identification.

Conclusions:

• dropkick offers a fast, automated, and reproducible solution for cell identification in scRNA-seq data.
• The tool enhances downstream analysis by improving the accuracy of cell identification and the recovery of biologically relevant signals.
• dropkick is compatible with popular single-cell Python packages, promoting widespread adoption in the research community.