Doublets detection using scrublet

To identify doublets from secondary motor cortex single-nucleus ATAC-seq datasets, we use single-cell RNA-seq doublets detection algorithm Scrublet³⁷. Briefly, Scrublet identifies doublets in the following manner: (1) Scrublet performs normalization, gene filtering, and principal components analysis (PCA) to project the high-dimension data to a low-dimension space; (2) Scrublet simulates doublets by adding the unnormalized counts from randomly sampled observed transcriptomes; (3) the simulated doublets are projected to the low-dimension embedding computed in step 1. The more neighbors of a cell are the simulated doublets, the more likely this cell is a “doublet”. Based on this idea, a KNN classifier was then used to estimate the doublet score for each cell.

Since Scrublet was designed for detecting doublets in single-cell RNA-seq, it is unclear whether it can be used for single-cell ATAC-seq. To examine this, we applied Scrublet to a single-cell ATAC-seq dataset of mixed human and mouse cells where the “ground-truth” doublets can be identified based on the alignment ratio to human and mouse genome. Compared to the ground truth, Scrubet can identify over 90% of the doublets in this dataset with ~90% accuracy (Supplementary Fig. ²⁶). This result suggests that although Scrubet was not developed for detecting doublets in single-cell ATAC-seq, it can find the doublets in scATAC-seq dataset with reasonable accuracy and sensitivity.