Solo: Doublet Identification in Single-Cell RNA-Seq via Semi-Supervised Deep Learning

Abstract

Highlights d Semi-supervised learning improves doublet classification in single-cell RNA sequencing d Combining our method with experimental approaches further improves accuracy

Benchmarking Computational Doublet-Detection Methods for Single-Cell RNA Sequencing Data

“…A likely reason for this discrepancy is how doublets are annotated in these real datasets. In hm-12k and hm-6k, doublets are annotated as the droplets that contain cells of two species, so all annotated doublets are heterotypic and easy to identify [8][9][10]14 . In contrast, doublets annotated in the other datasets may include homotypic doublets that are difficult to identify, posing a challenge to doublet-detection methods; or they may miss certain heterotypic doublets (e.g., if doublets are defined as the droplets that contain cells from two individuals, then heterotypic doublets formed by cells of different types within an individual would be missed), creating a downward bias in the calculation of detection accuracy (see further discussion in the Supplementary).…”

“…Some methods provide heuristic guidance to estimate the doublet rates or select the threshold on doublet scores. For example, DoubletFinder suggests using the rates of heterotypic doublets and Poisson doublet formation as the respective lower and upper bounds of the expected doublet rate 10,19 ; Scrublet recommends setting the doublet-score threshold in the middle of the two modes, which it expects to appear, in the doublet-score distribution 9 ; Solo sets the doublet-score threshold to 0.5 by default 8 . However, there lacks consensus or direct estimation of the doublet rate from scRNA-seq data.…”

“…Realizing the limitations of experimental strategies, researchers have attempted to tackle this doublet challenge from an alternative perspective: developing computational methods to detect doublets from already-generated scRNA-seq data 5 . So far, nine doublet-detection methods have been developed (with software packages and full-text manuscripts) based on distinct algorithmic designs [8][9][10][14][15][16][17] ( Table 1). Here is a brief summary of these methods except hybrid, which is a combination of two methods: bcds and cxds.…”

“…Here, we conducted the first comprehensive benchmark study of computational methods for doublet detection. We evaluated nine cutting-edge methods-doubletCells 17 , Scrublet 9 , cxds 14 , bcds 14 , hybrid 14 , Solo 8 , DoubletDetection 16 , DoubletFinder 10 , and DoubletDecon 15 -in three aspects. First, we compared their overall doublet detection accuracy using two criteria: the area under the precisionrecall curve (AUPRC) and the area under the receiver operating characteristic curve (AUROC), on a collection of 16 real scRNA-seq datasets containing experimentally annotated doublets.…”

Probing infectious disease by single-cell RNA sequencing: Progresses and perspectives

“…These doublets often consist of T cells and monocytes adhering to each other [110] . Some bioinformatic tools have been created, such as SOLO (a semi-supervised deep neural network model to identify doublets) [111] , Scrublet [112] and DoublettFinder [113] . To avoid losing crucial information relevant to the disease, careful consideration should be taken when deciding whether doublets should be excluded from the sequencing analysis.…”