Unbiased visualization of single-cell genomic data with SCUBI

• Published in: Cell reports methods 2022-01, Vol.2 (1), p.100135, Article 100135
• Main Authors: Hou, Wenpin, Ji, Zhicheng
• Format: Article
• Language: English
• Subjects: data mining; single-cell genomics; unbiased; visualization
• ISSN: 2667-2375; 2667-2375
• DOI: 10.1016/j.crmeth.2021.100135

Visualizing low-dimensional representations with scatterplots is a crucial step in analyzing single-cell genomic data. However, this visualization has significant biases. The first bias arises when visualizing the gene expression levels or the cell identities. The scatterplot only shows a subset of cells plotted last, and the cells plotted earlier are masked and unseen. The second bias arises when comparing the cell-type compositions across samples. The scatterplot is biased by the unbalanced total number of cells across samples. We developed SCUBI, an unbiased method that visualizes the aggregated information of cells within non-overlapping squares to address the first bias and visualizes the differences of cell proportions across samples to address the second bias. We show that SCUBI presents a more faithful visual representation of the information in a real single-cell RNA sequencing (RNA-seq) dataset and has the potential to change how low-dimensional representations are visualized in single-cell genomic data. [Display omitted] •Cell masking biases the visualization of low-dimensional representations•SCUBI addresses the masking bias by aggregating information in non-overlapping squares•Unbalanced cell numbers bias the visual comparisons across samples•SCUBI enables visualization of cell-type proportion and gene expression without bias With the increased number of cells and samples in single-cell genomic data, visualizing the low-dimensional representations with scatterplots is often biased. Two common types of biases happen either due to cells being masked by other cells or an unbalanced total number of cells when comparing across samples. These biases are often overlooked and rarely reported in the literature but may negatively impact downstream analyses or even lead to misinterpretation of data. We report these two types of biases and present a software tool, SCUBI, which overcomes these biases. The visualization bias of low-dimensional representations in single-cell genomic data is often overlooked. Hou et al. demonstrate two types of biases leading to data misinterpretation and present SCUBI, a tool that addresses these biases by aggregating information in non-overlapping squares or by controlling the total number of cells across samples.