PsiNorm: a scalable normalization for single-cell RNA-seq data

Abstract

Motivation Single-cell RNA sequencing (scRNA-seq) enables transcriptome-wide gene expression measurements at single-cell resolution providing a comprehensive view of the compositions and dynamics of tissue and organism development. The evolution of scRNA-seq protocols has led to a dramatic increase of cells throughput, exacerbating many of the computational and statistical issues that previously arose for bulk sequencing. In particular, with scRNA-seq data all the analyses steps, including normalization, have become computationally intensive, both in terms of memory usage and computational time. In this perspective, new accurate methods able to scale efficiently are desirable. Results Here we propose PsiNorm, a between-sample normalization method based on the power-law Pareto distribution parameter estimate. Here we show that the Pareto distribution well resembles scRNA-seq data, especially those coming from platforms that employ unique molecular identifiers (UMIs). Motivated by this result, we implement PsiNorm, a simple and highly scalable normalization method. We benchmark PsiNorm against seven other methods in terms of cluster identification, concordance and computational resources required. We demonstrate that PsiNorm is among the top performing methods showing a good trade-off between accuracy and scalability. Moreover PsiNorm does not need a reference, a characteristic that makes it useful in supervised classification settings, in which new out-of-sample data need to be normalized. Availability PsiNorm is implemented in the scone Bioconductor package and available at https://bioconductor.org/packages/scone/ Supplementary information Supplementary data are available at Bioinformatics online.

Depth normalization for single-cell genomics count data

“…These problems have significant implications for common tasks such as finding marker genes, as discussed above. Newer methods that explicitly couple statistical methods with software engineering considerations are needed; we examined several recent publications proposing new ideas but restricted the paper to widely used methods common in existing workflows (Brown et al 2021; Breda, Zavolan, and van Nimwegen 2021; Borella et al 2021; Bacher et al 2017). A detailed analysis and review of these methods is an important next step.…”

“…Despite the complexity of normalization in practice, much work on scRNAseq has focused on statistical details that, while important, are not necessarily the primary determinants of results. For example, the debate over whether gene-specific over-dispersion parameters should be used when computing Pearson residuals (Hafemeister and Satija 2019; Lause, Berens, and Kobak 2021; Hafemeister and Satija 2020; Choudhary and Satija 2022) ignores the fact that Pearson residuals are not the result of a monotonic transformation, and they create dense matrices that can lead to significant analysis limitations (Borella et al 2021). These problems have significant implications for common tasks such as finding marker genes, as discussed above.…”

“…While many methods have been proposed for single-cell RNA-seq normalization (Cole et al 2019; Tian et al 2019; You et al 2021; Lytal, Ran, and An 2020; Borella et al 2021; Ahlmann-Eltze and Huber 2021; Breda, Zavolan, and van Nimwegen 2021), the approach of equalizing depth for all cells, often to a “size factor” such as ten thousand (CP10k) or one million (CPM), followed by the application of a variance stabilizing transform like log plus one (log1p) is most popular. These methods are implemented in the widely used Seurat 1 and Scanpy (Wolf, Angerer, and Theis 2018) programs, but they do not explicitly model cell depth as a covariate.…”

Transcriptomic Mapping of Neural Diversity, Differentiation and Functional Trajectory in iPSC-Derived 3D Brain Organoid Models

“…We used a commonly applied normalization protocol depth scaling or count per million (CPM) normalization and subsequently log1p transformation. This protocol normalizes count data by a factor proportional to the count depth per cell [ 19 ]. While normalization is used to remove count sampling effects, the resulting dataset still contains unwanted variability due to dropout events, the expression effects of cell cycles, and technical inconsistencies such as batch differences.…”

The study of gene expression statistics

“…Therefore, the use of an appropriate data normalization method can minimize such a kind of errors and allow making a comparison between the results of gene expression obtained in different experiments. Using the assumption about the power-law distribution of gene expression in the works [14,15] the appropriate methods of normalization were developed.…”