Benchmarking principal component analysis for large-scale single-cell RNA-sequencing

Abstract

Background: Principal component analysis (PCA) is an essential method for analyzing single-cell RNA-seq (scRNA-seq) datasets, but for large-scale scRNA-seq datasets, computation time is long and consumes large amounts of memory. Results: In this work, we review the existing fast and memory-efficient PCA algorithms and implementations and evaluate their practical application to large-scale scRNA-seq datasets. Our benchmark shows that some PCA algorithms based on Krylov subspace and randomized singular value decomposition are fast, memory-efficient, and more accurate than the other algorithms. Conclusion: We develop a guideline to select an appropriate PCA implementation based on the differences in the computational environment of users and developers.

pipeComp, a general framework for the evaluation of computational pipelines, reveals performant single cell RNA-seq preprocessing tools

“…Since the various PCA approaches and implementations were recently benchmarked in a similar context [17], we focused on widely used approaches that had not yet been compared: Seurat's PCA, scran's denoisePCA, and GLM-PCA [40]. When relevant, we combined them with sctransform normalization.…”

“…In addition, we did not compare any of the alignment and/or quantification methods used to obtain the count matrix, which was for instance discussed in [18]. Some steps, such as the implementation of the PCA, were also not explored in detail here as they have already been the object of recent and thorough study elsewhere [14,17]. We also considered only methods relying on Euclidean distance, while correlation was recently reported to be superior [57] and would require further investigation.…”

“…A number of good comparison and benchmark studies have already been performed on various steps related to scRNAseq processing and analysis and can guide the choice of methodology [3][4][5][6][7][8][9][10][11][12][13][14][15][16][17][18][19][20][21]. However, these recommendations need constant updating and often leave open many details of an analysis.…”

“…Another missing aspect of current benchmarking studies is their limitation to capture all aspects of a scRNAseq processing workflow. Although previous benchmarks already brought valuable recommendations for data processing, some only focused on one aspect of data processing (e.g., [14]), did not evaluate how the tool selection affects downstream analysis (e.g., [17]) or did not tackle all aspects of data processing, such as doublet identification or cell filtering (e.g., [18]). A thorough evaluation of the tools covering all major processing steps is however urgently needed as previous benchmarking studies highlighted that a combination of tools can have a drastic impact on downstream analysis, such as differential expression analysis and cell-type deconvolution [3,18].…”

Impact of Data Preprocessing on Integrative Matrix Factorization of Single Cell Data

“…Linear regression using nonlinear iterative partial least-squares (NIPALS), eigen analysis, or singular value decomposition (SVD) are a few of the many ways to factorize or decompose a matrix. SVD is a basic matrix operation, and fast approximations of SVD, including IRLBA, are commonly applied to sc data [extensively reviewed by (45)]. (3) concatenating matrices and applying a matrix factorization, usually singular value decomposition (SVD); and (4) visualizing results.…”