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

• Published in: Genome Biology Vol. 21; no. 1; p. 9
• Main Authors: Tsuyuzaki, Koki, Sato, Hiroyuki, Sato, Kenta, Nikaido, Itoshi
• Format: Journal Article
• Language: English
• Published: London BioMed Central 20.01.2020; Springer Nature B.V; BMC
• Subjects: Accuracy; Algorithms; Animal Genetics and Genomics; Benchmarking; Benchmarking Studies; Bioinformatics; Biomedical and Life Sciences; Cell cycle; Cellular heterogeneity; Clustering; Computer applications; Data analysis; data collection; Datasets; Dimension reduction; Evolutionary Biology; Gene expression; genome; Genomics; guidelines; Human Genetics; Life Sciences; memory; Microbial Genetics and Genomics; Online/incremental algorithm; Pancreas; Plant Genetics and Genomics; Principal Component Analysis; Randomized algorithm; Ribonucleic acid; RNA; RNA-Seq - methods; sequence analysis; Single-Cell Analysis - methods; Single-cell RNA-seq; Software; Julia; Dimension reduction; Cellular heterogeneity; R; Out-of-core; Single-cell RNA-seq; Sparse data format; Online/incremental algorithm; Randomized algorithm; Principal component analysis; Python
• ISSN: 1474-760X; 1474-7596; 1474-760X
• DOI: 10.1186/s13059-019-1900-3

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.