There are so many papers performing PCA analysis with only top high variance genes(top 500 for example), and the plot seems good. Cells in different conditions separated from each other, but I am wondering is that reasonable? Since all the genes in my expression data can represent all the characteristic in my samples, only top 500 high variance selected for PCA analysis seems too artifical and could be on purpose just for separating samples out?
•
written
by
bioinfo_wen •
2 answers
One of the assumptions of several microarray and RNAseq analysis packages (e.g. limma, edgeR and DESeq2) is that the number of differentially expressed genes is not too large in relation to the total number of genes being measured. So although 500 is an arbitrary number, yes, it does make sense not to use all genes for these exploratory analyses.
If you did it with all of the genes, the plot would just be bigger. Performing PCA with only the high variance genes is to differentiate among those which are close to one another to begin with.
Log in to answer this question.
-
scRNAseq analysis with a few samples (n=2 per group)
written by Song •
-
RNAseq meta-analysis to identify “consistently expressed” genes
written by cgibbsm •
-
Number of accessible regions included by DESeq2 to generate the PCA plot with the nf-core ATACseq p…
written by Alexandre Eeckhoutte •
-
single-cell pseudotime analysis with monocle3
written by piotto •
-
What is the best practice for the analysis of single-cell RNA-seq of mouse and human mixed cells?
written by Dan •
-
Forum: 100% variance explained on one PC in DESeq2 PCA?
written by melatoninixo •
-
Why do we log transform and scale data prior to PCA for scRNA-seq analysis
written by Lucy •
-
Quality Check of Samples and Biological Replicates
written by doinelpierrot •
-
Batch effect in RNAseq
written by sure •
-
Principal Component Analysis: How many most variable genes should I take?
written by tlorin
I think ideally, you would want to use some kind of statistical method to find genes that vary significantly among all your samples. And then plot a PCA with only those genes. There really isn't any point to plot genes that don't vary significantly as they don't offer any information.
I am not sure if there are any good/easy methods for doing that right now, so that's probably why you see a lot of papers just plotting the top X most varying genes.