rna seq replicates not clustering

Question

[guillaume.rbt]

Hi all,

I'm currently doing a differential expression analysis on RNA-seq data. After counting the reads with salmon, and normalizing with deseq2, I've plotted the dispersion of my samples thanks to a PCA.

I have 3 replicates per conditions, and I was kind of waiting for them to cluster together, but I found out that there is not really a clustering of replicates, the samples being spread across the PCA.

I'm not really familiar with this kind of output, and I was wondering if not seing clustering of replicates could be normal (due to biological variation between replicates,even with the sam conditions), or if I should worried about something before going to DE analysis.

Thanks for your inputs,

Guillaume

[ATpoint]

Please give some details on the experimental setup: Cell line or primary samples, which organism, which treatment etc.

[guillaume.rbt]

I have 72 samples of Vitis vinifera leaves, with 4 changing treatments, and 3 biological replicates for each set of conditions. Mainly I wanted to know if a DE analysis can still be relevant with a low transcriptomic concordance between biological replicates.

[Kevin Blighe]

What have you actually input to the PCA functions, and which PCA functions have you used? Please show your exact code. Also, have you performed pre-filtering steps on the raw counts prior to normalisation?

[guillaume.rbt]

My input is the "vst" normalized table of counts, filtered for genes with no reads. I used the function plotPCA from the DEseq2 package.

[Kevin Blighe]

Try this code: A: PCA plot from read count matrix from RNA-Seq

You'll get a different answer. Why? - because the DESeq2 plotPCA function filters a large proportion of your genes based on variance prior to performing the PCA transformation.

[Devon Ryan]

You could also simply increase the ntop option to Inf (plotPCA(foo, ntop=Inf) or something like that).

[WouterDeCoster]

How to add images to a Biostars post

Answer 1

[Devon Ryan]

Don't read too much into PCA plots, they're telling you more about whether you have outlier samples than whether continuing with differential expression analysis is worth while.

[guillaume.rbt]

Ok thank for the tip, I will continue to explore the data then.