Question

Deseq2 using principle components on raw or log normalized counts as covariates

0

Entering edit mode

rowleyrolls • 0

@rowleyrolls-9115

Last seen 10.0 years ago

United States

Hi,

I am using svd to obtain the first principle components to use a covariates in differential expression testing in RNAseq data. Should I obtain the principle components on the raw counts or following log adjustment? thanks.

myData <- myCountTable
logMyData = log(myData + 1)
ruv1 = svd(logMyData - rowMeans(logMyData))$v[,1]
ruv2 = svd(logMyData - rowMeans(logMyData))$v[,2]

df = data.frame(condition,ruv1,ruv2)

ddsFullCountTable <- DESeqDataSetFromMatrix(
countData = myCountTable,
colData = df,
design = formula(~ ruv1 + ruv2 + condition))
dds <- ddsFullCountTable
dds <- DESeq(dds)

deseq2 rnaseq normalization • 1.0k views

ADD COMMENT • link updated 10.1 years ago by Michael Love 43k • written 10.1 years ago by rowleyrolls • 0

score 2 · Answer 1 · 2015-11-05

X * beta is on the log normalized count scale. I'd be careful, and suggest reading over the SVA paper first, or perhaps best to just use svaseq or RUVseq. The first PCs as you've calculated them might very well be correlated with condition, which would not produce the result you'd like. Methods like svaseq and RUVseq typically do something like SVD on the residuals in the log normalized count space.