FPKM quantile normalization

Question

[Sara]

Hi all, I have 250 samples from healthy and disease states. I want to integrate gene expression data into metabolic model and do flux balance analysis. Can I use FPKM directly for this work or should I normalize FPKM? For example in some publications I see that some researchers used quantile normalization of FPKM.

Any help is welcome

This question has been asked in bioconductor support too. Here is the link

[Kevin Blighe]

I see that some researchers used quantile normalization of FPKM

Who? Data on FPKM scale is already normalised, but not for cross-sample differences.

[Sara]

For example see this paper : https://www.pnas.org/content/115/50/E11874.short At the supplementary data, FPKM quantile normalization has been explained.

[WouterDeCoster]

I'm not familiar with metabolic modeling, but it's not because it has been published before that it is correct.

[Kevin Blighe]

Indeed, you may want to check on CrossValidated StackExchange about the feasibility of performing quantile normalisation on FPKM data. It does not feel right to me. A better transformation would be to Z-scores, via zFPKM package in R.

[ATpoint]

Maybe check quantro, a framework to test if your dataset fulfills the assumptions of quantile normlization.

Answer 1

[WouterDeCoster]

FPKM is not a perfect normalization method. I'd suggest you extract normalized counts from DESeq2.

[Sara]

Deseq normalization is a very good normalization methods in several studies but in metabolic modeling and integration of gene expression to metabolic network is not useful. Because it does not Normalize gene length.

[Kevin Blighe]

FPKM data is not suitable for cross-sample comparisons. DESeq2 can indeed perform adjustment for gene length when you import via tximport.

[Sara]

Many thanks Kevin, I have a look at tximport, but I confused. How can i do that? I have gene-level counts, not transcripts level.

[Sara]

I finally found the way to do this. Here is the R code for whoever needs it.

cts <- as.matrix(read.csv("Count_data.txt", row.names = 1, header= TRUE, sep="\t"))
coldata <- read.csv("coldata.txt", row.names = 1, sep="\t")
dds <- DESeqDataSetFromMatrix(countData = cts, colData = coldata, design =  ~ condition)
txdb <- makeTxDbFromGFF("gencode.v19.annotation.gtf_withproteinids", format="gtf", circ_seqs=character())
ebg <- exonsBy(txdb, by="gene")
rowRanges(dds) = GRangesList(ebg)
FPKM = fpkm(dds)

[Kristoffer Vitting-Seerup]

You probably need some additional transformation to make it suitable for modeling - the distribution of FPKM (even after inter-library normalization) is not standard!

[Sara]

Do you mean for example log transformation?

[Kristoffer Vitting-Seerup]

Yep - and some methods also require scaling. Also remember some ML methods have large problems with highly correlated features

[Kevin Blighe]

Sorry, I was flying across the ocean. Glad that you got the solution. Where are the gene lengths being stored in your code, though? - presumably from gencode.v19.annotation.gtf_withproteinids

[Kevin Blighe]

You asked the question here: https://support.bioconductor.org/p/116021/#116067

When you 'cross-post' to another website, please inform us so that our efforts are not in vain.

[Sara]

Yes, when I did not receive any reply from you, because I was in a hurry to do my project, I asked my question in Bioconductor support. Ok, I will link both question to each other. Thanks.

[ATpoint]

You should not expect live-support from a community driven by volunteers.

[Sara]

Many thanks Kevin. I downloaded the genecode version 19 gtf file and put it in my document. I am so thankful for your valuable advice to use Deseq2 normalization with gene adjustment.