Why some microarray data have negative value? How to fix it?

Question

[Morteza Razavi]

This is my data with negative values. ftp://ftp.ncbi.nlm.nih.gov/geo/series/GSE29nnn/GSE29745/matrix/GSE29745_series_matrix.txt.gz This is one color technology. What negative values mean in one color technology?

[Morteza Razavi]

Is it possible to use it beside other data when I use ComBat function from SVR (Surrogate Variable Analysis) package?

[Kevin Blighe]

I have moved your reply to become a comment. In the future, when you want to comment on an answer, click on the ADD COMMENT button beneath the answer.

Why do you need to use ComBat? Have you evidence that there exists surrogate variables in the data?

[Morteza Razavi]

I want to use it for a meta analysis. I have multiple data sets from multiple platforms with similar issues. And I don't know is it possible to do it with these negative values?

[Kevin Blighe]

I would only use ComBat as a final resort. ComBat was designed for microarray data, so, I imagine that this simple issue of negative values (which is common in microarray data) is managed by ComBat.

If you are actually combining multiple datasets together, it would at least help that they are matched on condition type, lab preparatory method, and array platform. If you have different conditions, prep. methods, and different array platforms, then you are going down a long, dark route...

For differential expression, you can typically include experiment (or batch, etc) as a covariate in the design formula. This will then adjust the statistical inferences based on the cross-experiment / batch differences.

If you aiming to use downstream tools, you may indeed need to directly remove batch effects.

[Morteza Razavi]

Thanks Kevin, I have made a matrix by combining 8 multiple microarray based on common Gene Symbols and my aim is to remove the batch effects and drawing a correlation heatmap to recognize the similarities between data sets. Also, I don't need the differential expression of genes. What are the other methods to do this except ComBat?

[Kevin Blighe]

You should generate a PCA bi-plot to see if there exists evidence of differences between the datasets. Please take a look at my PCAtools package: PCAtools: everything Principal Components Analysis (for now, you will have to install it with: devtools::install_github('kevinblighe/PCAtools')).

You can also generate a bi-plot using base R functions: A: PCA plot from read count matrix from RNA-Seq

Remember that you should not blindly use ComBat (or other batch-adjustment methods) without any justification.

Answer 1

[Kevin Blighe]

The data in that file is already normalised and log2 transformed, so, it is quite possible that many values will be negative. However, you can still use it for, e.g., differential expression analysis. Negative values are actually common on various microarray platforms.

This code will download the same file and plot a boxplot:

library(Biobase)
library(GEOquery)

gset <- getGEO("GSE29745", GSEMatrix =TRUE, getGPL=FALSE)    
if (length(gset) > 1) idx <- grep("GPL6480", attr(gset, "names")) else idx <- 1
gset <- gset[[idx]]

dev.new(width=4+dim(gset)[[2]]/5, height=6)
par(mar=c(2+round(max(nchar(sampleNames(gset)))/2),4,2,1))
title <- paste ("GSE29745", '/', annotation(gset), " selected samples", sep ='')
boxplot(exprs(gset), boxwex=0.7, notch=T, main=title, outline=FALSE, las=2)

Kevin

[linouhao]

hi, you said negative value is normal, and can I use limma to analyse it directly, because someone else said that the negative value should be removed because of backgroud substraction and log transform

and here do you think the picture you showed is comparable or not? thanks a lot