Batch Effect

[emolinari]

Hi guys,

I am currently struggling over an issue which I am not really sure how to solve.
In brief, I have run 4 sessions of PE seq on human cells (3 cohorts, Young, Old and Control). I have followed the Tophat-Cufflinks-Cuffmerge-Cuffdiff pipeline and visualized the data on Cummerbund.
When running the MDS plot I see 3 different clusters, and I can clearly assess that the samples cluster according to the sequencing session rather than the proper cohort. I've checked the quality of the data several times, and all the logs look ok. The only "weird" behavior is the properly mate reads rate, that for one group is on average 80%, for the other 72% and the third 60%.

Could this thing alone determine such a strange clustering or is rather a "batch effect"? Any suggestions for solving it?
BTW, I also have HTSeq counts of this data...do you thing I should use that on a different program, such as DeSeq or EdgeR???

Please help!!!

Manu

[dpryan]

That's a pretty classic batch effect. Use the SVA package (specifically, ComBat()) with DESeq2/edgeR/limma and you'll get more meaningful results. As to why this occurred, who knows. I've seen prominent library creation date batch effects before, so if the libraries were made on different dates then that could certainly be the original source of the problem.

[emolinari]

That's a pretty classic batch effect. Use the SVA package (specifically, ComBat()) with DESeq2/edgeR/limma and you'll get more meaningful results. As to why this occurred, who knows. I've seen prominent library creation date batch effects before, so if the libraries were made on different dates then that could certainly be the original source of the problem.

Thanks dpryan for your comment,
The libraries where indeed prepped on different dates, and cluster accordingly (with the only exception of one sample, that was prepped alone one day and clusters with the big group on the right). I will try to do what you suggest...still you don't think that the properly paired mates rate has an influence in determining the clustering? I am asking because maybe there is a way to try to fix it...

Thanks again!

[dpryan]

That could be the cause as well (my guess would be that it's not, but that's just a guess). Just subset the alignments to contain only properly paired alignments and then look at the PCA plot. If the clustering goes away then you know that's the cause and will have also solved the problem (though you'd be throwing information away, so you still might get slightly better results using ComBat()).

[emolinari]

That could be the cause as well (my guess would be that it's not, but that's just a guess). Just subset the alignments to contain only properly paired alignments and then look at the PCA plot. If the clustering goes away then you know that's the cause and will have also solved the problem (though you'd be throwing information away, so you still might get slightly better results using ComBat()).

I see what you mean, I'll try to that! A quick question though: should I use FPKM values or HTSeq?

[dpryan]

HTSeq (or featureCounts, which is faster).