Comparing two groups from different RNA-seq datasets

[dpryan]

There are still batch effects in RNAseq, though they're certainly less of an issue than in the microarray days. I happen to be looking at all of the publicly available mouse hippocampus RNAseq datasets at the moment and decided to create a little heatmap of the variance stabilised data, which you can find below. The datasets are color coded the same on the rows and columns to make life easier (there are 160 samples in the heatmap, so the labels are illegible). While there are obvious experimental differences in some of these datasets, there's still a lab batch-effect. Having said that, if you're interested in different organs or something like that then the difference due to that will be vastly greater than the batch-effect, so rskr's advise should hold-up quite well.

BTW, some of the red-colored samples are technical replicates that I never bothered merging, which is why they cluster the way they do.

[rskr]

Is it worth trying to normalize to housekeeping genes or something similar to that?

I wouldn't normalize anything, but using housekeeping genes to verify your statistical model might not be a bad idea.

[id0]

Is it worth trying to normalize to housekeeping genes or something similar to that?

[rskr]

Since with sequencing, there is much more confidence in the observations, which have a built in noise check during the mapping, and it is necessary to account for the overall differences in reads anyway comparing samples from different labs should be straight forward, unlike micro arrays, where batch effects were known to dominate in some cases requiring a high number of technical replicates.

After taking overall read depth into account, you would have to look for subtle effects like bias in GC annealing temperatures or PCR duplicates. Verifying the results in the lab may be difficult though, as with any meta-analysis.

[TiborNagy]

Well, try to download as many type A and type B datasets as you can and use it as technical repeats. A simple hierarchical clustering will show you discepancies.

[dpryan]

I'd be very hesitant to do such a comparison due to the uncontrolled batch effect. I suppose you could try to estimate what sort of batch effect exists by looking at the effect between only the B or A samples, but I'm not sure how well that would work in practice.

[id0]

I am interested in comparing samples of type A to samples of type B. There are several RNA-seq datasets in GEO that have either type A or type B, but not both. Is it possible to take samples from two different datasets and compare them? I am guessing most of the observed differences will be between the two labs and not between the two conditions. Is that a reasonable concern? Is there a proper way to deal with that?