Qualitative comparison of different RNA-Seq experiments

[TimK]

Hi,

I'm trying to do a qualitative comparison of publicly available RNA-Seq datasets from different bacteria. Given that the growth conditions are comparable is it possible, e.g. based on RPKM, to make a qualitative comparison of expression levels of conserved genes?

Cheers,
Tim

[Richard Finney]

What do you think the problems might be?

[TimK]

I'm not sure how to compare them. Usually I'd go with something like edgeR or similar for DE analysis. But in this case I'm not sure if that's suitable because it's different bacteria, different experimental setups etc. Therefore, I guess it's more a "gene A of orgA relatively highly expressed, in organismB it's not" ... but what normalization/approach to us? Does that make my problem more clear?

Tim

[Richard Finney]

You are, indeed, going to have the problems you mentioned.
The data is from various sources, prepared differently, using different methods, on different machines analyzing expression for sequences that , though conserved, are different sequences.

Comparing them with deep certainty about the results is going to be challenging.

Back in the day, I remember some guys using all the CEL files for Affymetrix expression chips and comparing expression between the various expressions from various sources in NIH's GEO (gene expression omnibus) database. A lot easier with a single chip type and single manufacturer's recommended procedure.

This EdegeR or DESEQ type approach are good for getting a measure of expression differences, that's not your problem.

The noise and biases of the various methods used to create the reads is your problem.

Perhaps limiting data to the best/most used protocols and sequencing machine might a good approach.

[Asaf]

Maybe you should compare pairs

You're going to stumble into many normalization issues here. I'm having trouble comparing experiments from the same lab in two different experimental settings, let alone different labs and bacteria.
I would try another approach using simple statistics. I assume you have enough replicates so you can pair the experiments, taking the two conditions in each bacteria, growth phase, lab etc. and in each pair determine, for each gene, if it was up or down regulated (you'll have to normalize them, using DESeq normalization or other method). Then, if you have about 15 (or even less, depends on the number of genes you're testing) such pairs you can run the sign test on each gene. You should get p-values with sufficient power even after FDR correction.

[mbblack]

If it were me, I would not even try to merge the two datasets. Analyze each independently just as you would any other cross-species comparison. Pull out your two independent gene lists and look for the overlap between them. Then you can make some comparisons of the commonly detected differential expression in each. First thing I'd want to see is do the two gene lists even rank order correlate at all. You can then do some stats if you wished to compare the independently derived gene lists.

[TimK]

Thanks everybody.
The "simple stats" solution is what I was thinking of, too. I will have a look into sign test and rank order conservation. Thanks for the help so far.