Hi All,
I am having a bit of trouble analysing some of my data. I've got data from two different experiments involving two different RNASeq runs of WT vs mutant isolates. As the experimental conditions, the library prep, and sequencing was the same, we assumed that we would be able to compare the WT isolates which were run in experiment run (2017) and the mutant isolates sequenced in (2019). My first question would be: is scientifically okay to compare these two data sets?
Secondly, I am getting different results each normalisation approach and DE analysis tool I use. I've normalised the data with several different tools, on a tmm, tss, mrn and rle, running the differential analysis with both EdgeR and DeSeq2 both with Trinity's DE tools, and through cloud-based application MeV (http://mev.tm4.org/#/welcome).
I've attached an image of the EdgeR differential expression matrix (P0.001) in which you can see that the left most (c45_0) and isolates (c622_2014_0, and c622_2015_0) towards the middle left are the WT isolates from 3 different sequence backgrounds to which I would be comparing experimentally generated mutants. Of course when normalising with TMM and then DE with EdgeR there's about 500 genes that are DE, but if I normalise with RLE(DeSeq) and run the data in either EdgeR or DeSeq2, I get far fewer genes (max 50) which is what I would expect when comparing the mutants generated to their WT.
Can someone let me know if:
- The WT and the mutant data I have can even be compared.
- If yes, what is the best way to normalise the data and what is the best tool to run the DE analysis with?
Thank you so much for your help!
I would say no since your conditions are fully confounded. You have no way of checking of the differences are the biological reality or a batch effect due to some technical issues that you cannot control. The thing with batch effects is that they are not intended and therefore impossible to control. I've seen myself that different experiments on the same specimen (same protocl, did all the pipetting myself) can give notably different results. There is unfortunately no computational tool that saves you posthoc from this. Batch correction would require replicates of all experimental groups in each batch.
Yeah, well this is somewhat normal. Different tools use different models etc. The main findings should hold though. I would still decide for one tool and then stick with it. Try to confirm results if possible using independent experiments or published data that are similar. Default settings also may differ between tools so be sure to read documentation carefully.