I'm having problems with RNA-seq data. I have two replicates for five different treatments (all on the same cell line) and also two replicates for the control. I used DESeq2 to perform the differential expression analysis and I obtain that all the treatments are very similar to each other (I mean that the percentage of differentially expressed genes is always around 1 to 3% for each treatment). Only for one treatment I found that the percentage of differentially expressed genes is around 20%. This was unexpected and I can't find out a method to evaluate if it is a biological or a statistical issue. Do you have any idea on which analysis I could perform to confirm that I carried out correctly the differential expression analysis? If it is a biological issue do you have any idea of a possible explanation of such a substantial difference?
A good first exploratory step in RNA-seq is to generate a PCA plot of all of your samples. That gives you a good preliminary idea of which samples are different, and relatively speaking which conditions are more or less different. For PCA I recommend the great PCAtools library. Make sure to explore all dimensions that have an appreciable percentage of variance explained.
Additionally, there are other QC plots to ensure data integrity, such as MA plots and p-value histograms. I would start with PCA plots first and depending on how those look go from there.
You should also include your code just so we can make sure there were no coding errors in your analysis too.
Thanks rpolicastro, for your answer. I have already generated the PCA plot and also a heatmap where I put the sample to sample distance and I see that the the two plots confirm the differences between this one treatment and all the other four. All the plots that I generate from these data are coherent but I think that it is obvious because I'm using always the same data to generate all the different plots. My question is how can I confirm that I carried out correctly the differential expression analysis?
My question is how can I confirm that I carried out correctly the
differential expression analysis?
You can post the code you used as noted by rpolicastro . Multiple pairs of eye can whet what you did. If you are primarily a statistician and are sure about your analysis then perhaps show the results to a biologist to see if the results look reasonable.
...percentage of differentially expressed genes is always around 1 to
3% for each treatment). Only for one treatment I found that the
percentage of differentially expressed genes is around 20%. This was
unexpected...
What was unexpected about it? These numbers are not inconsistent with what we know is possible biologically. What is it that you know about your experiment, that leads you to believe this is an odd result? Given the information in your question there's nothing incompatible with a biological explanation. The PCA plot sets these samples apart from the others, so a large DE response could make sense. If you look at the values for DE genes and non-DE between each condition and control, does the data makes sense to you? Is there a core set of controls or housekeeping genes that exhibit similar behavior/characteristics across the experiment?
Is there anything about the biology in your results that does make sense? Or does NOT make sense (i.e. you applied heat shock, but your cells are exhibiting a cold shock response)? Drawing a line between each condition and at least one or a few genes one might expect to respond should be doable. Is there a confounding factor you haven't mentioned? For instance, was everything controlled for except condition? (i.e. all cells and conditions prepared and applied on the same day in the same laboratory by the same person?) or is there a batch effect whereby the outlier condition was applied 3 months later in a different laboratory on the other side of the world (but maybe nobody mentioned that)?
A good first exploratory step in RNA-seq is to generate a PCA plot of all of your samples. That gives you a good preliminary idea of which samples are different, and relatively speaking which conditions are more or less different. For PCA I recommend the great PCAtools library. Make sure to explore all dimensions that have an appreciable percentage of variance explained.
Additionally, there are other QC plots to ensure data integrity, such as MA plots and p-value histograms. I would start with PCA plots first and depending on how those look go from there.
You should also include your code just so we can make sure there were no coding errors in your analysis too.
Thanks rpolicastro, for your answer. I have already generated the PCA plot and also a heatmap where I put the sample to sample distance and I see that the the two plots confirm the differences between this one treatment and all the other four. All the plots that I generate from these data are coherent but I think that it is obvious because I'm using always the same data to generate all the different plots. My question is how can I confirm that I carried out correctly the differential expression analysis?
You can post the code you used as noted by rpolicastro . Multiple pairs of eye can whet what you did. If you are primarily a statistician and are sure about your analysis then perhaps show the results to a biologist to see if the results look reasonable.