DESeq2 different experiments comparison
1
0
Entering edit mode
ribioinfo ▴ 100
@ribioinfo-9434
Last seen 3.6 years ago

Hi, I have two experiments in the first I have cells of type A treated and cells of type A not treated.

In the second I have cells of type B wild type and cell of type C wild type.

I would like to perform these comparisons: A treated VS A not treated and B wild type VS C wild type and compare the FC between the two analyses of some genes.

What is the best method to do this using DESeq2?

Should I perform two separate analyses with DESeq2 and then compare the FC?

Or should I load the four conditions in the same object and to use the contrast in order to do the comparisons?

Thank you.

Riccardo

 

deseq2 • 4.5k views
ADD COMMENT
0
Entering edit mode
I have biological replicates for every cell types.
ADD REPLY
0
Entering edit mode
Thank you for your answer. I would like to know if splitting the analysis in two and then compare the FC could be helpful when you cannot correct the batch effects or it has not advantages compared to perform the same analysis with all samples together. Thank you. Riccardo
ADD REPLY
0
Entering edit mode

I don't think there is an advantage to splitting in your case. It could really worsen the results for the 2 vs 2.

ADD REPLY
0
Entering edit mode

Thank you, I am sorry I meant in general when you have enough replicates not in my case.

 

ADD REPLY
0
Entering edit mode

It's a bit to vague for me to make a statement on, I guess the text i have in the FAQ is what i'd like to offer as advice for this situation.

ADD REPLY
0
Entering edit mode

Thanks, I did two plots of the dispersion;

Here there is plot of the first experiment: http://i.imgur.com/ttm5L46.jpg

Here there is plot of the second experiment: http://i.imgur.com/a5ngLtA.jpg

Putting the two experiments together I think that the shrinkage and dispersion of some genes are increased, am i right?

 

ADD REPLY
1
Entering edit mode
@mikelove
Last seen 2 hours ago
United States

We have a FAQ in the vignette about this: to compare all samples together, or to split out into two datasets. How many replicates do you have per group?

ADD COMMENT
0
Entering edit mode

Hi I have 4 replicates for the first experiment and 2 replicates for the second experiment. I saw the FAQ, but is it correct to create an object with all the samples together in my case? I think that in the case of the vignette you have  the same cell type and different treatments. In my case I have 2 experiments did by two different groups and I have also different cell types in the two experiments.

And I have also seen this topic DESeq/DESeq normalization on different experiments that suggest do not normalize all the samples together.

Maybe could I plot a PCA and check the variability between the samples and to choose if perform on analysis with all the samples or split the analysis?

Thanks.

Riccardo

ADD REPLY
0
Entering edit mode

So you have 4 + 4 + 2 + 2 = 12 samples total?

My tendency, because n=2 is bad for estimating within-group variance, that you put all the samples together. You should create a variable in the colData called 'group' which has levels Ac, At, B, C. You can then use the contrast argument of results to make comparisons of pairs, e.g. contrast=c("group","At","Ac"), etc.

ADD REPLY
0
Entering edit mode

Thank you, yes 12. 

In general if you have replicates you can follow the method of the PCA provided in the FAQ even if you have different experiments, different cell types etc?

In this topic DESeq/DESeq normalization on different experiments it is suggested to split the analysis because it refers to DESeq and not DESeq2?

Thank you.

Riccardo

 

ADD REPLY
0
Entering edit mode

You can put samples together. That's what we do when we put control and treated samples together, for example.

The question in that thread was, is there a benefit to normalizing sample together, which there isn't really. So if you're just interested in normalization, there's no reason to put the samples together to normalize, and then split them apart for the rest of the analysis.

However there is a benefit to estimating dispersion (a measure of within-group variance) across more samples, in the case that some of your groups have very small sample size (where n=2 is considered very small sample size). And then you can use results() to compare pairs of groups.

ADD REPLY
0
Entering edit mode

Thank you. Here I uploaded the rlog PCA blind = T of my data http://imgur.com/a/4hE9f.

As you can see on the left there is the first experiment and on the right the second experiment. There is a lot of variability in the PCA1 maybe due to batch effects. In this case is it better to perform two different analyses because of this variability or is it better to have a the problem of the dispersion (because I have only two replicates) ?

Thank you very much.

Riccardo

ADD REPLY
0
Entering edit mode

I would analyze the data all together. The differences across experiment are not a problem because they are modeled in the coefficients.

ADD REPLY
0
Entering edit mode

Thanks. Since I do not have shared samples across the experiments I cannot correct the batch effects so my colData will have only the column condition?

Thank you.

Riccardo

ADD REPLY
0
Entering edit mode

Yes, the difference across condition is taken care of with a single variable group/condition (whatever you choose to call it) that has levels Ac, At, B, C.

ADD REPLY

Login before adding your answer.

Traffic: 989 users visited in the last hour
Help About
FAQ
Access RSS
API
Stats

Use of this site constitutes acceptance of our User Agreement and Privacy Policy.

Powered by the version 2.3.6