huge number of DEGs from cuffdiff

[jamal]

Hi
Im working with RNA-seq data from solexa.I have 32 lanes each belonging to one sample and would like to test 16 vs 16.I have a problem when I run cuffdiff on the data I will get large number of DEGs like 3000 or 4000.I have got 72 DEGs out of this data with edgeR.
I have changed many option but still I have many genes.
Could you help me to get reasonable number of DEGs?
Thank you

[severin]

version

Hi
Im working with RNA-seq data from solexa.I have 32 lanes each belonging to one sample and would like to test 16 vs 16.I have a problem when I run cuffdiff on the data I will get large number of DEGs like 3000 or 4000.I have got 72 DEGs out of this data with edgeR.
I have changed many option but still I have many genes.
Could you help me to get reasonable number of DEGs?
Thank you

Hi Jamal, I would start by asking if you have the lastest version of Cuffdiff and to ask what threshold of FDR you are using for edgeR. I know in Cuffdiff they use an FDR threshold of 0.05

[jamal]

Hi Severin

Im using the newest version of cufflinks,and I have tried FDR 0.05 and 0.01.also Im trying different options which I think they can make less number of DEGs like minimum number of alignments in a locus,I have set it to even 5000 but still too many genes.that will be great if any body can help.
Thanks

[Simon Anders]

Hi Jamal

as you have very many replicates, you can try the following: Take one of your treatment groups of 16 samples, divide it up at random into two sub-groups of eight samples each. Then, use each of the tools you are considering to perform an eight-versus-eight comparison. As the samples are all replicates, the between group variance should be equal to the within-group variance, the tool should notice that, and hence report only very few results. If cuffdiff again reports a large number, you shouldn't believe it. (We had a similar discussion earlier here on SeqAnswers, and cufdiff failed this test back then; but that was an older version.)

For edgeR, it is crucial that you use 'tagwise dispersion' to make full use of your many replicates and use a rather low 'prior.n' value (see rule-of-thumb in the vignette). Did you do that? As author of DESeq, I'd like to suggest, of course, that you try our tool as well, but actually, I guess it will give quite similar results as edgeR.

You haven't told us much about ypur experimental design, but if each sample is a different individual from an outbred population, and the difference between your group is not something very drastic (e.g., if you compare non-related humans, who differ by some not that dramatic trait), it is entirely possible that 16 samples give you way to little power to detect more than the few genes you found.

[11xinqi]

Hi
Im working with RNA-seq data from solexa.I have 32 lanes each belonging to one sample and would like to test 16 vs 16.I have a problem when I run cuffdiff on the data I will get large number of DEGs like 3000 or 4000.I have got 72 DEGs out of this data with edgeR.
I have changed many option but still I have many genes.
Could you help me to get reasonable number of DEGs?
Thank you

Hi

Did you find out why you get this result? BC I am in the same situation. I have two groups, one is control, the other is treatment with a specific gene knockdown. each group has 3 replicates. I got 3000 genes DEGs from cuffdiff. But only around 70 genes from edgeR. A weird thing is that a lot of DEGs from cuffdiff only expressed in one replicate of one group. For example, gene A is only be expressed in one replicate of control group, but not in other two replicates and treatment group!

[11xinqi]

Hi Jamal

as you have very many replicates, you can try the following: Take one of your treatment groups of 16 samples, divide it up at random into two sub-groups of eight samples each. Then, use each of the tools you are considering to perform an eight-versus-eight comparison. As the samples are all replicates, the between group variance should be equal to the within-group variance, the tool should notice that, and hence report only very few results. If cuffdiff again reports a large number, you shouldn't believe it. (We had a similar discussion earlier here on SeqAnswers, and cufdiff failed this test back then; but that was an older version.)

For edgeR, it is crucial that you use 'tagwise dispersion' to make full use of your many replicates and use a rather low 'prior.n' value (see rule-of-thumb in the vignette). Did you do that? As author of DESeq, I'd like to suggest, of course, that you try our tool as well, but actually, I guess it will give quite similar results as edgeR.

You haven't told us much about ypur experimental design, but if each sample is a different individual from an outbred population, and the difference between your group is not something very drastic (e.g., if you compare non-related humans, who differ by some not that dramatic trait), it is entirely possible that 16 samples give you way to little power to detect more than the few genes you found.

Hi, Simon

I have a question. Would you please help me? I have two groups, one is control, the other is treatment with a specific gene knockdown. each group has 3 replicates. I got 3000 genes DEGs from cuffdiff. But only around 70 genes from edgeR. A weird thing is that a lot of DEGs from cuffdiff only expressed in one replicate of one group. For example, gene A is only be expressed in one replicate of control group, but not in other two replicates and treatment group! Do you know why cuffdiff take these genes only expressed in one replicate as DEGs? And why I get so different result from cuffdiff and edgeR?

Both groups are cow embryo. There are 3 different embryos in control group. The three treatment group contain 3 embryos with one gene knockdown. Do you think why I get just few DEGs is because each sample is a different individual from an outbred population, and the difference between your group is not something very drastic?

Joseph