too many Differential expressed genes using cuffdiff

[papori]

Hi,
i know this will be sticky question.. but i didnt find the answer..

i am using CuffDiff for DE between 2 condition.
i have 3 replicates for each.
The size of the number of reads that i have for each sample is variable..
3M,10M,20M,0.7M and so on..

when i used CuffDiff as:
cufflinks-1.3.0.Linux_x86_64/cuffdiff -N -o cuffdiff/2v10 cuffmerge/al/merged.gtf Tophat/case1/Trinity/Sample1/accepted_hits.sam,Tophat/case1/Trinity/Sample4/accepted_hits.sam,Tophat/case1/Trinity/Sample7/accepted_hits.sam Tophat/case1/Trinity/Sample2/accepted_hits.sam,Tophat/case1/Trinity/Sample5/accepted_hits.sam,Tophat/case1/Trinity/Sample8/accepted_hits.sam

i am getting 8000 genes in the list of DE.
it is not an usable number..
How can i reduce it?(Multiple Test as Bonferrony? BH?)

Thanks,
Pap

[severin]

Biological question

What was your biological question that you were testing. Did you control as many variables as possible to limit the sources of differential gene expression?

How did you normalize your libraries? Did all your replicates have such large library variation? What is your cut off for considering a gene differentially expressed? How many false positives do you expect given a histogram plot of your pvalues?


Lots of reasons why you may have gotten a large number of differentially expressed genes. This has become a very hot topic on this forum as of late with very little discussion. I would like to see this discussed more.

[billstevens]

I'm a newbie, but one thing I'm pretty sure you should do is plot the two conditions against each other in a scatter plot, as shown in the Nature Protocol paper. Most genes should line up on the 1:1 line, if not, there's something systematically wrong.

Secondly, sdriscoll, put up an interesting post about how his runs with variable read lengths resulted in skewed data. Take a peak at this post:

http://seqanswers.com/forums/showthread.php?p=70633#post70633

Thirdly, in regards to normalization, I don't know what species you are using, but I think it sounds like you don't have too much sequencing depth. Using upper-quartile normalization is supposed to help with getting more reliable data on lower-count genes.

Severin, I really would like further discussion in general. So most people use the getSig function with a p value of 0.05 and the false positive has a default of 0.05 as well. How do you feel about these parameters?

[papori]

What was your biological question that you were testing. Did you control as many variables as possible to limit the sources of differential gene expression?

How did you normalize your libraries? Did all your replicates have such large library variation? What is your cut off for considering a gene differentially expressed? How many false positives do you expect given a histogram plot of your pvalues?

Lots of reasons why you may have gotten a large number of differentially expressed genes. This has become a very hot topic on this forum as of late with very little discussion. I would like to see this discussed more.

i didnt normalized my libraries.. isnt the FPKM is the normalized?
my cutoff for DE is pval=0.05
i have a lot of false positive...

i had low quantity of rna so we use amplification step by NUGEN kit.
i think that it made a lot of noise..

Do you have any recommendation for "post effect"??

[sdriscoll]

In my opinion you still can't go completely on the differential expression results. You should sort your list of significant genes by fold change and then compare your list of results to the coverage information. Try making bedGraph files for your alignments and viewing them in the UCSC genome browser. You'll be able to make some decisions about your results based on how the coverage looks for those genes.

Also, as billstevens noted, pair-wise scatter plots are important.

If you want a second opinion try the htseq-count / DESeq pipeline and see what the results are. I find the results of the DESeq pipeline to make more sense than cuffdiff's results.