Cuffdiff Results confidence help

[himanshu04]

Dear Users,
I have a 2vs2 comparison model for Mouse data. I have aligned them with tophat2 and I am trying to find differential expression via Cuffdiff. I have a question for interpreting the Cuffdiff result. There are hardly 10 genes which pass the significance (FDR) using Cuffdiff. The other genes although are interesting and have a good p.value but have a very bad FDR . Do you trust those genes?. Also, some of the genes are very lowly expressed say 0.004 in one condition and 0.04 in other and will have a great fold change and p.value. I am having a tough time to interpret the results as I am new to this field .
I will highly appreciate the expert opinions in this case.
Awaiting your replies.
Thanks,
Himanshu

[pengchy]

I have the same questions. Take several examples:

gene_id sample_1        sample_2        status  value_1 value_2 log2(fold_change)       test_stat       p_value q_value significant
Gene00040       C       4h      OK      0       35.4111 inf     nan     0.07065 0.998432        no
Gene00054       C       4h      OK      7.56727 0       -inf    nan     5e-05   0.0160077       yes
Gene00109       C       4h      OK      445.777 0.302099        -10.5271        -7.85392        0.2447  0.998432

These genes were analyzed in one experiment. It is supprised that the gene with expressioin "7.56727 0" was significant, while "445.777 0.302099" was not!

There were many many examples like these. So whether the q_value is reliable?

[mattia]

I suppose you have to take in account two things:
1 - Variance between replicates;
2 - Number of replicates.

Generally, a statistical test gives you a significant p-value if the number of replicate is "high" (???) and the variance is low.
Check the value of FPKM both in sample1 and sample2 in both conditions.

[pengchy]

May be the question lies here, my experiment doesn't have any replicates. So, for the experiment without replicate, is it more reasonable to use the algorithm based on counts, such as edgeR and DESeq?

[mattia]

May be the question lies here, my experiment doesn't have any replicates. So, for the experiment without replicate, is it more reasonable to use the algorithm based on counts, such as edgeR and DESeq?

I cannot answer about edgeR, 'cause I've never used it but I'm quite sure that Deseq is able to work also without replicates.

[pengchy]

It seems himanshu04 has replicates (2 vs 2) and got the same question. Currently, I should be caution to use cufflinks/cuffdiff's results, and still struggle from many comments in SeqAnswer to get a clear thinking.

[himanshu04]

Hi ,
I have two replicates. Also the FPKM's in both conditions is sometimes really low say 0.004 in one amd 0.02 in other and so fold change is high and P value is good and for some genes its 1000 and 600 but tha p value is not so high . My question is what do I trust and how to proceed further to target a few genes computationally before jumping to biological validation.
Thanks,
Himanshu