[super0925]

You don't do a DE analysis without replicates. The best you can do is look at the ranked fold-changes (have a look at the GFold package, which tries to do this in a somewhat more useful way). Personally, I wouldn't waste my time without good reason.

I have tried Gfold on my dataset . It seems works well. However, my samples are strand. I use strand (Tophat setting) and unstrand (the defaulf of Tophat) to map my samples and get the sam/bam file separately. However, the Top 100 genes ranking by GFold is almost same between strand and unstrand.Is it normal?

[dpryan]

No, my concern is whether the genes from " res$pvalue[keep]" correspond to those from "assay(rld)" or not, since in one case you're subsetting an object and the other not. Without knowing what you did in the previous steps, it's impossible to say of the indices in "select" are appropriate or not. Presumably you did things such that this makes sense, but this is unclear.

[nbahlis]

thank you Ryan, do you mean I need to transform my data with rld before I select a subset?

[dpryan]

The general concept of that seems reasonable. My only concern would be if you're first subsetting "res$pvalue" and if you haven't already done this with "rld" (or the object you used to make it" then the "select" indices may not be comparable. Otherwise, that seems reasonable.

[nbahlis]

DEseq and Heatmap

Can someone please advise what is the best script to create a heatmap with only differentially expressed genes? I've tried the script below and I am not sure this is correct way to do it

> select <- order(p.adjust( res$pvalue[keep], method="BH" ) < .1 )
> heatmap.2( assay(rld) [ select, ], scale="row", cexRow=0.5, cexCol=0.75,
+ trace="none", dendrogram="column",
+ col = rev(heat.colors(25)))

thank you

[dpryan]

If you have replicates then there's no point in using GFold, so just exclude it. Aside from that, the best way to go forward is to make a relatively small list of candidates that aren't shared between the different packages and then see if they validate using other means (qPCR or whatever you prefer). Then you have an idea which package is better modeling your particular dataset. Try to pick candidates with relatively similar characteristics (i.e., fold-change and adjusted p-value) just to make it a more apples-to-apples comparison.

BTW, usually the most significant genes overlap pretty heavily among the various tools. It's normally at the margins of significance where you see differences (unsurprisingly).

[super0925]

It doesn't matter which one you pick, they'll all be only vaguely meaningful at best...that's the problem with not having replicates.

Sorry， I didn't say it clearly, what I mean is the real dataset with replicate. Of course the ranking of TOP1000 DE genes are different from cuffdiff/Deseq/edgeR/Gfold/limma...... So I said do I need to do 3-5 simultaneously and then compare the overlap? The method with the biggest overlap is the best one? Cheers.

[dpryan]

It doesn't matter which one you pick, they'll all be only vaguely meaningful at best...that's the problem with not having replicates.

[super0925]

You don't do a DE analysis without replicates. The best you can do is look at the ranked fold-changes (have a look at the GFold package, which tries to do this in a somewhat more useful way). Personally, I wouldn't waste my time without good reason.

I found that the ranking of DE genes list (e.g. Top 1000 up/down regulated genes) are different in DESeq,edgeR,Cuffdiff,baySeq,...... (Of course they are different), now including this Gfold method, which is not reply on P-value but ''generalized fold change''.
How do I final make the decision by ranking Significant genes? Each methods has their advantage and claimed their most reliable.

I remember last time I told you I am thinking to calculate the overlap over Top 1000 genes list among the 3-5 methods. And then decide to use the method with biggest overlap. Am I right? Or is this measurement too rude and naive ?Or any other methods from your suggestion?

[super0925]

I would recommend Gfold as well. Also I made Wrapper/GUI for it on iPlant Collaborative. I am going to test it out this or next week to see if it works properly on it.

Hi, dpryan and zapages, thank you both !

[Zapages]

You don't do a DE analysis without replicates. The best you can do is look at the ranked fold-changes (have a look at the GFold package, which tries to do this in a somewhat more useful way). Personally, I wouldn't waste my time without good reason.

I would recommend Gfold as well. Also I made Wrapper/GUI for it on iPlant Collaborative. I am going to test it out this or next week to see if it works properly on it.

[dpryan]

You don't do a DE analysis without replicates. The best you can do is look at the ranked fold-changes (have a look at the GFold package, which tries to do this in a somewhat more useful way). Personally, I wouldn't waste my time without good reason.

[super0925]

Have a look at my answer over on biostars to a similar question. That should tell you most of what you want to know (particularly given the included links and replies from others).

Hi dpryan
If I only have one sample in each condition, how do I do the DE analysis?
Cause I think I need at least 2 replicates in each condition and then doing the DE statistical by DESeq/edgeR/Cuffdiff. Am I right?

If I don't run Cufflinks but running Htseq-counts,
Could I use this kind of method to measure it?

gene1: (a1/s1)-(a2/s2)

a1 is the counts (i.e. the number of mapped reads) of this gene in sample1
s1 is the total mapped reads in sample 1
a2 is the counts (i.e. the number of mapped reads) of this gene in sample2
s2 is the total mapped reads in sample 2

Or if I use FPKM calculated by Cifflinks to measure it ? i.e. FPKM1-FPKM2?

[dpryan]

Have a look at my answer over on biostars to a similar question. That should tell you most of what you want to know (particularly given the included links and replies from others).

[super0925]

Don't let the perfect be the enemy of the "good-enough, let's finish analysing the data"

Have you used STAR before? Or is it better than Tophat? I read some blogs talking about the mapping rate of STAR is higher than Tophat. But I don't know is it (higher or not) is really key to downstream analysis.