Generally if you look at the summary(in edgeR) - It select up and down more than zero and less than zero. And personally I am also in favor of this. Because sometimes biologist are interested in normalized count of the DEG along with the log fold change. There's a very nice article about fold change and p-value(But this is for micro array]

http://www.biomedcentral.com/1471-2105/13/S2/S11

Thank you for all tour suggessions...

can you suggest a tool more sensitive than edgeR. Mine is paired data.

edgeR uses raw counts as input, I don't know why you point out Count Per Million (CPM), thats just the normalised output, making you able to compare the expression level on one gene at different times.

FPKM is much the same as CPM, only you also compensate for genes different length (total exon length), making you able to compare expression between genes, not only just the gene itself. This because longer genes will look higher expressed than lower one, if not compensatet for.

Limma and edgeR are developed by the same people (Gordon K Smyth), the Limma - package is for Micro Array and uses statistics for normal distributed continuous data. This does not hold true for RNAseq, because of the transcripts being discrete data. You should not use the Limma - package for RNAseq, and you don't have to, because edgeR uses Linear Mixed Models (limma) for RNAseq, so no problem.

DESeq is outdatet, if you want to, you could use DESeq2. The way to choose is maybe to take 10 genes that is only differently expressed by edgeR or DESeq2 , then analyse the genes with RT-qPCR and see what algorithm fits your data the most.

edgeR is the most sensitive tool, and you may use generalised linear models, paired data is handled with ease:

In your model.matrix, just make a column indicating the samples. Then incorporate samples as covariates in your design formula. R will then pick up that it is a paired design and will calculate the variance like a paired t-test.

When it comes to fold change cut-of, just divide into logFC > 0 and logFC < 0. Thats it, RNAseq has much lower noise than Micro Array, you don't have to be as stringent in your cut-of. Especially if you followed the filtering recommendations in the edgeR manual. But of course, the lower sample size the higher cut-off, and vice versa.

I would go for cutoffs of both 1.5 and 2 fold changes, and report them separately. 2 is more stringent. DESeq2 includes a different method of size factor normalization in comparison to DESeq 1, which might be useful in many cases.

The following nature protocol published last year includes both Edge R and DESeq (The low coverage filtering and many other elements of RNA-Seq analysis are included).

http://www.nature.com/nprot/journal/v8/n9/full/nprot.2013.099.html

You might want to use the latest version of DESeq.

There are various factors that affect the power of a statistical test, i.e., the probability of rejecting the null hypothesis of "no change in expression", when it is in fact true. Of course, we will want to have no error, i.e., a power of 1 or 100%; however, that is not generally possible. As we increase the sample size (i.e., the number of tags sampled) the power increases, thus a fold change of 1.5 which was not "significant" (at a given alpha) when sampling 1 million tags per library will possibly be highly significant (smaller alpha) if a sample of ten millions in employed for each library. It is indispensable to make a clear separation of what is "statistically significant" (at a given error alpha) and what is "biologically relevant". edgeR (or many other statistical packages employing distinct modeling methods and statistical tests) measure the mean difference between treatments comparing it with the "error" variation, measured by comparing the differences between replicates when the same treatment is employed. Even when your question is very important, it does NOT has an easy answer; i.e., there is not a "magic" number (1.5 or 2 in fold) which will make you to do the right decision about the biological relevance of a change of expression for all genes in all experiments. Statistics here give you only a number (P; alpha) which tell you "how likely is that the null hypothesis is true given the data". If this number is "very" small you can trust that the null hypothesis is false... Is better to take this into account that to blindly trust any specific number, in fold or frequency change.

To clarify on my previous post, I was suggesting 1.5 and 2 as arbitrary cutoffs, only for the changes which are statistically significant at the considered level of confidence. The p values should be adjusted on a considered false discovery rate.

I agree that there's no magic number that's why exploring multiple fold changes might help. A volcano plot might to get an idea of how the differential expression patterns are distributed among the p-values and the fold changes.

I would vote for DESeq as well.

You could always get clear pipeline example from their documents, either edgeR or DESeq, but please NOTE THE VERSION.

Fold change cut off itself is not a proper option here, because the variation pertained in the estimated fold change is expression level dependent, if no variation stabilization method was performed. Actually, the variation decreases as the increase of expression level. If you do not want to include low fold change genes into DE gene set, double cut off: FDR (or ajusted.p-value) cut off + fold change cut off should be used together.

I would vote for DESeq (or rather DESeq2) as well. We did a comprehensive evaluation of different tools for differential analysis of RNA-Seq data on gene level: http://genomebiology.com/2013/14/9/R95

The most common cutoff for miRNAs in the literature is 1.5 fold change. That'S because phisiologically even a minor change in these regulatory molecules can have an huge effect over a pathway.

I used both DESeq and edgeR. Although there are a few differences in the mathematical basis of both, they work with the same model, and in general the differences between them are small. I am still taking a look at DESeq2, and cannot tell much about it.

Both work with raw number of counts. This may seem a polemic issue, but for comparing gene expression between samples (as opposed to between genes) you don't really need any normalization other than the one provided by the packages (edgeR and DESeq). So forget about RPKM, FPKM, etc if this is the focus of your experiment.... Otherwise I would suggest using TPM as a method for normalization.

http://lynchlab.uchicago.edu/publications/Wagner,%20Kin,%20and%20Lynch%20%282012%29.pdf