While SAM can work with RNA-seq data, it is a non-parametric approach.  Other non-parametric tools include GFold, as well as a few other published ones that may not be available as R/BioConductor packages (I've seen at least one paper of a Bayesian non-parametric method that I do not recall mentioning it was available in BioConductor).

If you have 3 or more biological replicates, my preference would be for a parametric approach.  Many of the ones in R/BioConductor begin with raw count data - DESeq2, edgeR being perhaps the most popular two, but there are others as well.  You can also pick from any number of normalization approaches (RPKM - nor recommended as it has several well known flaws, or TMM, KDMM and others) and use standard ANOVA or other statistical tests.

At this point in the evolution of NGS differential expression analysis, there is no single clear best method or practice.  You will need to dig through some of the literature and decide for yourself how you wish to proceed.  Amadis already gave you one good comparative paper, and there are several others.  You can also look through the publications of the SEQC consortium at http://www.nature.com/nbt/collections/seqc/index.html.

Ultimately, you may find you really need to explore your data with several tools and algorithms to see how your data behaves with different normalizations and tests appropriate to those normalizations.  Beware of biases that can be introduced due to greatly unequal library sizes, or unequal  read counts passing QC and/or mapping cutoffs.

If i have understood correctly, you're looking for approaches to detect differentially expressed genes between two conditions. There are many published methods that do this type of analysis, and for some of them (such as DESeq) there's even no need to pre-process gene expression data.

I suggest to take a look at some publication reviewing those methods, like the one attached from Genome Biology. It is from 2013 but covers, in my opinion, many of the the most widely used methods for differential gene expression analysis.

http://genomebiology.com/2013/14/9/R95

While SAM can work with RNA-seq data, it is a non-parametric approach.  Other non-parametric tools include GFold, as well as a few other published ones that may not be available as R/BioConductor packages (I've seen at least one paper of a Bayesian non-parametric method that I do not recall mentioning it was available in BioConductor).

If you have 3 or more biological replicates, my preference would be for a parametric approach.  Many of the ones in R/BioConductor begin with raw count data - DESeq2, edgeR being perhaps the most popular two, but there are others as well.  You can also pick from any number of normalization approaches (RPKM - nor recommended as it has several well known flaws, or TMM, KDMM and others) and use standard ANOVA or other statistical tests.

At this point in the evolution of NGS differential expression analysis, there is no single clear best method or practice.  You will need to dig through some of the literature and decide for yourself how you wish to proceed.  Amadis already gave you one good comparative paper, and there are several others.  You can also look through the publications of the SEQC consortium at http://www.nature.com/nbt/collections/seqc/index.html.

Ultimately, you may find you really need to explore your data with several tools and algorithms to see how your data behaves with different normalizations and tests appropriate to those normalizations.  Beware of biases that can be introduced due to greatly unequal library sizes, or unequal  read counts passing QC and/or mapping cutoffs.

Ulli, io non sono esperto di questo, however there is a guy quite knowledgeable on this. Can contact him, he does RNaseq analyses every day.

[email protected]

I guess many scientists have problems with quantitative analysis of RNAseq data including myself. The simplest analysis may involve normalization throughthe total number of reads per contig and then if the distrubution seems normal, use statistical analysis in log2 mode similarly as in microarray analysis. This approach, however, needs various types of validations including qPCR and MALDI using e.g. peak area estimation. When R-square of the regression is reasonable (95). One can publish such analysis. This is what I have seen i some papers. In any case you need to have quite a few samples 5m and 5females etc at least. See what happens. Also depends how you map your genes (refmap), whether you have filtered out all errors, biases etc. 

Please keep in mind that with RNA-Seq, statistical test for differential expression should be performed at count level (NOT using measures that are derived after normalization by gene length such as FPKM/RPKM/TPM). Counts, of course, need to be normalized between libraries, and to our experience, simple median normalization works the best (we did a systematic study involving many factors -http://biorxiv.org/content/early/2014/10/17/010488 ). In the end of the day, you will be interested in a biological story, i.e. functional patterns of cellular response. Counterintuitive result of our above-mentioned study (that involved assessment of 6,912 combinations of technical and biological factors) was very low (close to zero, indeed) impact of low-level processing stage and lower than expected impact of the choice of DE testing method on the derived "functional signature of response". At the same time, choice of geneses enrichment method was critical. 

Agree with AM above. What I described works well with RNAseq when you sequence extracted transcripts of similar length, where you do not have to nebulize.

Some of the most frequently used approaches are generalized linear models (as implemented in edgeR and DESeq), with a bayesian approach (e..g, baySeq) being an alternative, as Michael Black and others noted above. I really like the edgeR software; both the vignette for edgeR and for DESeq are helpful and give some guidelines not only on how to use the software (i.e., commands) but also case studies that you can follow. As far as freely available command-line packages in R, these are great options, and relatively user-friendly. The GALAXY suite has an option for web-based analyses, along with several helpful tutorials (https://usegalaxy.org/). If you prefer a point-and-click type analysis, I think JMP Genomics may have implemented some options, but this is not free (and I'm less experienced with the package). Microarray analyses are similar but not ideal, because they are designed for continuous data (i.e., fluorescence intensity), whereas RNASeq is based on discrete counts.

Hi Ulrich, I have a couple of questions to better understand your question:

1) What type of files do you have and how did you obtain them (Do you have Fastq files, BAM and GTF files or Excel files with raw counts)?

2) Are you only interested in differential gene expression or are you also intereseted in differential transcripts expression?

Many great suggestions above. I could only add limma-voom method to the set of the linear model approaches: http://genomebiology.com/2014/15/2/R29

That is the method of choice for RNA-seq analysis in our Bioinformatics division - and that's the people that created edgeR and limma. So they should know, what they are doing, right? As far as I understand limma-voom controls the type I error rates slightly better than edgeR and DESeq.

But then again, if you are not familiar with R and Bioconductor (and unless you are keen to familiarise yourself) find a collaborator with an experience in RNAseq analysis. To be honest, you should find one before you even start an experiment ;)

Good luck!