Dear Yuan Xu

The previous answers are very good. I would like to explain about the microarray analysis in short to address your questions. The 'P' value is the statistical significance of the gene expression between the control and case groups. Usually we set a cut off of 0.05 for 'P' value. In case of FDR, it's better to set a cut off of 0.025. After cutting down based on 'P' value and fold change, you would get a short list of genes. Among these genes you could select the top 10 up regulated or top 10 down regulated genes for validation, the choice is based on your question of interest. R/Bioconductor helps a lot to address most of your questions.

best wishes.

Thank you very much for your helpful answers! I will look into all the tools mentioned and see how well they work with the data/aims of my project. Your answers have helped me build a clearer picture of the next steps of the experiment.

I like the NIA Array Analysis webtool. Here is the website.

http://lgsun.grc.nia.nih.gov/ANOVA/

Also, you seem woefully under-prepared in term of your statistics knowledge for this task. I would spend some time learning more statistics because you really need to understand what is a p-value and were to get one from.

Best of luck,

-Garrett

Indeed the p-values are important, however they usually mean that there is a significant expression of a particular gene, but not necessarily that the genes are over-expressed or under-expressed when comparing two conditions.

That is the meaning of the log2 fold-change; this value represents the logarithm base 2 of the ratio of the levels between the two conditions. Other parameter that is usually reported in Affymetrix’s arrays is the log-odd, which represents the logarithm of the probability that the gene is differentially expressed. When this value is zero, it means that there are 50% of chances that this gene is differentially expressed. There is a plot, called a “volcano plot”, where you can see the variation of the log-odds vs the log2 fold-change. This is an interesting plot, which has a “V” shape and the differentially expressed genes are located in upper-left corner (up-regulated) and the upper-right (down-regulated) corner on this graph.

There are some software packages (as Bioconductor) that can be used for plotting this kind of graph, but calculating the fold-changes and the log-odds from the Affymetrix data you may build this graph in software like Excel.

I do feel a bit underprepared for this project as I haven't had any teaching on bioinformatics prior to this project, although I have some statistics background (enough to understand what p-values are, at least!). I should have phrased my questions more exactly, sorry - in terms of the p-value, my understanding was that we wanted to show that, for genes that did have change in gene expression, there would be a *significant difference* in RNA quantity between treated and control groups. If the p-value was below the chosen critical value, then we would accept the gene, if not, then we would cut the gene from the list. I was more wondering about what kind of statistical test you would use in order to obtain the p-value (as I only know of the chi-squared, Students t test, ROC curves and Wilcoxon rank sum from previous projects, which don't seem to work in the context of this project). Thanks to the above contributors, I now have a list of statistical tools which are tailored for this purpose! I will also be seeking assistance from our department's statistician but that will have to wait until I obtain the data.

If it isn't too much to ask, could someone advise me on the usage of Bayes factors as an additional filter for final selection of genes of interest? From what I understand, the Bayes factor may be used as an alternative measure of "significance" or how strongly there is evidence for the null hypothesis. But I was recommended to use GATHER ( at http://gather.genome.duke.edu/) which appears to be a search engine of sorts which links to gene ontology terms and gives the Bayes factor for it. What exactly does this Bayes factor represent in the context of the gene ontology? I'm not sure if it's trying to say that "for gene ABCD123 which has this function X, at a Bayes factor of 20, this shows that there is good evidence that ABCD123 actually has the function X vs the null hypothesis that ABCD123 doesn't have the function X" or something else?

Also, whilst looking through http://lgsun.grc.nia.nih.gov/ANOVA/ I chanced upon GenMAPP http://www.genmapp.org/introduction.html which seems to be the kind of tool I was looking for - my supervisor has been telling me that Ingenuity IPA is a good tool to use (but we do not have access to the full version): how does GenMAPP compare to IPA in terms of functions? Also I heard that IPA has some function that can tell if there are contradictions in the downregulation/upregulation of genes from the data - can GenMAPP do this or will I have to manually check every gene for this?

There are a few steps involved. For Affymetrix expression arrays you will definitely want to do a number of quality control and a normalization step first and check the result of that normalization. We developed http://www.arrayanalysis.org for that which essentially uses a number of available statistical and mathematial approaches in Bioconductor/R developed by many groups in the community, streamlined and documented.

Next you will want to do statistics on a gene by gene level. The basic approaches do that already per array (there are multiple probes for a single probeset that detects a gene on an array). For simple designs www.arayanalysis.org can do that too. For more complex designs you will want to talk to a statistician (in fact you normally should do that before starting a study).

Supposing you have a list of genes and fold changes the hunt for meaning only starts. Pathway analysis was already mentioned and usually are very useful. We run WikiPathways which is a community resource to develop content and we will very soon update the Reactome content on that. There are several tools that use WikiPathways content (and sometimes other content too) to do pathway analysis. GenMAPP CS is a good one, and nice because of it's integration in Cytoscape. You might also want to look at our PathVisio (see http://www.pathvisio.org ) which gives you pathway statistics and visualization in core tool and which can be extended with many community developed plugins. Like GenMAPP it is free and open source. We actively collaborate with the GenMAPP team, and they are our main partner for the WIkiPathways project.

Next to pathway analysis you might also want to perform Gene Ontology analysis and possibly Disease Ontology analysis. There are many tools for Gene Ontology analysis and they all have their pro's and cons. They deal differently with leaves and branches in an ontology tree, which can yield you (very) different results. You might want to look at (again our, sorry) GO_elite which is on the GenMAPP site. http://www.genmapp.org/go_elite/ . It can run different algorithms and can work with other ontologies as just the Gene Ontology too.

Basically these pathway and ontology tools try to find the biology behind what you found more than some affected genes. Sometimes important processes contain a larger number of genes that are changed but not that much. Sometimes different individuals show different genes being affected but they are all involved in the same process (note that most tools do not (yet) deal with that possibility in an adequate manner.