I'd say, knowing only (mean) fold-changes and P values is not enough. You will need to know the sample sizes as well.

I totally agree with Jochen, but you can be satisfied by something more 'coarse grain' like calculating the Kendall Tau among the ranks of the entity (better fold change than P-values) of the changes induced by the drug (or the condition you want to test) in different experiments. You can weight the correlation entity by weigthing for the number of patients in the study (sample size) this will give usa sort of 'consensus' between studies. If you want to eliminate 'noise' of the not varying genes you can choose the rank in a way that no modified genes have the same rank.

Having reached such a geenral estimation of consensus (in some way analogous to a meta-analysis) you could look for the genes significantly modified in all the studies (if any) to select a strong invariant core of genes surely affected by the condition.

Alessandro, wouldn't it be possible to take the "most coarse grain" and simply look at the proportion of fold-changes going into one direction? The binomial test could then be used to test the null that this proportion is 0.5 (if the gene is not regulated, one would expect half of the fold-changes > 1, the other half < 1). The metaanalysis would not help to predict a reasonable effect size but at least give an indication if the gene might be regulated at all (and in which direction; up or down).

I agree with you, the point is that in the case of microarray you are not only interested in 'how many genes' are affected but even 'which geens are affected' this is why I sugegsted rank correlations and chi-square metrics on the superposition of affected genes between experiments...

Yes, ok. However, there was nothing said about "microarrays" in the question above. That's why I particularily thought about just one (or a few) genes being of interest anyway. There is a lot of literature containing information about some gene's regulation (for northern blots, qPCR, reporter assays...). But when there is whole-transcriptome data available, your idea using the correlation structure is good. I would guess that in publications showing results from microarrays the sample size is typically reported as well. Also, the raw data would be often availabe in data bases, so then there shouldn't be a restiction to work on fold-changes and p-values alone.

Gosh you are right !!!!! Great Jochen, Liyuan never mentioned microarrays in the question, this is very intriguing, I'm becoming so used to associate 'microarray' to the word 'fold-change' that I give it for granted.....mmmmh this is dangerous , Thank You Jochen !

You are wellcome :)

BTW，i know sample size, i can't get raw data.

Jochen Wilhelm · 113.55 · 272.25 · Justus-Liebig-Universität Gießen

I'd say, knowing only (mean) fold-changes and P values is not enough. You will need to know the sample sizes as well.

====

Thank you, i know sample size, i mean i can't get raw data.

Thank you both! I would like to use microarry data absolutely. But for many cases, i can only get fold change value and P value, i don't know if it is possible to do a meta-analysis

So if you do use microarray data, some correlational metric would bring you further, as Alessandro suggested. If you look for a particular gene and you know the sample size, then you have all you need. From the (log) fold-change, the p-value and the sample size you can derive the t-value, so you have the standard error, and since you know the sample size, you also have the variance.

Thank you Jochen! I'm a beginner on bioinformatics, your(and Alessandro's) anwsers are so valuable to me.

That's what RG is for, isn't it?

That's what RG is for definitively ! I think this kind of questions, dear Liyuan are the most suited for the potentialities of RG, while too specific questions of the kind '..which is the concentration of the molecule X is better to use for preparing the mixture J' or' which sofware is better for..' while potentially interesting, in my opinion could be solved by a look at the web or asking a colleague...

You are so nice! I have learned meta-analysis for just a very short time, Majority of cases that I can find form web focus on genetic data, such as result of GWAS, for expression data, most of pubulications utilized raw data of microarray, such as http://www.ncbi.nlm.nih.gov/pubmed/22513784 , so I confused and asked question. Do you know any publication that described meta-analysis with only P value and sampe size? If so, plz kindly introduce to me! Furthermore, I think I should pay attention to basic statistics knowledge, so that i can do judgement based on the core statistics questions.

BTW, this is my first question on RG, ^_^, celebrating!

This might be a starting point:

http://bioinformatics.oxfordjournals.org/content/25/20/2692.long

Thank you so much!