When We Can Carry Out Microarray Meta-Analysis?

Question

[Cheng Zhongshan]

Hi, dear members, I want to know whether someone has microarray meta-analysis experience. In my research, as mentioned in my previous question "How many genes differentially expressed in microarray can be seen as normal?", I want to combine microarray from different platforms, one microarray experiement was set in the sexual development time course (sexual development induction 0h, 24h, 48h, 72h, 96h and 144h; 0h was set as a control) of Gibberella zeae, a plant pathogen; the other microarray experiment is about a transcription factor mutant, it means a transcription factor in the strain was deleted, and it didn't produce any sexual organ after induction, the control of which is the normal sexual development strain, and the mRNA of them were extracted after 72h sexual induction. I actually want to combine these two platform microarray datasets.

I know many people using meta-analysis in this kind of situation. Can I do meta-analysis like this:

First, I would select differetially expressed genes in the two microarray platform separately by comparsion the treatment with their control.

Second, I get the fold changs of each treatment compared with their control. In the case of time course, I would get them like this: fdc1=24h/0h,fdc2=48h/0h, and so on. In the other case of transcription factor mutant, because this mutant didn't product sexual organ, can I see it as a control like the 0h control in the first microarray platform, then the fdc would be equal to (the control that produce sexual organ/the mutant).

Third, I want to use SAS or other software to compare the difference between the fold chang (first only at 72h) between the two different microarray platform, is this meaningful?

If I can compare them like the above, I would compare with the other time course such as 24h,48h,96h,144h. In this way, I would get genes fold chang are differentially expressed.

I would also want to compare this aquired gene set with the genes selected in the first step, maybe I can use venn diagram to get the intersection of these different expression genes along with the different fold chang genes. Are this procedures correct? I hope any of you can give me some suggestions about meta-analysis? Thanks very much!

[Nathan Harmston]

do you really mean combining over different platforms (i.e the microarrays a different (different probes/numbers of probes) or analysing over different experiments? Are both of these datasets normalised using the same method?

[Cheng Zhongshan]

Actually, I have already map the locus no their probset, and normorlized all the microarray data separately, I want to know how to compare them using meta-analysis, is my idea feasible?

[Cheng Zhongshan]

Actually, I have already mapped the locus no to their probsets with software SAS and Perl, and normorlized all the microarray data separately with CLC genomic workbench, I want to know how to compare them using meta-analysis, is my idea feasible?

[Cheng Zhongshan]

Here in my code

"TF134_1_3DAK,TF134_2_3DAK,WT1_3DAK,WT2_3DAK,H0_1,H0_2,H0_3,H0_4,H0_5,H72_1,H72_2,H72_3",

TF134 represents transcription factor mutant, and WT means the mutants' control, which would produce sexaul organ compared to the mutants of transcription factor. H0 and H72 represents microarray from another platform, and H0 means the mRNA was extracted at 0 hour of sexual induction, H72h means mRNA was extracted at 72 hours induction, so at this time point, this strain would produce sexual organ. Actually, I want to make my transcription factor mutant as a the control of 0H in the other platform, and set my analysis as this: my mutant and my WT(produce sexual organ), the 0 hour treatment microarray and 72 hour treatment microarray from the other platform.

RSadvance is for identification of genes that consistently up- or down- regulated by treatment, such as one medicine, compared to control. If people expect genes with high rank in ascending order (treatment vs. control) across different studies to be consistently down-regulated genes with high rank in desending order (treatment vs. control) to be consistently up-regulated. In the manual of RankProd, the author said 'Since we would expect candidate genes having relative consistant high rank in all studies, we prefer to use RSadvance to perform a three-original analysis treating each study as one origin. Additionally, RSadvance can used to identify genes up-regulated under one condition and down-regulated under another condition.

So what I really what to know is which is biological informative for me to use RankProd to analyze my data? Is identify genes upregualated in my mutant and down-regulated in 72h treatment at the same time meaningful? or is identify genes with consistent down- or up- regulated in my mutant compared with its control and the 0h treamt compared with the 72h treatment? Is the two the same?

Answer 1

[User 59]

The first answer that springs to mind is 'no you cannot compare the fold changes between the two platforms'. Even if you attempt to correct for the batch differences between the two platforms, the fact they are normalised independently is going to cause many issues.

For simple meta-analysis I lean towards using the RankProd method :

http://bioinformatics.oxfordjournals.org/cgi/content/short/22/22/2825

Which is certainly capable of doing what you want, as well as being able to work well with both poorly replicated, and noisy datasets.

However - you're going to end up comparing gene lists with gene lists, potentially from different experiments and there are already a number of approaches to this (LOLA, L2L etc.). I have a (not-exhaustive) list of papers for meta-analysis here.

Answer 2

[Cheng Zhongshan]

Thanks. I know RankProd and tried it previously, but I am not sure about the results. For example, in my case, I want to know which genes are significant affacted by the deletion of the transcription factor, by meta-analysis my microarray data with another time course microarray datasets. Is it feasible like this:

setwd("C:/Public")
Yang <- read.delim("all_microarray_(14034).txt",header=TRUE)
attach(Yang)
a <- data.frame(TF134_1_3DAK,TF134_2_3DAK,WT1_3DAK,WT2_3DAK,H0_1,H0_2,H0_3,H0_4,H0_5,H72_1,H72_2,H72_3,row.names=locus_no)
a.gnames <- Yang$locus_no
library(RankProd)
******I am not sure here, can you give me some suggestions according to what I really want to get!!!***********************
a.cl <- rep(c(1,0,0,1),c(2,2,5,3))
a.origin <- rep(c(1,2),c(4,8))
Rsum.adv.out <- RSadvance(a, a.cl,a.origin,num.perm = 100, logged =T,na.rm=FALSE, plot=FALSE, rand = 123,gene.names=a.gnames)
topGene(Rsum.adv.out, cutoff = 0.05, method = "pfp", logged =T,logbase = 2, gene.names = a.gnames)
a.plot <- plotRP(Rsum.adv.out, cutoff = 0.05)

[User 59]

RankProd I generally use to combine similar studies for additional power, you seem to be more interested in differences between the deletion of the transcription factor, which looks more like you want to contrast the results of the experiment with your own. It's hard to diagnose your R however without knowing your experiment a little better. I think you will probably want to use RPadvance rather than RSadvance however, I'm not sure there is an RSadvance in RankProd.