Dear List, I have microrarray data where condition is completely confounded with a time batch effect. When doing a PCA on the RMA normalized data, the first principal component separates clearly the two batches. What option do I have when I still want to compare different conditions across batches? As far as I understood I can't dissolve this batch effect neither with limma nor with e.g. ComBat. Shall I just go on with the limma analysis and keep in mind that some genes just might pop up due to batch effects? best, tefina

Hi Tefina, It sounds like you are in a very difficult situation (i.e. if the design is such that batch separates condition exactly). If the condition is nested within batch or vice versa you may be able to minimize the effect, but otherwise it may be a wash. If you can address an effect with limma, usually it will be possible to address it more effectively with ComBat, but if you cannot separate batch from condition then even something like SVA won't work. I don't suppose you have any replicates that were run within each batch? It's a long shot, but if you did, you could use those to estimate the batch effects. You might consider something like fRMA (instead of say RMA) to leverage other peoples' non-confounded designs. That's the best I can think of; hopefully one of the authors will happen upon your post and comment whether this makes sense given your situation. I would say something about experimental design but obviously you are well aware of its importance given your question. Hopefully you can use fRMA or tech reps to salvage it. Best of luck, --t On Thu, Mar 8, 2012 at 2:56 AM, tefina <tefina.paloma@gmail.com> wrote: > Dear List, > > I have microrarray data where condition is completely confounded with a > time > batch effect. When doing a PCA on the RMA normalized data, the first > principal > component separates clearly the two batches. > > What option do I have when I still want to compare different conditions > across > batches? > As far as I understood I can't dissolve this batch effect neither with > limma nor > with e.g. ComBat. > Shall I just go on with the limma analysis and keep in mind that some > genes just > might pop up due to batch effects? > > best, > tefina > > _______________________________________________ > Bioconductor mailing list > Bioconductor@r-project.org > https://stat.ethz.ch/mailman/listinfo/bioconductor > Search the archives: > http://news.gmane.org/gmane.science.biology.informatics.conductor > -- *A model is a lie that helps you see the truth.* * * Howard Skipper<http: cancerres.aacrjournals.org="" content="" 31="" 9="" 1173.full.pdf=""> [[alternative HTML version deleted]]

You might be able to salvage some of the biology, depending on the microarray that you used. If there are enough control probes on the array, you could use something like Surrogate Variable Analysis on the control probe values. However, do look into the types of control probes being used (they don't just have to be negative control probes). The idea would be that some portion of the differences across samples at these probes could capture some of the batch effect (given you wouldn't expect the experiment condition to be correlated with the control probe values). Then you could adjust for the surrogate variables generated by these control probes in your analysis of the regular probes. -Andrew [[alternative HTML version deleted]]

Dear Tefina, Since condition and time batch are completely confounded, strictly speaking there is no way do distinguish between the affect of the condition and the batch. So no software package can help. However, if you have negative controls, i.e. genes which should not be affected by condition (including manufacturer's control genes) you can try to use them to estimate (and at least partly remove) batch effect. Best regards, Moshe. > Dear List, > > I have microrarray data where condition is completely confounded with a > time > batch effect. When doing a PCA on the RMA normalized data, the first > principal > component separates clearly the two batches. > > What option do I have when I still want to compare different conditions > across > batches? > As far as I understood I can't dissolve this batch effect neither with > limma nor > with e.g. ComBat. > Shall I just go on with the limma analysis and keep in mind that some > genes just > might pop up due to batch effects? > > best, > tefina > > _______________________________________________ > Bioconductor mailing list > Bioconductor at r-project.org > https://stat.ethz.ch/mailman/listinfo/bioconductor > Search the archives: > http://news.gmane.org/gmane.science.biology.informatics.conductor > ______________________________________________________________________ The information in this email is confidential and intend...{{dropped:4}}

We clearly need more information about your experiment to help you with removing batch effects. Judging from your original email, it sounded like all of condition 1 was on batch 1 and all of condition 2 was run on batch 2 (ie this is perfect confounding). What exactly was run on Batch 1 (experiment 1) and what was run on Batch 2 (experiment 2)? In your approach outlined below, you might still miss a lot, or have a lot of false positives. But its hard to know without your experiment design. Also depending on your experiment, if you're working in a well understood system, like yeast or stem cells, biologists could probably give you a lists genes that should change with a given treatment. you could see the variation across batches at these genes to see if SVA or ComBat is removing real biology -Andrew Message: 14 Date: Mon, 12 Mar 2012 10:09:54 +0000 From: tefina <tefina.paloma@gmail.com> To: <bioconductor@stat.math.ethz.ch> Subject: Re: [BioC] batch effect confounded with condition Message-ID: <loom.20120312t104842-37@post.gmane.org> Content-Type: text/plain; charset="us-ascii" Thanks for all your answers. I think I will pursue the following strategy: I will analyse each experiment separately. Contrasts within each experiment should be fine. (As batch 1 concerns only experiment 1 and batch 2 concerns only experiment 2). Any comparisons between experiments ( = comparisons between batches) I will only do on the p value level. So I will only evaluate things like: do genes that show up in experiment 1 also show up in experiment 2? Doing this, I should be more or less on the safe side. What do you think? [[alternative HTML version deleted]]

Thanks for all your answers. I think I will pursue the following strategy: I will analyse each experiment separately. Contrasts within each experiment should be fine. (As batch 1 concerns only experiment 1 and batch 2 concerns only experiment 2). Any comparisons between experiments ( = comparisons between batches) I will only do on the p value level. So I will only evaluate things like: do genes that show up in experiment 1 also show up in experiment 2? Doing this, I should be more or less on the safe side. What do you think?

I have 2 conditions (lets call them conditon A and B) with 3 biolog. replicates each in batch 1. Additionally, I have 2 other conditions (conditions C and D) with 3 biolog. replicates in batch 2. Questions of interest are: - diff. expr. genes between A vs B, and C vs D - this can be done via LIMMA - common regulation of genes comparing AvsB and CvsD Obviously, the second question is the more difficult one. At the moment, I would just compare the lists of diff. expr. genes on the p-value level.