Hi,

We recently received a dataset of 48 rat samples on 6 Agilent G3 GEx Rat V2 chips. I have no experience with Agilent arrays so I’m really struggling with this one.

A couple of questions I was hoping someone could help me with:

1) This dataset contains samples from 3 different tissues that were randomized across all chips. I was wondering whether it is ok to read in all the data in at once and then normalize or whether it would be better to normalize per tissue? If that is even ok to do so..

2) I had some issues reading in the datasets that were provided to me so I used the command below to read in the files but I’m not sure if this is ok:

targets <- readTargets("targets.txt", row.names="FileName")
x <- read.maimages(targets$FileName, source="agilent", columns=list(R="F635 Median", Rb="B635 Median"), annotation=c("Row","Column","ID","Name"))

3) The chip contains ERCC spike in probes but I’m not sure how to handle these in the analysis. I tried to analyze the data using commands below

bc<-backgroundCorrect.matrix(x, method="normexp")
E <- normalizeBetweenArrays(bc, method="quantile")
ct <- factor(targets$Type)
design <- model.matrix(~0+ct)
colnames(design) <- levels(ct)
fit <- lmFit (E,design)
contrasts <- makeContrasts (treatment-control, levels=design)
contrasts.fit <- eBayes(contrasts.fit(fit, contrasts))
summary(decideTests(contrasts.fit, method="global"))
a<-topTable(contrasts.fit, coef=1)
write.table(a, file="treatment-control.txt", sep="\t")

But I guess that is not the correct way to do so because the spike ins turned up being differentially expressed.. Can anyone help me on this? Thanks so much..

Best, Annelies

It looks like these Agilent chips are green only, in which case you can just do

x <- read.maimages(targets$FileName, source="agilent", green.only = TRUE)

The remainder is in general what you would do if all you have are treatment and control samples. But you say that there are three tissues! In which case you wouldn't want to simply compare treatment and controls because that assumes no differences due to tissue (in which case why would you do separate tissues?).

Anyway, the easiest thing to do is to generate a composite factor that includes both the tissue and type, doing something like

targets$combo <- paste(targets$Type, targets$Tissue, sep = "_")
## here I assume the tissue is in a Tissue column
design <- model.matrix(~0+combo, targets)
colnames(design) <- gsub("combo", "", colnames(design))
contrast <- makeContrasts(tissue1_treatment - tissue1_control, tissue2_treatment - tissue2_control,
                          tissue3_treatment - tissue3_control, levels = design)
## and go on from there, testing each of the three coefficients

As for the spike-ins, I have yet to find a real reason to do anything with them. This is mainly due to the fact that (so far as I know) the spike-ins are added using some really small volume like 1 µl, which in the world of molecular biology is something that people think is actually a thing. But it's not! People who work in fields where you actually have to accurately aliquot small volumes like that will tell you that it's nearly impossible to do accurately, particularly with an air-aspirating pipettor, and especially if there are proteins or surfactants in the solution that will break the water tension. You can have more than 1 µl of fluid just clinging to the exterior of a pipettor tip. Even doing 50 µl is hard. You can get a five decimal scale and try to pipet 50 µl of RO water (the easiest thing to do) repeatedly, and you will likely be surprised at how hard it is to reproducibly get 50 ng of water from a Rainin or Eppendorf. Now maybe if you used a glass syringe...

Long story short, I would argue that you can safely ignore the ERCC spike-ins.