Using Bioconductor To Analyse Beadarray Data

Software Requirements

• R, get from [CRAN]
• Bioconductor, get from [Bioconductor]
• Bioconductor packages. Install as needed:
  • beadarray
  • limma

source("http://www.bioconductor.org/biocLite.R")
biocLite("PACKAGE")

Loading Data

• At a minimum you need the Probe Profile data (normally a txt file).
• For all R procedures first change directory to your working directory then next create a new script, and save all executed lines in that script file.
• Load the beadarray library, indictate dataFile (required), sampleSheet (normally a xls or csv file) and control set (Control Probe, normally a txt file)

data = "FinalReport_SampleProbe.txt"
controls = "ControlProbe.txt"
samplesheet = "Proj_54_12Aug09_WGGEX_SS_name.csv"
BSData = readBeadSummaryData(dataFile = data, qcFile= controls, sampleSheet=samplesheet)

• You may need to alter either the ProbeID or ControlID to fit the illuminaprobe column from the sampleprobe or controlprobe datasets.
• This fits the data into the BSData dataframe. Phenotype data can be accessed by pData(BSData) and expression data can be accessed by exprs(BSData).

Data Normalisation

• Microarray data is typically quantile normalised and log2 transformed:

BSData.quantile = normaliseIllumina(BSData, method="quantile", transform="log2")

• To examine the effects of normalisation on the dataset use boxplots:

boxplot(as.data.frame(log2(exprs(BSData))),las=2,outline=FALSE, ylab="Intensity (Log2 Scale)")
boxplot(as.data.frame(exprs(BSData.quantile)),las=2,outline=FALSE, ylab="Intensity (Log2 Scale)")

• Save these boxplots as postscript files.

Clustering Analysis

• This analysis will generate a euclidean distance matrix then a cluster analysis of that matrix and will show the distribution between replicates. Ideally similar treatments will cluster together.

d = dist(t(exprs(BSData.quantile)))
plot(hclust(d)

Differential Expression Analysis

• Normalised data can be analysed using the limma package for statistical differences
• First define groups for each treatment. If a samplesheet was provided correctly and had this information:

samples = pData(BSData)$Sample_Group

• Otherwise define these groups manually in the order that they were entered (check by looking at pData(BSData)

samples = c("Control", "Control", "Treatment1", "Treatment1, "Treatment2"...)

• Next the groups are used to set up a statistical design:

library(limma)
samples = as.factor(samples)
design = model.matrix(~0 + samples)
colnames(design) = levels(samples)
fit = lmFit(exprs(BSData.quantile), design)

• Now set up contrast matrices to define how you want the data analyses. For example you may want to compare some treatments to a control, as well as between some treaments. See the limma user guide for more information about specific analyses. When defining the contrast matrix use the sample group names as defined above.

cont.matrix = makeContrasts(Treatment1vsControl = Treatment1 - Control, Treatment2vsControl = Treatment2 - Control, Treatment1vsTreatment2 = Treatment1 - Treatment2, levels = design)
fit.cont = contrasts.fit(fit, cont.matrix)
ebFit = eBayes(fit.cont)

Generating a Venn Diagram for Differential Expression

• First define a cutoff criteria for inclusion.
  • One option is to use the decideTests function:

results = decideTests(ebFit, method="global")

• The relevant options are for method and adjust.method
  • method
    • default is "global", which allows for p-value comparasons
  • adjust.method, this defines the false-discovery rate adjustment:
    • default is "BH" for Benjami and Hochberg
    • other options are "none", "fdr" (same as BH), "holm" and "BY"