Illumina's Gene Expression Microarray Data Analysis using R

1. Experimental design and data

   • Platform: Illumina BeadChips

   • Design: patients, groups (markers), and chips
   • Files (txt files)
     • raw data: each gene corresponds to one row.
     • sample names and array barcodes
     • annotation file

2. Data preprocessing using lumi package

   1. Data input: using function lumiR or lumiR.batch
   2. Preprocessing
      • using encapsulating function lumiExpresso
      • Functions lumiB (background correction), lumiT (variance stabilizing transform), lumiN (normalization) and lumiQ (quality control), designed for preprocessing and quality control
   3. Filtering
      • remove the undetectable (unexpressed) genes based on detection pvalue threshold given by
        1. quantile of all p-values, e.g., 50% quantile if the half of total probes are not detectable
        2. false positive rate, e.g., threshold = 0.10 (p-values follow an uniform distribution under null hypothesis)
      • remove technical replicates and/or irrelevant patients
   4. Visualizing
      • using function plot, including density, boxplot, MAplot, pair, and sampleRelation. See the details using help("plot-methods").
      • boxplot and density plot of both raw and normalized intensities on log2 scale
   5. Clustering
      • Using function plotSampleRelation: estimate the sample relations based on selected probes (based on large coefficient of variance (mean/standard variance)). Two methods can be used: MDS (Multi-Dimensional Scaling) or hierarchical clustering methods. Example: plot(lumi.data.object, what='sampleRelation', cv.Th = 0.10)
      • Detect the outlier: The current outlier detection is based on the distance from the sample to the center (average of all samples after removing 10 percent samples farthest away from the center).

        Example: temp <- detectOutlier(lumi.data.object, ifPlot=TRUE); any(temp) #if FALSE, there does not exist an outlier.

      • Using function hclust (cluster samples using Euclidean distance)

        Exampe: X <- exprs(lumi.data.object); temp <- hclust(dist(t(X)), method="average"); plot(temp)

      • Using principal component analysis (PCA)

        Example: X <- exprs(lumi.data.object); temp <- prcomp(t(X), scale=TRUE); groupColors <- palette(rainbow(length(levels(group))))

        1. Clusters using two components: plot(temp$x[, 1:2], col=groupColors[group], pch=19, main="PCA"); legend("topright", levels(group), col=groupColors, pch=19)
        2. Clusters using three components: scatterplot3d(temp$x[, 1:3], color=groupColors[group], pch=19, main="PCA"); legend("topleft", levels(group)), col=groupColors, pch=19)

3. Statistical analysis of differential expressions using limma package

   1. Model design matrix generated using function model.matrix
      • define three factor variables: patient, marker (or group), and chip

      • unpaired design: design <- model.matrix(~ 0 + marker + chip)

      • paired design: the patient or sample effects may be different when measured twice or more.

        • design <- model.matrix(~ 0 + marker + chip + patient)

   2. Fitting linear models

      • fit <- lmFit(X, design)

      • X: a matrix of gene expressions, each row consists of expressions of one gene.
      • For gene i, fitting a linear model: x_i= design * b_i + e_i
   3. Fitting contrasts (e.g., 3 contrasts)

      • contrasts <- c("marker3-marker1", "marker3-marker2", "marker2-marker1")

      • contrast.matrix <- makeContrasts(contrasts = contrasts, levels=design)

      • fit1 <- contrasts.fit(fit, contrast.matrix)

   4. Empirical Bayes

      • fit2 <- eBayes(fit1)

   5. Generating a top table with adjusted p-values and combining with annotations of interest
      • topfit based on F-statistic

        • topfit <- topTable(fit2, number=nrow(X), adjust="BH")

      • topfit based on t-statistic for each contrast (e.g., contrast k)

        • topfit <- topTable(fit2, number=nrow(X), adjust="BH", coef=k)

      • combining with annotations and mean expressions
        • cbind(annotations, mean.expressions, topfit)

References

Du (2008) lumi: a pipeline for processing Illumina microarray, Bioinformatics.
Du et al. (2014) Using lumi, a package processing Illumina microarray
Du et al. (2014) Evaluation of VST algorithm in lumi package
Lin at al. (2008) Model-based variance-stabilizing transformation for Illumina microarray data, Nucleic Acids Res.
Smyth et al. (2014) limma: linear models for microarray data user’s guide
Smyth (2004) Linear models and empirical bayes methods for assessing differential expression in microarray experiments, Stat Appl Genet Mol Biol.