previous episode

Introduction to gene expression microarray analysis in R and Bioconductor

next episode

Microarray Data processing with RMA

Overview

Teaching: 20 min
Exercises: 20 min
Learning Objectives

  • Understand and explain Background correct, normalise and summarisation steps for microarray data

The processes of RMA

The RMA algorithm performs steps of data processing discussed in class.

  1. Background correction.
  2. Normalisation.
  3. Summarisation (calculating feature-level data)

These steps are performed in order, and yield a processed data set size that is considerably smallar than the data set prior to processing. While the starting object is an AffyBatch object, the result is an ExpressionSet. `

Plotting the signal densities from raw CEL files

The hist() functions from the oligo package show the signal densities of data from each sample in overlaid line plots.

plot of chunk affy-densities

As you can see, each sample does look like it comes from a similar kind of distribution for all six CEL files, but the exact distributions clearly differ. The prcessing steps of RMA address these issues to provide estimates for each feaature (transcript).

The need for background correction

Microarray data contains ubiquitous background noise, so an important first task is removal of background noise. Different algorithms deal with background differently. Most Affymetrix expression microarrays are designed with probe pairs, where one probe is a perfect match (PM) to the target and one is a single base mismatch (MM) to the target. Methods like Affymetrix’s own MAS5 algorithm use the difference between PM and MM hybridisation to addreess background noise. The RMA algorithm ignores the MM probes, and so must address background correction directly as a property of hybridisation.

The RMA model is to model the perfect match (PM) signal as a sum of the real signal for probe \(j\), from probeset \(k\), on array \(i\).

\[\text{PM}_{ijk} = \text{bg}_{ijk} + s_{ijk}\]

The key part of the model is to assume that the true signal is drawn from an exponential distribution

\[s_{ijk} \sim \text{Exp}(\lambda_{ijk})\]

while the background component (both optical noise and non-specific hybridisation) is drawn from a normal distribution that depends only on the array:

\[\text{bg}_{ijk} \sim \mathcal{N}(\beta_i,\sigma^2_i)\]

The background removal step of RMA estimates and removes this second component.

The gcrma algorithm differs from rma in assuming the background also depends on the GC content of the probes.

Background correcting... OK

plot of chunk bgcorrection

The need for normalisation

Sample normalisation in RMA is performed via quantile normalisation of the probe level data, as discussed in class. This allows samples to be compared to each other assuming the data arise from thesame parent distribution.

Quantile normalisation puts the data on a common empirical distribution.

oligo_normalised <- normalize(raw_nobg,method='quantile',which='pm')

Normalizing... OK

hist(oligo_normalised,lwd=2,xlab='log intensity', which='pm',
     main="CEL file densities after quantile normalisation")

plot of chunk oligo_normalization

Summarisation: from probe-level to feature-level data.

oligo_summarised <- rma(oligo_normalised,background=FALSE,normalize=FALSE)

Calculating Expression

hist(oligo_summarised)

plot of chunk oligo_summarised

The summarisation step reduces the size of the data for each sample to the number of measured transcripts (or genes, or exons, depending on the array). These are referred to as “features”, or “probesets”, and the resulting ExpressionSet is “feature-level” data. The summarisation step in rma is performed for each probe-set over all background-corrected, quantile-normalised samples using the median polish algorithm.

Performing RMA in R

Performing RMA in R is relatively straightforward. The function, rma() from the affy package takes in an AffyBatch object. The function basicRMA() from the oligo package takes in an ExpressionFeatureSet object. Hence, the following one line of code performs RMA normalisation, and returns an ExpressionSet object containing the background corrected, normalised, and summarised expression data.

gse33146_eset <- rma(gse33146_celdata)

Background correcting
Normalizing
Calculating Expression

gse66417_eset <- rma(gse66417_celdata)

Background correcting
Normalizing
Calculating Expression

Try it

If you get help on rma() using either help(rma) or ?rma, you can see how to run the same process without background correction or normalisation. Try it!

Manipulating an ExpressionSet object

Earlier, we have alluded to the data contained in an ExpressionSet object. We can access each of these pieces of data using the following functions:

Data Type of information Function
Annotation Chip information annotation()
PhenoData Phenotype data pData()
Expression Normalised gene expression exprs()
Experiment data Experimental information experimentData()
Feature data Probeset data featureData()

The phenotype data for each of the samples are retained after rma, so you should be able to see that information

You can also try, running RMA without background correction, or without quantile normalisation, and plot the densities.

Key Points

  • RMA, the most widely used processing algorithm for Affymetrix data, is implemented in R using the rma() function in the oligo or affy packages, depending on how the data was imported.

  • The steps of background correction, quantile normalisation, and summarisation are performed in order to obtain feature-level data

previous episode

next episode