Data normalisation for twocolour microarrays

1
Data normalisation for two-colour microarrays

• References for microarray analysis
• 1. Leung and Cavalieri. 2003. Fundamentals of
  cDNA microarray data analysis. Trends in
  Genetics 19 649-659
• 2. Hegde et al. (2000) A concise guide to cDNA
  microarray analysis. Biotechniques 29(3) 548-554
• Quackenbush (2002) Microarray data normalization
  and transformation. Nature genetics 32496-501
• Acknowledgements to John Quakenbush for slide
  material

2
Outline

• Why Normalise?
• Normalisation methods
• Available Normalisation packages
• DNMAD (GEPAS)
• MIDAS (TIGR)
• R

3
A two-colour normalisation
4
Why normalise?

• During probe preparations technical variations
  can be generated including
• Unequal amounts of cDNAs
• Differences in dye properties
• Differences in dye incorporation
• Differences in scanning
• Normalisation aims to correct for these variations

5
Assumptions

• Most global normalisation methods assume the two
  dyes are related by a constant factor
• RkG
• Most normalisation methods assume that the mean
  ratio is 1 or that the log of this ratio is 0 ie
  a gene does not change its expression under the
  condition being studied.

6
Remember ratio vs log ratio
Acy3, Bcy5
R
4
Gene1
3
2
1
Gene2
0
AB
Advantage of log transformation Treat
up-regulated and down-regulated genes
symmetrically
7
Log2(ratio) measures treat up- and
down-regulated genes equally
log2(1) 0 log2(2) 1 log2(1/2) -1
8
Normalisation methods

• Within array normalisation
• Small targeted arrays
• housekeeping genes
• internal spikes
• Larger randomly printed arrays
• total signal intensity
• linear regression
• non-linear regression and Lowess
• Between array normalisation

9
Normalisation to housekeeping genes

• Idea Some genes shouldnt be differentially
  expressed
• But what are these genes?
• Perhaps actin, ubiquitin, ribosomal RNAs etc
• Normalisation constants
• (Cy3)/(Cy5) for those genes or the median value
  of Cy3/Cy5 for several housekeeping genes

10
Normalisation to spikes or exogenous RNAs

• On the array place a number of sequences from a
  different organism - genes that have low homology
  to any gene in the organism of interest.
• Synthesize RNA for each of these genes by IVT
• Spike known quantities of these genes into known
  quantities of sample (NB- must start with same
  amt of sample RNA)
• Set the normalization constant to get the
  expected value of the ratios for the exogenous
  added genes

11
Normalisation to Total Signal

• Assume equal gene expression/signal in each
  sample
• Global/ non-selective microarray
• Normalisation constant
• S(Cy3)/S(Cy5) or
• S(Cy5)/S(Cy3) for the dye swap experiment
• For each gene, multiply the ratio by the
  normalization constant.

12
Linear regression of cy3 vs cy5

• Assume expression of majority of genes doesnt
  change between samples
• Scatter plot of raw data (green vs red)
• Median background subtraction from mean
  foreground? You can decide
• data is bunched up in the left hand corner
• solution log transformation

13
Linear regression of log(cy3) vs log(cy5)

• draw scatter plot of log(green) vs log(red)
• draw linear best fit line
• yaxb, where a0.878 b1.419 (red line)
• x normalisedaxb

14
Before after normalisation
Red- raw data, blue-normalised data
15
The RI or MA plot

• Checks if data exhibits an intensity-dependent
  structure
• Uncertainty in ratio measurements generally
  greater at lower intensities
• For RI Plot log2(R/G) vs. log2(RG)
• For MA Plot log2(R/G) vs (½ )log2(RG)
• Remember
• log (R/G) log(R) log(G)
• log (RG) log(R) log(G)

16
MA plots pre- and post normalisation
After normalisation, the fit of the data to the
horizontal line through 0 is much better
17
Example of good data RI plot
18
Example of bad data RI plot
Each print tip is coloured differently NB! The
data is curved so a straight line normalisation
is not a good idea!
19
Non-linear regression

• Global lowess (locally weighted scatterplot
  smoothing OR locally weighted linear regression)
• Print tip lowess
• 2D lowess (spatial bias)
• DNA spots on array must be arranged randomly

20
Applying a Global Lowess to a single sample

• Produce a MA plot (calculating the average log
  intensity and log ratio for each feature)
• Curve fits data far better than a straight line
  as the data is curved.
• Apply a lowess regression to the data
• Calculate the normalised log ratio for each
  feature according to the equation of the curve

21
How does Lowess work?

• Lowess performs a large number of local
  regressions in overlapping windows.
• Each regression is then then combined to form a
  smooth curve ? the Lowess data set.

22
Print-tip lowess

• Application of lowess to a sub-set of the
• data (ie print group or subgrid) (providing
  sufficient genes printed per tip)
• Advantage
• can correct for systematic spatial variation
• (inconsistencies between pins,
• variability in slide surface,
• slight local differences in hyb conditions)

Print tips normalised for mean around zero
Print-tip layout
23
Print-tip lowess
24
A potential problem Spatial effects

• In some experiments there is a spatial difference
  between the two channels, resulting in parts of
  the array being brighter in the cy3 and other
  parts in the cy5.
• A typical cause is slides with an uneven surface
  resulting in the different lasers being out of
  focus if they do not adjust as they scan.

25
Correcting for Spatial Effects

• 2d lowess regression - fits a 2 dimensional
  polynomial surface to the data
• Block-by-block lowess regression
• MA plots do not show gradients across slides,
  therefore a pseudo colour or false colour overlay
  is required for each slide

or ?
26
Normalisation methods

• Within array normalisation
• Small targeted arrays
• housekeeping genes
• internal spikes
• Larger random arrays
• total signal intensity
• linear regression
• non-linear regression and Lowess
• Between array normalisation

27
Between array normalisations

• Usually you will want to do more than one array,
  either as replicates or as additional samples.
• You will need to normalise between arrays if you
  are wanting to to compare the results.
• You can even normalise across different platforms
  (e.g. spotted and Affymetrix) and laboratories!

28
Main assumption

• The variation is due to experimental artefacts
  NOT biology!!
• If you expect large differences between samples
  you should not try and normalise!!
• Data needs to be centered ? means of log ratio0
• Data needs to be scaled ? s.d.1
• Data needs to be normally distributed ?
  distributions are identical

29
Scaling across slides

Equal spread of variation between slides
30
Normalization Approaches
The Solution(?)

• Can minimise normalisation
• by adjusting PMTs during scanning so that the
  mean/median of the ratios1
• The best technique is experiment dependent
• Check diagnostic plots carefully
• All analysis methods depend on good
• experimental design

31
Normalisation packages

• Genepix
• Excel
• DNMAD (GEPAS)
• MIDAS (TIGR)
• R

32
What you can do in Genepix

• Array quality control
• Scatter plots (cy3 vs cy5)
• MA plots
• Normalisation using housekeeping genes or
  internal spikes
• Global normalisation
• total intensity (mean/median of ratios1)
• linear regression (regression ratio1)

By default, flags are excluded from
normalisation
33
Excel

• What you can do in excel
• Scatter plots (cy3 vs cy5)
• MA plots
• Normalisation using housekeeping genes or
  internal spikes
• Normalisation using total intensity
• Linear regression normalisation

34
DNAMAD

• http//dnmad.bioinfo.cnio.es/
• Input GPR files
• Input array layout (eg Arabidopsis Arizona
  slides Main grid 12 rows X 4 columns, sub-grid
  26 rows X 25 columns)
• Normalisation options eg use flags, return flags
  as NA, use background subtraction, use global
  lowess
• Work through tutorial http//bioinfo.cnio.es/docu
  s/courses/dnmad/index.html

35
Output
Can down load normalised M values as txt file
(remember to multiple dye swaps by -1)
36
Box plot of M values by print-tip
37
MA plots pre- and post LOWESS
38
Slide1 diagnostic plots
39
Slide1 Image plots
Background and M plots also shown
40
Slide scale normalisation
Next send to Preprocessor Download normalised M
values (txt file)
41
Microarray Data Flow at TIGR
Image Analysis
.tiff Image File
Raw Gene Expression Data
Gene Annotation
Normalization / Filtering
Normalized Data with Gene Annotation
Expression Analysis
Data Entry / Management
Interpretation of Analysis Results
42
MIDAS is a Normalization and Filtering tool
for microarray data analysis!
Serves as a data pre-processor for clustering
analysis (MeV).
Requires Java Runtime environment and Windows XP
43
MIDAS data analysis methods

• 7 normalization/transformation methods

Total Intensity normalization
LOWESS (Locfit) normalization
Iterative linear regression normalization
Iterative log mean centering normalization
Ratio Statistics normalization
Standard deviation regularization
In-slide replicates analysis

• 8 filtering (quality control) methods

Low intensity filter
Slice analysis
Flip-dye consistency checking
Ratio Statistics Confidence Interval checking
Invalid-intensity checking
Signal/Noise checking
Spot QC flag checking
Cross-file-trim
44
Graphical scripting language
45
Graphical scripting language

• Read input files
• Define analysis
• pipeline and set
• parameters for
• each analysis module
• Write output files

• NB
• Input MEV files only (convert GPR files using
  Expression Converter)
• Click create PDF report

46
Using R
statistical microarray analysis (sma) module

• sma will normalise, compare slides, and do
  statistical tests on data
• Allows simultaneous multiple slide analysis
• To process the data
• load experiments into R
• describe slide printing configuration
• load experiments into a working data set
• Analyse data

47
Normalisation exercises

• Two exercises using DNMAD
• http//bioinfo.cnio.es/docus/courses/dnmad/index.h
  tml
• Using your own data
• Try linear regression normalisation (in excel) vs
  print-tip LOWESS (DNMAD and/or MIDAS)