1. Introduction to microarray
Bin Yao

2. Types of Microarray Affymetrix GeneChip (Oligo)
Spotted array (cDNA /Oligo)

3. Affymetrix GeneChip
in-situ Synthesis: photolithography and combinatorial chemistry.
Each probe set contain13-21 pairs of 25- mer oligo probes.
PM and MM

5. Spotted array
cDNA or Oligo are printed on glass slides
using arrayer

6. Procedures
Sample1 mRNA
Cy3
Cy5
Array
Sample2 mRNA

7. ADC
Image
PMT
Array
Laser
Data

8. Image quantification Pixel value
Image: 16 bits gray scale image. Range of value values. Signal>65535 is saturated.

9. Image segmentation: separate signal, background and contamination
Output data files: Spotted array
Signal Mean
Background Mean
Signal Median
Background Median
Signal Stdev
Background Stdev

10. Output data files: Affymetrix
.DAT: Pixel data
.CEL: Intensity information for a given probe on an array
.EXP: Experiment information
.CHP: Analysis result from a Microarray Suite analysis

11. Get gene expression value from probe level data
Consolidate 26 (13 PM data and 13 MM data) data into one gene expression value
MAS (4&5): Affymetrix algorithm
Gene expression=weighted average (PM-MM)
Dchip: model based expression index
PMij – MM ij = i j + εij
with invariant Set Normalization
RMA: robust multi-array average
Normalized log (PMij -BKG)=i+ j + εij
With quantile normalization

12. Data analysis What are problems for microarray data analysis?
Different sources of variance
Large number of genes (high false positives)
Small number of replicates (low sensitivity)

13. Data pre-processing
Background correction: Signal of a spot contains specific binding signal, non-specific binding signal and background signal.
Background estimation: local background, global background and negative control spots.
Data filtering: Low signal spots and contaminated spots.
Data transformation
Ratio is not symmetric.
0.5 2 1
2 fold decrease
2 fold increase
Log ratio is symmetric -1 1
Log2(2 fold decrease)
Log2(2 fold increase)
Multiplicative in ratioAdditive in logarithm log(A/B)=logA-logB

14. Fold change distribution
Log(fold) distribution

15. Sources of Variance Printing pin
Scanning (laser and detector, PMT, focus)
Hybridization (temperature, time, mixing, etc.)
Probe labeling
RNA preparation
Biological variability

16. Normalization
Many other effects (systematic errors) beside treatment effect can also change gene signal values. Normalization eliminates systematic errors so that gene signals can be compared directly.
Numerous normalization methods are available. How to choose?
Understand sources of variation in your data.
Understand assumptions behind each method.
Diagnostic plot

17. Normalization methods
Dividing by mean or median
Normalized signal =(signal of a spot on an array)/(mean|median intensity of all spots on the array)
This can be done for subset of genes e.g. excluding genes whose intensity is in top 10% or bottom 10% percentile to minimize the effect of outliers or differentially expressed genes.
Subtracting mean: Used for log transformed data
Z-transformation
Normalized signal =(signal of a spot –mean signal of the array)/signal standard deviation of the array

18. Normalization methods
Quantile normalization:

19. Intensity dependent normalization
Housekeeping gene
Normalized signal =(signal of a spot)/(signal of house keeping gene(s))
Intensity dependent normalization
Use local regression to correct non-linear intensity dependency.
2.000
3.000
4.000
-1.0
-0.5
0.0
0.5
-.5
Before Normalization
After Normalization

20. Which genes are differentially expressed?
One of goals of microarray experiment is to find lists of genes that are up or down regulated between treatments
Fold change:
Simple
Low sensitivity
High false positives

21. Hypotheses test
Take into consideration of both magnitude of the change and uncertainty of the measurement.
T-test: two-group comparison
Student t-test: assume equal variance, normal distribution.
Welch method: assume normal distribution, variance is not equal.
Wilcoxon and Mann-Whitney: Non-parametric, no assumption for distribution

22. Analysis of Variance (ANOVA):
Compare multiple groups: Which genes are differentially expressed at least in one condition. Post Hoc test finds the condition(s) that changes gene expression.
Tow- or higher-way ANOVA
One-way ANOVA test only one factor, treatment effect. In microarray there are more than one factors. Some of these are the factors that we are not interested but are not avoidable.
An ANOVA model for two-color microarray
Y=A+D+G+A*D+G*T
Where A=array effect, D=dye effect, G=gene effect, T=treatment effect, A*D=array gene interaction, G*T=gene treatment interaction (usually this is what we are interested)

23. Multiple test and p value adjustment
If the probability to make a false positive when doing t test for a single gene is p=0.05, for 5000 genes you can expect 5000x0.05=250 false positives.
To ensure the probability to make one mistake over the entire 5000 genes is still 0.05 (Family-wised error rate) p-value for each gene need to be adjusted.
Bonferroni adjustments: simple but conservative
p*=min{pxN,1} where p is the raw p value and N is the total number of tests.
Holm or step-down Bonferroni: less conservative
Wellfall and Young’s permutation: Take into consideration of possible correlations between genes. Slow
False discovery rate: Percentage of expected false positives in the gene list.

24. Cluster Analysis
First used by Tryon, 1939 to organize observed data into meaningful structures
Find genes have similar expression profile
Types of cluster analysis: Hierarchical cluster and k-means cluster

25. Hierarchical cluster
Dendrogram or tree shows hierarchical relationship.
Bottom up (agglomerative): Start from individual genes. Measure distance of all pairs of genes/nodes Joint the tow genes/nodes with shortest distance iterate until all genes are jointed

26. g1 g2 g4 g3 g1 g2 g3 g4 d1 d2 d3 d4 d5 d6 g12 g3 g4 d1’ d2’ d3’
Find minimum of {d1…d6}
Find minimum of {d1’…d3’} d1 g1 g2 g4 g3
g124 g3
d1’’
d2’
d1’’

28. K-means cluster: find k clusters that separate as far as possible.
Start from k random clusters and move elements between clusters to minimize the variability within clusters and maximize variability between clusters. Iterate until converged or specified number of iteration is reached.
Some methods are developed to estimate the number of cluster e.g Silhouette plot. However there is no completely satisfactory method for determining the number clusters.

29. Time

30. Distance measurement
Euclidean distance
distance(x,y) = A B C D

31. CCity-block (Manhattan) distance
distance(x,y) =
d(A,B)=a+b+c+d
Result is similar to Euclidean distance. Effect of single outlier is smaller
Both methods measure geometric distance a b c d

32. Angle distance
Euclidean distance does not take into account magnitude. Angle distance measure Angle distance between two vectors. Moving alone the lines do not change distance between A and B x y A B A’ B’ d d’ 
d(x,y)=
Angle distance

33. Measure how close are two genes change in same way.
Pearson correlation
Measure how close are two genes change in same way.
rxy is between –1 and 1. rxy <0 two genes change in opposite ways.
Distance is defined as 1- | rxy |
Spearman correlation
A non-parametric method, similar to Pearson correlation

34. Linkage Determine distance between clusters.
Single linkage (nearest neighbor)
Distance between two nodes is determined by the distance of the two closest objects (nearest neighbors) in the different nodes
Complete linkage (furthest neighbor)
Distances between nodes are determined by the greatest distance between any two objects ("furthest neighbors") in the different nodes.

35. Average (Centroid)
The centroid of a node is the average point in the multidimensional space. It is the center of the node. The distance between two clusters is determined as the distance between centroids.
Single linkage
Average linkage
Complete linkage

36. Self-Organizing Map
Self-Organizing Map (SOM) was introduced by Teuvo Kohonen in 1982.
In artificial neural network, neurons that forms an one or two dimensional elastic net lattice are trained with input data. neurons competes to approximate the density of the data. After the training is over, input data vectors map to n adjacent map neurons

37. Input layer
neurons
Neurons compete for the input pattern. The winner take all.
Winner and neighbors move toward the input pattern.
Neighborhood: Which neurons move with the winner.
Learning rate: How much dose the winner move each time.

38. Other methods Principle component analysis (PCA)
Reduce the dimensionality of the data matrix by finding new variables. Intended to narrow number of variables down to only those that are of importance.
Machine learning: Trained with data set with known classification. Predict or classify new data set. y’ x’ B x A y

40. Biological data mining
GeneOntology: Gene functions are classified into hierarchical structures. The top 3 are : molecular function, biological process and cellular component.
Tools using GO: Onto-Express, EASE, eGOn, GoSurfer
Pathway: KEGG, GeneMapp
Regulatory region analysis:
Tools for regulatory region analysis: Genomatix, Transfac
Gene network:
Tools for gene network: Pathway Assist, iHOP

41. Microarray Standard
MIAME: Minimal Information About a Microarray Experiment. Defining data standards
Information Required to Interpret and Replicate
Experimental Design
Array Design
Biological Samples
Hybridizations
Measurements
Data Normalization and Transformation

42. MIAME checklist: http://www.mged.org/Workgroups/
MIAME/miame_checklist.html
Public database
ArrayExpress (EBI)
GEO (NCBI)
CIBEX (DDBJ)
Other microarray database: BASE, SMD, Oncomine,
YMD