1. Clustering Analysis for Microarray Data
Ho Kim
School of Public Health,
Seoul National University

2. Microarrays 대량의 유전자에 대해 발현현상을 동시관찰
유전자의 regulation과 interaction의 이해에 기여
Full Yeast Genome in a Chip
(Brown Lab, Stanford. Univ.)

3. Microarray Experiment
cDNA Microarray Experiment using Apo AI

4. Microarray & Statistical Method
생물학적 의문점
-유전자 발현양상의 차이
-유전자 및 샘플의 분류
Experimental Design
Microarray Experiment
Image Processing
Normalization
Statistical problems in all the procedures
Validation
Estimation
Testing
Clustering
Classification
생물학적 해석 및 확인

5. Image Analysis-Practical Problems 1-
Comet Tails
Likely caused by insufficiently rapid immersion of the slides in the succinic anhydride blocking solution.

6. Image Analysis-Practical Problems 2-

7. Image Analysis-Practical Problems 3-
High Background
2 likely causes:
Insufficient blocking.
Precipitation of the labeled probe.
Weak Signals

8. Image Analysis-Practical Problems 4-
Spot overlap:
Likely cause: too
much rehydration
during post -
processing.

9. Steps in Images Processing
1. Addressing: locate centers
2. Segmentation: classification of pixels either as signal or background. using
seeded region growing).
3. Information extraction: for
each spot of the array,
calculates signal intensity
pairs, background and quality
measures.

10. Normalization Microarray자료에서 발현수준에 영향을 미치는
기술적 변이(systematic variation)를
찾아내서 제거함

11. M vs. A(Apo AI KO Experiment)
M = log2(R / G) A = log2(R*G) / 2

12. Lowess Regression Global lowess
Assumption: changes roughly symmetric at all intensities.

13. Print-tip-group Normalization
Assumption: For every print group, changes roughly symmetric
at all intensities.
log2R/G -> log2R/G – ci(A) = log2R/ (ki(A)G)
ci(A) is the loess fit to the M vs. A plot for the i-th grid

14. Effect of Location Normalization
Before normalization
After print-tip-group
Normalization
(just location)

15. Effect of location + scale Normalization
Just location

16. Cluster Analysis: intro
Genes with similar function yield similar expression patterns in microarray experiments
Searching for the groups (clusters) in the data, based on a measure or distance index of similarity or dissimilarity
cluster - a set of entities which are alike

17. Similarity
The similarity(dissimilarity) between two objects measures how different they are.
Actually we use distance function between two objects
Distance property
1. d(x, x) = 0
2. d(x, y) > 0
3. d(x, y) = d(y, x)

18. Kinds of Distance-1 Euclidean : Manhattan : Maximum, Binary…
(X2,Y2)
(X2,Y2)
(X1,Y1)
(X1,Y1)
Euclidean Distance
Manhattan Distance

19. Kinds of Distance-2
Correlation

20. K-Means Clustering
Select k initial seed
K is pre-deoermined

21. K-Means Clustering For each data point, assign cluster
Calculate cluster mean
Seed is changed to the mean of the cluster

22. K-Means Clustering
Repeats until seeds don’t change

23. K-Means Clustering
The result depends on initial seeds. As the dimension goes up, the problem gets more serious.
As the number of clusters K is changed, the cluster membership can change in arbitrary ways. That is, with say four clusters, the clusters need not be nested within the three clusters above.
K-means does not give a linear ordering of objects within a cluster

24. EXAMPLE

30. Hierarchical Clustering
find relatively homogeneous clusters of cases based on measured characteristics. 
Agglomerative(bottom-up)
starts with each case in a separate cluster
and then combines the clusters sequentially, reducing the number of clusters at each step until only one cluster is left. 
When there are N cases, this involves N-1 clustering steps, or fusions.
Divisive(top-down)
Reverse procedure of agglomerative

31. Dissimilarity measures
– single link: similarity of two most
similar members
– complete link: similarity of two least
– average link: average similarity
between members A B C D 2 5 3 9 7 4

32. Dissimilarity measures single linkage
D((AB), C) =min{(AC),(BC)} =min{5,3} =3
D((AB), D) =min{(AD),(BD)} =min{9,7} =7
D(C,D) = 4
min{3,7,4}=3(AB),C
D((ABC),D)=min{(AD),(BD),(CD)}=min{9,7,4}=4 4 A B C D 2 5 3 9 7 4 3 2
A B C D
Single

33. Dissimilarity measures complete linkage
D((AB), C) =max{(AC),(BC)} =max{5,3} =5
D((AB), D) =max{(AD),(BD)} =max{9,7} =9
D(C,D) = 4
min{5,9,4}=4(AB),(CD)
D((AB),(CD))=max{(AC),(AD),(BC),(BD)}=min{5,9,3,7}=9 A B C D 2 5 3 9 7 4 9 2 4
Complete

34. Dissimilarity measures average linkage
D((AB), C) =mean{(AC),(BC)} =mean{5,3} =4
D((AB), D) =mean{(AD),(BD)} =mean{9,7} =8
D(C,D) = 4 *
min{5,9,4}=4(AB),(CD)
D((AB),(CD))=mean{(AC),(AD),(BC),(BD)}=mean{5,9,3,7}=6 A B C D 2 5 3 9 7 4 6 2 4
Complete

35. Dendrogram Hierrarchical Clustering(Eisen et al. 1998)
Clustered display of data from time course of serum stimulation of primary human fibroblasts

36. EXAMPLE

38. Self Organizing Map(SOM)

39. SOM is a neural-network that is based on divisive clustering approach.
Microarray Data: High Dimensional
p genes, n treatments/times
→ n times p-dimensional points.
SOM: find clusters on 2(or 3)-D
surface in n-D space.

40. SOM Two layers : data layer & neuron layer
Data layer: input, want to find pattern
Neuron layer: collection of neuron
Neurons are connected with other neurons in neuron layer as well as data in data layer
Neuron layer (through learning) will make adaptation to data layer (reduce complexity and, easy to understand for human)

42. EXAMPLE

43. SOM parameters 1) Theta- the size of the neighborhood function
2) Phi- the amount a neuron within the neighborhood circumference will be moved towards the input-vector
3) Momentum- the amount to reduce the learning parameters theta and phi after each iteration
4) Tolerance- the circumference around each neuron that will define the set of input-vectors that belongs to each neuron
5) nxn neurons- the number of neurons in the lattice
6) Phi tolerance- a limit for how long the algorithm will keep running
7) Pause- the number of milliseconds to wait at each iteration
8) Updates before repaint- a way to control how many iterations the algorithm will perform before updating the plot

48. Example - Tamayo et al.(1999)
6 by 5 SOM.
The 828 genes that passed the variation filter were grouped into 30 clusters.
Each cluster is represented by the centroid (average pattern) for genes in the cluster.
Expression levels are shown on y-axis and time points on x-axis. Error bars indicate the SD of average expression. n indicates the number of genes within each cluster.

49. HL-60 SOM. HL-60 cells were treated with PMA for 0, 0
HL-60 SOM. HL-60 cells were treated with PMA for 0, 0.5, 4, or 24 hours,
and expression levels of more than 6,000 genes were measured at each time point.
The 567 genes passing the variation filter were grouped by a SOM.

50. Other methods,-partitioning
K-Means(Ruspini, 1970; centroid):
Iterative relocation clustering
PAM(Partitioning Around Medoids):
robust than K-Means
CLARA(Clustering Large Applications):
FUSSY: fractions of membership
MCLUST: probability-based model
Bayes factor for choosing k clusters
MDS(Multidimensional Scaling: Tamayo et al.,1999)
SOM(Self-Organizing Map: Tamayo et al.,1999)

51. Other methods, etc CLICK(Sharan and Shamir, 2000)
Gene Shaving(Hastie et al., 2000)
Gene Voting(Golub et al., 1999)
SVM(Support Vector Machine; Brown et al.,2000)

52. Software 상용화 S/W NAME COMPANY FEATURE ArraySuite Affymetrix
plots, fold changes
ImaGene
Biodiscovery
quantification of gene expression value, constant-factor normalization
GeneSight
background adjustment, clustering(hierarchical, SOM)
GeneSpring
Silicon
Genetics
normalization, clustering(hierarchical, SOM), fold-change
Spotfire
PCA, clustering, fold-change
Resolver
Rosetta
clustering, PCA, fold-change, plots
LifeArray
Incyte
clustering, PCA, fold-change
Expressionist
GeneData
clust, PCA, fold-change
GeneExpress
Gene Logic
IPLab MicroArray Suit
Scanalytics
clustering, image, fold-changes

53. 공개용 S/W NAME ORG FEATURE J-Express U Bergen clustering, PCA UCI/NCGR
t-test for fold-change
TreeView
Stanford
clustering, SOM, image analysis
(similarity, Cluster,Lawrence Berkeley Lab,  Eisen  Lab)
EPCLUST
EBI
clustering
SOM
Whitehead Inst.

54. Software - website link
Recently updated site for software of microarray data analysis
열린 강의실, 세미나자료