Microarray data analysis
Jonathan Pevsner, Ph.D.
Introduction to Bioinformatics pevsner@jhmi.edu
Johns Hopkins School of Public Health (260.602.01)
September 22, 2004
Copyright notice
Many of the images in this powerpoint presentation are from Bioinformatics and Functional Genomics by Jonathan Pevsner (ISBN 0-471-21004-8).
Copyright © 2003 by John Wiley & Sons, Inc.
These images and materials may not be used
without permission from the publisher. We welcome instructors to use these powerpoints for educational purposes, but please acknowledge the source.
The book has a homepage at http://www.bioinfbook.org including hyperlinks to the book chapters.
Schedule
Today : microarray data analysis (Chapter 7) Friday: computer lab (microarray data analysis) Monday: Protein analysis (Chapter 8)
Wednesday: Protein structure (Chapter 9)
Microarray data analysis
• begin with a data matrix (gene expression values versus samples)
Fig. 7.1 Page 190
Microarray data analysis
• begin with a data matrix (gene expression values versus samples)
Typically, there are many genes
(>> 10,000) and few samples (~ 10)
Fig. 7.1 Page 190
Microarray data analysis
• begin with a data matrix (gene expression values versus samples)
Preprocessing
Inferential statistics Descriptive statistics
Fig. 7.1 Page 190
Microarray data analysis: preprocessing
Observed differences in gene expression could be due to transcriptional changes, or they could be
caused by artifacts such as:
• different labeling efficiencies of Cy3, Cy5
• uneven spotting of DNA onto an array surface
• variations in RNA purity or quantity
• variations in washing efficiency
• variations in scanning efficiency
Page 191
Microarray data analysis: preprocessing
The main goal of data preprocessing is to remove the systematic bias in the data as completely as possible, while preserving the variation in gene expression that occurs because of biologically relevant changes in transcription.
A basic assumption of most normalization procedures is that the average gene expression level does not
change in an experiment.
Page 191
Data analysis: global normalization
Global normalization is used to correct two or more data sets. In one common scenario, samples are labeled with Cy3 (green dye) or Cy5 (red dye) and hybridized to DNA elements on a microrarray. After washing, probes are excited with a laser and detected with a scanning confocal microscope.
Page 192
Data analysis: global normalization
Global normalization is used to correct two or more data sets
Example: total fluorescence in Cy3 channel = 4 million units Cy 5 channel = 2 million units
Then the uncorrected ratio for a gene could show
2,000 units versus 1,000 units. This would artifactually appear to show 2-fold regulation.
Page 192
Data analysis: global normalization
Global normalization procedure
Step 1: subtract background intensity values (use a blank region of the array)
Step 2: globally normalize so that the average ratio = 1 (apply this to 1-channel or 2-channel data sets)
Page 192
Microarray data preprocessing
Some researchers use housekeeping genes for global normalization
Visit the Human Gene Expression (HuGE) Index:
www.HugeIndex.org
Page 192
Scatter plots
Useful to represent gene expression values from
two microarray experiments (e.g. control, experimental) Each dot corresponds to a gene expression value
Most dots fall along a line
Outliers represent up-regulated or down-regulated genes
Page 193
Scatter plot analysis of microarray data
Fig. 7.2 Page 193
Brain
Astrocyte Astrocyte
Fibroblast
Differential Gene Expression
in Different Tissue and Cell Types
ex pre ss ion le ve l high low
up
do w n
Expression level (sample 1)
E xp re s si o n le ve l ( sa m p le 2 )
Fig. 7.2 Page 193
Log-log
transformation
Fig. 7.3 Page 195
Scatter plots
Typically, data are plotted on log-log coordinates
Visually, this spreads out the data and offers symmetry raw ratio log2 ratio
time behavior value value
t=0 basal 1.0 0.0
t=1h no change 1.0 0.0
t=2h 2-fold up 2.0 1.0
t=3h 2-fold down 0.5 -1.0
Page 194, 197
expression level low high
up
down Mean log intensity
L o g r at io
Fig. 7.4 Page 196
SNOMAD converts array data to scatter plots http://snomad.org
-1 0 1
-1.0 -0.5 0.0 0.5 1.0
2-fold
2-fold
Log 10 (Ratio )
Mean ( Log10 ( Intensity ) )
EXP
CON
EXP
CON
EXP > CONEXP < CON
2-fold
2-fold
2-fold 2-fold
Linear-linear
plot Log-log
plot
Page 196-197
SNOMAD corrects local variance artifacts
-1 0 1
-1.0 -0.5 0.0 0.5 1.0
-1 0 1
-1.0 -0.5 0.0 0.5 1.0
2-fold
2-fold
Log 10 ( Ratio )
Mean ( Log10 ( Intensity ) )
robust local
regression fit residual
EXP > CONEXP < CON
Corrected Log10 ( Ratio ) [residuals]
Mean ( Log10 ( Intensity ) )
Page 196-197
SNOMAD describes regulated genes in Z-scores
-2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5
-2 -1 0 1 2
Corrected Log10 ( Ratio )
Mean ( Log10 ( Intensity ) )
2-fold
2-fold Locally estimated standard
deviation of positive ratios
Z= 1
Z= -1
Locally estimated standard deviation of negative ratios
-2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5
-10 -5 0 5 10
Local Log10 ( Ratio ) Z-Score
Mean ( Log10 ( Intensity ) )
Z= 5
Z= -5
-2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5
-2 -1 0 1 2
Corrected Log10 ( Ratio )
Mean ( Log10 ( Intensity ) )
2-fold
2-fold
Z= 2
Z= 1
Z= -1
Z= -2 Z= 5
Z= -5
Inferential statistics
Inferential statistics are used to make inferences about a population from a sample.
Hypothesis testing is a common form of inferential statistics. A null hypothesis is stated, such as:
“There is no difference in signal intensity for the gene expression measurements in normal and diseased samples.” The alternative hypothesis is that there is a difference.
We use a test statistic to decide whether to accept or reject the null hypothesis. For many applications,
we set the significance level to p < 0.05.
Page 199
Inferential statistics
A t-test is a commonly used test statistic to assess the difference in mean values between two groups.
t = = Questions
Is the sample size (n) adequate?
Are the data normally distributed?
Is the variance of the data known?
Is the variance the same in the two groups?
Is it appropriate to set the significance level to p < 0.05?
Page 199
x1 – x2
difference between mean values variability (noise)
Inferential statistics
Paradigm Parametric test Nonparametric Compare two
unpaired groups Unpaired t-test Mann-Whitney test Compare two
paired groups Paired t-test Wilcoxon test
Compare 3 or ANOVA
more groups
Table 7-2
Page 198-200
Inferential statistics
Is it appropriate to set the significance level to p < 0.05?
If you hypothesize that a specific gene is up-regulated, you can set the probability value to 0.05.
You might measure the expression of 10,000 genes and hope that any of them are up- or down-regulated. But you can expect to see 5% (500 genes) regulated at the p < 0.05 level by chance alone. To account for the
thousands of repeated measurements you are making, some researchers apply a Bonferroni correction.
The level for statistical significance is divided by the number of measurements, e.g. the criterion becomes:
p < (0.05)/10,000 or p < 5 x 10-6
Page 199
Page 200
Significance analysis of microarrays (SAM)
SAM -- an Excel plug-in (URL: page 202) -- modified t-test
-- adjustable false discovery rate
Fig. 7.7 Page 202
up-
regulated
down-
regulated
expected
observed
Fig. 7.7 Page 202
Descriptive statistics
Microarray data are highly dimensional: there are
many thousands of measurements made from a small number of samples.
Descriptive (exploratory) statistics help you to find meaningful patterns in the data.
A first step is to arrange the data in a matrix.
Next, use a distance metric to define the relatedness of the different data points. Two commonly used
distance metrics are:
-- Euclidean distance
-- Pearson coefficient of correlation
Page 203
Data matrix
(20 genes and 3 time points from Chu et al.)
Fig. 7.8 Page 205
3D plot (using S-PLUS software) t=0
t=0.5 t=2.0
Fig. 7.8 Page 205
Descriptive statistics: clustering
Clustering algorithms offer useful visual descriptions of microarray data.
Genes may be clustered, or samples, or both.
We will next describe hierarchical clustering.
This may be agglomerative (building up the branches of a tree, beginning with the two most closely related objects) or divisive (building the tree by finding the most dissimilar objects first).
In each case, we end up with a tree having branches and nodes.
Page 204
Agglomerative clustering
a b c d e
a,b
4 3
2 1
0
Fig. 7.9 Page 206 Adapted from Kaufman and Rousseeuw (1990)
a b c d e
a,b
d,e
4 3
2 1
0
Agglomerative clustering
Fig. 7.9 Page 206
a b c d e
a,b
d,e
c,d,e
4 3
2 1
0
Agglomerative clustering
Fig. 7.9 Page 206
a b c d e
a,b
d,e
c,d,e
a,b,c,d,e
4 3
2 1
0
Agglomerative clustering
…tree is constructed
Fig. 7.9 Page 206
Divisive clustering a,b,c,d,e
4 3 2 1 0
Fig. 7.9 Page 206
Divisive clustering
c,d,e
a,b,c,d,e
4 3 2 1 0
Fig. 7.9 Page 206
Divisive clustering
d,e
c,d,e
a,b,c,d,e
4 3 2 1 0
Fig. 7.9 Page 206
Divisive clustering a,b
d,e
c,d,e
a,b,c,d,e
4 3 2 1 0
Fig. 7.9 Page 206
Divisive clustering a
b c d e
a,b
d,e
c,d,e
a,b,c,d,e
4 3 2 1 0
…tree is constructed
Fig. 7.9 Page 206
divisive
agglomerative
a b c d e
a,b
d,e
c,d,e
a,b,c,d,e
4 3 2 1 0
4 3
2 1
0
Fig. 7.9 Page 206 Adapted from Kaufman and Rousseeuw (1990)
Fig. 7.8 Page 205
Fig. 7.10 Page 207
1
1 12
12
Agglomerative and divisive clustering
sometimes give conflicting results, as shown here
Fig. 7.10 Page 207
Cluster and TreeView
Fig. 7.11 Page 208
Cluster and TreeView
clustering
K means SOM PCAFig. 7.11 Page 208
Cluster and TreeView
Fig. 7.11 Page 208
Cluster and TreeView
Fig. 7.12 Page 208
Page 208
Fig. 7.12 Page 208
Fig. 7.12 Page 208
Two-way clustering
of genes (y-axis) and cell lines
(x-axis)
(Alizadeh et al., 2000)
Fig. 7.13 Page 209
Self-organizing maps (SOM)
To download GeneCluster:
http://www.genome.wi.mit.edu/MPR/software.html
Page 210
Self-organizing maps (SOM)
One chooses a geometry of 'nodes'-for example, a 3x2 grid
Formerly http://www.genome.wi.mit.edu/MPR/SOM.html
Fig. 7.15 Page 211
Self-organizing maps (SOM)
The nodes are mapped into k-dimensional space, initially at random and then successively adjusted.
Fig. 7.15 Page 211
Self-organizing maps (SOM)
Fig. 7.15 Page 211
Unlike k-means clustering, which is unstructured, SOMs allow one to impose partial structure on the clusters. The principle of SOMs is as follows.
One chooses an initial geometry of “nodes” such as a 3 x 2 rectangular grid (indicated by solid lines in the figure connecting the nodes). Hypothetical trajectories of nodes as they migrate to fit data during successive iterations of SOM algorithm are shown. Data points are represented by black dots, six nodes of SOM by large circles, and trajectories by arrows.
Fig. 7.15 Page 211
Self-organizing maps (SOM)
Neighboring nodes tend to define 'related' clusters.
An SOM based on a rectangular grid thus is analogous to an entomologist's specimen drawer in which
adjacent compartments hold similar insects.
1. Variation Filtering:
Data were passed through a variation filter to eliminate those genes showing no significant change in
expression across the k samples. This step is needed to prevent nodes from being attracted to large sets
of invariant genes.
2. Normalization:
The expression level of each gene was normalized across experiments. This focuses attention on the 'shape' of expression patterns rather than absolute levels of expression.
Two pre-processing steps essential to apply SOMs
Page 210
Principal component axis #2 (10%)
Principal component axis #1 (87%)
PC#3: 1%
C3 C4
C2
C1 N2
N3 N4 P1
P4
P2 P3
Lead (P) Sodium (N) Control (C) Legend
Principal components analysis (PCA),
an exploratory technique that reduces data dimensionality, distinguishes lead-exposed from control cell lines
Page 211
An exploratory technique used to reduce the dimensionality of the data set to 2D or 3D
For a matrix of m genes x n samples, create a new covariance matrix of size n x n
Thus transform some large number of variables into a smaller number of uncorrelated variables called principal components (PCs).
Principal components analysis (PCA)
Page 211
Principal components analysis (PCA): objectives
• to reduce dimensionality
• to determine the linear combination of variables
• to choose the most useful variables (features)
• to visualize multidimensional data
• to identify groups of objects (e.g. genes/samples)
• to identify outliers
Page 211
Page 212
http://www.okstate.edu/artsci/botany/ordinate/PCA.htm
Page 212
http://www.okstate.edu/artsci/botany/ordinate/PCA.htm
Page 212
http://www.okstate.edu/artsci/botany/ordinate/PCA.htm
Page 212
http://www.okstate.edu/artsci/botany/ordinate/PCA.htm
Fig. 7.16 Page 212
Fig. 7.16 Page 212
Chr 21
Use of PCA to demonstrate increased levels of gene expression from Down syndrome (trisomy 21) brain
Practice downloading a dataset (e.g. Chu et al. 1998) from www.dnachip.org
Try making scatter plots in Excel
Try loading the data into Avadis for advanced analyses