Transcript Microarrays
Knowledge Discovery in
Microarray
Gene Expression Data
Gregory Piatetsky-Shapiro
[email protected]
IMA 2002 Workshop on Data-driven Control and Optimization
Copyright © 2002 KDnuggets
Data Mining Methodology is Critical!
CRISP-DM methodology
Data Mining is a
Continuous
Process!
Following Correct
Methodology
is Critical!
Copyright © 2002 KDnuggets
2
IMA-2002 Workshop
Overview
Molecular Biology Overview
Microarrays for Gene Expression
Classification on Microarray Data
avoiding false positives
wrapper approach
Microarrays for Modeling Dynamic Processes
finding causal networks and clusters
Copyright © 2002 KDnuggets
3
IMA-2002 Workshop
Biology and Cells
All living organisms consist of cells.
Humans have trillions of cells. Yeast - one cell.
Cells are of many different types (blood, skin,
nerve), but all arose from a single cell (the
fertilized egg)
Each* cell contains a complete copy of the
genome (the program for making the organism),
encoded in DNA.
Copyright © 2002 KDnuggets
4
IMA-2002 Workshop
DNA
DNA molecules are long double-stranded chains;
4 types of bases are attached to the backbone:
adenine (A), guanine (G), cytosine (C), and
thymine (T). A pairs with T, C with G.
A gene is a segment of DNA that specifies how to
make a protein.
Human DNA has about 30-35,000 genes;
Rice -- about 50-60,000, but shorter genes.
Copyright © 2002 KDnuggets
5
IMA-2002 Workshop
Exons and Introns: Data and Logic?
exons are coding DNA (translated into a protein),
which are only about 2% of human genome
introns are non-coding DNA, which provide
structural integrity and regulatory (control)
functions
exons can be thought of program data, while
introns provide the program logic
Humans have much more control structure than
rice
Copyright © 2002 KDnuggets
6
IMA-2002 Workshop
Gene Expression
Cells are different because of differential gene
expression.
About 40% of human genes are expressed at one
time.
Gene is expressed by transcribing DNA into
single-stranded mRNA
mRNA is later translated into a protein
Microarrays measure the level of mRNA
expression
Copyright © 2002 KDnuggets
7
IMA-2002 Workshop
Molecular Biology Overview
Cell
Nucleus
Chromosome
Protein
Copyright © 2002 KDnuggets
Gene (mRNA),
single strand
8
Gene (DNA)
Graphics courtesy of the National Human Genome Research Institute
IMA-2002 Workshop
Gene Expression Measurement
mRNA expression represents dynamic aspects of
cell
mRNA expression can be measured with latest
technology
mRNA is isolated and labeled with fluorescent
protein
mRNA is hybridized to the target; level of
hybridization corresponds to light emission which
is measured with a laser
Copyright © 2002 KDnuggets
9
IMA-2002 Workshop
Gene Expression Microarrays
The main types of gene expression microarrays:
Short oligonucleotide arrays (Affymetrix);
cDNA or spotted arrays (Brown/Botstein).
Long oligonucleotide arrays (Agilent Inkjet);
Fiber-optic arrays
...
Copyright © 2002 KDnuggets
10
IMA-2002 Workshop
Affymetrix Microarrays
Raw image
1.28cm
50um
~107 oligonucleotides,
half Perfectly Match mRNA (PM),
half have one Mismatch (MM)
Raw gene expression is intensity
difference: PM - MM
Copyright © 2002 KDnuggets
11
IMA-2002 Workshop
Microarray Potential Applications
Biological discovery
new and better molecular diagnostics
new molecular targets for therapy
finding and refining biological pathways
Recent examples
molecular diagnosis of leukemia, breast cancer, ...
appropriate treatment for genetic signature
potential new drug targets
Copyright © 2002 KDnuggets
12
IMA-2002 Workshop
Microarray Data Analysis Types
Gene Selection
find genes for therapeutic targets
avoid false positives (FDA approval ?)
Classification (Supervised)
identify disease
predict outcome / select best treatment
Clustering (Unsupervised)
find new biological classes / refine existing ones
exploration
Copyright © 2002 KDnuggets
13
IMA-2002 Workshop
Microarray Data Mining Challenges
too few records (samples), usually < 100
too many columns (genes), usually > 1,000
Too many columns likely to lead to False positives
for exploration, a large set of all relevant genes is
desired
for diagnostics or identification of therapeutic
targets, the smallest set of genes is needed
model needs to be explainable to biologists
Copyright © 2002 KDnuggets
14
IMA-2002 Workshop
Data Preparation Issues (MAS-4)
Thresholding: usually min 20, max 16,000
For older Affy chips (new Affy chips do not have
negative values)
Filtering - remove genes with insufficient variation
e.g. MaxVal - MinVal < 500 and MaxVal/MinVal < 5
biological reasons
feature reduction for algorithmic
For clustering, normalize each gene (sample)
separately to Mean = 0, Std. Dev = 1
Copyright © 2002 KDnuggets
16
IMA-2002 Workshop
Classification
desired features:
robust in presence of false positives
understandable
return confidence/probability
fast enough
simplest approaches are most robust
advanced approaches can be more accurate
Copyright © 2002 KDnuggets
17
IMA-2002 Workshop
FALSE POSITIVES PROBLEM
Not enough records (samples), usually < 100
Too many columns (genes), usually >>1,000
FALSE POSITIVES are very likely because of
few records and many columns
Copyright © 2002 KDnuggets
18
IMA-2002 Workshop
Controlling False Positives
CD37 antigen
Class
178
105
4174
7133
1
1
2
2
Class
Avg
Std
1
2
2287.9
4457.5
1452.4
2010.3
Mean Difference between Classes:
T-value = -3.25
Significance: p=0.0007
Copyright © 2002 KDnuggets
19
IMA-2002 Workshop
Controlling False Positives with
Randomization
CD37 antigen
178
105
4174
7133
Randomized
Class
Class
1
1
2
2
Randomize
2
1
1
2
Randomization is
Less Conservative
Preserves inner
structure of data
Class
178
105
4174
7133
Copyright © 2002 KDnuggets
2
1
1
2
20
T-value = -1.1
IMA-2002 Workshop
Controlling false positives with
randomization, II
Gene
Class
178
105
4174
7133
1
1
2
2
Copyright © 2002 KDnuggets
Rand
Class
Randomize
500 times
2
1
1
2
Gene
Class
178
105
4174
7133
2
1
1
2
21
Bottom
1% T-value = -2.08
Select potentially
interesting genes at 1%
IMA-2002 Workshop
Controlling False Positives:
SAM (Statistical Analysis of Microarrays)
Tusher, Tibshirani, and Chu, Significance analysis
of microarrays …, PNAS, Apr 2001
SAM software available from Tibshirani web site
Copyright © 2002 KDnuggets
22
IMA-2002 Workshop
Feature selection approach
Rank genes by measure; select top 200-500
T-test for Mean Difference=
( Avg1 Avg2 )
( 1 / N1 2 / N 2 )
( Avg1 Avg2 )
Signal to Noise (S2N) =
( 1 2 )
Other: Information-based, biological?
Almost any method works well with a good
feature selection
Copyright © 2002 KDnuggets
24
IMA-2002 Workshop
Gene Reduction improves Classification
most learning algorithms looks for non-linear
combinations of features -- can easily find many
spurious combinations given small # of records
and large # of genes
Classification accuracy improves if we first reduce
# of genes by a linear method, e.g. T-values of
mean difference
Heuristic: select equal # genes from each class
Then apply a favorite machine learning algorithm
Copyright © 2002 KDnuggets
25
IMA-2002 Workshop
Wrapper approach to
select the best gene set
Select best 200 or so genes based on statistical measures
Test models using 1,2,3, …, 10, 20, 30, 40, ... genes with xvalidation. Select gene set with lowest average error
Heuristically, at least 10 genes overall
Error Avg for 10-fold X-val
30%
25%
20%
15%
10%
5%
0%
1
2
3
4
5
10
20
30
40
Genes per Class
Copyright © 2002 KDnuggets
26
IMA-2002 Workshop
Popular Classification Methods
Decision Trees/Rules
find smallest gene sets, but not robust false positives
Neural Nets - work well for reduced # of genes
K-nearest neighbor - robust for small # genes
TreeNet from authors of CART and MARS
networks of simple trees; very robust against outliers
Support Vector Machines (SVM)
good accuracy, does its own gene selection, but hard to
understand
...
Copyright © 2002 KDnuggets
27
IMA-2002 Workshop
Microarrays: An Example
Leukemia: Acute Lymphoblastic (ALL) vs Acute
Myeloid (AML), Golub et al, Science, v.286, 1999
72 examples (38 train, 34 test), about 7,000 genes
well-studied (CAMDA-2000), good test example
ALL
AML
Visually similar, but genetically very different
Copyright © 2002 KDnuggets
28
IMA-2002 Workshop
Results on the test data
Genes selected and model trained on Train set
ONLY!
Best Clementine neural net model used 10 genes
per class
Evaluation on test data (34 samples) gives
1 or 2 errors (94-97% accuracy),
Note: all methods give error on sample 66, believed to
be mis-classified by a pathologist
Copyright © 2002 KDnuggets
29
IMA-2002 Workshop
Multi-class Data Analysis
Brain data, Pomeroy et al 2002, Nature (415), Jan
2002
42 examples, about 7,000 genes, 5 classes
Photomicrographs of tumours (400x)
a, MD (medulloblastoma) classis
b, MD desmoplastic
c, PNET
d, rhabdoid
e, glioblastoma
Analysis also used Normal tissue, not
shown
Copyright © 2002 KDnuggets
30
IMA-2002 Workshop
Modeling with TreeNet
Build a model using top 3 genes from each class
Evaluate using cross-validation
Results: 95% accuracy:
1 error on training data, 1 on test
0.5
0.4
Risk
0.3
0.2
0.1
0.0
0
10
20
30
Number of Trees
Copyright © 2002 KDnuggets
31
IMA-2002 Workshop
TreeNet results for multi-class data
Class
MD
MGlio
Normal
PNET
Rhab
Learn
Cases
(Errors)
7 (0)
8 (0)
3 (0)
6 (1)
8 (0)
Test
Cases
3 (0)
2 (0)
1 (0)
2 (0)
2 (1)
Average cross-validation accuracy over 95%
Original authors had accuracy of about 85% using
nearest neighbor classifier.
Copyright © 2002 KDnuggets
32
IMA-2002 Workshop
Yeast SOM Clusters
Yeast Cell Cycle SOM.
www.pnas.org/cgi/content/full/96/6/2907
(a) 6 × 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 level of each gene was normalized to have mean = 0 and SD = 1
across time points. 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. Note that multiple clusters exhibit periodic behavior and that
adjacent clusters have similar behavior. (b) Cluster 29 detail. Cluster 29 contains 76
genes exhibiting periodic behavior with peak expression in late G1. Normalized
expression pattern of 30 genes nearest the centroid are shown. (c) Centroids for SOMderived clusters 29, 14, 1, and 5, corresponding to G1, S, G2 and M phases of the cell
cycle, are shown.
Copyright © 2002 KDnuggets
34
IMA-2002 Workshop
Yeast SOM Clusters
Copyright © 2002 KDnuggets
35
IMA-2002 Workshop
Discovery of causal processes
A long term goal of Systems Biology is to discover
the causal processes among genes, proteins, and
other molecules in cells
Can this be done (in part) by using data from
High Throughput experiments, such as
microarrays?
Copyright © 2002 KDnuggets
36
IMA-2002 Workshop
A Model of Galactose Utilization
(manually discovered)
T. Ideker, et al., Science 292 (May 4, 2001) 929-934.
Copyright © 2002 KDnuggets
37
IMA-2002 Workshop
Bayesian Causal Network Structure
P(GAL4)
P(GAL2 | GAL4)
P(Intracellular Galactose | GAL2)
Each variable is independent of
its distant causes given all of its
direct causes.
Thanks to Greg Cooper, U. Pitt
Copyright © 2002 KDnuggets
38
IMA-2002 Workshop
Bayesian Network Learned for Yeast
Hartemink et al, Combining Location and Expression Data for
Principled Discovery of Genetic Regulatory Network Models,
PSB 2002 psb.stanford.edu/psb-online
Copyright © 2002 KDnuggets
39
IMA-2002 Workshop
Future directions for Microarray Analysis
Algorithms optimized for small samples
Integration with other data
biological networks
medical text
protein data
Cost-sensitive classification algorithms
error cost depends on outcome (don’t want to miss
treatable cancer), treatment side effects, etc.
Copyright © 2002 KDnuggets
40
IMA-2002 Workshop
Integrate biological knowledge when analyzing
microarray data (from Cheng Li, Harvard SPH)
Right picture: Gene Ontology: tool for the unification of biology, Nature Genetics, 25, p25
Copyright © 2002 KDnuggets
41
IMA-2002 Workshop
GeneSpring Demo
Yeast data
Zoom all the way to bases
Yeast Cycle -- animation
Color -- expression strength
Copyright © 2002 KDnuggets
42
IMA-2002 Workshop
Acknowledgements
Sridhar Ramaswamy, MIT Whitehead Institute
Pablo Tamayo, MIT Whitehead Institute
Greg Cooper, U. Pittsburgh
Tom Khabaza, SPSS
Copyright © 2002 KDnuggets
43
IMA-2002 Workshop
Thank you!
Further resources on Data Mining:
www.KDnuggets.com
Contact:
Gregory Piatetsky-Shapiro: [email protected]
Copyright © 2002 KDnuggets
44
IMA-2002 Workshop