Clustering Large Data Sets in Gene expression analysis Daniel Weaver

1
Clustering Large Data Sets in Gene expression
analysisDaniel Weaver
2
Overview

• What is Gene Expression?
• Scientific questions and clustering techniques

3
The Central Dogma
Transcription
Translation
DNA ? RNA ? Protein

• The arrows represent the transfer or flow of
  information.
• DNA and RNA store information in a base-4 code
  (the four nucleotides).
• Proteins store information in a base-20 code (the
  20 amino acids).

4
Whats in a name?

• DNA?RNA Transcription
• because the information is exactly copied (or
  transcribed) from one base-4 system (DNA) to an
  equivalent base-4 system (RNA). Think of a monk
  transcribing a scroll.
• RNA?Protein Translation
• because the information is converted from a
  base-4 system (RNA) to a base-20 system
  (protein). Think of a monk translating a scroll
  into a new language.

5
What is a gene?

• A gene is a segment of DNA that contains all the
  information necessary to code for some function.
  
• A gene is also the unit of information that is
  transferred through Transcription and Translation.

6
Switching genes on (or off)

• Purpose to correctly control the amount of
  active functional (protein) product present in
  the cell or organism.

Promoter
Enhancer
Figure taken, with permission from Alberts et
al., Molecular Biology of the Cell
7
Presence vs. expression

• All cells have the same set of genes.
• Different cell types express different subsets of
  their genes.
• Constitutive genes are expressed in most cell
  types.
• Cell-type specific genes are expressed in only a
  few cell types.

A B C
A B C
8
Gene expression responds to the environment

• Changes to the cells internal or external
  environment can lead to changes in gene
  expression.
• Most human diseases manifest through a
  mis-regulation of gene expression

A B C
A B C
9
Microarrays and related technologies
10
Example - raw microarray data
more abundant in cell type A
more abundant in cell type B
equally abundant in both cell types
11
Interpreting raw data

• Most gene expression detection data sets are
  expressed as a ratio of RedGreen
  (experimentcontrol) signal.
• Frequently use a normalized log(redgreen) ratio
  for gene X
• Xi
• Such that the Euclidean length of X is 1.
• Interpreted raw data are tabulated in a
  Entity-by-Entity table, Genes-by-Experiments.

12
Gene-by-Experiment table

• Gene expression analysis is a variant of classic
  data mining looking for informative patterns in
  the rows and columns of this type of table.

13
Data volumes

• 120,000 genes in the human genome.
• Expression detection techniques can take from
  1-50 measurement simultaneously on each gene.
• Many, diverse Gene and Experiment attributes
• In 3-5 years, 105 data sets will be available
  for analysis
• Data volumes ranging from 10s of Gb to
• a few Tb

14
Analyzing Gene expression data

• What genes are (or are not) expressed?
• In different cells
• Under different external conditions
• In different disease states
• How much does their expression change?
• Does the change in expression correlate with
  other observed parameters?
• Handled with descriptive statistics

15

• Clustering and Classifying gene expression
• Scientific questions to be answered
• Clustering techniques that are being applied
• Lots of room and need for novel statistical and
  computational analyses

16
Clustering Gene expression data

• Functionally classify novel genes
• Identify co-regulated gene groups
• Identify diagnostic gene expression patterns

17
Functionally Classifying Genes

• Problem
• Genome sequencing projects identify many,
  previously unstudied genes.
• Can one use the genes expression patterns to
  cluster genes that have similar function?

18
Inputs and outputs

• Inputs
• A set of genes whose functional classification is
  know.
• A set of genes whose functional classification is
  unknown.
• Gene expression data sets for all the genes.
• Desired Output
• A best fit functional classification for each
  of the novel genes.

19
Examples

• Brown et al. (2000) PNAS 97(1), 262-267.
• Input
• Log normalized data from 79 experiments on 2,467
  genes
• Trained on 2/3 of the genes, tested on remaining
  3rd.
• Classifiers tried include Support Vector
  Machines and four machine learning algorithms
  (Parzen, FLD, C4.5, MOC1 )
• SVMs performed the best and using the kernel
• K(X,Y) (X?Y1)d (d1,2,or 3)
• This kernel transforms the data into higher
  dimensional space where it is easier to identify
  a separating hyperplane
• Sensitivity 0.6

20
Examples

• Hierarchical clustering, Average linkage (DeRisi
  et al)
• Cluster the genes
• Examine the clusters (through human intervention)
  to determine whether a cluster has a genes with
  known functions.

21
Co-regulated genes

• Problem
• Biological processes typically involve genes of
  many functional categories.
• Knowledge of what genes act coordinately can help
  direct drug development

Expression Group 1
Expression Group 2
Expression Group 3
22
Inputs and Outputs

• Inputs
• Gene expression data for all genes of interest
• (Information about the experimental conditions in
  which the gene expression data sets were
  collected)
• Desired Outputs
• Ordering of the input genes into sets of genes
  with related expression patterns

23
Examples

• Eisen et al. (1998) PNAS 95 14863-14868
• Input
• Log normalized data from 12 experiments on 2,467
  genes
• Performed pair-wise average linkage cluster
  analysis, using a modified Pearson correlation
  coefficient metric
• Gene that cluster together are displayed in a
  dendrogram wherein the branch lengths correlate
  to the degree of similarity

24
Examples

• Tavazoie et al. (1999) Nature Genetics
  22281-285.
• Inputs
• Variance-normalized data from 15 experiments on
  6,220 genes. Variance normalization is Xij (Xij
  Xi)/stdev(Xi) for gene i in experiment j.
• Used Euclidean distance as the metric and
  performed
• k-means clustering, programmed to find 10, 30,
  and 60 centroids.
• Gene clusters were shown to contain functionally
  related genes as expected.

25
Diagnostic expression patterns

• Problem
• Many diseases cannot be reliably distinguished
  through traditional techniques (microscopy,
  pathology, etc.)
• Given gene expression data from diseased tissue,
  is there a set of genes that correctly
  distinguishes the diseases (as judged by other
  criteria).

26
Inputs and Outputs

• Inputs
• Gene expression data for all genes (available)
• Information about the patients afflicted with the
  complex disease of interest.
• Desired output
• The minimal set of genes that accurately
  partitions the disease, i.e. the minimal
  diagnostic gene expression pattern.

27
Examples

• Alizadeh et al. (2000) Nature 403 503-511.
• Input
• Log normalized data from 96 experiments on 4,026
  genes (out of 17,856 measured).
• The 96 experiments were performed on cancer
  biopsies from patients with Diffuse Large B-cell
  Lymphoma (DLBCL).
• Pair-wise average linkage cluster analysis, using
  a modified Pearson correlation coefficient metric
  (Eisen et al., 1998).
• Two previously unknown DLBCL sub-types
  distinguished by small gene clusters (40 genes
  and 70 genes)
• Subtypes correspond to prognosis
• GC B-like ? 76 survivorship
• Activated B-like ? 16 survivorship
• (Overhead)

28
Summary

• Current techniques include supervised and
  unsupervised classification
• Three main scientific questions
• Functionally classifying genes
• Identifying co-regulated sets of genes
• Identifying diagnostic expression fingerprints
• Data sets are relatively small now, but growing
  rapidly.
• Classification draws from the expression data and
  from other domain knowledge.
• Lots of room and need for novel statistical and
  computational analyses

29
Further Reading

• Clustering Gene Expression Data
• Alizadeh, et al. (2000) Nature 403 503-511.
• Alon, et al. (1999) PNAS 96 6745-6750.
• Butte and Kohane. (2000) Proceedings of Pacific
  Sym. Biocomputing.
• Brown, et al. (2000) PNAS 97 262-267.
• Eisen, et al. (1998) PNAS 95 14863-14868.
• Iyer, et al. (1999) Science 283 83-87.
• Raychaudhuri, et al. (2000) Proceedings of
  Pacific Sym. Biocomputing.
• Roberts, et al. (2000) Science 287 873-880.
• Ross et al. (2000) Nature Genetics 24227-235.
• Scherf, et al. (2000) Nature Genetics 24
  236-244.
• Spellman, et al. (1998) Mol Biol Cell 9
  3273-3297.
• Tamayo, et al. (1999) PNAS 96 2907-2912.
• Tavazoie, et al. (1999) Nature Gen 22 281-285.
• Zhu and Zhang. (2000) Proceedings of Pacific Sym.
  Biocomputing.

30
Further Reading

• Other related gene expression papers
• Holstege, et al. (1998) Cell 95717-728.
• DeRisi et al. (1996) Nature Genetics 14457-460.
• Schena et al. (1995) Science 270467-470.
• DeRisi et al. (1997) Science 278680-686.
• Hilsenbeck et al. (1999) J. Natl. Cancer Inst.
  91453-459.

31
Expression Data sets

• European Bioinformatics Institute (EBI) (links to
  refs. 4,5,6,11)
• Main microarray page
• http//www.ebi.ac.uk/microarray/
• Microarray public data set page (this is a great
  portal site from which you can browse to many of
  the published data sets)
• http//industry.ebi.ac.uk/brazma/Data-mining/micr
  oarray.html
• National Human Genome Research Institute (NHGRI)
• Main page
• http//www.nhgri.nih.gov/DIR/LCG/15K/HTML/
• Data set down load page
• ftp//kronos.nhgri.nih.gov/pub/outgoing/olga/old/
• National Cancer Institute (NCI) (ref. 9 10)
• Main page
• http//discover.nci.nih.gov/
• Data set down load page
• http//discover.nci.nih.gov/nature2000/
• Lymphoma data set (ref. 1)
• Main page
• http//llmpp.nih.gov/
• Data set download page

32
Daniel Weaver