1 Decision tree based classifications of heterogeneous lung cancer data Student: Yi LI Supervisor: Associate Prof. Jiuyong Li Data: 15 th May 2009.

1

Decision tree based Decision tree based classificationclassificationss of of heterogeneous lung heterogeneous lung cancer datacancer data

Student: Yi LI Supervisor: Associate Prof. Jiuyong Li

Data: 15th May 2009

2

Outline Microarray dataMotivationsRelated workOur integrated frameworkExperiments Discussions Conclusion

3

Microarray data

Microarray rationale (Babu, 2004b)

4

Microarray data (con’d)Snapshot of DNA oligonucleotides

Reveal rich biological information: DNA sequences, cell structures &

cancer

Hugh amount of data:Number of attributes in thousands or

moreNumber of samples in hundreds or less

Microarray data (con’d)Microarray data (con’d)

5

its values

Gene name

Patient samples

A part of a microarray data set

Motivations Motivations Key goal: to find out reliable and

robust predictors (gene sets)

However, microarray studies addressing similar prediction tasks report different sets of predictive genes

6

Motivations (con’d)Motivations (con’d)Two-dimensional cluster analysis

+ leave-one-out cross-validation [van’t Veer et al. (2002)]

Cox’s proportional-hazards regression + clustering [Wang et al. (2005)]

7

Research questionResearch question

How to build up a framework to improve the prediction accuracy among heterogeneous microarray data sets?

8

Dilemma 1Dilemma 1

Usually a microarray data set contains

thousands of features, but with limited

number of samples.

It creates troubles to expect robust and

reliable classifiers.9

Related workRelated workCurse of data set sparsity + curse

of dimensionality [Somorjai et al. (2003)]

◦Use simple classifiers to show how those curses influence outcomes

◦Samples per feature ratio (SFR) in microarray data set is too small to expect robust classifiers.

◦Conventional solutions: feature redundant, apply classifiers that do not require feature space redundant.

10

Related workRelated workProbably approximately correct

sorting (PAC) [Ein-Dor et al. (2006)]

◦Use PAC to evaluate the robustness of results

◦Determine the number of samples that are required to achieve any desired level of reproducibility

11

Dilemma 2Dilemma 2

Heterogeneous microarray platforms,

differences in equipment and protocols, and

differences in the analysis methods may also

cause discordance between independent

experiments.

12

Related workRelated workCorrelation and concordance

calculations [Kuo et al. (2002)]Median rank scores + quantile

discretization + SVM [Warnat et al. (2005)]

◦Stanford type cDNA microarrays and Affymetrix oligonucleotide microarrays

13

Dilemma 3Dilemma 3

Eliminating the factors mentioned in

dilemma 1 & 2, the discrepancies between

studies still remain.

14

Related workRelated workExpand standard strategy to

multiple sets [Michiels et al. (2005)]

SVM-RFE + 5-fold cross-validation + joint-core [Fishel, I et al. (2007)]

◦There are many optimal predictive gene sets, which are strongly dependent on the subset of samples chosen for training.

15

Research goalResearch goal

Our purpose to build a robust and reliable

model to study heterogeneous microarray

data sets, to reduce study-specific biases, and

aiming to yield results which offer improved

reliability and validity.

16

Our integrated frameworkOur integrated framework

1. Classification on single data set◦ Standard classification◦ Single tree, Bagging & Random Forest

2. Classification on integrated data sets

◦ Low-level data integration◦ Single tree, Bagging & Random Forest

3. Classification on integrating models from multiple data sets

◦ High-level model integration◦ Integrated model based on two single trees

17

Our integrated framework Our integrated framework (con’d) (con’d)

18

Training set Test

set

Single tree Harvard

Michigan

Harvard + MichiganStanfor

d

Bagging

Random

Forest

Integrated

model

Single tree built upon

Harvard + single tree built

up Michigan

Available data setsAvailable data sets

Name Harvard Michigan Stanford

# of attributes

11657 6357 11985

# of samples 156 96 464

Data type Continuous Continuous

Continuous

Missing values?

No No Yes

Class (ADEN/normal)

139/17 86/10 41/5

Gene type Affymetrix ID

Affymetrix ID

Unknown 19

* All data sets are in .CSV format* Attribute names are denoted by gene probe names* All data sets are independent to each other

Available data sets Available data sets (con’d)(con’d)Harvard_Unique_probname.csvMichigan_Unique_probname.csv

◦Two columns: Probe & Gene Symbol◦Mapping files: maps probe names

with its corresponding gene symbols◦Multiple probe names may map to

one gene symbol

20

Data pre-processingData pre-processingGene name substitution◦R-programming language◦Remove missing values◦Remove duplicated genes

Remove all, including the 1st appeared one

◦Find out overlapping genes Find the common gene subsets between

Harvard and Michigan

21

Data pre-processing Data pre-processing (con’d)(con’d)◦Substitute gene symbols with probe

names H and M contain the same set of genes

(not same sequence) Stanford contains the same set, too

22

Data pre-processing Data pre-processing (con’d)(con’d)Feature selection◦Weka◦GainRatioAttributeEval > Ranker◦Select 100 highly ranked genes from

H & M, separately◦48 of them are overlapped, 52 genes

are unique

23

Data pre-processing Data pre-processing (con’d)(con’d)

24

◦3 parts: unique genes of H’, overlapping genes and unique genes of M’

◦H’ , M’ and S’: with gene set of the union parts above: H’: ‘?’s to indicate unique genes of M’ M’: ‘?’s to indicate unique gene of H’ S’: no missing values generated in this stage

Data pre-processing Data pre-processing (con’d)(con’d)Discretization ◦Mean value◦R-programming language◦Missing values

25

Data pre-processing Data pre-processing (con’d)(con’d)Handle incompatible format◦ARFF format◦Attribute section

Same sequence of attributes Same possible values with same

sequence

◦Data section Values must match their corresponding

data types

26

Experiments 1 Experiments 1 Weka Explorer Build single decision trees on data

sets◦Classify > Classifier > trees > J48◦Test options > Supplied test set

Build Bagging trees on data sets◦Classify > Classifier > meta > Bagging

Build Random Forest on data sets◦Classify > Classifier > meta >

RandomCommittee (Classifier >RandomForest)

27

Experiment 2Experiment 2MatlabBuild single trees upon H and M,

separatelyFor an unseen instance, do

prediction on two models, ◦if the predicted classes are the same,

then keep it as it is;◦otherwise, the class label with greater

confidence value wins.◦Accuracy = no. of correctly predicted /

total 28

Experiment resultsExperiment results

29

Experiments (con’d)Experiments (con’d)

30

Experiments (con’d)Experiments (con’d)

31

32

Major reference Babu, M. M. 2004b, “An introduction to Microarray data analysis” MRC Lab

page, visited on 15 June 2008, <http://www.mrc-lmb.cam.ac.uk/genomes/madanm/microarray/.>

Choi. J.K. et al. (2003) Combining multiple microarray studies and modeling interstudy variation. Bioinformatics, 19, i84-i90.

Ein-Dor, L. et al. (2005) Outcome signature genes in breast cancer: is there a unique set? Bioinformatics, 21, 171-178.

Ein-Dor, L. et al.(2006) Thousands of samples are needed to generate a robust gene list for predicting outcome in cancer. PNAS, 103, 5923-5928.

Fishel, I. et al. (2007) Meta-analysis of gene expression data: a predictor-based approach. Bioinformatics, Vol. 23, 1599-1606.

Jiang, H. et al. (2004) Joint analysis of two microarray gene expression data sets to select lung adenocarcinoma marker genes. BMC Bioinformatics, 5, 81.

Kuo, W.P. et al. (2002) Analysis of matched mRNA measurements from two different microarray technologies. Bioinformatics, 18, 405-412.

Michiels, S. et al. (2005) Prediction of cancer outcome with microarrays: a multiple random validation strategy, Lancet, 365, 488-492.

Rhodes, D. R. et al. (2002) Meta-analysis of microarrays: interstudy validation of gene expression profiles reveals pathway dysregulation in prostate cancer. Cancer Res., 62, 4427-4433.

Van’t Veer, L.J. (2002) Gene expression profiling predicts clinical outcome of breast cancer. Nature, 415, 530-536.

Wang, Y. et al. (2005) Gene-expression profiles to predict distant metastasis of lymph-node-negative primary breast cancer. Lancet, 365, 671-679.