1

Robust diagnosis DLBCL from gene Robust diagnosis DLBCL from gene expression data from different expression data from different

laboratories laboratories

Dimacs Workshop, June 22, 2005

Gyan Bhanot, IBM Research

2

Collaborators:

Gabriela Alexe1

Arnold Levine2,3

Gustavo Stolovitzky1 1 IBM2 IAS3 UMDNJ

3

Overview Overview

Motivation

Pattern-based meta-classifiers

Case study – compare data from two labs for

DLBCL vs FL diagnosis

4

Motivation Cancer is a genetic/proteomic disease

Genetic mutations/virus’s/radiation modify pathways

to create survival advantage for damaged cell

Gene Arrays are a way to study the variation of

mRNA levels between diseased and healthy cells.

This allows diagnosis and inference of pathways that

cause disease

5

Cancer diagnosis:Cancer diagnosis:

InputTraining (biomedical / proteomic, microarray) data:

k 2 classes (m samples) described by N >> features

OutputCollection of robust biomarkers, modelsRobust, accurate classifier /

tested on out-of-sample data

6

Data preprocessingCreating training and test dataNormalizationNoise estimation

Robust feature selectionFiltering Support set selection

Artificial Neural Networks

Support Vector Machines

Weighted Voting System (LAD)

k-Nearest Neighbors

Decision Trees(C4.5)

Logistic Regression

Pattern data(training)

Raw data(training)

Principal Components

META-CLASSIFIERValidation(test data)

Input data

Calibration

INDIVIDUAL CLASSIFIERS

INTERMEDIATE CLASSIFIERS

Classifier (Weighted Voting)

Biology-based feature selectionFilteringSupport set selection

7

Strategy of present paperStrategy of present paper

1. Transform original data to “pattern ” space

2. Find robust sets of biomarkers with significant collective

discriminatory power

3. Use many machine learning tools on original and pattern data

ANN, SVM, kNN, Weighted voting, Classification trees

4. Validate the results on data from a different lab

8

PatternsPatterns

Observed dataset System response

9

Positive patterns Negative patterns

Pattern basicsPattern basics

x1,..., xn Rn |i xi ,i I, x j j , j J

10

Individual classifiers usedIndividual classifiers used

SVM, ANN, WV, KNN, CART, LR

Trained / calibrated (leave-one-out):

raw data

pattern data

11

Application: Progression of Follicular Lymphoma (FL) to Diffuse Large B Cell Lymphoma (DLBCL) Gene Array data from different laboratories

Shipp et al. (2002) Nature Med.; 8(1), 68-74. (Whitehead Lab)

Stolovitzky G. (2005) In Deisboeck et al Complex Systems Science in BioMedicine (in press) (preprint: http://www.wkap.nl/prod/a/Stolovitzky.pdf). (DellaFavera Lab)

Alexe et al (2005) Artificial Intelligence in Medicine (in press)

12

Non-Hodgkin lymphomas

FL low grade non-Hodgkin lymphoma

t(14;18) translocation: over-expression of anti-apoptotic bcl2

25-60% FL cases evolve to DLBCL

DLBCL high grade non-Hodgkin lymphoma

< 2 years survival if untreated

Biomarkers: FL transformation to DLBCL• p53/MDM2 (Moller et al., 1999)

• p16 (Pyniol, 1998)

• p38MAPK (Elenitoba-Johnson et al., 2003)

• c-myc (Lossos et al., 2002)

13

Lymphoma datasetsLymphoma datasets

Data: WI (Shipp et al., 2002) Affy HuGeneFL

CU (DallaFavera Lab, Stolovitzky, 2005) Affy Hu95Av2

Samples:

WI: 58 DLBCL & 19 FL

CU: 14 DLBCL & 7 FL

Genes:

WI: 6817

CU: 12581

14

Data Preprocessing

50 % P calls, UL = 16000, LL = 20 2/1 stratify WI data to train/test. CU data testCompute SD per gene across samplesNormalize data to mean 0, SD 1 per geneGenerate 500 data sets using noise + k fold

stratified sampling + jacknife Find genes with high correlation to

phenotype using t-test or SNR. Keep genes that are in > 450/501 of datasets

15

Choosing Support SetsChoosing Support Sets

Create good patterns using small subsets of genes, validate using weighted voting with 10 fold cross validation

Sort genes by their appearance in good patterns

Select top genes to cover each sample by at least 10 patterns

16

The 30 genes that The 30 genes that

best distinguish best distinguish

FL from DLBCLFL from DLBCLG

ene

sym

bol

Ship

p e

t a

l.

Gene

s@

Work

t-te

st

p5

3 r

eg

ula

ted

Bio

log

ical

function

SEPP1 * * * oxidative stress

TXNIP * * metastases suppressor

DNASE1L3 * * apoptosis

CDH11 * * * cell adhesion

LUCA15 * apoptosis

GPR18 * * * signaling pathway

CLU * * * apoptosis

LY9 * * cell adhesion

RHOH * * T-cell differentiation

ELF2 transcription

CCNG2 * cell cycle

CR2 complement activation

CDKN2D * cell cycle

PPP2R5C * signal transduction

G18 cell growth

LY86 * apoptosis

ARPC1B cell motility

MCM7 * * * * cell cycle

BCL2A1 * * * apoptosis

IMPDH2 * * GMP biosynthesis

RRP45 * immune response

STAT1 NF-kappaB cascade

DLG7 * * * cell-cell signaling

SLC1A5 * * transport

TUBB2 * * microtubule movement

PSMA6 protein catabolism

PSMC1 * * * spinocerebellar ataxia

LGALS3 * * * sugar binding

CLTA * * transport

PAGA * * cell proliferation

17

Examples of FL and DLBCL patternsExamples of FL and DLBCL patterns

Pos Neg Pos Neg

P1 >- 1.13 >- 0.62 97 0 91 23

P2 0.91 >- 0.77 95 0 79 31

N1 >- 0.26 - 0.55 0 100 3 54

Training set Test set

Gene Symbol

Pattern

Prevalence (%)

GPR18 CLU DLG7 MCM7

WI training data:

Each DLBCL case satisfies at least one of the patterns P1 and P2

Each FL case satisfies the pattern N1 (and none of the patterns P1 and P2)

18

Pattern dataPattern data

WI training data

WI test data

Positive patterns

Negative

patternsD

LB

CL

FL

CU test data

19

Meta-classifier performanceMeta-classifier performance

Sensitivity (%)

Specificity (%)

Error rate(%)

Sensitivity (%)

Specificity (%)

Error rate(%)

ANN 0.08 94.74 92.31 5.88 82.35 84.62 17.02SVM 0.08 97.37 92.31 3.92 97.06 76.92 8.51kNN 0.09 97.37 100.00 1.96 91.18 84.62 10.64WV 0.07 92.11 92.31 7.84 94.12 76.92 10.64C4.5 0.06 94.74 84.62 7.84 94.12 69.23 12.77LR 0.07 97.37 84.62 5.88 94.12 69.23 12.77ANN 0.10 100.00 100.00 0.00 97.06 76.92 8.51SVM 0.10 100.00 100.00 0.00 97.06 76.92 8.51kNN 0.10 100.00 100.00 0.00 100.00 69.23 8.51WV 0.10 100.00 100.00 0.00 97.06 76.92 8.51C4.5 0.10 100.00 100.00 0.00 91.18 76.92 12.77LR 0.05 100.00 76.92 5.88 100.00 61.54 10.64

100.00 100.00 0.00 100.00 76.92 6.38Meta-classifier

Weight

Training Test

Tra

ined

on

raw

data

Tra

ined

on

patte

rn d

ata

Classifier

20

Error distribution: raw and pattern dataError distribution: raw and pattern data

0 10 20 30 40 50

Meta-classifier

Classifiers trained on pattern data

Classifiers trained on raw data

CU test dataWI test data

21

Biology Based MethodsBiology Based Methods

22

p53 related genes p53 related genes identified by filtering identified by filtering

procedure procedure

CCNB1 EPRS PMAIP1 E2F3MCM7 GSK3B ACAA2 MDM4BRCA1 COL6A1 E2F5* AMPD2BCL2A1 HRAS POLA RBBP4PPP2R4 SERPING1 HMGB2 CCNG2*EIF2S2 CCNA2 PSMB5 HARSCOMT CCT6A ACTA2 CASP6IARS PRKDC INSR RPS6KA1MPI CAD SNRPA GRP58ALAS1 TNFRSF1B G1P2 TP53MRPL3 ZNF184* IMPDH1 SMAD2NCF2 ALDOA MAP2K2 ATP5C1AARS KARS TOP2A TIMP3KIF11 MAD2L1 CXCL1 THBS2CDK4 GOT1 BAG1 MYCBPATP1B1 CDC25B TOP1 DTRCDC20 PSMA1 MAP4 TIMP3PRIM1 KIAA0101 FDFT1 CBSCDC2 PCNA MTA1 CDKN2D*TOP2A TCF3 CDKN1A RELACDK2 CYC1 HLAE*MYC UPP1 PLK1CCNE1 TOPBP1 CDK7

Gene symbol

FL FL DLBCL DLBCL progressionprogression

23

p53 pattern datap53 pattern data

WI data CU data

Positive patterns

Negative

patterns

DL

BC

LF

L

24

Examples of p53 responsive genes patternsExamples of p53 responsive genes patterns M

CM

7

CCN

B1

BCL

2A1

CCN

E1

KIA

A01

01

CD

C2

CBS

E2F

5

Pos

Neg Pos

Neg

P1 >- 0.66 >- 0.89 93 11 86 29P2 >- 0.66 >- 0.78 90 11 71 29P3 >- 0.8 >- 0.33 69 11 64 14N1 - 0.66 3 74 14 71N2 - 0.56 - 0.18 3 68 21 57N3 - 0.11 >0.11 3 68 7 71

Gene symbol

Patte

rn

Prevalence (%)

Training set Test set

WI data:Each DLBCL case satisfies one of the patterns P1, P2, P3Each FL case satisfies one of the patterns N1, N2, N3

25

p53 combinatorial biomarkerp53 combinatorial biomarker

77% FL & 21% DLBCL cases (3.7 fold) at most one gene over-expressed

79% DLBCL & 23% FL cases (3.4 fold)

at least two genes over-expressed

0

10

20

30

40

50

60

70

80

90

<= 1 >=2

# of over-expressed genes in DLBCL vs. FL

(p53, PLK1, CDK2)

% c

ases DLBCL

FL

Each individual gene: over- expressed in about 40-70% DLBCL & 20-40% FL

(specificity 50-60%, sensitivity 60-70%)

26

What are these genes?What are these genes?

Plk1 (stpk13): polo-like kinase serine threonine protein kinase 13, M-phase specific

cell transformation, neoplastic, drives quiescent cells into mitosis over-expressed in various human tumors Takai et al., Oncogene, 2005: plk1 potential target for cancer therapy, new

prognostic marker for cancer Mito et al, Leuk Lymph, 2005: plk1 biomarker for DLBCL

Cdk2 (p33): cyclin -dependent kinase: G2/M transition of mitotic cell cycle, interacts with cyclins A, B3, D, E

P53 tumor suppressor gene (Levine 1982)

27

Conclusions Conclusions

Pattern-based meta-classifier is robust against noise

Good prediction of FL DLBCL Biology Based Analysis also

possible Yields useful BiomarkerShould Study Biologically motivated

sets of genes build pathways

28Thank you for your attention !

<>

29

Artificial neural networksArtificial neural networks

Wij

Wjk

30

Support vector machinesSupport vector machines

H2 H1H

PositiveNegative

miz

xzbxw

xzbxw

w

i

i

i

zCbw

m

ii

,...,1 ,0

negative ~ ,1~positive ~ ,1~

2

1min

1

2

,

--

Find a maximum Find a maximum margin hyperplane margin hyperplane in pattern space in pattern space (Vapnik)(Vapnik)

miCyts

xxyy

ii

ii

ii

ijijiji

,...,2,1 ,0 ,0 ..

2

1min

-

(P) (D)

31

k-Nearest neighborsk-Nearest neighbors

Training data : samples in normalized peptide space Prediction for test data: The dominant class of the k-

nearest neighbors in Euclidean metric

Positive

NegativeNew case: Negative

32

Weighted votingWeighted voting

Pattern data: – each pattern P is a voter– weight = fraction of correctly classified cases

by the pattern– each test case: compute sum of weights of

triggered positive patterns and negative patterns

– classify by highest weight

33

Logistic regressionLogistic regression

Dataset of two phenotypes (e.g., cancer vs. non-cancer)

Transform into logit space y->ln(p/1-p)

Find phenotype predictor as a linear combination of data values in logit space

Insightful Miner

34

Decision trees / forestsDecision trees / forests

Find rules in training data:

– find root feature which best classifies samples by

phenotype

– iterate on each branch to find two new features

which best split each branch by phenotype

– if necessary prune weak support nodes

CART =Classification and Regression Trees (Breiman)

Many trees = forest