Benjamin Haibe-Kains Director, Bioinformatics and Computational Genomics Laboratory Scientific Advisor, Bioinformatics Core Facility Are pharmacogenomic.

Benjamin Haibe-Kains

Director, Bioinformatics and Computational Genomics LaboratoryScientific Advisor, Bioinformatics Core Facility

Are pharmacogenomic studies useful for developing predictors of drug response?

Non-Responders

Responders

D

C

A

B

Treat with conventional drugs

Treat with alternative drugs

Genomic data

Genomic predictive biomarkers

E

Benjamin Haibe-Kains QBBMM Conference 2013-09-20

• Predicting therapeutic response of patients based on their genomic profiles

Adapted from Luo et al. Cell, 2009

Therapeutic strategies in cancer


• Many drug compounds have been designed and many others are under development

• Success stories enabled to develop relevant therapeutic strategies and bring them to the clinic

• But the number of new (targeted) drugs being approved is dramatically slowing down

• Need for companion tests to identify patients who are likely to respond to targeted therapies

Anticancer therapies


• It is not sustainable to test thousands of compounds (and their combinations) in clinical trials

• One needs a different approach to screen the therapeutic potential of new compounds

• Cancer cell lines can be used as preclinical models:Cheap and high-throughputSimple models to investigate drugs’ mechanisms of

action Enable to build genomic predictors of drug response

Drug screening in preclinical models


Current studies

• Most studies investigated isolated, small pharmacogenomic datasets

• Very few have been validated in independent experiments and in clinical samples

• Some are sadly famous: Anil Potti’s scandal at Duke University [forensic Bioinformatics by Baggerly and Coombes]

The solution may lie in analyzing large collections of

cell lines from multiple datasets


Pharmacogenomic data

Resistant vs. sensitive cell lines


Large pharmacogenomic datasets


• The Cancer Cell Line Encyclopedia (CCLE) initiated by Novartis/Broad Institute

• 24 drugs• 1036 cancer cell lines

• Large-scale studies have been recently published in Nature

• The Cancer Genome Project (CGP) initiated by the Sanger Institute

• 138 drugs• 727 cancer cell lines

CGP CCLE

• Drugs: 15 drugs have been investigated both in CGP and CCLE

CCLECGP

256 471 565

• Cell lines: 471 cancer cell lines in common between CGP and CCLE

Paclitaxel Microtubules depolymerization inhibitor

PD-0325901, AZD6244 Mitogen-activated protein kinase kinase (MEK) inhibitor

AZD0530 (Saracatinib) Proto-oncogene tyrosine-protein Src inhibitor

Nutlin-3 Ubiquitin-protein ligase MDM2 inhibitor

Nilotinib BCR-ABL fusion protein inhibitor

17-AAG (Tanespamycin) Heat shock protein (Hsp90) inhibitor

PD-0332991 CDK4/6-Cyclin D inhibitor

PLX4720, Sorafenib RAF kinase inhibitors

Crizotinib, TAE684 ALK kinase inhibitors

Erlotinib, Lapatinib EGFR/HER2 kinase inhibitors

PHA-665752 Proto-oncogene c-MET kinase inhibitor


• Gene expression: ~12,000 genes were commonly assessed using Affymetrix HG-U133A and Plus2 chips

• Mutation: 68 genes were screened for mutations in both CGP and CCLE

• We used CGP data to train genomic predictors of drug response for the 15 drugs

• Gene expressions as input and IC50 as output

Genomic predictors of drug response

• We implemented five linear modeling approaches to build genomic predictors:• SINGLEGENE• RANKENSEMBLE• RANKMULTIV• MRMR• ELASTICNET


Validation framework


Genomic predictors of drug sensitivity (IC50)


CGP in 10-fold cross-validations



Trained on CGP, tested on CCLECommon cell lines



Trained on CGP, tested on CCLENew cell lines

• Given the poor performance of our predictors we decided to explore consistency between CGP and CCLE

• Different cell viability assays:• CGP: Cell Titer 96 Aqueous One Solution Cell (Promega)

amount of nucleic acids• CCLE: Cell Titer Glo luminescence assay (Promega)

metabolic activity via ATP generation

• Differences in experimental protocols including • range of drug concentrations tested• estimator for summarizing the drug dose-response curve

• Different technologies for measuring genomic profiles (gene expressions and mutations)


Consistency between CGP and CCLE

• Spearman correlation at different levels• Genomic data (gene expression)

• Drug sensitivity (IC50 and AUC)

• Gene-drug associations

Consistency measure


0 0.8 1

poor good

0.70.6

moderate substantial

Correlation

0.5

fair

• Cohen’s Kappa coefficient for mutations

Consistency of gene expression profiles

Benjamin Haibe-Kains 2013-09-20QBBMM Conference

Good correlation

Consistency of mutational profiles


Moderate agreement

Consistency of drug sensitivity (IC50)


Consistency of drug sensitivity (AUC)


Consistency of drug sensitivity


Poor

Fair

Moderate

• In 2010, GlaxoSmithKline tested• 19 compounds• on 311 cancer cell lines

• 194 cell lines in common with CGP and CCLE

• 2 drugs in common, Lapatinib and Paclitaxel

• CCLE and GSK used the same pharmacological assay (Cell Titer Glo luminescence assay, Promega)

GSK Cancer Cell Line Genomic Profiling Data


Comparison with GSK for Lapatinib


Comparison with GSK for Paclitaxel


Replicates in CGPSame assay, same protocol


Poor

Fair

Moderate

Significant gene-drug associationsFDR < 20%

Consistency of gene-drug associationsModel for gene-drug association:where Y = drug sensitivity

Gi = gene expression of gene i

T = tissue type


• To identify the most likely source of inconsistencies we intermixed the gene expressions and drug sensitivity measures between studies

• Original = [CGPg+CGPd] vs. [CCLEg+CCLEd]

• GeneCGP.fixed = [CGPg+CGPd] vs. [CGPg+CCLEd]

• GeneCCLE.fixed = [CCLEg+CGPd] vs. [CCLEg+CCLEd]

• DrugCGP.fixed = [CGPg+CGPd] vs. [CCLEg+ CGPd]

• DrugCCLE.fixed = [CGPg+CCLEd] vs. [CCLEg+CCLEd]

Source of inconsistencies


Source of inconsistencies


• Gene expressions used to be noisy but years of standardization enabled reproducible measurements

• Some more work needed to make variant calling more consistent but we will get there

• Drug phenotypes appear to be quite noisy though

• This prevents us to characterize drugs’ mechanism of action and to build robust genomic predictors of drug response

• Needs for standardization in terms of pharmacological assay and experimental protocol

• New protocols may be needed (combination of assays + more controls)

Take home messages


• Nehme Hachem• Rachad El-Badrawi• Simon Papillon-Cavanagh• Nicolas de Jay

• Jacques Archambault

Acknowledgements

• Hugo Aerts• John Quackenbush

• Andrew Beck• Andrew Jin• Nicolai Juul Birkbak

Thank you for your attention!

• Frank Emmert-Streib (Queen’s University, Ireland) and I are editing a Special Issue on Network Inference

• Your contributions are welcome!

One more thing …


Deadline: Sept 15

Appendix

• We implemented five linear models to build genomic predictors:

• SINGLEGENE: Univariate linear regression model with the gene the most correlated to sensitivity [-log10(IC50)]

• RANKENSEMBLE: Average of the predictions of the top 30 models

• RANKMULTIV: Multivariate model with the top 30 genes

• MRMR: Multivariate model with the 30 genes most correlated and less redundant

• ELASTICNET: Regularized multivariate model (L1/L2 penalization)

Modeling techniques



Consistency of gene expression profilesby tissue types

Consistency of drug sensitivityby tissue types


IC50AUC

Consistency of mutation-drug associationsModel for gene-drug association:where Y = drug sensitivity

Mi = presence of mutation in gene i

T = tissue type


Consistency of drug sensitivity calling


Drug sensitivity in CGP

IC50

AUC

Drug sensitivity in CCLE

IC50

AUC

IC50 in CGP and CCLE

AUC in CGP and CCLE

Benjamin Haibe-Kains Director, Bioinformatics and Computational Genomics Laboratory Scientific Advisor, Bioinformatics Core Facility Are pharmacogenomic.

Documents

benjamin haibekains

cgp ccle drugs

conventional drugs

drug compounds

drugs mechanisms of

cancer genome project

number of new targeted

cyclin d inhibitor plx4720