Microarray analysis as a prognostic and predictive tool: are we ready? Enzo Medico Laboratory of Functional Oncogenomics Institute for Cancer Research and Treatment University of Torino
Topics Platforms for gene expression profiling Breast cancer signatures From cell-based models to cancer classifiers
AFFYMETRIX GeneChip
45,000 gènes ! AFFYMETRIX GeneChip
The Probe Sets
Hybridization on the chip
Signal detection
Genomic raw data
Gene expression profiling by spotted microarrays Oligonucleotides or cDNAs Robotic printing
Gene expression profiling by spotted/dual colour microarrays RNA extraction, cDNA labelling Hybridization “Reference” RNA sample (pool) “Test” sample (tumour specimen)
Different platforms generate different data types Two-colourOne-colour Paired samplesIndependent samples Ref vs Sample – 1 Ref vs Sample - 2 Ref – 1 Ref - 2 Sample – 1 Sample - 2 vs
Topics Platforms for gene expression profiling Breast cancer signatures From cell-based models to cancer classifiers
61 years IDC Postmenopausal N - pT = 0.9 cm Grade 2 ER et PgR - HER2 - FA(E)C x 6 % Choices of 40 experts worldwide AUCUNCMFx6ACx4TAMAUTRE SHOULD ONE TREAT A SMALL (<1CM) ENDOCRINE UNRESPONSIVE TUMOR ?
WHO CAN BE SPARED THERAPY? WHICH THERAPY WILL WORK BEST? Prognostic factors neededPredictive factors needed THERAPY DECISION-MAKING FOR EARLY BREAST CANCER
Clinical Outcome ER- ER+ PNAS vol 98, no 19, , 2001 The “Intrinsic” Breast Cancer Signatures
PNAS vol 100, no 18, , 2003 ER- ER+ grade Clinical Outcome Confirmatory Study
Amsterdam’s Signature 312 patients 70 genes Rotterdam’s Signature 286 patients 76 genes Discovery of «poor prognosis signatures» for distant relapses
78 tumor samples 70-gene poor prognosis signature
Marc J Van de Vijver et al., NEJM, 347, 25, 2002 High RiskLow Risk 70-gene expression signature outperforms clinicopathological criteria
Lancet, 2005, 365, tumor samples
G3 G2 G1 Histologic Grade GG1 GG2 GG3 Genomic Grade Sotiriou et al., JNCI 2006 Poor inter observer reproducibility G2: difficult treatment decision making, under- or over-treatment likely Findings consistent across multiple data sets and microarray platforms More objective assessment Easier treatment decision-making High proportion of genes involved in cell proliferation !
Identify genes correlated with grade 1 vs grade 3 Grade 1Grade 3 Grade 1Grade 2Grade 3 Analyze on validation set (n = 125) Definition and validation of the Genomic Grade
Sorlie et al. PNAS 2001 Sotiriou et al. PNAS 2003 Van de Vijver et al. NEJM 2002 Central Pathology Review! Consistent Distribution of GG in Different Populations and Microarrays Platforms
GENE EXPRESSION SIGNATURE = POWERFUL PROGNOSTIC TOOL Highest priority = Transfer from bench to bedside HOW ?
Validation study… TRANSLATING MOLECULAR KNOWLEDGE INTO EARLY BREAST CANCER MANAGEMENT
THERAPY DECISION-MAKING FOR EARLY BREAST CANCER WHO CAN BE SPARED THERAPY? WHICH THERAPY WILL WORK BEST? Prognostic factors needed Predictive factors needed
Topics Platforms for gene expression profiling Breast cancer signatures From cell-based models to cancer classifiers
The Invasive Growth biological program Proliferation Differentiation, cell polarity, tubulogenesis Scattering and migration Survival and protection against apoptosis
MET SHH SMO PTCH1 GLI AFP CK19 ALB -AT TO MLP-29 liver stem/progenitor cells activate the invasive growth program in response to HGF Hedgehog pathway Liver lineage Liver differentiation -3+3 MLP29 / liver log2 ratio CTRL HGF 6h HGF 16h Day 1 Day 2 Day 4
Induced at 1h HGF/CTRL 1h 6h 24h EGF/CTRL 1h 6h 24h Induced at 6h Induced at 24h Suppressed at 24h HGF/CTRL 1h 6h 24h EGF/CTRL 1h 6h 24h Suppr. at 1h Suppressed at 6h Suppressed at 24h HGF/CTRL 1h 6h 24h EGF/CTRL 1h 6h 24h The Invasive Growth Transcriptional Program
Total NKI Breast cancer Dataset (311 samples - Agilent) Rotterdam Breast cancer Dataset (286 samples - Affymetrix) Statistical analysis Kaplan-Meier COX proportional hazard IG genes ranked by their individual performance (SNR over 1000 bootstraps) Number of genes in the classifier optimized and definition of the nearest mean classifier (NMC) Classifier construction and in silico validation using breast cancer microarray datasets
The Nearest Mean Classifier AVG Z Z3Z2Z1 AVG Y Y3Y2Y1 AVG X X3X2X1 Class A Sample 3 Sample 2 Sample 1 AVG X X6X5X4 AVG Z Z6Z5Z4 AVG Y Y6Y5Y4 Class B Sample 6 Sample 5 Sample 4 Gene Z Gene Y Gene X Training Group ATraining Group B
The Nearest Mean Classifier AVG Z Z3Z2Z1 AVG Y Y3Y2Y1 AVG X X3X2X1 Good Progn Class Sampl e 3 Sampl e 2 Sampl e 1 AVG X X6X5X4 AVG Z Z6Z5Z4 AVG Y Y6Y5Y4 Poor Prog Class Sampl e 6 Sampl e 5 Sampl e 4 Gene Z Gene Y Gene X Zs Ys Xs Test sample Gen eZ Gen eY Gen eX Group AGroup B Pearson correlation -> classification
Invasive growth genes classify breast cancer samples by their metastatic propensity
Validation on the Rotterdam dataset (286 breast samples, Wang et all., Lancet, 2005) Cox’s proportional hazards model IG 49 genes Cumulative Survival Time to relapse or last follow-up (months) NKI 49 genes Cumulative Survival Time to relapse or last follow-up (months) Poor prognosis Good prognosis Legend: 0 = Good prognosis samples 1 = Poor prognosis samples
Breast cancer expression profiling: towards an integrated approach to personalized therapy
Acknowledgments IRCC Laboratory of Functional Oncogenomics Tommaso Renzulli Claudio Isella Daniela Cantarella Barbara Martinoglio Roberta Porporato IRCC Gynaecological Oncology Daniela Cimino Luca Fuso Prof. Michele De Bortoli Prof. Piero Sismondi