Analytical challenges in genetic association studies David Meyre, Associate Professor, McMaster University (meyred@mcmaster.ca) HRM 728 Graduate Course: Genetic Epidemiology – November, 7th 2014 Li & Meyre., Int J Obes 2013 The march of technology 1980 single variant (100 SNPs) detailed study of individual genes (102 SNPs) 1990 regional studies (104 SNPs) 2000 2006 genome-wide association (5 105 SNPs) 3,5 106 SNPs (2007) 2011 Whole-genome sequencing (3 107 SNPs) A storm of data to deal with! Analytical challenges in genetic association studies . 426 positive findings in 127 genes but…. . only 22 genes associated with obesity-related phenotypes in > 5 studies Replication is challenging in genetic epidemiology Skepticism in the medical / scientific community Rankinen et al., Obesity 2006 Analytical challenges in genetic association studies I. Analytical challenges to find a true association in a discovery study (risk of false positive result) II. Analytical challenges to replicate a true positive association III. Guidelines for proper discovery and replication association study designs Analytical challenges to find a true association in a discovery study Are you ready for the Episode 1 of the saga! Analytical challenges in genetic association studies I. Lack of replication may occur because the original study reports a false positive result 1) The phenotype is not heritable Obesity is an heritable disease . 2 obese parents 10-fold increased risk for childhood obesity . Obesity has a strong genetic component: heritability 50- 85% (Stunkard et al., NEJM 1986; Wardle et al., AJCN 2008) Analytical challenges in genetic association studies I. Lack of replication may occur because the original study reports a false positive result 1) The phenotype is not heritable 2) Insufficient sample size Statistical power and sample size Effect sizes for obesity-associated common genetic variants are small (OR < 2) Statistical power and sample size MAF in controls 0.01 0.05 0.1 0.2 0.3 0.4 1.1 443,854 92,868 49,252 27,974 21,518 19,010 1.2 116,354 24,434 13,018 7,460 5,792 5,162 1.3 54,110 11,404 6,102 3,526 2,760 2,480 1.5 21,208 4,498 2,426 1,424 1,132 1,032 2.0 6,386 1,374 754 458 376 354 Allelic OR Table 1. Sample sizes needed in a case control design to detect significant association with a power of 90% and a two-sided P-value of 0.001 by odds ratio and allele frequency for risk allele. Calculations assume multiplicative effect on disease risk. Sample sizes presented are total number of cases and controls needed, assuming an equal number of cases and controls. GAD2 or the importance of a well-powered study .Association of the GAD2 promoter gene variant -243 A>G with morbid obesity (OR=1.05-1.58, P=0.01) using 575 cases and 646 controls . No prior statistical power calculation in the princeps study .Lack of confirmation of the association of the GAD2 promoter gene variant -243 A>G with morbid obesity (OR=0.90-1.36, P=0.28) in a metaanalysis of 1,252 cases and 1,800 controls Boutin et al., PLOS Biol 2003, Swarbrick et al., PLOS Biol 2005 Statistical power and rare variant analysis “We identified six highly correlated SNPs that show strong and comparable associations with risk of type 2 diabetes, but further refinement of these associations will require large sample sizes (>100,000) or studies in ethnically diverse populations.“ Fawcett et al., Diabetes 2010 Analytical challenges in genetic association studies I. Lack of replication may occur because the original study reports a false positive result 1) The phenotype is not heritable 2) Insufficient sample size 3) Lack of correction for multiple testing Multiple testing in the post-GWAS area 1 million polymorphisms! Bonferroni correction: Pcorrected = 0.05 / 1,000,000 = 5 x 10-8 2 SNP gene x gene interactions: Pcorrected = 1x 10-13 Multiple testing in the whole-exome/genome sequencing area 30 million polymorphisms 20,000 genes Bonferroni correction SNPs: Pcorrected = 0.05 / 30,000,000 = 1 x 10-9 Bonferroni correction genes: Pcorrected = 0.05 / 20,000 = 2.5 x 10-6 INSIG2: a GWA false positive association Science April 2006 Science January 2007 INSIG2 rs7566605 variant is INSIG2: lack of association with obesity associated with obesity in 3 independent designs (N=22,381) (ORmeta-analysis=1.05-1.42, P =0.008), far from the threshold of significance after multiple testing correction (P=5 x 10-7) Analytical challenges in genetic association studies I. Lack of replication may occur because the original study reports a false positive result 1) The phenotype is not heritable 2) Insufficient sample size 3) Lack of correction for multiple testing 4) Geographical population substructure Lactase persistence and population substructure LCT rs4988235 T allele frequency in UK Davey-Smith et al., EJHG 2009 Rare variants and founder effects -common SNP associated with adiponectin level in Fillipinos by GWAS -exon resequencing identified a rare coding variant (R221S) in LD with the common SNP strongly associated with adiponectin level -the mutation is found exclusively in Fillipinos Croteau-Chonka et al., HMG 2012 Analytical challenges in genetic association studies I. Lack of replication may occur because the original study reports a false positive result 1) The phenotype is not heritable 2) Insufficient sample size 3) Lack of correction for multiple testing 4) Geographical population substructure 5) Technological biases, lack of quality control procedure INS VNTR and association with childhood obesity, a technological bias? . Association of the INS VNTR variant with childhood obesity . Lack of association of the INS VNTR variant with childhood obesity . Genotyping by RFLP, a highly subjective method (Peters et al., CCM 2003) . Genotyping by TaqMan, a highly reliable method . Family-based design to enable a highstandard quality control procedure Le Stunff et al., Nat Genet 2000, Bouatia-Naji et al., Obesity 2008 Next generation sequencing and false-positive mutations . 10% of mutations are technological artifacts in next generation sequencing . The rate of false positive mutations is higher in ‘old’ DNA libraries Use of pedigrees, confirmation of mutations by Sanger resequencing New methods (Rain Dance technology) Bonnefond et al., PLOS One 2012 Analytical challenges in genetic association studies I. Lack of replication may occur because the original study reports a false positive result 1) The phenotype is not heritable 2) Insufficient sample size 3) Lack of correction for multiple testing 4) Geographical population substructure 5) Technological biases, lack of quality control procedure 6) Inappropriate statistical analysis Association and adjustement for confounding factors . Association between FTO intron 1 SNP and type 2 diabetes (OR=1.09-1.23, P= 5x 10-8) if adjustment for sex and age . Lack of association between FTO intron 1 SNP and type 2 diabetes (OR=0.96-1.10, P= 0.44) if adjustment for sex, age and BMI FTO is an obesity gene Inappropriate adjustment (or lack of adjustment) can lead to wrong conclusions Frayling et al., Science 2007 Analytical challenges in genetic association studies I. Lack of replication may occur because the original study reports a false positive result 1) The phenotype is not heritable 2) Insufficient sample size 3) Lack of correction for multiple testing 4) Geographical population substructure 5) Technological biases, lack of quality control procedure 6) Inappropriate statistical analysis Analytical challenges to replicate a true positive association Now the Episode 2 of the saga! Analytical challenges in genetic association studies II. Replication may be challenging even when the original result is a true positive association 1) Willingness to replicate the original study Lactase persistence and BMI variation Despite a convincing initial evidence of association between the LCT rs4988235 T variant and BMI (P=8 x 10-5) in 31,720 European individuals… Kettunen et al., HMG 2009 Lactase persistence and BMI variation Replication studies showed-up after 2-4 years… Correla et al., Obesity 2011 Analytical challenges in genetic association studies II. Replication may be challenging even when the original result is a true positive association 1) Willingness to replicate the original study 2) Winner’s curse effect and sample size in follow-up studies Obesity loci from GIANT and replication . Due to the small effect size of the SNPs on BMI variation, only a fraction of these associations replicates for obvious statistical power concerns (den Hoed et al., Diabetes 2010) Analytical challenges in genetic association studies II. Replication may be challenging even when the original result is a true positive association 1) Willingness to replicate the original study 2) Winner’s curse effect and sample size in follow-up studies 3) Gene x gene, gene x environment interactions Interactions between FTO SNP and physical activity .The effect of the rs9939609 SNP on obesity risk is decreased by 27% in physically active adults . No genotype x physical activity interaction on obesity risk in children Kilpelainen et al., PLOS Med 2012 Savage et al., Nat Genet 2002 Analytical challenges in genetic association studies II. Replication may be challenging even when the original result is a true positive association 1) Willingness to replicate the original study 2) Winner’s curse effect and sample size in follow-up studies 3) Gene x gene, gene x environment interactions 4) Heterogeneity (ethnic heterogeneity, phenotype heterogeneity) Ethnicity and linkage disequilibrium blocs SNP1 SNP2 SNP3 SNP4 SNP5 Icelandic French Asian African Distance (Kb) Causal SNP Proxy SNP Disease-associated LD block Ethnicity and SNP allele frequency . Intronic variation (rs2237892) in a new locus (KCNQ1) was strongly associated with T2D in Asian (OR: 1.26-1.42, 10-40< P-value < 10-12) . The association with T2D was nominally replicated in European descent populations (DIAGRAM: P=0.01), with similar OR but lower risk allele frequency (5-7% in European, 28-40% in Asian) Obesity, waist and BMI have a partially overlapping genetic architecture Analytical challenges in genetic association studies II. Replication may be challenging even when the original result is a true positive association 1) Willingness to replicate the original study 2) Winner’s curse effect and sample size in follow-up studies 3) Gene x gene, gene x environment interactions 4) Heterogeneity (ethnic heterogeneity, phenotyp heterogeneity) 5) Inheritance model (parent of origin effects, de novo mutations…) Analytical challenges in genetic association studies II. Replication may be challenging even when the original result is a true positive association 1) Willingness to replicate the original study 2) Winner’s curse effect and sample size in follow-up studies 3) Gene x gene, gene x environment interactions 4) Heterogeneity (ethnic heterogeneity, phenotyp heterogeneity) 5) Inheritance model 6) Subjective interpretation of data Subjective interpretation of data Is this glass half-full or half-empty? Analytical challenges in genetic association studies II. Replication may be challenging even when the original result is a true positive association 1) Willingness to replicate the original study 2) Winner’s curse effect and sample size in follow-up studies 3) Gene x gene, gene x environment interactions 4) Heterogeneity (ethnic heterogeneity, phenotyp heterogeneity) 5) Inheritance model 6) Subjective interpretation of data Guidelines for proper discovery and replication association study designs Enough time for the Episode 3 of the saga? Analytical challenges in genetic association studies III. Guidelines for proper discovery and replication association study designs Discovery 1) Study designs Gene discovery study designs General population N Lean Obese 1) Case control studies from extremes of the BMI tails 2) Quantitative trait studies in the whole population Correlation genotype / trait at a genetic locus Best approach (GIANT / GIANT extreme): BMI study in the whole population + analysis of the extremes of the BMI tails (genetic variance, effect size…) Berndt et al., Nat Genet 2013 Gene discovery study designs 3) Family-based association studies: allele transmission from parents to affected offsprings (imprinting, haplotypes….) 4) Cohort studies: correlation of a genotype with an incident disease event (gold standard) Gene discovery study designs General population N Lean Obese Normal weight 5) The case control case design: discovery of gene variants associated with leanness or with obesity (applications in drug design) The gain-of-function V103I and I251L variants in MC4R are associated with leanness French adults French children Italian children 16 cohorts: 5964 control and 6370 obese patients Swiss adults Ohshiro et al, 1999 Farooqi et al, 2000 Study Reference Jacobson et al, 2002 Jacobson et al, 2002 Miraglia del Giudice et al, 2002 Hinney et al, 2003 Marti et al, 2003 Valli-Jaakola et al, 2004 OR = 0.53, p-value = 4.26.10-5 Santini et al, 2004 Buono et al, 2005 Larsen et al, 2005 Summary 0.03 0.10 0.32 1.00 3.16 10.00 31.62 exp(Effect) -Meta-analysis in 39,879 subjects confirms an obesity-protective role of the V103I polymorphism (OR = 0.80; p-value = 0.002) -V103I et I251L are infrequent (0.41-2.24%) and induce a gain of function effect on the melanocorin 4 receptor (Xiang et al., Biochemistry 2006) Stutzmann et al., HMG 2007 Gene discovery study designs 6) Clinical trials, interventional studies: correlation of a genotype with response to intervention or treatment (lifestyle intervention, drug, surgery, smoking cessation, antipsychotic drug administration….) Analytical challenges in genetic association studies III. Guidelines for proper discovery and replication association study designs Discovery 1) Study designs 2) Phenotype How to chose a relevant obesity phenotype? Heritability for BMI: Heritability for type 2 diabetes: -h² = 0.48 at age 4 y. -h² = 0.69 (onset < 60 y.) -h² = 0.78 at age 11 y. -h² = 0.31 (onset < 75 y.) Haworth et al., Obesity 2008, Almgren et al., Diabetologia 2011 How to chose a relevant obesity phenotype? -clinically and biologically relevant -easy and inexpensive to measure -relevant in diverse ethnicities -minimal measurement error -minimal misclassification and reporting biases value of BMI to estimate the degree of adiposity questionable body fat content, body adiposity index are more relevant . Genome-wide association study for % fat mass in 36,000 subjects, replication of the best hits in 39,000 subjects . Three % fat mass-associated loci : FTO, IRS1, SPRY2 . Only one locus (FTO) out of three has been conclusively associated with BMI body mass index in literature Kilpelainen et al., Nat Genet 2011 Analytical challenges in genetic association studies III. Guidelines for proper discovery and replication association study designs Discovery 1) Study designs 2) Phenotype 3) Gene identification strategies Gene identification strategies AGNOSTIC APPROACH CANDIDATE GENE APPROACH -moderately successful -highly successful -novel disease causing mechanisms -previously known mechanisms -significance thresholds -strong selection criteria needed -lack of biological relevance -biological relevance HIGH-THROUGHPOUT CANDIDATE GENE APPROACH (pathway, expression, evolution…) Analytical challenges in genetic association studies III. Guidelines for proper discovery and replication association study designs Discovery 1) Study designs 2) Phenotype 3) Gene identification strategies 4) Genotyping methodology and quality control procedures Genotyping methodology and quality control -exclusion of low quality DNA (cases controls) -highly reliable genotyping technology -genotyping call rate (> 95%)\ -Hardy-Weinberg equilibrium (P > 0.005)\ -double genotyping concordance rate (> 99%) -MAF comparison in public databases -confirmation by a second method -association of SNPs in linkage disequilibrium -accurate experiments / data management and reporting (bar coding, automated processes, internal controls, flow charts….) -sex inconsistencies, hidden relatedness, ethnic outliers…. Analytical challenges in genetic association studies III. Guidelines for proper discovery and replication association study designs Discovery 1) Study designs 2) Phenotype 3) Gene identification strategies 4) Genotyping methodology and quality control procedures 5) Statistical analysis Statistical analysis -power calculation -limited number of hypotheses tested -multiple testing (FDR, Bonferroni…) -adjustment for confounding factors -caution with subgroup analyses -best fitting inheritance model -conditional analyses Analytical challenges in genetic association studies III. Guidelines for proper discovery and replication association study designs Discovery 1) Study designs 2) Phenotype 3) Gene identification strategies 4) Genotyping methodology and quality control procedures 5) Statistical analysis 6) Population stratification Population stratification -correction for self-reported ethnicity -exclusion of ethnic outliers -genomic control (Ancestry Informative Markers) -family-based association tests -case control matched for age, sex, geography… Analytical challenges in genetic association studies III. Guidelines for proper discovery and replication association study designs Replication 1) Systematic replication and reporting of promising associations Analytical challenges in genetic association studies III. Guidelines for proper discovery and replication association study designs Replication 1) Systematic replication and reporting of promising associations 2) Statistical power (Winner’s curse effect) Analytical challenges in genetic association studies III. Guidelines for proper discovery and replication association study designs Replication 1) Systematic replication and reporting of promising associations 2) Statistical power 3) Heterogeneity How to lower heterogeneity in replication studies? -same ethnicity / country -same study design -same ascertainment criteria -same phenotype -same genetic markers -same age window, same sex ratio -same inheritance model -same statistical analysis -same covariate adjustments Analytical challenges in genetic association studies III. Guidelines for proper discovery and replication association study designs Replication 1) Systematic replication and reporting of promising associations 2) Statistical power 3) Heterogeneity 4) Meta-analyses Analytical challenges in genetic association studies III. Guidelines for proper discovery and replication association study designs Replication 1) Systematic replication and reporting of promising associations 2) Statistical power 3) Heterogeneity 4) Meta-analyses 5) Additional studies Additional studies -worldwide contribution -extension to different study designs, ascertainment criteria -association with obesity endophenotypes -gene x environment interactions -fine-mapping, causative gene variants -functional experiments -biological insights FTO in 2007: ‘gene of unknown function in an unknown pathway’ 2014: > 740 articles published 1997: first identification of a monogenic obesity gene (LEP) 2007: first gene variant in FTO conclusively associated with obesity 2012: 40 monogenic (syndromic / non-syndromic) obesity genes, > 100 common gene variants conclusively associated with polygenic obesity ANY QUESTIONS? The French fair-play!