Pitfalls in Genetic Association Studies

Pitfalls in Genetic Association Studies M. Tevfik DORAK Paediatric and Lifecourse Epidemiology Research Group Sir James Spence Institute Newcastle University, U.K. Clinical Studies & Objective Medicine Bodrum, 15-16 April 2006 Incident or prevalent cases Comparable controls or convenience samples Population based? Confounding by ethnicity / Population stratification Confounding by locus / Linkage disequilibrium Statistical power Multiple comparisons No adjustment for known associations Effect modification by sex Different genetic models Overinterpretation LD ruled out? Biological plausibility (a priori hypothesis)? Replication / Consistency Publication bias Tabor, Risch & Myers. Nat Rev Genet 2002 (www) Internal Validity of an Association Study Avoid "BIAS" Control for "CONFOUNDING" Rule out "CHANCE" Internal Validity of an Association Study Avoid "BIAS" Be careful! Control for "CONFOUNDING" Matching, Stratification, MV Analysis Rule out "CHANCE" Large sample, Replication Special Kinds of Confounding in Genetic Epidemiology CONFOUNDING by Locus (LD) MV Analysis, LD Analysis CONFOUNDING by Ethnicity (Population Stratification) Matching, Stratification, MV Analysis Family-Based Association Studies Genomic Controls and Specially Designed GE Analysis Special Kinds of Confounding in Genetic Epidemiology CONFOUNDING by Locus (LD) MV Analysis, LD Analysis CONFOUNDING by Ethnicity (Population Stratification) Matching, Stratification, MV Analysis Family-Based Association Studies Genomic Controls and Specially Designed GE Analysis Mapping Disease Susceptibility Genes by Association Studies Plot of minus log of P value for case-control test for allelic association with AD, for SNPs immediately surrounding APOE (<100 kb) Martin, 2000 (www) Example: Linkage Disequilibrium HLA-B47 association with congenital adrenal hyperplasia (Dupont et al, Lancet 1977) HLA-B14 association with late-onset adrenal hyperplasia (Pollack et al, Am J Hum Genet 1981) Is congenital adrenal hyperplasia an immune system-mediated disease? Example: Linkage Disequilibrium HLA-B47 association with congenital adrenal hyperplasia is due to deletion of CYP21A2 on HLA-B47DR7 haplotype HLA-B14 association with late-onset adrenal hyperplasia is due to an exon 7 missense mutation (V281L) in CYP21A2 on HLA-B14DR1 haplotype Preliminary evidence of an association between HLADPB1*0201 and childhood common ALL supports an infectious aetiology Leukemia 1995;9(3):440-3 Evidence that an HLA-DQA1-DQB1 haplotype influences susceptibility to childhood common ALL in boys provides further support for an infection-related aetiology Br J Cancer 1998;78(5):561-5 Why not LD? Publication Bias Negative studies do not get published NIH Genetic Associations Database ? A different kind of publication bias? Preliminary evidence of an association between HLADPB1*0201 and childhood common ALL supports an infectious aetiology Leukemia 1995;9(3):440-3 Evidence that an HLA-DQA1-DQB1 haplotype influences susceptibility to childhood common ALL in boys provides further support for an infection-related aetiology Br J Cancer 1998;78(5):561-5 Special Kinds of Confounding in Genetic Epidemiology CONFOUNDING by Locus (LD) MV Analysis, LD Analysis CONFOUNDING by Ethnicity (Population Stratification) Matching, Stratification, MV Analysis Family-Based Association Studies Genomic Controls and Specially Designed GE Analysis Wacholder, 2002 (www) Population Stratification Marchini, 2004 (www) Cardon & Palmer, 2003 (www) Internal Validity of an Association Study Avoid "BIAS" Be careful! Control for "CONFOUNDING" Matching, Stratification, MV Analysis Rule out "CHANCE" Large sample, Replication Example: Functional Correlation Example: Replication HFE-C282Y Association in Childhood ALL 30 % 25 20 15 10 5 0 Patients Controls % 50 45 40 35 30 25 20 15 10 5 0 Patients Controls All patients Males Only Males (cALL) All patients Males Only Males (cALL) WELSH GROUP 117 patients - 415 newborns P = 0.005; OR = 2.8 (1.4 to 5.4) In cALL: P = 0.02; OR = 2.9 (1.4 to 6.4) SCOTTISH GROUP 135 patients - 238 newborns P = 0.0004; OR = 3.0 (1.7 to 5.4) In cALL: P < 0.0001; OR = 4.7 (2.5 to 8.9) Dorak et al, Blood 1999 Palmer LJ. Webcast (www) Multiple Comparisons & Spurious Associations Diepstra, Lancet 2005 (www) Genetic Models and Case-Control Association Data Analysis The data may also be analysed assuming a prespecified genetic model. For example, with the hypothesis that carrying allele B increased risk of disease (dominant model), the AB and BB genotypes are pooled giving a 2x3x2 table. This is particularly relevant when allele B is rare, with few BB observations in cases and controls. Alternatively, under a recessive model for allele B, cells AA and AB would be pooled. Analysing by alleles provides an alternative perspective for case control data. This breaks down genotypes to compare the total number of A and B alleles in cases and controls, regardless of the genotypes from which these alleles are constructed. This analysis is counter-intuitive, since alleles do not act independently, but it provides the most powerful method of testing under a multiplicative genetic model, where risk of developing a disease increases by a factor r for each B allele carried: risk r for genotype AB and r2 for genotype BB. If a multiplicative genetic model is appropriate, both case and control genotypes will be in Hardy–Weinberg equilibrium, and this can be tested for. A fourth possible genetic model is additive, with an increased disease risk of r for AB genotypes, and 2r for BB genotypes. This model shows a clear trend of an increased number of AB and BB genotypes, with the risk for AB genotypes approximately half that for BB genotypes. The additive genetic model can be tested for using Armitage’s test for trend. Lewis CM. Brief Bioinform 2002 (www) HLA-DRB4 Association in Childhood ALL 40 25 % 30 % Patients Controls * Patients 20 Controls Boys, n=64 20 15 Girls, n=53 10 10 5 * 0 DRB5 DRB3 DRB4 0 DRB5 DRB3 DRB4 Homozygosity for HLA-DRB4 family is associated with susceptibility to childhood ALL in boys only (P < 0.0001, OR = 6.1, 95% CI = 2.9 to 12.6 ) Controls are an unselected group of local newborns (201 boys & 214 girls) * Case-only analysis P = 0.002 (OR = 5.6; 95% CI = 1.8 to 17.6) This association extends to a DRB4-HSP70 haplotype (OR = 8.3; 95% CI = 3.0 to 22.9) This association has been replicated in Scotland and Turkey HLA-DRB4 ASSOCIATION ADDITIVE MODEL Linear Model Logit estimates Number of obs = 265 LR chi2(1) = 14.24 Prob > chi2 = 0.0002 Log likelihood = -139.37794 Pseudo R2 = 0.0486 -----------------------------------------------------------------------------caco | Common Odds Ratio Std. Err. z P>|z| [95% CI] -------------+---------------------------------------------------------------drb4add | 2.208651 .4734163 3.70 0.000 1.45103 - 3.36186 ------------------------------------------------------------------------------ Heterozygosity and Homozygosity Logit estimates Number of obs = 265 LR chi2(2) = 22.00 Prob > chi2 = 0.0000 Log likelihood = -135.49623 Pseudo R2 = 0.0751 -----------------------------------------------------------------------------caco | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------Wild-type | 1.00 (ref) Heterozygosity | 1.060652 .3557426 0.18 0.861 .549642 2.04676 Homozygosity | 6.258503 2.65464 4.32 0.000 2.72534 14.37211 ------------------------------------------------------------------------------ HLA-DRB4 - HSPA1B HAPLOTYPE ASSOCIATION EFFECT MODIFICATION Number of obs = 532 LR chi2(3) = 23.97 Prob > chi2 = 0.0000 Log likelihood = -268.27826 Pseudo R2 = 0.0428 -----------------------------------------------------------------------------caco | Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------sex | -.0299037 .2229554 -0.13 0.893 -.4668883 .4070808 hsp53 | 2.530033 .5929603 4.27 0.000 1.367852 3.692214 _IsexXhsp5~2 | -2.758189 .8812645 -3.13 0.002 -4.485436 -1.030943 _cons | -1.321474 .3517969 -3.76 0.000 -2.010984 -.6319651 -----------------------------------------------------------------------------Logit estimates CONFOUNDING BY SEX Number of obs = 532 LR chi2(2) = 11.99 Prob > chi2 = 0.0025 Log likelihood = -274.26995 Pseudo R2 = 0.0214 -----------------------------------------------------------------------------caco | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------hsp53 | 3.32777 1.191429 3.36 0.001 1.649684 6.712832 sex | .7693041 .1636106 -1.23 0.218 .5070725 1.167148 -----------------------------------------------------------------------------Adjusted for sex? Logit estimates HLA-DRB4 - HSPA1B HAPLOTYPE ASSOCIATION BOYS ONLY Number of obs = 265 LR chi2(1) = 22.41 Prob > chi2 = 0.0000 Log likelihood = -135.29119 Pseudo R2 = 0.0765 -----------------------------------------------------------------------------caco | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------hsp53 | 12.55392 7.444028 4.27 0.000 3.926876 40.13392 -----------------------------------------------------------------------------Logit estimates GIRLS ONLY Logit estimates Number of obs = 267 LR chi2(1) = 0.13 Prob > chi2 = 0.7205 Log likelihood = -132.98706 Pseudo R2 = 0.0005 -----------------------------------------------------------------------------caco | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------hsp53 | 0.796 .5189439 -0.35 0.726 .22181 2.856571 ------------------------------------------------------------------------------ The association is modified by sex Cardon & Bell. Nat Rev Genet 2001 (www) Current Criteria for Good Association Studies Cardon & Bell. Nat Rev Genet 2001 (www) Statistical checklist for genetic association studies - In a case-control study: Cases and controls derive from the same study base There are more controls than cases (up to 5-to-1, for increased statistical power) There are at least 100 cases and 100 controls - Statistical power calculations are presented - Hardy-Weinberg equilibrium (HWE) is checked and appropriate tests are used - If HWE is violated, allelic association tests are not used - Possible genotyping errors and counter-measures are discussed - All statistical tests are two-tailed - Alternative genetic models of association considered - The choice of marker/allele/genotype frequency (for comparisons) is justified - For HLA associations, a global test for association (G-test, RxC exact test) for each locus is used (if necessary, with correction for multiple testing) - Chi-squared and Fisher tests are NOT used interchangeably - P values are presented without spurious accuracy (with two decimal places) - Strength of association has been measured (usually odds ratio and its 95% CI) - In a retrospective case-control study, ORs are presented (as opposed to RRs) - Multiple comparisons issue is handled appropriately (this does not necessarily mean Bonferroni corrections) - Alternative explanations for the observed associations (chance, bias, confounding) are discussed http://www.dorak.info/hla/stat.html Multifactorial Etiology ROCHE Genetic Education (www) Models of gene–environment interactions Hunter, 2005 (www) Generating Protein Diversity from the 'Small' Genome Banks, 2000 (www) Generating Protein Diversity from the 'Small' Genome Alternative Splicing Can Generate Very Large Numbers of Related Proteins From a Single Gene Most extreme example is the Drosophila Dscam Gene: 12 x 48 x 33 x 2 = 38,016 alternative splice variants Black, Cell 2000 (www) Wojtowicz, Cell 2004 (www) DSCAM = Down syndrome cell adhesion molecule Generating Protein Diversity from the 'Small' Genome Alternative Splicing Can be Tissue or Cell-Specific Lodish et al. Molecular Cell Biology, 5th Ed, WH Freeman (www) Generating Protein Diversity from the 'Small' Genome mRNA editing (base modification) is a different mechanism of alternative splicing Lodish et al. Molecular Cell Biology, 5th Ed, WH Freeman (www) OMIM 107730 (www) Chen & Chan, 1996 (www) Wedekind, 2003 (www) RNA Editing in The Cell – NCBI Online (www) Integration of proteomics in genetic epidemiology studies would eliminate a lot of obstacles arising from the following Only <2% of the genome is protein-coding and most sequence variants are silent changes Even genome-wide sequence variant studies cannot identify the genomic counterparts of 1.5 million proteins Epigenetic changes, alternative transcription/splicing and posttranslational modifications cannot be predicted by study of sequence variants Genomic DNA studies does not take into account selective expression of genes in certain cell or tissues No genetic association is complete without demonstration of the functional relevance 50s Rule Genome - 1:50 coding Gene:Protein - 1:50 ratio Overall efficiency of pure genomic studies 1:2500 http://www.dorak.info