BSc Course: "Experimental design“ Genome-wide Association Studies Sven Bergmann Department of Medical Genetics University of Lausanne Rue de Bugnon 27 - DGM 328 CH-1005 Lausanne Switzerland work: cell:

Overview Population stratification Associations: Basics Whole genome associations Genotype imputation Uncertain genotypes New Methods

Overview Population stratification Associations: Basics Whole genome associations Genotype imputation Uncertain genotypes New Methods

6’189 individuals Phenotypes 159 measurement 144 questions Genotypes SNPs CoLaus = Cohort Lausanne Collaboration with: Vincent Mooser (GSK), Peter Vollenweider & Gerard Waeber (CHUV)

ATTGCAATCCGTGG...ATCGAGCCA…TACGATTGCACGCCG… ATTGCAAGCCGTGG...ATCTAGCCA…TACGATTGCAAGCCG… ATTGCAATCCGTGG...ATCGAGCCA…TACGATTGCACGCCG…ATTGCAAGCCGTGG...ATCTAGCCA…TACGATTGCAAGCCG… Genetic variation in SNPs (Single Nucleotide Polymorphisms)

Analysis of Genotypes only Principle Component Analysis reveals SNP-vectors explaining largest variation in the data

Example: 2PCs for 3d-data Raw data points: {a, …, z}

Example: 2PCs for 3d-data Normalized data points: zero mean (& unit std)!

Example: 2PCs for 3d-data Identification of axes with the most variance Most variance is along PCA1 The direction of most variance perpendicular to PCA1 defines PCA2

Ethnic groups cluster according to geographic distances PC1 PC2

PCA of POPRES cohort

Overview Population stratification Associations: Basics Whole genome associations Genotype imputation Uncertain genotypes New Methods

Phenotypic variation:

What is association? chromosomeSNPstrait variant Genetic variation yields phenotypic variation Population with ‘ ’ allele Distributions of “trait”

Quantifying Significance

T-test t-value (significance) can be translated into p-value (probability)

Association using regression genotypeCoded genotype phenotype

Regression analysis X Y “response” “feature(s)” “intercept” “coefficients” “residuals”

Regression formalism (monotonic) transformation phenotype (response variable) of individual i effect size (regression coefficient) coded genotype (feature) of individual i p(β=0) error (residual) Goal: Find effect size that explains best all (potentially transformed) phenotypes as a linear function of the genotypes and estimate the probability (p-value) for the data being consistent with the null hypothesis (i.e. no effect)

Overview Population stratification Associations: Basics Whole genome associations Genotype imputation Uncertain genotypes New Methods

Whole Genome Association

Current microarrays probe ~1M SNPs! Standard approach: Evaluate significance for association of each SNP independently: significance

Whole Genome Association significance Manhattan plot observed significance Expected significance Quantile-quantile plot Chromosome & position GWA screens include large number of statistical tests! Huge burden of correcting for multiple testing! Can detect only highly significant associations ( p < α / #(tests) ~ )

GWAS: >20 publications in 2006/2007 Massive!

Genome-wide meta-analysis for serum calcium identifies significantly associated SNPs near the calcium-sensing receptor (CASR) gene Karen Kapur, Toby Johnson, Noam D. Beckmann, Joban Sehmi, Toshiko Tanaka, Zolt á n Kutalik, Unnur Styrkarsdottir, Weihua Zhang, Diana Marek, Daniel F. Gudbjartsson, Yuri Milaneschi, Hilma Holm, Angelo DiIorio, Dawn Waterworth, Andrew Singleton, Unnur Steina Bjornsdottir, Gunnar Sigurdsson, Dena Hernandez, Ranil DeSilva, Paul Elliott, Gudmundur Eyjolfsson, Jack M Guralnik, James Scott, Unnur Thorsteinsdotti, Stefania Bandinelli, John Chambers, Kari Stefansson, G é rard Waeber, Luigi Ferrucci, Jaspal S Kooner, Vincent Mooser, Peter Vollenweider, Jacques S. Beckmann, Murielle Bochud, Sven Bergmann

Current insights from GWAS: Well-powered (meta-)studies with (ten-)thousands of samples have identified a few (dozen) candidate loci with highly significant associations Many of these associations have been replicated in independent studies

Current insights from GWAS: Each locus explains but a tiny (<1%) fraction of the phenotypic variance All significant loci together explain only a small (<10%) of the variance David Goldstein: “~93,000 SNPs would be required to explain 80% of the population variation in height.” Common Genetic Variation and Human Traits, NEJM 360;17

The “Missing variance” (Non-)Problem Why should a simplistic (additive) model using incomplete or approximate features possibly explain anything close to the genetic variance of a complex trait? … and it doesn ’ t have to as long as Genome-wide Association Studies are meant to as an undirected approach to elucidate new candidate loci that impact the trait!

1.Other variants like Copy Number Variations or epigenetics may play an important role 2.Interactions between genetic variants (GxG) or with the environment (GxE) 3.Many causal variants may be rare and/or poorly tagged by the measured SNPs 4.Many causal variants may have very small effect sizes 5.Overestimation of heritabilities from twin-studies? So what do we miss?

Overview Population stratification Associations: Basics Whole genome associations Genotype imputation Uncertain genotypes New Methods

Intensity of Allele G Intensity of Allele A Genotypes are called with varying uncertainty

Some Genotypes are missing at all …

… but are imputed with different uncertainties

… using Linkage Disequilibrium! Markers close together on chromosomes are often transmitted together, yielding a non-zero correlation between the alleles. Marker 123n LD D

Conclusion Genotypic markers are always measured or inferred with some degree of uncertainty Association methods should take into account this uncertainty

Two easy ways dealing with uncertain genotypes 1.Genotype Calling: Choose the most likely genotype and continue as if it is true (p 11 =10%, p 12 =20% p 22 =70% => G=2) 2.Mean genotype: Use the weighted average genotype (p 11 =10%, p 12 =20% p 22 =70% => G=1.6)

Overview Associations: Basics Whole genome associations Population stratification Genotype imputation Uncertain genotypes New Methods

1.Improve measurements: - measure more variants (e.g. by UHS) - measure other variants (e.g. CNVs) - measure “molecular phenotypes” 2.Improve models: - proper integration of uncertainties - include interactions - multi-layer models How could our models become more predictive?

Towards a layered Systems Model We need intermediate (molecular) phenotypes to better understand organismal phenotypes

Organisms Data types Conditions Developmental Physiological Environmental Experimental Clinical –Genotypic data –Gene expression data –Metabolomics data –Interaction data –Pathways The challenge of many datasets: How to integrate all the information?

Network Approaches for Integrative Association Analysis Using knowledge on physical gene-interactions or pathways to prioritize the search for functional interactions

Using expression data from Hapmap panel we computed all pairwise interactions for several expression values: The strongest interactions do not occur for SNPs with high (marginal) significance! Bioinformatics Advance Access published online on April 7, 2010 Bioinformatics, doi: /bioinformatics/btq147

Transcription Modules reduce Complexity SB, J Ihmels & N Barkai Physical Review E (2003) /ExpressionView

Association of (average) module expression is often stronger than for any of its constituent genes

Overview Associations: Basics Whole genome associations Population stratification Genotype imputation Uncertain genotypes New Methods

Analysis of genome-wide SNP data reveals that population structure mirrors geography Genome-wide association studies elucidate candidate loci for a multitude of traits, but have little predictive power so far Future improvement will require –better genotyping (CGH, UHS, …) –New analysis approaches (interactions, networks, data integration) Take-home Messages: