Title: Genome-wide association studies
1Genome-wide association studies
- Misha Kapushesky
- Slides Johan Rung, EBI
- St. Petersburg Russia 2010
2Overview
- Methods for genome-wide association studies
- Montreal GWAS for Type 2 Diabetes
- GWAS results - context and caveats
3Study coverage
- Associating phenotype/disease state to genetic
variation - Cost per genotype has decreased
- Instead of a candidate gene approach, just scan
the entire genome - SNP microarrays covering up to 5M SNPs on one
chip - Increased sample sizes
4Recombination
5Linkage disequilibrium
Two markers on the genome are inherited together
more often than would be expected by chance This
leads to high correlation between nearby markers
in its haplotype block
6Haplotypes and genotype tagging
7Association studies
- Linkage disequilibrium enables association
studies, because of detection by proxy - not
every variant need to be typed
8Study power
9Study power
- The power of a study is to correctly predict a
true positive - To calculate this, you need
- risk model
- genotype relative risk
- allele frequency
- number of cases and controls
- population penetrance
- Acceptable rate of false positives
10How many SNPs should be tested? Studies of small
regions revealed linkage disequilibrium blocks in
which common SNPs are highly correlated (usually
lt10,00030,000 base pairs in African populations
or 30,00050,000 base pairs in the newer European
and Asian populations) (22). This motivated the
HapMap Project (www.hapmap.org 12), which has
validated approximately 4 million SNPs, including
2.8 million of the estimated 10 million common
SNPs in major world populations, while creating
competition among biotechnology companies to
develop high-throughput genotyping technologies.
Sequencing and genotyping studies showed that
sets of 500,000 (European populations) to
1,000,000 (African populations) SNPs could "tag"
(serve as proxies for) approximately 80 of
common SNPs (23).
11Quality controls
- Call rates for samples and SNPs
- Exclusion of low frequency SNPs
- Exclusion of SNPs out of Hardy-Weinberg
Equilibrium - Clean (or take into account) population
stratification
12Hardy-Weinberg Equilibrium
- If the alleles A and B have frequencies p and q,
you would expect the following genotype
frequencies - AA p2
- AB 2pq
- BB q2
13Hardy-Weinberg Equilibrium
- When observed genotype frequencies deviate from
the ones expected under HWE, this is indicative
of - population stratification
- different mutation rates between males and
females - different fitness between alleles
- genotype calling problems
- true association at the locus
14Binary or real-valued phenotypes
- Binary traits are typically disease state labels
(case or control) - Real-valued traits are quantitatively measured
phenotypes - blood sugar
- lipids
- height
- BMI
- gene expression
15Molecular vs disease phenotypes
- Disease phenotypes are the result of combinations
of molecular phenotypes in the body - Progression with time
- Precision of phenotype measurement
16Molecular vs disease phenotypes
- Many physiological phenotypes involved in
disease dynamics
17Molecular vs disease phenotypes
Molecular phenotypes can give more precise
information about disease state
18Association statistics
- Association statistics for binary traits are most
often based on a c2-statistic, based on the
genotype count table, or a logistic regression
model - c2-statistic summarizes independence between
disease state and genotype
19Association statistics
- For aa in cases, you would expect
- The sum of the squares of the differences is
c2-distributed
aa aA AA Sum
Cases r0 r1 r2 R
Controls s0 s1 s2 S
Count n0 n1 n2 N
20Regression
- For real-valued phenotypes, use linear regression
- For binary phenotypes, use logistic regression
21Population stratification
- Population stratification occurs when groups or
subpopulations within your sample are more
related than would be expected by random - This introduces correlations and inflates
association p-values and need to be corrected for
22Genomic control
23Eigenstrat
24Imputation
- Using a reference population (like HapMap or 1000
genomes) we can infer the genotype of SNPs that
were not tested - IMPUTE or MACH commonly used
- Yields probabilistic genotypes that need special
treatment
25Imputation
Wu et al, Nat. Genet. 41, 991-995, 2009
26Montreal GWAS
27Type 2 diabetes
- Blood glucose levels are regulated by insulin
release - Increased blood glucose levels triggers release
of insulin, that signals to the cells in muscle
for glucose intake - Through b-cell dysfunction or insulin resistance,
insulin regulation is impaired, leading to
increased glucose levels and eventually type 2
diabetes
28Type 2 diabetes
29Genetics of type 2 diabetes
- Before GWAS, T2D genetics was studied with
linkage studies and candidate gene approaches - Results in particular for MODY variants, caused
by disruptions of single genes - Genome-wide association studies and SNP arrays
made it possible to study complex diseases - Five large GWAS for T2D in 2007
- DIAGRAM meta-analysis in 2008
30Montreal GWAS
- Part of a larger T2D project at McGill and Genome
Quebec - After initial planning for candidate gene
genotyping, we switched to a GWAS strategy
31Multi-stage GWAS
- Two main strategies for increasing study power
- Meta-analyses increase effective sample size by
combining results from different studies - Multi-stage approaches scan the whole genome with
relatively low power, followed by focusing in on
the hits with higher power - Maximizing power in a single study in a
cost-effective way
32Multi-stage GWAS
33Study design
Fasting glucose Normoglycemic individuals Stage
1 French (N654) Stage 2 rs560887
(N9,353) Previously published, Science, May 2007
Fasting glucose Normoglycemic individuals Stage
1 French (N654) Stage 2 rs560887
(N9,353) Previously published, Science, May 2007
Fasting glucose Normoglycemic individuals Stage
1 French (N654) Stage 2 rs560887
(N9,353) Previously published, Science, May 2007
Fasting glucose Normoglycemic individuals Stage
1 French (N654) Stage 2 rs560887
(N9,353) Previously published, Science, May 2007
Stage 1 Genome-wide scan - 392,365 SNPs French
(N1,376) 679 cases, 697 controls
Stage 1 Genome-wide scan - 392,365 SNPs French
(N1,376) 679 cases, 697 controls
Stage 1 Genome-wide scan - 392,365 SNPs French
(N1,376) 679 cases, 697 controls
Stage 1 Genome-wide scan - 392,365 SNPs French
(N1,376) 679 cases, 697 controls
Stage 1 Genome-wide scan - 392,365 SNPs French
(N1,376) 679 cases, 697 controls
Stage 1 Genome-wide scan - 392,365 SNPs French
(N1,376) 679 cases, 697 controls
Stage 1 Genome-wide scan - 392,365 SNPs French
(N1,376) 679 cases, 697 controls
Stage 1 Genome-wide scan - 392,365 SNPs French
(N1,376) 679 cases, 697 controls
Stage 1 Genome-wide scan - 392,365 SNPs French
(N1,376) 679 cases, 697 controls
Stage 1 Genome-wide scan - 392,365 SNPs French
(N1,376) 679 cases, 697 controls
Focused Stage 2 - 16,273 SNPs French
(N4,977) 2,245 cases, 2,732 controls
Fast-track confirmation - 57 SNPs French
(N5,511) 2,617 cases, 2,894 controls Previously
published, Nature, Feb 2007
Focused Stage 2 - 16,273 SNPs French
(N4,977) 2,245 cases, 2,732 controls
Fast-track confirmation - 57 SNPs French
(N5,511) 2,617 cases, 2,894 controls Previously
published, Nature, Feb 2007
Focused Stage 2 - 16,273 SNPs French
(N4,977) 2,245 cases, 2,732 controls
Fast-track confirmation - 57 SNPs French
(N5,511) 2,617 cases, 2,894 controls Previously
published, Nature, Feb 2007
Focused Stage 2 - 16,273 SNPs French
(N4,977) 2,245 cases, 2,732 controls
Fast-track confirmation - 57 SNPs French
(N5,511) 2,617 cases, 2,894 controls Previously
published, Nature, Feb 2007
Focused Stage 2 - 16,273 SNPs French
(N4,977) 2,245 cases, 2,732 controls
Focused Stage 2 - 16,273 SNPs French
(N4,977) 2,245 cases, 2,732 controls
Fast-track confirmation - 57 SNPs French
(N5,511) 2,617 cases, 2,894 controls Previously
published, Nature, Feb 2007
Focused Stage 2 - 16,273 SNPs French
(N4,977) 2,245 cases, 2,732 controls
Fast-track confirmation - 57 SNPs French
(N5,511) 2,617 cases, 2,894 controls Previously
published, Nature, Feb 2007
Focused Stage 2 - 16,273 SNPs French
(N4,977) 2,245 cases, 2,732 controls
Fast-track confirmation - 57 SNPs French
(N5,511) 2,617 cases, 2,894 controls Previously
published, Nature, Feb 2007
Focused Stage 2 - 16,273 SNPs French
(N4,977) 2,245 cases, 2,732 controls
Fast-track confirmation - 57 SNPs French
(N5,511) 2,617 cases, 2,894 controls Previously
published, Nature, Feb 2007
Focused Stage 2 - 16,273 SNPs French
(N4,977) 2,245 cases, 2,732 controls
CASE-CONTROL T2D ASSOCIATION
Focused Stage 3 - 28 SNPs Danish (N7,698) 3,334
cases, 4,364 controls
Focused Stage 3 - 28 SNPs Danish (N7,698) 3,334
cases, 4,364 controls
Stage 4 population effect study - 1 SNP
(rs2943641) Population based study
samples French (N3,351), Finnish (N5,183),
Danish (N5,824)
Stage 4 population effect study - 1 SNP
(rs2943641) Population based study
samples French (N3,351), Finnish (N5,183),
Danish (N5,824)
QT ASSOCIATION IN POPULATIONS
34Stage 1 samples
- French individuals 690 cases, 670 controls
- Criteria for cases
- T2D
- First degree relative with T2D
- Non-obese (BMI lt 31 kg/m² , 25.8 2.8 kg/m²)
- Controls from DESIR, a prospective French cohort
- Normal glucose tolerance for the 9 years of the
study
35Stage 1 SNPs
- Tested on Illumina Human1 (100k) and HumanHap300
(300k) - 392,935 unique SNPs from the combined arrays
36Stage 1 results
37Fast-track validation
- Top 57 fast-tracked and tested on a Sequenom
panel on 2,617 cases, 2,894 controls - Relaxed criteria for cases
- BMI lt 35 kg/m² (28.9 3.7 kg/m²)
- Sladek et al., Nature 445, 881-885, 2007
38Results
SNP Chr Position pMAX Closest gene
rs7903146 10 114748339 1.5 x 10-34 TCF7L2
rs13266634 8 118253964 6.1 x 10-8 SLC30A8
rs1111875 10 94452862 3.0 x 10-6 HHEX
rs7923837 10 94471897 7.5 x 10-6 HHEX
rs7480010 11 42203294 1.1 x 10-4 LOC387761
rs3740878 11 44214378 1.2 x 10-4 EXT2
rs11037909 11 44212190 1.8 x 10-4 EXT2
rs1113132 11 44209979 3.3 x 10-4 EXT2
39SLC30A8
Chimienti et al. Biometals 18313
40HHEX
41HHEX controls pancreatic development
Hex homeobox gene-dependent tissue positioning is
required for organogenesis of the ventral
pancreas. Bort (2004) Heart induction by Wnt
antagonists depends on the homeodomain
transcription factor Hex. Foley (2005) The
homeobox gene Hex is required in definitive
endodermal tissues for normal forebrain, liver
and thyroid formation. Martinez Barbera (2000)
Habener Endocrinology 1461025
42Stage 2
- Top 5 of GWAS hits were selected for design of a
focused Stage 2 - Control for population bias with EIGENSTRAT
- iSelect array with 16,405 SNPs, tested on 2,245
cases, 2,732 controls (French) - Analysis with EIGENSTRATand selection of 28 SNPs
for a focused Stage 3
43QC
Exclusion criterion Samples
Call rate lt 95 27
Continental stratification 296
Sex mismatch 64
Related individuals 70
Total 457
Chromosome SNPs Failed HWE Failed MAF Successful
TOTAL 16,360 48 43 16,273
44EIGENSTRATcorrection
filters for MAF, HWE, call rate
filters for MAF, HWE, call rate and r2
45Results - stage 1 vs stage 2
46Results - taking out known loci
47(No Transcript)
48Stage 3
- The top 28 SNPs were tested using a Sequenom
panel in 7,700 Danish cases and controls - We confirm association of TCF7L2, WFS1, CDKAL1
and find one new association rs2943641 near IRS1
49rs2943641
- We studied the effect of variation in rs2943641
on T2D risk and metabolic phenotypes in general
populations - DESIR 3,351 French adults
- Inter99 5,183 Danish adults
- NFBC 1986 5,824 Finnish adolescents
50Metabolic traits
- A variety of indexes to capture b-cell function
and insulin resistance - HOMA-B and HOMA-IR based on fasting levels of
glucose and insulin - For Inter99, we had access to OGTT data and could
calculate other measures of insulin response - time course data
- AUC
- corrected insulin response (CIR)
- disposition indexes
51Oral Glucose Tolerance Test
52Metabolic traits 1
Metabolic trait Cohort rs2943641 rs2943641 rs2943641 P add P dom P rec
Metabolic trait Cohort C/C C/T T/T P add P dom P rec
Age NFBC 1986 16 16 16
Age DESIR 47.1 9.8 47.5 9.9 47.6 10.1
Age INTER99 44.9 7.9 45.4 7.8 45.2 7.6
Sex NFBC 1986 1062/1092 1153/1208 322/346
Sex DESIR 645/728 728/812 216/222
Sex INTER99 776/942 974/1070 307/354
BMI (kg/m2) NFBC 1986 21.3 3.8 21.3 3.7 21.1 3.5 0.24 0.43 0.21
BMI (kg/m2) DESIR 24.5 3.7 24.4 3.5 24.4 3.4 0.55 0.63 0.61
BMI (kg/m2) INTER99 25.6 3.9 25.4 4.1 25.7 4.2 0.57 0.094 0.24
Fasting plasma glucose (mmol/l) NFBC 1986 5.13 0.41 5.14 0.40 5.13 0.41 0.77 0.62 0.90
Fasting plasma glucose (mmol/l) DESIR 5.21 0.44 5.20 0.42 5.18 0.43 0.05 0.32 0.07
Fasting plasma glucose (mmol/l) INTER99 5.31 0.40 5.31 0.41 5.33 0.39 0.66 0.93 0.32
Fasting serum insulin (pmol/l) NFBC 1986 78.7 48.6 76.8 44.5 71.7 32.1 0.001 0.03 0.0009
Fasting serum insulin (pmol/l) DESIR 50.6 32.9 48.4 29.7 49.1 29.1 0.05 0.003 0.76
Fasting serum insulin (pmol/l) INTER99 38.8 24.7 36.4 21.9 37.6 23.3 0.018 0.0043 0.49
53Metabolic traits 2
HOMA-B NFBC 1986 141 95.1 136 80.1 131 91.6 0.006 0.05 0.009
HOMA-B DESIR 109 87.0 103 64.8 108 92.2 0.16 0.006 0.24
HOMA-B INTER99 75.2 65.6 68.3 42.2 71.0 49.9 0.005 0.0011 0.32
HOMA-IR NFBC 1986 2.52 1.63 2.47 1.58 2.29 1.06 0.007 0.07 0.005
HOMA-IR DESIR 1.95 1.35 1.86 1.20 1.88 1.17 0.03 0.004 0.95
HOMA-IR INTER99 1.54 1.00 1.44 0.89 1.49 0.95 0.026 0.0058 0.59
Insulin 30 INTER99 300 183 277 172 281 169 0.0019 8.1 x 10-4 0.14
Insulin 120 INTER99 176 138 163 127 162 124 0.0059 0.011 0.057
AUC insulin INTER99 22000 13800 20300 12900 20500 12700 6.9 x 10-4 2.2 x 10-4 0.12
Glucose 30 INTER99 8.19 1.53 8.17 1.56 8.22 1.50 0.72 0.34 0.55
Glucose 120 INTER99 5.51 1.11 5.51 1.11 5.47 1.15 0.54 0.99 0.23
AUC glucose INTER99 182 101 181 102 180 99.5 0.44 0.48 0.59
AUC insulin / AUC glucose INTER99 32.5 17.4 30.1 16.2 30.6 16.1 6.0 x 10-4 1.6 x 10-4 0.13
CIR INTER99 1140 4210 1000 1130 1000 1060 0.045 0.066 0.17
ISI INTER99 0.151 0.095 0.16 0.098 0.156 0.096 0.026 0.0058 0.59
Disp. Index (CIR ISI) INTER99 180 1610 147 220 143 174 0.73 1.0 0.50
54IRS1 locus - rs2943641
55IRS1
- G972R is a missense polymorphism in IRS1 that is
known to impair insulin signalling (rs1801278)
(Almind 1993) - G972R associated to insulin resistance and
insulin release (Clausen 1995, Sesti 2001) - In mice, IRS1 disruption causes disrupted insulin
action, both in target tissues and in b-cells
(Nandi 2004) - Also linked to insulin resistance, glucose
intolerance, islet hyperplasia (Tamemoto 1994,
Araki 1994, Terauchi 1997, Withers 1998) - G972R not conclusively associated to T2D (Florez
2004, Florez 2007, Jellema 2003, Zeggini 2004) - We detect no epistasis between rs2943641 and
G972R in DESIR or NFBC, only nominal significance
in Inter99 - Evidence for link between rs2943641 and IRS1?
56rs2943641 - IRS1 protein association
57rs2943641 - IRS1 protein association
rs2943641 CC rs2943641 CT rs2943641 TT PAdd PDom PRec
n (male/female) 74 (35/39) 88 (51/37) 28 (10/18)
Age (years) 42.5 17.1 43.5 16.9 43.2 17.6
BMI (kg/m2) 25.0 3.8 24.9 3.9 25.3 4.1 0.3 0.7 0.2
Rd insulin clamp (mg/kgFFM/min) 10.4 3.5 11.0 3.2 11.7 3.7 0.2 0.2 0.4
Di (x 10-7) 1.7 1.1 1.8 1.3 1.8 1.1 0.8 0.8 0.9
IRS-1 protein basal (AU) 296.7 167.7 314.0 155.1 413.1 227.6 0.03 0.3 0.009
IRS-1 protein insulin (AU) 276.6 143.6 280.9 156.4 313.3 147.9 0.3 0.7 0.2
IRS-1-associated PI3K activity basal (AU) 25.0 12.6 26.6 15.4 30.1 17.2 0.3 0.4 0.4
IRS-1-associated PI3K activity insulin (AU) 47.1 29.9 56.6 32.1 72.2 41.3 0.001 0.02 0.002
58Conclusions
- The multi-stage study detected T2D risk loci that
were later confirmed in other cohorts (SLC30A8,
HHEX) - Variation in rs2943641 is associated to
- T2D risk
- increased insulin levels
- impaired insulin sensitivity
- IRS1 protein levels
- IRS1 activity in insulin signaling pathway
- Study provided a full story from GWAS scan to
functional evidence thanks to rich phenotyping
59Paper
Rung et al., Nature Genetics, 41, 1110-1115, 2009
60Acknowledgements
Rosalie Frechette Valérie Catudal Philippe
Laflamme Stephane Cauchi Christian Dina David
Meyre Christine Cavalcanti-Proença Anders
Albrechtsen Torben Hansen Knut Borch-Johnsen Torst
en Lauritzen Marjo-Riitta Järvelin Jaana
Laitinen Emmanuelle Durand Paul Elliott Samy
Hadjadj Michel Marre
Alexander Montpetit Charlotta Pisinger Barry
Posner Anneli Pouta Marc Prentki Rasmus
Ribel-Madsen Aimo Ruokonen Anelli Sandbaek Jean
Tichet Martine Vaxillaire Jorgen
Wojtaszewski Allan Vaag
- Johan Rung
- Rob Sladek
- Philippe Froguel
- Oluf Pedersen
- Constantin Polychronakos
- Ghislain Rocheleau
- Alexander Mazur
- Lishuang Shen
- David Serre
- Philippe Boutin
- Daniel Vincent
- Alexandre Belisle
- Samy Hadjadj
- Beverley Balkau
- Barbara Heude
- Guillaume Charpentier
- Tom Hudson
- Sebastien Brunet
- François Bacot
61GWAS into context
- Complexity of interactions in biological
systems...
62Complexity
63B
G
C
A
D
F
E
A
B
64Redundancy
65Network structure
- Biological networks have a scale-free structure
Log( genes)
Most genes have few connections
Few genes have many connections
Log(edges)
66Signal propagation
- The structure of biological networks result in
robustness against random errors - Most mutations, even knockouts, can go by
unnoticed because of redundancy and network
wiring - Low probability to knock out a hub
67Common diseases
- What is most common - disease cause by many
variants with low effect, or few rare variants
with strong effects? - GWAS so far have by necessity focused on common
variants - Many known rare variants associated with common
diseases - or phenotypes that may contribute and
progress to disease
68Common disease / common variant
- The hypothesis that most common diseases are
caused by a large number of variants, common in a
general population, but each adding just a small
risk - GWAS results find many loci for common complex
diseases, with small risk - But... GWAS detected loci so far only explain a
very small fraction of the observed variation
69Rare variants
- With improved and lower cost sequencing, we can
address rare variants - Not just SNPs
- Utility of extreme cohorts
- Ex. A new highly penetrant form of obesity due
to deletions on chromosome 16p11.2 (Nature Feb
4, 2010)
70Polygenic contributions
- Groups of non-genomewide significant SNPs proven
to be associated with phenotype - Individual SNPs can not be inferred, just group
action - Supports the idea of many weak variants
responsible for effect - Ex. Common polygenic variation contributes to
risk of schizophrenia and bipolar disorder
(Nature 460, 748-752)
71Meta-analysis caveats
- Meta-analysis on heterogeneous data
- Phenotypes
- Quality control
- Platforms
- Genotype calling
- Analysis
72Future directions for GWAS
- Sequencing is cheaper and yielding higher quality
data - Better basis for studying and detecting rare
variants and their effect on diseases or
phenotypes - Copy number variants
- Genetic interactions, GxE interactions
- More samples gt higher power
73Future directions for GWAS
- Complex phenotypes
- Association of genetic loci to
- genome-wide expression levels
- protein levels
- metabolite levels
74Future directions for GWAS
- More data shared gt better quality of results
- As in other branches of science, data sharing,
transparency and openness should be promoted
75Resources
- Analysis software packages
- PLINK - http//pngu.mgh.harvard.edu/purcell/plink
/ - Abel - http//mga.bionet.nsc.ru/yurii/ABEL/
- MERLIN - http//www.sph.umich.edu/csg/abecasis/mer
lin/ - Imputations
- IMPUTE - http//mathgen.stats.ox.ac.uk/impute/impu
te.html - MACH - http//www.sph.umich.edu/csg/abecasis/MACH/
- Population structure
- Eigenstrat - http//genepath.med.harvard.edu/reic
h/Software.htm - EMMA(X) - http//genetics.cs.ucla.edu/emmax/index.
html - Meta-analysis
- METAL - http//www.sph.umich.edu/csg/abecasis/META
L/ - GWAMA - http//www.well.ox.ac.uk/gwama/index.shtml
- Data
- EGA - http//www.ebi.ac.uk/ega/