Handling and analyzing data from a genome-wide association study
Laura Scott
Biostatistics DepartmentCenter for Statistical Genetics
University of Michigan
Outline
• Storing large amounts of genotype data
• Quality control
• Generating initial association analysis
• Viewing results
• Imputation of missing SNP genotypes
• Storing results and planning specialized analysis
Genotype data is huge
• 500,000 SNPs * 2000 cases + controls = 1,000,000,000 genotypes!
• Need compact ways to store data– If store each genotype as 00, 01, or 11 will
have file that looks like Person 100011001010001001001100….
000000100001000010001000….
Genotype data is huge
• 500,000 SNPs * 2000 cases + controls = 1,000,000,000 genotypes!
• Need compact ways to store data– If store each genotype as 00, 01, or 11 will
have file that looks like SNP 100011001010001001001100….
000000100001000010001000….
Genotype data is huge
• 500,000 SNPs * 2000 cases + controls = 1,000,000,000 genotypes!
• Need compact ways to store data– If store each genotype as 00, 01, or 11 will
have file that looks like 100011001010001001001100….
000000100001000010001000….
– Total file space for 300K SNPs: 4 Gigabytes– Largest chromosome file: .4 Gigabytes
Need to do extensive planning for genotype data before it arrives
• Chromosome datasets are too large for SAS and other commonly used analytic packages
• Need programs to select and write out genotype data in multiple formats
• Tests of procedures with large-scale trial datasets
Gather other data needed for analysis
• SNP information– Chromosome– Position
• SNP annotation– Gene– Function
• Translation of called allele to a standard allele– Example forward strand of given genome build
How good is the data?
• Identify and remove bad samples and SNPs
• Compute summary statistics– Percent successfully genotyped samples– Average genotyping success rate– Duplicate sample error rate– Non-Mendelian inheritance error rates (errors
not consistent with normal transmission of chromosomes in family members)
Identify bad samples and remove
• Poor quality samples– Sample genotype success rate < 95 to 97.5%– Greater proportion of heterozygous genotypes
than expected
• Related individuals (if independent samples) – Based on pair-wise comparisons of similarity of
genotypes
• Sample switches– Wrong sex
• Regions of homozygosity in cell line
Identify poor quality SNPs and remove
• Expected proportions of genotypes are not consistent with observed allele frequency (Hardy Weinberg Equilibrium (HWE))
• HWE p-value < 10-4 to 10-6
• Look for deviation from expected distribution of p-values under the null
• Genotyping success rate < 95%• Duplicate sample or Non-Mendelian error rate
is elevated• Differential missingness in cases and controls
Programs are available for large scale quality control analysis
• Plink– Duplicate error rates, sample relatedness, HWE,.. – Develop by Shaun Purcell– http://pngu.mgh.harvard.edu/~purcell/plink/
• GAINQC Software used for the quality control analysis of the GAIN project – Duplicate error rates, sample relatedness, HWE,..– Developed by Shyam Gopalakrishnan and Goncalo
Abecasis– Available from [email protected]
Initial analysis is straightforward once have everything in place
• Case/control association– Use test that is not affected by deviations
from HWE• Cochran-Armitage test for trend• Equivalent to score test in logistic regression
• TDT or other family-based test
• Quantitative trait association
Programs are available for large scale case-control or family-based analysis• Plink
– Case/control, tdt, quantitative traits– Develop by Shaun Purcell– http://pngu.mgh.harvard.edu/~purcell/plink/
• Merlin– Quantitative traits in independent samples or families,
ability to impute genotypes for untyped individuals based on genotyped family members
– Developed by Goncalo Abecasis– http://www.sph.umich.edu/csg/abecasis/Merlin/
Are the results believable?
• Are stronger associations correlated with poorer quality control measures?
• Is there a strong deviation from expected distribution of p-values?
• Is there confounding from differences in the genetic origins of case and control samples (population stratification)?– Genomic control– Eigenstrat analysis
Seeing from many different angles is believing (sometimes)
• Plink graphical output
• User added custom tracks in the UCSC browser– http://genome.ucsc.edu/– http://genome.ucsc.edu/goldenPath/help/
hgTracksHelp.html#CustomTracks
• Homemade graphes
FUSION T2D association
Many different ways to display similar data
Zeggini et al. (2007) Science 316: 13361341; Diabetes Genetic Initiative (2007) Science 316:1331-1336; Scott et al., (2007) Science 316:1341-1345
Getting more for your genotyping dollars: Imputation of SNP genotypes
• Impute/predict genotypes for:– Missing data within genotyped markers– Untyped markers
• Uses haplotype structure of existing sample such as HapMap samples to infer data for samples with sparser marker set
Observed genotypes
Observed Genotypes
. . . . A . . . . . . . A . . . . A . . .
. . . . G . . . . . . . C . . . . A . . .
Reference Haplotypes
C G A G A T C T C C T T C T T C T G T G CC G A G A T C T C C C G A C C T C A T G GC C A A G C T C T T T T C T T C T G T G CC G A A G C T C T T T T C T T C T G T G CC G A G A C T C T C C G A C C T T A T G CT G G G A T C T C C C G A C C T C A T G GC G A G A T C T C C C G A C C T T G T G CC G A G A C T C T T T T C T T T T G T A CC G A G A C T C T C C G A C C T C G T G CC G A A G C T C T T T T C T T C T G T G C
Study Sample
HapMap
Gonçalo Abecasis
Identify match among reference
Observed Genotypes
. . . . A . . . . . . . A . . . . A . . .
. . . . G . . . . . . . C . . . . A . . .
Reference Haplotypes
C G A G A T C T C C T T C T T C T G T G CC G A G A T C T C C C G A C C T C A T G GC C A A G C T C T T T T C T T C T G T G CC G A A G C T C T T T T C T T C T G T G CC G A G A C T C T C C G A C C T T A T G CT G G G A T C T C C C G A C C T C A T G GC G A G A T C T C C C G A C C T T G T G CC G A G A C T C T T T T C T T T T G T A CC G A G A C T C T C C G A C C T C G T G CC G A A G C T C T T T T C T T C T G T G C
Gonçalo Abecasis
Phase chromosomes, impute missing genotypes
Observed Genotypes
c g a g A t c t c c c g A c c t c A t g gc g a a G c t c t t t t C t t t c A t g g
Reference Haplotypes
C G A G A T C T C C T T C T T C T G T G CC G A G A T C T C C C G A C C T C A T G GC C A A G C T C T T T T C T T C T G T G CC G A A G C T C T T T T C T T C T G T G CC G A G A C T C T C C G A C C T T A T G CT G G G A T C T C C C G A C C T C A T G GC G A G A T C T C C C G A C C T T G T G CC G A G A C T C T T T T C T T T T G T A CC G A G A C T C T C C G A C C T C G T G CC G A A G C T C T T T T C T T C T G T G C
Gonçalo Abecasis
Imputing genotype data allows much more thorough analysis
• Allows testing of untyped variation
• Allows easy combination of data across genotyping platforms
• Provides complete data for analysis with multiple SNPs
Imputed data takes care to generate, analyze and understand
• Requires large scale computing resources• Need to assess quality of imputation
– Compare imputed gentoypes to actual genotypes
• Error rates are higher than for genotyped SNPs• Works less well for rarer alleles• Best to take account of uncertainty imputed
SNPs in analysis– Need ways to take into account fractional genotype
counts
Imputation programs are available
• IMPUTE– Developed by Jonathan Marchini– Nature Genetics, Advance online publication– http://www.stats.ox.ac.uk/~marchini/#software
• Mach 1.0, Markov Chain Haplotyping– Developed by Goncalo Abecasis– http://www.sph.umich.edu/csg/abecasis/MACH/
Need to store results and prepare for large scale specialized analysis
• System to store, view, merge results– SQL database– Plink
• Testing speed of specialized analyses in different statistical packages
• Potential development of software to run large scale specialized analysis
Summary: Ideally what needs to happen before getting the data
• Ability to store, select and write out genotype data in multiple formats for quality control and association analysis
• Identification of primary quality control and analysis programs
• Systems to store, view, merge results• Adequate computing resources to do
intensive computing • Testing of standard and specialized
processes with large-scale trial datasets
FUSION study
U Michigan
Michael BoehnkeKaren ConneelyCharles Ding William DurenTerry GliedtLarry HuAnne JacksonXiao-Yi LiAndrew SkolHeather StringhamPeggy WhiteCristen WillerFang XiangRui Xiao
National Public Health Institute Helsinki
Jaakko TuomilehtoTimo Valle
USCRichard BergmanThomas BuchananRichard Watanabe
Randall PruimCalvin College
Kimberly DohenyElizabeth Pugh
CIDR
Francis CollinsLori BonnycastlePeter ChinesMichael ErdosNarisu NarisuL. Prokunina-OlssonNancy RiebowAndrew SprauAmy SwiftMaurine Tong
NHGRI / NIH
Karen MohlkeKyle GaultonJason Luo Li Qin
UNC-Chapel Hill
U Michigan
Gonçalo AbecasisYun LiJun DingPaul Scheet