7 What Have We Learned from GWAS?
165
incorporate genotypes at SNPs that are not completely certain in the association with
phenotype are relatively well understood, at least for common variation (Marchini
and Howie 2010). Both the utilization of imputation to test markers for association
in GWAS and the fact that consistent replication and validation using further direct
genotyping of imputed variants show empirically that imputation as a practice for
common variation does work and helps to improve power to associate variants with
phenotype.
7.2.5 Collecting Evidence Across Studies via Meta-Analysis
Once directly genotyped and imputed SNPs have been collected across multiple
studies for a specific trait, it is natural and statistically desirable to pool evidence
across studies into a single assessment of association. The most direct way
to perform this analysis is direct with genotypes, in a stratified analysis (e.g.,
the Cochran-Mantel-Haenszel procedure). However, this analysis approach is not
directly feasible when informed consent and patient protections for genotype data
do not allow the transmission of clinical data to external actors. In this case, it is
desirable to summarize association statistics study to study, by estimates of the
effect of the SNP alleles to phenotype and the associated error as well as by pvalues, and meta-analyze those data. The meta-analysis of association data in this
way raised a number of details that needed to be worked out to obtain faithful
and unconfounded estimates of association across studies. Some of those issues
were determining the most powerful and appropriate statistical models used for
meta-analysis; controlling for the “forward and reverse-strand ambiguous” nature
of A/T and C/G polymorphisms and ensuring that the same allele for each SNP
is tested for each contributing study; thresholds to filter out SNPs passing into
meta-analysis based on frequency, quality of imputation, and sample sizes; or how
stratification and genomic control ought to be applied to the subsequent data.
While the technical details did take time to work out (de Bakker et al. 2008),
ultimately, these and other issues surrounding the analysis were resolved effectively.
The combination of methodology and workflow to appropriately combine data to
maintain power and discover new loci has had a tremendous impact, not least
of all in terms of the number of discoveries as a meta-analytical consortium
for glycemic, cardiovascular, autoimmune-mediated, and psychiatric disease grow
larger in sample sizes as additional studies are added. The details and issues
surrounding how to amass information across multiple studies in a systematic and
comparable way are a critical step for complex trait studies to collectively analyze
a dozen of studies together, and those details will be all the more important as new
technologies and studies amass ever-increasing detail of genetic variation in patient
populations.
165
incorporate genotypes at SNPs that are not completely certain in the association with
phenotype are relatively well understood, at least for common variation (Marchini
and Howie 2010). Both the utilization of imputation to test markers for association
in GWAS and the fact that consistent replication and validation using further direct
genotyping of imputed variants show empirically that imputation as a practice for
common variation does work and helps to improve power to associate variants with
phenotype.
7.2.5 Collecting Evidence Across Studies via Meta-Analysis
Once directly genotyped and imputed SNPs have been collected across multiple
studies for a specific trait, it is natural and statistically desirable to pool evidence
across studies into a single assessment of association. The most direct way
to perform this analysis is direct with genotypes, in a stratified analysis (e.g.,
the Cochran-Mantel-Haenszel procedure). However, this analysis approach is not
directly feasible when informed consent and patient protections for genotype data
do not allow the transmission of clinical data to external actors. In this case, it is
desirable to summarize association statistics study to study, by estimates of the
effect of the SNP alleles to phenotype and the associated error as well as by pvalues, and meta-analyze those data. The meta-analysis of association data in this
way raised a number of details that needed to be worked out to obtain faithful
and unconfounded estimates of association across studies. Some of those issues
were determining the most powerful and appropriate statistical models used for
meta-analysis; controlling for the “forward and reverse-strand ambiguous” nature
of A/T and C/G polymorphisms and ensuring that the same allele for each SNP
is tested for each contributing study; thresholds to filter out SNPs passing into
meta-analysis based on frequency, quality of imputation, and sample sizes; or how
stratification and genomic control ought to be applied to the subsequent data.
While the technical details did take time to work out (de Bakker et al. 2008),
ultimately, these and other issues surrounding the analysis were resolved effectively.
The combination of methodology and workflow to appropriately combine data to
maintain power and discover new loci has had a tremendous impact, not least
of all in terms of the number of discoveries as a meta-analytical consortium
for glycemic, cardiovascular, autoimmune-mediated, and psychiatric disease grow
larger in sample sizes as additional studies are added. The details and issues
surrounding how to amass information across multiple studies in a systematic and
comparable way are a critical step for complex trait studies to collectively analyze
a dozen of studies together, and those details will be all the more important as new
technologies and studies amass ever-increasing detail of genetic variation in patient
populations.