94
R. E. Graff et al.
SNPs are genotyped with a less expensive genotyping platform in the remaining
samples. The procedure prioritizes the most promising SNPs for evaluation in
additional stages and can pinpoint associated regions for fine mapping. The optimal
division of samples across stages depends on a number of factors, but in general,
the most efficient approach entails the inclusion of approximately one-third to onehalf of the samples in the initial stage and the remaining samples in follow-up
stages (Skol et al. 2006, 2007). The number of noteworthy SNPs that should be
tested depends on the sample sizes in the respective stages, the number of falsenegative results that one is willing to accept, and whether or not one wishes to
incorporate SNP information (e.g., proximity to the nearest gene or likelihood of
being functional) (Chen and Witte 2007; Roeder and Wasserman 2009; Roshan et
al. 2011; Thomas et al. 2009). Ideally, at least 1% of the first stage SNPs should be
typed in the second stage (Skol et al. 2006). One must also decide whether the early
follow-up stages should be treated as part of a replication or joint analysis.
5.2.2.3 Limitations
Despite their numerous strengths, GWAS carry several notable limitations. First, it
is important to note that most variants discovered via GWAS are only associated
with, and not causal for, disease. Even when an association is real and statistically
reproducible in other datasets, another untyped variant in LD with the associated
SNP may still be the causal variant. Determining the factors underlying results
can be extremely challenging and require separate fine-mapping and mechanistic
studies. That many of the associations detected to date are not in gene regions can
make the findings yet more complicated (Buniello et al. 2019). These issues limit
our understanding of the biological basis of results and our ability to implement
preventive or therapeutic measures.
Second, findings from GWAS thus far account for only a limited amount of
disease heritability (Maher 2008; Nolte et al. 2017). Most SNPs detected by GWAS
show a small magnitude of effect. That said, as sample sizes for GWAS are
increasing, studies are detecting and replicating a larger number of trait-associated
variants. That we are now also able to examine essentially the entirety of common
variation across the genome (at least indirectly) allows us to explain an increasing
proportion of heritability. So too does our ability to assess the contribution of rare
variants. The polygenic model of heritability is becoming increasingly accepted;
many risk variants with small effect sizes are thought to underlie disease risk.
Finally, GWAS have not yet sufficiently distinguished between individuals with
low- and high-risk disease. In general, screening tests based on SNPs detected by
GWAS to date may have low positive (and negative) predictive value for disease
and thus limited utility in a diagnostic setting (Kraft et al. 2009; Ware 2006). As
more SNPs are discovered, however, combining them into polygenic risk scores
(PRS) efficiently summarizes individuals’ genetic susceptibility profiles, thereby
improving phenotypic prediction (Torkamani et al. 2018). PRS have the potential
to personalize risk estimates and improve the discriminatory ability of screening
tests (Mavaddat et al. 2019; Toland 2019). For example, a 2015 study created a
risk score of 105 SNPs that was strongly associated with prostate cancer risk among
R. E. Graff et al.
SNPs are genotyped with a less expensive genotyping platform in the remaining
samples. The procedure prioritizes the most promising SNPs for evaluation in
additional stages and can pinpoint associated regions for fine mapping. The optimal
division of samples across stages depends on a number of factors, but in general,
the most efficient approach entails the inclusion of approximately one-third to onehalf of the samples in the initial stage and the remaining samples in follow-up
stages (Skol et al. 2006, 2007). The number of noteworthy SNPs that should be
tested depends on the sample sizes in the respective stages, the number of falsenegative results that one is willing to accept, and whether or not one wishes to
incorporate SNP information (e.g., proximity to the nearest gene or likelihood of
being functional) (Chen and Witte 2007; Roeder and Wasserman 2009; Roshan et
al. 2011; Thomas et al. 2009). Ideally, at least 1% of the first stage SNPs should be
typed in the second stage (Skol et al. 2006). One must also decide whether the early
follow-up stages should be treated as part of a replication or joint analysis.
5.2.2.3 Limitations
Despite their numerous strengths, GWAS carry several notable limitations. First, it
is important to note that most variants discovered via GWAS are only associated
with, and not causal for, disease. Even when an association is real and statistically
reproducible in other datasets, another untyped variant in LD with the associated
SNP may still be the causal variant. Determining the factors underlying results
can be extremely challenging and require separate fine-mapping and mechanistic
studies. That many of the associations detected to date are not in gene regions can
make the findings yet more complicated (Buniello et al. 2019). These issues limit
our understanding of the biological basis of results and our ability to implement
preventive or therapeutic measures.
Second, findings from GWAS thus far account for only a limited amount of
disease heritability (Maher 2008; Nolte et al. 2017). Most SNPs detected by GWAS
show a small magnitude of effect. That said, as sample sizes for GWAS are
increasing, studies are detecting and replicating a larger number of trait-associated
variants. That we are now also able to examine essentially the entirety of common
variation across the genome (at least indirectly) allows us to explain an increasing
proportion of heritability. So too does our ability to assess the contribution of rare
variants. The polygenic model of heritability is becoming increasingly accepted;
many risk variants with small effect sizes are thought to underlie disease risk.
Finally, GWAS have not yet sufficiently distinguished between individuals with
low- and high-risk disease. In general, screening tests based on SNPs detected by
GWAS to date may have low positive (and negative) predictive value for disease
and thus limited utility in a diagnostic setting (Kraft et al. 2009; Ware 2006). As
more SNPs are discovered, however, combining them into polygenic risk scores
(PRS) efficiently summarizes individuals’ genetic susceptibility profiles, thereby
improving phenotypic prediction (Torkamani et al. 2018). PRS have the potential
to personalize risk estimates and improve the discriminatory ability of screening
tests (Mavaddat et al. 2019; Toland 2019). For example, a 2015 study created a
risk score of 105 SNPs that was strongly associated with prostate cancer risk among