Reading GWAS Association Signals and Polygenic Risk Scores
Use multiple-testing arithmetic, linkage disequilibrium and the distinction between a PRS percentile and disease probability to interpret genomic findings.
Publication & corrections: infostoa · Editorial standards
Individual author and specialist reviewer information is not provided for this article.
1. What GWAS compares
A genome-wide association study (GWAS) identifies statistical associations between genomic variants and diseases or traits. It can compare variant frequencies in cases and controls or analyze a continuous trait. Multiple variants and environmental factors contribute to complex traits, but the same number of variants does not act identically in every disease.
The 2007 Wellcome Trust Case Control Consortium study used arrays in a British sample, with approximately 2,000 cases for each of seven diseases and approximately 3,000 shared controls. Results from this design should be distinguished from subsequent studies with different samples or technologies.
2. Stringent p values and clusters of signals
Testing many hypotheses produces chance signals. For illustrative arithmetic, suppose one million tests have no true associations and each uses p<0.05. The expected false-positive count is 1,000,000×0.05=50,000. This is an expectation under hypothetical conditions, not an actual GWAS result or a claim that one million SNPs are independent. The corresponding Bonferroni calculation is 0.05/1,000,000=5×10⁻⁸. This is the representative genome-wide threshold used in Visscher and colleagues' 2017 review, not a mandatory threshold for every study. The actual 2007 paper's abstract reported 24 independent association signals at p<5×10⁻⁷.
Linkage disequilibrium (LD) is nonrandom correlation among variants. Nearby SNPs inherited together can produce several signals in one associated region. The lead SNP with the smallest p value is therefore not automatically causal, nor is the nearest gene automatically responsible. Correlated SNPs should not all be described as independent tests. Fine-mapping and functional data can narrow candidates, but the evidence provided by each method must be assessed separately.
3. Moving from discoveries to a PRS
A polygenic risk score (PRS) combines information from multiple variants to summarize genetic risk differences. Estimated variant effects from a discovery sample are applied to other people's genomic data; SNP selection and effect estimation differ across scores. A list of significant SNPs is different from a predictor whose performance has been evaluated in an independent sample.
A PRS percentile is not a disease probability. An illustrative result at the 90th percentile describes a position in a specified reference population's score distribution, not a 90% chance of disease. Absolute risk requires separately connecting the disease, a timeframe, baseline risk and clinical information such as age. NHGRI likewise distinguishes a score based on genomic information from other risk factors.
When reading a result, check the discovery sample, disease and threshold, the independent population used to validate the score, and differences between that population and the situation of interest. Across ancestry groups, LD and variant frequencies can differ, so predictive performance should not be assumed to transfer unchanged. The future projections or clinical-adoption descriptions in the 2017 review and 2020 NHGRI guide should not be imported as current clinical standards.