Genome-wide association studies of polygenic risk score-derived phenotypes may lead to inflated false positive rates

Abstract

Gouveia and colleagues (2022) 1 conducted a genome-wide association study (GWAS) of a polygenic risk score (PRS)-derived phenotype (N = 37,784), in which they identified 246 independent loci and 473 lead SNPs. This is an enormous increase compared to the most recent and largest GWAS of AD 2 (N = 1,126,563), which identified 38 loci. Here we show that the applied approach by Gouveia and colleagues may lead to an inflated false positive rate.In this approach, beta-estimates from a recent GWAS of Alzheimer's disease (AD) 3 were used to construct PRSs in the European UK Biobank 4 sample, using pruning and thresholding 5 with a p-value threshold of 5%. Next, a new case-control phenotype was constructed based on the bottom and top 5% of the PRS distribution, removing 90% of their initial sample. Lastly, a GWAS was conducted on this new PRS-derived phenotype. The authors reasoned that by enriching the sample for individuals with known AD-associated variants, you may also enrich for unknown AD-associated variants. Our major concern is that the applied approach used the same single-nucleotide polymorphisms (SNPs) to construct, as well as to predict the phenotype. In other words, the phenotype was partly regressed on itself, which can inflate test statistics.We performed simulations roughly emulating the approach (see Methods). In short, we simulated individual phenotypes under a liability threshold model and genotypes that loosely reflect the genetic architecture of AD 2,3,6 (excluding the APOE locus) including 170,000 independent SNPs of which 1200 were causal and 168,800 were non-causal (null-SNPs). We then simulated a discovery sample such that the PRS explains approximately 5% of the phenotypic variance on the liability scale (N = 366,771). We ran a GWAS of AD in this discovery sample and used the estimated betas to construct a PRS in a target sample (N = 300,000). We then selected individuals in the top and bottom 5% of the PRS distribution (N = 30,000) and ran a second GWAS on this new PRS-derived case-control phenotype. The target cohort overlapped to varying degrees with the discovery cohort (i.e. 0%, 50%, and 100%), noting the AD GWAS summary statistics used by Gouveia and colleagues (2022) 1 also contained the UK Biobank.Our results show highly inflated false positive rates in the GWAS of the PRS-derived phenotype (see Fig. 1 and Supplementary Table ). Across all null-SNPs and when there is no overlap between discovery and target cohort, the false positive rate was 0.0024 (s.e.m. = 1 × 10 -5 ), which constitutes a 48,000-fold increase compared to a well-controlled false positive rate of 5 × 10 -8 (see Supplementary Fig. 1 for α = 0.05). This inflation is driven by null-SNPs that were used to construct the PRS-derived phenotype. The false positive rate of these null-SNPs was equal to 0.05 (s.e.m. = 0.0002, a 1 × 10 6 -fold increase) when there was no overlap, while null-SNPs which were not used to construct the PRS-derived phenotype did not show any inflation. We also looked at the number of false positive associ...

Reply to: Genome-wide association studies of polygenic risk score-derived phenotypes may lead to inflated false positive rates

“…

In their Matters Arising article, Uffelmann et al 1 present simulation experiments suggesting that the approach used in 2 , where genome-wide association studies are based on polygenic risk score-derived phenotypes, may lead to inflated positive rates. We acknowledge the importance of confirming the increased number of false positive results by performing simulation experiments but have reservations about how these impact our interpretation of data and subsequent conclusions derived from the results in the original study.Our study determined each individual's polygenic risk score (PRS) for AD in the UK Biobank dataset.

…”