A second approach, pruning and thresholding, was used to build an additional 24 candidate GPSs. Pruning and thresholding scores were built using a p-value and LD-driven clumping procedure in PLINK version 1.90b (--clump).35 In brief, the algorithm forms clumps around SNPs with association p-values less than a provided threshold. Each clump contains all SNPs within 250kb of the index SNP that are also in LD with the index SNP as determined by a provided r2 threshold in the LD reference. The algorithm iteratively cycles through all index SNPs, beginning with the smallest p-value, only allowing each SNP to appear in one clump. The final output should contain the most significantly disease-associated SNP for each LD-based clump across the genome. A GPS was built containing the index SNPs of each clump with association estimate betas (log odds) as weights. GPSs were created over a range of p-value (1, 0.5, 0.05, 5×10−4, 5×10−6, 5×10−8) and r2 (0.2, 0.4, 0.6, 0.8) thresholds, for a total of 24 pruning and thresholding-based candidate scores for each disease. The resulting GPS for a p-value threshold of 5×10−8 and r2 of < 0.2 was denoted the ‘GWAS significant variant’ derivation strategy.
