We focused on GWAS data from height7 that identified 697 genome-wide significant variants in 423 loci, and conducted the summary-based TWAS over all genes in these loci using YFS and NTR as expression training data. Because the YFS individuals had been phenotyped for height and not tested in the GWAS, we could directly evaluate whether selected genes were associated with phenotype. At each locus, we considered three strategies for selecting a single causal gene: 1) the gene nearest to the most significantly associated SNP; 2) the gene for which the index SNP is the strongest eQTL in the training data; 3) the most significant TWAS gene. For each strategy, we then constructed a risk-score using the genetic value of expression for the selected gene weighted by the corresponding TWAS Z-score. The same procedure was then re-evaluated using TWAS values trained in the NTR cohort (which introduces additional noise due to heterogeneity between the cohorts, Supplementary Table 5).We separately used GCTA to estimate the heritability of height explained by all of the genes selected by each algorithm by constructing a GRM
