Of the existing algorithms 54-57 we chose HYST to conduct the pathway analyses because: (a) it has adequate control of the type I error rate, (b) good power (i.e., it is at least as powerful as - and in some circumstances - more powerful than alternative set-based tests 14, 16), (c) it is computationally efficient (i.e., it does not rely on permutations/simulations to account for LD between the SNPs when computing the p-values), (d) it exploits all the available information (i.e., the SNP p-values and their LD structure) to compute the pathway p-value, and thus does not require arbitrarily chosen p-value thresholds for selecting the list of input genes, and (e) importantly, it uses up to date annotation information on variants, genes and pathways from bioinformatics databases. Yet, despite the differences in terms of assumptions and the algorithm's power, we note that in large samples as the one we used, results obtained based on alternative pathway-based approaches converge 56 (see also Supplementary Tables S8 and S9 for results obtained using an alternative overrepresentation test as implemented in Reactome 23). In
