Fig. 4: Power of RSS-NET to identify gene-level associations from GWAS summary statistics.

We used a B cell-specific regulatory network and real genotypes of 348,965 genome-wide SNPs to simulate individual-level GWAS data under four scenarios: a θ = 0, σ² = 0; b θ > 0, σ² = 0; c θ = 0, σ² > 0; d θ > 0, σ² > 0. Using the simulated individual-level data we computed single-SNP association statistics, on which we compared RSS-NET with gene-level association components of RSS-E¹⁶ and Pascal²⁶. RSS-E is a special case of RSS-NET assuming σ² = 0, and RSS-E-baseline is a special case of RSS-E assuming θ = 0. Pascal includes two gene scoring options: maximum-of-χ² (“max”) and sum-of-χ² (“sum”). Given a network, Pascal and RSS-E-baseline do not leverage any network information, RSS-E ignores the edge information, and RSS-NET exploits the full topology. Each scenario contains 200 datasets and each dataset contains 16,954 autosomal protein-coding genes for testing. We defined a gene as "trait-associated'' if at least one SNP j within 100 kb of the transcribed region of this gene had non-zero effect (β_j ≠ 0). For each gene in each dataset, RSS methods produced posterior probabilities that the gene was trait-associated (P₁), whereas Pascal methods produced association P-values; these statistics were used to rank the significance of gene-level associations. The first row of each panel displays ROC curves and AUROCs for all methods, with dashed diagonal lines indicating random performance (AUROC = 0.5). The second row of each panel displays precision-recall (PRC) curves and areas under PRC curves (AUPRCs) for all methods, with dashed horizontal lines indicating random performance. For both AUROC and AUPRC, higher value indicates better performance. Simulation details and additional results are provided in Supplementary Figs. 7, 8.