Abstract
<title>Abstract</title> <p> Ischemic heart disease (IHD) carries a substantial polygenic component, but candidate-gene and polygenic-score evidence remains scarce in South Asians. This Pakistani case–control study evaluated eight candidate variants, their aggregation into conventional and epistasis-weighted polygenic scores, and machine-learning classifiers. 612 participants (306 IHD cases, 306 controls) were genotyped for <italic>ADAMTS7</italic> rs3825807, <italic>ADAMTS13</italic> rs2301612, <italic>APOE</italic> rs769452, <italic>AGT</italic> rs699, <italic>APOB</italic> rs676210, <italic>MMP9</italic> rs3918242, <italic>MTHFR</italic> rs1801133 and <italic>ZC3HC1</italic> rs11556924, with covariate-adjusted single-variant testing. Unweighted and effect-size-weighted polygenic risk scores (PRS) and a three-tier epistasis-weighted candidate-gene PRS (EW-cgPRS) were built, a lipid quantitative-trait scan performed, and four machine-learning classifiers benchmarked. Six of eight variants were nominally associated with IHD; four survived Bonferroni correction across 40 inheritance-model tests: <italic>ADAMTS13</italic> (additive OR 6.85, 95% CI 3.57–13.12), <italic>ADAMTS7</italic> (recessive OR 6.00, 2.69–13.40), <italic>APOE</italic> (dominant OR 4.44, 2.12–9.26) and <italic>MMP9</italic> (recessive OR 14.49, 3.40–61.80). The weighted PRS separated cases from controls (cross-validated AUC 0.861), with a monotonic case-prevalence gradient across strata (top- versus bottom-tertile OR 59.2). Neither epistatic nor functional-annotation weighting improved on the classic score; the interaction tier's apparent superiority reflected covariate access, not epistasis. All four classifiers achieved near-perfect discrimination (AUC 0.98–0.999), traced to lipid-based ascertainment of controls rather than genetic signal. Several candidate variants are associated with IHD in this population, and a simple weighted score captures a reproducible genetic gradient; epistatic or functional enrichment adds nothing. The near-perfect machine-learning performance reflects study design, not clinical utility, and the largest single-variant effects warrant cautious interpretation given the modest sample size, Hardy–Weinberg deviations, and single-population design. </p>