Abstract
Inherited genetics represents an important contributor to risk of esophageal adenocarcinoma (EAC), and its precursor Barrett's esophagus (BE). Genome-wide association studies have identified ∼30 susceptibility variants for BE/EAC, yet genetic interactions remain unexamined. To address challenges in large-scale G×G scans, we combined knowledge-guided filtering and machine learning approaches, focusing on genes with (1) known/plausible links to BE/EAC pathogenesis (n = 493) or (2) prior evidence of biological interactions (n = 4,196). Approximately 75 × 106 SNP×SNP interactions were screened via hierarchical group lasso (glinternet) using BEACON GWAS data. The top ∼2,000 interactions retained in each scan were prioritized using p values from single logistic models. Identical scans were repeated among males only (78%), with two independent GWAS datasets used for replication. In overall and male-specific primary replications, 11 of 187 and 20 of 191 interactions satisfied p < 0.05, respectively. The strongest evidence for secondary replication was for rs17744726×rs3217992 among males, with consistent directionality across all cohorts (Pmeta = 2.19 × 10-8); rs3217992 "T" was associated with reduced risk only in individuals homozygous for rs17744726 "G." Rs3217992 maps to the CDKN2B 3' UTR and reportedly disrupts microRNA-mediated repression. Rs17744726 maps to an intronic enhancer region in BLK. Through in silico prioritization and experimental validation, we identified a nearby proxy variant (rs4841556) as a functional modulator of enhancer activity. Enhancer-gene mapping and eQTLs implicated BLK and FAM167A as targets. The first systematic G×G investigation in BE/EAC, this study uncovers differential risk associations for CDKN2B variation by BLK genotype, suggesting novel biological dependency between two risk loci encoding key mediators of tumor suppression and inflammation.