Knowledge-driven multi-locus analysis reveals gene-gene interactions influencing HDL cholesterol level in two independent EMR-linked biobanks

Abstract

Genome-wide association studies (GWAS) are routinely being used to examine the genetic contribution to complex human traits, such as high-density lipoprotein cholesterol (HDL-C). Although HDL-C levels are highly heritable (h(2)∼0.7), the genetic determinants identified through GWAS contribute to a small fraction of the variance in this trait. Reasons for this discrepancy may include rare variants, structural variants, gene-environment (GxE) interactions, and gene-gene (GxG) interactions. Clinical practice-based biobanks now allow investigators to address these challenges by conducting GWAS in the context of comprehensive electronic medical records (EMRs). Here we apply an EMR-based phenotyping approach, within the context of routine care, to replicate several known associations between HDL-C and previously characterized genetic variants: CETP (rs3764261, p = 1.22e-25), LIPC (rs11855284, p = 3.92e-14), LPL (rs12678919, p = 1.99e-7), and the APOA1/C3/A4/A5 locus (rs964184, p = 1.06e-5), all adjusted for age, gender, body mass index (BMI), and smoking status. By using a novel approach which censors data based on relevant co-morbidities and lipid modifying medications to construct a more rigorous HDL-C phenotype, we identified an association between HDL-C and TRIB1, a gene which previously resisted identification in studies with larger sample sizes. Through the application of additional analytical strategies incorporating biological knowledge, we further identified 11 significant GxG interaction models in our discovery cohort, 8 of which show evidence of replication in a second biobank cohort. The strongest predictive model included a pairwise interaction between LPL (which modulates the incorporation of triglyceride into HDL) and ABCA1 (which modulates the incorporation of free cholesterol into HDL). These results demonstrate that gene-gene interactions modulate complex human traits, including HDL cholesterol.

Figure 1. Flow diagram overview of the analysis plan.

For the single locus analysis in the Marshfield PMRP cohort, a genome-wide association study was performed for both median and modeled HDL-C (see Phenotyping section in Methods). For the multilocus analysis, the Biofilter was used to generate putative multilocus interaction models that were tested for association with median HDL-C in the Marshfield PMRP cohort. The Vanderbilt BioVU cohort was used for replication. See the Methods and Results sections.

Figure 2. Summary of genome-wide association results.

This plot shows the –log10(P-values) from linear mixed effects regression model using the natural log tranformed median HDL-C level in the Marshfield PMRP sample, adjusted for smoking, age, age², BMI, BMI², and gender. The red line indicates a Bonferroni-corrected significance threshold. Genes described in the results and the discussion sections are highlighted in green: ADIPOQ (chromosome 3), LPL (chromosome 8), TRIB1 (chromosome 8), APOA1/C3/A4/A5 (chromosome 11), LIPC (chromosome 15), and CETP (chromosome 16).

Figure 3. Association results for the TRIB1 region.

This shows the association results for the Tribbles 1 Homolog (TRIB1) region, for median adjusted HDL-C (left) and modeled HDL-C (right) in the Marshfield PMRP cohort. Color scale displays the degree of linkage disequilibrium (r²) between markers. Blue line shows recombination rate from HapMap CEU. Gene location is shown along the horizontal axis of each panel. This plot shows that while the statistical significance of the effect of TRIB1 on HDL-C levels is less compelling when adjusting for triglyceride concentration, the association is much stronger when using the modeled HDL-C phenotype (even though this phenotype has far fewer samples than the median adjusted HDL-C phenotype). SNP rs4871603 in TRIB1 was associated with median HDL-C at p = 7.06e-4 and with modeled HDL-C at p = 2.61e-6 without adjusting for triglyceride concentration. After adjusting for triglycerides, the p-values for median and modeled HDL-C become less significant (p = 0.296 and p = 0.0056, respectively, data not displayed).