(P21-003-25) A New Approach to Refine Genetic Predictors of Dietary Traits for Use in Mendelian Randomization

Objectives: While Mendelian Randomization (MR) may help infer causal diet-disease links, many diet-related genetic variants (dGVs) are invalid to use in MR, as they violate a core assumption that they have no effect on diet-disease confounders (i.e., pleiotropy). We aimed to develop an approach to refine dGVs to those with "direct" effects on diet; and to compare the accuracy of MR using refined versus unrefined sets of dGVs for predicting known effects of whole grain (WG) intake on LDL-cholesterol (LDL-C).

Methods: Individual-level data were used to systematically test dGVs for statistical evidence of pleiotropy using percent change in variant-diet effects (β_change) when adding adjustment for common diet-disease confounders (smoking, alcohol use, physical activity, income, education, BMI, waist-to-hip). Refined sets included dGVs robust to adjustment, defined by β_change < 20% in a full covariate-adjusted model. We used individual-level data from the UK Biobank (UKB; N ~ 360,000 European adults) and 43 dGVs (P < 5e-8) from a GWAS of white vs. WG bread (i.e., low WG) intake in UKB. We calculated covariate-adjusted effects with linear regression and causal effects of low WG intake on LDL-C (mg/dL) with inverse-variance weighted (IVW) two-sample MR.

Results: In fully-adjusted models, 19 of 43 dGVs (44%) had evidence of pleiotropy. Covariate adjustment identified dGVs with plausible mechanisms of pleiotropy, including variants mapped to known BMI genes (FTO, MC4R, TMEM18) or genes related to socioeconomic status. MR of low WG intake on LDL-C using the unrefined dGV set showed nonsignificant effects with counterintuitive directions (IVW, β [95% CI]: -0.20 [-0.75, 0.34]; P = 0.47). In contrast, MR using the refined dGV set (23 variants) successfully identified significant and directionally expected effects (IVW, β [95% CI]: 0.11 [0.05, 0.18]; P = 5e-4), with less heterogeneity in refined vs. unrefined dGV sets. Applying this approach to other established diet-outcome pairs yielded mixed benefits, potentially due to differences in confounding structure.

Conclusions: MR using a refined dGV set identified known effects of WG intake on LDL-C while MR using an unrefined set did not, indicating the potential value of this covariate adjustment approach to address confounding bias and improve the accuracy of MR of diet-health associations.

Funding Sources: ADA Grant 7-21-JDFM-005