Incorporating Dietary Information to Enhance Polygenic Prediction Models with Applications to Body Mass Index and Type 2 Diabetes

Polygenic predictors can enhance screening for biomedical conditions, such as metabolism-related traits and diseases, but explain limited phenotypic variance and face implementation challenges in non-European populations. On the other hand, dietary quality and other sociocultural factors are well established metabolic risk factors that remain under-investigated in risk stratification models. In this study, we developed and evaluated risk stratification model combining polygenic predictors and diet-based models for body mass index (BMI) and type 2 diabetes (T2D). Using 5,368 Native Hawaiians from the Multiethnic Cohort (MEC-NH) with genetic data, we integrated large-scale cross-ancestry GWAS summary statistics to develop polygenic score (PGS) models with better prediction accuracies (partial-R2 [SE] = 0.12 [0.04] for BMI; liability-R2 [SE] = 0.09 [0.04] for T2D) than using GWAS information from single ancestry (partial-R2 = 0.07-0.09 for BMI and liability-R2 = 0.05-0.07 for T2D) or in combination with GWAS from MEC-NH (partial-R2 = 0.05-0.06 for BMI and liability-R2 = 0.01-0.05 for T2D). Moreover, machine learning models trained on 520 dietary variables available from 14,346 MEC-NH individuals substantially explained BMI variation (partial-R2 [SE] = 0.12 [0.01]) and enhanced prediction when combined with PGS (adjusted-R2 = 0.29). The best performing diet score model for BMI was associated with multiple chronic diseases in the same cohort, potentially mediated via inflammatory and lipid pathways. Both PGS and dietary scores provided significant, complementary information for predicting BMI and T2D. For populations with limited genetic studies like Native Hawaiians, integrating external GWAS data or non-genetic dietary information can improve risk stratification models.