Diet scores boost genetic prediction of BMI and diabetes in Native Hawaiians

Combining genetic risk scores with machine-learning-derived dietary scores substantially improves the prediction of body mass index in Native Hawaiians, according to a new study published in Genes & Nutrition. Researchers working with the Multiethnic Cohort found that a model integrating both predictors explained 29% of the variation in BMI, compared with 21% for either the polygenic score or the diet score alone. For type 2 diabetes, the combined model offered only a marginal improvement over genetics alone, but the work still demonstrates, in quantitative terms, how much predictive power a modifiable environmental factor such as diet can add to genomic risk stratification in a population that has long been underserved by genetic research.

The study focused on Native Hawaiians, a population with a high prevalence of obesity and type 2 diabetes but limited representation in genome-wide association studies. Because most large genetic studies have been conducted in people of European ancestry, polygenic scores tend to perform poorly when transferred to other populations, a well-documented problem that has hindered clinical implementation. The team, led by researchers including Charleston W. K. Chiang and Song-Yi Park, set out to test whether dietary information, collected in detail over decades in the Multiethnic Cohort, could complement genetic prediction in this setting.

The cohort includes 14,346 Native Hawaiian participants with dietary data from a baseline questionnaire administered between 1993 and 1996, of whom 5,374 were genotyped. The dietary assessment was unusually rich: a 26-page questionnaire centered on a quantitative food frequency questionnaire with more than 180 food items, from which the researchers computed a total of 520 dietary variables covering nutrients, individual foods, food groups, and dietary pattern scores. A calibration study showed good validity of the questionnaire against repeated 24-hour dietary recalls.

On the genetic side, the team built cross-ancestry polygenic scores using PRS-CSx, a Bayesian method that jointly models GWAS summary statistics from multiple populations while accounting for population-specific patterns of linkage disequilibrium. For BMI, they combined summary statistics from East Asian-ancestry studies of roughly 256,450 individuals, European-ancestry studies of about 700,000 individuals, and an in-house GWAS of 4,148 Native Hawaiians genotyped on the Illumina MEGA array. Performance was evaluated in a held-out validation set of 1,226 Native Hawaiians genotyped on a different array, the Global Diversity Array.

The best-performing BMI score combined only the East Asian and European GWAS data, achieving a partial R² of 0.12 in the held-out validation set, meaning it explained about 12% of variation in sex-normalized, age-adjusted BMI. That outperformed single-population models, which explained between roughly 3% and 9%. Notably, adding the Native Hawaiian GWAS to the mix actually hurt performance, likely because the in-house GWAS, with only about 4,000 participants, was underpowered and introduced noise. A similar pattern held for type 2 diabetes, where the East Asian plus European model reached a liability R² of about 0.09, compared with 0.01 to 0.07 for single-population and other cross-ancestry models.

The dietary scores were built with a deliberately data-driven strategy. The researchers applied three machine-learning algorithms—elastic net regularization, random forest, and neural networks—to all 520 dietary variables, training the models in a dietary training set of 8,372 Native Hawaiians who had survey data but no genetic data, and evaluating them in a held-out validation set. For both BMI and type 2 diabetes, the random forest model performed best. A one standard deviation increase in the BMI diet score was associated with a 0.38 standard deviation increase in BMI, and the score explained approximately 12% of BMI variability after adjusting for daily energy intake—essentially the same amount of variance captured by the polygenic score.

When the two predictors were combined in a joint regression model among participants with both genetic and dietary data, the results were striking for BMI. The model containing both scores achieved an adjusted R² of 0.29, versus 0.21 for models with either score alone, a difference that was highly statistically significant. The partial R² of each predictor barely diminished in the joint model, indicating that the genetic and dietary scores capture complementary, nearly additive information. No significant interaction between the two scores was detected. Sensitivity analyses treating the diet score as quartiles yielded consistent results.

For type 2 diabetes, the picture was more sobering. The polygenic score was clearly the stronger predictor, with an odds ratio of 1.71 per standard deviation and an AUC of 0.73, while the diet score alone yielded an odds ratio of 1.17 and an AUC of 0.70. The combined model improved the fit statistically, but the gain was marginal and, by the authors’ own assessment, likely negligible in practice. The diet score for diabetes explained only around 1% of liability-scale variance. The authors suggest several reasons: dietary patterns change over time, baseline diet is a poor predictor of disease developing years later, type 2 diabetes classification may involve misclassification, and people with prediabetic conditions may have already shifted toward healthier eating, confounding the diet–disease relationship.

The BMI diet score also turned out to be associated with more than adiposity. In participants with genetic data who were not used to develop the score, each standard deviation increase was linked to higher odds of developing type 2 diabetes (21%), stroke (20%), chronic kidney disease (20%), ischemic heart disease (17%), and anemia (12%), based on Medicare claims linkages. Causal mediation analysis showed that BMI itself mediated 32% to 52% of these associations—52% for type 2 diabetes, 36% for ischemic heart disease, 34% for anemia, and 32% for chronic kidney disease. Stroke was the exception, with a non-significant indirect effect through BMI, suggesting its association with the diet score operates through other pathways. Substantial direct effects of the diet score remained after accounting for BMI across all five outcomes.

To probe those BMI-independent pathways, the researchers examined four biomarkers measured in a subset of roughly 1,900 participants: C-reactive protein, LDL, HDL, and triglycerides. After correcting for 24 tests, no mediation effects reached the Bonferroni significance threshold of 0.0021, though suggestive trends emerged, including possible roles for CRP in type 2 diabetes and HDL in chronic kidney disease and stroke. Triglycerides partially mediated the diet score’s effect on incident type 2 diabetes, accounting for about 13% of the total effect. The authors caution that these biomarkers come from a single cross-sectional measurement, cannot establish causality, and are available for only a fraction of the cohort, so longitudinal studies are needed.

The study has other important limitations. The diet score models depend on having the same set of dietary variables available, which limits portability: when the researchers attempted to harmonize the Multiethnic Cohort’s dietary phenotypes with those of the UK Biobank, variables matched only about 30% to 57% of the time depending on how strictly they were interpreted. Dietary patterns are also likely specific to each ethnicity, so transferability to other groups remains untested. And because dietary information was collected in adulthood, reverse causation—people changing their diets in response to early, undiagnosed disease—cannot be ruled out, although the BMI diet score did predict BMI measured roughly ten years after baseline.

Even so, the implications are considerable. For populations with limited genetic data, the study shows that leveraging large-scale GWAS data from other ancestries can substantially improve polygenic score performance, and that diet information explains a meaningful share of phenotypic variation that the best available genetic predictors miss. The authors also note a practical benefit for genetic research itself: adjusting for diet scores in GWAS of relatively small, underrepresented cohorts could remove non-genetic variance and improve statistical power. More broadly, the findings argue that personalized metabolic risk assessment should integrate both genetic and dietary information, with dietary patterns tailored to the population being served—a framework the researchers say can, in principle, be extended to other populations and other traits.

Subject of Research: Integrating dietary and genetic data to improve prediction of BMI and type 2 diabetes in Native Hawaiians

Article Title: Incorporating dietary information to enhance polygenic prediction models with applications to body mass index and type 2 diabetes

Article References: Lee, E. Y., Dinh, B. L., Tang, J., Streicher, S. A., Wang, X., Biswas, S., Tian, H., Yu, X., Taparra, K., Naseri, T., Viali, S., Weeks, D. E., Carlson, J. C., Haiman, C. A., Marchand, L. L., Maskarinec, G., Wilkens, L. R., Park, S.-Y., & Chiang, C. W. K. (2026). Incorporating dietary information to enhance polygenic prediction models with applications to body mass index and type 2 diabetes. Genes & Nutrition, 21(1), Article 20. https://doi.org/10.1186/s12263-026-00811-1

Image Credits: AI Generated

DOI: 10.1186/s12263-026-00811-1

Keywords: polygenic scores, dietary scores, body mass index, type 2 diabetes, Native Hawaiians, Multiethnic Cohort, machine learning, PRS-CSx, risk prediction, mediation analysis, genetic ancestry, precision medicine