Integration of epidemiological and blood biomarker analysis links haem iron intake to increased type 2 diabetes risk

Abstract

Dietary haem iron intake is linked to an increased risk of type 2 diabetes (T2D), but the underlying plasma biomarkers are not well understood. We analysed data from 204,615 participants (79% females) in three large US cohorts over up to 36 years, examining the associations between iron intake and T2D risk. We also assessed plasma metabolic biomarkers and metabolomic profiles in subsets of 37,544 (82% females) and 9,024 (84% females) participants, respectively. Here we show that haem iron intake but not non-haem iron is associated with a higher T2D risk, with a multivariable-adjusted hazard ratio of 1.26 (95% confidence interval 1.20-1.33; P for trend <0.001) comparing the highest to the lowest quintiles. Haem iron accounts for significant proportions of the T2D risk linked to unprocessed red meat and specific dietary patterns. Increased haem iron intake correlates with unfavourable plasma profiles of insulinaemia, lipids, inflammation and T2D-linked metabolites. We also identify metabolites, including L-valine and uric acid, potentially mediating the haem iron-T2D relationship, highlighting their pivotal role in T2D pathogenesis.

Extended Data Fig. 1 |. Identification of intermediate metabolites underlying the heme iron-type 2 diabetes association.

Panel (a) presents the four criteria for selecting intermediate metabolites. Panel (b) shows the association between heme iron intake and type 2 diabetes risk with and without adjusting for the potential mediating metabolite. Data are HRs (central points) and 95% CIs (error bars) for T2D per 1 mg/d heme iron intake, estimated by multivariable-adjusted Cox regression. All analyses were conducted among participants with available metabolomics data (n = 9,024).

Extended Data Fig. 2 |. Sensitivity analysis for metabolomics data.

Panel (a) compares associations for metabolites and type 2 diabetes risk among all participants (n = 9024) versus those among controls only (n = 4,795). Panel (b) compares associations for heme iron intake and metabolites among all participants (n = 9024) versus those among controls only (n = 4,795).

Extended Data Fig. 3 |. Flow chart of study population.

The flow chart illustrates the selection process of the study population included in this study.

Fig. 1 |. Study overview.

a, The three sections of statistical analyses designed for the present study. We first examined the association between iron intake and risk of incident T2D from three large prospective cohorts (n = 204,615). Then, we assessed the relationship between iron intake and 12 metabolic biomarkers in a subset of 37,544 participants. Finally, we integrated metabolomics data to investigate the interplay between haem iron intake, metabolites and T2D risk in a subset of 9,024 participants. b, The distribution of age, BMI and sex across these three populations. The centre lines in the boxes of age represent medians, the box heights represent the interquartile range and the dots represent outliers and whiskers represent ranges (excluding outliers). The figure was created with BioRender.com.

Fig. 2 |. Association between haem iron intake and T2D and its attributable proportion in diet–T2D associations.

a, Dose–response relationship for T2D, evaluated by the restricted cubic spline analysis. The reference levels were set to the median haem iron values of the first quintile. The vertical dotted lines indicate median values of each haem iron quintile. The solid line indicates HRs, and the dashed lines depict 95% CI. P for linearity is 6.5 × 10⁻²⁴. b, Subgroup analysis for T2D. The data are HRs (central points) and 95% CIs (error bars) for T2D per 1 mg per day haem iron intake, estimated by multivariable-adjusted Cox regression. A high waist–hip ratio was defined as >0.90 for men and >0.85 for women according to the World Health Organization. All results were stratified by age and calendar year and adjusted for BMI, race, family history of diabetes, menopausal status and post-menopausal hormone use (NHS and NHS2 only), multivitamin use, smoking status, alcohol drinking, physical activity, baseline history of hypertension, baseline history of hypercholesterolaemia, intakes of total energy, magnesium, trans fat and cereal fibre, ratio of polyunsaturated fat intake to saturated fat intake and glycaemic load. Covariates, including subgrouping variables, were updated every 2–4 years. c, Pearson correlation coefficients for haem iron intake and red meat and dietary patterns. d, Associations of a one standard deviation increment (per 1-s.d.) in red meat and dietary patterns with T2D with and without adjusting for haem iron intake (evaluated by multivariable-adjusted Cox regression), as well as the proportion explained by haem iron intake underlying these associations (estimated by mediation analysis). We pooled data from three cohorts for the mediation analysis. All results were stratified by age, calendar year and cohort, and adjusted for BMI, race, family history of diabetes, current post-menopausal hormone use, multivitamin use, smoking status, alcohol drinking, physical activity, baseline history of hypertension, baseline history of hypercholesterolaemia and intake of total energy. The analyses in a–d were conducted in the epidemiological dataset (n = 204,615). All statistical tests were two sided.

Fig. 3 |. Percentage of differences in metabolic biomarker concentrations associated with iron intake.

a, Results for different types of iron intake (n = 37,544), adjusted for age at blood draw, sex, case–control status, fasting status, BMI, race, menopausal status and post-menopausal hormone use (NHS and NHS2 only), multivitamin use (only for analyses of haem iron and dietary iron), aspirin use, smoking status, alcohol drinking, physical activity, baseline history of hypertension, baseline history of hypercholesterolaemia, intakes of total energy, magnesium, trans fat and cereal fibre, ratio of polyunsaturated fat intake to saturated fat intake and glycaemic load. The data are percentage differences (bars) and 95% CIs (error bars) derived from multivariable-adjusted linear regression coefficients. b, Comparison of associations for haem iron intake and biomarkers among all participants (n = 37,544) versus those among controls only (n = 23,873), with the same covariates adjustments as in a. The units for iron intake in the analyses approximate to the difference between the median values of the highest and lowest quintiles. All statistical tests were two sided, and a Bonferroni correction was applied for biomarkers.

Fig. 4 |. Plasma metabolomic biomarkers underlying the association between haem iron intake and T2D.

a, Associations of 50 metabolites selected in the T2D metabolomic score (present in at least 800 out of 1,000 iterations) with 12 metabolic biomarkers (n = 6,915) from multivariable-adjusted linear regression. b, Association between the T2D metabolomic score and risk of developing T2D (n = 9,024). The solid line indicates log HRs estimated by multivariable-adjusted Cox regression, and the dashed lines depict 95% CI. P for linearity is 4.7 × 10⁻⁷¹. c, Association between haem iron intake and T2D metabolomic score (n = 9,024). The data are multivariable-adjusted linear regression coefficients (central points) and 95% CIs (error bars) for standardized T2D metabolomic score comparing higher quintiles of haem iron intake to the lowest. P trend is 6.8 × 10⁻⁸. d, Association between 17 potential mediating metabolites and T2D risk from multivariable-adjusted Cox regression (n = 9,024). e, Association between the potential mediating metabolites and haem iron intake from multivariable-adjusted linear regression (n = 9,024). The height of the bars in d and e indicates the size of beta coefficients. f, Associations of mediating metabolites and 12 metabolic markers from multivariable-adjusted linear regression (n = 6,915). Analyses for 12 metabolic biomarkers and for haem iron-metabolite were adjusted for the same covariates as those in Fig. 2; analyses for metabolite-T2D were adjusted for the same covariates as those in Table 2. All statistical tests were two sided, and a Bonferroni correction was applied for metabolites. CE, cholesterol ester; DAG, diacylglycerol; LPC, lysophosphatidylcholine; LPE, lysophosphatidylethanolamine; PC, phosphatidylcholine; PC-A, phosphatidylcholine-A; PE, phosphatidylethanolamine; SM, sphingomyelin.