Interplay of genetic predisposition, plasma metabolome and Mediterranean diet in dementia risk and cognitive function

Abstract

Alzheimer's disease (AD) and AD-related dementias (AD/ADRD) have a substantial genetic basis, with APOE4 homozygotes increasingly recognized as a distinct genetic subtype. To identify genotype-specific metabolic pathways and modifiable risk factors, we integrated genetic, plasma metabolomic and dietary data from 4,215 women and 1,490 men in prospective cohorts. Here we show that the associations of 57 metabolites with dementia risk varied by APOE4 genotype or other AD/ADRD risk variants. For example, cholesteryl esters and sphingomyelins were most strongly associated with increased dementia risk in APOE4 homozygotes, whereas inverse associations with glycerides were specific to this genotype. Dimethylguanidino-valeric acid was more strongly associated with dementia risk among carriers of the rs2154481-C allele (APP). Adherence to the Mediterranean diet more effectively modulated dementia-related metabolites in APOE4 homozygotes, suggesting targeted prevention strategies. Incorporating metabolomic data modestly improved dementia risk prediction, particularly during early follow-up. Mendelian randomization analysis identified 19 putative causal relationships between metabolites and cognitive outcomes, including protective effects of 4-guanidinobutanoate, carotenoids and N⁶-carbamoylthreonyladenosine. These findings reveal genotype-dependent metabolic profiles of cognitive health and support precision nutrition approaches for ADRD prevention.

Fig. 1. Prospective cohort studies examining the interrelationship of genetics, plasma metabolomics, MedDiet, cognitive function and dementia risk.

a, Prospective follow-up of 4,215 women in the NHS from 1989 to 2023. Genetic and metabolomic profiles were generated from blood samples collected at baseline. Detailed demographic, lifestyle, dietary, medical history and medication use data were collected via questionnaires. Dementia cases were ascertained through the follow-up as a composite endpoint of incident dementia and death due to dementia. In addition, a telephone-based neuropsychological assessment battery was administered longitudinally from 1995 to 2008 to assess cognitive function in a subset of 1,037 participants. A total of 1,490 men from the HPFS were included as a replication cohort (Extended Data Fig. 1a). b, Distribution of plasma metabolites (n = 401). The outer circle represents the variation of each metabolite, with a gradient in gray indicating the coefficient of variation. The inner circle displays the mean relative abundance of each metabolite, shown as a gradient in blue. The innermost circle color codes represent the different HMDB superclasses defined based on chemical structural similarities. c, Overall genetic structure associated with individual metabolites. Each dot represents an individual and is colored by APOE4 genotype, showing no clear pattern between the overall population substructure and APOE4 genotype. The metabolites with the highest Pearson’s correlations with the top two genetic PCs from each metabolite superclass are included on the plot as arrows, colored by their superclass (see legend for b). The arrowhead coordinates represent the correlation coefficients of the metabolites with genetic PC1 and PC2. d, Associations between established genetic risk factors for AD/ADRD and dementia risk. The lines indicate cumulative incidence across APOE4 genotypes and tertiles of the PRS of ADRD (excluding the APOE region) over the follow-up period, with shaded areas representing 95% CIs and P values from the log-rank test annotated. Consistent with the curves, unadjusted hazard ratios (HRs) were estimated using Cox proportional hazards (PH) model; covariate-adjusted HRs with 95% confidence intervals (CIs) are provided in Supplementary Table 4. Person time was accrued from baseline until the earliest occurrence of an incident dementia case, dementia death or the end of follow-up. No adjustment was made for multiple comparisons, because this was a hypothesis-driven analysis. e, A wide range of adherence to the MedDiet, as assessed by a dietary index and intake levels of food and nutrient components of MedDiet. All analyses and distributions were based on data from 4,215 NHS participants. All statistical tests were two sided. MAG, monoacylglycerol; TAG, triacylglycerol. Panel a created using BioRender.com. Source data

Fig. 2. Associations of plasma metabolomic profiles, dementia risk and cognitive function differ according to individuals’ genetic predisposition to AD/ADRD.

a, Significant variation in the association between metabolites and dementia risk across different genotypes. Left, in the two heatmaps, the color gradient denoting the HR for dementia risk per 1-s.d. increment in metabolite levels among individuals with different genetic predispositions, as defined by APOE4 genotype or PRS of ADRD (including the APOE variants), estimated using Cox PH model. Only metabolites with FDR < 0.05 for their interaction terms with genotype are displayed in the heatmaps. Right, the color gradient in the two heatmaps representing the product of the β coefficient and the −log₁₀(FDR) of the interaction term between each metabolite and genotype, as defined by the APOE4 genotype or other common AD/ADRD genetic variants, from Cox PH model. In all heatmaps, associations or interactions with FDR < 0.05 are indicated by double asterisks (**) and those with FDR < 0.25 by a single asterisk (*). Metabolites are grouped according to HMDB superclasses. The analyses were conducted among 4,215 NHS participants. b, Gene–metabolite interactions related to dementia risk widely distributed across metabolite superclasses and genotypes. The Manhattan plot displays metabolome-wide interaction results, represented by −log₁₀(FDR) values for interaction terms from Cox PH models. Each dot represents a metabolite colored by the direction of interaction and grouped by HMDB superclass. Top, for the APOE4 genotype, the data point with the lower FDR between heterozygous (diamond) and homozygous (square) APOE4 interactions included for each metabolite. Bottom, for other common AD/ADRD variants, the most significant interaction across all 73 variants shown for each metabolite. The analyses were conducted among 4,215 NHS participants. c, Selected associations between metabolites and dementia risk with FDR for interaction <0.05, stratified by genotype. The first row presents stratified HRs and 95% CIs for dementia risk per 1-s.d. increment in metabolite level, categorized by the APOE4 genotype, with FDR values for interaction terms between APOE4 heterozygosity and homozygosity annotated (using the noncarrier as the reference group). The second row displays stratified results (HR and 95% CIs per 1-s.d. increment in metabolite level) by AD/ADRD variants, with FDR values for interaction terms with the variant effect allele dosage annotated. Genotype groups were defined based on rounded allele dosages. Results for the rs1800978 GG genotype group are excluded due to data sparsity. The analyses were conducted across 4,215 NHS participants. d, Consistency of metabolite–APOE4 interaction results across models with dementia risk and cognitive function as dependent variables. Each dot represents a metabolite with significant APOE4 interactions, colored by the HMDB superclass. Pearson’s correlation coefficients in the β coefficients for interaction terms between metabolites and APOE4 carrier status estimated from Cox PH models, with dementia as the dependent variable, and from generalized linear models, with cognitive function scores as the dependent variable, are annotated on each figure. APOE4 carriers were not further divided into heterozygotes and homozygotes due to data sparsity among homozygotes with non-missing values for each metabolite in the cognitive function subset. Dementia risk analyses were conducted among 4,215 NHS participants and cognitive function analyses among 1,037 NHS participants. All statistical tests were two sided. DAG, diacylglycerol; PC, phosphatidylcholine; PE, phosphatidylethanolamine; TAG, triacylglycerol. Source data

Fig. 3. MedDiet adherence is associated with cognitive health and plasma metabolites in an APOE4-dependent manner.

a, Higher adherence to the MedDiet prospectively associated with a lower risk of dementia and enhanced cognitive performance, as assessed by the telephone-based neuropsychological assessment battery (TICS). For dementia risk analysis, a restricted cubic spline Cox PH model estimated HRs and 95% CIs across varying levels of the MedDiet index, using 0 as the reference. The P value from a likelihood ratio test comparing the model without the MedDiet index and the model with its spline term is annotated. For the TICS score analysis, a generalized linear model estimated the adjusted TICS score and corresponding 95% CI across MedDiet index levels, with P values annotated. The analyses were conducted among NHS participants with cognitive and dietary data (n = 86,740 for dementia analysis and n = 16,244 for cognitive function analysis). b, The protective association between adherence to the MedDiet and risk of dementia most pronounced among APOE4 homozygotes. Stratified HR and 95% CIs for dementia risk per a 1-unit increment in the MedDiet index score, categorized by the APOE4 genotype, were estimated from Cox PH models, with stratified P values annotated (unadjusted for multiple comparisons in the hypothesis-driven analysis). The analyses were conducted among NHS participants with genetic, dietary and dementia outcome data (n = 16,497). c, Strong association between adherence to MedDiet and the overall plasma metabolome from an RF model to classify individuals in the top versus the bottom quartile of the MedDiet index based on plasma metabolites. For the RF classification, the dataset was randomly divided into training (60%) and test (40%) sets. The ROC curve for the test set is shown, with the AUC and 95% CI annotated on the plot. The analyses were conducted among 4,215 NHS participants. d, Associations between MedDiet adherence and plasma metabolite levels differing by APOE4 genotype. The heatmap shows β coefficients representing a 1-s.d. increment in the MedDiet index from a generalized linear model, with plasma metabolite levels as the dependent variable. The analyses were conducted among 4,215 NHS participants. e, Select associations between the MedDiet index and plasma metabolite levels with P < 0.05 from the likelihood ratio test for the interaction between APOE4 genotype and MedDiet index in relation to metabolites, using a generalized linear model stratified by APOE4 genotype. Covariate-adjusted residuals of metabolites are shown along with fitted linear regression lines, 95% CIs and P values for interaction. These results were not adjusted for multiple testing. The analyses were conducted across 4,215 NHS participants. All statistical tests were two sided. Source data

Fig. 4. Integrating genetic variation with plasma metabolites and MedDiet enhances the prediction of dementia risk and cognitive status.

a, The inclusion of genetic factors improving dementia risk prediction using Cox PH model, with an additional modest enhancement when plasma metabolites also included. Time-dependent ROC curve analyses were conducted for dementia risk over both the entire follow-up period and the first 15 years of follow-up. The baseline model predictors included age, family history of dementia, education level, smoking status, history of depression or regular antidepressant use and MedDiet index. The PRS of ADRD excluded variants in the APOE region (see Methods for selection of metabolite predictors). b, Plasma metabolites among the top contributors for predicting dementia risk as quantified by the SHAP value. Feature contributions were evaluated for Cox PH model to predict overall and 15-year dementia risk, including the full list of predictors. SHAP values were calculated for each category of predictors by summing the SHAP value of all predictors in that category. Features were ranked by the SHAP value from the highest to the lowest for predicting the overall and 15-year dementia risk. c, Integration of genetic and metabolomic data enhancing cognitive status prediction within APOE4 subgroups. The heatmap displays AUCs from an RF model classifying participants in the highest versus the lowest tertile of the overall TICS score. In subgroup analyses by APOE4 carrier status, APOE4 genotype was excluded as a predictor. For all analyses, the NHS dataset (n = 4,215) was randomly divided into training (60%) and test (40%) sets; models were fitted on the training set and evaluated on the test set. All results shown are from the test set. Source data

Fig. 5. Genetics enables the inference of putative causal relationships between plasma metabolites and cognitive outcomes.

a, Schematic of the two-sample MR and colocalization analyses. Genetic instruments were selected for 657 metabolites and 133 metabolite ratios from a published GWAS. Summary statistics for overall dementia, AD, vascular dementia and cognitive performance were also obtained from published GWASs (Methods). Two-sample MR was performed to identify putative causal relationships between metabolites or ratios and cognitive outcomes, followed by colocalization analysis for the causal associations with an FDR < 0.05. b, Identification of numerous putative causal interrelationships of various metabolites, metabolite ratios and cognitive outcome using genetic instruments. The chord diagram displays causal relationships with an FDR < 0.05 from MR analyses using Wald ratio, inverse variance-weighted or MR Egger methods, where each arc represents an identified link between a metabolite or ratio and a cognitive outcome. Arcs and nodes are color coded by the HMDB superclasses. c, Colocalization analyses strengthening evidence of causality, suggesting that the identified putative causal relationships between metabolites or ratios and cognitive outcomes have potential shared causal variants and biology. Putative causal relationships, represented by the odds ratios (ORs) or β coefficients with 95% CIs, are shown for associations with FDR < 0.05 and colocalization signals. Bayesian colocalization analysis was performed within the ±500-kb region around a genetic instrument (Methods). If multiple instruments were used, a causal association was reported if the metabolite or ratio and cognitive outcomes colocalized at least one genetic locus. Colocalization signals were reported for a locus if the conditional probability of colocalization, PP.H4/(PP.H3 + PP.H4), was >70%, where PP.H3 is posterior probability that the two traits have independent causal variants and PP.H4 is the posterior probability that the two traits share a single causal variant. d, Regional genetic association plots providing evidence of potential shared causal variants affecting both metabolites or ratios and cognitive outcomes at specific genetic loci. The plots display genetic association results for metabolites or ratios and cognitive outcomes at three colocalized loci with PP.H4/(PP.H3 + PP.H4) > 70%. Each plot is annotated with the genetic instrument and dots are color coded according to their linkage disequilibrium with the instrumental variant. The −log₁₀(P) values for both metabolites or ratios and cognitive outcomes were obtained from the original GWASs. The sample sizes for the original GWASs are as follows: metabolites or ratios (n = 8,299), cognitive performance (n = 257,841), dementia (5,933 cases and 166,584 controls), AD (90,338 cases and 1,036,225 controls) and vascular dementia (881 cases and 211,508 controls) (Methods). All statistical tests were two sided. Panel a created using BioRender.com. Source data

Extended Data Fig. 1. Overview of the Health Professionals Follow-Up Study as a replication cohort.

(a) We prospectively followed 1,490 men in Health Professionals Follow-Up Study (HPFS) from 1993 through 2023. Genetic and metabolomic profiles were generated from blood samples collected at baseline. Detailed demographic, lifestyle, dietary, medical history and medication use data were collected via questionnaires. Dementia cases were ascertained through the follow-up as a composite endpoint of incident dementia and death due to dementia. (b) Associations between established genetic risk factors for Alzheimer’s disease and related dementias (ADRD) and dementia risk. Lines indicate cumulative incidence across APOE4 genotypes and tertiles of polygenic risk score (PRS) of ADRD (excluding the APOE region) over the follow-up period, with shaded areas representing 95% confidence intervals (CIs) and P from the log-rank test annotated. Consistent with the curves, unadjusted hazard ratios (HRs) were estimated using Cox proportional hazards models; covariate-adjusted HRs with 95% CIs are provided in Supplementary Table 4. Person-time was accrued from baseline until the earliest occurrence of an incident dementia case, dementia death or the end of follow-up. No adjustment was made for multiple comparisons, as this was a hypothesis-driven analysis. The analyses were conducted among 1,490 HPFS participants. All statistical tests were two-sided. Panel a created using BioRender.com. Source data

Extended Data Fig. 2. Associations between genetic factors and 401 metabolites in Nurses’ Health Study.

(a) Correlations between genetic principal components (PCs) and metabolites: PC1 (top panel) and PC2 (bottom panel). The dashed lines denote false discovery rate (FDR) = 0.05, adjusted for multiple comparisons, based on Pearson’s correlation tests. (b) Associations of APOE4 heterozygosity (top panel) and homozygosity (bottom panel) with metabolites, with APOE4 noncarrier as the reference group from Cox proportional hazards models. Models were adjusted for age, the top 4 genetic PCs, and genotyping platforms. The dashed lines denote nominal P = 0.05 (none with FDR < 0.05). All analyses were conducted among 4,215 NHS participants. All statistical tests were two-sided. Source data

Extended Data Fig. 3. Number of metabolites with significant interactions with genetic factors in relation to dementia risk in Nurses’ Health Study by Human Metabolome Database superclass.

(a) Number of metabolites with significant interactions with APOE4 genotype. (b) Number of metabolites with significant interactions with other common AD/ADRD variants. Bars indicate the number of associations that reached nominal significance (P < 0.05) based on the Cox proportional hazards models, and those that remained significant after adjustment for multiple comparisons (false discovery rate <0.05). All statistical tests were two-sided. Source data

Extended Data Fig. 4. Consistency of metabolite-APOE4 interactions in Nurses’ Health Study and Health Professionals Follow-Up Study.

Each dot represents a metabolite with significant APOE4 interactions in Nurses’ Health Study (NHS; n = 38 out of 49 available in Health Professionals Follow-Up Study [HPFS]), colored by the Human Metabolome Database superclass (see Fig. 2 for legend). Pearson’s correlation coefficients in the β coefficients for interaction terms between metabolites and APOE4 carrier status estimated from Cox proportional hazards models with dementia as the dependent variable in NHS and HPFS is annotated on the figure. APOE4 carriers were not further stratified into heterozygotes and homozygotes due to data sparsity among homozygotes with non-missing values for each metabolite in HPFS. Analyses were conducted among 4,215 NHS and 1,490 HPFS participants. All statistical tests were two-sided. Source data

Extended Data Fig. 5. Associations of Mediterranean diet index score with dementia risk and cognitive function by genetic factors in Nurses’ Health Study.

(a) Associations between Mediterranean diet (MedDiet) index score and cognitive function scores by APOE4 genotype from generalized linear models. (b) Associations between MedDiet index score and dementia risk by tertiles of ADRD PRSs from Cox proportional hazards models. (c) Associations between MedDiet index score and cognitive function scores by tertiles of ADRD PRSs from generalized linear models. Stratified β coefficients (for cognitive function) and hazard ratios (HRs; for dementia risk), along with 95% confidence intervals (CIs) per one-unit increase in the MedDiet index score, are shown by APOE4 genotype or PRS categories. The analyses were conducted among NHS participants with genetic, dietary, and cognitive outcome data (n = 16,497 for dementia analyses and n = 3,770 for cognitive function analyses). Source data

Extended Data Fig. 6. Replication of the associations of Mediterranean diet index score with plasma metabolites and dementia risk by APOE4 genotype in Health Professionals Follow-Up Study.

(a) Higher adherence to the Mediterranean diet (MedDiet) was prospectively associated with a lower risk of dementia. A restricted cubic spline Cox proportional hazards (PH) model estimated hazard ratios (HRs) and 95% confidence intervals (CIs) across varying levels of the MedDiet index. The P value from the likelihood ratio test comparing the model without the MedDiet index and the model with its spline term is annotated. This analysis was conducted among Health Professionals Follow-Up Study (HPFS) participants with dementia and dietary data (n = 43,500). (b) The protective association between adherence to the MedDiet and risk of dementia was most pronounced among APOE4 homozygotes. Stratified HR and 95% CIs for dementia risk per one-unit increment in the MedDiet index score, categorized by APOE4 genotype, were estimated from Cox PH models, with stratified P annotated (unadjusted for multiple comparisons in the hypothesis-driven analysis). The analyses were conducted among HPFS participants with genetic, dietary, and dementia outcome data (n = 9,828). (c) Strong association between adherence to MedDiet and the overall plasma metabolome profile identified by a random forest (RF) model to classify individuals in the top versus bottom quartile of the MedDiet index based on plasma metabolites. For the RF classification, the dataset was randomly divided into training (60%) and test (40%) sets. The receiver operating characteristic curve for the test set is shown, with the area under the curve (AUC) and 95% CI annotated on the plot. The analyses were conducted among 1,490 HPFS participants. (d) Associations between MedDiet adherence and plasma metabolite levels are consistent in NHS and HPFS and differ by APOE4 genotype. The heatmap shows β coefficients representing a 1-s.d. increment in the MedDiet index from a generalized linear model, with plasma metabolite levels as the dependent variable, restricting to the 254 metabolites available in both cohorts. In the metabolite analysis, APOE4 carriers were not further stratified into heterozygotes and homozygotes due to data sparsity among homozygotes with non-missing values for each metabolite in HPFS. The analyses were conducted among 4,215 NHS and 1,490 HPFS participants. All statistical tests were two-sided. Source data

Extended Data Fig. 7. Correlation coefficients between principal components of metabolites and the Mediterranean diet index, along with its individual components in Nurses’ Health Study.

Metabolite principal component (PC) plot with correlation coefficients between metabolite PC1 and PC2 and Mediterranean diet (MedDiet) score and its components. Each dot represents an individual and is colored by the MedDiet score. MedDiet score and its components are shown on the plot as arrows; the coordinates of the arrow heads represent their Pearson’s correlation coefficients with metabolite PC1 and PC2. The analyses were conducted among 4,215 participants in Nurses’ Health Study. Source data

Extended Data Fig. 8. Associations of glycerides with the Mediterranean diet index in Nurses’ Health Study.

Each dot represents a metabolite (diglycerides on the left panel and triglycerides on the right panel), color-coded by the direction and significance level of its association with the Mediterranean diet index. The size of each dot represents the effect size of the association. The P value is unadjusted for multiple comparisons given the exploratory nature of the analysis. The analyses were conducted among 4,215 NHS participants. All statistical tests were two-sided. Source data

Extended Data Fig. 9. Harrell’s C index for predicting dementia risk from models with different predictors in Nurses’ Health Study.

APOE4 genotype was not included as a predictor in the subgroup analysis within APOE4 carrier and noncarrier subgroups. The full NHS dataset (n = 4,215) was randomly split into training (60%) and test (40%) sets. All prediction models were fitted in the training set and were evaluated in the test set. All results shown are from the test set.

Extended Data Fig. 10. Integrating genetic variation with plasma metabolites and Mediterranean diet enhances the prediction of dementia risk in the Health Professionals Follow-Up Study.

(a) The inclusion of genetic factors improved dementia risk prediction using Cox proportional hazards (PH) model, with an additional modest enhancement when plasma metabolites were also included. The baseline model predictors included age, family history of dementia, profession, smoking status, history of depression or regular antidepressant use, and Mediterranean diet (MedDiet) index. The polygenic risk score (PRS) of Alzheimer’s disease and related dementias excluded variants in the APOE region. (b) Plasma metabolites are among the top contributors for predicting dementia risk as quantified by the Shapley Additive Explanations (SHAP) value. Feature contributions were evaluated for the Cox PH model for predicting dementia risk including the full list of predictors. SHAP values were calculated for each category of predictors by summing the SHAP value of all predictors in that category. Features were ranked by the SHAP value from the highest to the lowest for predicting dementia risk. We did not assess 15-year dementia risk in the Health Professionals Follow-Up Study (HPFS) due to the limited number of dementia cases. For all analyses, the HPFS dataset (n = 1,490) was randomly divided into training (60%) and test (40%) sets; models were fitted on the training set and evaluated on the test set. All results shown are from the test set. Source data