Plasma metabolites link dietary patterns to stroke risk

Abstract

Objective:

While dietary intake is linked to stroke risk, surrogate markers that could inform personalized dietary interventions are lacking. We identified metabolites associated with diet patterns and incident stroke in a nested cohort from the REasons for Geographic and Racial Differences in Stroke (REGARDS) study.

Methods:

Levels of 162 metabolites were measured in baseline plasma from stroke cases (n=1,198) and random controls (n=904). We examined associations between metabolites and a plant-based diet pattern previously linked to reduced stroke risk in REGARDS. Secondary analyses included three additional stroke-associated diet patterns: a Mediterranean, Dietary Approaches to Stop Hypertension (DASH), and Southern diet. Metabolites were tested using Cox proportional hazards models with incident stroke as the outcome. Replication was performed in the Jackson Heart Study (JHS). Inverse odds ratio-weighted mediation was used to determine whether metabolites mediated the association between a plant-based diet and stroke risk.

Results:

Metabolites associated with a plant-based diet included the gut metabolite indole-3-propionic acid (β=0.23, 95% CI [0.14, 0.33], p=1.14×10⁻⁶), guanosine (β=−0.13, 95% CI [−0.19, −0.07], p=6.48×10⁻⁵), gluconic acid (β=−0.11, 95% CI [−0.18, −0.04], p=2.06×10⁻³), and C7 carnitine (β=−0.16, 95% CI [−0.24, −0.09], p=4.14×10⁻⁵). All of these metabolites were associated with both additional diet patterns and altered stroke risk. Mediation analyses identified guanosine (32.6% mediation, p=1.51×10⁻³), gluconic acid (35.7%, p=2.28×10⁻³), and C7 carnitine (26.2%, p=1.88×10⁻²) as mediators linking a plant-based diet to reduced stroke risk.

Interpretation:

A subset of diet-related metabolites are associated with risk of stroke. These metabolites could serve as surrogate markers that inform dietary interventions.

Introduction

Dietary patterns are consistently recognized as an important modifiable risk factor for stroke. For instance, a plant-based diet^1,2, the Dietary Approaches to Stop Hypertension (DASH) diet³, and a traditional Mediterranean diet⁴ have each been associated with reduced stroke risk. Conversely, a Southern dietary pattern—characterized by a high intake of added fats, fried foods, processed meat, and sugary beverages—is associated with an elevated risk of stroke². Accordingly, an emphasis on maintaining a healthy diet is an important element of an overall strategy for primary and secondary stroke risk reduction⁵.

Dietary intake can affect host disease states either through the direct absorption of nutrients that impart biological activity, or indirectly through changes in the gut microbiome, which is emerging as an important mediator of human health. Diet plays a pivotal role in shaping the gut microbiome^6,7 and its function, and changes in diet can induce microbial shifts in the short term^8,9. Intestinal microorganisms metabolize a diverse array of diet-derived compounds, in turn generating a variety of products that are released locally and into circulation. Accordingly, gut microbial metabolites are implicated in a wide range of host disease states, including immune function¹⁰, cardiovascular health¹¹, and cognition and memory¹².

Despite the well-documented association between diet patterns and stroke risk, there are no known surrogate markers of stroke risk that reflect dietary patterns or gut microbial composition. The discovery of markers that could guide or track dietary changes would facilitate the development of personalized targets for stroke prevention.

The REasons for Geographic and Racial Differences in Stroke (REGARDS) cohort is a longitudinal study of Black and White adults aimed at evaluating the causes of stroke¹³. The design of REGARDS is suitable for studying the long-term risks or benefits conferred by dietary patterns and for identifying markers that could guide future dietary interventional studies. We previously found metabolites associated with incident ischemic stroke in REGARDS after adjustment for traditional vascular risk factors¹⁴. Because plasma metabolite signatures can be modified by diet¹⁵, we hypothesized that some of these novel metabolite markers could represent a link between dietary intake and stroke risk.

In this study, we aimed to identify metabolites associated with dietary patterns and test whether these metabolites were also associated with incident stroke in REGARDS and an independent observational cohort, the Jackson Heart Study (JHS)¹⁶. We also sought to determine whether specific metabolites were mediators of the altered stroke risk associated with dietary intake. Taken together, our findings highlight the links between diet, circulating metabolites, and stroke risk. These results suggest strategies that could ultimately guide future preventive studies, including specific, individualized dietary recommendations¹⁷.

Materials and Methods:

Study population

The REGARDS study is an ongoing prospective cohort study that enrolled 30,239 non-Hispanic Black and White participants beginning in 2003. The methods and study design have been described in detail elsewhere^13,18. Consented participants provided clinical, demographic, and lifestyle information over the phone at the time of entry into the study. Exclusion criteria were medical conditions that would prevent long-term participation; a history of malignancies or active cancer treatment; living in a nursing home; an inability to communicate in English; and race other than Black or White. Race classification was self-reported during a phone interview. During a subsequent in-person visit 2–3 weeks later, baseline fasting plasma samples were collected¹⁹. EDTA blood samples were stored on ice until centrifuged, and subsequent plasma aliquots were stored in a central laboratory at −80°C until sample processing and measurement of metabolites.

Participants were contacted every 6 months by telephone to ascertain hospitalizations and healthcare encounters for stroke. Medical records were obtained and reviewed centrally by physicians to confirm the diagnosis of stroke; the data review included neuroimaging and other diagnostic records. If the imaging data were unavailable or the records were deemed to be insufficient, a questionnaire was completed using protocols developed for previous stroke clinical trials^20,21 and observational studies²². At least two physician adjudicators reviewed all the available information and classified stroke events. In cases of disagreement, additional adjudicators reviewed the event. A stroke event was recorded if all reviewers agreed on the occurrence of stroke and stroke type (ischemic vs. hemorrhagic). Stroke was defined as a focal neurological deficit lasting >24 hours or non-focal neurological symptoms consistent with ischemic or hemorrhagic stroke on neuroimaging, as defined by the World Health Organization¹⁸.

Stroke cases included all participants who experienced an incident stroke from the time of original enrollment through April 1, 2019. Individuals with a self-reported history stroke or transient ischemic attack (TIA) prior to enrollment were excluded. Participants in the control cohort were selected using stratified random sampling (20 strata based on age, race, and gender) to generate a representative subsample of the entire REGARDS cohort. This was done for efficiency (i.e., measuring biomarkers in the entire cohort of ~30,000 participants is impractical and most likely unnecessary). This unmatched random cohort has been utilized in previous REGARDS studies on stroke risk factors^23–26. All participants provided written informed consent, and institutional review boards at participating institutions approved the study design and methodology.

Targeted metabolomics

A targeted metabolomics approach was selected to measure a predefined panel of metabolites, as previously described^27–30. Briefly, polar metabolites were extracted using protein precipitation from 30 μL of EDTA plasma. Sample extraction was carried out over ice. Isotopic standards were included for quality control monitoring, including proline (¹³C₅, ¹⁵N), glutamine (¹³C₅, ¹⁵N₂), deuterated leucine-d10, and phenylalanine-d8. Isotopic standards were obtained from Cambridge Isotope Laboratories (Tewksbury, MA). The extracted metabolites from the supernatant were separated on Xbridge Amide columns (2.1×100 mm 3.5 μm, Waters, Milford, MA), applying previously described methods using dual Infinity II 1290 HLPC pumps and a 6495 QQQ tandem mass spectrometer (Agilent, Santa Clara, CA). The QQQ mass spectrometer design was optimized to quantify 162 plasma metabolites involved in well-known, biologically important pathways. Human pooled plasma samples were also extracted and injected after every 10 samples for quality control assessments. All peaks were integrated and reviewed using MassHunter QQQ Quantitative Analysis software (Agilent). Following peak integration, metabolites were normalized to the nearest pooled plasma samples using standard approaches^27–30.

Covariates

Demographic covariates, including age, race, and sex, were determined from self-report. Clinical variables included smoking status, systolic blood pressure (SBP), hypertension (HTN), diabetes mellitus (DM), cardiovascular disease (CVD), left ventricular hypertrophy (LVH), and atrial fibrillation (AF). Smoking status was determined by self-report. During the in-home visit, systolic blood pressure (SBP) was measured twice and the average of two measurements was used. HTN was defined as a SBP ≥140 mm Hg, a diastolic blood pressure (DBP) ≥90 mm Hg, or the self-reported use of antihypertensive medications. DM was defined as current use of diabetes medications (e.g., insulin or oral glucose-lowering agents) or a blood glucose concentration of ≥126 mg/dL or ≥200 mg/dL in the fasted and non-fasted states, respectively. Cardiovascular disease (CVD) was defined as a self-reported history of myocardial infarction or coronary revascularization procedure, or evidence of a prior myocardial infarction in a baseline electrocardiogram (ECG). LVH was classified by review of the baseline ECG, and AF was determined from medical history or the presence of AF on the baseline ECG.

Two dietary patterns (plant-based and Southern) were previously derived in the entire REGARDS cohort using exploratory and confirmatory factor analysis on Food Frequency Questionnaire (FFQ) data, as described elsewhere^2,31. Briefly, participant responses to the Block 1998 FFQ were used to assess the mean daily intake of 107 individual food items at baseline. Five factors were derived from these intakes, and each factor was annotated as a dietary pattern based on the food factor loadings that contributed most to that pattern. Factor scores were assigned to participants, which reflected their adherence to each dietary pattern.

Participants were also scored on adherence to other healthy diet patterns derived using simpler scoring systems. For the Mediterranean diet, Block 1998 FFQ data were used to derive a 9-point score as described elsewhere^4,32. Higher scores reflected a greater intake of vegetables, fruits, legumes, cereals, fish, and monounsaturated lipids, along with lower dairy and meat intake and moderate alcohol consumption. As reported previously, FFQ data were also used to derive a DASH diet score based on each participant’s intake of fruits, vegetables, whole grains, nuts/legumes, low-fat dairy, red/processed meats, sweetened beverages, and sodium³³. Subjects with incomplete or implausible FFQ results were not assigned scores for any dietary patterns³¹.

To account for potential variability in metabolomics profiling results based on the sample storage duration³⁴, the length of time between the collection of baseline plasma samples and metabolomics measurements was calculated for all participants.

Statistical methods

Statistical analyses were performed in R version 3.6 and Stata Stata/MP 15.1. In all statistical analyses, the Benjamini-Hochberg procedure was used to control the false discovery rate (FDR), which refers to the expected rate of false positive results when conducting multiple comparisons. Statistical significance was set at FDR<0.05.

Due to the skewed distribution of metabolite levels, the data for each metabolite underwent rank-based inverse normal transformation (INT) prior to analyses¹⁴. Briefly, sample measurements were mapped to the probability scale by replacing the observed values with fractional ranks, then transformed into Z-scores. Sensitivity analyses in which natural log transformation was substituted for INT produced the same overall results for regression analyses described in the Results.

Our primary analysis of diet-metabolite associations focused on the plant-based dietary pattern, which was previously derived and validated within the entire REGARDS cohort³¹. To further evaluate whether metabolites correlated with a plant-based diet were associated with healthy dietary intake, these metabolites were also correlated with scores for the Southern, Mediterranean, and DASH diet patterns. In subsequent analyses, we carried forward metabolites associated with both the plant-based diet and two or more additional diet patterns (FDR<0.05). Since the plant-based diet was associated with reduced stroke risk in not only the entire REGARDS cohort² but also the nested case-control cohort (see Results), this dietary pattern was used for causal mediation analysis.

To account for the stratified random sampling, control participants were assigned survey weights to make statistics computed from the random cohort more representative of the overall REGARDS cohort^23–26. Weighted linear regressions assessing metabolite levels as a function of diet pattern scores were adjusted for age, race, sex, and time to measurement. Factor scores for the plant-based and Southern patterns were z-scored prior to linear regression analyses, while adherence to the Mediterranean and DASH diets was assessed using previously described scoring systems^2,4,35.

To evaluate the effect of diet patterns and metabolites on incident stroke risk, weighted Cox proportional hazards models were adjusted for age, sex, race, time to measurement, and age-by-race interaction, similar to prior studies on stroke in REGARDS^23–26. Stroke risk models were not adjusted for additional vascular stroke risk factors to avoid confounding by conditions associated with dietary intake. Consistent with previous studies, factor scores for the Southern and plant-based diet patterns in Cox models were grouped into weighted quartiles^2,36. As described elsewhere, control participants who developed a stroke during the observation period were weighted as controls prior to the stroke date and cases for an instantaneous period (~0 days) at the time of stroke^14,23–26. P-values for metabolites in Cox regressions from the REGARDS cohort were meta-analyzed with p-values from similar Cox models in an independent replication cohort (see “Replication cohort” section).

Mediation effects were assessed using the inverse odds ratio weighting (IORW) approach³⁷ by treating metabolites as potential mediators between the plant-based pattern and stroke risk. For mediation analysis, plant-based diet was treated as a binary exposure, with weighted quartiles used to assign high (Q3–4) or low (Q1–2) intake. Bootstrapping (n=500) was used to determine 95% confidence intervals for direct and indirect effects. Total and direct effects were estimated via weighted Cox proportional hazards models, and the indirect effect was estimated by subtracting the direct effect from the total effect.

Replication cohort

The Jackson Heart Study (JHS) is a prospective, community-based epidemiological study of cardiovascular disease that enrolled 5,306 Black participants from 2000 to 2004. The study design, recruitment, and data collection for the JHS have been described elsewhere³⁸. Metabolomics data were available for 127 incident stroke cases and 1,825 participants without stroke. Similar to REGARDS, metabolomics measurements in the JHS were performed using targeted LC-MS/MS, as previously reported³⁹. In total, there were 126 metabolites measured in the JHS that were also measured in the REGARDS cohort.

Metabolites found to be correlated with stroke-linked diet patterns in REGARDS were evaluated for their association with incident stroke using Cox proportional hazards models in both REGARDS and JHS. Cox models for JHS were adjusted for age and sex; since all participants were Black, race was not included as a covariate. Given the limited number of incident stroke cases in JHS, we performed a study-level meta-analysis for metabolites shared in common between the REGARDS and JHS cohorts. P-values for each metabolite in Cox models from the REGARDS and JHS cohorts were combined into a single p-value using Fisher’s method for independent tests prior to multiple comparisons correction. Hazard ratios and confidence intervals for metabolites in each cohort were reported separately.

Results:

Study population

The REGARDS cohort enrolled 30,239 community-dwelling participants, and metabolomic analyses were performed on the nested stroke case-control cohort (n=2,373), which included all incident stroke cases through April 1, 2019 and randomly selected controls stratified on age, sex, and race. The levels of 162 prespecified metabolites were measured for each participant in baseline plasma samples. After excluding subjects with a self-reported history of stroke, there were 1,198 incident stroke cases and 904 random controls without incident stroke (Fig 1). Of these participants, dietary data was available for 822 stroke cases and 630 controls. Table 1 shows the baseline characteristics of the cohort, in which individuals who developed a stroke had a higher prevalence of traditional risk factors compared to control participants.

Figure 1:

Participant flow diagram. Baseline metabolomics data were obtained from incident stroke cases and random controls within the REasons for Geographic and Racial Differences in Stroke (REGARDS) study. Participants with a prior stroke history were excluded from the cohort.

Table 1: Baseline characteristics of incident stroke cases and random controls in the REasons for Geographic and Racial Differences in Stroke (REGARDS) study.

Values outside and inside the parentheses represent characteristics of participants with metabolomics data (n=2,102) and the subset with dietary data (n=1,452) respectively. Values for the random controls were weighted to account for stratified sampling. Data for controls with an incident stroke event (n=70) were included in both groups (see Methods).

	Random controls n=904 (n=630)	Incident stroke n=1,198 (n=822)
Age, years ± SD	65±9 (64±9)	69±9 (70±9)
Female sex (%)	45.3% (44.7%)	51.2% (52.4%)
Black race (%)	39.8% (34.7%)	40.6% (32.0%)
Current smoker (%)	13.6% (13.8%)	15.9% (14.9%)
Systolic blood pressure, mm Hg ± SD	127±16 (126±16)	132±17 (132±17)
Hypertension (%)	72.2% (70.6%)	84.8% (82.9%)
Diabetes mellitus (%)	20.5% (17.9%)	28.2% (24.8%)
Cardiovascular disease (%)	15.8% (15.3%)	28.4% (27.6%)
Left ventricular hypertrophy (%)	7.6% (7.5%)	14.9% (14.0%)
Atrial fibrillation (%)	8.8% (8.1%)	13.7% (14.0%)

Metabolites and stroke-associated dietary patterns

We first identified metabolites correlated with a healthy, plant-based dietary pattern previously associated with lower stroke risk in REGARDS². Following Benjamini-Hochberg correction, six metabolites were associated with adherence to a plant-based diet in regression models adjusted for age, sex, race, and the time between sample collection and metabolite measurements (see Methods) (Fig 2A to B). The gut microbial metabolite indole-3-propanoic acid (IPA) was the leading metabolite that was positively correlated with this dietary pattern (β= 0.23, 95% CI [0.14, 0.33], p=1.14×10⁻⁶), as was glyceric acid (β= 0.14, 95% CI [0.06, 0.22], p=4.10×10⁻⁴). In contrast, C7 carnitine (β= −0.16, 95% CI [−0.24, −0.09], p=4.14×10⁻⁵), guanosine (β= −0.13, 95% CI [−0.19, −0.07], p=6.48×10⁻⁵), and gluconic acid (β= −0.11, 95% CI [−0.18, −0.04], p=2.06×10⁻³) were negatively correlated with a plant-based diet.

Figure 2:

Metabolites associated with a plant-based diet pattern. A) β coefficients for each metabolite ranked in ascending order. Selected metabolites correlated with adherence to the plant-based diet are labeled. B) Coefficients and 95% confidence intervals for metabolites correlated with the plant-based diet. A positive correlation coefficient implies a positive association between metabolite abundance and adherence to the diet pattern. Weighted linear regression models assessed normalized metabolite levels as a function of diet factor scores, adjusted for age, race, sex, and time to measurement.

To further validate these findings, we examined associations between a Mediterranean diet or DASH diet and the metabolites correlated with a plant-based diet (Table 2). Similar to the plant-based diet, the Mediterranean and DASH diets are healthy patterns associated with reduced stroke risk^3,4,32,40. However, these patterns were not originally developed and validated within the REGARDS study population and are derived using simpler scoring systems. Of the six candidate metabolites associated with a plant-based diet, five were similarly correlated with adherence to a Mediterranean-style diet. These metabolites included C7 carnitine (β= −0.18, 95% CI [−0.25, −0.10], p=1.16×10⁻⁵), IPA (β= 0.15, 95% CI [0.06, 0.24], p=5.90×10⁻⁴), and guanosine (β= −0.11, 95% CI [−0.18, −0.04], p=3.53×10⁻³). The DASH diet score was likewise associated with all six metabolites, including guanosine (β= −0.17, 95% CI [−0.24, −0.11], p=8.87×10⁻⁷), C7 carnitine (β= −0.18, 95% CI [−0.26, −0.09], p=3.37×10⁻⁵), gluconic acid (β= −0.12, 95% CI [−0.20, −0.04], p=2.11×10⁻³), IPA (β= 0.19, 95% CI [0.09, 0.28], p=9.16×10⁻⁵), and glyceric acid (β= 0.15, 95% CI [0.07, 0.22], p=1.50×10⁻⁴).

Table 2: Metabolites associated with a plant-based dietary pattern.

All metabolite-diet correlations are shown for metabolites associated with the plant-based diet. Multiple hypothesis testing correction accounted for 162 metabolites for the plant-based diet and six metabolites for the three additional diet patterns.

	Plant-based		Mediterranean		DASH		Southern
Metabolite	β coefficient (95% CI)	P-value	β coefficient (95% CI)	P-value	β coefficient (95% CI)	P-value	β coefficient (95% CI)	P-value
IPA	0.23 (0.14, 0.33)	1.14E-06*	0.15 (0.06, 0.24)	5.90E-04*	0.19 (0.09, 0.28)	9.16E-05*	−0.22 (−0.31, −0.14)	5.16E-07*
C7 carnitine	−0.16 (−0.24, −0.09)	4.14E-05*	−0.18 (−0.25, −0.10)	1.16E-05*	−0.18 (−0.26, −0.09)	3.37E-05*	0.08 (−0.01, 0.16)	6.85E-02
Guanosine	−0.13 (−0.19, −0.07)	6.48E-05*	−0.11 (−0.18, −0.04)	3.53E-03*	−0.17 (−0.24, −0.11)	8.87E-07*	0.14 (0.07, 0.21)	1.99E-04*
Glyceric acid	0.14 (0.06, 0.22)	4.10E-04*	0.08 (0.0–0.16)	4.11E-02*	0.15 (0.07, 0.22)	1.50E-04*	−0.13 (−0.21, −0.05)	1.43E-03*
Gluconic acid	−0.11 (−0.18, −0.04)	2.06E-03*	−0.09 (−0.16, −0.01)	2.23E-02*	−0.12 (−0.20, −0.04)	2.11E-03*	0.19 (0.11, 0.27)	2.22E-06*
SAH	−0.11 (−0.18, −0.04)	1.10E-03*	−0.08 (−0.16, 0.00)	5.96E-02	−0.09 (−0.17, −0.01)	2.11E-02*	0.03 (−0.06, 0.12)	4.78E-01

^*False discovery rate (FDR) < 0.05.

Abbreviations: IPA, indole-3-propionic acid; SAH, S-Adenosyl-L-homocysteine.

Although fewer studies have focused on unhealthy dietary patterns, a Southern diet has previously been associated with higher incidence of stroke in REGARDS². Accordingly, we found that four of the metabolites associated with a Southern dietary pattern inversely mirrored the findings with the plant-based pattern (Table 2). Greater adherence to a Southern diet pattern was associated with lower levels of IPA (β= −0.22, 95% CI [−0.31, −0.14], p=5.16×10⁻⁷) and glyceric acid (β= −0.13, 95% CI [−0.21, −0.05], p=1.43×10⁻³). However, Southern diet intake correlated with higher levels of gluconic acid (β= 0.19, 95% CI [0.11, 0.27], p=2.22×10⁻⁶) and guanosine (β= 0.14, 95% CI [0.07, 0.21], p=1.99×10⁻⁴).

In total, five metabolites—IPA, guanosine, gluconic acid, glyceric acid, and C7 carnitine—were associated with both a plant-based dietary pattern and two or more additional stroke-associated dietary patterns. These five diet-associated metabolites (DAMs) were carried forward in subsequent analyses. Notably, the DAMs were not only associated with Southern, Mediterranean, and DASH dietary patterns, but also among the top metabolites associated with each of these patterns (Table S1).

Diet-associated metabolites and incident stroke risk

We next examined whether DAMs were associated with incident stroke. Using Cox proportional hazards models adjusted for age, sex, race, time to measurement, and age-by-race interaction, we found that guanosine (HR=1.44, 95% CI [1.29, 1.60], p=1.51×10⁻¹⁰), gluconic acid (HR=1.28, 95% CI [1.15, 1.42], p=7.14×10⁻⁶), and C7 carnitine (HR=1.11, 95% CI [1.01, 1.22], p=2.89×10⁻²) were linked to elevated stroke risk in REGARDS (Table 3). Guanosine’s association with higher ischemic stroke risk in REGARDS has been reported previously and is known to be partially independent of Framingham stroke risk factors¹⁴.

Table 3: Diet-associated metabolites and incident stroke risk in the REasons for Geographic and Racial Differences in Stroke (REGARDS) and Jackson Heart Study (JHS) cohorts.

Hazard ratios (HRs) represent the HR of incident stroke per standardized unit of the baseline level of each metabolite. Cox models were survey weighted to account for the case-cohort design and p-values were combined using Fisher’s method. REGARDS models were adjusted for age, race, sex, age*race, and time to measurement. JHS models were adjusted for age and sex.

	REGARDS		JHS
Metabolite	HR (95% CI)	P-value	HR (95% CI)	P-value	Combined p-value
Guanosine	1.44 (1.29–1.60)	1.51E-10*	-	-	1.51E-10**
Gluconic acid	1.28 (1.15–1.42)	7.14E-06*	1.15 (0.94–1.39)	1.68E-01	1.75E-05**
C7 carnitine	1.11 (1.01–1.22)	2.89E-02*	1.16 (0.97–1.40)	9.97E-02	1.97E-02**
Glyceric acid	0.91 (0.82–1.01)	6.60E-02	0.86 (0.72–1.03)	9.68E-02	3.87E-02**
Indole-3-propionic acid	0.92 (0.84–1.01)	9.01E-02	0.83 (0.70–0.99)	3.77E-02	2.27E-02**

^*False discovery rate (FDR) < 0.05 within REGARDS.

^**FDR < 0.05 when incorporating combined p-values.

All the DAMs were measured in the Jackson Heart Study (JHS) metabolomics cohort with the exception of guanosine. For the remaining metabolites, we performed a study-level meta-analysis of results from Cox models in the REGARDS and JHS cohorts (see Methods). Gluconic acid remained associated with higher stroke risk following meta-analysis with Cox models in the JHS (JHS HR=1.15, 95% CI [0.94, 1.39]; meta-p=1.75×10⁻⁵). C7 carnitine also remained associated with increased stroke risk following meta-analysis (JHS HR=1.16, 95% CI [0.97, 1.40]; meta-p=1.97×10⁻²). In contrast, IPA was associated with decreased stroke risk after meta-analysis (REGARDS HR=0.92, 95% CI [0.84, 1.01]; JHS HR=0.83, 95% CI [0.70, 0.99]; meta-p=2.27×10⁻²), as was glyceric acid (REGARDS HR=0.91, 95% CI [0.82, 1.01]; JHS HR=0.86, 95% CI [0.72, 1.03]; meta-p=3.87×10⁻²).

Causal mediation analysis

Since all five DAMs correlated with a plant-based diet were also associated with incident stroke—in either the REGARDS cohort or following meta-analysis with the JHS—we next examined whether these metabolites could serve as mediators linking this dietary pattern to altered stroke risk. We first verified the association between adherence to a plant-based diet and reduced incident stroke risk, which has been previously reported in the entire REGARDS cohort², within the nested stroke case-control cohort. The plant-based diet pattern was similarly associated with reduced stroke risk (HR=0.72, 95% CI [0.57, 0.91], p=6.50×10⁻³) in the nested cohort and was therefore suitable for causal mediation analyses (Table 4).

Table 4: Metabolite mediators of the plant-based diet and incident stroke.

Direct and indirect effects were calculated using inverse odds ratio-weighted mediation analysis. Cox regression coefficients are shown non-exponentiated. Listed metabolites were associated with both adherence to the plant-based diet pattern and risk of incident stroke.

Metabolite	Total effect (95% CI)	Direct effect (95% CI)	Indirect effect (95% CI)	% mediation
Indole-3-propionic acid	Quartile 3,4 vs. Quartile 1,2: −0.33 (−0.57, −0.09), p-value=6.50E-03	−0.29 (−0.54, −0.04), p-value=2.30E-02	−0.04 (−0.13, 0.06), p-value=4.33E-01	-
C7 carnitine		−0.24 (−0.49, 0.00), p-value=5.30E-02	−0.09 (−0.16, −0.01), p-value=1.88E-02	26.2%
Guanosine		−0.22 (−0.47, 0.02), p-value=7.70E-02	−0.11 (−0.17, −0.04), p-value=1.51E-03	32.6%
Glyceric acid		−0.29 (−0.54, −0.04), p-value=2.00E-02	−0.04 (−0.11, 0.03), p-value =2.93E-01	-
Gluconic acid		−0.21 (−0.46, 0.03), p-value=9.00E-02	−0.12 (−0.19, −0.04), p-value=2.28E-03	35.7%

We found that guanosine levels partially mediated the inverse association between a plant-based diet and incident stroke risk (32.6% mediation, indirect effect p=1.51×10⁻³), as did gluconic acid (35.7% mediation, indirect effect p=2.28×10⁻³) and C7 carnitine (26.2% mediation, indirect effect p=1.88×10⁻²). However, a significant mediation effect was not observed for either IPA or glyceric acid (Table 4).

Discussion:

In this study, we characterized associations between dietary patterns and circulating metabolites, while examining the significance of these associations in terms of stroke risk. We first identified five key metabolites correlated with multiple diet patterns previously linked to altered stroke risk. A plant-based diet was negatively correlated with metabolites associated with elevated stroke risk (guanosine, gluconic acid, and C7 carnitine) but positively correlated with metabolites associated with reduced stroke risk (IPA and glyceric acid). Similar consistency between the direction of metabolite-diet and metabolite-stroke associations was observed for additional diet patterns associated with lower (DASH, Mediterranean) and higher (Southern) stroke risk. Finally, we found that guanosine, gluconic acid, and C7 carnitine each partly mediated the lower stroke risk associated with adherence to a plant-based diet.

Recent studies have highlighted metabolites associated with dietary intakes, including the DASH⁴¹ and Mediterranean⁴² diet scores. Additional cohort studies have examined metabolite signatures associated with diet patterns derived from dimensionality reduction analyses on FFQ data¹⁶. However, our study is notable for focusing on multiple diet patterns and examining associations specifically with stroke. Taken together, these findings suggest that plasma metabolites can serve as intermediate markers that connect diet patterns to stroke risk.

Metabolomic profiling can offer insight into downstream mechanisms by which diet patterns may relate to stroke risk. These mechanisms might not only converge on known risk factors—such as cardiovascular disease, diabetes mellitus, or hypertension—but also suggest new potentially contributing pathways. For example, because IPA is produced exclusively from tryptophan metabolism by certain human gut microorganisms⁴³, associations between elevated IPA levels and the plant-based, Mediterranean, and DASH diet patterns suggest that a healthy diet may promote the enrichment of commensal gut microbes that protect against stroke. This finding is consistent with previous work suggesting that IPA has a protective role in limiting inflammation, preventing oxidative injury, and reducing cardiovascular disease⁴⁴.

The identification of metabolite mediators potentially linking dietary intake to stroke also raises intriguing questions. Notably, dietary nucleosides are absorbed in the gut and can have a variety of metabolic, immunologic, and microbial effects⁴⁵. While the significance of guanosine in particular remains unclear, some evidence suggests guanosine can contribute directly to endothelial injury⁴⁶. Meanwhile, gluconic acid can stimulate gut bacteria to produce butyrate, a short-chain fatty acid which has important effects on energy balance, inflammation, and immune function⁴⁷. Carnitines—including C7 carnitine—are also known to have numerous functions in metabolic and cardiovascular health⁴⁸.

This study has several strengths, including a large number of incident stroke cases and controls ascertained in a prospective and biracial cohort. Another strength of this study is the targeted tandem mass spectrometry platform for acquiring metabolomics data. While not maximizing the number of metabolites detected, the tandem quadrupole design enabled quantification of a sizable panel of biologically relevant metabolites with higher specificity and sensitivity¹⁴.

However, there are also several limitations. Exploration of potential causal relationships between diet patterns, metabolite levels, and stroke risk was limited by the observational study design. The use of a random control cohort—rather than matched stroke cases and controls—may have also magnified the effect of potential confounders, though a random cohort avoids the risk of overmatching on non-confounding variables. Furthermore, while the results from mediation analyses may suggest common pathways linked to stroke risk, this approach does not imply causality. Finally, there are limitations to self-reported dietary intake (e.g. energy underreporting), which may not accurately reflect true intake⁴⁹.

Although prior work has identified metabolite signatures associated with specific food intakes⁵⁰, our analysis highlights specific metabolites linked to stroke risk. Given the limitations of self-reported diet data, metabolite markers could serve as readouts of adherence to dietary patterns that predict elevated or reduced stroke risk. It is also possible that metabolite markers—which integrate activity in biological pathways modulated by dietary intake—could better capture changes in stroke risk than “upstream” alterations in dietary choices. Further studies are warranted to explore whether insights from metabolomics studies can guide personalized dietary interventions for stroke prevention.

Summary for Social Media:

What is the current knowledge on the topic?

Dietary intake is recognized as an important risk factor for stroke. However, despite the association between diet and stroke risk, there are no known biomarkers of stroke risk that reflect dietary choices.

What question did this study address?

We sought to identify associations between metabolite levels and adherence to dietary patterns previously linked to altered stroke risk. We also explored whether diet-associated metabolites were associated with the future likelihood of stroke.

What does this study add to our knowledge?

We identified metabolites associated with stroke-linked dietary patterns and found that these metabolites were likewise associated with stroke. We determined that several metabolites could serve as intermediate markers connecting diet to stroke risk.

How might this potentially impact on the practice of neurology?

Although prior work has identified metabolites associated with food intakes, our analysis highlights particular diet-associated metabolites linked to stroke risk. These findings could be used to guide personalized dietary interventions.

Draft Tweet:

Bhave et al. identify circulating metabolite markers that may link diet to stroke risk. These surrogate markers could inform personalized dietary interventions.