Metabolite profiling and cardiovascular event risk: a prospective study of 3 population-based cohorts

Abstract

Background: High-throughput profiling of circulating metabolites may improve cardiovascular risk prediction over established risk factors.

Methods and results: We applied quantitative nuclear magnetic resonance metabolomics to identify the biomarkers for incident cardiovascular disease during long-term follow-up. Biomarker discovery was conducted in the National Finnish FINRISK study (n=7256; 800 events). Replication and incremental risk prediction was assessed in the Southall and Brent Revisited (SABRE) study (n=2622; 573 events) and British Women's Health and Heart Study (n=3563; 368 events). In targeted analyses of 68 lipids and metabolites, 33 measures were associated with incident cardiovascular events at P<0.0007 after adjusting for age, sex, blood pressure, smoking, diabetes mellitus, and medication. When further adjusting for routine lipids, 4 metabolites were associated with future cardiovascular events in meta-analyses: higher serum phenylalanine (hazard ratio per standard deviation, 1.18; 95% confidence interval, 1.12-1.24; P=4×10(-10)) and monounsaturated fatty acid levels (1.17; 1.11-1.24; P=1×10(-8)) were associated with increased cardiovascular risk, while higher omega-6 fatty acids (0.89; 0.84-0.94; P=6×10(-5)) and docosahexaenoic acid levels (0.90; 0.86-0.95; P=5×10(-5)) were associated with lower risk. A risk score incorporating these 4 biomarkers was derived in FINRISK. Risk prediction estimates were more accurate in the 2 validation cohorts (relative integrated discrimination improvement, 8.8% and 4.3%), albeit discrimination was not enhanced. Risk classification was particularly improved for persons in the 5% to 10% risk range (net reclassification, 27.1% and 15.5%). Biomarker associations were further corroborated with mass spectrometry in FINRISK (n=671) and the Framingham Offspring Study (n=2289).

Conclusions: Metabolite profiling in large prospective cohorts identified phenylalanine, monounsaturated fatty acids, and polyunsaturated fatty acids as biomarkers for cardiovascular risk. This study substantiates the value of high-throughput metabolomics for biomarker discovery and improved risk assessment.

Keywords: amino acids; biological markers; fatty acids; metabolomics; risk factors.

Figure 1

Overview of the study design and statistical analyses conducted.

Figure 2

Metabolite associations with future cardiovascular events. Hazard ratios of 68 metabolite measures with incident cardiovascular disease during 15-year follow-up in the FINRISK study (n=7256, 800 events). Hazard ratios are per 1-SD log-transformed metabolite concentration and adjusted for age, sex, blood pressure, smoking, diabetes, geographical region, and cardiovascular medication. Errorbars denote 95% confidence intervals. ★: P<0.0007 (multiple testing correction).

Figure 3

Meta-analysis of metabolite biomarkers for future cardiovascular events. Hazard ratios of biomarkers with incident cardiovascular events in three population-based studies and meta-analysis (n=13441; 1741 events during 12–23 years follow-up). Analyses are adjusted for age, sex, blood pressure, smoking, diabetes, geographical region, cardiovascular medication as well as total and HDL cholesterol. Hazard ratios are per 1-SD log-transformed metabolite concentration and errorbars denote 95% confidence intervals. I² indicates heterogeneity of meta-analysis. Metabolites associated with cardiovascular events at P<0.0007 (multiple testing correction) are shown here; associations with P<0.05 are listed in Supplemental Figure 6. The four biomarkers highlighted in bold are independent of each other and were included in the risk prediction score.

Figure 4

Phenylalanine interaction with age. Hazard ratios of phenylalanine with incident cardiovascular disease in different baseline age groups for the FINRISK study. Analyses were adjusted for age, sex, blood pressure, smoking, diabetes, geographical region, cardiovascular medication as well as total and HDL cholesterol. Hazard ratios are per 1-SD higher log-transformed phenylalanine concentration (approximately corresponding to 14 μmol/L). Errorbars denote 95% confidence intervals. The dashed line denotes the hazard ratio for the full age range. The continuous interaction of phenylalanine with age is shown for all three cohorts in Supplemental Figure 8.

Figure 5

Phenylalanine associations with vascular mortality. Hazard ratios of phenylalanine with death from cardiovascular disease, coronary heart disease and stroke in the three cohorts and meta-analysis (n=13815). Analyses were adjusted for age, sex, blood pressure, smoking, diabetes, geographical region, cardiovascular medication as well as total and HDL cholesterol. Hazard ratios are per 1-SD higher log-transformed phenylalanine concentration (approximately corresponding to 14 μmol/L). Errorbars denote 95% confidence intervals. The numbers of coronary and stroke deaths sum to a slightly higher number than the cardiovascular deaths, since we have considered all causes of death written in the death certificate.

Figure 6

Consistency between NMR and LC-MS for biomarker associations with incident cardiovascular disease. Metabolites overlapping between metabolomics platforms and nominally associated with incident CVD in the present study (P<0.05, Supplemental Figure 6) are shown. Left panel: Biomarker associations with CVD risk observed in the present study based on NMR (white diamonds) compared with those obtained in the Framingham Offspring Study based on LC-MS (red squares; n=2289, 466 events). Right panel: Biomarker associations with CVD risk in a case-cohort subset of the FINRISK study (n=679, 305 events) profiled both by NMR (black circles) and LC-MS (red circles). Hazard ratios are per 1-SD higher log-transformed metabolite concentration. Errorbars denote 95% confidence intervals. All associations were adjusted for age, sex, blood pressure, smoking, diabetes status, geographical region, cardiovascular medication as well as total and HDL cholesterol. LC-MS-based associations were further adjusted for batch. The corresponding age- and sex-adjusted biomarker associations are shown in Supplemental Figure 10. *Associations of omega-6 fatty acids were compared with lysophosphatidylcholine- and cholesterolester-linoleic acid in the Framingham Offspring Study, and with total linoleic acid in the FINRISK subset. † Associations of DHA were compared with lysophosphatidylcholine- and cholesterolester- DHA in the Framingham Offspring Study. ^¶ DHA ratio was scaled to the total fatty acid concentration quantified by NMR for both platforms. The DHA ratio was not measured in the Framingham Offspring Study.