Serum Metabolomic Profiling of All-Cause Mortality: A Prospective Analysis in the Alpha-Tocopherol, Beta-Carotene Cancer Prevention (ATBC) Study Cohort

Abstract

Tobacco use, hypertension, hyperglycemia, overweight, and inactivity are leading causes of overall and cardiovascular disease (CVD) mortality worldwide, yet the relevant metabolic alterations responsible are largely unknown. We conducted a serum metabolomic analysis of 620 men in the Alpha-Tocopherol, Beta-Carotene Cancer Prevention Study (1985–2013). During 28 years of follow-up, there were 435 deaths (197 CVD and 107 cancer). The analysis included 406 known metabolites measured with ultra-high-performance liquid chromatography/mass spectrometry–gas chromatography/mass spectrometry. We used Cox regression to estimate mortality hazard ratios for a 1-standard-deviation difference in metabolite signals. The strongest associations with overall mortality were N-acetylvaline (hazard ratio (HR) = 1.28; P < 4.1 × 10⁻⁵, below Bonferroni statistical threshold) and dimethylglycine, 7-methylguanine, C-glycosyltryptophan, taurocholate, and N-acetyltryptophan (1.23 ≤ HR ≤ 1.32; 5 × 10⁻⁵ ≤ P ≤ 1 × 10⁻⁴). C-Glycosyltryptophan, 7-methylguanine, and 4-androsten-3β,17β-diol disulfate were statistically significantly associated with CVD mortality (1.49 ≤ HR ≤ 1.62, P < 4.1 × 10⁻⁵). No metabolite was associated with cancer mortality, at a false discovery rate of <0.1. Individuals with a 1-standard-deviation higher metabolite risk score had increased all-cause and CVD mortality in the test set (HR = 1.4, P = 0.05; HR = 1.8, P = 0.003, respectively). The several serum metabolites and their composite risk score independently associated with all-cause and CVD mortality may provide potential leads regarding the molecular basis of mortality.

Leading causes of mortality worldwide include tobacco use, overweight, hypertension, hyperglycemia, and physical inactivity (1), with these risk factors contributing to a higher risk of cardiovascular disease (CVD), diabetes, and cancer. However, the underlying biological mechanisms and biochemical actions that could serve as therapeutic or preventive targets are not completely understood. Advances in laboratory technologies of liquid and gas chromatography, mass spectrometry and nuclear magnetic resonance have enabled population-based metabolomic studies to quantify a broader spectrum of low–molecular weight metabolites in biospecimens, including serum. Such metabolomic profiles reflect influence of exogenous and endogenous exposures and, when coupled with health status, may offer insight into biochemical pathways involved in disease pathogenesis and mortality.

Very limited prospective population data exist relating circulating metabolites to overall mortality. One targeted study found that 4 out of 106 plasma biomarkers were associated with all-cause and CVD mortality in the Estonian Biobank and FINRISK study (n = 17,345) (2), but these were primarily large proteins or lipoproteins with known CVD functions (i.e., α-1-acid glycoprotein, albumin, and very-low-density lipoprotein particle size). Another untargeted serum metabolomic analysis identified that 9 out of 204 metabolites (including cotinine, mannose, and γ-glutamyl-leucine) were associated with all-cause mortality among 1,887 African Americans in the Atherosclerosis Risk in Communities (ARIC) Study (3).

To evaluate serum metabolites independently associated with mortality risk, we conducted an untargeted, prospective, serum metabolomic analysis of overall mortality in the Alpha-Tocopherol, Beta-Carotene Cancer Prevention (ATBC) Study cohort, with up to 28 years of outcome follow-up.

METHODS

Study population

Details of ATBC Study have been documented elsewhere (4). Briefly, a total of 29,133 male smokers aged 50–69 years were recruited into the trial from 1985 to 1988 in southwest Finland. Participants were randomly assigned to receive α-tocopheryl-acetate (50 mg), β-carotene (20 mg), both, or placebo daily for 5–8 years. At a baseline, presupplementation visit, fasting blood samples were collected and stored at −70°C; risk-factor questionnaires were completed; and height, weight, and total and high-density lipoprotein cholesterol were measured (4). The ATBC Study was approved by institutional review boards at the US National Cancer Institute and the Finnish National Public Health Institute. All participants provided written informed consent.

The present analysis is based on control participants without cancer (cancer-free at the index date) previously selected for one of 5 metabolomic studies nested within the ATBC Study (Web Figure 1 (available at https://academic.oup.com/aje): metabolomic set 1 (n = 186), metabolomic set 2 (n = 38), metabolomic set 3 (n = 67), metabolomic set 4 (n = 131), and metabolomic set 5 (n = 198)) (5–7). After excluding duplicate samples, our study included 620 men.

Outcome assessment

All-cause mortality, CVD-related mortality, and cancer-related mortality were ascertained through December 31, 2013, using the Causes of Death registry, Statistics Finland. All-cause mortality was defined as death from any cause, CVD mortality was defined where the underlying cause of death was CVD (International Classification of Diseases (ICD), ninth or tenth revision: ICD-9 codes 390–459 or ICD-10 codes I00–I99), and malignant neoplasms as underlying cause of death defined cancer mortality (ICD-9 codes 140–208 or ICD-10 codes C00–C96).

Metabolite assays

Fasting serum metabolites were measured using high-resolution, accurate-mass ultrahigh-performance liquid chromatography/mass spectroscopy and gas chromatography/mass spectroscopy at Metabolon, Inc. (Durham, North Carolina). Methodologic details of sample preparation, quality control, data extraction, and compound identification were described previously (8, 9). Metabolites with fewer than 10 nonmissing values within each metabolomic set were excluded, with 906 metabolites identified in at least one of the 5 metabolomic sets. After further excluding metabolites missing from 2 or more of the metabolomic sets, 406 metabolites remained in the final analysis. Of these, 406 metabolites were categorized into one of 8 chemical classes: amino acids, carbohydrates, cofactors and vitamins, energy metabolites, lipids, nucleotides, peptides, and xenobiotics (Web Table 1). Quality-control samples (9%) were assigned to each batch to evaluate technical reliability, and a coefficient of variation was calculated (median coefficient of variation = 9% (interquartile range, 4–20)) (5–7). In previous studies, we and others have evaluated the within-individual variability of metabolites over time. These studies have found that the median intraclass correlation coefficient of metabolites, based on samples separated by between 4 months and 2 years, was approximately 0.5 (10–12).

Statistical analysis

We batch-normalized each metabolite by dividing by the batch median. Undetected values (missing values) within each metabolite were imputed to the minimum value. The metabolite levels were then processed through log-transformation and normalization. Within each of the 5 data sets, we examined the association between the metabolite level and all-cause, CVD, and cancer mortality using Cox proportional hazard regression, using attained age as the time scale. For Cox regression, among subjects included in the nested case-control studies, start date was the index date at which the individually matched case was diagnosed with cancer. Among subjects included in the evaluation of vitamin supplementation (5), the start date was the baseline enrollment date. We thus removed from our analysis “immortal” person-time (13), which is the person-time during which an event case could not have occurred. In the models, we adjusted for age at blood collection (continuous), body mass index (calculated as weight (kg)/height (m)²; continuous), number of cigarettes per day (continuous), total cholesterol (continuous), high-density lipoprotein cholesterol (continuous), history of hypertension (elevated blood pressure), history of diabetes mellitus, and serum creatinine (continuous). ATBC intervention group (as a categorical variable) was omitted because it was not associated with baseline serum levels. We additionally adjusted for physical activity and dietary factors (total energy intake, fruit intake, vegetable intake, and red meat consumption) as potential confounder factors in the models, and they did not change the effect of any metabolite remarkably, thus they were not included in the final model. We then performed a fixed-effect meta-analysis to obtain single estimated hazard ratios and 95% confidence intervals to describe the association between each metabolite level with all-cause, CVD, and cancer-related mortality. We also fitted crude models for our top metabolite signals that adjusted only for age at blood collection (continuous) in order to evaluate and not overadjust for potential mediators. To account for multiple testing (14), a Bonferroni-corrected threshold of statistical significance was defined as P < 4.1 × 10⁻⁵ (across tests for 406 metabolites and 3 outcomes).

Each metabolomic set was divided into a training set and test set (70% and 30%, respectively). Using only the former, we identified metabolites (false discovery rate (FDR) < 0.1) associated with all-cause mortality (number of metabolites = 12), CVD mortality (n = 12), and cancer mortality (n = 0; no further training-test analysis). In each training set, we performed a Cox regression with all qualifying metabolites and then used fixed-effects meta-analysis to obtain a single set of coefficients. In the test set, we constructed a metabolite risk score, a linear sum of metabolite levels weighted by their corresponding coefficients. Then the metabolite risk score was normalized and used as both a continuous variable (per standard deviation (SD)) and categorized quartiles to estimate the associations with each outcome (all-cause or CVD mortality), using Cox proportional hazard regression (attained age as time scale). All models adjusted for multiple covariates as described above. If the number of covariates exceeded the size of the training set (or the model fit did not converge) for one of the 5 data sets, results from that training set were not included in the meta-analysis.

Pathway analyses assessed the associations between chemical classes and subclasses of metabolites and mortality. For each pathway, we created a single measure of significance, a P value based on Fisher’s statistic (e.g., sum of log-P values) to combine the P values for the score statistics from the Cox regression (after adjustment for multiple covariates). Because of the correlation between metabolites, we calculated the P value based on Fisher’s statistic using a parametric bootstrap (15, 16). For each bootstrap replication, we generated a vector of score test statistics from a multivariate normal distribution with mean 0 and estimated covariance matrix (15). Fisher’s statistic was recalculated for each replication, and the reported P value for each pathway is the proportion of the 10⁵ permutations where the permuted statistic is more extreme than the observed value.

Correlations between top metabolites for the 3 outcomes were estimated using Pearson’s coefficient. Metabolites with r values greater than 0.5 or lower than −0.5 were considered highly positively or negatively correlated, respectively. We used SAS, version 9.3 (SAS Institute, Inc., Cary, North Carolina), and R, version 3.2.3 (R Foundation for Statistical Computing, Vienna, Austria), for all analyses. All reported P values were 2-sided.

RESULTS

Baseline characteristics of the study population in each metabolomic set are presented in Table 1. A total of 620 participants were included from the 5 metabolomic sets, with a follow-up period of up to 28 years (median, 10.7 years (interquartile range, 5.7–18.5 years)), during which there were 435 deaths, including 197 CVD deaths and 107 cancer deaths. We observed no meaningful differences in these characteristics across sets.

Table 1.

Baseline Characteristics of 620 Men in 5 Metabolomic Sets in the Alpha-Tocopherol, Beta-Carotene Cancer Prevention Study, Finland, 1985–2013^a

Characteristic	Metabolomic Set 1 (n = 186)	Metabolomic Set 2 (n = 38)	Metabolomic Set 3 (n = 67)	Metabolomic Set 4 (n = 131)	Metabolomic Set 5 (n = 198)
	Median (IQR)	Median (IQR)	Median (IQR)	Median (IQR)	Median (IQR)
Age at baseline, years	58 (53–61)	60 (56–63)	56 (53–62)	57 (54–61)	59 (55–64)
Follow-up time, years	20.2 (13.7–26.9)	9.0 (5.4–14.3)	6.2 (3.7–9.7)	8.5 (4.6–13.5)	8.4 (3.6–13.5)
No. of deaths
All-cause mortality	128	29	39	94	145
CVD mortality	63	13	16	53	52
Cancer mortality	34	9	10	20	34
BMIb, mean	26.0	26.0	26.3	26.2	26.0
Smoking, cigarettes/day	20 (15–25)	18 (10–20)	20 (15–25)	20 (15–24)	20 (13–25)
Serum HDL cholesterol, mmol/L	1.1 (1.0–1.3)	1.2 (1.0–1.3)	1.1 (1.0–1.4)	1.2 (1.0–1.4)	1.1 (1.0–1.3)
Serum total cholesterol, mmol/L	6.3 (5.6–7.1)	6.3 (5.6–6.8)	6.3 (5.3–7.1)	6.4 (5.6–6.9)	6.2 (5.4–6.8)
History of hypertension, %	16.1	10.5	10.5	21.4	17.7
History of diabetes mellitus, %	4.3	2.6	1.5	3.1	4.0
Physically active, %	24.2	21.1	23.9	24.4	19.2
Dietary intake per day
Total energy, kcal	2,648 (2,231–3,063)	2,532 (2,060–3,122)	2,614 (2,255–3,084)	2,601 (2,222–3,024)	2,598 (2,070–3,066)
Fruit, g	121 (62–200)	131 (72–182)	124 (73–207)	110 (63–185)	98 (57–170)
Vegetables, g	113 (74–162)	116 (58–158)	106 (65–167)	102 (73–149)	98 (66–146)
Red meat, g	68.5 (52.2–95.5)	63.1 (49.6–83.6)	69.6 (49.5–90.7)	68.3 (49.5–88.9)	67.0 (49.0–88.1)

Abbreviations: BMI, body mass index; CVD, cardiovascular disease; HDL, high-density lipoprotein; IQR, interquartile range.

^a Data for continuous variables are shown as median (IQR); otherwise as indicated.

^b Calculated as weight (kg)/height (m)².

Metabolites associated with all-cause mortality

After adjustment for multiple covariates, metabolites associated with all-cause mortality with a FDR of ≤0.05 are shown in Table 2, sorted by P value. The amino acid N-acetylvaline, which achieved statistical significance at the Bonferroni-corrected threshold for multiple tests, along with dimethylglycine, 7-methylguanine, C-glycosyltryptophan, taurocholate, and N-acetyltryptophan, showed the strongest signals, being positively associated with all-cause mortality after adjusting for conventional risk factors (meta-analysis hazard ratios (HRs) per SD metabolite increase were 1.23–1.32, and 2.02 × 10⁻⁵ ≤ P < 1.38 × 10⁻⁴; all subsequent HRs are fully adjusted). The next most significant metabolites related to higher overall mortality included erythronate, 4-androsten-3β,17β-diol disulfate 1, N-acetylmethionine, and 5,6-dihydrothymine (all P < 9 × 10⁻⁴). In addition, increased serum metabolites in long-chain fatty-acid metabolism (e.g., palmitoleate, myristoleate, and docosadienoate), purine metabolism (e.g., N1-methylguanosine and N2,N2-dimethylguanosine), and benzoate metabolism (e.g., 3-hydroxycotinine glucuronide, 3-methyl catechol sulfate 1 and 2) were associated with an increased all-cause mortality (Table 2). Associations for the top metabolite signals and all-cause mortality in the crude models, which adjusted only for age at blood collection, did not differ materially from the multivariable models (data not shown).

Table 2.

Associations Between All-Cause Mortality and Serum Metabolites^a in the Alpha-Tocopherol, Beta-Carotene Cancer Prevention Study, Finland, 1985–2013

Metabolite	Chemical Class	Chemical Subclass	HRb	95% CI	P Value
N-Acetylvalinec	Amino acid	Valine, leucine, and isoleucine metabolism	1.28	1.14, 1.43	2.02 × 10−5
Dimethylglycine	Amino acid	Glycine, serine, and threonine metabolism	1.26	1.13, 1.40	4.56 × 10−5
7-Methylguanine	Nucleotide	Purine metabolism, guanine-containing	1.31	1.15, 1.51	8.77 × 10−5
C-Glycosyltryptophan	Amino acid	Tryptophan metabolism	1.32	1.15, 1.51	9.18 × 10−5
Taurocholate	Lipid	Primary bile acid metabolism	1.23	1.10, 1.36	1.33 × 10−4
N-Acetyltryptophan	Amino acid	Tryptophan metabolism	1.24	1.11, 1.39	1.38 × 10−4
Erythronate	Carbohydrate	Aminosugar metabolism	1.23	1.09, 1.39	6.62 × 10−4
4-Androsten-3β,17β-diol disulfate 1	Lipid	Sterol/steroid	1.21	1.08, 1.36	7.46 × 10−4
N-Acetylmethionine	Amino acid	Cysteine, methionine, SAM, taurine metabolism	1.23	1.09, 1.39	7.62 × 10−4
5,6-Dihydrothymine	Nucleotide	Pyrimidine metabolism, thymine-containing	1.28	1.11, 1.48	9.46 × 10−4
Hexanoylcarnitine	Lipid	Carnitine metabolism	1.20	1.08, 1.34	1.17 × 10−3
Palmitoleate (16:1n7)	Lipid	Long-chain fatty acid	1.20	1.08, 1.35	1.26 × 10−3
5-Dodecenoate (12:1n7)	Lipid	Medium-chain fatty acid	1.20	1.08, 1.35	1.29 × 10−3
N-Acetylphenylalanine	Amino acid	Phenylalanine and tyrosine metabolism	1.20	1.07, 1.34	1.33 × 10−3
Myristoleate (14:1n5)	Lipid	Long-chain fatty acid	1.20	1.07, 1.35	1.37 × 10−3
N1-Methylguanosine	Nucleotide	Purine metabolism, guanine-containing	1.23	1.08, 1.39	1.37 × 10−3
Docosadienoate (22:2n6)	Lipid	Long-chain fatty acid; polyunsaturated fatty acid (n3 and n6)	1.18	1.06, 1.32	2.70 × 10−3
Taurochenodeoxycholate	Lipid	Primary bile acid metabolism	1.22	1.07, 1.39	2.73 × 10−3
Homocitrulline	Amino acid	Urea cycle; arginine and proline metabolism	1.24	1.08, 1.44	2.81 × 10−3
3-Hydroxycotinine glucuronide	Xenobiotics	Tobacco metabolite	1.26	1.08, 1.47	2.86 × 10−3
N-Formylmethionine	Amino acid	Cysteine, methionine, SAM, taurine metabolism	1.20	1.06, 1.36	3.49 × 10−3
3-Methyl catechol sulfate 1	Xenobiotics	Benzoate metabolism	1.18	1.05, 1.32	3.86 × 10−3
4-Acetamidobutanoate	Amino acid	Guanidino and acetamido metabolism; polyamine metabolism	1.23	1.07, 1.41	3.88 × 10−3
γ-Glutamyltryptophan	Peptide	γ-Glutamyl amino acid	1.19	1.06, 1.34	3.96 × 10−3
Asparagine	Amino acid	Alanine and aspartate metabolism	0.86	0.78, 0.96	4.55 × 10−3
N-Acetyltyrosine	Amino acid	Phenylalanine and tyrosine metabolism	1.16	1.05, 1.29	5.19 × 10−3
Xylose	Carbohydrate	Pentose metabolism	1.17	1.05, 1.31	5.31 × 10−3
3-Methyl catechol sulfate 2	Xenobiotics	Benzoate metabolism	1.22	1.06, 1.39	5.43 × 10−3
N2,N2-Dimethylguanosine	Nucleotide	Purine metabolism, guanine-containing	1.20	1.05, 1.36	5.58 × 10−3
3-Hydroxybutyrate (BHBA)	Lipid	Ketone bodies	1.17	1.05, 1.30	5.78 × 10−3
4-Vinylphenol sulfate	Xenobiotics	Benzoate metabolism	1.17	1.05, 1.31	6.20 × 10−3

Abbreviations: CI, confidence intervals; HDL, high-density lipoprotein; HR, hazard ratios; SAM, S-adenosylmethionine; SD, standard deviation.

^a Metabolites were natural log-transformed and standardized (mean = 0, variance = 1). All 620 participants were included in each test, except for metabolites 5,6-dihydrothymine, taurochenodeoxycholate, homocitrulline, and 3-hydroxycotinine glucuronide (missing metabolic data from set 1 (n = 186), thus a total of 434 participants were included in these tests), for metabolites N-acetyltryptophan and 3-methyl catechol sulfate (missing metabolic data from set 2 (n = 38): a total of 582 participants were included in these tests). False discovery rate of ≤0.05.

^b For each study set, we used attained age as time metric in the Cox proportional hazards regression model, and we adjusted for age at baseline, body mass index, number of cigarettes per day, total cholesterol, HDL cholesterol, history of hypertension (elevated blood pressure), history of diabetes mellitus, and serum creatinine. The reported HR (per SD) and P value were obtained from meta-analysis, which was conducted using a fixed-effects model to pool the study sets estimates.

^cN-acetylvaline achieved statistical significance after Bonferroni correction for multiple tests.

We combined the 12 metabolites identified from the training set with an FDR of <0.1 (Table 3) to yield a metabolite risk score and observed an elevated mortality in the test set for men having a higher score (HR = 1.38 per SD, 95% confidence interval: 1.08, 1.75; Table 3). For the categorized metabolite risk score, individuals in the third and fourth quartiles showed 94% and 73% higher mortality, respectively, than those in the lowest quartile (P for trend = 0.05; Table 3).

Table 3.

Hazard Ratios for the Association of Metabolite Risk Score With All-Cause Mortality in the 30% Test Sets in the Alpha-Tocopherol, Beta-Carotene Cancer Prevention Study, Finland, 1985–2013

Quartile of Metabolite Risk Scorea	Event	Participant	Person-years	HRb	95% CI
1	27	44	690.4	1.00	Referent
2	25	41	570.7	1.30	0.71, 2.38
3	36	45	474.7	1.94	1.08, 3.48
4	37	43	400.6	1.73	0.91, 3.30
P value for trend				0.05
Metabolite risk score as continuous variable (per SD)				1.38	1.08, 1.75

Abbreviations: CI, confidence intervals; FDR, false discovery rate; HDL, high-density lipoprotein; HR, hazard ratios; SD, standard-deviation.

^a The risk score was generated by summing the top 12 metabolites (FDR of ≤0.1) identified in the test set, which were weighed by their regression coefficients in the training set. The risk score is calculated as 0.073 × (log-N-acetylvaline) + 0.085 × (log-C-glycosyltryptophan) + 0.07 × (log-dimethylglycine) + 0.084 × (log-N-acetylputrescine) + 0.077 × (log-erythronate) + 0.07 × (log-ADSGEGDFXAEGGGVR) + 0.071 × (log-mannose) + 0.082 × (log-taurochenodeoxycholate) + 0.065 × (log-taurocholate) + 0.095 × (log-5,6-dihydrothymine) + 0.078 × (log-N-acetylmethionine) + 0.085 × (log-acisoga).

^b Hazard ratio for all-cause mortality generated from Cox proportional hazards regression, and adjusted for age at randomization, body mass index, number of cigarettes per day, history of diabetes, serum cholesterol, HDL, history of hypertension (elevated blood pressure), serum creatinine, and metabolomic sets.

Metabolites associated with CVD mortality

Metabolites related to CVD mortality (after adjustment for multiple covariates) with a FDR of ≤0.05 are shown in Table 4, sorted by P value. We found that the following were related to elevated CVD mortality (per SD, HR = 1.38–1.62, and 8.4 × 10⁻⁶ ≤ P < 3.8 × 10⁻⁴): higher serum amino acids C-glycosyltryptophan, 3-(4-hydroxyphenyl)lactate, N-acetylvaline, and dimethylglycine; nucleotide 7-methylguanine; and lipids 4-androsten-3β,17β-diol disulfate, taurocholate, and taurochenodeoxycholate. Of these, P values for C-glycosyltryptophan, 7-methylguanine, and 4-androsten-3β,17β-diol disulfate achieved statistical significance at the Bonferroni-corrected threshold. The next most significant metabolites related to higher CVD mortality included the peptide ADSGEGDFXAEGGGVR, 4-acetamidobutanoate, erythronate, N-acetylphenylalanine, and cortisol, with asparagine being inversely associated (all P < 10⁻³) (Table 4). The crude model estimates that adjusted only for age at blood collection were similar to the multivariable-adjusted associations (data not shown).

Table 4.

Associations of Cardiovascular Disease-Related Mortality and Serum Metabolites^a in the Alpha-Tocopherol, Beta-Carotene Cancer Prevention Study, Finland, 1985–2013

Metabolite	Chemical Class	Chemical Subclass	HR b	95% CI	P Value
C-Glycosyltryptophanc	Amino acid	Tryptophan metabolism	1.60	1.30, 1.96	8.4 × 10−6
7-Methylguaninec	Nucleotide	Purine metabolism, guanine-containing	1.62	1.30, 2.03	1.9 × 10−5
4-Androsten-3β,17β-diol disulfate 1c	Lipid	Sterol/steroid	1.49	1.24, 1.80	2.1 × 10−5
3-(4-Hydroxyphenyl)lactate	Amino acid	Phenylalanine and tyrosine metabolism	1.43	1.20, 1.71	6.4 × 10−5
Taurocholate	Lipid	Primary bile acid metabolism	1.38	1.18, 1.62	7.5 × 10−5
N-Acetylvaline	Amino acid	Valine, leucine, and isoleucine metabolism	1.41	1.19, 1.68	8.6 × 10−5
Taurochenodeoxycholate	Lipid	Primary bile acid metabolism	1.52	1.22, 1.91	2.7 × 10−4
Dimethylglycine	Amino acid	Glycine, serine, and threonine metabolism	1.38	1.15, 1.65	3.8 × 10−4
ADSGEGDFXAEGGGVR	Peptide	Fibrinogen cleavage peptide	1.38	1.15, 1.66	4.4 × 10−4
4-Acetamidobutanoate	Amino acid	Guanidino and acetamido metabolism; polyamine metabolism	1.55	1.21, 1.98	4.6 × 10−4
Erythronate	Carbohydrate	Aminosugar metabolism	1.40	1.15, 1.70	7.6 × 10−4
N-Acetylphenylalanine	Amino acid	Phenylalanine and tyrosine metabolism	1.33	1.13, 1.57	7.9 × 10−4
Cortisol	Lipid	Sterol/steroid	1.38	1.14, 1.68	9.3 × 10−4
Asparagine	Amino acid	Alanine and aspartate metabolism	0.76	0.65, 0.90	9.8 × 10−4
4-Hydroxyhippurate	Xenobiotics	Benzoate metabolism	1.35	1.12, 1.62	1.5 × 10−3
N1-Methylguanosine	Nucleotide	Purine metabolism, guanine-containing	1.39	1.13, 1.71	1.6 × 10−3
5-Dodecenoate (12:1n7)	Lipid	Medium-chain fatty acid	1.31	1.11, 1.56	1.9 × 10−3
2-Hydroxyglutarate	Lipid	Fatty acid, dicarboxylate	1.38	1.13, 1.70	1.9 × 10−3
γ-Glutamyltryptophan	Peptide	γ-Glutamyl amino acid	1.33	1.11, 1.60	2.4 × 10−3
N-Acetyltyrosine	Amino acid	Phenylalanine and tyrosine metabolism	1.29	1.09, 1.52	2.4 × 10−3
N-Acetyltryptophan	Amino acid	Tryptophan metabolism	1.30	1.10, 1.55	2.5 × 10−3
N-Acetylmethionine	Amino acid	Cysteine, methionine, SAM, taurine metabolism	1.38	1.12, 1.69	2.7 × 10−3
Palmitoleate (16:1n7)	Lipid	Long-chain fatty acid	1.29	1.09, 1.53	2.7 × 10−3
5-Methylthioadenosine (MTA)	Amino acid	Polyamine metabolism	1.33	1.10, 1.60	2.8 × 10−3
1-Linoleoylglycerophosphoethanolamine	Lipid	Lysolipid	1.30	1.09, 1.55	3.4 × 10−3
N6-Carbamoylthreonyladenosine	Nucleotide	Purine metabolism, guanine-containing; purine metabolism, adenine-containing	1.39	1.11, 1.73	3.9 × 10−3
Andro steroid monosulfate 1	Lipid	Sterol/steroid	1.29	1.08, 1.53	4.0 × 10−3
4-Androsten-3β,17β-diol disulfate 2	Lipid	Sterol/steroid	1.31	1.09, 1.58	4.5 × 10−3
Mannose	Carbohydrate	Fructose, mannose, galactose, starch, and sucrose metabolism	1.29	1.08, 1.54	4.6 × 10−3
Glycine	Amino acid	Glycine, serine, and threonine metabolism	0.81	0.70, 0.94	4.7 × 10−3
1-Oleoylglycerophosphoethanolamine	Lipid	Lysolipid	1.28	1.07, 1.52	5.4 × 10−3
7-α-hydroxy-3-oxo-4-cholestenoate (7-HOCA)	Lipid	Sterol/steroid	1.27	1.07, 1.50	5.6 × 10−3
Malate	Energy	Krebs cycle/tricarboxylic acid cycle	1.29	1.07, 1.54	6.2 × 10−3
3-Hydroxybutyrate (BHBA)	Lipid	Ketone bodies	1.28	1.07, 1.53	6.4 × 10−3
γ-CEHC glucuronide	Cofactors and vitamins	Tocopherol metabolism	1.26	1.07, 1.49	6.6 × 10−3
Pseudouridine	Nucleotide	Pyrimidine metabolism, uracil-containing	1.30	1.08, 1.58	6.9 × 10−3
Acetoacetate	Lipid	Ketone bodies	1.27	1.06, 1.51	7.6 × 10−3
N6-Acetyllysine	Amino acid	Lysine metabolism	1.25	1.06, 1.48	7.6 × 10−3
Pregnen-diol disulfate	Lipid	Sterol/steroid	1.27	1.06, 1.51	8.1 × 10−3
Myristoleate (14:1n5)	Lipid	Long-chain fatty acid	1.25	1.06, 1.47	8.2 × 10−3
Glycochenodeoxycholate	Lipid	Primary bile acid metabolism	1.24	1.06, 1.46	8.4 × 10−3
Histidine	Amino acid	Histidine metabolism	0.80	0.68, 0.95	8.7 × 10−3
1-Linoleoylglycerophosphoinositol	Lipid	Lysolipid	1.25	1.06, 1.48	8.8 × 10−3

Abbreviations: CI, confidence intervals; CVD, cardiovascular disease; HDL, high-density lipoprotein; HR, hazard ratios; SAM, S-adenosylmethionine; SD, standard-deviation.

^a Metabolites were natural log-transformed and standardized (mean = 0, variance = 1). All 382 participants were included in each test (analysis included referent individuals (n = 185) and individuals with CVD-related death (n = 197)), except for metabolites taurochenodeoxycholate (missing metabolic data from set 1: a total of 261 participants were included in the test) and for metabolites ADSGEGDFXAEGGGVR, 2-hydroxyglutarate, N-acetyltryptophan, and γ-CEHC glucuronide (missing metabolic data from set 2: a total of 360 participants were included in these tests). False discovery rate of ≤0.05.

^b For each study set, we used attained age as time metric in Cox proportional hazards regression models that adjusted for age at baseline, body mass index, number of cigarettes per day, total cholesterol, HDL cholesterol, history of hypertension (elevated blood pressure), history of diabetes mellitus, and serum creatinine. The reported HR (per SD) and P value were obtained from meta-analysis, which was conducted using a fixed-effects model to pool the study sets estimates.

^c Achieved statistical significance after Bonferroni correction for multiple tests.

For the CVD analysis, the metabolite risk score was based on the following 12 metabolites identified in the training set: C-glycosyltryptophan, dimethylglycine, N-acetylvaline, 4-androsten-3β,17β-diol disulfate 1, stearoyl-linoleoyl-GPPE, ADSGEGDFXAEGGGVR, 4-hydroxyhippurate, mannose, 3-hydroxyhippurate, N-acetylputrescine, asparagine, and oleic acid ethanolamide (Table 5). In the test set, we found an increased mortality in men with a higher metabolite risk score (HR = 1.83, 95% confidence interval: 1.28, 2.62) (Table 5). For the categorized metabolite risk score, individuals in the highest quartile experienced over four times the mortality risk when compared with those in the lowest quartile (HR = 4.35, P for trend: 0.003, Table 5).

Table 5.

Hazard Ratios for the Association of Metabolite Risk Score With Cardiovascular Disease-Related Mortality in the 30% Test Sets in the Alpha-Tocopherol, Beta-Carotene Cancer Prevention Study, Finland, 1985–2013

Quartile of Metabolite Risk Scorea	Event	Participant	Person-Years	HRb	95% CI
1	11	32	462.9	1.00	Referent
2	15	32	450.5	2.13	0.87, 5.23
3	15	32	429.1	1.95	0.79, 4.80
4	21	32	265.2	4.35	1.78, 10.6
P value for trend				0.003
Metabolite risk score as continuous variable (per SD)				1.83	1.28, 2.62

Abbreviations: CI, confidence intervals; CVD, cardiovascular disease; HDL, high-density lipoprotein; HR, hazard ratios; SD, standard deviation.

^a The risk score was generated by summing the top 12 metabolites (false discovery rate of ≤0.1) identified in the test set, which were weighed by their regression coefficients in the training set. The risk score is calculated as 0.124 × (log-C-glycosyltryptophan) + 0.112 × (log-dimethylglycine) + 0.109 × (log-N-acetylvaline) + 0.112 × (log-4-Androsten-3β,17β-Diol Disulfate 1) + 0.145 × (log-stearoyl-linoleoyl-GPPE) + 0.117 × (log-ADSGEGDFXAEGGGVR) + 0.113 × (log-4-hydroxyhippurate) + 0.107 × (log-mannose) + 0.111 × (log-3-hydroxyhippurate) + 0.141 × (log-N-acetylputrescine) − 0.1 × (log-asparagine) + 0.177 × (log-oleic acid ethanolamide).

^b Hazard ratio for CVD mortality generated from Cox proportional hazards regression and adjusted for age at randomization, body mass index, number of cigarettes per day, history of diabetes, serum cholesterol, HDL, history of hypertension (elevated blood pressure), serum creatinine, and metabolomic sets.

Metabolites associated with cancer-related mortality

After adjustment for multiple covariates, no metabolite was associated with cancer mortality at either the FDR of ≤0.1 or Bonferroni threshold. Metabolites associated with cancer-related mortality with a nominal P value of <0.05 are shown in Table 6. The amino acids dimethylglycine, N-acetylvaline, levulinate, and N-acetylmethionine were the top metabolites positively associated with cancer-related mortality (per SD, HR = 1.42–1.62, and 6.7 × 10⁻⁴ ≤ P < 6.1 × 10⁻³), as were the tobacco metabolites hydroxycotinine and cotinine N-oxide. By contrast, the amino acids indolepropionate and 3-phenylpropionate and the peptide glutamine-leucine were inversely associated with cancer mortality (per SD, HR = 0.63–0.70, and 8.4 × 10⁻⁴ ≤ P < 6.2 × 10⁻³) (Table 6). As expected, the tobacco metabolites hydroxycotinine and cotinine N-oxide showed stronger associations with cancer-related mortality in the age-adjusted model (per SD, HR = 1.83 and 1.72, P = 6.8 × 10⁻⁵ and 5.0 × 10⁻⁵, respectively). For the other top signals, associations with cancer mortality were not materially different in the crude models (data not shown).

Table 6.

Associations of Cancer-Related Mortality and Serum Metabolites (P < 0.05)^a in the Alpha-Tocopherol, Beta-Carotene Cancer Prevention Study, Finland, 1985–2013

Metabolite	Chemical Class	Chemical Subclass	HRb	95% CI	P Value
Dimethylglycine	Amino acid	Glycine, serine, and threonine metabolism	1.61	1.22, 2.11	6.7 × 10−4
Indolepropionate	Amino acid	Tryptophan metabolism	0.63	0.48, 0.83	8.4 × 10−4
N-Acetylvaline	Amino acid	Valine, leucine, and isoleucine metabolism	1.62	1.22, 2.16	9.4 × 10−4
Levulinate (4-oxovalerate)	Amino acid; xenobiotics	Valine, leucine, and isoleucine metabolism; food component/plant	1.42	1.12, 1.79	3.4 × 10−3
3-Phenylpropionate (hydrocinnamate)	Amino acid	Phenylalanine and tyrosine metabolism	0.70	0.55, 0.89	3.5 × 10−3
N-Acetylmethionine	Amino acid	Cysteine, methionine, SAM, and taurine metabolism	1.60	1.14, 2.24	6.1 × 10−3
Glutamine-leucine	Peptide	Dipeptide	0.70	0.54, 0.90	6.2 × 10−3
Glycocholate sulfate	Lipid	Secondary bile acid metabolism	1.42	1.10, 1.83	7.0 × 10−3
N-Acetylphenylalanine	Amino acid	Phenylalanine and tyrosine metabolism	1.50	1.11, 2.02	7.8 × 10−3
N-Acetyltryptophan	Amino acid	Tryptophan metabolism	1.46	1.10, 1.93	8.9 × 10−3
Hexanoylcarnitine	Lipid	Carnitine metabolism	1.42	1.08, 1.86	1.1 × 10−2
O-Cresol sulfate	Amino acid	Phenylalanine and tyrosine metabolism	1.48	1.09, 2.00	1.1 × 10−2
Ergothioneine	Xenobiotics	Food component/plant	0.74	0.58, 0.94	1.3 × 10−2
Hydroxycotinine	Xenobiotics	Tobacco metabolite	1.54	1.10, 2.18	1.3 × 10−2
Betaine	Amino acid	Glycine, serine, and threonine metabolism	1.44	1.08, 1.93	1.4 × 10−2
Erythronate	Carbohydrate	Aminosugar metabolism	1.45	1.08, 1.95	1.4 × 10−2
Cotinine N-oxide	Xenobiotics	Tobacco metabolite	1.45	1.08, 1.96	1.4 × 10−2
Andro steroid monosulfate 1	Lipid	Sterol/steroid	1.42	1.07, 1.87	1.5 × 10−2
1-Oleoylglycerol (1-monoolein)	Lipid	Monoacylglycerol	1.40	1.07, 1.84	1.5 × 10−2
2-Stearoylglycerophosphoethanolamine	Lipid	Lysolipid	0.77	0.62, 0.95	1.6 × 10−2
4-Vinylphenol sulfate	Xenobiotics	Benzoate metabolism	1.36	1.06, 1.75	1.6 × 10−2
C-Glycosyltryptophan	Amino Acid	Tryptophan metabolism	1.43	1.06, 1.92	1.8 × 10−2
5,6-Dihydrothymine	Nucleotide	Pyrimidine metabolism, thymine-containing	1.53	1.07, 2.19	2.0 × 10−2
2-Ethylphenylsulfate	Xenobiotics	Benzoate metabolism	1.38	1.05, 1.81	2.0 × 10−2
N-(2-Furoyl)glycine	Xenobiotics	Food component/plant	1.31	1.03, 1.68	3.0 × 10−2
Succinate	Energy	Krebs cycle/tricarboxylic acid cycle	1.31	1.03, 1.67	3.1 × 10−2
Homostachydrine	Xenobiotics	Food component/plant	1.30	1.02, 1.65	3.4 × 10−2
Xylose	Carbohydrate	Pentose metabolism	1.29	1.02, 1.64	3.7 × 10−2
Phenylalanylglycine	Peptide	Dipeptide	1.31	1.01, 1.70	4.0 × 10−2
Phenol sulfate	Amino Acid	Phenylalanine and tyrosine metabolism	1.27	1.01, 1.59	4.2 × 10−2
4-Hydroxyhippurate	Xenobiotics	Benzoate metabolism	1.31	1.00, 1.72	4.7 × 10−2
Hydroquinone sulfate	Xenobiotics	Drug	1.32	1.00, 1.73	5.0 × 10−2

Abbreviations: CI, confidence intervals; HDL, high-density lipoprotein; HR, hazard ratios; SAM, S-adenosylmethionine; SD, standard deviation.

^a Metabolites were natural log-transformed and standardized (mean = 0, variance = 1). All 292 participants were included in each test (analysis included referent individuals (n = 185) and individuals with cancer-related death (n = 107)), except for metabolites 2-stearoylglycerophosphoethanolamine and 5,6-dihydrothymine (missing metabolic data from set 1: a total of 200 participants were included in these tests) and for metabolites indolepropionate, glutamine-leucine, N-acetyltryptophan, erythronate, 1-oleoylglycerol (1-monoolein), and homostachydrine (missing metabolic data from set 2: a total of 274 participants were included in these tests).

^b For each study set, we used attained age as time metric in Cox proportional hazards regression models that adjusted for age at baseline, body mass index, number of cigarettes per day, total cholesterol, HDL cholesterol, history of hypertension (elevated blood pressure), history of diabetes mellitus, and serum creatinine. The reported HR (per SD) and P value were obtained from meta-analysis, which was conducted using a fixed-effects model to pool the study sets estimates.

Pathway analysis

In the metabolite pathway analysis, we found that the amino-acid chemical class and primary-bile-acids subclass were significantly associated with all-cause and CVD mortality below the Bonferroni-corrected P value threshold (Web Tables 2–5). After Bonferroni correction, however, no chemical class or subclass was associated with cancer mortality (Web Tables 6 and 7).

The pathway analysis also showed that fructose/mannose/galactose/starch/sucrose metabolism, phenylalanine/tyrosine metabolism, glycine/serine/threonine metabolism, purine/guanine metabolism, tryptophan metabolism, and branched-chain amino acids metabolism were top chemical subclasses associated with all-cause and CVD mortality (Web Tables 3 and 5). The subclasses benzoate metabolism and tobacco metabolism ranked as the top pathways related to all-cause and cancer mortality but not for CVD mortality (Web Tables 3, 5 and 7). Of note, all the models adjusted for multiple mortality risk factors, including smoking. The correlation of the top metabolites for the 3 outcomes are presented in Web Figures 2–4. Higher positive correlations were observed among chemical subclasses of lipid fatty acids (myristoleate (14:1n5), 5-dodecenoate (12:1n7), palmitoleate (16:1n7)), bile acids (taurochenodeoxycholate, taurocholate, glycochenodeoxycholate), sex steroids (pregnen-diol disulfate, 4-androsten-3β,17β-diol disulfate 1, 4-androsten-3β,17β-diol disulfate 2), purine nucleotides (N1-methylguanosine, N2,N2-dimethylguanosine), benzoate xenobiotics (3-methyl catechol sulfate 1, 3-methyl catechol sulfate 2, 4-vinylphenol sulfate), and tobacco xenobiotics (cotinine N-oxide, hydroxycotinine) (Web Figures 2–4). We also show the metabolite-by-metabolite correlation matrix that constituted metabolite risk scores of all-cause and CVD deaths in Web Figures 5 and 6. Among the 12 metabolites included in the risk score for all-cause mortality, the pairwise correlation ranged from −0.05 to 0.84 (Web Figure 5), and among the 12 metabolites included in the risk score for CVD, the correlated ranged from −0.01 to 0.48 (Web Figure 6). The top positive correlation was seen within chemical subclass bile acids (taurochenodeoxycholate, taurocholate) and benzoate xenobiotics (3-hydroxyhippurate and 4-hydroxyhippurate), respectively.

DISCUSSION

In this study, we prospectively investigated associations between >400 serum metabolites and all-cause, CVD, and cancer mortality among 620 men during a median follow-up of 11 years. The all-cause-mortality–related metabolite N-acetylvaline, and CVD-mortality–related metabolites C-glycosyltryptophan, 7-methylguanine, and 4-androsten-3β,17β-diol disulfate yielded the strongest signals that exceeded the multiple-comparisons statistical threshold. By contrast, no metabolite was associated with cancer mortality at an FDR of <0.1. Validating in the test set the risk score that was based on the 12 top metabolites revealed that individuals with higher metabolite scores had elevated risks for all-cause and CVD mortality. Of note, men in the highest risk-score quartile experienced quadrupled CVD mortality compared with those in the lowest score quartile.

Serum N-acetylvaline was a top metabolite signal positively associated with all-cause and CVD mortality (it was also a top signal for cancer mortality). This metabolite is in the branched-chain amino acids, valine/leucine/isoleucine metabolism pathway. Our data also showed that serum N-acetylvaline was strongly correlated with serum valine (r = 0.44; P = 1 × 10⁻³⁰). Branched-chain amino acid metabolites play a role in human health outcomes including cardiovascular disease, stroke, insulin resistance, diabetes, and pancreatic cancer (17–23), and they are associated with obesity and physical activity (24, 25). On the other hand, elevated N-acetyl amino acids, including N-acetylvaline, may indicate disruptions in acetylation activity that could influence cell homeostasis through histone-chromatin function and gene regulation (26–28). Whether some of these factors mediate the increased mortality–higher circulating N-acetylvaline (and other N-acetyl amino acids) associations, or a direct biological action influencing risk of death, will require further study.

Also strongly related to mortality was the tertiary amine dimethylglycine, which can be produced from betaine during the transfer of a methyl group from homocysteine to methionine, a reaction catalyzed by betaine-homocysteine methyltransferase (29). Serum dimethylglycine was significantly correlated with serum betaine in our data (r = 0.26; P = 3 × 10⁻¹¹). Elevated plasma dimethylglycine has been associated with mortality risk, and it may enhance risk prediction of all-cause and CVD-related mortality, particularly among coronary heart disease patients (30). Dimethylglycine was also independently associated with incident acute myocardial infarction and improved outcome prediction among patients with stable angina (31). Regarding cancer, higher fecal dimethylglycine has been related to colorectal cancer in China (32), and urinary dimethylglycine has been correlated with clinical stage of hepatocellular carcinoma in West Africa (33). By contrast, circulating dimethylglycine was unrelated to colorectal cancer or prostate cancer risk in nested case-control studies (34–36).

Higher serum C-glycosyltryptophan and 7-methylguanine were associated with increased overall and CVD mortality (e.g., odds of overall and CVD mortality increased, respectively, by 30% and 60% with each 1-SD log-metabolite increase). C-glycosyltryptophan (also known as C-mannosyltryptophan) is a tryptophan glycoconjugate that has been used as a biomarker of kidney function (37–39) and is related to infectious burden and increased inflammation (40). It is strongly correlated with age and has been related to methylation of the promoter region of WDR85, a gene that may regulate diphthamide synthesis, important for RNA translation, cell cycle, and embryonic development, thereby supporting a role for C-glycosyltryptophan in aging and human development (41). It is possible that its role in aging and inflammation and its relationship to kidney function may partially account for the positive association we observed for overall and CVD mortality. 7-Methylguanine is a by-product of DNA methylation damage repair that is used as a marker of exposure to methylating agents and is potentially related to cancer and aging (42). Tumor tissues exhibit elevated 7-methylguanine levels (43, 44), which might reflect decreased defense against intracellular reactive oxygen species (43–45).

Taurocholate, taurochenodeoxycholate, and the primary bile acid metabolism pathway were also significantly associated with all-cause and CVD mortality. Taurocholate, the bile acid conjugate of cholic acid and taurine, like taurine itself, has been related to lower risks of hypertension, stroke, and other atherosclerotic diseases (46), while hepatic uptake of taurocholate may decline with age (47). Experimental intestinal infusion of taurocholate is related to Akt signaling pathway activation, a possible determinant of cellular senescence (48–50). Other biologically plausible metabolites associated with increased risk of CVD mortality include lower serum histidine and glycine. Histidine may be independently inversely associated with age (51), and it can be metabolized to carnosine, a known antioxidant characterized as an “anti-aging” biochemical based on suppression of oxidative damage, glycation of proteins, and scavenging toxic age-related molecules (52). Glycine is the precursor of several molecular species, including purines and glutathione, and a substantial body of evidence supports its beneficial role in cytoprotection, antioxidation, antiinflammation responses, and metabolic regulation (53–57). Increased CVD mortality was also related to elevated serum mannose, consistent with findings from the Atherosclerosis Risk in Communities study (3), as well as to lysolipids (e.g., 1-linoleoyl-glycerophosphoethanolamine), which are considered important cell-signaling molecules that contribute to regulation of cell differentiation, growth, proliferation, and invasion (58–62), and to steroid hormones in the androgen pathway (e.g., 4-androsten-3β,17β-diol disulfate 1).

Although we identified no metabolites significantly associated with cancer mortality in the model that adjusted for multiple covariates, tobacco metabolism was the top associated pathway, a finding consistent with population studies showing excess cancer mortality is associated with both tobacco smoking and higher circulating cotinine concentrations (63–68). It is noteworthy but not unexpected that the crude models, not adjusting for tobacco smoking, showed strong cancer mortality associations for tobacco metabolites, including cotinine N-oxide and hydroxycotinine.

Several mortality-related metabolites we identified have been associated with chronic aging (41, 51). Considering multiple cellular actions, transformations, and cumulative cellular damage that occur across the life course, with cumulative health deterioration and eventually death, it is biologically plausible that the 2 related but distinct biological traits may be contributed to and regulated by several common molecular functions and biochemical pathways. Also, several of the metabolites we identified are associated with known epidemiologic risk factors related to mortality. These include, for example, physical activity and mannose (24); hypertension risk and tRNA-specific modified nucleoside N2,N2-dimethylguanosine and tricarboxylic-acid-cycle intermediate malate (69); several tobacco smoke-related metabolites (70); type 2 diabetes/hyperglycemia and glycine (71–73), mannose (71, 72), and ketone bodies 3-hydroxybutyrate and acetoacetate (71); and body mass index–related biochemicals asparagine, 3-(4-hydroxyphenyl)lactate, histidine, and glycine (amino acids), mannose (carbohydrate), and hexanoylcarnitine and 7-HOCA (lipids) (74, 75). Discovery of these specific risk factor-associated metabolites in relation to mortality both validates the clinical risk association and affords hypothesis-generating exploration of underlying biological mechanisms for the factor-outcome associations. Further, with regard to potential clinical value, the elucidation of possible underlying biological mechanisms of action for the risk factor–mortality association affords a more precise understanding with potential therapeutic/preventive implications. Second, the metabolomic approach may identify novel biochemicals associated with heretofore unknown risk exposures.

Limitations and strengths of the present study deserve consideration. Even though the study was not large, substantial and highly statistically significant associations were discovered. All participants were Finnish, aged 50–69 years, male, and smokers, which limits generalizability of our findings when considering other populations (e.g., women, younger individuals, and those of other ethnicities). The analysis was restricted to known compounds that were found in at least 4 of 5 study subsets, making it possible that other associations with mortality exist for excluded or unnamed metabolites that were not evaluated. Although all models adjusted for potential confounding factors such as serum creatinine, body mass index, and history of diabetes, it is possible that metabolites-mortality associations were partly mediated by subclinical diseases, such as renal insufficiency, hepatic dysfunction, or insulin resistance. The reported hazard ratios reflect the association between mortality and a single measure of each metabolite; the association with average lifetime levels are likely to be stronger given their documented within-person variability over time (10–12). Important strengths of the study were its prospective nature, with up to 28 years of follow-up, permitting examination of metabolite profiles years prior to the mortality outcomes, and validated mortality ascertainment from national registries that had little or no loss-to-follow-up.

In summary, we identified a panel of circulating metabolites and their composite risk score that were prospectively independently associated with all-cause and CVD-related mortality and substantiated by pathway analyses. The metabolomic traits were related to branched-chain amino acid metabolism, DNA repair, primary bile acid and androgen metabolism, aging, inflammation, and tobacco smoking. Additional prospective investigations in more diverse populations are warranted to reexamine these associations, which, if replicated, will require elucidation of deeper underlying biological mechanisms. Translation to potential therapeutic and preventive targets should also be pursued.

Abbreviations

ATBC

Alpha-Tocopherol, Beta-Carotene Cancer Prevention

CVD

cardiovascular disease

FDR

false discovery rate

HR

hazard ratio

ICD

International Classification of Diseases

SD

standard deviation