A multi-metabolite signature robustly predicts long-term mortality in the PREDIMED trial and several US cohorts

Abstract

Metabolome-based biomarkers contribute to identify mechanisms of disease and to a better understanding of overall mortality. In a long-term follow-up subsample (n = 1878) of the PREDIMED trial, among 337 candidate baseline plasma metabolites repeatedly assessed at baseline and after 1 year, 38 plasma metabolites were identified as predictors of all-cause mortality. Gamma-amino-butyric acid (GABA), homoarginine, serine, creatine, 1-methylnicotinamide and a set of sphingomyelins, plasmalogens, phosphatidylethanolamines and cholesterol esters were inversely associated with all-cause mortality, whereas plasma dimethylguanidino valeric acid (DMGV), choline, short and long-chain acylcarnitines, 4-acetamidobutanoate, pseudouridine, 7-methylguanine, N6-acetyllysine, phenylacetylglutamine and creatinine were associated with higher mortality. The multi-metabolite signature created as a linear combination of these selected metabolites, also showed a strong association with all-cause mortality using plasma samples collected at 1-year follow-up in PREDIMED. This association was subsequently confirmed in 4 independent American cohorts, validating the signature as a consistent predictor of all-cause mortality across diverse populations.

Keywords: All-cause mortality; Biomarkers; Metabolomic; Metabolomic signature; Plasma metabolites.

Fig. 1.

Scheme of the study development. Using a combination of two PREDIMED case-cohorts as the main data, the association between baseline plasma metabolites and all-cause mortality was addressed through metabolome-wide association approaches and also feature selection by an elastic net penalized Cox regression model, which make it possible to create a metabolomic signature (score) using these selected metabolites, separating training sets from testing sets, with special care to avoid any overfitting. Further analysis was implemented using an unbiased metabolic signature, evaluating the prediction of all-cause mortality with Cox hazard multivariable models. Additionally, an assessment of this metabolomic signature in relation to premature mortality was conducted using rate advancement period (RAP) and survival analyses. Internal validation of a plasma 38-multimetabolite score was conducted using another sample of PREDIMED (PREDIMED specimens collected after 1 year of follow-up instead of baseline). External validations were implemented in American cohorts, using 13 combined studies of Nurses’ Health Studies I and II and Health Professionals Follow-Up Study (HPFS), and a case-control study of coronary heart disease nested in 2 combined cohorts from the Women’s Health Initiative. For Elastic Net Penalized Cox Regression, α and λ are fitting parameters to control the L1-norm and L2-norm penalization of the model, R(t_i): risk set at time t_i. For Metabolomic Signature, x_i: abundance of each metabolite; β_i: weight of each metabolite to the metabolomic score, n: number of selected metabolites. PREDIMED: PREvención con DIeta MEDiterránea; NHS-I: Nurseś Health Study I; NHS-II: Nurseś Health Study II; HPFS: Health Professionals Follow-Up Study.

Fig. 2.

Boxplots of baseline continuous variables in each extreme quintile of the metabolomic score (first row) and stratified by sex (second row).

Fig. 3.

Associations between each metabolite and mortality. a. Volcano plot for the individual associations between each metabolite and long-term all-cause mortality; The top 40 associations (20 inverse and 20 direct hazard ratios) of individual metabolites with mortality and their respective HRs (95 % CI) per +1 SD increment are shown in the upper left graph and the upper right table. b. Scatter plot to compare the results of the beta coefficients [log(HR)] of Cox models for each metabolite by sex. Direct and inverse significant associations for both male and female (FDR < 0.05) with individual metabolite and all-cause mortality are represented with red and blue, respectively c. d. & e. Scatters plot for log(HR) of all-cause mortality vs. log(HR) of each cause of death (cancer, CVD and other causes). Direct and inverse significant associations for each cause of death (FDR < 0.05) with individual metabolite and all-cause mortality are represented in red (direct) and blue (inverse). All p-values were corrected using the Benjamini & Hochberg FDR < 0.05 correction; In the upper right table, we indicate the non-included or non-available metabolites: ^˟not selected for the 38-multi-metabolomic signature associated with all-cause mortality; *not available in NHS-I, NHS-II and HPFS; **not available in WHI, ***not available in NHS-I, NHS-II, HPFS and WHI. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 4.

Multivariable-adjusted model showing independent associations between baseline variables, including the plasma LOFO-multi-metabolite score, and all-cause mortality, adjusted for the conventional risk factors shown in the table and graph. Hazard ratio (HR) for quintiles of the LOFO-multi-metabolite score, adjusted for all the variables shown in the graph: age, sex, baseline glucose level (per 20 mg/dL), waist-to-height ratio (3 categories: <0.6, 0.6 to <0.75 and ≥0.75), smoking (3 categories: never, former and current), alcohol intake (3 categories: <5 g/d, 5 to 15 g/d and ≥15 g/d), hypertension [HT] at baseline (yes or no), educational level (low or high), family history of premature coronary heart disease [CHD] (yes or no), dyslipidemia diagnosed at baseline (yes or no), body mass index [BMI] (3 categories:<30, 30–35, ≥ 35 kg/m²), total energy intake (per 500 kcal/d) and leisure-time physical activity (per 100 MET-min/d). In addition, the shown model was also adjusted for baseline adherence to the Mediterranean diet adherence screener (0–14 points) and for the randomized group (3 categories: olive oil, nuts and low fat). HR: Hazard ratio; CI: Confidence interval. CHD: Coronary heart disease.

Fig. 5.

Cumulative all-cause mortality by quintiles of the baseline LOFO-multi-metabolite score. Long-term mortality rates in PREDIMED by quintiles (Q1 to Q5) of the baseline multi-metabolomic score, adjusted for potential confounders (the same variables shown in the footnote of Table 2) using inverse probability weighting (IPW). Quintiles 2 and 3 were merged because they mostly overlapped.