A Metabolomic Aging Clock Using Human Cerebrospinal Fluid

Abstract

Quantifying the physiology of aging is essential for improving our understanding of age-related disease and the heterogeneity of healthy aging. Recent studies have shown that, in regression models using "-omic" platforms to predict chronological age, residual variation in predicted age is correlated with health outcomes, and suggest that these "omic clocks" provide measures of biological age. This paper presents predictive models for age using metabolomic profiles of cerebrospinal fluid (CSF) from healthy human subjects and finds that metabolite and lipid data are generally able to predict chronological age within 10 years. We use these models to predict the age of a cohort of subjects with Alzheimer's and Parkinson's disease and find an increase in prediction error, potentially indicating that the relationship between the metabolome and chronological age differs with these diseases. However, evidence is not found to support the hypothesis that our models will consistently overpredict the age of these subjects. In our analysis of control subjects, we find the carnitine shuttle, sucrose, biopterin, vitamin E metabolism, tryptophan, and tyrosine to be the most associated with age. We showcase the potential usefulness of age prediction models in a small data set (n = 85) and discuss techniques for drift correction, missing data imputation, and regularized regression, which can be used to help mitigate the statistical challenges that commonly arise in this setting. To our knowledge, this work presents the first multivariate predictive metabolomic and lipidomic models for age using mass spectrometry analysis of CSF.

Keywords: Aging clock; Biomarker; Cerebrospinal fluid; Metabolomics.

Figure 1.

Distribution of age (top) as well as a comparison of missingness between profiles, split in 3 approximately equal-sized age groupings (bottom). GOT = globally optimized targeted.

Figure 2.

Separation of sex and age using the combined result of PCA and PLS-DA on the data set of combined profiles. The x-axis displays the first principal component from PCA, while the y-axis displays the first PLS component discriminating sex at birth on the space orthogonal to x-axis. For PLS-DA, the first 2 principal components are plotted, with confidence ellipses (assuming t distribution). In parentheses, the axis names contain the percent variation (of the data set) explained by that component. PCA = principal component analysis; PLS-DA = partial least squares discriminant analysis.

Figure 3.

Predicted vs chronological age for controls in each profile, withR2, RMSE, and MAE reported. Numbers in parentheses represent the performance of the mean model for comparison. Points are the average of the predictions for the 5 imputations to estimate missing values, with the error bars representing the most extreme predicted values from the imputations. We also include the y = x line. Points above the line correspond to overestimates of a subject’s chronological age, while points below the line correspond to underestimates. GOT = globally optimized targeted; MAE = mean absolute error; RMSE = root mean squared error.

Figure 4.

The leave-one-out performance of the untargeted model only looking at matched controls (top) and the performance of the model on the AD/PD cohort (bottom). The circled point is the greatest outlier (in terms of RMSE) for PD. Numbers in parentheses refer to the performance of the model which predicts the mean age for each subject. AD = Alzheimer’s disease; MAE = mean absolute error; PD = Parkinson’s disease; RMSE = root mean squared error.

Figure 5.

The marginal linear relationships between chronological age and the targeted metabolites overlapping the tryptophan metabolism pathway from the Small Molecule Pathways Database (top) and the untargeted metabolites identified by Mummichog to be associated with the carnitine shuttle (bottom). The x-axis contains the standardized relationship between metabolite concentration and age, expressed as a t statistic. Boxplots display the distribution of the statistic across untargeted metabolites in the case when Mummichog identifies the proposed compound in more than one metabolite.