Metabolic correlates of prevalent mild cognitive impairment and Alzheimer's disease in adults with Down syndrome

Abstract

Introduction: Disruption of metabolic function is a recognized feature of late onset Alzheimer's disease (LOAD). We sought to determine whether similar metabolic pathways are implicated in adults with Down syndrome (DS) who have increased risk for Alzheimer's disease (AD).

Methods: We examined peripheral blood from 292 participants with DS who completed baseline assessments in the Alzheimer's Biomarkers Consortium-Down Syndrome (ABC-DS) using untargeted mass spectrometry (MS). Our sample included 38 individuals who met consensus criteria for AD (DS-AD), 43 who met criteria for mild cognitive impairment (DS-MCI), and 211 who were cognitively unaffected and stable (CS).

Results: We measured relative abundance of 8,805 features using MS and 180 putative metabolites were differentially expressed (DE) among the groups at false discovery rate-corrected q< 0.05. From the DE features, a nine-feature classifier model classified the CS and DS-AD groups with receiver operating characteristic area under the curve (ROC AUC) of 0.86 and a two-feature model classified the DS-MCI and DS-AD groups with ROC AUC of 0.88. Metabolite set enrichment analysis across the three groups suggested alterations in fatty acid and carbohydrate metabolism.

Discussion: Our results reveal metabolic alterations in DS-AD that are similar to those seen in LOAD. The pattern of results in this cross-sectional DS cohort suggests a dynamic time course of metabolic dysregulation which evolves with clinical progression from non-demented, to MCI, to AD. Metabolomic markers may be useful for staging progression of DS-AD.

Keywords: Alzheimer's disease; Down syndrome; carbohydrate metabolism; energy metabolism; fatty acid metabolism; lipid metabolism; metabolism; metabolomics; mild cognitive impairment.

FIGURE 1

Differentially expressed metabolite features. Volcano plots showing differential expression (DE) of individual features for each of the three comparisons: cognitively unaffected and stable (CS) versus Down syndrome‐mild cognitive impairment (DS‐MCI; A), DS‐MCI versus Down syndrome‐Alzheimer's disease (DS‐AD; B), and CS versus DS‐AD (C). We enforced false discovery rate (FDR) q < 0.05, but no fold change criterion for DE. There were no DE features for the CS versus DS‐MCI comparison, 17 DE features for the DS‐MCI versus DS‐AD comparison, and 163 DE features for the CS versus DS‐AD comparison. The red horizontal line represents the cut‐off for FDR and red circles represent DE features in each plot

FIGURE 2

Features selected by the machine learning algorithms. This figure shows the group distributions of the nine features selected by the least absolute shrinkage selection operator (LASSO) feature selection algorithm for the cognitively unaffected and stable (CS) versus Down syndrome‐Alzheimer's disease (DS‐AD) comparison (A) and the five features selected by the support vector machine (SVM) for the Down syndrome‐mild cognitive impairment (DS‐MCI) versus DS‐AD comparison. The boxplots show the distribution of metabolite abundances for each group with each participant represented as a solid circle. The solid line in each box represents the median while the lower and upper boundaries of the box reflect the first and third quartiles. The whiskers reflect the minimum and maximum values. The horizontal red line in each panel represents the optimum cut‐off for sensitivity and specificity in a univariate receiver operating characteristic area under the curve (ROC AUC). Panels with red outlines are the metabolites definitively identified by MS/MS and are listed by name in Table 2. Panels without red outlines could not be definitively identified by MS/MS

FIGURE 3

Classification performance using putative metabolites. Receiver operating characteristic area under the curve (ROC AUC) for the classification models using the nine unidentified features for the cognitively unaffected and stable (CS) versus Down syndrome‐Alzheimer's disease (DS‐AD) comparison (A) and the five features for Down syndrome‐mild cognitive impairment (DS‐MCI) versus DS‐AD comparison (B). For the CS versus DS‐AD comparison, the left panel shows strong classification using a logistic regression model with 10‐fold cross validation (ROC AUC = 0.868), the middle panel shows similar performance for the same model using a more rigorous 100‐fold Monte Carlo cross validation procedure (ROC AUC = 0.855), and the right panel shows consistent classification performance using an alternate support vector machine (SVM) classification algorithm (ROC AUC = 0.859). In the DS‐AD versus DS‐MCI comparison (B) the left panel shows strong classification performance using the logistic regression model with 10‐fold cross validation (ROC AUC = 0.891), the middle panel shows similar performance with 10‐fold Monte Carlo resampling approach (ROC AUC = 0.881) and the right panel shows strong SVM performance (ROC AUC = 0.885)

FIGURE 4

Classification performance using definitively identified metabolites. These receiver operating characteristic (ROC) plots show the classification model performance using the seven MS/MS definitively identified metabolites for the cognitively unaffected and stable (CS) versus Down syndrome‐Alzheimer's disease (DS‐AD) comparison (A) and the two definitively identified metabolites for the DS‐AD versus Down syndrome‐mild cognitive impairment (DS‐MCI) comparison (B). The classification performance from these reduced set of metabolites is not significantly different from the larger sets used in Figure 3. The consistency of the ROC AUC across the resampling schemes (10‐fold CV and 100‐fold Monte Carlo CV) and classification models (logistic regression and support vector machine) shows the overall stability of the models and argues against overfitting