Metabolic network failures in Alzheimer's disease: A biochemical road map

Abstract

Introduction: The Alzheimer's Disease Research Summits of 2012 and 2015 incorporated experts from academia, industry, and nonprofit organizations to develop new research directions to transform our understanding of Alzheimer's disease (AD) and propel the development of critically needed therapies. In response to their recommendations, big data at multiple levels are being generated and integrated to study network failures in disease. We used metabolomics as a global biochemical approach to identify peripheral metabolic changes in AD patients and correlate them to cerebrospinal fluid pathology markers, imaging features, and cognitive performance.

Methods: Fasting serum samples from the Alzheimer's Disease Neuroimaging Initiative (199 control, 356 mild cognitive impairment, and 175 AD participants) were analyzed using the AbsoluteIDQ-p180 kit. Performance was validated in blinded replicates, and values were medication adjusted.

Results: Multivariable-adjusted analyses showed that sphingomyelins and ether-containing phosphatidylcholines were altered in preclinical biomarker-defined AD stages, whereas acylcarnitines and several amines, including the branched-chain amino acid valine and α-aminoadipic acid, changed in symptomatic stages. Several of the analytes showed consistent associations in the Rotterdam, Erasmus Rucphen Family, and Indiana Memory and Aging Studies. Partial correlation networks constructed for Aβ_1-42, tau, imaging, and cognitive changes provided initial biochemical insights for disease-related processes. Coexpression networks interconnected key metabolic effectors of disease.

Discussion: Metabolomics identified key disease-related metabolic changes and disease-progression-related changes. Defining metabolic changes during AD disease trajectory and its relationship to clinical phenotypes provides a powerful roadmap for drug and biomarker discovery.

Keywords: Acylcarnitines; Alzheimer's disease; Biochemical networks; Biomarkers; Branched-chain amino acids; Dementia; Metabolism; Metabolomics; Metabonomics; Pharmacometabolomics; Pharmacometabonomics; Phospholipids; Precision medicine; Serum; Sphingomyelins; Systems biology.

Fig. 1

Clustering of pairwise metabolite correlations and association results with clinical variables. (A) Heat map of Spearman correlations between the residuals of metabolite concentrations on the single metabolites. Metabolites are clustered using hierarchical clustering using the Euclidean distance metric. The clustering assigns metabolites to their biochemical class: amino acids, biogenic amines, short-chain and long-chain acylcarnitines, lyso-lipids, PC, and SM. Significant clusters of acylcarnitines are outlined in blue and amines outlined in brown. (B) Association results of the regression analyses. The distribution of association results of metabolites with clinical variables mirrors the correlation structure of the metabolites. Abbreviations: α-AAA, α-aminoadipic acid; AD, Alzheimer’s disease; C0, free carnitine; Cx:y, acylcarnitines; Cx:y-OH, hydroxylacylcarnitines; Cx:y-DC, dicarboxylacylcarnitines; CN, cognitively normal; lysoPC, lyso-glycero-phosphatidylcholines (a = acyl); MCI, mild cognitive impairment; Path. Aβ_1–42, pathological Aβ_1–42; PC, glycero-phosphatidylcholines (aa = diacyl, ae = acyl–alkyl); SDMA, symmetric dimethylarginine; SM, sphingomyelin; SMx:y, sphingomyelins; SM (OH) x:y, N-hydroxylacyloylsphingosyl-phosphocholine; T4-OH-Pro, trans-4-hydroxyproline.

Fig. 2

Relationship between serum metabolites, clinical diagnosis, and Aβ_1–42 status. Serum PC ae 44:4 (A), PC ae 44:4 (B), and C18 (C) concentrations stratified by clinical diagnosis and CSFAβ_1–42–defined groups. The concentration of each metabolite is shown for each diagnosis red: CN, green: MCI, blue: AD and by N. Abeta: normal concentrations of Aβ_1–42 (>192 pg/mL), and Path. Abeta: pathological concentrations of Aβ_1–42 (<192 pg/mL), Y-axes are values for each metabolite. Scatter plot for ADAS-Cog13 and serum valine values (D). Black line and shading are the regression line and 95% confidence interval. (D and E) Correlations between valine levels and cognitive decline in ADNI-1 and Rotterdam, respectively. Abbreviations: α-AAA, α-Aminoadipic acid; ADAS-Cog13, Alzheimer’s Disease Assessment Scale–Cognition; ADNI-1, Alzheimer’s Disease Neuroimaging Initiative–1; C0, free carnitine; Cx:y, acylcarnitines; Cx:y-OH, hydroxylacylcarnitines; Cx:y-DC, dicarboxylacylcarnitines; lysoPC, lyso-glycero-phosphatidylcholines (a = acyl); PC, glycero-phosphatidylcholines (aa = diacyl, ae = acyl–alkyl); SDMA, symmetric dimethylarginine; SMx:y, sphingomyelins; SM (OH) x:y, N-hydroxylacyloylsphingosyl-phosphocholine; T4-OH-Pro, trans-4-hydroxyproline.

Fig. 3

Longitudinal associations for SM C20:2. (A) Cox hazards model of the association of conversion from MCI to AD. Black line: 1st tertile, red line: 2nd tertile, green line: 3rd tertile. Analysis was conducted using quantitative values, and stratification by tertiles was used only for graphical representation. (B) Association between baseline concentrations of SM 20:2 and longitudinal cognitive (ADAS-Cog13) and imaging (MRI: brain ventricular volume) changes during follow-up. Lines represent trajectories on subjects on the 25th percentile (black line), 50th percentile (red line), 75th percentile (green line) of baseline SM 20:2. Y-axes are ADAS-Cog13 score (left) and ventricular volume (right). Trajectories for these values are calculated based on the studied mixed-effects models. Abbreviations: AD, Alzheimer’s disease; ADAS-Cog13, Alzheimer’s Disease Assessment Scale–Cognition; MCI, mild cognitive impairment; MRI, magnetic resonance imaging.

Fig. 4

Network model showing metabolic pathways correlated with the temporal evolution of biomarkers and clinical variables in AD. (A) Partial correlation network. Gaussian graphical model of metabolite concentrations showing reconstructed metabolic pathways and highlighting of the different modules involved in the steps along the temporal evolution of biomarkers and clinical variables in AD. Nodes in the network represent the metabolites, and edges (lines) illustrate the strength and direction of their partial correlations. Only partial correlations significant after Bonferroni correction for all possible edges are included. Labels show the major classes of metabolites included in our study. Gray circles outline the modules highlighted in panel B. (B) Schematic diagram of the model of temporal evolution of biomarkers in AD, modified from Jack and Holtzman [30], augmented with colored versions of the network from panel A. In these networks, nodes are highlighted according to the strength and direction of the metabolite’s association with the respective clinical trait with blue as positive and red as negative (networks in temporal order from left to right: pathological Aβ_1–42, T-tau, SPARE-AD, and ADAS-Cog13). Significant associations are colored in dark blue/bright red, and weaker (but at least nominally significant at 0.05) associations are displayed in fainter colors. Modules of metabolites implicated in the respective trait are highlighted by circles colored by their first occurrence in the temporal order following the color scheme of the time sequence on the bottom. The partial correlation network for Aβ_1–42 (panel A) highlighted direct correlations with short- and medium-chain SM and PC with ether bonds suggesting a role for membrane structure and function, contact sites, and membrane signaling in amyloid pathology. There was a different pattern for tau (panel B) with highlighted metabolites with long-chain acylcarnitines and SM implicated in lipid metabolism showing association with T-tau level. The SPARE-AD and ADAS-Cog13 partial correlation networks were very similar suggesting associations of brain atrophy and cognitive decline with metabolic changes in BCAAs and short-chain acylcarnitines that have been implicated in mitochondrial energetics as well as additional changes in lipid metabolism. Abbreviations: AD, Alzheimer’s disease; ADAS-Cog13, Alzheimer’s Disease Assessment Scale–Cognition; BCAA, branched-chain amino acid; PC, glycero-phosphatidylcholines (aa = diacyl, ae = acyl–alkyl); SM, sphingomyelin; SPARE-AD, Spatial Pattern of Abnormalities for Recognition of Early AD.

Fig. 5

Coexpression subnetwork with direct and indirect interconnections between select metabolites. A coexpression subnetwork focused on three metabolites also identified in the Rotterdam data set (PC ae C40:3, valine, and SM C20:2) was generated from the primary network (Supplementary Fig. 2). The subnetwork shows these three metabolites have high correlations (red edges lines) and lower correlations (green edges lines) to multiple modules via direct and indirect interconnections. Each module is denoted by a color representing a robust set of coregulated metabolites in interconnected biochemical pathways, for example, orange module contained a subset of amines, green module consists of long-chain acylcarnitines; teal, brown, and blue modules contained exclusively PC and lysoPC; red module contained SM and PC; gray module contained short-chain acylcarnitines and other amines. Each node represents a metabolite. The edge (line) opacity is proportional to the Pearson correlation, that is, lighter means weaker correlation value and darker means stronger correlation. The intermodule edges represent correlations and potentially indirect interactions among metabolites and biochemical pathways. The coexpression network captures all significant associations between metabolites and reveals a global correlation structure and interconnections among different modules that adds to our understanding of the disease network. Abbreviations: lysoPC, lyso-glycero-phosphatidylcholines (a = acyl); PC, glycero-phosphatidylcholines (aa = diacyl, ae = acyl–alkyl); PC ae, ether-containing PC; SM, sphingomyelin.