Metabolomic and lipidomic signatures in autosomal dominant and late-onset Alzheimer’s disease brains

Abstract

Introduction:

The identification of multiple genetic risk factors for Alzheimer’s disease (AD) suggests that many pathways contribute to AD onset and progression. However, the metabolomic and lipidomic profiles in carriers of distinct genetic risk factors are not fully understood. The metabolome can provide a direct image of dysregulated pathways in the brain.

Methods:

We interrogated metabolomic signatures in the AD brain, including carriers of pathogenic variants in APP, PSEN1, and PSEN2 (autosomal dominant AD; ADAD), APOE ɛ4, and TREM2 risk variant carriers, and sporadic AD (sAD).

Results:

We identified 133 unique and shared metabolites associated with ADAD, TREM2, and sAD. We identified a signature of 16 metabolites significantly altered between groups and associated with AD duration.

Discussion:

AD genetic variants show distinct metabolic perturbations. Investigation of these metabolites may provide greater insight into the etiology of AD and its impact on clinical presentation.

1 |. BACKGROUND

Alzheimer’s disease (AD), the most common form of dementia, is a heterogeneous and complex disease characterized by the accumulation of amyloid (Aβ) plaques and neurofibrillary tangles in the brain. AD may develop as familial or sporadic. Recent advancements in AD diagnosis and treatment could benefit from a comprehensive multiomic approach to studying diverse biological processes, including metabolism.^1,2 Pathological changes in AD begin decades before AD diagnosis.³ Therefore, metabolomic changes linked to AD pathology could precede disease onset and be highly informative for predictive models and preventative medicine. Metabolic decline is one of the first physiological changes detected in patients with mild cognitive impairment (MCI) due to AD.⁴ Changes in lipid and energy metabolism are proven hallmarks of AD, but there are also reports of impairments in neurotransmitter, urea cycle, purine, polyamine, and bile acid metabolisms.⁵ Current symptomatic treatments (cholinesterase inhibitors and memantine) target deficits in neurotransmitters to minimize cognitive decline.⁶

The dysregulation of sphingolipids and glycerophospholipids in blood samples from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) and blood and brain samples from the Baltimore Longitudinal Study of Aging (BLSA) cohorts have been previously reported.^7–9 These metabolites allowed discrimination between AD and controls with high accuracy, sensitivity, and specificity.¹⁰ Blood and brain endophenotype scores summarized the relative importance of each metabolite to the severity of AD pathology and disease progression. Furthermore, Stamate et al. used machine learning classifiers to demonstrate that a panel of plasma metabolites has the potential to match the area under the curve of well-established cerebrospinal fluid (CSF) biomarkers when used to classify AD versus healthy individuals.¹¹ Pathway analysis with the top 20 predictive metabolites indicated that the nitrogen pathway was overrepresented. Though much progress has been made in determining the specific metabolic changes in biospecimens from AD patients, the metabolomic landscape has yet to be fully understood.

AD is highly heritable and can be caused by autosomal dominant genetic variants in the amyloid precursor protein (APP), presenilin1 and −2 (PSEN1 and PSEN2) genes, or associated with risk factors in multiple other loci, including apolipoprotein E (Apo E) and triggering receptor on myeloid cells 2 (TREM2).^12–14 The singularities of downstream effects of the complex AD genetic etiology are currently poorly understood. Pathogenic genetic variants in APP, PSEN1, and PSEN2 support the amyloid cascade hypothesis.¹⁵ TREM2 interacts with APOE, Aβ, and other lipids, mediating the recruitment of microglia to Aβ plaques.^16,17 Rare variants in TREM2 may impair microglial function, contributing to AD pathology.¹⁸ Apo E is a critical player in lipid metabolism, transport, and homeostasis in the brain, and the ɛ4 allele is the main common genetic risk factor for AD. Previous studies reported a more pronounced impairment of mitochondrial energy production in females carrying APOE ɛ4,¹⁹ suggesting that genetic risk factors contribute to AD pathology through distinct mechanisms. However, the metabolomic changes associated with AD pathology and most genetic factors are currently unknown.

We sought to systematically investigate the metabolic signature of major AD genetic risk factors by interrogating the metabolomes of carriers of autosomal dominant AD (ADAD) pathogenic mutations, APOE and TREM2 risk variant carriers, symptomatic non-carrier AD, presymptomatic individuals with AD neuropathological change but no or minimal decline of cognition, and cognitively unimpaired controls without AD neuropathology. Our analysis uncovered common profiles across genetically categorized brains, and metabolites and lipids specific to distinct genetic factors.

2 |. METHODS

2.1 |. Cohorts

2.1.1 |. Washington University School of Medicine

Archived fresh-frozen post-mortem parietal cortical tissue samples were obtained from the Charles F. and Joanne Knight Alzheimer Disease Research Center Brain Bank (Knight ADRC) and the Dominantly Inherited Alzheimer Network (DIAN) at Washington University School of Medicine (WUSM). Samples were obtained with informed consent, and the WUSM Institutional Review Board approved the study. Samples were categorized based on neuropathological and genetic information: AD diagnosis with variants in ADAD genes, carriers of TREM2 risk variants (TREM2), sporadic AD (sAD), neuropathological changes but no clinical symptoms (Presymptomatic), or no or minimal neuropathological AD lesions (CO). (Table 1, Table S1, Supplementary Text). These cohorts have been described previously.^20–28 Data from the Knight ADRC donors are available at NIAGADS accession number ng0013. Data from the DIAN cohort can be requested directly from DIAN (https://dian.wustl.edu/).

TABLE 1.

Summary statistics for the six datasets included in this study

						Braak(Tau)				CDR
		Condition	N	Agea	%Fe	%APOE4+	0–III	IV–VI	Not avail.	< = 0.5	1–2	3	PMI
Discovery	WUSM (parietal brain)	ADAD	25	54 ±13.9	32%	36%	0	20	5	0	0	25	13.3
		sAD	305	84 ± 8.8	62%	56%	23	195	87	0	23	282	12.5
		TREM2	21	84 ± 7.5	52%	57%	3	12	6	0	1	20	14.1
		Presymptomatic	15	87 ± 9.7	60%	53%	3	8	4	15	0	0	13.6
		Control	26	88 ±9.1	69%	12%	20	1	5	26	0	0	10.0
Replication	ROSMAP (DLPFC brain Metabolon)	sAD	233	89 ±2.9	73%	38%	15	218	0				8.1
		Control	94	86 ± 4.7	59%	10%	94	0	0				7.7
	ROSMAP (DLPFC brain p180)	sAD	43	88 ± 2.8	19%	29%	3	40	0				9.6
		Control	23	87 ± 3.8	35%	7%	23	0	0				8.7
	ROSMAP (serum)	sAD	36	87 ± 3.8	81%	8%
		Control	55	87 ± 4.4	69%	0%
	ADNI1 (serum)	sAD	184	75 ± 7.5	48%	66%
		Control	224	76 ± 5.0	48%	27%
	ADNIGO/2 (serum)	sAD	137	74 ± 8.2	42%	66%
		Control	181	73 ± 6.4	51%	28%

Abbreviations: CDR, Clinical Dementia Rating; PMI, post mortem interval; sAD, sporadic Alzheimer’s disease.

^aAge is age at death for brain samples, and age at baseline for serum samples.

2.1.2 |. Religious Orders Study and Memory and Aging Project (ROSMAP)

Data from the ROSMAP study was generated by the Duke Metabolomics and Proteomics Shared Resource, a member of the ADMC.^19,29–31 Details of the ROSMAP study design and methods have been described previously.^32,33 The following samples were analyzed: 36 sAD and 55 CO serum samples, 233 sAD and 94 CO dorsolateral prefrontal cortex (DLPFC) samples quantified with Metabolon, and 43 sAD and 23 CO DLPFC samples quantified with Biocrates p180 (Table 1, Supplementary Text). An additional grouping of 223 AD and 154 CO participants was considered based on consensus clinical diagnosis only (Table S2). The ROSMAP data are publicly available on Synapse as syn26007829 and syn26007830.

2.1.3 |. Alzheimer’s Disease Neuroimaging Initiative (ADNI)

Data from the ADNI was also generated by the Duke Metabolomics and Proteomics Shared Resource.^19,29,30 Additional information is available elsewhere.^34,35 In total, 321 sAD and 405 CO serum samples were analyzed from ADNI1 and ADNIGO/2 cohorts (Table 1, Supplementary Text). The ADNI data can be accessed at https://adni.loni.usc.edu.

2.2 |. Metabolite quantification

2.2.1 |. Metabolon Precision Metabolomics platform

Data from the Knight ADRC, DIAN, and one ROSMAP cohort were generated on the Metabolon Precision Metabolomics platform (Metabolon, Inc., Morrisville, USA). For the Knight ADRC and DIAN cohorts, 50 mg frozen parietal cortical tissue samples were analyzed (Supplementary Text, Table S3).

2.2.2 |. Biocrates AbsoluteIDQ p180 platform

The remaining ADNI and ROSMAP datasets were quantified by the Biocrates AbsoluteIDQ p180 platform, which measures approximately 180 metabolites (Biocrates Life Science AG, Innsbruck, Austria). Of these, 85 could be matched with those quantified by Metabolon based on Human Metabolome Database Identifier (HMDB ID) (Table S3).

2.3 |. Quality control

2.3.1 |. Knight ADRC and DIAN cohorts

The quantification platform’s consistency was evaluated with 13 technical replicate samples. The 815 metabolites annotated with structural identities were kept for further analyses. We performed filtering based on missingness rate; 188 metabolites missing over 20% of readings were excluded (Supplementary Text, Figure S1). Raw readings were log10-transformed to approximate a normal distribution. Readings outside 1.5× interquartile range for each metabolite were excluded, and means were adjusted to zero.

No samples were excluded due to missingness. Principal component analysis was performed on scaled and imputed data provided by Metabolon to identify outlier samples using the FactoMineR package³⁶ (Figure S2). The final dataset consisted of 627 metabolites in 392 samples (Table S4).

2.3.2 |. Replication datasets

We used the above-described procedure for cleaning and QC of the ADNI and ROSMAP datasets (Supplementary Text). After QC, the ADNI1 dataset included 149 metabolites in 408 samples, the ADNIGO/2 dataset included 157 metabolites in 318 samples, the ROSMAP serum dataset included 162 metabolites in 91 samples, and the ROSMAP p180 brain dataset included 157 metabolites in 66 samples. The ROSMAP Metabolon brain dataset included 595 metabolites in 327 samples (Table 1).

2.4 |. Statistical analyses

Association analyses were conducted using linear regression in R software version 3.6.³⁷ Metabolite levels were modeled for sAD, ADAD, and TREM2 compared to CO, corrected by sex, age at death (AAD), and post-mortem interval (PMI). Based on age at AD onset (AAO), we also performed a stratified analysis of early-onset AD (EOAD, AAO < 65 years) and late-onset AD (LOAD, AAO ≥ 65 years) in the WUSM sAD group. Associations with APOE ε4 carrier status were tested within sAD with the same covariates. When comparing ADAD and EOAD with CO, AAD was not included as it is collinear with status. To investigate age-associated metabolites whose differential abundance may have been associated with the difference in AAD between ADAD and CO, linear regression was used to test each metabolite’s association with AAD in the sAD and CO status groups. The metabolome-wide false discovery rate (FDR)-adjusted significance (q-value) threshold was established as q < 0.05. We employed an analysis of covariance (ANCOVA) to compare the effect sizes for metabolites in sAD and TREM2 relative to their effects in ADAD (Figure S3). We performed additional ANCOVA tests with individuals matched by Clinical Dementia Rating (CDR),³⁸ Braak tau, and Braak Aβ.

2.4.1 |. Eigenmetabolite analysis

We calculated the first principal component of readings for metabolites that were differentially abundant in multiple groups, similar to the “eigengene” concept in the WGCNA package.³⁹ We tested the differences in the eigenmetabolite profile between groups. Within the sAD group, we tested for association with CDR, Braak tau stage, and disease duration (Figure S3).

2.4.2 |. Replication of association findings

PMI was not applicable as a covariate for the follow-up analyses in serum, and age at baseline was used rather than AAD. A meta-analysis was also performed combining the three serum datasets, using the same linear model additionally corrected by study.

2.4.3 |. Pathway analysis

Pathway and network analyses were performed using MetaboAnalyst^40,41 and IMPaLA pathway over-representation analysis,⁴² and the Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway database^43–45 (Supplementary Text, Table S5).

2.4.4 |. Web browser

Our results are available through a public browser at http://web.hararilab.org/Metabolomics (Supplementary Text). Source code for the browser and scripts employed to QC, clean, and analyze the data are available at http://github.com/HarariLab/Metabolomics.

3 |. RESULTS

3.1 |. Study design

In this study, we performed a metabolomics analysis of parietal cortical tissue from participants of the DIAN and Knight ADRC cohorts (Figure 1). We determined the metabolomic profile of 392 participants, including three AD genetic subgroups: autosomal dominant AD (ADAD), carriers of risk variants in TREM2 (TREM2), and sporadic AD (sAD). Detailed phenotypic information included genetic risk factors, CDR, and Braak staging for tau and Aβ. Out of 880 metabolites quantified, 627 passed quality control (Table S4). We tested differential abundance using linear models and compiled the differentially abundant metabolites into a distinct profile between AD genetic groups. We also conducted pathway analysis. We analyzed 393 brain samples and 817 serum samples from the ROSMAP and ADNI cohorts to validate our results.

FIGURE 1.

Study design. Parietal cortical tissue from donors to the Knight ADRC and DIAN were analyzed on the Metabolon Precision Metabolomics platform: autosomal dominant AD (ADAD, n = 25), sporadic AD (sAD, n = 305), TREM2 (n = 21), presymptomatic (n = 15), and healthy control (CO, n = 26). After quality control, 627 metabolites were tested for differential abundance via linear modeling. A metabolic profile was generated from 16 metabolites in common between groups. Pathway analysis was performed on the differentially abundant metabolites, and a web browser was created to share the data and results. Findings were validated in five independent datasets: dorsolateral prefrontal cortex (DLPFC) tissue from the ROSMAP cohort analyzed on the Metabolon platform (sAD n = 233, CO n = 94), as well as four datasets quantified using the Biocrates p180 platform: DLPFC from the ROSMAP cohort (sAD = 43, CO = 23), serum from the ROSMAP cohort (sAD n = 36, CO n = 55), serum from the ADNI1 cohort (sAD n = 184, CO n = 224), and serum from the ADNIGO/2 cohort (sAD n = 137, CO n = 181). The ROSMAP Metabolon dataset was found to have 506 metabolites in common with the Knight ADRC/DIAN cohort after quality control. The p180 platform was found to have 85 metabolites in common with the Metabolon platform.

3.2 |. β-citrylglutamate shows differential abundance in sporadic AD and ADAD

Association analysis indicated that ADAD had the most distinct brain metabolomics profile (131 significant metabolites; Table S6). The profiles for TREM2 and sAD cases showed fewer differences, with three and five significant metabolites, respectively (Figure 2). The Super Pathways represented in the ADAD-associated metabolites were amino acids (48 metabolites), carbohydrates (12), cofactors and vitamins (9), energy (2), lipid (30), nucleotide (12), peptide (12), and xenobiotics (6). We tested metabolite association with AAD to rule out confounding effects between ADAD and CO (Supplementary Text, Table S7).

FIGURE 2.

Association analysis in WUSM dataset. Volcano plots for (A) sAD versus CO, (B) ADAD versus CO, and (C) TREM2 versus CO. (D) Venn diagram of overlapping significant metabolites in each group. Box plots for abundance of top metabolites: (E) β-citrylglutamate, (F) α-tocopherol, (G) ergothioneine, and (H) N-acetylputrescine.

Pathway analysis indicated overrepresented pathways in amino acid metabolism (including glutamate, glutathione, tryptophan, lysine, and histidine metabolisms), sphingolipid metabolism, and novel associations with vitamin pathways (Tables S8 and S9).

β-citrylglutamate (BCG) was significantly differentially abundant in all three genetic groups. Administration of fluoxetine, a commonly prescribed SSRI antidepressant, has been shown to significantly affect serum BCG levels.⁴⁶ Pharmacological records indicated that five AD participants from the Knight ADRC took fluoxetine during their most recent clinical assessment. However, we ruled out the confounding effects of fluoxetine on the association between BCG levels and AD (Supplementary Text, Table S10).

3.3 |. APOE ɛ4 carrier status shows nominal associations with metabolites

In an association analysis of APOE ε4 carriers versus non-carriers in the sAD group, no metabolites were significantly associated after FDR correction, though 25 metabolites were nominally significant (P < .05) (Table S11).

3.4 |. Follow-up in independent datasets

To independently validate our results, we tested the differential abundance of metabolites in serum and DLPFC samples from the ADNI1, ADNIGO/2, and ROSMAP cohorts. We identified 44 differentially abundant metabolites with consistent effect direction in our ADAD analysis and at least one independent dataset. Among the replicated metabolites were α-tocopherol, BCG, and serotonin (Table S12).

Of the seven significant analytes in both serum and brain cohorts, opposite direction of effect between tissue types was identified among five analytes (Supplementary Text).

3.5 |. A metabolic profile associated with AD duration

To investigate whether ADAD, TREM2, and sAD showed varying magnitudes of alteration in their metabolomic profiles, we selected the 16 metabolites with call rates above 80% that were significantly associated with ADAD after metabolome-wide FDR correction (q-value < 0.05) and nominally associated with both sAD and TREM2 (P-value < .05) (Table 2, Figure S3). Ergothioneine, serotonin, BCG, CDP-ethanolamine, and α-tocopherol were statistically significant after metabolome-wide FDR correction in sAD or TREM2 (Table 2). These 16 metabolites were considered for the identification of a metabolic profile. In the ROSMAP cohort, the differential abundance of nine of the 16 remaining metabolites was replicated after metabolome-wide FDR correction (Table 2). Additionally, four were nominally significant, and one (3-hydroxy-2-ethylpropionate) did not pass QC.

TABLE 2.

Effects and significance for the 16 metabolites that were significant after metabolome-wide FDR correction in the ADAD versus CO comparisons and at least nominally significant in the sAD versus CO and TREM2 versus CO comparisons in the WUSM cohort

Metabolite	ADAD vs. CO			sAD vs. CO		TREM2 vs. CO		ROSMAP(DLPFC Metabolon)
	Effect	P-value	q-valuea	Effect	P-value	Effect	P-value	Effect	P-value
Aspartate	0.160	1.1 × 10−3	1.1 × 10−2	0.085	2.2 × 10−3	0.123	6.0 × 10−3	0.042	1.1 × 10−2
γ-glutamylthreonine	0.263	3.2 × 10−3	2.3 × 10−2	0.171	5.4 × 10−4	0.278	7.1 × 10−4	0.095	5.1 × 10−4
β-citrylglutamate	−0.217	4.9 × 10−5	2.2 × 10−3	−0.148	2.1 × 10−4	−0.293	7.8 × 10−7	−0.062	3.7 × 10−4
Glutamate	0.086	7.1 × 10−3	3.9 × 10−2	0.048	1.7 × 10−2	0.071	6.5 × 10−3	0.060	1.1 × 10−8
N-acetylglutamate	−0.207	1.7 × 10−5	1.0 × 10−3	−0.080	2.7 × 10−3	−0.122	3.2 × 10−3	−0.033	6.8 × 10−4
Ergothioneine	−0.255	8.3 × 10−3	4.3 × 10−2	−0.213	8.0 × 10−5	−0.274	7.5 × 10−3	−0.092	2.5 × 10−2
3-hydroxy-2-ethylpropionate	0.177	2.0 × 10−3	1.7 × 10−2	0.099	4.4 × 10−2	0.195	8.7 × 10−3
1,5-anhydroglucitol (1,5-AG)	0.310	6.6 × 10−6	7.3 × 10−4	0.131	1.1 × 10−2	0.155	4.3 × 10−2	0.088	4.4 × 10−3
2-methylcitrate/homocitrate	−0.171	1.3 × 10−3	1.2 × 10−2	−0.087	1.3 × 10−2	−0.118	2.9 × 10−2	−0.048	1.1 × 10−2
Glutarate (C5-DC)	0.252	7.3 × 10−5	2.3 × 10−3	0.121	5.7 × 10−3	0.170	1.6 × 10−2	0.090	1.5 × 10−5
CDP-choline	−0.095	9.6 × 10−4	1.0 × 10−2	−0.040	4.6 × 10−2	−0.093	1.6 × 10−3	0.042	7.5 × 10−3
CDP-ethanolamine	−0.102	6.1 × 10−4	7.7 × 10−3	−0.054	2.8 × 10−3	−0.128	1.8 × 10−4	−0.024	8.2 × 10−2
Glycerophosphoinositol	−0.164	3.4 × 10−4	5.3 × 10−3	−0.095	1.3 × 10−3	−0.098	2.7 × 10−2	−0.021	4.5 × 10−2
Nicotinamide (vitamin B3)	−0.066	6.8× 10−3	3.7 × 10−2	−0.032	1.9 × 10−2	−0.056	9.4 × 10−3	0.008	0.33
α-tocopherol (vitamin E)	−0.120	4.1 × 10−5	2.0 × 10−3	−0.065	1.5 × 10−3	−0.115	1.8 × 10−5	−0.123	3.0 × 10−5
Retinol (vitamin A)	−0.287	1.3 × 10−5	9.2 × 10−4	−0.156	8.0 × 10−4	−0.192	1.4 × 10−3	−0.196	2.9E-04

Note: Highlighted in bold are metabolites with significant metabolome-wide FDR-adjusted q-values in each group (see Table S6 for metabolome-wide sAD, TREM2, and ROSMAP q-values).

^aq-value represents metabolome-wide FDR adjusted significance for 627 metabolites.

To determine whether genetic status would influence the degree of metabolic perturbation, we compared the magnitude of the effects of the 16 chosen metabolites across the three genetic groups. We ascertained through an ANCOVA test that the TREM2 and ADAD groups showed greater metabolic perturbations than the sAD group (P = 4.37 × 10⁻⁴) (Figure 3B). This finding persisted when individuals were matched by CDR (P = 2.54 × 10⁻⁰²) as well as Braak stage for tau (P = 1.18 × 10⁻⁰³) and Aβ (P = 2.20 × 10⁻⁰³).

FIGURE 3.

Metabolic profile consisting of 16 metabolites which passed FDR correction in ADAD versus CO and were at least nominally significant in sAD versus CO and TREM2 versus CO. (A) Heatmap showing relative abundance for each metabolite in the profile. Participants are divided by disease status group: healthy controls (CO), neuropathology but no cognitive impairment (Presymptomatic), sporadic AD (sAD), carriers of TREM2 risk variants (TREM2), and carriers of Mendelian mutations (ADAD). The 30 individuals with a mild metabolic AD profile (MMAD) as identified by hierarchical clustering are indicated within the red box. Annotations show Clinical Dementia Rating and Braak stages for tau accumulation. (B) Comparison of effects for the 16 metabolites in each model. The x-axis shows the effect of each metabolite in the ADAD versus CO model, while the y-axis shows the effects in the sAD versus CO (blue) and TREM2 versus CO (green) models. (C) Boxplot showing distribution of the first principal component for the 16-metabolite profile among each of the status groups.

We generated an “eigenmetabolite” from these 16 metabolites, representing the average metabolic profile for each individual.³⁹ We initially tested that the eigenmetabolite recapitulates the differences identified by the individual metabolite tests. The AD groups showed significantly lower eigenmetabolite values than CO, and ADAD showed significantly lower values than sAD (Figure 3C). Lower values were associated with a longer disease duration (P = 1.86 × 10⁻⁰²), higher Braak tau stage (P = 4.52 × 10⁻⁵), and higher CDR (P = 8.64 × 10⁻⁸) within sAD. In ROSMAP, the eigenmetabolite was associated with AD clinical diagnosis (P = 2.73 × 10⁻³, Figure 4) and disease duration (P = 2.68 × 10⁻⁰²), but not with Braak tau stage in sAD (P = .55).

FIGURE 4.

Distribution of metabolic eigenmetabolite profile between clinical diagnosis groups calculated with 15 metabolites on ROSMAP Metabolon data. 3-hydroxy-2-ethylpropionate was not included because it did not pass quality control in the ROSMAP dataset.

Hierarchical clustering analyses in the WUSM cohort identified 30 sAD individuals with eigenmetabolite values not significantly different from CO (P = .66) (Figure 3A). We termed this group Mild Metabolic AD profile (MMAD) as these individuals showed lower CDR (P = 8.02 × 10⁻⁶) and Braak tau (P = 1.22 × 10⁻⁴) compared to sAD. MMAD individuals also showed a shorter disease duration (P = 4.42 × 10⁻²) than the rest of sAD.

Eigenmetabolite values for the presymptomatic group were not significantly different from CO and MMAD (P = .18 and P = .06, respectively) (Figure 3C). As expected, the presymptomatic group showed significantly lower CDR (P = 2.47 × 10⁻⁰³) than the MMAD group. The groups did not differ in Braak tau stage (P = .92).

Super Pathways associated with these 16 metabolites included amino acid, phospholipid, and vitamin metabolisms. We found negligible confounding effects of vitamin supplementation (Supplementary Text).

3.6 |. Early onset AD shows pronounced metabolic perturbations

We performed a stratified analysis of 54 early onset AD (EOAD, AAO < 65) and 249 late-onset AD (LOAD, AAO >= 65) cases. EOAD individuals showed earlier age at death (72.3 ± 6.5 years) than LOAD (85.0 ± 7.1years) (P < 2.2 × 10⁻¹⁶), but also a significantly longer disease duration (13.13 ± 4.8 years compared to 9.1 ± 4.4 years for LOAD, P = 2.88 × 10⁻⁷). EOAD tended to have significantly higher CDR (P = 1.45 × 10⁻⁴) and Braak tau stage (P = 3.09 × 10⁻⁵) than LOAD. EOAD showed 40 differentially abundant metabolites compared to control after metabolome-wide FDR correction (Table S13). Of these, 24 metabolites were also associated with AAD in WUSM sAD, but none were associated with AAD in ROSMAP. Furthermore, 27 of these 40 metabolites showed significantly different levels in ADAD. Tryptophan betaine was significantly different from CO in LOAD and EOAD and was associated with AAD. We could not study EOAD in the ROSMAP cohort as AAO for all individuals in that cohort was greater than 71 years.

4 |. DISCUSSION

We performed a metabolomics analysis of parietal cortical tissue from the DIAN and Knight ADRC cohorts to characterize three genetically defined AD subgroups: sAD, ADAD, and carriers of risk variants in TREM2. We analyzed detailed phenotypic information, including genetic risk factors and clinical, pharmacological, and neuropathological variables.

We found a significantly different metabolic profile in ADAD compared to healthy individuals, with 131 significant metabolites related to multiple pathways (Tables S8 and S9). The parietal cortex of ADAD individuals has been reported to have a higher burden of neurofibrillary tangles (NFT) than that of sAD individuals.⁴⁷ Of TREM2 and sAD, only sAD showed metabolite differences unique to its category (tryptophan betaine and N-acetylputrescine). Tryptophan betaine is an N-methylated form of tryptophan, which is the serotonin precursor and has been found dysregulated in MCI-AD.⁴⁸ N-acetylputrescine is an acetyl-CoA-ated putrescine and a GABA precursor found to build up in stable MCI but not in AD, where putrescine is preferentially metabolized to other polyamines.⁴⁹ We also observe depleted N-acetylputrescine levels in sAD in our data (Figure S4) which supports previously reported findings of lower GABA levels in AD⁵⁰.

Among the metabolites identified as differentially abundant in sAD and ADAD were BCG and ergothioneine. Both showed lower concentrations in AD compared to control. BCG acts as an iron carrier to activate aconitase activity.⁵¹ BCG had the lowest abundance in ADAD, followed by TREM2 and sAD. This observation could be associated with lower activation of aconitase and lower energetic metabolism. BCG is also a component of glutamate metabolism, and BCG levels can be increased by the selective serotonin reuptake inhibitor (SSRI) fluoxetine.⁴⁶ We determined that BCG associations in the Knight ADRC cohort are not confounded by fluoxetine administration. Ergothioneine levels were also reduced in sAD and ADAD. Ergothioneine, a naturally occurring amino acid produced by fungi, has antioxidant and anti-inflammatory properties.^52,53 Ergothioneine blood levels in humans decline with age and decline faster in individuals with cognitive impairment compared to age-matched individuals with no cognitive impairment.⁵⁴ Ergothioneine protected against loss of memory and learning abilities in mice treated with an intracerebroventricular injection of Aβ1–40.⁵⁵

Neurotransmitters, especially serotonin, have been shown to play a role in processing APP and reducing the generation of Aβ₄₂ through activation of the ERK signaling cascade⁵⁶. The KEGG serotonergic synapse pathway (map04726) also demonstrates the relationship between 5-HT4 receptor activation by serotonin and production of neuroprotective sAPPα (Figure S5). In our study, serotonin levels were significantly decreased in sAD and ADAD compared to control and nominally decreased in TREM2. This effect was replicated in both serum and DLPFC tissue. SSRIs, which increase serotonin levels in the brain, show promise for reducing Aβ accumulation in the brain and CSF. Studies in APP/PS1 transgenic mice showed that the SSRI citalopram caused a 50% reduction in brain amyloid plaque load, and escitalopram, citalopram’s S-isomer, reduced interstitial fluid Aβ by 25%.^56,57 A controlled clinical trial of cognitively normal adults showed that escitalopram could decrease CSF Aβ₄₂ levels in humans.⁵⁸ Our results corroborate the association of low serotonin with AD and suggest that this effect, and potentially the benefit of serotonin modulation via SSRIs, spans all three genetic groups.

Our discovery dataset identified a set of 16 metabolites that were consistently differentially abundant between the three AD groups and control. When we compared the magnitude of the differential abundance of the metabolites among the genetic groups, the effects were greatest in ADAD. The TREM2 group exhibited a smaller effect than ADAD; the smallest effect was observed in sAD. The first principal component, or eigenmetabolite, of these 16 metabolites was distinct between the AD groups and healthy individuals and between sAD and ADAD. The eigenmetabolite was additionally associated with disease duration in sAD in the Metabolon data of both the Knight ADRC and ROSMAP cohorts. Further studies in independent cohorts of ADAD and TREM2 carriers with regional quantitative neuropathological information can shed light on the association of the eigenmetabolite with neuropathology.⁵⁹ We observed that the eigenmetabolite did not significantly differ between presymptomatic and healthy individuals. In addition, we identified a set of sAD cases (mild metabolic AD profile/MMAD) with a metabolic profile close to that of healthy individuals. This group showed significantly lower CDR and Braak tau stages than the rest of sAD. The presymptomatic and MMAD groups showed no significant difference in the metabolic eigenmetabolite or Braak tau but showed significantly different CDR. Together, these observations suggest that the metabolic profile could be implicated in AD status and duration.

We also performed a stratified analysis comparing early- and late-onset AD and found that EOAD shared more metabolic perturbations with ADAD than did LOAD. Further research is needed to replicate this finding in an independent dataset of EOAD individuals. We cannot conclude whether this observation is due to AAO or other factors such as CDR, disease duration, or neuropathology burden.

This study identified differences in metabolite abundance specific to and common among genetically defined AD subgroups. We replicated our main findings in five independent datasets. Common differences in metabolite levels allowed us to generate a metabolic profile associated with disease status and duration and identify a subset of AD cases with a profile similar to CO (MMAD). Metabolomics of the brain can identify metabolic signatures specific to AD genetic subgroups. These metabolites may support the creation of “metabolomics scores” to assess disease status. Limitations of our study include the sample size of some groups, for example, TREM2 and CO, and the lack of an age-matched control group for ADAD. We could not find associations with the Braak Aβ stage, possibly because of reduced statistical power as scores were not available for all participants. Direct replication of our results in ADAD individuals was unachievable due to a lack of independent ADAD datasets. Similarly, this dataset does not provide enough statistical power to compare metabolic profiles for specific autosomal dominant variants. Unlike previous studies, we did not find significant associations between APOE ε4 carriers and non-carriers in cases of sporadic AD. Our ability to replicate findings in other tissues, such as blood serum, was possibly limited due to different platforms used by other studies. In future studies, we will extend our analysis to serum metabolomics data and seek replication of our findings to facilitate further identification of novel biomarkers for AD.

HIGHLIGHTS.

• APP/PSEN1/PSEN2 and TREM2 variant carriers show distinct metabolic changes.

• A total of 133 metabolites were differentially abundant in AD genetic groups.

• β-citrylglutamate is differentially abundant in autosomal dominant, TREM2, and sporadic AD.

• A 16-metabolite profile shows differences between Alzheimer’s disease (AD) genetic groups.

• The identified metabolic profile is associated with duration of disease.

RESEARCH IN CONTEXT.

1. Systematic review:

We reviewed the available literature through traditional sources (e.g., PubMed, Google Scholar). Most studies of metabolomics in the AD brain have focused on sporadic AD or the APOE locus. There is a gap in knowledge regarding the metabolomic effects of autosomal dominant pathogenic mutations and other AD risk genetic variants.

2. Interpretation:

There are 133 differentially abundant metabolites among three genetically-defined groups compared to healthy controls, and a 16-metabolite profile that was distinct between groups. β-citrylglutamate is differentially abundant in all three groups compared to control. AD genetic variants (APP, PSEN1, PSEN2, and TREM2) produce distinct metabolic profiles compared to sporadic AD while sharing some metabolic perturbations. AD-associated metabolites could serve as treatment targets or aid diagnosis or risk determination.

3. Future directions:

Validation of our findings in blood and cerebrospinal fluid, and correlations with clinical variables in longitudinal cohorts to determine the clinical utility of these metabolites in diagnostic testing, risk prediction, or disease progression.