Abstract

Introduction

   Altered metabolism may occur years before clinical manifestations of
   Alzheimer's disease (AD). We used untargeted metabolomics on the
   cerebrospinal fluid of patients with mild cognitive impairment (MCI) to
   uncover metabolomic derangements.

Methods

   CSF from 92 normal controls and 93 MCI underwent untargeted
   metabolomics using high‐resolution mass spectrometry with liquid
   chromatography. Partial least squares discriminant analysis was used
   followed by metabolite annotation and pathway enrichment analysis
   (PES). Significant features were correlated with disease phenotypes.

Results

   We identified 294 features differentially expressed between the two
   groups and 94 were annotated. PES showed that sugar regulation
   (N‐glycan, P = .0007; sialic acid, P = .0014; aminosugars, P = .0042;
   galactose, P = .0054), methionine regulation (P = .0081), and tyrosine
   metabolism (P = .019) pathways were differentially activated and
   significant features within these pathways correlated with multiple
   disease phenotypes.

Conclusion

   There is a metabolic signature characterized by impairments in sugars,
   methionine, and tyrosine regulation in MCI. Targeting these pathways
   may offer new therapeutic approaches to AD.

   Keywords: Alzheimer's disease, cerebrospinal fluid, metabolism, mild
   cognitive impairment

1. INTRODUCTION

   Alzheimer's disease (AD) is characterized by a complex set of molecular
   pathways that begin decades before symptoms start.[30] ^1 , [31]^2
   Changes in proteins, lipids, and many other molecular networks have
   been described.[32] ^3 , [33]^4 The overlapping and interaction of
   these networks can obscure the root pathogenic mechanisms when not
   fully accounted for in molecular or analytical methods. The level of
   complexity in these networks is becoming more evident as the cumulative
   knowledge of AD pathogenesis has increased in the last decade.
   Disentangling these complexities is becoming more feasible due to the
   significant advances in high throughput technologies[34] ^5 coupled
   with novel bioinformatic tools including those developed by our
   team.[35] ^6 Recent examples applying the high throughput measurements
   of thousands of metabolites coupled with advanced bioinformatic
   approaches has comprehensively described molecular alterations and
   pathways in multiple diseases.[36] ^7 , [37]^8 , [38]^9 , [39]^10 We
   apply these advances in investigating underlying metabolic changes in
   AD.

   Metabolomic research focuses on examining metabolites, small molecules
   (typically <1500 Da) that are end products of multiple biological
   pathways and processes. The human metabolome is estimated to contain
   ≈150,000 or more of such metabolites and a large fraction are still
   unidentified.[40] ^11 Metabolomics aid in identifying downstream
   perturbations from the genetic and post genetic pathways reflecting a
   functional signature of biochemical activities that are closer to the
   phenotypical changes.[41] ^12 Our work uses high resolution untargeted
   metabolomic approaches, which can be a powerful tool in describing
   novel and previously unknown pathways involved in AD pathogenesis.

   Brain hypometabolism has been reported in symptomatic AD as well as
   before the onset of cognitive symptoms.[42] ^13 Preliminary studies
   have suggested the existence of multiple metabolic changes in this
   prodromal stage.[43] ^14 , [44]^15 However, many previous studies have
   either included small samples for cerebrospinal fluid (CSF)
   analyses,[45] ^16 , [46]^17 , [47]^18 used targeted approaches, or
   focused on plasma profiling.[48] ^16 , [49]^17 , [50]^18 , [51]^19 In
   this study, we conducted a case‐control untargeted high resolution
   metabolomic study on the CSF of a larger sample, relative to published
   CSF studies to date, of normal cognition (NC) and mild cognitive
   impairment (MCI), a prodromal state for AD. We aimed at investigating
   the alterations between NC and MCI in the metabolome and metabolic
   pathways using an established high resolution metabolomic biospecimen
   and data analysis pipeline. We further explored the association of
   these metabolic alterations with multiple disease phenotypes related to
   cognition, CSF amyloid beta 1‐42 (Aβ42) and tau biomarkers, and brain
   magnetic resonance imaging (MRI) measures.

2. METHODS

2.1. Participant description

   Data for the current analysis were drawn from the baseline assessment
   of participants in the Brain Stress Hypertension and Aging program
   (B‐SHARP) at Emory University. B‐SHARP participants undergo baseline
   cognitive assessments, neuroimaging, and lumbar punctures and are
   subsequently enrolled clinical studies. This analysis used data from
   the 185 participants enrolled from March 2016 to January 2019 who had
   CSF obtained during their baseline evaluations. The protocol was
   approved by the Emory University Institutional Review Board prior to
   recruitment. Each participant provided a written informed consent.

HIGHLIGHTS

     * Metabolic signature is detectable in amnestic mild cognitive
       impairment (MCI), a prodromal state for Alzheimer's disease.
     * This signature includes dysregulation in sugars, methionine, and
       tyrosine metabolism.
     * S‐adenosylmethionine is under‐ and S‐adenosylhomocysteine is
       overexpressed in MCI.

RESEARCH IN CONTEXT

    1. Systematic review: The authors searched PUBMED and Google Scholar
       for previous reports of metabolomics and Alzheimer's disease (AD).
       Search terms included: mild cognitive impairment, Alzheimer's
       disease “AND” metabolism, metabolomics. This search resulted in the
       following findings: Prior studies have either included small
       samples, used targeted approaches, or focused on plasma profiling.
       In this study, we conducted a case‐control untargeted high
       resolution metabolomic study on the cerebrospinal fluid of a larger
       sample of normal cognition and mild cognitive impairment (MCI).
    2. Interpretation: We discovered that multiple pathways, including
       pathways in sugar, methionine and homocysteine, and tyrosine
       metabolism were dysregulated in AD. Further, features that were
       significantly different between MCI and normal cognition had
       different patterns of association with cognitive, neuroimaging, and
       amyloid and tau biomarkers.
    3. Future direction: These pathways offer new potential targets for
       AD.

   The sample includes community‐dwelling adults 50 years or older with NC
   or amnestic MCI. Potential study participants were identified either
   through a referral from the Goizueta Alzheimer's Disease Research
   Center at Emory or through strategic community partnerships with grass
   roots health education organizations, health fairs, advertisements, and
   mail out announcements. An appropriate study informant, defined as an
   individual who has regular contact with the participant at least once a
   week (in person or telephone), was also identified for each
   participant. The potential study participant attended a screening
   visit, during which they underwent cognitive testing. A study physician
   also performed a clinical evaluation, cognitive interview, and a lumbar
   puncture (LP).

2.2. Cognitive diagnosis and exclusionary criteria

   Amnestic MCI categorization was done using modified Peterson criteria.
   This modification included using the Montreal Cognitive Assessment
   (MoCA)[52] ^20 instead of Mini‐Mental State Exam.[53] ^21 MCI criteria
   included subjective memory complaints, a MoCA < 26, Clinical Dementia
   Rating (CDR) score, memory sum of boxes=0.5,[54] ^22 education adjusted
   cutoff score on Logical Memory delayed recall of the Wechsler Memory
   Scale,[55] ^23 and preserved Functional Assessment Questionnaire
   (FAQ)<=7.[56] ^24 Individuals with amnestic MCI are at high risk for
   progression into dementia due to AD and hence may be considered a
   prodromal state for AD.[57] ^25

   NC was defined as having no significant memory complaints beyond those
   expected for age, a MoCA score >26 points, a CDR score of 0 (including
   0 on the Memory Box score), and preserved FAQ <=7. Participants were
   excluded if they had a history of stroke in the past 3 years, were
   unwilling or unable to undergo study procedures including MRI and LP,
   did not have a study informant, had a clinical diagnosis of dementia of
   any type, or abnormal serum thyroid stimulating hormone (>10) or B12
   (<250).

2.3. Cognitive assessment and biomarker measurements

   Demographics (age, sex, education), anthropometrics (weight and
   height), medical diagnosis, and medications were collected at baseline
   by interview. Cognitive assessment included those described above plus
   Trail Making Tests (TMT Part A and B) a measure of executive function
   and Hopkins Verbal Learning Test (HVLT) for episodic memory. Cognitive
   assessment was performed by trained personnel supervised by the study
   neuropsychologist. After a fast of at least 6 hours, CSF samples were
   collected via LP using 24G Sprotte atraumatic spinal needles. Samples
   were collected in sterile polypropylene tubes, separated into 0.5cc
   aliquots, and stored at −80°C. Samples were subsequently shipped to and
   analyzed by the Biomarker Research Laboratory at the University of
   Pennsylvania (Dr. Leslie Shaw).[58] ^26 CSF biomarkers: Aβ, t‐tau, and
   p‐tau were measured using the multiplex with the multiplex xMAP Luminex
   platform (Luminex Corp, Austin, TX) with Innogenetics (INNO‐BIA
   AlzBio3; Ghent, Belgium; for research use–only reagents) immunoassay
   kit–based reagents. The test–retest reliabilities are 0.98, 0.90, and
   0.85 for t‐tau, Aβ, and p‐tau181p, respectively.[59] ^26

2.4. MRI brain imaging

   Brain magnetic resolution imaging (MRIs) were also completed at Emory
   University (3.0 Tesla Trio MRI scanner, Siemens Medical Solutions,
   Malvern, PA). Anatomical images were acquired using high‐resolution
   three‐dimensional (3D) magnetization‐prepared rapid acquisition with
   gradient echo (MPRAGE). Images were then digitally saved for offline
   processing. Hippocampal volume and other volumetric measurements were
   calculated using the free surfer package with manual supervision.
   Quality checks were performed for each scan. Left and right hippocampal
   volumes were obtained and combined to derive the total hippocampal
   volume and cortical thickness. Intra‐cranial volume (ICV, mm^3) was
   also derived from this analysis. Volumetric measurements using free
   surfer has been shown to provide similar estimates to a fully manual
   procedure.[60] ^27 We used ICV‐adjusted hippocampal volume to reflect
   the degree of neurodegeneration for each participant.[61] ^28

2.5. Untargeted metabolomic high‐resolution metabolomics approaches and
pipeline

   Our metabolomic approaches used an established pipeline developed at
   the Clinical Biomarker Laboratory, led by Dr. Dean Jones (diagrammatic
   representation of this pipeline is included in Figure S1 in supporting
   information). High‐resolution metabolomics (HRM) was completed using
   established methods by an analyst blinded to sample identity.[62] ^8 ,
   [63]^29 Briefly, CSF samples were prepared and analyzed in batches of
   20. Prior to analysis, CSF aliquots were removed from storage at −80°C
   and thawed on ice. A 65 µL aliquot of CSF was then treated with 130 µL
   of liquid chromatography‐mass spectrometry (LC‐MS) grade acetonitrile,
   equilibrated for 30 minutes on ice and centrifuged (16.1 × g at 4°C)
   for 10 minutes to remove precipitated proteins. The supernatant was
   added to an autosampler vial and maintained at 4°C until analysis.
   Sample extracts were analyzed using LC and Fourier transform
   high‐resolution mass spectrometry (Dionex Ultimate 3000, Q‐Exactive HF,
   Thermo Scientific). For each sample, 10 µL aliquots were analyzed in
   triplicate using hydrophilic interaction liquid chromatography (HILIC)
   with electrospray ionization source operated in positive mode. This use
   of complementary chromatography phases and ionization polarity has been
   shown to improve detection of endogenous and exogenous chemicals.[64]
   ^30 Analyte separation was accomplished by HILIC using a 2.1 mm × 100
   mm × 2.6 µm Accucore HILIC column (Thermo Scientific) and an eluent
   gradient (A = 2% formic acid, B= water, C= acetonitrile) consisting of
   an initial 1.5 minutes period of 10% A, 10% B, 80% C, followed by
   linear increase to 10% A, 80% B, 10% C at 6 minutes and then held for
   an additional 4 minutes, resulting in a total runtime of 10 minutes per
   injection. Mobile phase flow rate was held at 0.35 mL/min for the first
   1.5 minutes, increased to 0.5 mL/min, and held for the final 4 minutes.

   The high‐resolution mass spectrometer was operated in full scan mode at
   120,000 resolution and mass‐to‐charge ratio (m/z) range 85–1275. Probe
   temperature, capillary temperature, sweep gas, and S‐Lens RF levels
   were maintained at 200°C, 300°C, 1 arbitrary units (AU), and 45 AU,
   respectively, for both polarities. Positive tune settings for sheath
   gas, auxiliary gas, sweep gas and spray voltage setting were 45 AU, 25
   AU, and 3.5 kV, respectively. Raw data files were extracted and aligned
   using apLCMS[65] ^31 with modifications by xMSanalyzer.[66] ^32
   Uniquely detected ions consisted of accurate mass m/z, retention time
   and ion abundance, referred to as m/z features. Data filtering was
   performed to remove m/z features with median coefficient of variation
   within technical replicates ≥75%. Additionally, only samples with
   Pearson correlation within technical replicates ≥0.7 were used for
   downstream analysis. Feature intensities for triplicates were median
   summarized with the requirement that at least two replicates had
   non‐missing values. Batch‐effect correction was performed using
   ComBat.[67] ^33

2.6. Metabolome‐wide association analysis

   A feature was retained for further analysis if at least 90% of the
   subjects had non‐zero intensity reading in either MCI or NC groups.
   After exclusion, the missing values for a feature were imputed as half
   of the lowest signal detected for that feature across all samples.
   After data filtering, all intensity values were log[2] transformed to
   reduce heteroscedasticity and quantile normalized to reduce systematic
   errors due to technical and other non‐biological factors.
   Metabolome‐wide association analysis was conducted using partial least
   squares discriminant analysis (PLS‐DA) implemented in the mixOmics[68]
   ^34 R package and features were selected based on the variable
   importance for projection (VIP) criteria. P‐values were obtained for
   each feature using a permutation test. A 1000‐permutation approach was
   performed by randomly shuffling the group labels of subjects and
   performing feature selection using PLS‐DA at each iteration.[69] ^30
   Multiple testing correction was performed using Storey and Tibshirani
   false discovery rate (FDR) adjustment.[70] ^35 Discriminatory features
   were selected using the thresholds of VIP ≥ 2, permutation derived P <
   .05, and FDR < 0.1. Only features that passed all three criteria were
   considered significantly different between the two groups. Manhattan
   plot was used to visualize the pattern of differential expression
   across all features with respect to retention time. Fold change of
   log[2] transformed intensity values was calculated for each feature as
   the difference between the average intensity of the two groups, log[2]
   FC=average[NC]‐average[MCI]. Two‐way hierarchical clustering analysis
   (HCA) was used to visualize the clustering pattern of discriminatory
   features and samples.

2.7. Pathway analysis

   Pathway enrichment analysis was performed using mummichog (v2.0.6),
   which uses both m/z and retention time, and included discriminatory
   features that met the following criteria: VIP ≥ 1.5, P < .05, and FDR
   <0.1. A lower VIP was used to increase enrichment within the pathway
   and prevent information loss.[71] ^10 , [72]^36 Detailed descriptions
   of mummichog computational procedures were previously published for
   V1.0.[73] ^37 Discriminatory features detected in the pathways were
   further tested for differential expression between the NC and MCI
   groups using Wilcoxon rank sum test.

   Additional pathway analyses were performed using Cystoscape‐based
   metabolomic pathway analysis and visualization using Metscape.[74] ^38
   MetScape is a plug‐in for Cytoscape, an open source software platform
   for visualizing complex networks and provides a method to use
   experimental data leveraged with bioinformatic databases of
   metabolites, genes, and pathways to display them in the context of
   system networks. We used that tool to provide complementary information
   on possible metabolic differences between MCI and NC.

2.8. Metabolite annotation and identification

   Metabolite annotation and identification was performed using MS/MS,
   comparison with in‐house library of confirmed metabolites, and using
   xMSannotator[75] ^33 with the Human Metabolome Database[76] ^34 (HMDB).
   Discriminatory features that were associated with the significantly
   enriched pathways and had P < .05 using the Wilcoxon rank sum test were
   selected for MS/MS analysis. For MS/MS, samples were analyzed using a
   Thermo Fusion Orbitrap high‐resolution (120,000 mass resolution) mass
   spectrometer (Thermo Fisher Scientific, San Diego, CA) operated in
   positive ion mode with 5‐minutes HILIC column chromatography and
   similar source conditions used for the untargeted metabolic profiling.
   Prior to analysis, CSF proteins were precipitated using
   acetonitrile:water (2:1 vol/vol) and allowed to sit on ice for 30
   minutes. The supernatant was then carefully pipetted for MS/MS
   analysis. The tandem mass spectrometry data were processed using the
   xcmsSet and xcmsFragments functions in XCMS[77] ^39 to extract the
   MS/MS fragments associated with each parent mass and the experimental
   spectra were compared to in‐silico fragmentation using MetFrag[78] ^40
   or the spectra available from mzCloud ([79]https://www.mzcloud.org/).

   We further annotated and confirmed identities of the selected
   metabolites using an in‐house library of metabolites that have been
   previously confirmed by comparing the retention time and MS/MS of the
   metabolic feature with authentic standards. Additionally, we performed
   computational annotation using xMSannotator[80] ^33 (v1.3.2) with the
   HMDB[81] ^34 (v3.5). xMSannotator uses adduct/isotope patterns,
   correlation in intensities across all samples, retention time
   difference between adducts/isotopes of a metabolite, and network and
   pathway associations for associating m/z features with known
   metabolites and categorizing database matches into different confidence
   levels.[82] ^41 This multi‐step annotation process reduces the number
   of false matches compared to only m/z‐based database search. Metabolite
   identification levels were assigned using an adapted version of the
   criteria proposed by Schymanski et al.: (1) confirmed by MS/MS and
   co‐elution with authentic standards (level 1); (2) confirmed by MS/MS
   and matches with online databases or in‐silico predicted spectra (level
   2); (3) confirmed by MS/MS at the chemical class level, but no evidence
   for a specific metabolite (level 3); (4) computationally assigned
   annotation using xMSannotator (medium or high confidence) (level 4);
   (5) accurate mass match (level 5).[83] ^42

2.9. Association of discriminatory features with other disease phenotypes

   Discriminatory metabolites associated with significantly enriched
   pathways were then tested for associations with three AD phenotypical
   or endophenotypic areas: cognitive performance (MoCA for global
   function, TMT A and B for executive function, and HVLT‐delayed recall
   for episodic memory), neuroimaging (hippocampal volume and cortical
   thickness as indicators of neurodegeneration) and CSF AD biomarkers
   Aβ[1‐42] and total and phosphorylated tau using Spearman's correlation
   analyses. A heatmap was used to visualize the correlation patterns
   between significant metabolic features and these measures.

3. RESULTS

3.1. Participants

   Of the 185 participants who provided CSF, 93 were MCI and 92 were
   normal controls. The basic clinical characteristics of the sample are
   provided in Table [84]1. The MCI group were older (P = .007) and had
   higher levels of tau and p‐tau (both P < .0001), but not Aβ (P =
   .6).They also had lower cognitive performance in all measures as
   expected and lower hippocampal volume (P < .0001).

TABLE 1.

   Characteristics of the overall sample by the two groups, normal
   cognition and mild cognitive impairment (MCI)
   Characteristic Overall (n = 185) Normal cognition (n = 92) Mild
   cognitive impairment (n = 93) P value
   Age (years)
   Mean ± SD (N) 64.4 ± 8.2 (185) 62.7 ± 7.1 (92) 66.1 ± 8.9 (93) .0071
   Sex
   Female 116 (62.7) 62 (67.4) 54 (58.1) .19
   Male 69 (37.3) 30 (32.6) 39 (41.9)
   Race
   White 117 (63.2) 63 (68.5) 54 (58.1) .24
   Black or African American 65 (35.1) 27 (29.3) 38 (40.9)
   Other 3 (1.6) 2 (2.2) 1 (1.1)
   Education (years)
   Mean ± SD (N) 16.3 ± 2.9 (179) 16.6 ± 3.0 (92) 15.9 ± 2.9 (87) .23
   Smoking status
   Never 44 (37.9) 32 (43.8) 12 (27.9) .0014
   Current 21 (18.1) 6 (8.2) 15 (34.9)
   Past 51 (44.0) 35 (47.9) 16 (37.2)
   EtOH consumption
   Current 88 (89.8) 66 (93.0) 22 (81.5) .09
   Never or remote 10 (10.2) 5 (7.0) 5 (18.5)
   Body mass index (kg/m^2)
   Mean ± SD (N) 27.3 ± 5.5 (178) 27.4 ± 5.0 (92) 27.3 ± 6.0 (86) .66
   Systolic blood pressure (mm Hg)
   Mean ± SD (N) 128.9 ± 18.1 (179) 129.6 ± 19.3 (92) 128.1 ± 16.8 (87)
   .80
   Diastolic blood pressure (mm Hg)
   Mean ± SD (N) 75.1 ± 12.4 (179) 76.9 ± 12.1 (92) 73.1 ± 12.5 (87) .037
   Pulse rate, beats per min
   Mean ± SD (N) 67.8 ± 10.8 (179) 66.9 ± 10.0 (92) 68.8 ± 11.5 (87) .30
   Hypertension
   Yes 79 (49.7) 52 (57.8) 27 (39.1) .020
   No 80 (50.3) 38 (42.2) 42 (60.9)
   High cholesterol
   Yes 70 (44.6) 43 (48.9) 27 (39.1) .22
   No 87 (55.4) 45 (51.1) 42 (60.9)
   Diabetes mellitus
   Yes 21 (13.1) 11 (12.2) 10 (14.3) .70
   No 139 (86.9) 79 (87.8) 60 (85.7)
   Heart disease
   Yes 13 (8.2) 6 (6.7) 7 (10.1) .43
   No 146 (91.8) 84 (93.3) 62 (89.9)
   Congestive heart failure
   Yes 4 (2.5) 1 (1.1) 3 (4.3) .20
   No 155 (97.5) 89 (98.9) 66 (95.7)
   Depression
   Yes 47 (29.4) 20 (22.2) 27 (38.6) .024
   No 113 (70.6) 70 (77.8) 43 (61.4)
   Atrial fibrillation or arrythmias
   Yes 20 (12.7) 11 (12.4) 9 (13.2) .87
   No 137 (87.3) 78 (87.6) 59 (86.8)
   MOCA, score
   Mean ± SD (N) 24.3 ± 3.7 (162) 26.6 ± 2.6 (92) 21.3 ± 2.8 (70) <.0001
   HVLTR, delayed recall
   Mean ± SD (N) 8.1 ± 3.2 (161) 9.7 ± 2.0 (92) 6.0 ± 3.4 (69) <.0001
   Trail Part A (seconds)
   Mean ± SD (N) 39.3 ± 16.6 (162) 34.9 ± 11.1 (92) 45.1 ± 20.6 (70) .0008
   Trail Part B (seconds)
   Mean ± SD (N) 108.8 ± 67.1 (161) 83.3 ± 41.0 (92) 142.8 ± 79.3 (69)
   <.0001
   Ab42 (pg./dl)
   Mean ± SD (N) 255.5 ± 83.2 (183) 256.0 ± 61.1 (91) 255.1 ± 100.8 (92)
   .61
   tau (pg./dl)
   Mean ± SD (N) 60.4 ± 35.8 (183) 48.4 ± 20.6 (91) 72.2 ± 43.2 (92)
   <.0001
   Ptau (pg/dl)
   Mean ± SD (N) 15.7 ± 9.5 (180) 12.3 ± 6.5 (90) 19.0 ± 10.9 (90) <.0001
   Total hippocampal volume
   Mean ± SD (N) 7303 ± 1046 (139) 7654 ± 881.8 (80) 6828 ± 1071 (59)
   <.0001
   [85]Open in a new tab

   Abbreviations: EtOH, ethyl alcohol; HVLTR, Hopkins Verbal Learning
   Test‐Revised; MOCA, Montreal Cognitive Assessment; SD, standard
   deviation.

3.2. MWAS results

   Overall, 13,064 features were detected, and 8043 features met the data
   filtering criteria and were used for downstream analyses. Using PLS‐DA,
   294 discriminatory features were identified using the predefined
   criteria (Figure [86]1A). Of those, 107 features were underexpressed
   and 187 features were overexpressed in MCI patients relative to NC, as
   shown in Figure [87]1A. Two‐way HCA using the 294 discriminatory
   features identified 19 clusters of samples indicating clinical and
   metabolic heterogeneity within the MCI group (Figure S2 in supporting
   information). Clusters 13, 9, and 15 (blue box) primarily comprised the
   MCI samples. Seventy clusters comprising features with similar
   abundance levels across samples were identified. Of the 294 features,
   94 were successfully matched to known metabolites in HMDB using
   xMSannotator with an annotation confidence score of medium or high
   (Table S1 in supporting information).

FIGURE 1.

   FIGURE 1
   [88]Open in a new tab

   A, Manhattan plot shows the variable importance for projection (VIP)
   and mass‐to‐charge ratio (m/z) of 8043 features. A total of 294
   features were significantly different between mild cognitive impairment
   (MCI) cases (n = 93) and controls (n = 92) by partial least squares
   discriminant analysis (PLS‐DA) using a VIP measure of 2.0 or greater
   (threshold indicated by horizontal line). One hundred eighty seven
   metabolic features increased (red dots) and 107 decreased (blue dots)
   in MCI patients compared to controls are indicated. Metabolite classes
   detected at different retention time segments are annotated in the
   boxes. B, Pathways altered in MCI compared to normal controls. Pathway
   analysis was performed using Mummichog 2.0.6 on the 1049 features
   identified by PLS‐DA with a VIP ≥ 1.5

3.3. Pathway analysis

   To enhance the coverage of metabolites for pathway enrichment analyses
   and to prevent information loss, 1049 discriminatory features were
   included using the less stringent criteria of VIP >1.5, P < .05, and
   FDR <0.1. We identified 13 pathways that were perturbed between the MCI
   and normal control groups, which are shown in Figure [89]1B. The top
   four pathways were related to bioenergetics and glucose metabolism:
   N‐glycan (P = .0007), sialic acid (P = .0014), amino‐sugars (P =
   .0042), and galactose (P = .0054) metabolism. Keratan sulfate (P =
   .0173), methionine (P = .0081), cyanocobalamin (P = .0106), tyrosine (P
   = .0193), purine (P = .0352) and biopterine (P = .0275) were also
   differentially activated between the two groups. Within the enriched
   pathways that were significantly different between NC and MCI, 15
   features with an identification confidence of 1 to 5 were
   differentially expressed and are shown in Table [90]2. These features
   were then included in the Metscape analysis and visualization, leading
   to a signature that includes increased expressions of features related
   to sugar metabolism/bioenergetics, homocysteine, tyrosine and biopterin
   pathways, and lower expression of methionine. The final networks with
   relevant signature features are provided in Figure [91]2. The box plots
   for these 15 features are also provided in Figure [92]3. The complete
   list of features in these analyses is provided in Table S2 in
   supporting information.

TABLE 2.

   Results of the pathway enrichment analysis with the significant
   features and associated pathways in the normal versus MCI groups
   m/z time (s) Feature name (KEGG compound name) Pathway(s)[93] ^a Fold
   change[94] ^d VIP Wilcoxon P Metabolite identification level[95] ^c
   Adduct
   173.0434 52 L‐Ribulose ([96]C00508)[97] ^b Tyrosine metabolism; Purine
   metabolism −0.1642 3.30 0.0001 3 M+Na[1+]
   205.0682 62 D‐Sorbitol ([98]C00794 )[99] ^b Galactose metabolism −0.124
   1.97 0.0025 5 M+Na[1+]
   365.1054 164 Maltose ([100]C00208 )[101] ^b Sialic acid metabolism;
   Galactose metabolism −0.2469 2.81 0.0055 5 M+Na[1+]
   385.1303 73 S‐Adenosylhomocysteine ([102]C00021) Methionine and
   cysteine metabolism; Vitamin B12 (cyanocobalamin) metabolism; Urea
   cycle/amino group metabolism; Tyrosine metabolism −0.2034 2.15 0.0167 1
   M+H[1+]
   517.9829 81 7,8‐Dihydroneopterin 3′‐triphosphate ([103]C04895)[104] ^b
   Biopterin metabolism −0.6569 2.32 0.0217 5 M+Na[1+]
   255.1076 68 Galactosylglycerol ([105]C05401) Sialic acid metabolism;
   Galactose metabolism −0.4977 2.72 0.0231 4 M+H[1+]
   260.0538 57 N‐Acetyl‐D‐glucosamine 6‐phosphate ([106]C00357)[107] ^b
   Aminosugars metabolism −0.1686 2.05 0.0248 5 M+H[1+]
   223.0826 53 Salsolinol 1‐carboxylate ([108]C06160) Tyrosine metabolism
   −0.374 2.22 0.0271 4 M+H[1+]
   708.2568 255 N‐acetyl‐alpha‐D‐glucosamine ([109]C00043)[110] ^b
   N‐Glycan degradation; Keratan sulfate degradation −0.308 2.15 0.0289 4
   M+H[1+]
   399.1444 145 S‐Adenosylmethionine ([111]C00019) Methionine and cysteine
   metabolism; Vitamin B12 (cyanocobalamin) metabolism; Urea cycle/amino
   group metabolism; Tyrosine metabolism 0.3296 2.17 0.0334 1 M+H[1+]
   384.1499 231 beta‐D‐Galactosyl‐1,4‐N‐acetyl‐D‐glucosamine
   ([112]C00611)[113] ^b N‐Glycan Degradation; Aminosugars metabolism;
   Galactose metabolism −0.1902 2.02 0.0357 5 M+H[1+]
   221.042 61 Vanillylmandelic acid ([114]C05584) Tyrosine metabolism
   −0.5378 1.92 0.0365 4 M+Na[1+]
   799.6688 37 Levothyroxine ([115]C01829) Tyrosine metabolism −0.7153
   2.31 0.0416 4 M+Na[1+]
   277.0894 67 3‐beta‐D‐Galactosyl‐sn‐glycerol ([116]C03692)[117] ^b
   Sialic acid metabolism; Galactose metabolism −0.1336 2.18 0.0474 4
   M+Na[1+]
   244.0797 49 GlcNAc ([118]C00140)[119] ^b N‐Glycan degradation; Sialic
   acid metabolism; Aminosugars metabolism; Galactose metabolism; Keratan
   sulfate degradation; Hyaluronan Metabolism −0.0803 2.22 0.0507 5
   M+Na[1+]
   [120]Open in a new tab
   ^a

   Some compounds were matched or involved in multiple pathways.
   ^b

   These features had multiple chemical names or were matched to multiple
   metabolites in the databases (we report the KEGG compound name involved
   in the significant pathway).
   ^c

   Description of metabolite identification levels (adapted from
   Schymanski et al.[121] ^42 ):

   Level 1: confirmed by MS/MS and co‐elution with authentic standards

   Level 2: confirmed by MS/MS and matches with online databases or
   in‐silico predicted spectra

   Level 3: confirmed by MS/MS at the chemical class level, but no
   evidence for a specific metabolite

   Level 4: computationally assigned annotation using xMSannotator (medium
   or high confidence)

   Level 5: accurate mass match
   ^d

   (log2; NC vs MCI) ‐ve: lower in NC +ve: higher in NC.

   Abbreviations: KEGG, Kyoto Encyclopedia of Genes and Genomes; MCI, mild
   cognitive impairment; MS/MS, tandem mass spectrometry; m/z,
   mass‐to‐charge ratio; NC, normal control; VIP, variable importance for
   projection

FIGURE 2.

   FIGURE 2
   [122]Open in a new tab

   Overall metabolomic networks and related features comprising the
   metabolomic signature in the cerebrospinal fluid of individuals with
   mild cognitive impairment (MCI). The image was obtained using Metscape
   Plug‐in for Cytoscape using the 15 features included in the MCI
   signature along with the fold change and P‐value and the
   Compound‐Reaction‐Enzyme‐Gene option selected. The Network is then
   built from the underlying data by finding compounds that participate in
   reactions that are catalyzed by enzymes that are encoded by genes

FIGURE 3.

   FIGURE 3
   [123]Open in a new tab

   Box plots for the 15 differentially expressed features identified in
   the cerebrospinal fluid of individuals with mild cognitive impairment.
   Green box plots are in the methionine pathway, brown in the sugar
   metabolism pathways, purple in the tyrosine metabolism pathway, and
   blue in the biopterin pathway

3.4. Correlation with disease phenotype

   We then explored the associations between these signature features with
   disease phenotypes. These results are shown in Figure [124]4. Increased
   expression of bioenergetics and glucose metabolism were associated with
   higher tau and ptau but also with lower cognitive performance,
   hippocampal volume, and cortical thickness. Sugar metabolism
   dysregulations were associated with increased tau, and ptau. Further, 5
   of these features were associated with decreased cortical thickness;
   hippocampal volume; and cognitive performance on MoCA, TMT, and delayed
   recall. S‐adenosylhomocysteine (SAH) was associated with lower MoCA
   scores and decreased cortical thickness. In the tyrosine pathway,
   salsolinol‐1‐carboxylate was associated with higher tau and ptau
   whereas Vanillylmandelic acid (VMA) was associated with lower MoCA
   score. Finally, features in the biopterin pathway were not associated
   with any disease phenotype. Detailed results are provided in Table S3
   in supporting information. Correlation with demographics are also
   provided in Table S4 in supporting information.

FIGURE 4.

   FIGURE 4
   [125]Open in a new tab

   Correlations between significant features and disease phenotypes.
   Spearman correlation coefficients. Red indicates positive correlations
   and blue indicates negative correlations. Correlations with P ≥0.05 are
   marked in gray. ^**These Features had multiple chemical names or were
   matched to multiple metabolites in the databases (we report the KEGG
   compound name involved in the significant pathway)

4. DISCUSSION

   This study of untargeted HRM identified a CSF signature of amnestic MCI
   characterized by dysregulation of multiple pathways including sugar,
   homocysteine/methionine, and tyrosine metabolism. Multiple features
   within this signature were associated with increased total tau and ptau
   biomarkers and lower scores on cognitive measures, hippocampal volume,
   and cortical thickness.

   Although targeted metabolomic approaches have been used in multiple
   tissues and samples such as Biocrates AbsoluteIDQ‐p180 kit,[126] ^43
   untargeted metabolomics in the CSF have been less common. The latter
   can be complementary to prior targeted platforms and plasma analyses
   and may identify new markers or new potential therapeutic targets in
   AD. The recent advance in MS technology and related bioinformatics have
   enhanced the potential for the application of metabolomics in AD.[127]
   ^44 Indeed, over the last decade reports using untargeted metabolomics
   have suggested a significant previously unrecognized metabolic
   derangement in AD post mortem brains, plasma, and to a lesser extent
   CSF.[128] ^45 , [129]^46 Our study adds to these reports by confirming
   and expanding on previously described impaired pathways or identifying
   new ones. We discuss these in the next sections.

   Multiple studies have suggested an association between AD and impaired
   glucose metabolism that may be pronounced in those with type 2 diabetes
   and insulin resistance.[130] ^47 , [131]^48 Prior
   fluorodeoxyglucose‐positron emission tomography (FDG‐PET) scans have
   suggested decreased brain metabolism across the spectrum of AD.[132]
   ^49 , [133]^50 Our study suggests that in the CSF of those with MCI,
   there was evidence for dysregulation of multiple glucose metabolism
   pathways and related increase in glucose metabolism byproducts. Taken
   together, the increase in CSF features of sugar metabolism pathways
   coupled with the previously reported brain hypometabolism may in part
   be explained by a lower brain glucose uptake, for example, secondary to
   glucose uptake transporter impairment,[134] ^51 , [135]^52 leading to
   increased CSF levels. An alternative explanation is that the possible
   central insulin resistance reported in AD is associated with increases
   in metabolic by‐products in the brain and CSF. This is further
   supported by our observation that these increased metabolic features
   are associated with greater tau measures and with lower performance on
   cognitive assessments, hippocampal volume, and cortical thickness.
   Taken together, our finding may offer support for multiple sugar
   metabolism pathways as therapeutic targets in AD.

   Our observation that alterations in pathways related to methionine and
   homocysteine metabolism is of great interest. Specifically,
   S‐adenosylmethionine (SAM) was underexpressed and SAH was overexpressed
   in CSF of the MCI participants. SAM is a key molecule in the methionine
   cycle involved in nucleic acid and protein metabolism and synthesis.
   SAH is formed by demethylation of SAM. Prior reports suggest that SAM
   is decreased and SAH is increased in CSF of AD and related to tau
   biomarkers.[136] ^53 However, in this study only SAM was related to
   additional disease phenotypes including cognitive measures and cortical
   thickness. Nevertheless, this untargeted approach suggests that
   homocysteine‐methionine pathways are dysregulated in the prodromal
   stages of AD.

   We identified perturbations in tyrosine pathways with overlapping
   features in the purine, methionine, and homocysteine pathways,
   including SAM, SAH, VMA, and thyroxine. These cycles are involved in
   catecholamine and serotonin neurotransmitter systems and might be
   altered in AD.[137] ^54 A prior CSF analysis in a smaller number of MCI
   using a targeted metabolomic approach suggested a similar finding of
   impairments in methionine and tyrosine pathways.[138] ^55 Despite the
   difference between the groups in this pathway, there were minimal
   associations with the other disease measures.

   There are multiple advantages to this study including the untargeted
   and advanced bioinformatic approaches, which allowed us to consider
   many pathways and features, the comparably larger number of samples
   with CSF, and the availability of multiple additional disease
   phenotypes that offer greater confidence in the associations with MCI.

   The limitations include the cross‐sectional design and the number of
   identified features that could not be matched to known metabolites or
   matched to multiple metabolites, which is a major bottleneck in
   untargeted metabolomics.[139] ^6 Even with identified metabolites, the
   certainty of feature identity is another limitation to untargeted
   metabolomic approaches. We used MS/MS with an in‐house library of
   confirmed metabolites in the Clinical Biomarker Lab where these
   analyses were performed using authentic standards to confirm these
   identities and we provide a standard scale of confidence in all our
   results. This coupled with advanced bioinformatics tools for metabolite
   identification and annotation enhanced the reliability of the identity
   of our metabolites compared to many prior untargeted studies.

   Clinical translations of these findings are important. The key pathways
   that are perturbed in AD are potential targets for existing or new drug
   developments. For example, insulin and other antidiabetic agents may
   address the sugar metabolism abnormalities identified in this
   analysis.[140] ^56 , [141]^57 Drugs that may restore balance between
   SAM and SAH or enhance tyrosine metabolism may also be of relevance in
   drug development of AD.[142] ^58

5. CONCLUSION

   In this untargeted metabolomic study of CSF, we identified a metabolic
   signature characterized by impairments in sugar metabolism and
   methionine, homocysteine, and tyrosine pathways in MCI. These offer
   insight into the metabolic alterations that occur in predementia stages
   of AD and offer potential therapeutic targets.

CONFLICTS OF INTEREST

   The authors have no competing interests to declare.

Supporting information

   Supplementary information
   [143]Click here for additional data file.^ (591.8KB, docx)

ACKNOWLEDGMENTS