Plasma lipidomics analysis finds long chain cholesteryl esters to be associated with Alzheimer's disease

Abstract

There is an urgent need for the identification of Alzheimer's disease (AD) biomarkers. Studies have now suggested the promising use of associations with blood metabolites as functional intermediate phenotypes in biomedical and pharmaceutical research. The aim of this study was to use lipidomics to identify a battery of plasma metabolite molecules that could predict AD patients from controls. We performed a comprehensive untargeted lipidomic analysis, using ultra-performance liquid chromatography/mass spectrometry on plasma samples from 35 AD patients, 40 elderly controls and 48 individuals with mild cognitive impairment (MCI) and used multivariate analysis methods to identify metabolites associated with AD status. A combination of 10 metabolites could discriminate AD patients from controls with 79.2% accuracy (81.8% sensitivity, 76.9% specificity and an area under curve of 0.792) in a novel test set. Six of the metabolites were identified as long chain cholesteryl esters (ChEs) and were reduced in AD (ChE 32:0, odds ratio (OR)=0.237, 95% confidence interval (CI)=0.10-0.48, P=4.19E-04; ChE 34:0, OR=0.152, 95% CI=0.05-0.37, P=2.90E-04; ChE 34:6, OR=0.126, 95% CI=0.03-0.35, P=5.40E-04; ChE 32:4, OR=0.056, 95% CI=0.01-0.24, P=6.56E-04 and ChE 33:6, OR=0.205, 95% CI=0.06-0.50, P=2.21E-03, per (log2) metabolite unit). The levels of these metabolites followed the trend control>MCI>AD. We, additionally, found no association between cholesterol, the precursor of ChE and AD. This study identified new ChE molecules, involved in cholesterol metabolism, implicated in AD, which may help identify new therapeutic targets; although, these findings need to be replicated in larger well-phenotyped cohorts.

OPEN

Citation: Transl Psychiatry (2015) 5, e494; doi:10.1038/tp.2014.127

http://www.nature.com/tp

Web End =www.nature.com/tp

ORIGINAL ARTICLE

Plasma lipidomics analysis nds long chain cholesteryl esters to be associated with Alzheimers disease

P Proitsi1, M Kim2, L Whiley2, M Pritchard1, R Leung1, H Soininen3, I Kloszewska4, P Mecocci5, M Tsolaki6, B Vellas7, P Sham8, S Lovestone9, JF Powell1,11, RJB Dobson1,10,11 and C Legido-Quigley2,11

There is an urgent need for the identication of Alzheimers disease (AD) biomarkers. Studies have now suggested the promising use of associations with blood metabolites as functional intermediate phenotypes in biomedical and pharmaceutical research. The aim of this study was to use lipidomics to identify a battery of plasma metabolite molecules that could predict AD patients from controls. We performed a comprehensive untargeted lipidomic analysis, using ultra-performance liquid chromatography/mass spectrometry on plasma samples from 35 AD patients, 40 elderly controls and 48 individuals with mild cognitive impairment (MCI) and used multivariate analysis methods to identify metabolites associated with AD status. A combination of 10 metabolites could discriminate AD patients from controls with 79.2% accuracy (81.8% sensitivity, 76.9% specicity and an area under curve of 0.792) in a novel test set. Six of the metabolites were identied as long chain cholesteryl esters (ChEs) and were reduced in AD (ChE 32:0, odds ratio (OR) = 0.237, 95% condence interval (CI) = 0.100.48, P = 4.19E 04; ChE 34:0, OR = 0.152, 95% CI = 0.050.37,P = 2.90E 04; ChE 34:6, OR = 0.126, 95% CI = 0.030.35, P = 5.40E 04; ChE 32:4, OR = 0.056, 95% CI = 0.010.24, P = 6.56E 04 and ChE 33:6, OR = 0.205, 95% CI = 0.060.50, P = 2.21E 03, per (log2) metabolite unit). The levels of these metabolites followed the trend control4MCI4AD. We, additionally, found no association between cholesterol, the precursor of ChE and AD. This study identied new ChE molecules, involved in cholesterol metabolism, implicated in AD, which may help identify new therapeutic

targets; although, these ndings need to be replicated in larger well-phenotyped cohorts.

Translational Psychiatry (2015) 5, e494; doi:10.1038/tp.2014.127; published online 13 January 2015

INTRODUCTIONA better understanding of the biological mechanisms underlying

Alzheimers disease (AD) is required. AD is a devastating illness that currently affects over 496,000 people in the UK (http://www.alz.org.uk

Web End =www.alz.org. http://www.alz.org.uk

Web End =uk ). It is one of the major challenges for health care in the 21st century and with estimated longer life expectancy, the number of demented patients worldwide are expected to reach 81.1 million in 2040.1 The lack of current treatments for AD and the lack of a denite and early diagnosis highlights the absence of a comprehensive understanding of the biological mechanisms underlying the changes that occur during the process of neurodegeneration. This underscores the urgent need for biomarkers that would lead to novel treatment strategies and improve the lives of those affected.2 Current sulcal cerebrospinal uid or brain imaging biomarkers are costly, can cause discomfort to the patient and are impractical at large scale. For example, the Medicare Evidence Development and Coverage Advisory Committee concluded that there was too little evidence to show that amyloid positron emission tomography (PET) scans improve AD outcomes or that its benets outweigh the harms and high cost.3 A blood-based biomarker could act as a screening tool to identify

at-risk individuals for further investigation or recruitment into clinical trials.

Studies have now demonstrated the potential of using blood metabolites, the repertoire of molecules (size o10001500 Da)4 present in cells and tissue,5 as functional intermediate phenotypes in biomedical and pharmaceutical research.6 Metabolic phenotyping and its subset lipidomics is the rapidly evolving eld of the comprehensive measurement, in a non-targeted manner, of ideally all endogenous metabolites in a biological sample.7 These small-to-medium molecules are the nal products of interactions between gene expression, protein expression and the cellular and external environment, and represent an essential aspect of the phenotype of an organism.8 In addition, it is now known that there is a communication between the brain and the periphery and is increasingly believed that the damage to the bloodbrain barrier caused by AD increases theoretically the chance of metabolites crossing to the brain.2,9,10 Coupled with the fact that blood is relatively easily accessible, plasma metabolites are an ideal source of noninvasive biomarkers and a molecular footprint of disease.

Recently, a number of non-targeted blood metabolomic studies (liquid chromatography-mass spectrometry (LC-MS) or direct infusion mass spectrometry (MS)) in AD have emerged,

1Institute of Psychiatry, Psychology and Neuroscience, King's College London, London, UK; 2Institute of Pharmaceutical Science, King's College London, London, UK; 3Department of Neurology, Kuopio University Hospital and University of Eastern Finland, Kuopio, Finland; 4Department of Old Age Psychiatry & Psychotic Disorders, Medical University of d, d, Poland; 5Section of Gerontology and Geriatrics, Department of Medicine, University of Perugia, Perugia, Italy; 6Memory and Dementia Center, 3rd Department of Neurology, Aristotle University of Thessaloniki, Thessaloniki, Greece; 7Department of Internal and Geriatrics Medicine, INSERM U 1027, Gerontopole, Hpitaux de Toulouse, Toulouse, France;

8Department of Psychiatry, State Key Laboratory of Brain and Cognitive Sciences, and Centre for Genomic Sciences, Li Ka Shing Faculty of Medicine, the University of Hong Kong, Hong Kong; 9Department of Psychiatry, University of Oxford, Warneford Hospital, Oxford, UK and 10National Institute for Health Research Biomedical Research Centre for Mental Health and Biomedical Research Unit for Dementia at South London and Maudsley NHS Foundation, London, UK. Correspondence: Dr P Proitsi, Department of Basic and Clinical Neuroscience, Institute of Psychiatry, Psychology and Neuroscience, King's College London, De Crespigny Park, London SE5 8AF, UK.

E-mail: mailto:[email protected]

Web End [email protected]

11These senior authors contributed equally to this work.

Received 2 June 2014; revised 28 September 2014; accepted 26 October 2014

Plasma cholesterol metabolites in AD P Proitsi et al

2 highlighting the role of lipid compounds, such as sphingolipids,11

bile acids,12 desmosterol13 and phosphatidylcholines (PCs).1416 In a LC-MS/NMR screen followed by subsequent quantication, we previously identied three lipid PC molecules (PC16:0/20:5, PC16:0/22:6 and 18:0/22:6) that were progressively diminished in subjects with mild cognitive impairment (MCI) and AD patients.16

Interestingly, the latter two were found to be markers in a recent study where a panel of 10 lipids showed 90% area under curve (AUC) for preconversion to amnestic MCI or AD.14

Lipid metabolism has been extensively implicated in the pathogenesis of AD through cell biological17,18 and genetic

studies.1922 Further support comes from brain tissue metabolomic studies, which have shown changes in lipid compounds such as lysobisphosphatidic acid, sphingomyelin, sphingolipids and desmosterol in the brains of AD patients.2326

In this study, we performed untargeted lipidomics utilizing plasma samples from 36 AD patients, 40 healthy controls and 48 individuals with MCI with the aim of discovering new molecules that could diagnose AD patients from controls and MCI and improve our current knowledge of molecules associated with AD and their underlying biology.

MATERIALS AND METHODS Patient sample collection

This study utilized 124 age and gender matched plasma samples (36 AD patients, 48 MCI individuals and 40 controls) from the Dementia Case Register (DCR) at Kings College London and the EU funded AddNeuroMed study.16,27 Normal elderly control subjects were recruited from non-related family members of AD patients, care-givers relatives, social centres for the elderly or GP surgeries and had no evidence of cognitive impairment. AD and MCI subjects were recruited primarily from local memory clinics, and as such the MCI cohort was expected to be composed largely of subjects with a likely AD end point. Overlapping samples from this study were previously used by Whiley et al.16 to measure the relative amounts of three PCs to a PC internal standard in plasma. Relevant ethics board approved the study and informed consent was obtained for all subjects. Each patient was required to fast for 2 h before sample collection and 10 ml of blood was then collected in tubes coated with sodium ethylenediaminetetraacetic acid to prevent clotting. Whole blood was centrifuged to form a plasma supernatant, which in turn was removed and placed at 80 C until further use.

LipidomicsSample treatment. Sample treatment has been described elsewhere,4,16

20 l of plasma was added to a glass HPLC vial containing a 400 l glass insert (Chromacol, Welwyn Garden City, UK). Ten microlitres of high purity water and 40 l of MS-grade methanol were added to each sample, followed by a 2 min vortex mix to precipitate proteins. Then, 200 l of methyl t-butyl ether (containing 10 g ml1 of internal standard triacylglycerol 45:0) was added, and the samples were mixed via vortex at room temperature for 1 h. After the addition of 50 l of high purity water, a nal sample mixing was performed before centrifugation at 3000 g for 10 min. The upper, lipid-containing, methyl t-butyl ether phase was then injected onto the LC-MS system directly from the vial by adjustment of the instrument needle height (17.5 mm from bottom).

The LC-MS-MS method employed has previously been published2,16 and

has shown to measure amounts of 44500 metabolite species, particularly lipids. Instrumentation included a Waters ACQUITY UPLC and XEVO QTOF system (Waters, Milford, MA, USA).

Chromatographic separation was achieved using an Agilent (Palo Alto, CA, USA) Poroshell 120 EC-C8 column (150 mm 2.1 mm, 2.7 m), maintained at 60 C. A gradient was used consisting of 10 mM ammonium formate in water (A) and 10 mM ammonium formate in methanol (B). The solvent was delivered at a ow rate of 0.5 ml min1. The gradient consisted of 0 min (75% B), 23 min (96% B), 36 min (96% B), 36.5 min (100% B),41.5 min (100% B), 42 min (75% B), 51 min (75% B).

The XEVO QTOF was operated in the positive ion mode with a capillary

voltage of 2.5 kV and a cone voltage of 60 V. The desolvation gas ow was 500 l h1 and the source temperature was 120 C. All analyses were acquired using the lock spray setting; leucine enkephalin was used as lock

mass (m/z 556.2771 and 278.1141). Data were collected in the centroid mode over the mass range m/z 1001000 with an acquisition time of 1 s per scan.

Samples were analysed in a randomized order, in four batches, with pooled plasma sampled (quality control (QC) samples) at regular intervals throughout the run (n = 20). LC-MS raw data were aligned and normalized to total mean area, using Waters MarkerLynx software (Waters). Further validation of metabolite concentrations, which were associated with disease status took place using Waters QuanLynx software (Waters) and calculating peak ratios for metabolites that were over the LOQ (limit of quantication) and the area under the peak of the internal standard.

Statistical analysisData pre-treatment. Metabolites detected in o80% of each of the diagnostic groups and the four batches were excluded from analyses.

Metabolite distributions were inspected using histograms and the ShapiroWilks test was used to test for normality of the metabolite distributions. The distribution of a large number of metabolites was skewed and the data were further log2 transformed. The Empirical Bayes method ComBat28 was used to correct for batch effects in each metabolite. Principal components analysis was then used for the detection of outliers and to check whether the 20 QC samples clustered together. Missing data points were imputed using KNN (k-nearest neighbours, k = 10), separately for each disease phenotype (impute). Metabolite correlations were visualized using a heatmap (ggplot2). Analysis of variance was used to test for differences in the levels of continuous variables between the three diagnostic groups, followed by Tukeys honest signicant differences post hoc test between pairs of groups when the results of analysis of variance were signicant at the Po0.05 level. Pearsons 2 was used to test for differences in frequency of the categorical variables between the diagnostic groups. All the analyses took place in R.3.01.

Single analyte analysis. Analytes were centred around their mean. Logistic regression (glm) was used to investigate the association of each metabolite individually with disease status (AD versus controls). Logistic regression models were adjusted for the number of APOE 4 alleles, age, gender and batch. False discovery rate (FDR) correction (0.05) was applied to correct for multiple testing. Secondary models were run to test whether the association of metabolites with disease status was modied by the presence of the APOE 4 allele by testing for interactions.

Multivariate analysis. A Random Forest approach (using rf and rfe in the package CARET) was used to develop an AD versus control classier. Due to the large number of variables and their highly correlated structure, the following approach was used to achieve high performance with a minimal variable set. AD cases and controls were divided into a training (2/3) and an independent test data set (1/3) such that the training set comprised equal numbers of each diagnostic group. The training set was further divided into 100 bootstrap sample sets comprising a bootstrap training set (75%) used to build the Random Forest model and a bootstrap test set (25%) used to test the model. In each model, the default setting for ntree = 500 was used and the optimum mtry number after 100 bootstraps was chosen and tted to the whole training data set. The AUC was used to test for the performance of each classier.

In each of the 100 bootstrap iterations, each variable was assigned a variable importance score. The ranks were summed for each metabolite over the 100 bootstraps providing an indication of the predictive power of each variable. We then selected the top 10% analytes on the basis of their summed variable importance rankings and ran the Random Forest with recursive feature elimination (rfe), that is, backward elimination on these selected variables using a second round of 100 bootstraps. The boostrapping was repeated keeping the top 50 to 10 analytes in steps of ve, and down to two analytes in steps of one. For each subset of predictors, the mean bootstrap testing performance was calculated, and, on the basis of this, the optimal number of variables was identied using the sizeTolerance function. This takes into account the whole prole during rfe and picks a subset size that is small without sacricing too much performance. The optimal number of variables was then used to build a nal model in the complete training data, which was tested with the independent test set (Model 1).

The nal model was also tested on the MCI sample to determine whether it would classify MCI as cases or controls.

Translational Psychiatry (2015), 1 8

Plasma cholesterol metabolites in AD P Proitsi et al

3

We then repeated the steps above including APOE 4 genotype (0,1 or 2 alleles) during the initial Random Forest bootstrapping step to assess its variable importance and the predictive ability of the metabolites in the presence of APOE.

RESULTSThe sample demographics are displayed in Table 1 and

Supplementary Figure 1 gives a detailed account of the QC steps. In brief, 1878 molecular features were extracted from the 124 samples. Following QC, metabolites detected in o80% of the three diagnostic groups and the four batches were removed and 573 features were left for analysis. The distribution of 490% of the 573 features deviated from the normal, reecting in some cases different batch distributions. The features therefore underwent log2 transformation and batch effect correction, resulting in 495% metabolites following a normal distribution. Principal components analysis showed that all 20 QC samples clustered together (Supplementary Figure 2), verifying reproducible results across the batches and highlighting the presence of one outlier (AD case), which was removed. KNN imputation was performed on the 123 samples with the 573 features.

Single analyte analyses resultsLogistic regression analyses for AD versus controls, adjusting for the number of APOE 4 alleles, age, gender and batch, indicated that 95 analytes were associated with AD at P-value o0.05 and 41 analytes at q-value o0.05. Results for all 573 analytes and all pairwise comparisons are provided in Supplementary Table 1. Secondary models investigating for APOE 4 specic associations by testing for interactions between analytes and the APOE 4 allele indicated that the association of 29 analytes with AD seemed to be modied for the APOE 4 at P-value o0.05;

However, none of these associations were signicant at q-value o0.05 (results not presented).

Multivariate analysis resultsRandom Forest was performed on the training data set using the 573 features (100 bootstraps). Supplementary Figure 3 displays the variable importance after 100 bootstraps, with the dotted line showing a slight leveling off in the importance measure and which corresponds to the top 10% features (n = 57). Random Forest with rfe on the training data set showed that the highest mean training performance was for a model with 25 features. However,

to choose a model with high accuracy while reducing the number of features as low as possible, a model with 10 molecules was chosen (AUC = 0.867) (Supplementary Figure 4). We tted the Random Forest model with the 10 selected variables on the whole training data set, which predicted the training data set with82.35% accuracy (sensitivity = 87.5%, specicity = 77.8%, AUC =0.826) and could classify our test data set with 79.2% accuracy,81.8% sensitivity, 76.9% specicity, a positive predictive value of75.0%, a negative predictive value of 83.3% and an AUC of 0.792. When we repeated the analysis including APOE, APOE had very

low variable importance (222 out of 573) and was not selected forward. In addition, APOE was not selected in the nal model after forcing it during rfe together with the top 10% of the selected analytes from the 100 Random Forest bootstraps.

Results for the Random Forest model are displayed in Figure 1. Nine of the 10 metabolites of the classier were associated with AD in single analyte regression analysis at qo0.05 (Supplementary Table 1 and Table 2) and none of them was shown to interact with the APOE 4 genotype at Po0.05.

ROC plot: Random Forest for the independent test dataset

1.0

0.8

True positive rate

0.6

7 identified metabolites

10 metabolites Training dataset: Accuracy=82.35%

Sensitivity=87.5% Specificity=77.8%

Auc=0.826

Test dataset: Accuracy=79.2%

Sensitivity=81.8% Specificity=76.9%

Auc=0.792

Traning dataset: Accuracy=82.35%

Sensitivity=82.65% Specificity=82.1%

Auc=0.824

Test dataset: Accuracy=66.67%

Sensitivity=69.23% Specificity=62.64%

Auc=0.664

0.4

0.2

0.0

0.0 0.2 0.4 0.6 0.8 1.0

False positive rate

Figure 1. Receiver operator curve (ROC) for the independent test data set for the selected 10 metabolites after recursive feature elimination for AD versus controls and for the seven identied metabolites, and summaries of the classier models for the training and test data sets. AD, Alzheimers disease; AUC, area under curve.

Table 1. Sample demographics

Comprehensive plasma LC-MS lipidomics (n = 124)

Control MCI AD ANOVA F(df1,df2) or x2(df) tests

N 40 48 36Age (mean, s.d.) 78.46 (6.7) 78.96 (5.6) 78.14 (7.7) F = 0.180 (2,121), P = 0.836 MMSE (mean, s.d.) 29.00 (1.1) 26.94 (1.9) 21.49 (4.8) F = 65.46(2,121), Po2.0E 16a

Female/male (N) 17/18 26/22 21/19 2 = 0.259 (2), P = 0.879 APOE 4 alleles (0/1/2) (N) 33/7/0 28/12/3 15/15/6 2 = 12.812 (2), Po16.5E 03b

Batch (1/2/3/4) (N) 9/10/9/12 14/11/10/13 9/8/10/9 2 = 1.256 (2), P = 0.974 Diabetesc 1 3 3 2 = 0.749 (2), P = 0.688 Smokingd 6 0 9 2 = 2.729 (2), P = 0.255 Statins 10 20 14 2 = 3.552 (2), P = 0.169 Samples (LON) 40 28 36 NA Samples (EUR) 0 20 0 NA

Abbreviations: AD, Alzheimers disease; ANOVA, analysis of variance; EUR, samples obtained from the non-London AddNeuroMed European centres; LC-MS, liquid chromatography-mass spectrometry; LDN, samples obtained from London AddNeuroMed- and DCR-based patients; MCI, mild cognitive impairment; MMSE, Mini-Mental State Examination score; NA, not available. aTukeys honest signicant differences post hoc tests: AD versus Control Po1.0E 17; AD versus

MCI P = 4.02E 13; MCI versus Control P = 3.36E 03. bAD versus Control P = 7.5E 04; AD versus MCI P = 0.1364; MCI versus Control P = 0.099. cDiabetes information was available for 34 AD patients, 29 controls and 43 MCI. dInformation on smoking was available for 34 AD patients, 29 controls and 6 MCI.

Translational Psychiatry (2015), 1 8

Plasma cholesterol metabolites in AD

P Proitsi et al

4

Table 2. List of metabolite molecules selected by the Random Forest classier (AD versus elderly health controls)

Metabolite molecule Peak height data normalized to whole mean only Peak ratios normalized to internal standard

OR 95% CI P-value OR L95 95% CI P-value

Mass/z 367 0.124 0.030.39 1.31E 03 0.115 0.030.36 8.09E 04 ChE 32:0 0.251 0.100.52 7.51E 04 0.237 0.100.48 4.19E 04 ChE 34:0 0.151 0.050.38 3.14E 04 0.152 0.050.37 2.90E 04 ChE 34:6 0.231 0.080.53 1.56E 03 0.126 0.030.35 5.40E 04 ChE 32:4 0.141 0.040.43 1.75E 03 0.056 0.010.24 6.56E 04 ChE 33:6 0.218 0.070.55 3.15E 03 0.205 0.060.50 2.21E 03 Mass/z 628 3.669 1.2013.93 3.49E 02 NA NA NA Mass/z 906 0.210 0.070.50 1.40E 03 0.226 0.090.48 3.73E 04 Mass/z 315 5.084 1.7816.88 4.12E 03 NA NA NA ChE 40:4 0.362 0.180.67 2.56E 03 0.279 0.120.55 6.43E 04 Cholesterol NA NA NA 0.316 0.025.01 4.21E 01

Abbreviations: AD, Alzheimers disease; ChE, cholesteryl ester; CI, condence interval; NA, not available; OR, odds ratio. Results are presented for peak height and peak ratio (measurement normalized to internal standard) after covariate adjustment. The metabolites are ordered according to their importance during recursive feature elimination. All metabolites except for mass 628 were associated with AD in single analyte regression analysis (qo0.05). Results for semiquantied mass 628 and mass 315 are not presented as they were not consistently above the limit of quantication. Results are also presented for the APOE 4 allele and for cholesterol peak ratio. The association of APOE 4 was OR = 5.620, 95% CI = 2.2716.29, P = 5.48E 04 per 4 allele.

m/z 369

O

R O

H

Fatty acid chain Cholesterol head group

Cholesterol

R1

O

O O

O

O

R2

R2

O

P

P

O

N+

O

O

O

HO

Phosphatidylcholine

Lecithin: cholesterol acyltransferase (LCAT)

OH

O

N+

O O

O

O

R1

O

Cholesteryl ester Lysophosphatidylcholine

Figure 2. Cholesteryl ester molecules structure (a) and synthesis from cholesterol catalysed by lecithin: cholesterol acyl transferase (LCAT) (b). ChE is characterized by the presence of cholesterol head group in ESI(+) of 369.

Identication and measurement of putative biomarkersSix molecules were identied using MS/MS fragmentation patterns (Figure 2a) and found to be cholesteryl esters (ChEs) with very long chain fatty acids,29,30 synthesized from cholesterol (Figure 2b). These were m/z 866 (ChE 32:0), m/z 894 (ChE 34:0), m/z 882 (ChE 34:6), m/z 856 (ChE 32:4), m/z 868 (ChE 33:6) and m/z 970 (ChE 40:4). In addition, one molecule was proposed as an oxidized form of desmosterol following the expected fragmentation pattern (m/z 367). The six identied ChE molecules and the desmosterol-related molecule predicted the training data set with82.35% accuracy (sensitivity = 82.61%, specicity = 82.14%, AUC =

0.824) and could classify the test data set with 67% accuracy,69.23% sensitivity, 62.64% specicity, a positive predictive value of69.23%, a negative predictive value of 63.64% and an AUC of 0.670.

Relative quantication was produced for eight of the molecules, which were consistently over the limit of quantication (LOQ). Univariate association analysis of the eight peak ratios revealed that the associations of all of them were strengthened. We, additionally, measured cholesterol, the precursor of ChE, nding no association with AD (Table 2).

Figure 3 plots the levels of the metabolites and cholesterol in the three diagnostic groups, which highlights the overall decrease pattern in metabolite levels from healthy controls to MCI to AD. To further investigate the association of the metabolites with cognition, we investigated their correlation with MMSE, which showed a weak association (Supplementary Figure 5).

When we tested our classier on individuals with MCI, 22 were classied as AD and 18 as controls.

Finally, logistic regression analyses including statin use and smoking status as covariates produced identical results (not presented).

DISCUSSIONWe performed lipidomics on plasma samples from 36 late-onset

AD patients, 40 healthy controls and 48 individuals with MCI to identify metabolites that differentiate between AD patients and healthy controls. Univariate analysis identied 41 analytes associated with AD (qo0.05). We then applied Random Forest and a backwards elimination approach to identify a reproducible metabolic signature to classify AD patients and we found a combination of 10 metabolite molecules that classied AD patients in the independent test data set with 79% accuracy. This is one of the largest non-targeted plasma studies to date to use such a systematic analysis pipeline, which included assessment of the model in an independent data set (1/3 of sample), to identify new targets linked to AD and provide improvement to current diagnostic classiers.1116 Although APOE was associated with AD in single analyte analyses, its variable importance in the presence of the metabolites during classication was low and was not included in the classier probably due to the metabolites already capturing the information that APOE provides. It should be noted that the single metabolite logistic regression results indicated that the metabolites provided information over and above APOE.

Translational Psychiatry (2015), 1 8

Plasma cholesterol metabolites in AD P Proitsi et al

5

Mass/z 367

(ANCOVA p=9.20E-04)

ChE 32:0 (ANCOVA p=4.80E-04)

ChE 34:6

(ANCOVA p=8.75E-04)

ChE 33:6 (ANCOVA p=1.37E-03)

Plasma (log2) ChE 33:6

3

Plasma (log2) Mass/z 367

Plasma (log2) ChE 32:0

2

-2

2

-2

2

-2

2

Plasma (log2) ChE 34:6

1

1

1

1

0

0

0

0

-1

-1

-1

-1

-2

-2

-2

CONTROL MCI AD

CONTROL MCI AD

CONTROL MCI AD

Diagnosis

Diagnosis

Diagnosis

CONTROL MCI AD

Diagnosis

CONTROL MCI AD

Diagnosis

ChE 32:4

(ANCOVA p=9.66E-04)

ChE 34:0

(ANCOVA p=7.56E-05)

ChE 40:4 (ANCOVA p=4.53-04)

Plasma (log2) ChE 40:4

Mass/z 906

(ANCOVA p=9.92-04)

2

1.5

2

2

Plasma (log2) ChE 32:4

Plasma (log2) ChE 34:0

Plasma (log2) Mass/z 906

1.0

1

1

1

0.5

0

0.0

0

0

-0.5

-1

-1

-1

-1.0

-1.5

CONTROL MCI AD

CONTROL MCI AD

CONTROL MCI AD

Diagnosis

Diagnosis

Diagnosis

Cholesterol (ANCOVA p=5.86E-01)

Plasma (log2) cholesterol

0.4

0.2

0.0

-0.2

-0.4

CONTROL MCI AD

Diagnosis

Figure 3. Boxplots depicting change in the level of the Random Forest measured molecules which were consistently above the level of quantication (LOQ) and that of cholesterol in the three diagnostic groups. All molecules were decreased in AD compared with controls after putative biomarker measurement. A decrease is also observed in MCI compared with controls. The ANCOVA P-value for the difference in metabolite levels in the three groups is displayed above each graph. AD, Alzheimers disease; ANCOVA, analysis of covariance; MCI, mild cognitive impairment.

Most of the metabolites in the classier were reduced in AD patients. We observed that their levels in MCI subjects were more variable but overall followed the control4MCI4AD trend with some of them being different at Po0.05 between AD and MCI or

MCI and controls. To investigate this further, we plotted the correlation of metabolites with MMSE, which is a more informative continuous surrogate marker for disease status and observed weak association between the metabolic signature and cognition.

We subsequently extracted single metabolite measures and observed that their associations with AD were strengthened (Table 2). Six molecules were identied as ChEs, molecules which have not been previously associated with AD and one as a potentially oxidized form of desmosterol, which could predict the test data set with 67% accuracy highlighting both their good predictive value and also the importance of the unidentied molecules in AD. We, additionally, measured cholesterol, as it is the precursor of ChE, and wanted to explore the chemical path differences with AD, nding no association. This suggested that the association with AD is specic to metabolites synthesized from cholesterol rather than to cholesterol itself.

A heatmap of the measured molecules together with cholesterol is shown in Figure 4. We have previously reported three PCs to be decreased in AD versus controls16 using overlapping samples. Although these PCs were not included in our classier, their raw values were decreased in our data set (Po0.05) with one

of them (mass 780 which is PC16:0/20:5) being associated with AD at qo0.05. These molecules were also recently identied as markers of phenoconversion to either amnestic MCI or AD.14 We

have therefore included these three PCs in the heatmap to compare their biological variation with the newly discovered and biochemically related ChEs, which highlights the positive correlation between the metabolites identied here, cholesterol and the three PCs; therefore, it is likely that the information the PCs provide is captured by the analytes selected in the rfe step.

Role of cholesterol and its derivativesCholesterol esters are largely synthesized in plasma by the transfer of fatty acids to cholesterol from PC, a reaction catalysed by the enzyme LCAT (lecithin: cholesterol acyl transferase). Free cholesterol can be taken up by APOE-containing lipoproteins, such as HDL, but is conned to the outer surface of the particle. The esterication of cholesterol to ChE ensures that more cholesterol is packaged into the interior of lipoproteins and this increases their capacity, allowing more efcient cholesterol transport through the bloodstream. LCAT has a preference for plasma 16:018:2 or 18:0 18:2 PCs, therefore, connecting the PCs identied by Whiley et al.16 and the ChEs identied here through a one-step enzymatic reaction (Figure 2b). Interestingly, LCAT is also expressed in the brain by astrocytes, and, together with APOE and ABCA1, has a key role in the maturation of glial-derived nascent lipoproteins.

Translational Psychiatry (2015), 1 8

Plasma cholesterol metabolites in AD

P Proitsi et al

6

Color key

-1 -0.5 0 0.5 1

Value

ChE 34:0

ChE 32:0

ChE 34:6

ChE 33:6

ChE 40:4

ChE 32:4

PC 16:0/22:6

PC 18:0/22:6

PC 16:0/20:5

Cholesterol

Mass/z 367

Mass/z 906

Mass/z 367

Cholesterol

PC 16:0/20:5

PC 18:0/22:6

PC 16:0/22:6

ChE 32:4

ChE 40:4

ChE 33:6

ChE 34:6

ChE 32:0

ChE 34:0

Mass/z 906

Figure 4. Heatmap of the eight measured metabolites selected during Random Forest classication, following recursive feature elimination, which were consistently above the level of quantication (LOQ). Cholesterol and the three phosphatidylcholines (PCs) previously published by Whiley et al.16 are also included. AD, Alzheimers disease; ANCOVA, analysis of covariance; MCI, mild cognitive impairment.

In other tissues, cholesterol is esteried by acyl-coenzyme A: cholesterol acyltransferase 1 and 2 (ACAT1 and ACAT2). ACAT uses acyl-CoA as a source for the acyl chains. Recent evidence suggests a strong link of ACAT1 with amyloid deposition. Pharmacological inhibition of ACAT in AD mice resulted in diminished amyloid plaque burden in their brains and improved cognitive function25,31 and genetic ablation of Acat1 in AD mice diminished the levels of A42, decreased the amyloid plaque burden and full-length human APP (human amyloid precursor protein), and improved the cognitive function of the mice.32 Finally, a recent gene therapy study showed that adeno-associated virus targeting Acat1 for gene knockdown delivered to the brains of AD mice decreased the levels of total brain amyloid-, oligomeric amyloid- and full-length human APP to levels similar to complete genetic knockdown of Acat1.33

Approximately one-third of the ChE is transferred from HDL to APOB-containing lipoproteins, such as VLDLs, in exchange for triglycerides by ChE transfer protein. Thus most of the cholesterol in circulating lipoproteins is ChE produced from HDL lipoproteins by the LCAT-catalysed reaction. This process results in lower HDL cholesterol and indirectly decreases HDL size as frequently observed in type 2 diabetes34 and a deciency of ChE transfer protein is associated with increased HDL and decreased LDL levels, a prole typically antiatherogenic.35,36 Interestingly, an elevation of cholesterol esters (chains 14:20) was observed in the brains of AD patients and of three transgenic familial AD mouse models.24

Lower desmosterol levels have been previously found in the plasma and brains of AD patients compared with healthy controls.13,26 Desmosterol is a precursor of cholesterol and seladin (DHCR24), which governs the metabolism of desmosterol to cholesterol in specic brain areas has been shown to counteract the -secretase cleavage of APP and the formation of amyloid-.37,38 An unknown molecule with similar structure and the same mass as desmosterol shows the highest correlation with cholesterol here.

Strengths and limitations

To our knowledge, ours is the rst study to implicate ChEs in AD.

We should highlight, however, that these molecules are biochemically related to metabolites previously associated with AD initiation and progression and are, therefore, biologically relevant. Although lipid molecules have been previously implicated in AD, results are not always consistent. It is worth noting that even if previously published studies go under the umbrella of plasma lipidomics and/or metabolomics, differences are considerable in terms of patient cohort selection (such as differences in patient groups selection, in age and in dementia severity), in terms of technical treatment (such as sampling/storage conditions and fasting state before sample collection) and in clinical data availability. As an example, sample groups can differ from a simple case versus control design11 to including MCI1216 and in one case including time points with pre- and post-conversion patients.14 Sample groups will inuence the selectivity of results, as well as patient numbers, which ranged from o50 in early studies11 to the low hundreds in more recent studies.1316 In addition, LC-MS ngerprinting methods can be qualitative or/and quantitative, hence the information can be relative amounts11,12,15

or a combination of relative and absolute amounts for chosen metabolites.13,14,16 Another factor for differences observed

between studies would be the different data processing tools applied for data mining and statistical methods used to produce AD biomarker predictive models. Some studies showed predictive models of AD based on one to three molecules13,16 or a battery

ranging from three to 10 markers and have used a wide range of statistical analysis methods ranging from simple single metabolite approaches11 to methods using multivariate regression approaches and resampling techniques,15 and validation of the panels in independent data sets.14 Here, we have used a well-characterized AD cohort matched for age and gender and performed a careful and systematic analysis pipeline to extract metabolites associated with AD, by performing bootstrapping to avoid over-tting and validating our results in an unseen data set. After the molecules associated with AD were identied, an additional method of metabolite quantication in all the samples was performed, which was possible due to internal standards and a run separating lipids, which acquired data for 2 h for each sample.

Our study also suffers from potential limitations. Although ours is one of the largest AD metabolomics studies to date, the sample size is still modest and replication is required in larger cohorts. In addition, this study suffers from limitations inherent to AD case control studies, such as the possibility that some of the healthy elderly controls may already carry pathology, and that some of the clinically diagnosed AD may be pathologically non-AD dementias. Through follow-up data on the individuals used in this study, however, we know that all AD patients used for our analysis maintained the diagnosis of AD as did all controls for at least 3 years from their baseline visit. We also believe that achieving such high performance in both training and test data sets for an AD casecontrol data set highlights the efcacy of the classier, as AD diagnostic classiers rarely achieve such high performance; we, therefore, believe that having additional pathology information would only increase its performance. Furthermore, we need to acknowledge that individuals with MCI are more heterogeneous in pathology; however, the MCI subjects used in this study were recruited primarily from local memory clinics and were, therefore, expected to be composed largely of subjects with likely an AD end point.

Finally, due to the large number of comorbidities in old age, we have to acknowledge that our metabolite signal may not be AD specic but it could it be associated with overall ill health and other comorbid conditions, making it potentially a not good biomarker for recruitment in clinical trials. Investigating and

Translational Psychiatry (2015), 1 8

Plasma cholesterol metabolites in AD P Proitsi et al

7

comparing the metabolic proles of other disorders would increase the specicity of our panel.

CONCLUSIONSIn this study, we used a Random Forest approach and identied a combination of 10 metabolites, which predicted AD with near 80% accuracy. We subsequently identied six of the metabolites to be ChEs, molecules not previously implicated in AD, which are connected to PCs through a one-step enzymatic reaction. The newly identied molecules were reduced in AD patients compared with controls. All these, combined with the lack of association between cholesterol and AD, suggest that it is the dysregulation of specic steps in cholesterol metabolism, rather than cholesterol itself, that is responsible for these observations and suggest novel targets for future work. These ndings need to be replicated in larger well-phenotyped cohorts, which will be possible in the near future through the large biomarker consortia being set up. In addition, information on pathology status, such as amyloid marker cerebrospinal uid, PET or brain atrophy measurements, through magnetic resonance imaging (MRI), would provide more precise phenotypes for biomarker discovery and would capture different stages of disease pathology, and, the comparison of metabolite levels between MCI patients who converted to AD and those who remained stable would provide us with metabolites, which are associated with disease initiation. Finally, integrating additional types of biological modalities, such as protein, gene expression and genotype information, will help investigate the origin of ChE dysregulation in AD.

CONFLICT OF INTEREST

The authors declare no conict of interest.

ACKNOWLEDGMENTS

We thank the individuals and families who took part in this research. We acknowledge the use of the computational Linux cluster and the Biomedical Research Centre Nucleus Informatics Team supported by National Institute for Health Research (NIHR) Mental Health Biomedical Research Centre and Dementia Unit at South London and Maudsley NHS Foundation Trust and (Institute of Psychiatry, Psychology and Neuroscience) Kings College London. This work was also supported by the 7th Framework Programme of the European Union (ADAMS project, HEALTH-F4-2009-242257). AddNeuroMed was funded through the EU FP6 programme. PP is an Alzheimers Society post-doctoral fellow. The funding bodies had no role in the design, collection, analysis, interpretation of the data, writing of the manuscript and the decision to submit the manuscript.

REFERENCES

1 Ferri CP, Prince M, Brayne C, Brodaty H, Fratiglioni L, Ganguli M et al. Global prevalence of dementia: a Delphi consensus study. Lancet 2005; 366: 21122117.

2 Whiley L, Legido-Quigley C. Current strategies in the discovery of small-molecule biomarkers for Alzheimer's disease. Bioanalysis 2011; 3: 11211142.

3 Samson K. MEDICARE panel: Too little evidence to show amyloid PET scans improve Alzheimer outcomes. Neurol Today 2013; 13: 3642.

4 Whiley L, Godzien J, Ruperez FJ, Legido-Quigley C, Barbas C. In-vial dual extraction for direct LC-MS analysis of plasma for comprehensive and highly reproducible metabolic ngerprinting. Anal Chem 2012; 84: 59925999.

5 Lindon JC, Holmes E, Nicholson JK. Metabonomics in pharmaceutical R&D. FEBS J 2007; 274: 11401151.

6 Suhre K, Shin SY, Petersen AK, Mohney RP, Meredith D, Wagele B et al. Human metabolic individuality in biomedical and pharmaceutical research. Nature 2011; 477: 5460.

7 Suhre K, Meisinger C, Doring A, Altmaier E, Belcredi P, Gieger C et al. Metabolic footprint of diabetes: a multiplatform metabolomics study in an epidemiological setting. PLoS One 2010; 5: e13953.

8 Barba I, Fernandez-Montesinos R, Garcia-Dorado D, Pozo D. Alzheimer's disease beyond the genomic era: nuclear magnetic resonance (NMR) spectroscopy-based metabolomics. J Cell Mol Med 2008; 12: 14771485.

9 Lunnon K, Ibrahim Z, Proitsi P, Lourdusamy A, Newhouse S, Sattlecker M et al. Mitochondrial dysfunction and immune activation are detectable in early Alzheimer's disease blood. J Alzheimers Dis 2012; 30: 685710.

10 Zipser BD, Johanson CE, Gonzalez L, Berzin TM, Tavares R, Hulette CM et al. Microvascular injury and blood-brain barrier leakage in Alzheimer's disease. Neurobiol Aging 2007; 28: 977986.

11 Han X, Rozen S, Boyle SH, Hellegers C, Cheng H, Burke JR et al. Metabolomics in early Alzheimer's disease: identication of altered plasma sphingolipidome using shotgun lipidomics. PLoS One 2011; 6: e21643.

12 Greenberg N, Grassano A, Thambisetty M, Lovestone S, Legido-Quigley C. A proposed metabolic strategy for monitoring disease progression in Alzheimer's disease. Electrophoresis 2009; 30: 12351239.

13 Sato Y, Suzuki I, Nakamura T, Bernier F, Aoshima K, Oda Y. Identication of a new plasma biomarker of Alzheimer's disease using metabolomics technology. J Lipid Res 2012; 53: 567576.

14 Mapstone M, Cheema AK, Fiandaca MS, Zhong X, Mhyre TR, Macarthur LH et al. Plasma phospholipids identify antecedent memory impairment in older adults. Nat Med 20: 4154182014.

15 Oresic M, Hyotylainen T, Herukka SK, Sysi-Aho M, Mattila I, Seppanan-Laakso T et al. Metabolome in progression to Alzheimer's disease. Transl Psychiatry 2011; 1: e57.

16 Whiley L, Sen A, Heaton J, Proitsi P, Garcia-Gomez D, Leung R et al. Evidence of altered phosphatidylcholine metabolism in Alzheimer's disease. Neurobiol Aging 2014; 35: 271278.

17 Reitz C. Dyslipidemia and the risk of Alzheimer's disease. Curr Atheroscler Rep

2013; 15: 307.

18 Reitz C, Mayeux R. Alzheimer disease: epidemiology, diagnostic criteria, risk factors and biomarkers. Biochem Pharmacol 2014; 88: 640651.

19 Harold D, Abraham R, Hollingworth P, Sims R, Gerrish A, Hamshere ML et al. Genome-wide association study identies variants at CLU and PICALM associated with Alzheimer's disease. Nat Genet 2009; 41: 10881093.

20 Hollingworth P, Harold D, Sims R, Gerrish A, Lambert JC, Carrasquillo MM et al. Common variants at ABCA7, MS4A6A/MS4A4E, EPHA1, CD33 and CD2AP are associated with Alzheimer's disease. Nat Genet 2011; 43: 429435.

21 Lambert JC, Heath S, Even G, Campion D, Sleegers K, Hiltunen M et al. Genome-wide association study identies variants at CLU and CR1 associated with Alzheimer's disease. Nat Genet 2009; 41: 10941099.

22 Lambert JC, Ibrahim-Verbaas CA, Harold D, Naj AC, Sims R, Bellenguez C et al. Meta-analysis of 74,046 individuals identies 11 new susceptibility loci for Alzheimer's disease. Nat Genet 2013; 45: 14521458.

23 Astarita G, Jung KM, Vasilevko V, Dipatrizio NV, Martin SK, Cribbs DH et al. Elevated stearoyl-CoA desaturase in brains of patients with Alzheimer's disease. PLoS One 2011; 6: e24777.

24 Chan RB, Oliveira TG, Cortes EP, Honig LS, Duff KE, Small SA et al. Comparative lipidomic analysis of mouse and human brain with Alzheimer disease. J Biol Chem 2012; 287: 26782688.

25 Hejazi L, Wong JW, Cheng D, Proschogo N, Ebrahimi D, Garner B et al. Mass and relative elution time proling: two-dimensional analysis of sphingolipids in Alzheimer's disease brains. Biochem J 2011; 438: 165175.

26 Wisniewski T, Newman K, Javitt NB. Alzheimer's disease: brain desmosterol levels. J Alzheimers Dis 2013; 33: 881888.

27 Lovestone S, Francis P, Strandgaard K. Biomarkers for disease modication trials--the innovative medicines initiative and AddNeuroMed. J Nutr Health Aging 2007; 11: 359361.

28 Johnson WE, Li C, Rabinovic A. Adjusting batch effects in microarray expression data using empirical Bayes methods. Biostatistics 2007; 8: 118127.

29 Butovich IA. Fatty acid composition of cholesteryl esters of human meibomian gland secretions. Steroids 2010; 75: 726733.

30 Butovich IA, Uchiyama E, McCulley JP. Lipids of human meibum: mass-spectrometric analysis and structural elucidation. J Lipid Res 2007; 48: 22202235.

31 Hutter-Paier B, Huttunen HJ, Puglielli L, Eckman CB, Kim DY, Hofmeister A et al. The ACAT inhibitor CP-113,818 markedly reduces amyloid pathology in a mouse model of Alzheimer's disease. Neuron 2004; 44: 227238.

32 Bryleva EY, Rogers MA, Chang CC, Buen F, Harris BT, Rousselet E et al. ACAT1 gene ablation increases 24(S)-hydroxycholesterol content in the brain and ameliorates amyloid pathology in mice with AD. Proc Natl Acad Sci USA 2010; 107: 30813086.

33 Murphy SR, Chang CC, Dogbevia G, Bryleva EY, Bowen Z, Hasan MT et al. Acat1 knockdown gene therapy decreases amyloid-beta in a mouse model of Alzheimer's disease. Mol Ther 2013; 21: 14971506.

34 de Vries R, Borggreve SE, Dullaart RP. Role of lipases, lecithin:cholesterol acyltransferase and cholesteryl ester transfer protein in abnormal high density lipo-protein metabolism in insulin resistance and type 2 diabetes mellitus. Clin Lab 2003; 49: 601613.

Translational Psychiatry (2015), 1 8

Plasma cholesterol metabolites in AD P Proitsi et al

8

35 Barter P, Rye KA. Cholesteryl ester transfer protein inhibition to reduce cardiovascular risk: where are we now? Trends Pharmacol Sci 2011; 32: 694699.

36 Barter PJ, Brewer HB Jr, Chapman MJ, Hennekens CH, Rader DJ, Tall AR. Cholesteryl ester transfer protein: a novel target for raising HDL and inhibiting atherosclerosis. Arterioscler Thromb Vasc Biol 2003; 23: 160167.

37 Greeve I, Kretzschmar D, Tschape JA, Beyn A, Brellinger C, Schweizer M et al. Age-dependent neurodegeneration and Alzheimer-amyloid plaque formation in transgenic Drosophila. J Neurosci 2004; 24: 38993906.

38 Crameri A, Biondi E, Kuehnle K, Lutjohann D, Thelen KM, Perga S et al. The role of seladin-1/DHCR24 in cholesterol biosynthesis, APP processing and Abeta generation in vivo. EMBO J 2006; 25: 432443.

This work is licensed under a Creative Commons Attribution 4.0 International License. The images or other third party material in this article are included in the articles Creative Commons license, unless indicated otherwise in the credit line; if the material is not included under the Creative Commons license, users will need to obtain permission from the license holder to reproduce the material. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/

Web End =http://creativecommons.org/licenses/ http://creativecommons.org/licenses/by/4.0/

Web End =by/4.0/

Supplementary Information accompanies the paper on the Translational Psychiatry website (http://www.nature.com/tp)

Translational Psychiatry (2015), 1 8