Association of blood lipids with Alzheimer s disease: A comprehensive lipidomics analysis

Transcription

1 Fetured Article Assocition of blood lipids with Alzheimer s disese: A comprehensive lipidomics nlysis Q6 P. Proitsi,,1, M. Kim b,1, L. Whiley b, A. Simmons,c, M. Sttlecker,c, L. Velyudhn,d, M. K. Lupton e, H. Soininen f, I. Kloszewsk g, P. Mecocci h, M. Tsolki i, B. Vells j, S. Lovestone k, J. F. Powell,2, R. J. B. Dobson,c,2, C. Legido-Quigley b,2 Q1 Q2 King s College London, Institute of Psychitry, Psychology & Neuroscience, London, UK b King s College London, Institute of Phrmceuticl Science, London, UK c NIHR Biomedicl Reserch Centre for Mentl Helth nd Biomedicl Reserch Unit for Dementi t South London nd Mudsley NHS Foundtion Trust d Oxles NHS Foundtion Trust e QIMR Berghofer Medicl Reserch Institute, Brisbne, Austrli f Deprtment of Neurology, Kuopio University Hospitl nd University of Estern Finlnd, Kuopio, Finlnd g Deprtment of Old Age Psychitry & Psychotic Disorders, Medicl University of Lodz, Lodz, Polnd h Section of Gerontology nd Geritrics, Deprtment of Medicine, University of Perugi, Perugi, Itly i Memory nd Dementi Centre, Aristotle University of Thessloniki, Thessloniki, Greece j Deprtment of Internl nd Geritrics Medicine, INSERM U 1027, Gerontopole, H^opitux de Toulouse, Toulouse, Frnce k Deprtment of Psychitry, University of Oxford, Wrneford Hospitl, Oxford, UK Abstrct Keywords: 1. Bckground Introduction: The im of this study ws to (1) replicte previous ssocitions between six blood lipids nd Alzheimer s disese (AD) (Proitsi et l 2015) nd (2) identify novel ssocitions between lipids, clinicl AD dignosis, disese progression nd brin trophy (left/right hippocmpus/entorhinl cortex). Methods: We performed untrgeted lipidomic nlysis on 148 AD nd 152 elderly control plsm smples nd used univrite nd multivrite nlysis methods. Results: We replicted our previous lipids ssocitions nd reported novel ssocitions between lipids molecules nd ll phenotypes. A combintion of 24 molecules clssified AD ptients with.70% ccurcy in test nd vlidtion dt set, nd we identified lipid signtures tht predicted disese progression (R , test dt set) nd brin trophy (R , ll test dt sets except left entorhinl cortex). We puttively identified number of metbolic fetures including cholesteryl esters/triglycerides nd phosphtidylcholines. Discussion: Blood lipids re promising AD biomrkers tht my led to new tretment strtegies. Ó 2016 The Authors. Published by Elsevier Inc. on behlf of the Alzheimer s Assocition. This is n open ccess rticle under the CC BY-NC-ND license ( Alzheimer s disese; Dementi; Brin trophy; smri; Rte of cognitive decline; Lipidomics; Metbolomics; Biomrkers; Mchine lerning; Multivrite; Clssifiction; Rndom forest Alzheimer s disese (AD) is devstting illness nd one of the mjor public helth chllenges of the 21st century. The lck of effective tretments nd erly dignosis highlights the importnce of the identifiction of 1 These uthors contributed eqully to the mnuscript. 2 These senior uthors contributed eqully to the mnuscript. Q3 Corresponding uthor. Tel.: ---; Fx: ---. E-mil ddress: petroul.proitsi@kcl.c.uk Alzheimer s & Dementi - (2016) 1-12 noninvsive biomrkers, for erly dignosis nd disese progression. Blood metbolites hve recently emerged s promising AD biomrkers [1 4]. They re smll molecules which could theoreticlly cross the lredy compromised AD blood-brin brrier [5]; they re esily ccessible, nd they represent n essentil spect of the phenotype of n orgnism nd moleculr fingerprint of disese progression [6,7]. They cn therefore id erly dignosis, recruitment into trils nd my help identify new therpeutic trgets / Ó 2016 The Authors. Published by Elsevier Inc. on behlf of the Alzheimer s Assocition. This is n open ccess rticle under the CC BY-NC-ND license (

2 2 P. Proitsi et l. / Alzheimer s & Dementi - (2016) A number of blood metbolomic studies hve highlighted the role of lipid compounds, such s phosphtidylcholines (PCs) in AD [1 4]. We previously identified three PCs tht were diminished in mild cognitive impirment (MCI) individuls nd AD ptients [4] nd were further ssocited with poorer memory performnce nd decresed brin function during ging [8]. We further performed lipidomics nlysis nd identified 10 metbolites tht predicted AD in n unseen test dt set with 79% ccurcy [9]; six nlytes were puttively identified s cholesteryl esters (ChoEs), molecules relted to PCs, nd were reduced in MCI nd AD. Here, we performed lipidomics nlysis in smple of 142 AD ptients nd 135 helthy controls iming to (1) replicte our previous ssocitions [9] nd (2) discover new lipids nd combintions of lipids ssocited with clinicl AD dignosis nd AD endophenotypes, such s the rte of cognitive decline nd brin trophy mesures. This is to our knowledge the most comprehensive blood lipidomics study to dte to identify lipid signtures ssocited with AD nd AD endophenotypes, improving our current knowledge of molecules ssocited with AD. 2. Methods 2.1. Ptient smple collection This study used 148 AD ptients nd 152 controls from the Dementi Cse Register t King s College London nd the EU-funded AddNeuroMed study [10]. All individuls with AD ptients met criteri for either probble (NINCDS- ADRDA, DSM-IV) or definite (CERAD) AD. All nonpopultion individuls who were controls were screened for dementi using the MMSE or ADAS-cog or were determined to be free from dementi t neuropthologic exmintion or hd Brk score 2.5. Dignosis ws confirmed by pthologic exmintion for proportion of cses nd cognitively norml elderly controls. All AD cses hd n ge of onset 60 yers, nd controls were 60 yers t exmintion. A totl of 102 AD cses nd 104 controls hd HDL-c, LDL-c, TC, nd TG serum levels (mmol/l) vilble. Nonoverlpping individuls from these cohorts hve been previously reported [9]. Ech individul ws required to fst for 2 hours before smple collection, nd 10 ml of blood ws collected in tubes coted with sodium ethylenediminetetrcetic cid to prevent clotting. Whole blood ws centrifuged t 2000 g for 10 minutes t 4 C to seprte plsm, which ws removed nd stored t 280 C. All smples were centrifuged within pproximtely 2 hours of collection Lipidomics Smple tretment hs been described elsewhere [4,9,11] nd is explined in detil in Supplementry Methods 1. Briefly, 20 ml of plsm ws dded to glss HPLC vil contining 400-ml glss insert (Chromcol, UK). Ten microliters of high purity wter nd 40 ml of MS grde methnol were dded to ech smple, followed by 2-minute vortex mix to precipitte proteins; 200 ml of Methyl tert- Butyl Ether (MTBE) contining 10 mg/ml of internl stndrd Tripentdecnoin (TG45:0) ws dded, nd the smples were mixed vi vortex t room temperture for 1 hour. After ddition of 50 ml of high purity wter, finl smple mixing ws performed before centrifugtion t 3000 g for 10 minutes. The upper, lipid-contining, MTBE phse ws then injected onto the LC-MS system directly from the vil by djustment of the instrument needle height (17.5 mm from bottom). Lipidomics ws performed by Wters ACQUITY UPLC nd XEVO QTOF system. The method hs previously been published [4,12] nd hs been shown to quntitte.4500 metbolite species (Supplementry Methods 1). Smples were nlyzed in rndomized order, in four btches, with pooled plsm smpled (QC) t regulr intervls throughout the run (n 5 30 for both positive nd negtive ioniztion). Fetures were extrcted from netcdf files using the R pckge XCMS [13] which performed filtrtion, pek identifiction, mtching of peks cross smples, nd retention time correction. Positive nd negtive ioniztion mode dt were extrcted seprtely nd quntile normlized Structurl mgnetic resonnce imging Volumes of whole brin nd the hippocmpi nd entorhinl cortices were obtined using FreeSurfer from 123 subjects (53 AD ptients nd 70 Controls) who hd undergone smri. Regions were normlized by intrcrnil volume [14]. The volumetric dt were not used to id in the clinicl dignosis of AD. Detiled informtion regrding dt cquisition, pre-processing, nd qulity control ssessment hs been described elsewhere [15,16]. Before nlyses, smri mesures were stndrdized to hve men of 0 nd stndrd devition (SD) of Clcultion of rte of cognitive decline The ROD ws vilble for 118 AD ptients with nlyte dt nd hs been described elsewhere [17]. The ROD ws bsed on longitudinl mini mentl stte exmintion (MMSE) ssessments [18], nd only smples with t lest three MMSE mesures were included in the clcultion using liner mixed effect models. After covrite djustment [17], the slope coefficient for ech smple ws used s the ROD defined s the chnge in MMSE per dy Sttisticl nlysis Qulity control Dt QC hs been previously described [9] nd included filtering of fetures nd individuls, dt trnsformtion, btch effect correction, outlier detection, nd imputtion (Supplementry Methods 2.1 nd Supplementry Figure 1). All nlyses took plce in R

3 P. Proitsi et l. / Alzheimer s & Dementi - (2016) Q Single-nlyte sttisticl nlysis Logistic regression investigted the ssocition of ech metbolite with clinicl AD dignosis nd liner regression the ssocition with cognitive decline nd smri mesures. Logistic regression nd liner regression models for the smri mesures were djusted for ge t smpling, gender, presence of the polipoprotein E (APOE) ε4 llele, btch, nd study site. For the ROD models, covrite djustment ws only pplied for btch s the rest of the covrites were included in the ROD clcultion [17]. smri mesurements were not djusted for dignosis to llow identifiction of fetures ssocited with brin trophy cused by AD. Flse discovery rte (FDR) correction (0.05) ws pplied to correct for multiple testing ( fdrtool ). Secondry models investigted whether ny ssocitions were modified by the APOE 4 llele or by gender. Logistic regression results (summry sttistics) for the positive ioniztion metbolites were combined with the results (summry sttistics) from the Proitsi et l dt set [9] using inverse vrince weighted fixed effect met-nlysis ( metfor ). The published dt set [9] ws restricted to 576 fetures extrcted using Mss-Lynx, nd therefore, the nlysis presented here includes lrge number of previously unreported molecules extrcted using XCMS. All ssocitions re reported s the chnge per one metbolite stndrd devition (SD) Multivrite sttisticl nlysis A rndom forest (RF) clssifier pproch (using rf nd rfe in CARET ) ws used to develop clinicl dignosis clssifier s previously described [9] (Supplementry Methods 2.2). Briefly, AD cses nd controls were divided Tble 1 Smple demogrphics into trining dt set (2/3 of the smple) mtched for ge, gender, nd site nd n independent dt set (rest 1/3 of the smple). An RF model ws built on the trining dt set (100 bootstrps), nd in ech itertion, ech vrible ws ssigned vrible importnce (VI) score. The summed VI rnks provided n indiction of the predictive power for ech vrible, nd the top 10% molecules were selected for RF with recursive feture elimintion (rfe; 100 bootstrps) from 250 down to two fetures. For ech subset of predictors, the men bootstrp testing performnce ws clculted, nd the optiml number of vribles ws identified using sizetolernce tht picks subset of vribles tht is smll without scrificing too much performnce. Subsets of vribles within 2.5% nd 5% of the optimum performnce were exmined nd used to build finl models in the complete trining dt, which were tested on the test set. The finl model ws lso tested in the Proitsi et l dt set [9] which ws used s vlidtion dt set, fter excluding metbolites in the negtive ioniztion mode. The re under curve (AUC) ws used to test the performnce of ech clssifier. Receiver opertor curves (ROCs) were plotted using ROCR. Models including APOE ε4 nd the six fetures in Proitsi et l [9] were lso tested. Rndom forest regression (RFR) models were built for cognitive decline nd smri mesures following the sme strtegy s for clinicl dignosis. The dt set ws split rndomly into trining (2/3 of the dt) nd test set (1/3 of the dt) for ech endophenotype such tht the trining nd test dt sets were strtified for ech endophenotype nd contined equl representtion of ech site. Age, gender, nd APOE ε4 presence were included in the model development for the smri models, nd the root men squred error (RMSE) ws used to evlute the performnce of the models. AD (N 5 142) Controls (N 5 135) Difference between AD ptients nd controls Age, men (SD) 77 (6.5) 74 (5.9) t (270), P vlue ! Gender (mles/femles) 48/87 47/95 c (1), P vlue APOE ε4 llele (bsence/presence) 54/81 99/43 c (1), P vlue ! MMSE, men (SD; Rnge) 20.1 (4.6; 10 27) 29.2 (0.9; 27 30) t (143), P vlue, 2.0! ROD (per yer), men (SD) y (1.26) NA NA Entorhinl cortex right, men (SD) zx (0.0003) (0.0003) t (99), P vlue ! Entorhinl cortex left, men (SD) zx (0.0002) (0.0004) t (86), P vlue ! Hippocmpus right, men (SD) zx (0.0004) (0.0003) t (100), P vlue 5 1.3! Hippocmpus left, men (SD) zx (0.0004) (0.003) t (104), P vlue, 2.0! Men HDL-c (SD), mmol/l k 1.58 (0.37) 1.55 (0.38) b (SE ), P vlue Men LDL-c (SD), mmol/l k 3.42 (1.01) 3.07 (0.82) b (SE ), P vlue Men TC (SD), mmol/l k 5.69 (1.17) 5.29 (1.01) b (SE ), P vlue Men TG (SD), mmol/l k 1.64 (1.04) 1.52 (0.67) b (SE ), P vlue Sttins (yes/no) 38/97 34/108 c (1), P vlue Abbrevitions: AD, Alzheimer s disese; MMSE, mini-mentl stte exmintion score; ROD, rte of cognitive decline; SD, stndrd devition. Differences in the mens/frequencies of clinicl/demogrphic vribles were tested using t test t(df), x2(df) test, or liner regression nlyses fter djusting for ge, gender, the APOE ε4 llele, nd study site. y Rte of decline dt ws vilble for subset of AD ptients (N 5 118). z smri dt were vilble for subset of study prticipnts (N [AD 5 53, controls 5 70]). x Normlized to intrcrnil volume. k Serum HDL cholesterol, LDL cholesterol, totl cholesterol, nd triglyceride levels were vilble for subset of study prticipnts (N [AD 5 102, Controls 5 106])

4 4 P. Proitsi et l. / Alzheimer s & Dementi - (2016) web 4C=FPO Fig. 1. Associtions of previously reported molecules (Proitsi et l 2015) with clinicl AD dignosis in the current dt set nd ssocitions of puttively nnotted molecules, selected through rndom forest nlyses, with the respective phenotype. (A) Assocition of Mss 856 with clinicl AD dignosis; (B) Assocition of Mss 866 with clinicl AD dignosis; (C) Assocition of Mss 868 with clinicl AD dignosis; (D) Assocition of Mss 882 with clinicl AD dignosis; (E) Assocition of Mss 894 with clinicl AD dignosis; (F) Assocition of Mss 970 with clinicl AD dignosis; (G) Assocition of Mss 882 (2) (PC 40:4) with clinicl AD dignosis; (H) Assocition of Mss 948 (1) TG (57:1) with clinicl AD dignosis; (I) Assocition of Mss 919 (1) TG 50:2 with Hippocmpus (Right); (J) Assocition of Mss 943 (1) (ChoE/TG) with Hippocmpus Left; (K) Assocition of Mss 367 (sterol) with Entorhinl Cortex (Right); (L) Assocition of Mss 816 (1) with Entorhinl Cortex (Left); (M) Assocition of Mss 771 (1) PC 36:3 with the rte of cognitive decline (ROD). The P vlues displyed re for the univrite regressions fter djusting for covrites. All molecules re scled to hve men of 0 nd stndrd devition of

5 P. Proitsi et l. / Alzheimer s & Dementi - (2016) web 4C=FPO 3. Results A totl of 2539 positive ioniztion nd 358 negtive ioniztion fetures were initilly extrcted from 300 individuls. Fig. 1. (Continued) After QC, 2216 positive nd 289 negtive ioniztion fetures from 277 individuls (142 AD cses nd 135 controls) were used in subsequent nlyses. Of these, 53 AD ptients nd 70 controls hd smri dt vilble, nd 118 AD ptients hd

6 6 P. Proitsi et l. / Alzheimer s & Dementi - (2016) ROD dt vilble. Smple demogrphics re displyed in Tble Univrite nlyses results Logistic regression nlyses were initilly used to investigte the ssocition of ech lipid with AD. We then performed fixed-effects met-nlyses between the results of this dt set nd our previously published dt set [9],using fixed-effects met-nlyses. Briefly, 425 fetures were ssocited with AD t P vlue,.05inthisdtset;of these, 87 fetures pssed correction for multiple testing t Q vlue,0.05. After met-nlysis, 377 fetures were ssocited with AD t P vlue,.05 nd 125 t Q vlue,0.05. All six fetures from Proitsi et l [9] were ssocited with AD t Q vlue,0.05 in met-nlysis (Fig. 1 (A F) nd Tble 2). Liner regression investigted the ssocition of ech lipid with brin trophy nd the rte of cognitive decline. A totl of 266 fetures were ssocited with the ROD t P vlue,.05, but none pssed multiple testing correction. A totl of 181 fetures were ssocited with right hippocmpus volume nd 224 were ssocited with left hippocmpus volume; only six fetures were ssocited with left hippocmpus t Q vlue,0.05. Finlly, 156 nd 124 fetures were ssocited with EC volume (left nd right, respectively) t P vlue,.05, but no ssocitions pssed correction for multiple testing. Results for ll logistic nd liner regression nlyses re provided in Supplementry Tble 1. Overll, most lipids were reduced in AD compred to controls (54 of the 87 fetures ssocited t Q-vlue,0.05 were reduced in AD). Additionlly, we observed substntil overlp between fetures ssocited with clinicl AD dignosis nd brin trophy (Supplementry Figure 2). We further investigted whether the APOE ε4 nd gender modified the ssocitions between lipids nd clinicl AD dignosis. APOE ε4 modified the ssocition of 231 fetures with AD, nd gender modified the ssocition of 191 fetures with AD (P vlue,.05); none of these ssocitions ws significnt t Q-vlue, There were only 3 individuls with the ε2/ε4 genotype, nd there were therefore no differences in lipids levels between ε4 nd non-ε4 crriers fter excluding ε2/ε4 individuls Multivrite nlysis results A RF pproch ws used to identify pnel of molecules ssocited with clinicl AD dignosis. After n initil RF pre-selection step on the trining dt set, the top 10% lipids (250 fetures), in terms of their vrible importnce, were selected (fter 100 Bootstrps). Furthermore, rndom forest with recursive feture elimintion (RF-rfe) on the trining dt set showed tht the best trining performnce ws for model with 240 fetures. To choose model with high ccurcy while reducing the number of fetures s low s possible, 5% tolernce RF-rfe model (25 fetures) ws fitted on the whole trining dt set (AUC, 0.87) nd clssified the test dt set with 73% ccurcy (Supplementry Figure 3 nd Tble 3). The model ws then fitted on the trining dt set, excluding one negtive ioniztion mode nlyte nd clssified the test trining dt set with 74% ccurcyndtheproitsietl[9] vlidtion dt set with 71% ccurcy. There ws no increse in ccurcy when covrites nd the fetures from Proitsi et l [9] were dded to the models (Tble 3). Rndom forest regressions using the sme pipeline were pplied to the ROD nd brin trophy mesures. After RFR-rfe on the trining dt set, the lowest men RMSE for ROD ws for model with 40 fetures. The 5% tolernce model of the lowest RMSE model (10 fetures) ws fitted to the whole trining dt set (R ) nd predicted the test dt set with R (Tble 4). For right hippocmpus, the lowest RMSE ws with model with 70 fetures tht included ge. A 5% tolernce model (12 fetures) ws fitted to the trining dt set (R ) nd predicted the test dt set with R (Tble 3). For left hippocmpus, the lowest trining RMSE ws with 100 fetures tht lso included ge. The 5% tolernce model (12 fetures) ws fitted to the trining dt set (R ) nd predicted the test dt set with R (Tble 3). The performnce of the models ws lmost identicl when ge ws excluded. For the right EC, the lowest men trining RMSE ws with 70 fetures; 5% tolernce model (12 fetures) ws fitted to the trining dt set (R ) nd predicted the test dt set with R Finlly, for left EC, model with 90 fetures hd the lowest RMSE, nd 5% tolernce model (12 fetures) ws fitted to the trining dt set (R ) nd predicted the test dt set with R (Tble 3). Results of ll 2.5% models re presented in Supplementry Tbles 2 nd 3, nd the list of molecules included in ech clssifier is found in Supplementry Tble 1. The strength of ssocition between selected fetures nd ech model is shown in Fig. 2, nd the scled VI of ech lipid fter RF-RFE/ RFR-RFE for ech phenotype is shown in Supplementry Figure Lipid nnottion nd puttive identifiction We opted to nnotte the top fetures, in terms of VI from ech model nd fetures selected in more thn one model, using our in-house lipid dtbse nd MS/MS frgmenttion ptterns [4,11,12]. These fetures were nnotted s minly LCTs nd ChoEs, some were PCs nd sterol. Fig. 1 (G M) nd Tble 2 present the univrite ssocitions of these molecules with the respective phenotypes. The ssocition of the nnotted molecules with ll phenotypes is shown in Supplementry Figure 5. The rw intensity counts for ech AD ssocited lipid cross AD nd controls, long with the coefficients of vrition (reltive stndrd devition [RSD])

7 P. Proitsi et l. / Alzheimer s & Dementi - (2016) Tble 2 List of puttively identified metbolite molecules selected by the six rndom forest models Logistic regression nlysis Clinicl AD dignosis Liner regression nlysis of the pooled smples (QCs) re shown in Supplementry Tble Discussion m/z (ioniztion mode) Puttive metbolite molecule This is to our knowledge, the lrgest nontrgeted blood lipidomics study in AD to dte. Here, we expnded our Present study dt set Proitsi et l 2015 dt set Met-nlysis OR 95% CI P vlue OR 95% CI P vlue OR 95% CI P vlue 882 (2) PC 40: E204 NA NA NA NA NA NA 948 (1) TG 57: E E E (1) ChoE/TG y E E E (1) ChoE/TG y E E E (1) ChoE/TG y E E E (1) ChoE/TG y E E E (1) ChoE/TG y E E E (1) ChoE/TG y E E E204 Bet 95% CI P vlue NA NA NA NA NA NA Hippocmpus (right) 919 (1) TG 50: E205 NA NA NA NA NA NA Hippocmpus (Left) 943 (1) ChoE/TG E204 NA NA NA NA NA NA Entorhinl Cortex 367 (1) Sterol to E202 NA NA NA NA NA NA (Right) Entorhinl Cortex 816 (1) NA E202 NA NA NA NA NA NA (Left) ROD 771 (1) PC 36:3 z to E204 NA NA NA NA NA NA Abbrevitions: AD, Alzheimer s disese; ChoE, cholesteryl ester; CI, confidence intervl; m/z, mss-to-chrge rtio; OR, odds rtio; PC, phosphtidylcholine; ROD, rte of cognitive decline; TG, Triglyceride. NOTE. The six rndom forest models were for the clinicl AD dignosis, ROD, hippocmpus (R/L), nd entorhinl cortex (R/L) phenotypes. The ssocition of ech molecule is presented with the respective phenotype (i.e., primry phenotype of ssocition). The ssocition of the six molecules previously reported by Proitsi et l 2015 with AD is lso presented. Q vlue,0.05. y Fetures identified by Proitsi et l 2015 nd for the Proitsi et l., dt set semiquntified vlues re presented. z PC 36:3 hs m/z 770, nd m/z 771 is its C13 isotope. ChoE/TG indictes co-elution of ChoE nd TG molecules. recent work [9], nd using univrite nd multivrite pproches, we replicted the ssocitions between six previously reported blood lipids nd AD [9] nd reported their ssocition with brin trophy. We further identified combintions of lipids tht clssified AD ptients with reltively good ccurcy when tested in both test nd vlidtion dt set (.70%), nd combintions of molecules tht Tble 3 Rndom forest clssifier model results (clinicl AD dignosis) for the trining dt set nd predictions on the test dt set nd the Proitsi et l dt set Model (5% tolernce) Trining dt set (N 5 179) Test dt set (N 5 98) Vlidtion dt set (N 5 75) Sens. Spec. AUC Acc. Sens. Spec. AUC PPV NPV Acc. Sens. Spec. AUC PPV NPV Covrites only fetures y NA NA NA NA NA NA 24 fetures z fetures y 1 covrites NA NA NA NA NA NA 24 fetures z 1 covrites fetures y 1 6 ChoE/TG x NA NA NA NA NA NA 24 fetures z 1 6 ChoE/TG x fetures z 1 covrites NA NA NA NA NA NA 1 6 ChoE/TG x 24 fetures z 1 covrites 1 6 ChoE/TG x Abbrevitions: Acc, ccurcy; AUC, re under the curve; ChoE, cholesteryl ester; NPV, negtive predictive vlue; PPV, positive predictive vlue; Sens, sensitivity; Spec, specificity; TG, triglyceride. Age, sex, ε4. y 5% tolernce model including negtive ioniztion molecule. z 5% tolernce model excluding negtive ioniztion molecule. x Six fetures identified by Proitsi et l

8 8 P. Proitsi et l. / Alzheimer s & Dementi - (2016) Tble 4 Rndom forest regression model results for the trining dt set nd predictions on the test dt set for ech AD endophenotype Phenotype Hippocmpus (right) Hippocmpus (Left) Entorhinl cortex (right) Entorhinl cortex (left) Model (5% tolernce) Trin dt set (n 5 93) predicted chnges in disese progression (R for test dt set) nd brin trophy (R for ll test dt sets except for left EC). Although these signtures cnnot be used for dignostic purposes yet, they suggest importnt biologicl mechnisms ssocited with AD Identifiction nd role of lipids in AD Test dt set (n 5 28) RMSE R 2 RMSE R 2 Covrites only fetures y Covrites only fetures y Covrites only , fetures Covrites only , fetures ROD 10 fetures z Abbrevitions: RMSE, root men squred error; ROD, rte of cognitive decline. Age, sex, ε4. y Age ws included in the finl model. There ws no difference in either trin or test dt set performnce when ge ws excluded. z Covrites were lredy included in the clcultion for the ROD. We puttively identified two PC molecules; dditionlly, ChoEs nd triglycerides (TGs) were tenttively nnotted due to chromtogrphic coelution, nd finlly, we puttively nnotted molecule s sterol. The higher MS-MS sensitivity chieved here enbled the detection of number of dditionl lipids tht co-eluted with ChoE; these were nnotted s TGs (Tble 2). The ssocition of PCs with AD nd cognition hs been extensively described [4,8]. Here, one of the molecules most strongly ssocited with AD is puttive PC (PC 40:4), nd the top lipid in the ROD model is lso puttive PC (PC 36:3). In contrst to the sme species of molecules, we hve previously identified, both PCs re incresed in AD, nd PC 36:3 is ssocited with fster ROD. Although most studies to dte hve reported reduction of PC levels in AD, n increse in CSF PCs hs been observed in AD compred to control brins [19] nd recently in AD-like ptients bsed on their CSF Amyloid-bet42, Tu, nd Phospho- Tu-181 levels [20]. A recent study lso reported prllel increse of PCs contining sturted nd short-chin ftty cids in serum from AD ptients [21]. These suggest deregultion in the biosynthesis, turnover, nd cyl chin remodeling of phospholipids, in ccordnce with incresed phospholipid brekdown due to PLA2 [21] overctivtion. We hve lso reported ssocitions with low-chin nd very-low-chin triglycerides (LCTs/VLCTs; ftty cid chin length.16 crbons). One of the most interesting findings ws tht due to the higher MS-MS sensitivity chieved in this study, we were ble to observe puttive VLCTs tht were coeluting with ChoEs (Tble 2). We hve previously reported on the synthesis of ChoEs [9]; briefly, it tkes plce by trnsfer of ftty cids from PC to cholesterol, rection ctlyzed by lecithin cholesterol cyl trnsferse in plsm nd by cyl-coenzyme A: cholesterol cyl trnsferse 1 nd 2 (ACAT1 nd ACAT2) in other tissues, including the brin. The ssocition of LCTs/VLCTs with AD is noteworthy. Although overll TGs re seen s risk fctors for mny disorders including crdiovsculr disese (CVD) nd type 2 dibetes (T2D), numerous investigtions point to the diverse role of TGs with different chin lengths. It is known for exmple tht medium-chin triglycerides (MCTs) nd LCTs hve different metbolic pthwys in digestion nd bsorption [22]. Moreover, lthough LCTs of lower crbon number nd double bond content hve been ssocited with incresed CVD [23] nd T2D risk [24], LCTs with higher crbon number nd double bond content, like the ones here, hve been ssocited with decresed risk of T2D [24], wheres no ssocitions between T2D nd totl triglyceride levels were observed in the sme individuls [24]. Furthermore, decresed concentrtion of LCTs nd n incresed concentrtion of VLCTs hve been ssocited with longevity [25]. These findings re prticulrly interesting s most vegetble oils re comprised of long-chin ftty cids; however, only MTCs hve to our knowledge been implicted in AD, lthough findings re controversil [26]. When we previously investigted the ssocition of totl cholesterol nd TGs with AD in overlpping individuls using Mendelin rndomiztion, we found no evidence for n ssocition with AD [27]. Additionlly, we observed no difference in serum triglycerides, totl cholesterol, LDL cholesterol, nd HDL cholesterol between AD ptients nd controls for rndom subset of study prticipnts in this study (AD 5 102, controls 5 106) tht hd serum lipid mesures vilble for the sme visit, s well s no difference in the frequency of AD ptients nd controls who were tking sttins. On the other hnd, recent study reported n overlp between genes involved in elevted plsm lipid levels nd inflmmtion nd the risk for AD [28]. All these highlight the relevnce of investigting smller lipid frctions s they highlight specific steps in their biosynthesis nd metbolism tht my be ssocited with AD. Finlly, we observed n ssocition between most phenotypes nd feture of m/z 367. We previously described molecule with the sme mss nd similr retention time to be reduced in AD [9]. The molecule discovered here ws ssocited with n incresed risk for AD in both dt sets nd with reduced brin volume nd ws included in the ERC nd clinicl dignosis models. We believe this feture is frgment nd sterol, specificlly n isomer of desmosterol. Desmosterol is precursor of cholesterol nd seldin (DHCR24), which governs the metbolism of desmosterol

9 P. Proitsi et l. / Alzheimer s & Dementi - (2016) web 4C=FPO to cholesterol in specific brin res. Desmosterol hs been shown to inhibit b-secretse clevge of APP, nd the formtion of myloid-b nd lower desmosterol levels hs been found in the plsm nd brins of AD ptients compred to controls [29 32] Clssifier Phenotype AD-CTL XN XN XN XN XN XN AD-CTL AD-CTL (Proitsi et l) AD-CTL Met-nlysis ROD HIP_L HIP_R ERC_L ERC_R bet AD-CTL (Proitsi et l) AD-CTL & ERC_R AD-CTL & HIP_R ERC_L ERC_L & ERC_R ERC_R HIP_L HIP_L & HIP_R HIP_R HIP_R & ERC_R Fig. 2. Hetmp of the univrite ssocitions between fetures selected during rndom forest nlyses for ech phenotype. The color of ech box represents the univrite logistic regression bet coefficient (log[or]) for clinicl dignosis or the univrite scled liner regression bet coefficients for the rte of cognitive decline nd brin trophy, fter djusting for covrites. The strs on ech box represent the strength of the ssocition: P vlue,.05; P vlue,.01; P vlue,.001; P vlue,.0001; P vlue, The order of the metbolite molecules on the y-xis is bsed on hierrchicl clustering using the metbolites pirwise correltions. XN denotes negtive ioniztion mode feture. Abbrevitions: AD-CTL, Clinicl AD dignosis; ROD, rte of cognitive decline; HIP_L, left hippocmpus; HIP_R, right hippocmpus; ERC_L, left entorhinl cortex; ERC_R, right entorhinl cortex. ROD Although the ssocition of berrnt lipid metbolism in AD pthogenesis is undisputed [33 35]; t this stge, the mechnisms by which these chnges in lipids might occur in AD re uncler. One possibility involves AD selective ltertions to circulting lipid metbolism. However, nother

10 10 P. Proitsi et l. / Alzheimer s & Dementi - (2016) possibility reltes to cellulr lipid production. A number of phospholipids re synthesized within specilized region of the endoplsmic reticulum (ER) tht is closely ssocited with mitochondri, the mitochondri-ssocited ER membrnes (MAM). The close ssocition of MAM to mitochondri fcilittes C 21 nd phospholipid exchnge between the two orgnelles [36 38]. Recent studies hve shown tht MAM contcts re dmged in AD [39 42]. Becuse ERmitochondri contcts re required for the synthesis of certin lipids [36 38], such chnges my ffect lipid metbolism nd led to some of the chnges described here. Indeed, different APOE lleles hve been shown to influence MAM [43] Strengths nd limittions Here, we hve used lrge well-chrcterized AD cohort nd creful nd systemtic nlysis pipeline. Through bootstrpping, we hve reduced over-fitting, nd subsequently, we vlidted our results in n unseen dt set for ech phenotype nd n dditionl vlidtion dt set for clinicl AD dignosis. Our AD dignostic clssifier chieved 88% ccurcy on the trining dt set (summry of 100 bootstrps) nd predicted the test nd vlidtion dt sets with.70% ccurcy. Our trining dt set comprised of individuls mtched on ge nd gender. On the other hnd, the test dt set consisted primrily by femles, nd AD ptients were significntly older thn controls. Additionlly, the vlidtion dt set [9] included AD ptients nd controls of UK origin only, older thn the individuls of the current dt set. These findings highlight robustness in the model. For the AD endophenotypes, the Rndom Forest Regression models hd very good performnce on ll trining dt sets. Although the performnce dropped significntly in the test dt sets, we observed R for ll phenotypes except for Left EC (R ). The drop in performnce cn be ttributed to over-fitting of the trining dt sets nd the smller number of individuls with ROD/brin trophy mesures. The poor performnce of the Left EC is in greement with our univrite nlyses tht highlighted weker ssocitions with the EC for the whole smple; however, it is in contrst to the overll right-to-left symmetry in AD [17]. A limittion of this study is tht we were not ble to decipher the exct ftty cid chin structure of some fetures. Owing to the higher MS-MS sensitivity, we observed number of puttive ChoEs nd TGs co-eluting, which is commonly observed in lipidomics studies due to hundreds of lipids detected in one nlysis; to minimize co-elution problems, our chromtogrphic run is 2 hours long using ultr pressure chromtogrphy [11,44]. Additionlly, lthough this is the lrgest AD lipidomics study to dte, we cknowledge tht the smple size is still modest nd further repliction is required, especilly for the ROD nd brin trophy phenotypes. Moreover, lthough we hd informtion on the ROD, this clcultion ws bsed on the MMSE, which is crude mesure of mesurement of cognition. Furthermore, the present study did not contin n MCI cohort or informtion on conversion to MCI/AD, nd therefore, we do not know whether these fetures re ssocited with initition of AD. This study dditionlly suffers from limittions inherent to AD cse-control studies, such s the lrge number of comorbidities in old ge, the possibility tht some of the elderly controls my lredy crry pthology, nd tht some of the cliniclly dignosed AD my be pthologiclly non-ad dementis. Finlly, this study lcks informtion on BMI nd body ft distribution tht could potentilly explin some of the differences between AD ptients nd controls. However, through the longitudinl nture of these cohorts, we know tht ll the AD ptients used for our nlysis mintined the dignosis of AD s did ll controls for t lest 3 yers from their bseline visit. Additionlly, our informtion on disese progression nd brin trophy provide us with more precise phenotypes tht cpture different stges of disese pthology including the erly preclinicl stges. Given the good performnce of these models, we believe tht enrichment with dditionl individuls nd pthology informtion would increse their performnce. Finlly, lthough we did not hve BMI informtion for our cohort, we observed no difference in sttin use or serum lipids between AD cses nd controls. 5. Conclusion In conclusion, the findings of this study deepen our knowledge of AD disese mechnisms nd emphsize the importnce of investigting in detil different lipid frctions in dementi reserch. As it is not known whether the observed chnges in lipid levels re cuslly relted to or re just mrker of chnges in lipoprotein dynmics nd composition, studies tht ddress cuslity re essentil, s the success of trgeting specific molecules nd identifying potentilly cusl pthwys menble to intervention is predicted on these molecules being on the cusl pthwy. Finlly, integrting dditionl types of biologicl modlities such s protein, gene expression, nd genotype informtion my increse the fit of these models nd help us to understnd more bout the biologicl context in which these molecules operte. Acknowledgments The uthors thnk the individuls nd fmilies who took prt in this reserch. The uthors thnk Professor Chris Morris for his input on Mitochondril ER Membrnes (MAM). We would like to cknowledge the use of the computtionl Linux cluster nd the Biomedicl Reserch Centre Nucleus Informtics Tem supported by Ntionl Institute for Helth Reserch (NIHR) Mentl Helth Biomedicl Reserch Centre nd Dementi Unit t South London nd Mudsley NHS Foundtion Trust nd (Institute of Psychitry) King s College London. This work ws supported by the Ntionl Institute for Helth Reserch (NIHR) Mentl Helth Biomedicl Reserch Centre

11 P. Proitsi et l. / Alzheimer s & Dementi - (2016) nd Dementi Unit t South London nd Mudsley NHS Q5 Foundtion Trust nd [Institute of Psychitry] King s College London nd the 7th Frmework Progrmme of the Europen Union (ADAMS project, HEALTH-F ). Add- NeuroMed ws funded through the EU FP6 progrmme. Petroul Proitsi is n Alzheimer s Society Post-Doctorl Fellow. Richrd JB Dobson nd Cristin Legido-Quigley re prtilly supported from the Innovtive Medicines Inititive Joint Undertking under EMIF grnt greement No , resources of which re composed of finncil contribution from the Europen Union s Seventh Frmework Progrmme (FP7/ ) nd EFPIA compnies in-kind contribution. The funding bodies hd no role in the design, collection, nlysis, interprettion of the dt, writing of the mnuscript nd the decision to submit the mnuscript. Supplementry dt Supplementry dt relted to this rticle cn be found t RESEARCH IN CONTEXT 1. Systemtic review: The uthors reviewed the literture using PubMed nd reported key publictions. There is pressing need to identify noninvsive Alzheimer s disese (AD) biomrkers, nd blood metbolites re promising biomrkers tht could id erly dignosis nd ultimtely led to the development of more effective interventions. Recent blood metbolomic studies hve highlighted the role of lipid compounds in AD. However, most studies re smll nd reltively heterogeneous. 2. Interprettion: This study replicted previous ssocitions between blood lipids nd AD nd reported novel ssocitions between blood lipids nd clinicl AD dignosis, the rte of cognitive nd brin trophy. These findings deepen our knowledge of AD disese mechnisms nd suggest novel trgets for future work. 3. Future directions: Results of this study could be complemented with protein nd genetic dt. Future studies should ddress whether these chnges re cuslly relted to AD or re just mrker of chnges in lipoprotein dynmics nd composition. References [1] Mpstone M, Cheem AK, Findc MS, Zhong X, Mhyre TR, McArthur LH, et l. Plsm phospholipids identify ntecedent memory impirment in older dults. Nt Med 2014;20: [2] Oresic M, Hyotylinen T, Herukk SK, Sysi-Aho M, Mttil I, Seppnn-Lkso T, et l. Metbolome in progression to Alzheimer s disese. Trnsl Psychitry 2011;1:e57. [3] Trushin E, Dutt T, Persson XM, Mielke MM, Petersen RC. Identifiction of ltered metbolic pthwys in plsm nd CSF in mild cognitive impirment nd Alzheimer s disese using metbolomics. PLoS One 2013;8:e [4] Whiley L, Sen A, Heton J, Proitsi P, Grci-Gomez D, Leung R, et l. Evidence of ltered phosphtidylcholine metbolism in Alzheimer s disese. Neurobiol Aging 2014;35: [5] Zipser BD, Johnson CE, Gonzlez L, Berzin TM, Tvres R, Hulette CM, et l. Microvsculr injury nd blood-brin brrier lekge in Alzheimer s disese. Neurobiol Aging 2007;28: [6] Lindon JC, Holmes E, Nicholson JK. Metbonomics in phrmceuticl R&D. FEBS J 2007;274: [7] Suhre K, Shin SY, Petersen AK, Mohney RP, Meredith D, Wgele B, et l. Humn metbolic individulity in biomedicl nd phrmceuticl reserch. Nture 2011;477: [8] Simpson BN, Kim M, Chung YF, Beson-Held L, Kitner-Triolo M, Krut M, et l. Blood metbolite mrkers of cognitive performnce nd brin function in ging. J Cereb Blood Flow Metb 2016; 36: [9] Proitsi P, Kim M, Whiley L, Pritchrd M, Leung R, Soininen H, et l. Plsm lipidomics nlysis finds long chin cholesteryl esters to be ssocited with Alzheimer s disese. Trnsl Psychitry 2015;5:e494. [10] Lovestone S, Frncis P, Kloszewsk I, Mecocci P, Simmons A, Soininen H, et l. AddNeuroMed the Europen collbortion for the discovery of novel biomrkers for Alzheimer s disese. Ann N Y Acd Sci 2009;1180: [11] Whiley L, Godzien J, Ruperez FJ, Legido-Quigley C, Brbs C. In-vil dul extrction for direct LC-MS nlysis of plsm for comprehensive nd highly reproducible metbolic fingerprinting. Anl Chem 2012;84: [12] Whiley L, Legido-Quigley C. Current strtegies in the discovery of smll-molecule biomrkers for Alzheimer s disese. Bionlysis 2011;3: [13] Smith CA, Wnt EJ, O Mille G, Abgyn R, Siuzdk G. XCMS: processing mss spectrometry dt for metbolite profiling using nonliner pek lignment, mtching, nd identifiction. Anl Chem 2006;78: [14] Westmn E, Aguilr C, Muehlboeck JS, Simmons A. Regionl mgnetic resonnce imging mesures for multivrite nlysis in Alzheimer s disese nd mild cognitive impirment. Brin Topogr 2013; 26:9 23. [15] Simmons A, Westmn E, Muehlboeck S, Mecocci P, Vells B, Tsolki M, et l. MRI mesures of Alzheimer s disese nd the Add- NeuroMed study. Ann N Y Acd Sci 2009;1180: [16] Simmons A, Westmn E, Muehlboeck S, Mecocci P, Vells B, Tsolki M, et l. The AddNeuroMed frmework for multi-centre MRI ssessment of Alzheimer s disese : experience from the first 24 months. Int J Geritr Psychitry 2011;26: [17] Sttlecker M, Kiddle SJ, Newhouse S, Proitsi P, Nelson S, Willims S, et l. Alzheimer s disese biomrker discovery using SOMAscn multiplexed protein technology. Alzheimers Dement 2014;10: [18] Wilson RS, Arnold SE, Schneider JA, Kelly JF, Tng Y, Bennett DA. Chronic psychologicl distress nd risk of Alzheimer s disese in old ge. Neuroepidemiology 2006;27: [19] Wlter A, Korth U, Hilgert M, Hrtmnn J, Weichel O, Hilgert M, et l. Glycerophosphocholine is elevted in cerebrospinl fluid of Alzheimer ptients. Neurobiol Aging 2004;25: [20] Kol T, Klvins K, Seppi D, Kemmler G, Humpel C. Sphingomyelin SM(d18:1/18:0) is significntly enhnced in cerebrospinl fluid smples dichotomized by pthologicl myloid-bet42, tu, nd phospho-tu-181 levels. J Alzheimers Dis 2015;44: [21] Gonzlez-Dominguez R, Grci-Brrer T, Gomez-Ariz JL. Combintion of metbolomic nd phospholipid-profiling pproches for the study of Alzheimer s disese. J Proteomics 2014;104: