Comparative Lipidomic Analysis of Mouse and Human Brain with Alzheimer Disease*

Background: Lipid dyshomeostasis has been linked to Alzheimer disease (AD). Results: Lipidomic analyses of brain tissue from AD patients reveal region-specific changes in multiple bioactive lipids, some of which are phenocopied in AD mouse models. Conclusion: Lipid anomalies observed in AD may be linked to pathogenesis, including endolysosomal dysfunction. Significance: This study highlights the hypothesis-generating potential of lipidomics and its applicability to other diseases. Lipids are key regulators of brain function and have been increasingly implicated in neurodegenerative disorders including Alzheimer disease (AD). Here, a systems-based approach was employed to determine the lipidome of brain tissues affected by AD. Specifically, we used liquid chromatography-mass spectrometry to profile extracts from the prefrontal cortex, entorhinal cortex, and cerebellum of late-onset AD (LOAD) patients, as well as the forebrain of three transgenic familial AD (FAD) mouse models. Although the cerebellum lacked major alterations in lipid composition, we found an elevation of a signaling pool of diacylglycerol as well as sphingolipids in the prefrontal cortex of AD patients. Furthermore, the diseased entorhinal cortex showed specific enrichment of lysobisphosphatidic acid, sphingomyelin, the ganglioside GM3, and cholesterol esters, all of which suggest common pathogenic mechanisms associated with endolysosomal storage disorders. Importantly, a significant increase in cholesterol esters and GM3 was recapitulated in the transgenic FAD models, suggesting that these mice are relevant tools to study aberrant lipid metabolism of endolysosomal dysfunction associated with AD. Finally, genetic ablation of phospholipase D2, which rescues the synaptic and behavioral deficits of an FAD mouse model, fully normalizes GM3 levels. These data thus unmask a cross-talk between the metabolism of phosphatidic acid, the product of phospholipase D2, and gangliosides, and point to a central role of ganglioside anomalies in AD pathogenesis. Overall, our study highlights the hypothesis generating potential of lipidomics and identifies novel region-specific lipid anomalies potentially linked to AD pathogenesis.

which encodes a protein involved in cholesterol metabolism and lipid transport (3). Importantly, a variety of genes have been recently linked to LOAD through genomewide association studies and are directly or indirectly connected to lipid metabolism or cellular membrane dynamics (4,5).
Strong evidence now links AD with aberrant lipid homeostasis, in part for the following reasons (6,7). First, neuronal lipid composition regulates the trafficking and/or activity of the key membrane-bound proteins controlling A␤ levels, including APP, BACE1, and presenilins. For instance, ␥-secretase activity is regulated by membrane levels of cholesterol and sphingomyelin (SM) (8). Second, A␤ exerts its cytotoxic effects primarily by perturbing cellular membranes, in part by modulating the activity of phospholipases, such as PLA 2 (9), PLC (10), and phospholipase D 2 (PLD 2 ) (11). Third, lipids, such as ganglioside GM1, modulate the pathogenic potential of A␤ by affecting its propensity to aggregate (12). It is thus clear that an in-depth analysis of brain lipid composition changes in AD would provide a basis for further investigation of cell signaling and metabolic pathways that are disrupted in AD. Advances in lipidomics, particularly electrospray ionization-mass spectrometry, have facilitated the precise profiling of lipid species within tissues, including low abundance lipid classes that participate in cell signaling and membrane trafficking (13). Several studies focusing on defined families of lipids have made use of this technology to analyze lipid composition from AD brain tissue, showing changes in plasmalogen phosphatidylethanolamine (pPE) (14), sulfatide (Sulf) (15), ceramide (Cer), galactosylceramide (GalCer), cholesterol (16), and alkylacylglycerophosphocholine (17). Our own study has shown that a specific molecular species of phosphatidic acid (PA) previously linked to neurodegeneration in the fly (18) accumulates in the forebrain of old Swedish APP (swAPP) mice (11).
In this study, we present the lipidomic analysis of postmortem brain tissue obtained from patients with LOAD, specifically regions that typically are either moderately (prefrontal cortex, PFC) or severely (entorhinal cortex, ERC) affected in AD as well as a control region (cerebellum, CRB) that is largely unaffected by this disorder (19,20). We also provide the lipidome of forebrain tissue obtained from three commonly used transgenic mouse models of FAD, namely (i) PSEN1 M146V (PS1) (21), (ii) the Swedish mutant of APP K670N,M671L (APP) (22), and (iii) PS1-APP double transgenic (23) lines. Using a combination of normal phase and reverse phase LC-MS approaches, we analyzed 26 different lipid subclasses including lysophospholipids, glycerophospholipids, sphingolipids, glycerolipids, and sterols, covering over 300 lipid species. Our data show that multiple lipid subclasses are dysregulated in both the human AD tissue and AD mouse models, uncovering novel lipid pathways that may be linked to abnormal cellular activity and thus potential targets for further studies.

EXPERIMENTAL PROCEDURES
Annotation of Lipid Species-Glycerophospholipids, lysophospholipids, and diacylglycerol (DAG) were annotated as Ͻlipid subclassϾ Ͻtotal fatty acyl chain lengthϾ:Ͻtotal number of unsaturated bondsϾ. Sphingolipids were annotated as Ͻlipid subclassϾ Ͻsphinogoid base residueϾ/Ͻfatty acyl res-idueϾ. Triacylglycerol (TAG) was annotated as ϽTAGϾ Ͻtotal fatty acyl chain lengthϾ:Ͻtotal number of unsaturated bondsϾ/Ͻfatty acyl residueϾ). Cholesterol ester (CE) was annotated ϽCEϾ Ϫ Ͻfatty acyl residueϾ.
Human Brain Tissue-Frozen postmortem brain tissue samples were obtained from the New York Brain Bank at Columbia University Medical Center. Three different brain regions were sampled from different subjects: PFC, ERC, and CRB. Morphologically, each dissected region consisted predominantly of gray matter with minimal associated underlying white matter. AD was defined as "high probability of AD" based on NIA, National Institutes of Health-Reagan Institute (NIA-RI) criteria. Controls were free of major neuropathological diagnoses, and either did "not meet" or met "low probability of AD" by NIA-RI criteria. Further sample information, including a semiquantitative assessment of neuronal loss, neurofibrillary tangles, amyloid neuritic plaques, and Lewy body burdens, for the PFC and ERC, is presented in supplemental Dataset S3. On average, the ERC exhibited higher levels of pathology than the PFC in LOAD samples.
AD Mouse Models-A mixed population of wild-type, PS1, APP, and PS1-APP mice were described previously (21)(22)(23)(24), aged between 9 and 11.5 months, were used for full lipidomics analysis in this study. An older group of wild-type and APP mice aged between 14 and 15 months were used for targeted PA analysis. SwAPP mice with the Pld2 ϩ/ϩ and Pld2 Ϫ/Ϫ genotypes with an average age of 12 months were described before (11) and used to examine the changes in sphingolipids and cholesterol ester. All mice were sacrificed by cervical dislocation, the forebrain dissected immediately and frozen in liquid nitrogen before storage at Ϫ80°C.
Analysis of Lipids Using High Performance Liquid Chromatography-Mass Spectrometry-Lipid extracts were prepared using a modified Bligh/Dyer extraction procedure, spiked with appropriate internal standards, and analyzed using an Agilent 1200 HPLC system coupled with an Applied Biosystem Triple Quadrupole/Ion Trap mass spectrometer (3200Qtrap). Further cross-lab validation of our approach can be found in our previous work (25). Separation of individual lipid classes of polar lipids by normal phase HPLC was carried out using a Phenomenex Luna 3 silica column (inner diameter 150 ϫ 2.0 mm) with the following conditions: mobile phase A (chloroform:methanol:ammonium hydroxide, 89.5:10:0.5), B (chloroform:methanol:ammonium hydroxide:water, 55:39:0.5:5.5); flow rate of 300 l/min; 5% B for 3 min, then linearly changed to 30% B over 24 min and maintained for 5 min, and then linearly changed to 70% B over 5 min and maintained for 7 min. Then, the gradient was changed back to the original ratio over 5 min and maintained for 6 min for column re-equilibration. Multiple reaction monitoring transitions were set up for quantitative analysis of various polar lipids (26,27). Individual lipid species were quantified by referencing to spiked internal standards. TAG, DAG, and CE were analyzed using a modified version of reverse phase HPLC-electrospray ionization-MS described previously (28). Briefly, separation of TAG and CE from polar lipids was carried out on an Agilent Zorbax Eclipse XDB-C18 column (inner diameter 4.6 ϫ 150 mm) using an isocratic mobile phase chloroform, methanol, 0.1 M ammonium acetate (100:100:4) at a flow rate of 250 l/min. TAG were calculated as relative contents to the spiked d 5 -TAG 48:0 internal standard (CDN isotopes), whereas cholesterol esters were normalized to corresponding d 6 -C18 cholesterol ester (CDN isotopes). DAG species were quantified using 4-methyl 16:0 diether DAG as an internal standard (Avanti Polar Lipids). Free cholesterol was analyzed by reverse phase HPLC atmospheric pressure chemical ionization MS, utilizing the same XDB-C18 column as above but with chloroform:methanol, 1:1 (v/v), as the mobile phase at a flow rate of 500 l/min (29). The instrument was operated in the positive mode with a vaporizer temperature of 500°C and corona current of 3 A. Free cholesterol was quantified relative to spiked d 6 -cholesterol (CDN isotopes) internal standard.
Data Analysis and Presentation-Lipid levels for each sample were calculated by summing up the total number of moles of all lipid species measured by all three LC-MS methodologies, and then normalizing that total to mol %. The final data are presented as mean mol % with error bars showing mean Ϯ S.E. All lipid species in the human samples and more than 98% of the lipid species measured in the mouse samples were found to have equal variance (data not shown). For the human samples, a two-tailed Student's t test was used for statistical analysis. The mouse data were analyzed by one-way analysis of variance followed by post hoc multiple comparisons Tukey range test as indicated accordingly. In all cases, *, p Ͻ 0.05; **, p Ͻ 0.01; ***, p Ͻ 0.001. For all bar graphs, error bars represent the mean Ϯ S.E.
Comparative Lipid Profiling of FAD Mouse Models Forebrain Brain Tissue-To determine whether any of the lipid changes seen in LOAD are phenocopied in the brain of commonly used mouse models of FAD, we characterized the forebrain lipid composition of WT, PS1, APP, and PS1-APP mice within an age range (i.e. 9 -11.5 months) where transgenic mice have been  Table 1. The absolute level of individual lipid species is presented in supplemental Dataset S1. characterized for various biochemical, morphological, electrophysiological, and behavioral phenotypes. Although the PS1 mice exhibit subtly increased A␤42 levels without noticeable pathology (21), the APP mice show significantly elevated A␤ levels and develop numerous plaques between 9 and 13 months (22). Comparatively, A␤ production is markedly increased in the PS1-APP mice, which begin developing a severe plaque burden by 6 -7 months (23). Reflecting in part the severity of A␤ burden, synaptic and behavioral impairments are least apparent in the PS1 mice and highly exacerbated in the PS1-APP mice (23,24).
Lipid Dysregulation in FAD Mouse Models-All three AD transgenic mouse models showed distinctive lipid profiles com-pared with the WT mice (Fig. 2, 4, supplemental Fig. S5 and Dataset S2). Fatty acyl chain remodeling was evident, with an overall decrease in long-chain phospholipid (C Ͼ 40), and a corresponding increase in medium-to short-chain phospholipids (C Ͻ 36) (Fig. 3G). Related to these changes, a decrease in fatty acyl chains carrying four polyunsaturated bonds was also replaced by an increase in monounsaturated and saturated fatty acyls (Fig. 3H). In the APP and PS1-APP mice but not the PS1 mice, there was a similar decrease in long-chain sphingolipids (C Ͼ 40) that was balanced by increases in short and medium length species (C ϭ 36 to 38) (Fig. 3I).
Similar to the other transgenic mice, the APP mutant showed decreased levels of PS (ϳ33%), PI (ϳ41%), PG (ϳ25%), LBPA (ϳ43%), and LPE (ϳ20%) with an overall increase in PC (ϳ10%). A decrease in various molecular species of PA (e.g. 32:1 and 34:2) was also observed. These data are in contrast to the specific increase of PA 34:2 seen in a different strain of Swedish APP transgenic mice that were also older in age (11). To determine whether this was an age-related phenomenon, we conducted a targeted PA analysis of 14 -15-month-old APP mice of the same genetic background and found that PA 34:2, and to a lower extent other molecular species of PA, were indeed elevated in these mice relative to 9 -11-month-old APP mice (supplemental Fig. S6). Similar to the PS1 mutant, an increase in SM was observed in APP mice, reflecting the selective accumulation of medium-chain length SM d18:1/18:1, d18:1/18:0, and d18:0/18:0. This may have resulted from either increased catabolism of GluCer and/or Sulf, both of which showed significant decreases (ϳ10 and 26%, respectively). Finally, the APP mice showed a trend for increasing total levels of GM3 and CE as represented by specific molecular pools such as GM3 d18:1/ 18:0 and CE-22:6 (supplemental Fig. S5).
PLD 2 Ablation Causes Normalization of GM3 Levels-Lipidomics can be harnessed to identify lipid pathways that are specific to perturbations made to different biological systems (34,35). Combining our current data set with pharmacological treatments or genetic models that affect pathophysiological phenotypes of AD may identify bona fide targets of the disease. We recently reported that swAPP transgenic mice lacking PLD 2 , which hydrolyzes PC to generate PA, are protected from the synaptotoxic and memory-impairing actions of A␤ in vitro and in vivo (11). This protection occurs despite the production of high amounts of A␤ indicating that the toxic effects of A␤ may be neutralized. To understand the molecular basis of this protection, we conducted a targeted lipidomic analysis of the forebrain of these mice, which were slightly older than those used in Figs. 2-4 (see "Experimental Procedures"). Although most sphingolipid classes appear to be unaltered in the mutant mice, profound alterations were observed for GM3 and CE (Fig.  5A). In the Pld2 Ϫ/Ϫ only mice, most molecular species of GM3 were slightly reduced in the forebrain. In contrast, expression of the swAPP transgene alone produced a significant increase in GM3 and CE, similar to the PS1-APP animals (Fig. 2). Most importantly, when swAPP expression and PLD 2 ablation were combined, GM3 but not CE levels were normalized, suggesting that ganglioside metabolism may be correlated to the pathogenicity associated with the swAPP model and its rescue in the Pld2 Ϫ/Ϫ background (Fig. 5B).

DISCUSSION
In the present study, we conducted an in-depth lipidomic analysis of AD brain tissue that covered over 300 lipid species from 26 lipid subclasses. We observed that different regions of the human brain undergo heterogeneous changes in lipid metabolism in AD-affected patients. Although the CRB lipid profile was largely unaffected, significant changes were observed in the diseased PFC and ERC ( Fig. 1 and supplemental  Fig. S1). Between these two tissues, the PFC seemed to carry more severe lipid alterations, with decreased levels of LPC, LPCe, and PE and elevated levels of Cer, GluCer, GalCer, and DAG. In contrast, the ERC only showed significant increases in LBPA, SM, GM3, and CE levels. However, there were also numerous changes to individual lipid species in different subclasses, suggesting that fatty acid carbon chain specificity may play an important role in AD pathogenesis ( Fig. 4 and supplemental Figs. S2 and S3). In the PFC, this included various molecular species from the ePC, SM, Sulf, and CE, whereas ERC displayed variability in ePC, PE, and TAG. Importantly, the differences in lipid signatures suggest that these two subregions encounter different aspects or stages of AD pathogenesis, as supported by functional and pathological analyses (19,20). As expected, neuronal loss was more obvious in the ERC than in the PFC (supplemental Dataset S3). Because neuronal loss is typically associated with gliosis, the lipidome of regions undergoing significant neuronal loss, such as the ERC, may reflect in part changes in the neuron/glia ratio.
Because mouse models are used extensively to study AD pathogenesis, we set out to define whether specific changes in lipid metabolism in brain tissue with LOAD are phenocopied by mouse models of FAD and if these pathways can be targeted for further studies. Granted, this type of comparative analysis faces inherent challenges, including species-specific differences and disease etiology (i.e. FAD versus LOAD), and are thus not expected to be a perfect model for studying LOAD. With these caveats in mind, we observed dramatic lipid changes in all three transgenics that were not seen in humans. Specifically, a variety of anionic phospholipids and sphingolipids, such as Sulf, were decreased in transgenic models. Furthermore, genotype-specific lipid changes were observed and in several cases, the alterations present in the PS1-APP mice reflected an additive or synergistic effect of bigenic FAD gene expression. For instance, PS1-APP mice exhibited a striking elevation of CE and GM3, which was present to a lower extent in the APP mice and virtually absent from PS1 mice ( Fig. 2 and supplemental Fig. S5).
Our analysis points to dysregulation of specific lipids potentially involved in key anomalies seen in AD including synaptic dysfunction and neurotransmission deficits (36). For instance, DAG metabolism, which is highly aberrant in diseased PFC, has been implicated in a number of synaptic processes including the regulation of dendritic spine morphology, plasticity, and neurotransmitter release (37). Importantly, overactivation of PKC, a target of DAG, interferes with PFC-regulated working memory (38), a key feature of AD. DAG accumulation in  the PFC likely occurs through A␤-stimulated hydrolysis of PI(4,5)P 2 by PLC (10), based on the specific accumulation of polyunsaturated DAG 38:4 (supplemental Fig. S2), a primary metabolite of PI(4,5)P 2 (30). Compounding this effect may be the loss of DAG kinase activity because synaptic targeting of DAG kinase depends on PSD95, whose levels are decreased in AD-affected brain tissue (11,37). Furthermore, aberrant accumulation of sphingolipids in the PFC may also contribute to dysregulation of synaptic vesicle fusion and trafficking of NMDA receptors (39). This may reflect the fact that A␤42 activates neutral sphingomyelinase, which produces Cer (8), and that sphingolipid enzymes are significantly altered in this tissue (40). Overall, our data suggest that aberrant DAG and sphingolipid metabolism are likely to participate in synaptic dysfunction in diseased PFC.
Another important observation is the LBPA increase in diseased ERC. LBPA is enriched specifically within late endosomes, where it plays a role in the biogenesis of multivesicular bodies (41). The enrichment of LBPA is thus consistent with a dysfunction of the late endosomes/multivesicular bodies. Aberrant endosomes have been previously reported in Down syndrome and AD (42) and growing evidence, most recently through genome wide association (4,5) and gene expression (43) studies, indicates that defects in the endolysosomal pathway are linked to AD. An emerging hypothesis in the field is that defects in the endolysosomal pathway can affect amyloidogenesis, as organelles along this pathway contain high levels of APP as well as functionally active pools of BACE1 and ␥-secretase (44,45). Additionally, lysosomal defects alter the clearance of autophagosomes, which are double-membrane organelles of the autophagy pathway that mediate the elimination of protein aggregates, including Tau (45). Notwithstanding other possible functional implications (see below), the ERC accumulation of SM, GM3, and CE adds strong support to the endolysosomal dysfunction hypothesis. Indeed, SM and GM3 traffic within and between the plasma membrane and endosomes (46), whereas intracellular CE, normally found in lipid droplets, has also been detected in late endosomes (47). Of particular relevance, the accumulation of SM and gangliosides impairs lysosomal degradation of APP-CTFs, which enhances amyloidogenesis (48). A likely secondary implication of SM and GM3 accumulation is the propensity for ganglioside-enriched membrane microdomains at the cell surface and in endosomes to seed A␤ oligomerization (12). Taken together, the overall lipid profile changes seen in the ERC of AD patients suggest a pathogenic cascade that is reminiscent of lipid traffic jams described in lipid storage diseases such as Niemann-Pick types A and C (49).
Although lipid changes in FAD models are in most aspects vastly different from the LOAD samples, our lipidomic analysis suggest that the APP and PS1-APP mice are likely suitable models to study the role of ganglioside (GM3) and CE accumulation in AD, particularly because these anomalies are also observed in the ERC of AD patients (Fig. 4). A key question is whether the reversal of these lipid changes can ameliorate the AD-relevant phenotypes of the transgenic models, including enhanced amyloidogenesis. A partial answer, as well as an important "proof of concept" validation of our analysis, is that the pharmacological inhibition and genetic ablation of acyl-coenzyme A:cholesterol acyltransferase 1, a CE-synthesizing enzyme, lower A␤ production and improves memory related tasks in transgenic models of AD (32). Our own studies suggest that the protective role of PLD 2 ablation in APP transgenic mice may be related to gangliosides, as suggested by the normalization of GM3 levels to those found in control mice (Fig. 5). This raises important questions about how PLD and its product PA are related to ganglioside metabolism and perhaps endolysosomal dysregulation in AD (11,50). Additionally, because PLD 2 ablation mitigates the signaling, but not the biogenesis of A␤ (11), we hypothesize that the GM3 elevation observed in the APP models (which is corrected in the Pld2 knock-out) occurs downstream of A␤. Interestingly, the normalization of GM3 levels in the FAD mouse lacking PLD 2 occurs independently of CE, suggesting that CE elevation in FAD mice may not be driven by increased A␤ signaling per se. Overall, our data suggest that there are multiple ways to relieve lipid traffic jams that may be detrimental to neural cells in AD pathogenesis.
Finally, our study suggests that AD is associated with a more general perturbation of membrane properties. The length and degree of saturation are important determinants of many membrane characteristics including membrane thickness, fluidity, local curvature, and molecular packing, which in turn regulate the activities of membrane-bound enzymes (51). Intramembrane proteolysis is likely to be exquisitely regulated by membrane thickness, with major implications for ␥-secretase-mediated cleavage of APP (52). Interestingly, FAD mouse forebrains and to a lower extent diseased PFC and ERC, exhibited a relative increase in short-chain lipids that was accompanied by a reduction in polyunsaturated phospholipids (Fig. 3). Similar results have been described in synaptosomal and detergent-resistant membranes due to aging (53) and AD (54). It is well known that long-chain polyunsaturated phospholipids display an aversion to highly ordered lamellar microdomains, whereas short-chain monounsaturated or saturated phospholipids interact favorably with sphingolipids and cholesterol in lipid rafts (6,55). These features thus point to an increased propensity of membranes from the AD-affected brain to form lipid rafts, which may have consequences on the amyloidogenic processing of APP (6,7).
A fundamental challenge in understanding a complex disease using a systems based approache, such as lipidomics, is the identification of pathophysiologically relevant changes among a myriad of changes. Here, we have capitalized on the exquisite sensitivity of LC-MS to characterize the lipidome of brain tissue from AD-affected humans and commonly used FAD mouse models. We describe novel region-specific lipid changes in human AD tissue that may contribute to AD etiology. We also identified overlaps in lipid dysregulation between mice and humans, suggesting that the FAD mouse model may be suitable for the study of specific aspects of AD-associated lipid dysregulation. Our data thus provides a frame of reference for lipids that can be targeted for therapeutic interrogation either through pharmacological or enzymatic manipulations, to improve learning and memory in the context of AD. In particular, PLD 2 isoform-specific inhibitors (56) may prove to be valuable therapeutic agents in AD and thus require further evaluation. Our study also provides further incentive to systemati-cally analyze the various animal models of AD available in the biomedical community as well as to expand these types of profiling to other AD-vulnerable brain regions, so that a better understanding of AD-linked lipid dysfunction can be achieved.