Background
Lipids are important components of the brain that play a critical role in the membrane formation of neuronal cells, and participate in essential physiological functions such as cellular transport, energy storage, in addition to acting to modulate transmembrane proteins and signaling molecules, promoting effective signal transduction, and regulating gene expression [
1‐
3]. In recent years, growing evidence from both animal models and humans’ studies has identified that abnormal lipid metabolites were associated with the molecular mechanisms underlying Alzheimer’s disease (AD) pathophysiology beyond amyloid plaques and neurofibrillary tangles [
4‐
9]. In fact, altered plasma lipid profiles have appeared to exacerbate cognitive decline, subsequently increasing the risk of the incidence of AD in nondemented elderly adults [
7,
10‐
14]. Specifically, recent biological and neuroimaging data have indicated that the dysfunctional composition of lipid rafts, primarily located in membrane microdomains and serving as an important platform for signal processing, may contribute to AD pathophysiology [
2,
11]. Cholesterol, as a major component of lipid rafts, is thought to be involved in amyloid precursor protein (APP) processing and β-amyloid (Aβ) overproduction characterized as a key feature of AD pathophysiology [
12], while gemfibrozil, a fibric acid agent commonly used to treat hyperlipidemias in clinic, significantly reduces amyloid pathology and reverses memory deficits in APP-PSEN1ΔE9 mice [
15], a murine model that mimics AD-like pathology and cognitive decline. As changes of lipoprotein in the blood can be detected prior to cognitive decline, it is of considerable interest to know whether lipid pathway-based metabolites substantially contribute to AD pathophysiology [
16]. However, to date, it remains unclear how lipid metabolites, cerebral spinal fluid (CSF) biomarkers, and brain function are linked or interacted with the progression of cognitive decline in preclinical or clinical AD patients.
Brain network integrity plays an instrumental role in the regulation of high-order cognitive function. Resting-state networks (RSNs), which measure temporal correlation depend on intrinsic blood oxygenation level dependent (BOLD) signals within large-scale systems and provide a powerful tool to investigate network integrity between structurally segregated and functionally specialized brain regions at the system level [
17]. Importantly, the spatial–temporal evolution of RSNs has been found to be tightly associated with neural correlates of cognitive impairment observed in preclinical and clinical AD patients [
18‐
21], including default mode network (DMN), executive control network (ECN), salience network (SAN), attention network (AN), and visuospatial network (VIS), suggesting that changes in distributed networks at a large-scale system level could predict clinical progression and neurodegeneration [
18,
22,
23]. Recently, particular attention to network integrity has shifted towards investigation of molecular pathological changes invoked in intrinsic large-scale network dynamics supporting diverse cognitive function [
24]. Specifically, increasing evidence has demonstrated that neural correlates of the disrupted connectivity of RSNs in cognitively healthy individuals with brain amyloidosis or AD-related genetic risk factors were similar to abnormalities observed in symptomatic AD [
25‐
27]. As such, it may be possible that RSNs could be used as an intermediate phenotype linking downstream cognitive decline and upstream molecular cascading events underlying AD pathophysiology. Dysregulation of lipids is substantially associated with the disrupted architecture of RSNs and is directly involved in the molecular and cellular changes underlying AD pathophysiology [
28,
29]. For example, high serum cholesterol has been associated with decreased cortical and hippocampal volumes in cholesterol-fed rabbits [
30] and disrupt functional connectivity of the SAN in the non-demented elderly [
28]. Increased low-density lipoprotein cholesterol (LDL-C) levels causes a detrimental effect to posterior cingulate gray matter volumes and verbal memory [
31], while elevated high-density lipoprotein cholesterol (HDL-C) provides protection against hippocampal atrophy and AD [
32,
33]. From our previous work, we previously reported that the effects of the accumulation of genetic variants of cholesterol-pathway molecules produces widespread effects on cortico-subcortical-cerebellar spontaneous brain activity in amnestic mild cognitive impairment (aMCI) patients [
34]. These findings suggest that several, distinct lipidomic signatures influence brain network integrity and subsequently contribute to AD. However, it remains unclear how altered lipid metabolites impinge on the dynamic spatiotemporal patterns of RSNs as AD progresses. Indeed, in the context of lipid-centric gene and protein changes, evaluation of the potential effects of lipid abnormalities that affect dynamic brain network trajectory and CSF biomarkers, subsequently leading to cognitive decline, are beneficial in order to capture a more holistic picture of the processes of AD.
In the present study, a new approach was used that combining large-scale brain networks with canonical correlation analysis (CCA) to explore the effects of lipid metabolic disturbance on the dynamic trajectory of ten RSNs changes and molecular biomarkers that promote cognitive decline following AD progression. Firstly, the relationship between lipid-centric gene variants and proteins, CSF biomarkers, and cognitive performance across the AD spectrum (ADS) was examined. Secondly, the dynamic trajectory of large-scale network changes was identified both within- and between ten predefined RSNs from cognitive normal (CN) healthy to mild AD stage individuals. Thirdly, the potential associations between lipid-related gene variants and proteins, and the dynamic trajectory of large-scale network connectivity, CSF biomarkers, and cognitive performance were explored using CCA. Fourth, a support vector machine (SVM) model of machine learning was used to distinguish ADS patients from CN subjects. Finally, path analysis with structural equation modeling (SEM) was used to test the effects of lipoproteins on large-scale RSNs, CSF biomarkers, and cognitive performance. Taken together, these findings provided an integrated view of lipid metabolite abnormalities exacerbated cognitive decline and increased the risk of AD occurrence via mediating large-scale brain network integrity and promoting neuropathological processes.
Methods
Participants
All data at baseline were extracted from the Alzheimer's disease Neuroimaging Initiative (ADNI) database (
http://adni.loni.usc.edu) prior to January 20th, 2020. Data for lipid gene and protein expression, CSF biomarkers and that of imaging quality control of a total of 124 subjects incorporating 51 cognitive normal (CN), 26 early amnestic mild cognitive impairment (EMCI), 26 late mild cognitive impairment (LMCI) and 21 mild Alzheimer’s disease (AD) subjects were included in the present study (Table
1). Detailed inclusion and exclusion criteria were provided in Additional file
1.
Table 1
Demographic data, lipid pathway-based genotypes, cerebrospinal fluid biomarkers, and global cognitive performance across the AD spectrum
Age (years) | 74.08 ± 5.79 | 70.04 ± 6.87 | 70.81 ± 7.14 | 71.81 ± 7.77 | 0.051 |
Gender (F/M) | 30/21 | 14/12 | 11/15 | 9/12 | 0.447* |
Education (years) | 16.31 ± 2.59 | 15.27 ± 2.51 | 16.31 ± 2.51 | 15.14 ± 2.76 | 0.157 |
Multiple protective genes |
CLU T status (TC + TT/CC) | 38/13 | 17/9 | 18/8 | 13/8 | 0.710* |
LDLR A status (AG + AA/GG) | 38/13 | 16/10 | 15/11 | 11/10 | 0.240* |
LRP1 T status (TC + TT/CC) | 15/36 | 8/18 | 5/21 | 10/11 | 0.214* |
PICALM A status (AG + AA/GG) | 27/24 | 15/11 | 16/10 | 11/10 | 0.884* |
Multiple risk genes |
APOE ε4 status (+ / −) | 15/36 | 14/12 | 12/14 | 15/6 | 0.008* |
SORL1 G status (TG + GG/TT) | 19/32 | 12/14 | 12/14 | 6/15 | 0.548* |
CETP A status (AG + AA/GG) | 46/5 | 24/2 | 23/3 | 19/2 | 0.974* |
ABCA1 G status (AG + GG/AA) | 45/6 | 25/1 | 21/5 | 19/2 | 0.369* |
BIN1 C status (TC + CC/TT) | 27/24 | 17/9 | 15/11 | 11/10 | 0.739* |
Cerebrospinal fluid biomarkers |
Aβ (pg/ml) | 192.79 ± 50.17bc | 183.61 ± 50.66d | 168.77 ± 50.80 | 140.40 ± 43.59 | 0.001 |
Tau (pg/ml) | 68.53 ± 34.14c | 79.32 ± 51.89d | 86.01 ± 52.19e | 129.29 ± 61.42 | < 0.001 |
pTau (pg/ml) | 34.18 ± 16.58bc | 39.60 ± 24.73d | 48.42 ± 33.50 | 55.23 ± 26.13 | 0.005 |
Global cognitive performance |
MMSE | 28.84 ± 1.16abc | 27.92 ± 2.13d | 27.73 ± 1.54e | 22.67 ± 2.50 | < 0.001 |
ADAS-Cog | 10.76 ± 6.53abc | 14.19 ± 6.58d | 16.96 ± 5.32e | 35.81 ± 8.99 | < 0.001 |
Demographic data such as age, gender and years of education were enrolled in this study. Based on the cholesterol metabolism pathway, nine candidate genes were selected: CLU, LDLR, LRP1, PICALM, SORL1, CETP, ABCA1, BIN1 and APOE (Tables
1 and
2). Hardy–Weinberg equilibrium test for each allele was calculated with chi-square test. In addition, thirty-eight lipid metabolic biomarkers were obtained. The detailed acquisition and selection procedures of plasma lipids were available in the Additional file
1. Further, CSF biomarkers including Amyloid-β 1 to 42 peptide (Aβ), total tau (Tau) and tau phosphorylated at the threonine 181 (pTau) were collected. The MMSE and Alzheimer’s Disease Assessment Scale-Cognitive Subscale (ADAS-Cog) were used to measuring global cognitive function.
Calculation of polygenic scores
Genes were divided into two categories: protective or hazardous, depending on the value of odds ratio (OR) for each gene. For OR values > 1, the locus was defined as hazardous, while for OR < 1 there were defined as protective variants. Polygenic scores were defined as a sum of ORs of multiple loci. The concept of relative risk score (RRS) utilized in the present study was defined as genetic risk score (GRS) minus genetic protective score (GPS). Due to the strong risk effect of the APOE genotype, GRS was calculated with the APOEε4 (GRS) and without the APOEε4 genotype (GRS_n), respectively. Consequently, RRS was also separated into RRS with APOEε4 (RRS) and RRS without APOEε4 (RRS_n). Gene information was detailed in Additional file
1: Table S1.
Functional network construction
Resting-state functional MRI image acquisition and preprocessing procedures were described in Additional file
1. The atlas of Power et al. [
35] was used to partition the brain of each participant into 226 cortical and subcortical areas. Subsequently, network connectivity was calculated within 10 RSNs as defined by previous fMRI studies [
35,
36]. We also calculated network connectivity between all pairs of the 10 RSNs, as well as between each RSN and all other RSNs (i.e., one-versus-all-others). The detailed construction of the network is shown in Additional file
1.
Statistical analysis
Comparisons between groups used one-way analysis of variance for continuous variables and chi-square tests for categorical variables. The significance level was set at p < 0.05. Post hoc analysis with least significance difference (LSD) correction (p < 0.05) was used to compare differences between two groups. All statistical analyses were conducted using SPSS v25 software (SPSS, Inc., Chicago, IL, USA).
To investigate correlations among polygenic scores (including GPS, GRS, RRS, GRS_n and RRS_n), lipid metabolites in the blood, CSF biomarkers, and cognitive performance in AD spectrum individuals, linear and binomial nonlinear regression analyses were employed, after controlling the covariates of age, gender, and years of education. The significance level was set at p < 0.05.
Each network metric (within-, one-versus-all-others-, and pairwise between-network connectivity) was compared across groups using generalized linear model analysis adjusted for age, gender, and education as covariates. All p values were adjusted for multiple comparisons (10 within-network metrics + 10 one-versus-all-others-network metrics + 45 pairwise between-network metrics = 65 comparisons) by controlling false discovery rate. Post hoc analysis was then performed to determine the significance of specific comparisons with network-based statistics (NBS) among groups (p < 0.01, FDR correction).
Additionally, the CCA was used to identify relationships between brain network connectivity measures and clinical phenotypes, CSF biomarkers, lipid related genetic variants, and lipoproteins in the serum of AD spectrum patients. Given a significant CCA mode, Pearson’s correlation was used to assess the correlation between the CCA mode and the corresponding set of original variables of which it consisted. Finally, the correlation coefficients were visualized using the radar plots in Fig.
5. Details on CCA were described in Additional file
1.
Support vector machine classification
SVM was used in this study to classify AD spectrum in MATLAB based on a library (LIBSVM) [
37]. The LIBSVM classifier algorithm was applied within Leave-one-out cross-validation (LOOCV). Grid search method and Gaussian radial basis function (RBF) kernels were used for parameter optimization. Post hoc analysis revealed nineteen lipoproteins and three gene scores were associated with network connectivity. Then, we performed Pearson correlation to find the functional connections which were correlated with all nineteen lipoproteins and all three gene scores. P values of correlation coefficient < 0.05 was considered statistically significant. Those functional connection were used in the classification by SVM. In order to quantify the performance of the final machine learning model, the accuracy, sensitivity, specificity, and area under the curve (AUC) were calculated to reduce the impact of deviations in the distribution of the training and testing sets. In addition, the accuracy (ACC) of testing set was assessed by permutation test with 1,000 epochs as described in previous studies [
38].
Path analysis
We further used SEM to examine the relationship among variables in radar plots (Additional file
1: Fig. S4). All variables in the radar plots were observed variables. Moreover, we constructed four variables (dyslipidemia, pathology, brain function, and cognition) as latent variables. Hypothesized relationships were constructed among variables based on the results of post hoc analysis. The causal path relationship of the 5 latent variables constituted the SEM structural model, and the relationship between latent variables and their corresponding observed variables constituted the SEM measurement model. SEM was conducted using IBM SPSS Amos version 22 statistical software (Amos Development Co., Armonk, NY, USA). For the hypothesized relationships, t-tests and path coefficients were determined using a bootstrapping approach with a sampling of 5000. The goodness of fit was assessed by chi square/degree of freedom ratio (CMIN/DF), root-mean-square error of approximation (RMSEA), goodness-of-fit index (GFI), adjusted GFI (AGFI), Tucker-Lewis Index (TLI), normed fit index (NFI), comparative fit index (CFI), and incremental (IFI). The significance level was set 0.05 in this study.
Discussion
This is the first study focusing on the potential of lipid-related genes and proteins to influence the dynamic trajectory of large-scale RSNs, CSF biomarkers, and cognitive decline in ADS patients using a CCA approach. The present study shed mechanistic light on the role of lipid metabolites disturbance in promoting large-scale RSNs disruption and accelerating CSF biomarker deposition and subsequently caused cognitive decline in ADS individuals. These findings provided novel insight for uncovering the neural link between lipid metabolites and cognitive decline at a large-scale network level and expanding our understanding of the mechanisms underlying AD pathophysiology.
Although it is not well-established that potential relationships between lipid metabolites and AD exist, the majority of studies have reported that abnormal lipid metabolites apparently increased the risk of cognitive decline and substantially contribute to the development and progression of AD [
3,
7,
40,
41]. Recently, a meta-analysis reported that high midlife total serum cholesterol significantly increases the risk of late-life AD, and may correlate with the onset of AD pathology [
42]. A prospective study with a large-cohort sample in which 22,623 participants were recruited established that the concentration of cholesterol esters relative to total lipids in large HDL and the total cholesterol to total lipid ratio in very large VLDL significantly increased the risk of incidence of dementia [
7]. We also found that lipid metabolites, including genes and lipoproteins were markedly associated with CSF biomarkers and cognitive impairment, also supporting the hypothesis that lipid metabolic dysfunction substantially contributes to AD pathophysiology via interference through progressive neuropathological changes of CSF biomarkers and declining cognitive function across the entire ADS.
Disrupted network integrity, including abnormal structural and functional network connectivity, was preferentially targeted by specific genetic variants or molecular pathology in preclinical AD, or mapped the clinical phenotype with disease progression and supported the recent description of the theoretical framework and empirical evidence of AD [
24,
43]. As such, brain network integrity emerged as potential intermediate biomarkers to bridge upstream determinants (gene and molecular pathology) and downstream effects (clinical phenotypes) [
23,
24]. However, the complicated association that the dynamic spatiotemporal patterns of brain network integrity linking molecular pathology and cognitive decline in ADS individuals remains largely unclear. According to cascading network failure theory, distinct DMN subsystems representing differential spatiotemporal evolution correspond with the AD pathophysiological response, and differentially affected by AD pathological biomarkers including Aβ deposition and tau tangles, subsequently leading to stereotypic network-based cognitive decline in ADS patients [
18,
22]. This study firstly described the progressive changes in spatiotemporal network patterns within the DMN system in ADS patients. We then further identified changes in dynamic trajectory within- and between networks reflected by the active capability of network inner cohesion and connectors beyond the DMN across the entire ADS. More attention should be focused on whether such changes in dynamic trajectory in large-scale RSNs are cascading failure or not. In contrast, a proportion of the networks represented enhanced network inner cohesion or exhibited network connector roles as the disease progressed. Compelling evidence has been reported that a gradual decrease in connectivity of RSNs is associated with amyloid deposition that accelerates disease progression, while the commonly observed increase in connectivity of RSNs also found in preclinical and prodromal AD patients has been interpreted as a compensatory phenomenon supporting better performance on cognitive tasks [
44,
45]. However, this enhanced network connectivity is the consequence of transient compensation to network disruption or an adaptive response to AD pathophysiological processes that still require identification through additional study.
Importantly, the dynamic changes in large-scale networks over the course of disease that were also observed were significantly affected by lipid-related genetic variants and lipoproteins, CSF biomarkers, and cognitive function, confirming the biological nature of the predictable correlation with network connectivity by linking upstream molecular pathology and downstream clinical phenotype to the preclinical stage of AD. Furthermore, post hoc analysis was performed to trace the source of these system-level correlations and identified that distinctive connectivity within- and between networks was specifically related to the effects of accumulated lipid-related genetic variants or lipoproteins, neuropathological biomarkers, in addition to cognitive decline. Due to changes in lipids often prior to molecular pathology and cognitive decline, we hypothesized that compromised large-scale networks and CSF biomarkers may mediate the effects of lipid metabolites on cognitive decline with progression of AD. Previously, structural atrophy or functional decoupling of RSNs were, at least partly, ascribed to abnormalities in lipid metabolites which suggested that lipid metabolites may be a vulnerable molecular substrate of large-scale RSNs [
28,
34]. More importantly, disturbed lipid metabolites and dynamic brain network changes occurred prior to measurable amyloid deposition and tau tangles related to ageing [
18], while lipid pathway genetic variants, including APOEε4 genotype and lipoproteins markedly enhanced the disruption of brain network architecture in preclinical AD patients [
34,
46] and even in the cognitively normal elderly [
47,
48]. In addition, carriers of the APOEε4 allele, the strongest risk factor for sporadic late-onset AD, represented a specific phenotype in which the relationship with brain networks preceded any measurable systems or molecular level changes in cognitively normal subjects [
49‐
51]. Furthermore, cholesterol-related genetic risk scores were associated with hypometabolism in AD-affected brain regions, even when controlling for the effects of APOE ε4 gene dose [
52].
In this regard, we putatively identified that the dynamic trajectory changes of large-scale networks observed in this study may be induced because of a lipid-driven pathological interaction with Aβ abnormal deposition and tau-related neurofibrillary tangles and then promoted cognitive decline to dementia. From the path analysis, we found that dyslipidemia directly influenced brain function network reorganization leading to cognitive impairment or indirectly affected the CSF biomarkers levels and subsequently caused cognitive decline.
Another interesting finding of the present study was the SVM classifier model. This SVM classifier achieved a relatively high performance and implies that a significantly important role of lipid metabolism in the onset and neuropathology of AD. Lipid associated neuroimaging biomarkers would serve as a good potential biomarker for ADS diagnoses and an invention target to early prevent AD incidence.
Several limitations should be noted. Firstly, the 38 lipid-related genetic variants and lipoproteins selected in this multimodal cross-sectional study may underestimate the potential of lipid metabolites for the early detection and diagnosis of AD. Lipidomic approaches should be considered in order to explore the pathogenesis of AD, because this provides a new tool to investigate the association between blood-based genetic variants or changes in lipoproteins in the serum or plasma and the pathological mechanisms of CNS disorders. Secondly, longitudinal studies should be performed to explore the potential biomarkers of lipid metabolites in AD pathophysiology, validate the neural links between changes in lipids and neuropathology, and determine the causal contributions of lipid metabolite disturbance and disrupted network integrity, in addition to cognitive decline in ADS patients.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.