APOE-ε4 Shapes the Cerebral Organization in Cognitively Intact Individuals as Reflected by Structural Gray Matter Networks.

Gray matter networks (GMn) provide essential information on the intrinsic organization of the brain and appear to be disrupted in Alzheimer's disease (AD). Apolipoprotein E (APOE)-ε4 represents the major genetic risk factor for AD, yet the association between APOE-ε4 and GMn has remained unexplored. Here, we determine the impact of APOE-ε4 on GMn in a large sample of cognitively unimpaired individuals, which was enriched for the genetic risk of AD. We used independent component analysis to retrieve sources of structural covariance and analyzed APOE group differences within and between networks. Analyses were repeated in a subsample of amyloid-negative subjects. Compared with noncarriers and heterozygotes, APOE-ε4 homozygotes showed increased covariance in one network including primarily right-lateralized, parietal, inferior frontal, as well as inferior and middle temporal regions, which mirrored the formerly described AD-signature. This result was confirmed in a subsample of amyloid-negative individuals. APOE-ε4 carriers showed reduced covariance between two networks encompassing frontal and temporal regions, which constitute preferential target of amyloid deposition. Our data indicate that, in asymptomatic individuals, APOE-ε4 shapes the cerebral organization in a way that recapitulates focal morphometric alterations observed in AD patients, even in absence of amyloid pathology. This suggests that structural vulnerability in neuronal networks associated with APOE-ε4 may be an early event in AD pathogenesis, possibly upstream of amyloid deposition.

Introduction

Alzheimer’s disease (AD) is characterized by a distinctive pattern of brain atrophy along with progressive decline in multiple cognitive domains (Braskie and Thompson, 2013). Core pathological hallmarks of AD are cerebral extracellular deposition of beta-amyloid (Aβ) fibrils and intraneuronal aggregation of hyperphosphorylated tau protein (Frisoni et al., 2017). Abnormally low cerebral glucose metabolism, as assessed with 2-[18F]fluoro-2-deoxy-D-glucose (FDG) positron emission tomography (PET), is also present across several regions including temporo-parietal and posterior cingulate cortices (Alexander et al., 2002; Mosconi, 2013). Yet, preclinical signatures of the disease in asymptomatic individuals are already evident at least two decades prior to cognitive impairment, opening a temporal window, which is critical to delay the onset of dementia (Sperling et al., 2011). The preclinical course of AD is influenced by multiple genetic markers (Escott-Price et al., 2015; Tan et al., 2017). Among those, apolipoprotein E (APOE)-ε4 represents the greatest genetic risk factor for AD, by lowering the age of onset in a dose-dependent manner (Corder et al., 1993; Farrer et al., 1997). The ε4 allele codes the E4 isoform of the APOE protein, which is associated to a deficient neuronal cholesterol delivery, as well as poor clearance of extracellular Aβ, therefore facilitating the formation of fibrillary plaques (Liu et al., 2013). In cognitively unimpaired individuals, APOE-ε4 has been related to faster rates of cerebral Aβ accumulation (Reiman et al., 2009; Jansen et al., 2015), as well as lower brain glucose metabolism (Reiman et al., 2005). These effects appear to be dose-dependent, that is, proportional to the number of APOE-ε4 alleles (see Fouquet et al., 2014 for a review). With respect to brain structure, some previous studies have reported significantly different gray matter volumes in healthy ε4-carriers compared with noncarriers (NC) in regions typically associated to AD degeneration, such as the hippocampus, frontal, and temporal areas (Honea et al., 2009; Alexander et al., 2012; Ten Kate et al., 2016). We have recently reported dose-dependent effects of the risk allele on posterior hippocampal volume and on other areas in a sample of cognitively unimpaired individuals that was genetically enriched for the presence of APOE-ε4 homozygotes (Cacciaglia et al., 2018a). However, the adoption of a univariate statistical approach prevented us to investigate the role of APOE-ε4 on the structural organization of the brain.

The adoption of multivariate methods for the study of cerebral anatomy allows the detection of gray matter networks (GMn), which are expression of the coordinated volumetric variability across several brain areas. Unlike univariate voxel-based or region of interest analyses, this approach provides essential information on the intrinsic organization of cerebral morphology and it may thus capture subtle differences not yet observable with traditional approaches (Alexander-Bloch et al., 2013; Guo et al., 2014). GMn have been studied in relation to cognitive abilities in healthy people (Eckert et al., 2010; Yoon et al., 2017), as well as in psychiatric (Xu et al., 2009; Kasparek et al., 2010) and neurologic (Steenwijk et al., 2016) disorders. For example, AD patients show significantly reduced structural covariance within the regions of the default mode network (DMN) (Seeley et al., 2009; Spreng and Turner, 2013). Yet, most of the studies investigating GMn have relied on the selection of a seed region to identify common cerebral networks, such as the DMN, salience, or the executive control (Evans, 2013). While this approach has the advantage of studying brain systems that are known to regulate cognitive functions, it prevents from uncovering additional hidden sources of variance, which may be subject to preclinical alterations in asymptomatic individuals.

Therefore, the goal of the present study was to investigate GMn as a function of the APOE genotype using a data-driven, spatially unbiased approach, to provide further insights on the structural vulnerability associated to the genetic risk for AD. We assessed both within- and between-networks structural covariance, capitalizing on a large sample (n = 533) of cognitively unimpaired individuals, which was enriched for the genetic risk of AD by hosting an unprecedented number of APOE-ε4 homozygotes for a single-site imaging study (n = 65). This sample characteristic allowed us to test different models of genetic penetrance enabling to pinpoint distinctive cerebral features as a function of the allelic load. In addition, to better characterize GMn, we determined their association with age and cognitive performance in the entire sample. In order to rule out the influence of amyloid pathology in our findings, we repeated the analyses on a subsample of participants that was amyloid-negative.

Materials and Methods

Study Participants

Participants were enrolled to the ALFA (ALzheimer and FAmilies) study (Clinicaltrials.gov Identifier: NCT01835717), a research platform aiming at the identification of pathophysiological alterations in preclinical AD. The ALFA cohort entangles 2743 cognitively unimpaired individuals, with a Clinical Dementia Rate score of 0, most of them being first-order descendants of AD patients (Molinuevo et al., 2016). Subjects with a psychiatric diagnosis were excluded from the study. Additional exclusion criteria were described in detail previously (Molinuevo et al., 2016). After APOE genotyping, which is described in detail in the Supplementary Material, all participants homozygous for the ε4 allele as well as carriers of the ε2 allele were invited to undergo magnetic resonance imaging (MRI) scanning along with ε4-heterozygotes (ε4HET) and NC matched for age and sex. This recruitment strategy resulted in 576 study participants, out of which 43 had to be discarded due to MRI incidental findings or poor image quality, leading to a final sample of 533 individuals. For the statistical analyses, participants were pooled according to the cumulative presence of the ε4 allele, that is, NC, ε4HET and ε4-homozygotes (ε4HMZ). The study was approved by the local ethics committee and conducted according to the principles expressed in the Declaration of Helsinki, with all subjects providing written informed consent.

Image Data Acquisition and Preprocessing

MRI was conducted with the same MR equipment and in a single site for each study participant (3 T General Electric scanner, GE Discovery MR750 W). Structural 3D high-resolution T1-weighted images were collected using a fast spoiled gradient-echo sequence with the following parameters: voxel size = 1 mm3 isotropic, repetition time (TR) = 6.16 ms, echo time (TE) = 2.33 ms, inversion time (TI) = 450 ms, matrix size = 256 × 256 × 174, and flip angle = 12°. The new segment function implemented in Statistical Parametrical Mapping software (SPM 12, Wellcome Department of Imaging Neuroscience, London, UK) was used to segment gray matter in native space and locate all images into a common reference space for subsequent normalization, using a 12-affine parameter transformation. Segmented GM images were used to generate a reference template of the sample, which was warped into a standard Montreal Neurological Institute (MNI) space using the high-dimensional DARTEL toolbox (Ashburner, 2007). The generated flow fields and normalization parameters were then implemented to normalize the native GM images to the MNI space. In order to preserve the native local amount of GM volume, we applied a modulation step, where each voxel signal’s intensity was multiplied by the Jacobian determinant derived from the normalization procedure (Good et al., 2001). Quality control of normalization was assured by checking the sample homogeneity with the computational anatomy toolbox (CAT12) (http://dbm.neuro.uni-jena.de/cat/) using nonsmoothed data, which did not return errors in the registration procedure in any subject. Finally, images were spatially smoothed with a 6-mm full-width at half maximum Gaussian kernel. Total intracranial volume (TIV) was computed by summing the segmented GM, white matter (WM), and cerebrospinal fluid (CSF) volumes for each individual.

Assessment of the Gray Matter Networks

After preprocessing, GMn were assessed from normalized and smoothed images using source-based morphometry (SBM) (Xu et al., 2009), implemented in the GIFT software (http://mialab.mrn.org/software/gift/index.html). SBM uses independent component analysis (ICA) to decompose signal intensity across the images into sources of common variance. This procedure uses variance in the data to detect brain regions that covary in terms of gray matter volume, and therefore represent a structural network. We used the Infomax neural network algorithm, which iteratively minimizes the mutual information of the network outputs to identify naturally grouping and maximally independent sources (Bell and Sejnowski, 1995). The number of independent components was estimated data-driven on the basis of the computation of the correspondent latent principal components. This was achieved using the minimum description length criteria (Li et al., 2007). To increase component reliability and consistency, we used the ICASSO algorithm (Himberg et al., 2004) and repeated ICA 250 times with bootstrapping and random initialization. Each individual gray matter image was converted into a 1D vector and arrayed into a matrix of number of subject by each voxel’s intensity. This group image matrix was then decomposed into a mixing matrix, representing the weights of each component for each subject, and a source matrix, representing the relationship between each component and brain voxels. The mixing matrix contained the loadings for each component, which were submitted to group statistical analyses. After components estimation, we removed those with poor clustering as revealed by a quality index Iq < 0.8 indicating the difference between intracluster and extracluster similarity (Himberg et al., 2004). Remaining components were visually inspected to select those of biological relevance. In so doing, we removed components containing sharp edges near to the brain parenchyma, or primarily appearing in nongray matter areas (i.e., WM or ventricles), as previously recommended (Xu et al., 2009).

Neuropsychological Evaluation

We evaluated episodic memory (EM) and processing speed (PS), two cognitive domains that have been included in the preclinical Alzheimer cognitive composite (Donohue et al., 2014) and for which we have already delineated the underlying cerebral morphology in a subsample of the present study (Cacciaglia et al., 2018b, 2019). EM was assessed with the Spanish adapted version of the memory binding test (Gramunt et al., 2016), an instrument that was developed for the detection of subtle impairments in the healthy population (Buschke, 2014). A description of the administration procedure is available at the online Supplementary Material. PS was assessed with the coding subtest of the Wechsler Adult Intelligence Scale-Fourth Edition (Wechsler, 2012), where participants are given keys that match numeric digit spanning from 1 to 9 with a symbol. The task is to write down the correct symbol aside a list of numbers as quickly as possible.

CSF Sampling and Analysis

CSF was sampled on average 4.11 years after MRI data acquisition. Fresh CSF samples were collected in 15 mL polypropylene tubes (Sarstedt catalog #62.554), the supernatant aliquoted into 0.5 mL polypropylene tubes (Sarstedt catalog #72.730.005), and frozen within 2 h after lumbar puncture. Aliquots were placed into long-term storage boxes and stored at −80 °C until shipment on dry ice for analysis. CSF samples were measured using the Elecsys β-amyloid (1–42) (Bittner et al., 2016), and the Elecsys phosphotau (181P) and Elecsys total-tau immunoassays for CSF on a cobas e 601 analyzer (software version 05.02) at the Clinical Neurochemistry Laboratory, University of Gothenburg, Sweden. Individuals with CSF Ab42 values below 1098 pg/mL were categorized as negative (Schindler et al., 2018).

Statistical Analyses

The impact of APOE-ε4 on GMn was determined separately for within- and between-networks structural covariance. In the former case, we conducted an analysis of covariance (ANCOVA) with each component’s factor loadings modeled as dependent variable, while including age, sex, and TIV as covariates. For the between-networks analysis, we repeated the same ANCOVA and additionally modeled, for any structural component as dependent variable, the interaction between APOE-ε4 and any other component. This interaction term captures the modulatory effect of APOE-ε4 on the structural coupling between any two components. In all models, APOE-ε4 group membership was entered as independent fixed factor. To this respect, we tested three different genetic models of penetrance: additive, dominant, and recessive. Briefly, an additive model predicts an incremental response of the quantitative trait according to the allelic load, whereas a dominant model predicts a common response to one or two copy of the risk allele (i.e., ε4-carriers vs. NC). Finally, a recessive model predicts a common response to zero or one copy of the risk allele (i.e., NC and ε4HET vs. ε4HMZ) (Clarke et al., 2011). To assess the linear association between age and GMn, two-tailed Pearson’s correlation was performed between each structural component’s factor loading and age. Lastly, to detect the association between GMn and cognitive performance, two-tailed partial correlation controlling for age and sex was performed between each IC factor loading and cognitive scores in both EM and PS. All the results were considered significant at P < 0.05, after correction for multiple testing using false discovery rate (FDR) (Benjamini and Hochberg, 1995). Analyses were repeated in the amyloid-negative subsample.

Results

Participants’ Demographics and Cognitive Data

Table 1 summarizes demographic information stratified across APOE-ε4 subgroups along with cognitive scores. NC, ε4HET, and ε4HMZ did not significantly differ in years of education, proportion of males/females, or TIV. Similarly, there were no significant differences in any cognitive variable after adjusting for covariates. However, ε4HMZ were significantly younger than both NC and ε4HET. For this reason, age was included as covariate in all subsequent analyses. The range of the assessed demographic and cognitive variables is provided in the Supplementary Material (see Supplementary Table 1).

Table 1

Sample characteristics

	Total sample (n = 533)	NC (n = 261)	ε4HET (n = 207)	ε4HMZ (n = 65)	P-value
Age (y)	57.58 (7.43)	57.93 (7.53)	58.22 (7.41)	54.14 (6.18)	<0.01
Sex (F/M)	320/213	166/95	113/94	41/24	0.13
Education (y)	13.64 (3.56)	13.69 (3.62)	13.66 (3.53)	13.38 (3.46)	0.82
TIV (cm3)	1490.11 (146.07)	1484.96 (155.17)	1495.41 (137.73)	1493.91 (135.20)	0.72
TPR	24.18 (4.46)	23.91 (4.81)	24.23 (4.20)	25.14 (3.69)	0.21a
TFR	16.54 (5.17)	16.45 (5.22)	16.29 (5.08)	17.66 (5.15)	0.60a
TDPR	23.90 (4.60)	23.65 (4.87)	23.94 (4.40)	24.82 (3.98)	0.27a
TDFR	16.89 (5.18)	16.58 (5.35)	16.90 (4.78)	18.11 (5.62)	0.29a
Coding	65.50 (15.20)	65.67 (15.83)	64.26 (14.85)	68.78 (13.28)	0.63a

Note: NC: noncarriers; ε4HET: ε4-heterozygotes; ε4HET: ε4-homozygotes; TIV: total intracranial volume; M: mean; SD: standard deviation; TPR: total paired recall; TFR: total free recall; TDPR: total delayed paired recall; and TDFR: total delayed free recall. Except for the categorical variable sex, numbers represent mean and SD.

Corrected for age, sex, and years of education.

APOE-ε4 and Gray Matter Networks: Within Network Analysis

Our ICA retrieved 45 ICs, which were reduced to 16 biologically relevant GMn of interest, upon applying the criteria described above. Figure 1 shows surface rendering of four exemplary GMn projected onto a mesh template, along with a chord diagram exemplifying the correlations among all networks in the entire sample.

Figure 1

Structural GMn. (A) Surface rendering of four exemplary GMn identified in our sample with structural ICA. In each network, warm colors (red–yellow) indicate areas that are significantly related among each other in terms of GM volume. Cold colors (green–blue) denote the anticorrelation network, that is, brain regions that display a negative correlation with the areas in the positive correlation network. (B) Chord diagram illustrating the correlation matrix among all the 16 structural networks. Each sector along the circle, measured by arbitrary bin units in canvas coordinates, represents one structural network, with the length denoting the total amount of connections of each network. The thickness of inner ribbons indicates the strength of the correlation, while the sign is encoded by the color. (C) Volume rendering of the 16 structural components projected over axial slices, at their respective global maximum standardized coordinate.

The genotypic recessive model yielded a significant main effect of APOE-ε4 on IC29, where ε4HMZ showed higher structural covariance compared with both NC and ε4HET (F4,533 = 8.71; P = 0.003; and PFDR = 0.044) (Fig. 2). This network (IC29) consisted of prominently right-lateralized regions including parietal, inferior frontal, as well as inferior and middle temporal regions (Table 2). In addition, the left angular gyrus and precuneus, as well as the left middle and inferior temporal cortex were included (Table 2). All of these regions were previously included within a composite region of interest known as AD-signature, which undergoes focal atrophy early in the course of AD (Bakkour et al., 2009, 2013; Dickerson et al., 2009, 2011). Figure 2 shows the spatial overlay between IC29 and a reconstruction of the AD-signature, which was performed on the basis of the regions provided in Dickerson et al. (2009). The additive model retrieved a nominally significant difference in the same component (IC29), which however did not survive correction for multiple testing (F4,533 = 4.35; P = 0.013; and PFDR = 0.239). No other ICs showed significantly different values among the three genotype groups.

Figure 2

APOE-ε4 homozygotes showed altered structural brain covariance. (A) Volume rendering of IC29 overlaid on the AD-signature composite region of interest, projected over sagittal slides. Brown areas indicate the spatial overlay between the two rendered volumes, which was evident in the inferior parietal, frontal as well as middle and inferior temporal areas and the precuneus. (B–D) Boxplot charts showing significantly heightened structural covariance of IC29 in APOE-ε4 homozygotes compared with both NC and heterozygotes. Τhe width of each box indicate the interquartile range, with the horizontal line showing the median. Data from each individual are shown in a separate vertical column. NC: noncarriers; ε4HET: APOE-ε4 heterozygotes; and ε4HMZ: APOE-ε4 homozygotes.

Table 2

Brain regions included in IC29

Brain region	Peak intensity	Laterality	k	Peak MNI coordinates
				x	y	z
Cerebellar Crus 2	8.48	L	8022	-42	−51	−34
Precuneus	6.13	R	12 126	18	−60	19
Angular gyrus	4.93	R	3132	43	−42	39
Superior parietal	3.87	R	680	21	−48	60
Temporal pole	3.41	R	3921	46	9	−36
Superior frontal	3.31	L	1314	−13	0	66
Middle frontal	3.23	R	409	30	42	21
Superior parietal	3.01	L	211	−19	−70	55
Middle frontal	2.96	L	141	−24	4.5	49
Middle temporal	2.78	L	757	−42	−61	13
Inferior temporal	2.73	L	106	−55	−31	−18
Precuneus	2.65	L	220	−9	−60	45
Middle temporal	2.48	R	296	54	−51	0
Orbitofrontal	2.47	R	351	7.5	40	−4
Inf. front. (Tri)	2.41	R	1033	49	22	4
Middle frontal	2.38	R	267	36	9	43
Inf. front. (Orb)	2.07	L	103	−27	16	−25

Note: k: cluster size, indicates the number of contiguous voxels within each cluster. MNI: Montreal Neurological Institute.

APOE-ε4 and Gray Matter Networks: Between Network Analysis

Of the tested genotypic models, the dominant one yielded a significant effect of APOE-ε4 (F1,533 = 8.23; P = 0.004; and PFDR = 0.048) in modulating the linear association between IC10 and IC15 (Fig. 3). IC10 included both medial and lateral aspects of the superior frontal cortex bilaterally, with the inclusion of the middle cingulate gyrus, while IC15 entangled the frontal poles as well as the bilateral middle temporal gyrus (Fig. 3). Additionally, a nominally significant effect was observed under the same genotypic model and in the same direction for IC02 and IC17 (Fig. 3), which however did not survive correction for multiple testing (F1,533 = 6.84; P = 0.009; and PFDR = 0.094). Overall, these results indicate that, when compared with NC, both ε4HET and ε4HMZ displayed reduced structural covariance between these couples of networks.

Figure 3

APOE-ε4 carriers display reduced connectivity between two pairs of structural networks. (A) Chord diagrams computed for each APOE subgroups highlighting the two network couplings (IC10/IC15 and IC02/IC17), which were affected by the APOE-ε4 genotype. The length of each outer sector is measured by arbitrary bin units in canvas coordinates and encodes the total amount of connections of a given network. The thickness inside each chord diagram encodes the strength of correlation. NC: noncarriers; ε4HET: APOE-ε4 heterozygotes; and ε4HMZ: APOE-ε4 homozygotes. (B) Surface rendering of the binarized IC10 and IC15, where an FDR-surviving significant effect of APOE-ε4 was found. (C) Group scatterplots with marginal densities reported aside each axes, showing the modulatory effect of APOE-ε4 on the linear association between two pairs of component. Shaded areas around the fitted regression lines indicated 90% confidence interval.

Complementary Analysis in CSF Aβ42 Negative Individuals

To determine to what extent the observed effects were specific to APOE genetic variance, we reanalyzed data from a subsample of individuals who were classified as negative to CSF Aβ42 in a retrospective analysis. While this approach is blind to the actual Aβ42 values when MRI data were acquired, it allows to identify those who did not have amyloid pathology at that time. Of the entire sample (n = 533), CSF was available for 121 individuals. Of those, 71 were Aβ-negative (see Supplementary Table 2 for a sample description). This analysis showed that, in the subsample with available CSF, ε4HMZ continued to have significantly greater structural covariance in IC29 (F3,121 = 16.82; P = 0.001) (Fig. 2), and, most importantly, this was confirmed when we restricted the analysis to those with negative CSF Aβ42 (F3,71 = 7.18; P = 0.009) (Fig. 2), indicating that these effects were independent on amyloid. Conversely, the modulatory role of APOE-ε4 on the structural coupling between IC10 and IC15 observed in the entire sample was not confirmed when inspected in amyloid negative subjects (F5,71 = 0.15; P = 0.71), suggesting that this effect may be related to the presence of amyloid.

Gray Matter Networks and Their Association with Cognitive Performance

We found significantly negative linear association between all GMn and age, with the exception of IC20, which entangled the bilateral superior parietal cortex. Partial correlations did not reveal significant relationships between any GMn and EM. However, for PS, we observed significantly positive associations in four networks as well as significantly negative associations in two networks (Table 3).

Table 3

Linear correlations between GMn and PS in the entire sample

	Component	r-value	P	P FWE
Positive correlations
	IC02	0.13	0.004	0.02
	IC13	0.14	0.001	0.02
	IC15	0.15	0.001	0.01
	IC16	0.12	0.001	0.01
Negative correlations
	IC29	−0.12	0.006	0.02
	IC30	−0.12	0.002	0.02

Discussion

We have reported that APOE-ε4 risk variant for AD exerts a significant impact on intrinsic GMn in cognitively unimpaired middle-aged individuals.

Our approach of retrieving naturally grouping sources of structural covariance within a sample hosting an unprecedented number of ε4HMZ (n = 65) allowed us to unmask a latent structural vulnerability imaging phenotype associated to this risk variant. Within-network analysis revealed that cognitively intact ε4HMZ had increased structural covariance in one network (IC29), which included prominently right lateralized parietal, inferior frontal, as well as inferior and middle temporal regions. In addition, this network entangled the left angular gyrus, precuneus, as well as the left middle and inferior temporal cortex. IC29 exhibited the least degree of covariance with the rest of networks, indicating that its variability was not related to the topological patterns in the rest of the brain. Moreover, it was negatively associated to PS performance across the entire sample. Overall, this suggests that IC29 captures an intrinsic structural organization of the brain that expresses a pathological process. In support of this, we have observed that nearly all the regions included in such a network are part of a previously described composite region of interest, which undergoes focal atrophy early in the course of AD, the so-called “AD-signature” (Bakkour et al., 2009, 2013; Dickerson et al., 2009, 2011). Other research groups detected selective cortical thinning in the same areas of the AD-signature both in AD and MCI patients (Singh et al., 2006). These brain regions are also typically subject to hypometabolism in preclinical (Besson et al., 2015) and full-blown AD dementia (Alexander et al., 2002; Mosconi, 2013), as well as in cognitively unimpaired APOE-ε4 carriers compared with NC (Reiman et al., 2005; Mosconi et al., 2008). It is worth noting that the temporal and parietal areas included in IC29 have been previously described as “cortical hubs,” that is, heteromodal regions interconnecting spatially distributed and functionally specialized brain systems, which, possibly because of their higher metabolic consumption rates, are especially vulnerable to the AD neurodegenerative process (Buckner et al., 2009; Sepulcre et al., 2012). Finally, IC29 overlaps with the anterior subdivision of the DMN (Xu et al., 2016; Staffaroni et al., 2018), which is known for being vulnerable in AD (Buckner et al., 2005; Seeley et al., 2009).

Since its definition, the AD-signature has been validated as effective diagnostic and prognostic marker. Some studies demonstrated that AD-signature cortical thinning outperforms hippocampal volume in predicting conversion from MCI to AD (Dickerson et al., 2011; Dickerson and Wolk, 2013). Dickerson and Wolk (2013) reported the superior sensitivity of AD-signature cortical thinning over both amyloid and tau fluid biomarkers in predicting AD dementia onset within 1 year from baseline assessment. One study demonstrated that cortical thinning in these regions was related to suprathreshold CSF tau levels in cognitively unimpaired individuals (Wang et al., 2015). In the same study, the authors reported that AD-signature cortical thinning was negatively related to visuospatial processing (Wang et al., 2015), which is consistent with the negative association we observed between IC29 factor loadings and PS. Notably, even though the AD-signature was initially detected on the basis of cortical thickness data, the same regions were retrieved when comparing AD patients with asymptomatic individuals adopting a voxel-wise method, a technique similar to the one used in the present study (Schwarz et al., 2016). In line with this, exploratory studies employing voxel-based morphometry found that, among MCI patients, only those who later progressed to AD dementia showed, compared with healthy controls, volume reductions in AD-signature areas (Chételat et al., 2005; Bozzali et al., 2006; Whitwell et al., 2008). In our previous study conducted on the same sample as the one of the present report, we did not find volume difference among APOE-ε4 subgroups in these regions (Cacciaglia et al., 2018a). Hence, our data suggest that an increased structural covariance within the AD-signature areas in individuals exposed to exceptionally high risk for AD, such as APOE-ε4 homozygotes, may predate cortical thinning as well as hypometabolism in these regions. Moreover, an exaggerated connectivity among these brain areas may facilitate the patterned spreading of AD proteinopathies within this network, which is well-known for being particularly vulnerable to AD pathology (Zhou et al., 2012; Ossenkoppele et al., 2019).

Importantly, we replicated this finding in a subsample of individuals who had no signs of incipient amyloid accumulation as revealed by negative CSF Aβ42 biomarker, indicating that it was specifically due to APOE genetic variance. This strongly suggests that ε4HMZ harbor a structural brain vulnerability that precedes amyloidosis and supports earlier findings of APOE-ε4 related effects on neurodevelopment (Shaw et al., 2007; Wolf et al., 2013; Chang et al., 2016). In line with this, some studies have observed AD-signature cortical thinning in association with CSF tau levels in asymptomatic, amyloid negative, and subjects (Wang et al., 2015). In addition, APOE-ε4 has been found to have an impact on temporoparietal FDG uptake in cognitively healthy individuals who were Aβ negative (Jagust et al., 2012), suggesting that its effects on neurodegeneration markers may occur earlier or independently on amyloid.

Nevertheless, the cross-sectional nature of our study prevents us to draw this conclusion and our data foster upcoming studies to assess GMn in longitudinal settings.

It is worth noting that the relatively high prevalence of amyloid positive individuals in the subset with available CSF data (n = 171, 29% with positive CSF Aβ42) is not surprising, given that our sample was enriched for the genetic risk of AD, according to the selection criteria employed in the ALFA study (Molinuevo et al., 2016).

We should note that IC29 was predominantly right-lateralized, with relatively few and regions appearing in the left hemisphere. Our data, however, indicate that the magnitude of this component was associated with the presence of ε4HMZ. The relative poor representation of this subgroup in the entire sample (65 out of 533 individuals) may have caused some loss of signal, resulting in a spatial inhomogeneity. It is possible that a more balanced sample with an equal number of NC and heterozygotes as homozygotes would lead to a bilateral distribution. On the other hand, as noted above, the spatial topology of IC29 overlaps with the anterior subdivision of DMN (Xu et al., 2016; Staffaroni et al., 2018). This is in general not surprising given that networks of structural and functional covariance share the same topology (Alexander-Bloch et al., 2018, Smith et al., 2019). In cognitively healthy individuals, the DMN was found to be mostly left-lateralized with aging shifting this lateralization to the right side (Agcaoglu et al., 2015). Interestingly, age-related right-lateralization of the anterior DMN predicts poorer cognitive outcomes (Banks et al., 2018), supporting our data on the negative relationship between IC29 and PS. Thus, the predominantly right-lateralized topology of IC29 may represent an additional indicator of an aging-related pathological process, encapsulated by this component.

When performing the between-networks analysis, we found that compared with NC, APOE-ε4 carriers had significantly decreased structural covariance between IC10 and IC15. The former component included medial and lateral aspects of the superior frontal cortex bilaterally, with the inclusion of the middle cingulate gyrus, while the latter entangled the frontal poles as well as the anterior cingulate cortex and bilateral middle temporal gyrus (Fig. 3). All of these regions constitute preferential target of early Aβ deposition (Klunk et al., 2004; Myers et al., 2014; Grothe et al., 2017). Notably, unlike for the previous finding on the increased covariance within IC29, we could not replicate this result when restricting the analysis to amyloid-negative subjects. This might be due to the reduced statistical power given the lower sample size or be associated to a putative role of amyloid. In support of the latter, earlier studies have shown that cortical Aβ deposition impairs synaptic transmission and consequently reduces the local functional connectivity between the affected circuits (Lacor et al., 2007; Palop and Mucke, 2010). Thus, it is possible that incipient cortical Aβ deposition associated with the presence of APOE-ε4 allele impairs the structural covariance between these two networks in asymptomatic individuals. This result may represent an AD intermediate phenotype in a way that it reflects a network disconnection, in line with the “disconnection hypothesis” of AD (Delbeuck et al., 2003; Reid and Evans, 2013; Sepulcre et al., 2013), possibly due to impairment in synaptic communication. Finally, it is worth noting that the reduced covariance between IC10 and IC15 was significant under a genotypic dominance assumption, indicating that the effect was common to the heterozygote and homozygote groups, compared with NC. In other words, this neuroimaging phenotype was detectable already in a subgroup, the ε4HET, which is moderately exposed to the genetic risk of AD, compared with the ε4HMZ (Liu et al., 2013). This suggests that the two imaging findings reported here may reflect distinct levels of the genetic risk for AD, which appear to be hierarchically organized.

Taken together, our results indicate that APOE-ε4 shapes the cerebral organization in middle-age asymptomatic individuals, even in individuals with no amyloid pathology, in a way that recapitulates focal morphometric alterations observed in MCI and AD patients. This provides new insights for the structural vulnerability associated to this highly penetrant risk variant. Our study promotes the use of structural covariance networks as a valuable AD intermediate phenotype, which may be subject to change earlier than cortical thinning or volume loss. Future studies shall track GMn in longitudinal designs along with more extensive biomarkers screening to further address potential interaction between AD-related pathology and APOE genetic variance.

Click here for additional data file.