Characterizing the structural covariance network in AD-susceptible single nucleotide polymorphisms and the correlations with cognitive outcomes

Background The clinical manifestations of Alzheimer disease (AD) are related to brain network degeneration, while genetic differences may mediate network change patterns. A number of AD-susceptible loci have been reported using genome-wide association studies, however, how they modulate intracerebral volume and relationships to cognitive outcomes remains to be established. We hypothesized that different genotype groups may modulate large-scale brain networks independently or interact with apolipoprotein E4 (ApoE4) status to determine neurobehavior test scores. Gray matter structural covariance networks were constructed in 324 patients with AD using T1 magnetic resonance imaging with independent component analysis (ICA). We assessed 15 genetic loci (rs9349407, rs3865444, rs670139, rs744373, rs3851179, rs11136000, rs3764650, rs610932, rs6887649, rs7849530, rs4866650, rs3765728, rs34011, rs6656401, rs597668) using the additive, recessive and dominant model on clinical outcomes. Statistical analysis was performed to explore the independent role of each locus, interactions with ApoE4 status, and relationships to brain ICA network integrity score. We used Cognitive Abilities Screening Instrument (CASI) total score or short-term memory (STM) subscore as the major outcome factors, adjusted for covariates of education, disease duration and age. Clinically, the CD2AP G allele showed a protective role in CASI-total and CASI-STM scores independently or via interactions with non-ApoE4 status, while the CR1 A genotype group was associated with lower STM independently of ApoE4 status. Three loci showed synergic interactions with ApoE4: BIN 1 T allele and MS4A6A G allele with non-ApoE4 status, and FTMT G allele with ApoE4 status. The network integrity scores revealed 9 signicant ICA networks (anterior and posterior hippocampus, right temporal, right or left thalamus, inferior cerebellum, medial cerebellum, default mode network, frontal attention network) that correlated with cognitive scores, in which only the ApoE4 and MS4A6A genotype group was independently related to the hippocampus network. Genetic loci of MS4A6A, BIN1, CD2AP, CD33, CLU, BIN1, P73, EXOC3L2, CR1 and MS4AE4 exerted network inuence independently or via interactions with ApoE4 status. This study suggests that AD-susceptible loci may exert clinical signicance independently, through interactions with ApoE4 status or by modulating ICA networks to determine cognitive outcomes. measures of The SNP-brain co-variation between genotype while integrity can used as a dependent variable to test the genetic interactions. In this study, we tested whether SCNs constructed using ICA could serve as an endophenotype for cognitive test outcomes in AD patients. Based on GWAS results, we assessed 15 AD-associated SNPs belonging to the following genetic loci: CD2AP (rs9349407), CD33 (rs3865444), MS4AE4 (rs670139), BIN1 (rs744373), PICALM (rs3851179), CLU (rs11136000), ABCA7 (rs3764650), MS4A6A (rs610932), FTMT (rs6887649), SPTLC1 (rs7849530), Intergenic SNP (rs4866650), p73 (rs3765728), FGF1 (rs34011), CR1 (rs6656401) and EXOC3L2 (rs597668). We tested the independent effect of each SNP on cognitive outcomes and evaluated whether the associations were related to interactions with the ApoE4 genotype. The clinical signicance of individual SNPs or interactions between the SNP and ApoE4 on SCN was calculated and validated by correlating the network integrity score and cognitive scores. Puried amplicons were then subjected to primer extension using an iPLEX Gold Reagent Kit. Primer extension was performed with a cycling program of 94°C for 30 s, followed by 40 cycles of 94°C for 5 s, and 5 cycles of 52°C for 5 s and 80°C for 5 s within 40 cycles, followed by a nal extension at 72°C for 3 min. The extended reaction products were puried by cation exchange resins and then spotted onto a 384-format SpectroCHIP II array using a MassArray Nanodispenser RS1000. Mass determination was done on a MassARRAY Compact Analyzer. The resulting spectra were processed and alleles called with MassARRAY Typer 4.0 with model-based cluster analysis to analyze the genotypes of the SNPs. We tested 15 SNPs (rs9349407, rs3865444, rs670139, rs744373, rs3851179, rs11136000, rs3764650, rs610932, rs6887649, rs7849530, rs4866650, rs3765728, rs34011, rs6656401, rs597668). The risk alleles and minor allele frequencies (MAFs) are listed in Supplementary Table 1. The ApoE genotype was determined using rs7412 and rs429358. ApoE4 carriers were dened as those with one or two E4 alleles.

measures of interest. The SNP-brain relationships modelled by ICA can be used to test the co-variation of GM between genotype groups, while the network integrity score can be used as a dependent variable to test the genetic interactions.

Methods
This study was conducted in accordance with the Declaration of Helsinki and was approved by the Institutional Review Board of Chang Gung Memorial Hospital. The study participants were treated at the Cognition and Aging Center, Department of General Neurology, Kaohsiung Chang Gung Memorial Hospital.
The multi-disciplinary team was composed of behavior neurologists, psychiatrists, neuropsychologists, neuroradiologists and experts in nuclear medicine. We enrolled patients with AD who were diagnosed according to the International Working Group-2 criteria (Dubois, Feldman et al. 2014), and further con rmed by amyloid imaging (TW-ADNI: http://tadni.cgmh-mi.com/home) if the consensus panel did not agree on the diagnosis. All of the patients were in a stable condition under acetylcholine esterase inhibitor treatment from the time of diagnosis. The exclusion criteria were a past history of clinical stroke, a negative amyloid scan, a modi ed Hachinski ischemic score >4 and depression. After checking the inclusion and exclusion criteria, a total of 324 subjects (152 males, 172 females) were included and underwent imaging and genetic tests.

Clinical and Neurobehavioral Assessments
After enrolment, the demographic data of each patient were recorded. We also recorded the time of the rst symptom. A trained neuropsychologist administered the neurobehavioral tests using Cognitive Abilities Screening Instrument (CASI) total scores as a global assessment of cognitive function. The CASI contains nine subdomains. We used attention, verbal uency, abstract thinking, and mental manipulation subdomain scores to assess executive function (Huang, Chang et al. 2013), and orientation, short-and long-term memory, language ability, and drawing as non-executive domains. As the salient feature of AD is short-term memory (STM) impairment, we used the CASI-STM subscore as the major outcome for genetic correlations.
Genotyping SNP genotyping was performed using MassARRAY technology with iPLEX Gold chemistry (Agena Bioscience, San Diego, CA, USA). The PCR primers and single-base-extension primers were designed using Assay Design Suite v2.0 software. The genotyping analysis was performed using an iPLEX Gold Reagent Kit according to the manufacturer's instructions. Brie y, 1 µl of DNA sample (10 ng/µl) was subjected to 5 µl of PCR reaction containing 0.2 units of Taq polymerase, 2.5 pmol each of the PCR primers and 25 mM each of the dNTPs. Thermocycling was started at 94°C for 2 min followed by 45 cycles of 94°C for 30 s, 56°C for 30 s and 72°C for 1 min, and a nal extension was done at 72°C for 1 min. Unincorporated dNTPs were dephosphorylated by 0.3 U of shrimp alkaline phosphatase. Puri ed amplicons were then subjected to primer extension using an iPLEX Gold Reagent Kit. Primer extension was performed with a cycling program of 94°C for 30 s, followed by 40 cycles of 94°C for 5 s, and 5 cycles of 52°C for 5 s and 80°C for 5 s within 40 cycles, followed by a nal extension at 72°C for 3 min. The extended reaction products were puri ed by cation exchange resins and then spotted onto a 384-format SpectroCHIP II array using a MassArray Nanodispenser RS1000. Mass determination was done on a MassARRAY Compact Analyzer. The resulting spectra were processed and alleles called with MassARRAY Typer 4.0 with model-based cluster analysis to analyze the genotypes of the SNPs. We tested 15 SNPs (rs9349407, rs3865444, rs670139, rs744373, rs3851179, rs11136000, rs3764650, rs610932, rs6887649, rs7849530, rs4866650, rs3765728, rs34011, rs6656401, rs597668). The risk alleles and minor allele frequencies (MAFs) are listed in Supplementary Table 1. The ApoE genotype was determined using rs7412 and rs429358. ApoE4 carriers were de ned as those with one or two E4 alleles.

Image Acquisition
Magnetic resonance images were acquired using a 3.0T magnetic resonance imaging (MRI) scanner (Excite, GE Medical Systems, Milwaukee, WI, USA). All MRI images were performed within 3 months of cognitive test scores. High-resolution structural images were acquired for spatial normalization using the following protocols: a T1-weighted, inversion-recovery-prepared, three-dimensional, gradient-recalled acquisition in a steady-state sequence with a repetition time/echo time/inversion time of 8,600 ms/minimal/450 ms, a 256 × 256 mm eld of view, and a 1-mm slice sagittal thickness with a resolution of 0.5 × 0.5 × 1 mm3.

Data Analysis for Neuroimaging Biomarkers
Image preprocessing and statistical analysis were performed using SPM12 (Wellcome Trust Centre of Cognitive Neurology, University College London, UK, http://www. l.ion.ucl.ac.uk/spm/). The T1 images were reoriented, realigned, and normalized using the standard Montreal Neurological Institute space. The images were then segmented into GM and white matter. Related tissue segments were used to create a custom template using the diffeomorphic anatomical registration using exponentiated lie algebra approach, which is one of the highest ranking registration methods in patients with AD (Cuingnet, Gerardin et al. 2011). The modulated and warped images were then smoothed using a Gaussian kernel of 8 mm full width at half maximum.
The preprocessed spatial normalized modi ed T1 images from the patients were concatenated to form a subject series and entered into the ICA process.
Spatial ICA was carried out using Multivariate Exploratory Linear Optimized Decomposition into Independent Components software package version 3.15 (http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/MELODIC). The resulting independent components were z-transformed and visualized using a threshold of z > 1.96 (p < 0.05). Differences in ICA intensities between groups in each SNP were calculated. In addition, interactions between individual SNPs and ApoE4 status were also modelled using a general linear model, corrected for age, educational level, gender and disease duration (years). To understand the clinical signi cance of the identi ed SCN in the patients, we also calculated partial correlations between the extracted SCN intensity and the clinical scores by setting the signi cance value at p < 0.05 using Bonferroni correction for multiple comparisons and covariates of age, educational level and disease duration (years).

Statistical Analysis
Clinical and laboratory data were expressed as mean ± standard deviation. The Student's t test was used to compare continuous variables, and the chi-square test was used for categorical variables. Linear regression models of associations between the cognitive scores and 15 selected AD-related SNPs were analyzed, adjusting for education, age and duration of disease. All statistical analyses were conducted using SPSS software (SPSS version 22 for Windows®, SPSS Inc., Chicago, IL). Statistical signi cance was set at p < 0.05.

Effect Of Independent Genotype Groups On Cognitive Outcomes
Two SNPs showed independent or synergic effects on cognitive measures. The CD2AP G allele was associated with CASI-total scores (Table 2) on the additive or dominant model, and the G allele was associated with higher scores. The CD2AP G allele (on the additive or dominant model) was also associated with a higher STM score (Table 3). The association between the CD2AP G allele and cognitive scores showed interactions with the non-ApoE4 genotype in CASI-total score (Table 2) and STM score (Table 3). In the recessive model, the CR1 A allele was associated with a lower STM score (Table 3), while the risk of a detrimental effect was independent of ApoE4 status. Of note, the MAF of the A allele in CR1 was 0.029.

Network Topography And Clinical Signi cance Validation
To understand the effects of SCN and SNP genotype groups, a total of 20 ICA components were constructed. For each SNP, we compared network integrity score between SNP genotype groups (Supplementary Table 2). The strati cation of each SNP genotype group was based on the number of cases and the balance of Hardy-Weinberg equilibrium. The spatial extent of 15 ICA modes showing genotypic differences is shown in Fig. 1.
Using network integrity scores, we correlated each signi cant ICA with cognitive tests (Table 4) to establish network signi cance. The CASI scores were classi ed as indicating general cognitive performance (CASI-total score), executive domains and non-executive domains. The anterior hippocampus, posterior hippocampus, right temporal, right thalamus, left thalamus, inferior cerebellum default mode network and frontal attention network were signi cantly related to CASI-total scores. For correlations with executive domains scores, the network patterns overlapped with the general performance network except for default mode network and inferior cerebellum. The topography of non-executive networks overlapped with the CASI-total score network. Meanwhile, the network integrity scores of supplementary motor area (SMA) and postcentral gyrus were inversely correlated with STM.   Numbers under the cognitive test name represent partial correlation coe cient, adjusted for age, disease duration and education.
A number of networks showed SNP-ApoE4 interactions (Table 4, Fig. 1A-D, G, H, M). These networks included cardinal networks of AD, including anterior and posterior hippocampus, right temporal, postcentral gyrus, frontal executive network and supplementary motor cortex. All preselected SNPs interacted with ApoE4 status in the abovementioned networks except CLU. Among the 15 networks, only basal ganglia (Fig. 1A), lateral cerebellum (Fig. 1J) and cingulate network (Fig. 1O) showed no correlations with clinical scores.

Major Findings
In this study, we explored the clinical signi cance of 15 AD-susceptible loci by constructing SCNs using an ICA approach in patients with AD. Our results re ected complex interactions among preselected SNPs, brain network integrity and cognitive outcomes in which the mechanisms were via different pathways. First, the SCN served as the endophenotype in predicting individual genotype groups in cognitive outcomes. For SCN as the endophenotype, the genotype group either exerted an independent role or synergistic interactions with ApoE4 status on cognitive measures. The SCNs showing clinical signi cance included the hippocampal axis, temporal, thalamus, default mode network and frontal attention network. As the SCNs were each associated with different cognitive domains, the identi cation of genotype-SCN relationships may help to understand the neurobiology in AD. Second, for genetic-clinical relationships, our results suggested that the main genotype effect on cognitive measures was via interactions with ApoE4 status. This included protective CD2AP G allele and MS4A6A G allele with non-ApoE4 interactions, and risk allele of BIN1 T allele and FTMT G allele with ApoE4 status. Of note, independent roles of genotype groups (CD2AP and CR1) on cognitive outcomes were also found. However, as CR1 risk allele A was a minor allele with a low MAF, the CR1 nding here should be interpreted with caution.
ApoE4 Modulated Hippocampal SCN and Determined the Salient Feature of AD By de nition, the SCN is based on the similarity of the same microstructural variations and thus may be in uenced by factors in uencing underlying structures such as the expression of common genetic traits during development. The independent role of ApoE4 on the hippocampal and SMA SCN was established in this study. As the hippocampal SCN intensity score was also correlated with STM score in our AD patients, this nding supports the role of ApoE4 in linking salient cognitive and biosignature features in AD.
Traditional cytoarchitectonic distribution of the hippocampus (Frederickson, Klitenick et al. 1983) is a convoluted GM structure encompassing three architectonically distinct regions: the fascia dentata, the CA region (which can be subdivided into CA1-CA4 elds), and the subicular complex. In this study, two hippocampus SCNs in ICA (anterior versus posterior hippocampus) showed ApoE4 genotype group differences. The SCN of the hippocampus axis associated with ApoE4 showing the anterior-posterior axis is consistent with task-related activities or connectivity patterns (Colombo, Fernandez et al. 1998, Przezdzik, Faber et al. 2019. A similar anterior-posterior organization has recently been reported, suggesting that the gene expression is linked to the distinct molecular gradient in the hippocampus (Vogel, La Joie et al. 2020). Among the 15 preselected SNPs and ApoE4, only ApoE4 showed an independent role in the entire hippocampal axis. The association with cognitive test scores in hippocampal SCN intensity scores and the interactions between hippocampus and cortical/subcortical structures demonstrate its tight integration within the large-scale degenerative system, so that hippocampus-ApoE4 is the most important factor in predicting AD functional severity.
ApoE4 Interacted with MS4A, CR1, PICALM, p73, BIN1 and EXOC3L2 on the Hippocampal Axis Our results showed the role of MS4A6A (rs610932) on posterior hippocampus axis. In addition, MS4A4E and MSA4A6A also interacted with ApoE4 in the entire hippocampal SCN, similar to that reported in a smaller AD cohort (Chang, Mori et al. 2019). A previous study reported interactions between MS4A and CLU (Lambert, Ibrahim-Verbaas et al. 2013) or CD33 on conferring the risk of AD. Common variants of MS4A6A (rs610932), along with MS4A4E (rs670139), CD33 (rs3865444) and CD2AP (rs9349407), CLU and PICALM have been associated with memory decline (Hollingworth, Harold et al. 2011, Naj, Jun et al. 2011, Karch, Jeng et al. 2012). Our results emphasize the role of the hippocampus and MS4A epistasis on the clinical features of AD.
The SNPs that interacted with the ApoE4 allele on the hippocampal axis in this study may explain the mechanisms of AD susceptibility. CR1 on chromosome 1 (rs6656401, odds ratio [OR] = 1.21, 95% con dence interval [CI] 1.14-1.29, p = 3.7 × 10 − 9 ) has been shown to encode complement component (3b/4b) receptor 1 (Lambert, Heath et al. 2009) which is known to participate in the clearance of β amyloid peptides in AD pathology. By facilitating the clearance of Aβ peptides through enhanced translocation, increased micro-vessel PICALM (phosphatidylinositol binding clathrin assembly protein) (rs3851179, OR = 0.86, p = 1.3 × 10 − 9 ) has been shown to reduce the development of AD (Harold, Abraham et al. 2009). The SNP rs3765728 within tumor protein p73 (Tao, Sun et al. 2014) has been associated with neuron survival, and has been found to have a synergetic effect with APOE on the risk of AD. In addition, rs744373 near BIN1 (OR = 1.13; 95% CI, 1.06-1.21 per copy of the minor allele; p = 1.59 × 10 − 11 ) has been related to AD with possible mechanisms involving tau-related cascade (Franzmeier, Rubinski et al. 2019). Moreover, rs597668 is near EXOC3L2 (OR, 1.18; 95% CI, 1.07-1.29; p = 6.45 × 10 − 9 ), and mutations of this gene may be associated with AD.
Dominant Model Suggested Protection of the CD2AP rs9349407 G Allele in Non-E4 Carriers The role of rs9349407 as a risk SNP in Han Chinese people has been reported , but the ndings have been inconsistent (Jiao, Liu et al. 2015). One meta-analysis suggested that rs9349407 C is a risk allele for AD susceptibility in East Asian, American, Canadian, and European populations . In our additive or dominant model, AD patients with the rs9349407 G allele had higher CASI total and STM scores, especially the non-E4 carriers. The minor allele C of rs9349407 has been associated with neuritic plaque burden in pathology, which may explain why non-E4 carriers with the G allele may have higher cognitive test scores (Shulman, Chen et al. 2013).
In the current study, we only enrolled patients with clinical AD, and we tested whether risk or protective SNP alleles have an effect on cognitive test scores or SCNs. Although not all of our patients with AD received amyloid scans, our exclusion criteria reduced the possibility of non-AD pathologies. rs9349407 is a polymorphism in the CD2AP gene which translates the scaffolding molecule for signal transduction. Loss of function of CD2AP has been linked to enhanced Aβ production, tau-induced neurotoxicity, abnormal neurite structure modulation and reduced blood-brain barrier integrity, which has been implicated in AD pathogenesis (Dubey, Gulati et al. 2018, Ramos de Matos, Ferreira et al. 2018).

Modulation of Frontal Attention Network by the CR1 rs6656401 A Allele on Cognitive Outcomes
As STM is the salient feature in AD, it was used as a dependent variable in SNP genetic model analysis. In addition to rs9349407, we also identi ed the independent role of rs6656401 A allele on lower STM scores, consistent with a previous meta-analysis of a greater risk on the minor allele A in AD (Shen, Chen et al. 2015). CR1 is an AD susceptibility locus that also in uences AD-related traits on neuritic plaque deposition and in episodic memory decline. The coding variant in the long homologous repeat D region of the CR1 gene, rs4844609 (Ser1610Thr) has been associated with episodic memory decline and accounts for the known effect of SNP rs6656401 (Keenan, Shulman et al. 2012). In SCN analysis in this study, the effect may have been partially modulated by interactions with ApoE4 on the hippocampal axis.
Three studies have con rmed the association between AD susceptibility and the rs6656401 A allele in Chinese patients (Zhang, Yu et al. 2010, Chen, Kao et al. 2012, Jin, Li et al. 2012, although another study showed no association (Li, Shi et al. 2011). rs6656401 is in the CR1 gene, and complement system activation in the clearance of amyloid has been proposed to be a possible mechanism of the risk associated with rs6656401 (Zhang, Yu et al. 2010). Although the independent role of rs6656401 minor allele on lower STM was con rmed in this study, it is worth noting that the A allele was a minor allele and the MAF was only 0.029. Therefore, the relationship with lower STM may be due to the small sample size of the A allele. To understand the effect, we also checked SCN intensity and explored whether the relationships between the rs6656401 A allele and cognitive test results were modulated by SCN. As shown in our correlation analysis, the frontal attention SCN intensity was signi cantly different between A carriers and G allele, and the intensity was also related to the executive and non-executive domains. Therefore, we suggest that the frontal attention network is the endophenotype of the link between rs6656401 and cognitive outcomes.

Effect of FTMT on STM in AD via ApoE4 and Default Mode Network
The intergenic SNP (rs6887649) is 10 KB upstream of ferritin mitochondrial gene (FTMT), which has been shown to modify the association between amyloid positivity and baseline ventricular volume (Hohman, Koran et al. 2014). From our study, the effect of FTMT on STM is via interaction with ApoE4, while it also modulates the default mode network. The default mode network is regarded to be an early neuroimaging bio-signature (Chang, Huang et al. 2015), and a recent report suggested that the default mode network may be comprised of multiple, spatially dissociated but interactive components (Andrews-Hanna, Reidler et al. 2010), of which two subsystems are of particular interest. Our SCN result of an association with FTMT was consistent with cores in the posterior cingulate cortex and anterior medial prefrontal cortex that is known as the "dorsal medial prefrontal cortex subsystem" (or midline core subsystem).

Limitations And Methodological Considerations
This study has several limitations. First, we enrolled subjects with early stage AD, and we did not include a control group. As the SNPs were preselected from GWAS results showing signi cance in AD susceptibility, the inclusion of a control group may have helped to elucidate whether these SNPs exerted similar GM modulation patterns in healthy elderly subjects as in those with AD. However, as these SNPs each exert different functional activities on the pathogenetic mechanisms in AD, the use of a pure AD population may help to maximize the effect of each SNP on regional GM networks. As a structural covariance matrix is de ned by estimating the inter-regional correlations of cortical volumes between all possible pairs of regions de ned by anatomy, SCN construction relies on both the spatial patterns of morphometric and signal similarities. Given the differences in brain morphometry in controls, it would be di cult to match all in uential factors. Therefore, we focused on the initial hypothesis testing the endophenotypic role of SCN in cognitive outcomes. Another limitation is the estimation of the number of components for ICA analysis. Most studies have used 12 to 30 components in structural networks or resting state networks. In this study, we constructed 20 networks and ltered the clinical signi cance of a network using correlation analysis with cognitive measures and by matching with the ICA template (https://brainmap.org/icns/). From a methodological aspect, the group-wise structural covariance analysis relied on the morphological properties of each voxel with the rest of the brain across a group of participants. The registration of a single participant's structural data to the template involves linear and non-linear deformation that can result in inaccuracies in subregions. In this report, we emphasized the well-characterized network to explain the SNP effect where the network signi cance was established by correlations with cognitive measures. Finally, the use of 15 AD-susceptible SNPs to validate the underlying pathological mechanisms may have oversimpli ed the genetic interactions. We only tested interactions between each SNP and ApoE4 status as ApoE4 remains the strongest predictor. The interpretations of interactions between SNPs and SCN were based on a literature review of possible alterations in functional pathways and may not fully explain the in vivo situation. Therefore, the ndings should be interpreted with caution.

Conclusion And Future Perspectives
In this study, we aimed to elucidate the major modes of structure variation in AD-susceptible SNPs and assess the SNP co-plastic properties. Our ndings demonstrated AD-related SNP effects that may in uence the SCN independently or synergistically with ApoE4. The use of SCN as an endophenotype allowed us to assume the independent and synergistic role of putative SNPs to predict cognitive measures. The complex interplay among these SNPs in our study suggests that a hierarchical order of SNPs modulate brain networks.

Ethics approval and consent to participate
This study was approved by the Chang Gung Memorial Hospital Ethics Committee following the standards for medical research in humans recommended by the Declaration of Helsinki. All participants gave signed, informed consent for the cognitive tests described, inclusion of samples, neuroimaging and demographic data.

Consent for publication
Not applicable Availability of data and materials The data analyzed during the current study are available from the corresponding author on reasonable request.

Competing interests
The authors declare that they have no competing interests.