Brain structure and allelic associations in Alzheimer's disease

Abstract Background Alzheimer's disease (AD), the most prevalent form of dementia, affects 6.5 million Americans and over 50 million people globally. Clinical, genetic, and phenotypic studies of dementia provide some insights of the observed progressive neurodegenerative processes, however, the mechanisms underlying AD onset remain enigmatic. Aims This paper examines late‐onset dementia‐related cognitive impairment utilizing neuroimaging‐genetics biomarker associations. Materials and Methods The participants, ages 65–85, included 266 healthy controls (HC), 572 volunteers with mild cognitive impairment (MCI), and 188 Alzheimer's disease (AD) patients. Genotype dosage data for AD‐associated single nucleotide polymorphisms (SNPs) were extracted from the imputed ADNI genetics archive using sample‐major additive coding. Such 29 SNPs were selected, representing a subset of independent SNPs reported to be highly associated with AD in a recent AD meta‐GWAS study by Jansen and colleagues. Results We identified the significant correlations between the 29 genomic markers (GMs) and the 200 neuroimaging markers (NIMs). The odds ratios and relative risks for AD and MCI (relative to HC) were predicted using multinomial linear models. Discussion In the HC and MCI cohorts, mainly cortical thickness measures were associated with GMs, whereas the AD cohort exhibited different GM‐NIM relations. Network patterns within the HC and AD groups were distinct in cortical thickness, volume, and proportion of White to Gray Matter (pct), but not in the MCI cohort. Multinomial linear models of clinical diagnosis showed precisely the specific NIMs and GMs that were most impactful in discriminating between AD and HC, and between MCI and HC. Conclusion This study suggests that advanced analytics provide mechanisms for exploring the interrelations between morphometric indicators and GMs. The findings may facilitate further clinical investigations of phenotypic associations that support deep systematic understanding of AD pathogenesis.


| INTRODUC TI ON
Alzheimer's disease (AD) is by far the most common form of dementia among the elderly. 1,2 Late onset Alzheimer's disease (LOAD), defined by the onset of symptoms after age 65, is sporadic, nonfamilial AD. 3,4 Genetic studies have provided significant insights on the molecular basis of AD, but the mechanisms underlying AD onset and progression remain largely unexplained. While the underlying causes of LOAD are still unknown, there is evidence from familial aggregation, transmission pattern, and twin studies that AD has a substantial genetic component that has an estimated heritability of 58%-79%, and the lifetime risk of AD among first-degree relatives of patients may be twice that of the general population. 5,6 Recent genome-wide association studies (GWASs), which examine associations of AD diagnosis with genetic markers (single-nucleotide polymorphism [SNP]) across the genome, have discovered more than 20 AD gene variants that confer genetic risk. 7,8 These findings improve the understanding of risks and causes for AD, and may guide diagnosis and therapy on a patient-specific basis. 9 However, casecontrol GWAS cannot completely characterize the exact roles of the identified genetic susceptibility loci in the pathophysiology of AD. Joint analysis of genetic and neuroimaging data could uncover the genetic mechanism in the disease's underlying biology. 5 This study investigates holistically the significance of multi-gene patterns associated with neuroimaging markers (NIMs) of AD using imaging and genomic data of the Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort. 10 Several single-nucleotide polymorphisms (SNPs) present in apolipoprotein E (APOE) gene have previously been associated with neuroimaging measures in both cognitively healthy control (HC) or impaired (such as mild cognitive impairment [MCI] and AD dementia) patients. [11][12][13] However, a single gene or a few imaging measures may be insufficient to understand the multiple mechanisms and imaging manifestations of the complex disease.
Recent and ongoing advances in neuroimaging and genetics, including high-throughput genotyping techniques, have made it possible to scan populations with multimodality neuroimaging to collect genome-wide data and to study the influence of genetic variation on the brain structure and function.
In this work, we related high-throughput neuroimaging-derived phenotypes of brain structure to the clinical states of AD, and then associated the significant AD-related NIMs with GWAS-supported susceptibility genetic variants for AD to obtain true system-level gene-brain associations in dementia. Specifically, we used structural brain imaging to obtain biomarkers of a wide variety of brain morphological properties, allelic data to capture genotypic variation, and functional connectivity to evaluate imaging-genetic-phenotypic variation. We present a neuroimaging genetics framework that uses a whole-genome-and-whole-brain strategy to systematically evaluate genetic effects on neuroimaging phenotypes to discover quantitative trait (QT) loci. QT association studies have been shown to have increased statistical power and thus decreased sample size requirements. 14,15 In addition, neuroimaging phenotypes may be closer to the underlying biological etiology of the disease, making it easier to identify underlying genes. The methodology proposed in this paper is based on the identification of strong associations between regional neuroimaging phenotypes as QTs and SNP genotypes as QT loci.
The genetics of AD are complex because the practical effects may be weak, albeit statistical effects could still be strong, samplesizes are often unbalanced (number of cases ≪ genomic markers [GMs]), and considerable difficulties with result replication and validation. 16,17 Large-scale GWAS shows promise in untangling the genetic footprint of this neurodegenerative disease. Considering the limited sample size in the ADNI cohort (n = ~1200), we used the AD-related genetic variants identified by the largest (n = ~450,000) case-control GWAS in AD to date 8 instead of performing GWAS on the ADNI cohort.
We hypothesized that there exist significant relationships between the AD-related NIMs and the GWAS-supported susceptibility genetic variants for AD. Several prior studies have been conducted on the relationship between neuroimaging phenotypes and genetic variants. [18][19][20] However, few reports have previously performed functional analyses of neuroimaging genetics. This study expands the knowledge about dementia phenotypes using modern neuroimaging genetics and the network analysis to explore the relationships between genetic, phenotypic, and NIMs.
Multinomial linear models of clinical diagnosis showed precisely the specific NIMs and GMs that were most impactful in discriminating between AD and HC, and between MCI and HC. For ADNI-1 participants, quality assessment data available on the LONI IDA 21 (MRI MPRAGE Ranking: MRIMPRANK) was used to select the higher-rated scan for each visit. Accelerated scans were removed and 3T scans were chosen over 1.5T scans where possible. For those with multiple scans after this filtering, the scan with the larger Image ID was kept. For the ADNI-GO and ADNI-2 participants, accelerated scans were removed and among the remaining participants with more than one scan, the scan with the larger Image ID was kept. After these filtering steps were completed, the scans were combined into a final set of scans consisting of 1242 participants with only one scan each. The ADNI database query yielded a pool of 1242 volunteers with both brain MRI and genetic data from the study phases of ADNI-1, ADNI-GO and ADNI-2. Three subjects were discarded due to failed MRI processing. We further removed 110 participants who did not have CEU ancestry (ancestry from Northern and Western Europe) and employed the remaining 1129 subjects for identifying the AD-related NIMs. In the final imaginggenetics association analysis another 103 subjects were excluded as they did not pass the genetic data quality control. We ended up with 1026 subjects who successfully completed all imaging and genetics processing protocols. For each participant, clinical severity of dementia was assessed using an annual semi-structured interview, which yielded an overall Clinical Dementia Rating (CDR) score and the CDR Sum of Boxes (CDR-SB). In addition, the Mini-Mental State Examination (MMSE) and a neuropsychological battery were also recorded.
All the ADNI participants included in this study were those with a baseline age of 65

| MRIprocessingandanalysis
The MRI acquisition protocols can be found on the ADNI website (https://adni.loni.usc.edu) and have been previously described elsewhere. 22,23 Baseline structural MRI scans of the ADNI subjects were processed for reconstructing cortical surfaces, brain parcellation and extracting morphological phenotypes using the FreeSurfer (v6.0) software package (https://surfer.nmr.mgh.harva rd.edu/). 24 The FreeSurfer processing includes motion correction and averaging of volumetric T1-weighted images, 25 removal of non-brain tissue, 26 automated Talairach transformation, brain volume segmentation, 27,28 intensity normalization, 29 tessellation of the boundary between gray matter and white matter, automated topology correction 30 and surface deformation. 31 Once the cortical models are completed, a number of deformable procedures were performed for further data processing and analysis, including surface inflation, registration to a spherical atlas using individual cortical folding patterns to match cortical geometry across subjects, 32 and finally creation of a variety of surface-based data including maps of surface area, cortical thickness, curvature features, etc.
For each subject, 1380 imaging-derived biomarkers were extracted using FreeSurfer, including measures of surface area, volume, thickness, standard deviation of thickness, mean curvature, Gaussian curvature, folding index, curvature index and/or gray matter/white matter contrast for different cortical, subcortical, and white matter regions.
All the imaging phenotypes were adjusted for age, gender, education, handedness, and intracranial volume (ICV) using linear mixedeffects regressions. ANOVA tests were then performed to find the NIMs associated with AD diagnosis at the GWAS significance level p < 5 × 10 −8 in the cohort of N = 1129 subjects with a CEU ancestry.

| Geneticsdataprocessing
Genetic SNP data was downloaded from the ADNI database (https:// www.loni.usc.edu/ADNI) through the LONI imaging data archive (IDA) interface (https://ida.loni.usc.edu/pages/ acces s/genet icData. jsp) onto the LONI Cranium high-performance computing (HPC) cluster. The processing resulted in a single dataset containing the genetics information of all 1026 participants. The ADNI-1 genetics data was downloaded as PLINK bed/bim/bam files in the hg18 (build 36) format. The genome build was converted from hg18 to hg19 using liftOver, as described in (https://www.nature.com/artic les/ nprot.2015.077). The ADNI-GO and ADNI-2 genetics data, which is in the hg19 (build 37) format, was downloaded as PLINK bed/bim/ bam files for sets 1-9 and as intensity data CSV files for sets 10-15. The intensity data CSV files were converted to PLINK files at a multiple GenCall Score (GC) threshold of 0.15 based on the procedure described in Ref. [13].
Used genetic imputation, we harmonized the genetic data across the different ADNI studies. To prepare for imputation, population stratification analysis was first used to remove all of the non-CEU participants. We used PLINK for population stratification. PLINK relies on genome-wide average proportion of alleles shared between any two individuals to cluster subjects into homogeneous subsets and perform classical multidimensional scaling (MDS) to visualize substructure and provide quantitative indices of population genetic variation. Next, we used the "HRC or 1000G Imputation preparation and checking" tool (HRC-1000G-checkbim-v4.2.9) from the McCarthy Group to conduct common preimputation checks, such as strand, reference allele assignment, and frequency differences (https://www.well.ox.ac.uk/~wrayn er/ tools). Imputation was completed using the Michigan Imputation Server v1.0.4 (Sept. 14, 2018) (https://imput ation server.sph. umich.edu/index.html). This imputation, based on Minimac3, 33 was completed with the accompanying quality control offered by the service. The reference panel used was the HRC r1.1 2016, phasing was competing using Eagle, and quality control was based on the European population. The ADNI-1 input consisted of 694 samples and 568,933 SNPs, 10,583 of which were excluded for imputation due to being monomorphic or having a SNP call rate of <90%. The ADNI-GO/2 input consisted of 723 samples and 696,245 SNPs, 38,605 of which were excluded due to being monomorphic or having a SNP call rate of <90%. The output of the Michigan Imputation Server was in the minimac3 output format, including both info and dosage files. The HRC-imputed data for the ADNI-1 and ADNI-GO/2 datasets were merged with bcftools (https://samto ols.github.io/bcfto ols/bcfto ols.html). After this imputation and combination, the sites were filtered to only include those with R 2 > 0.6 and with a minor allele frequency >0.5% using bcftools and tabix (https://samto ols.github.io/bcfto ols/bcfto ols. html, http://www.htslib.org). Finally, the data was converted into the pgen format (PLINK 2 binary format) using PLINK 2.00 alpha (https://www.cog-genom ics.org/plink/ 2.0/).

| Thepipelinecomputationalenvironment
To manage the large and complex raw and derived data, design and execute the end-to-end processing protocols, and to track provenance, we employed the LONI Pipeline. [34][35][36][37] The Pipeline is a graphical workflow environment facilitating the collaborative design, execution, validation, visualization, modification and sharing of complex heterogeneous computational protocols.
To promote reproducible open-science development and validation, we designed a Pipeline workflow that represents an end-to-end computational protocol for high-throughput data preprocessing.
The pipeline workflow includes skull-stripping, volumetric registration, brain anatomical parcellation, extraction of volume and cortical thickness and between group statistical analyses of shape regional differences. The output of the pipeline workflow is a collection of 3D scenes illustrating the statistically significant regional anatomical differences between the study cohorts.
Rank-ordering the complete collection of NIMs, we chose the 200 most salient NIMs which provided the highest discrimination between the AD and HC groups. These 200 NIMs were derived from all structural imaging data using the workflow and are based on the automated ROI extractions generated by FreeSurfer. Finally, the pipeline workflow, computed the most significant genotypic discriminants among AD, MCI and HC subjects. The 200 NIMs were then associated with the top 29 SNPs, which were chosen by the PLINK. 8

| Analyticalprotocol
The end-to-end data analysis protocol was implemented via the Pipeline graphical workflow environment and involved the following steps (1) Imputation on the Michigan Server using IMPUTE2, a genotype imputation and haplotype phasing program based on ideas from, 38 and (2) Beagle is a software package for phasing genotypes and for imputing ungenotyped markers. 39 The SOCR statistical computing infrastructure (https://SOCR.umich.edu) [40][41][42][43][44][45] was utilized to implement and execute the end-to-end computational statistics protocol, which included multivariate linear modeling and general parametric and non-parametric statistical analyses.
The imputation protocol relied on the default setting. Following quality control (QC) and imputation, there were 1026 subjects with a European ancestry remaining with minimac3 outputs, including both info and dosage files. Next, we combined the ADNI-1 and ADNI-GO/2 imputed data into a single PLINK file. ADNI1 and ADNIGO/2 are genotyped using different chips, so we imputed the arrays prior to their integration and ran GWAS on the combined array. The resulting data were filtered to include only the subjects with a CEU ancestry and contained the following four tensors: We extracted the AD-related GMs and NIMs for the network analysis. Genotype dosage data for AD-associated SNPs were extracted from the imputed ADNI genetics dosage data using samplemajor additive (0/1/2) coding. A set of 29 SNPs were selected, representing a subset of independent SNPs found to be highly associated with AD in a recent AD meta-GWAS study by Jansen et al., 8 which met the MAF threshold determined by the imputation process. The genotyping data of these SNPs are included in the genetic. markers.associated.with.ADdx.CEU data object along with SNP IDs, 8   with.ADdx.ANOVA tensor. These NIMs were extracted and used to generate association heatmaps (SNPs by diagnosis and Imaging biomarker by diagnosis).

| Neuroimaging-geneticsassociation analytics protocol
More details are provided in the Appendix S1 (Methods).

| Networkanalysisbetween29genomicand 200 neuroimaging markers
The WGCNA (Weighted graph correlation network analysis) R package (version 1.68) 8 was used to perform network analysis using 29 genomic and 200 neuroimaging markers. The aim of this analysis was to connect the NIMs having similar patterns observed from the GMs. WGCNA was originally developed to find the network of co-expressed genes based on their expression patterns in multiple conditions. Specifically, for each of three cohorts, such as AD, MCI, and HC, the values of the NIMs and GMs were measured, and correlation coefficient values between the two kinds of markers were calculated. The correlation coefficient values constitute the correlation coefficient matrix (CCM) between the two kinds of markers. The CCM was then converted to matrices representing an unsigned adjacency matrix (using soft thresholding of 7) and a topology overlap matrix (using a score threshold of 0.05).
The NIMs with similar genetic patterns were predicted by WGCNA using the converted matrices. The predicted networks for each of the AD, MCI, and HC cohort were visualized using Cytoscape (version 3.7.1). 46

| StatisticallysignificantMCI/HCandAD/HC odds ratios
Using multinomial linear modeling of diagnosis, we studied the associations between the three individual cohorts (HC, MCI and AD dementia). The differences of the 200 NIMs and 29 SNPs between HC, MCI, and AD dementia cohorts. The results of a 3-way ANOVA (ROI, Dx, SNP) may be less interpretable compared to a multinomial linear modeling (Outcome = Dx). We computed the odds ratios (ORs) and relative risks (RRs) for AD and MCI, relative to HC.
The MCI and AD effects quantified the metrics "relative to HC." These represent extensions of the binary outcome in logistic regression, but reflect 3 categorical outcomes (HC, MCI, AD), which may also be analyzed via more general multi-nominal linear modeling. In general, to assess statistical significance a customary false-positive rate of α = 0.05 may be used for many different tests.
However, in many GWAS studies, it's common that a correction for multiple comparison (e.g., false discovery rate, family wise error rate), or other strategies are used to control the false positive rate of significance. 47

| ADgeneticandimagingmarkersforthe network analysis
We extracted the AD-related genotypes and NIMs for the network analysis. Genotype dosage data for AD-associated SNPs were extracted from the imputed ADNI genetics dosage data using sample-major additive (0/1/2) coding. The genotypes, SNPs, chromosomes, and positions can be found in 8 Table 2.
To streamline the analyses, we chose the top 200 NIMs corresponding to the lowest p-values for the discrimination between HC and AD subjects to correlate with 29 GMs, Table S1.

| Associationanalysisamonggeneticand neuroimaging biomarkers
In-house R-scripts were developed to generate three heatmaps for each of the three cohorts. These plots represent the association analyses (SNPs*diagnosis, NIMs*diagnosis), see Figure 1 and Table S2.  Table 3A.
Additional details are provided in Figure 1A and Table S2A.  Table 3B.
Additional details are provided in Figure 1B and Table S2B. NIMs were significantly associated with the 11 markers among the 29 GMs at the level of 'p < 0.01'. There were three ROIs (1 thickness, 1 volume, and 1 pct) which were associated with three GMs at the level of 'p < 0.001', Table 3C.
Additional details are provided in Figure 1C and Table S2C.

| Networkanalysis
The interactions between the 29 genes and the 200 NIMs were also explored using network analysis for each of the three cohorts, HC, MCI and AD. We did not find strong network patterns solely within the 29 genes themselves, see Figure 2. However, we found strong networking patterns within the HC group. Three types of NIMs were divided into thickness, volume, and proportion of white to gray matter (pct). These three measurement groups were networking separately under the control of the 29 genes in the HC group, regardless of the p-values in association analysis which is described earlier in the association analysis section.
In the MCI cohort, the 200 NIMs were not divided as in the HC subjects. All the NIMs intermingled without a clear subgroup clustering. These three measurement groups were networking diffusely without any grouping under the control of the 29 genes, regardless of the p-values in association analysis, which was described earlier in the association analysis section, see Figure 2B.
In the AD cohort, we found some patterns between the 29 genes and the 200 NIMs. Three types of NIMs were divided into thickness, volume, and pct.
These three measurement groups independently tracked neuroimaging-genetic associations for the 29 genes in the AD subjects. The protocol for estimating the corresponding association pvalues was described earlier in the association analysis section.
In this model, the 29 SNPs did not appear to be significantly associated with AD, relative to HC, see Figure S1.  Figure S3 shows us predicted diagnosis depiction.

| SNPselection
There are two options for identifying GMs associated with AD.  8 Genotype dosage data for AD-associated SNPs were extracted from the imputed ADNI genetics dosage data using sample-major additive coding. Such 29 SNPs were selected, representing a subset of independent SNPs found to be associated with AD in the recent AD meta-GWAS study.

| Neuroimaging-geneticsassociation(priorto network analysis)
Several prior neuroimaging-genetics GWAS studies analyzed and reported specific NIMs as QTs. 5,11,13 There are some similarities with our study, however previous reports do not employ network analysis.
In our HC cohort, CD2AP (rs9381563), MS4A6A (rs2081545) and SUZ12P1 (rs8093731) were associated with thickness measures, and EPHA1 (rs11763230) and CASS4 (rs6014724) were associated with volume measures, especially in the occipital lobe. In general, however, cortical thickness measures were more significantly associated with the GMs compared to volume measures.

CD2AP (CD2 [cluster of differentiation 2] associated protein)
loss of function is linked to enhanced Aβ metabolism, tau-induced neurotoxicity, abnormal neurite structure modulation and reduced blood-brain barrier integrity. 49 CD2AP is expressed in both neuron and microglia in the brain and postulated to be involved in immune system regulation. 46 To the best of our knowledge, this is the first report of neuroimaging association of CD2AP gene to date. MS4A6A ABCA7 gene is strongly expressed in the hippocampus subfield CA1 61 and showed a significant association with hippocampal atrophy 56 and gray matter density. 62 Like this, ABCA7 is quite well known gene that is associated with some NIMs in AD.
In the AD cohort, some metrics of thickness, pct and volume (3rd ventricle) were significantly associated with the GMs. ECHDC3 (rs11257238) was significantly associated with a thickness measure.
The ECHDC3 (enoyl CoA hydratase domain containing 3) is known to be involved in fatty acid biosynthesis in mitochondria. 63 Desikan et al. reported that the gene expression of ECHDC3 was changed in opposite directions in the AD cohort, 64 and in our study, the gene was significantly associated with a thickness measure in the AD cohort. This gene may need to be further investigated in the future neuroimaging genetics studies. The CASS4 was significantly associated with a pct measure in the AD cohort, which was different from the HC and MCI cohorts. The SUZ12P1 was significantly associated with the 3rd ventricle, however, ventricular volume may also inversely reflect cortical volume changes. Finally, SUZ12P1 and CASS4 were significantly and differentially associated with some ROIs in the HC, MCI, and AD cohorts.
Cortical thickness is a signature marker for memory functioning across the adult lifespan. Among asymptomatic healthy individuals, the degree of cortical thinning predicts progression to clinical AD. 65,66 Moreover, our study suggests that cortical thickness may be an important measure of early detection of cognitive impairment progressing from HC to MCi and eventually leading to AD pathogenesis. [67][68][69] Previous findings suggest that (1) cortical thickness and cortical surface area are independent, both globally and regionally; and (2) gray matter volume is tracked by both metrics, even though cortical thickness is less influential than surface area. 70,71 We found meaningful associations between 200 NIMs and 29 GMs for cortical thickness and regional ROI volume measures. However, cortical surface area appears as a less sensitive measure in individualized analysis within each cohort. Further studies are necessary to determine the intricate relationships between regional morphometry metrics, such as cortical volume, surface area, and thickness, specific genotypic markers, such as SNPs, and different clinical phenotypes.
On the other hand, our research finding suggests that cortical thickness may represent an important factor for tracking and discriminating subtle differences between HC and MCI cohorts. Yet, the importance of this association may not yet be extended to the AD cohort. 5 Our current research provides evidence that cortical thickness measures are important early on, prior to dementia onset, but their importance may taper off after dementia diagnosis.

| Networkanalysis
To obtain the most reliable networks of 200 NIMs, threshold values for the network prediction were carefully chosen to meet the scale-free topology criterion as used in other recent studies based on biological network analysis. [74][75][76] No specific network patterns were identified jointly within all the 29 genes, which was contrary to our initial expectations. Hence, we performed network analyses using individual GMs.
We found specific network patterns within the respective HC and AD groups. The NIMs were divided into three thickness, volume and pct metrics under the control of 29 GMs. In the MCI cohort, transient intermingled stages without the patterning were traced and compared with the HC and AD cohorts. We found some specific ADproper networking patterns (HC and MCI patterns were subtracted) as well between the GMs and the NIMs. In general, the two types of NIMs were divided into thickness and volume measures. Pct measures were not highly impactful in this AD-proper networking, Figure 2D.
This may imply that in the demented AD cohort, pct measures are not homogeneously tracking the observed disease pathogenesis.

| StatisticallysignificantORsandRRsinMCI/ HCandAD/HCcomparisons
We also calculated the ORs to study the relationships of AD and MCI to HC in terms of AD pathogenesis. In AD RR, rh_G_occipital_mid-dle_foldind (cortical folding index) is protective, but in MCI RR, it appears detrimental rather than protective. We aimed to identify statistically significant associations between regional, diagnostic, and genetic effects, using multinomial linear modeling and network analysis.
Basically, we suggest that similar genetic and epigenetic mechanisms continue to impact the structure and function of the brain throughout life. Early on, both genetics and experience guide neocortical and brain patterning, and these mechanisms continue to impact the maintenance of cortical areas and their boundaries as well as physiological area function throughout adulthood. Late in life, similar genetic mechanisms may be involved in the breakdown of brain microstructure as in early development, can either advance or ameliorate the deleterious effects of aging. 5 These ideas can be reflected in the pathogenesis of AD as well.

| Limitationsandfuturedirections
Potential limitations of this study reflect the relatively small sample size to analyze genetic influence on NIMs. The ADNI sample was not collected under a perfect epidemiological ascertainment strategy and the sample size was relatively small for a GWAS study, which may affect the generalizability of the findings. Because of our restricted statistical power, we were forced to constrain our analysis to SNPs that have been previously reported in Janssen et al. We used 29 SNPs from the Janssen et al. that are derived from the other diagnosis system from ADNI research subjects, which may also affect the generalizability of the findings.
For the neuroimaging genetics study, we used imputation tools to unify several separated ADNI data and to increase as much as possible the sample size of the genetic and neuroimaging ADNI data.
This allowed us to aggregate the ADNI data and generate computable multimodal data objects including homogenous NIMs and GMs.
We did not manually inspect the brain scans of all participants (this is done by ADNI QC), to avoid potential subjective rater bias for location, size, or etiology of MRI-evident infarcts in the QC protocol.
So, there is a potential that minor WMHI effects may play a role in our analyses.
The sample only contained mild AD patients (CDR = 1), a relatively narrow range of illness, and is thus not fully representative of the HC-MCI-AD spectrum. At this point in time, ADNI does not collect gene expression/RNAseq data, and we could not complete a full network analysis in terms of neuroimaging genetics due to lack of available data.
Despite these challenges, the results are encouraging, and the proposed analytic framework appears to have a potential for enabling the discovery and localization of phenotypic imaging-genetics associations. We believe that imaging-genetics techniques offer important clues for the formulation of advanced methods of early detection, monitoring, and treatment of dementia. This neuroimaging-genetics study provides valuable clues to dementia onset and the prospective pathogenic trajectory. Our results are promising for untangling the intricate interrelations between brain anatomy and genetic phenotypes. Network analysis using neuroimaging measures and genotypic biomarkers provides cues to the structure of various deep brain-networks and assists with interpreting structural imaging-genomics association with disease. Further studies are necessary to reveal any specific mechanistic associations between GMs and NIMs and discover triggers or buffers of complex AD pathogenic traits.