MAPT haplotype–stratified GWAS reveals differential association for AD risk variants

Abstract Introduction MAPT H1 haplotype is implicated as a risk factor for neurodegenerative diseases including Alzheimer's disease (AD). Methods Using Alzheimer's Disease Genetics Consortium (ADGC) genome‐wide association study (GWAS) data (n = 18,841), we conducted a MAPT H1/H2 haplotype–stratified association to discover MAPT haplotype–specific AD risk loci. Results We identified 11 loci—5 in H2‐non‐carriers and 6 in H2‐carriers—although none of the MAPT haplotype–specific associations achieved genome‐wide significance. The most significant H2 non‐carrier–specific association was with a NECTIN2 intronic (P = 1.33E‐07) variant, and that for H2 carriers was near NKX6‐1 (P = 1.99E‐06). The GABRG2 locus had the strongest epistasis with MAPT H1/H2 variant rs8070723 (P = 3.91E‐06). Eight of the 12 genes at these loci had transcriptome‐wide significant differential expression in AD versus control temporal cortex (q < 0.05). Six genes were members of the brain transcriptional co‐expression network implicated in “synaptic transmission” (P = 9.85E‐59), which is also enriched for neuronal genes (P = 1.0E‐164), including MAPT. Discussion This stratified GWAS identified loci that may confer AD risk in a MAPT haplotype–specific manner. This approach may preferentially enrich for neuronal genes implicated in synaptic transmission.


INTRODUCTION
Tauopathies, a class of neurodegenerative disorders, are characterized by neurofibrillary tangles (NFTs) in the brain due to pathological aggregation of hyperphosphorylated microtubule-associated protein tau The FTDP-17 splice-site mutations within MAPT demonstrated that an imbalance in the ratio of 3R and 4R tau isoforms is sufficient to cause disease. [2][3][4] Further association studies revealed that the locus can be divided into two major haplotypes: H1 and H2. MAPT falls within the largest known block of linkage disequilibrium (LD) in the human genome, spanning ≈1. 8 Mb. There is a 900 kb inversion of the H2 haplotype with respect to the H1 haplotype, covering a region encompassing several genes, including MAPT, IMP5, CRHR1, and NSF.
The inversion results in a reduced recombination between the inverted H2 and non-inverted H1 haplotypes.
The common MAPT haplotype H1 shows robust association with risk for the primary tauopathies PSP 5 and CBD, 6 as well as Parkinson disease (PD), which is not considered as a tauopathy. 7 MAPT H1 haplotype-tagging single-nucleotide polymorphisms (SNPs) were identified among the top PSP 8 and PD genome-wide association study (GWAS) 9 signals. In addition, MAPT H1 haplotype shows considerable variation 10,11 and leads to H1-subhaplotypes, where H1c, has been implicated in the risk of PSP, CBD, AD, and PD. 12,13 MAPT H2 haplotype has been associated with reduced risk for several neurodegenerative disorders. 14,15 Although MAPT is a compelling candidate for neurodegenerative disease susceptibility, evidence of association of AD with the MAPT H1 and H2 haplotypes have produced equivocal results. 12,16,17 This may in part be due to limited sample sizes, and therefore limited power for most MAPT haplotype association studies in AD. In a large study from Genetic and Environmental Risk for Alzheimer's Disease (GERAD1) consortium, 18 the MAPT H2 haplotype-tagging variant was found to have association with reduced AD risk. In a study of >20,000 individuals from Mayo Clinic and the Alzheimer's Disease Genetics Consortium (ADGC), we identified associations with both reduced AD risk and reduced brain MAPT levels with the H2 haplotype. 14  • Many of the genes at these loci are differentially expressed in AD brains.
• The MAPT haplotype-stratified approach identified genes in synaptic transmission networks.

RESEARCH IN CONTEXT
1. Systemic review: Comprehensive review of the literature shows that the microtubule-associated protein tau gene (MAPT) is a strong candidate for neurodegenerative disease susceptibility. The MAPT H2 haplotype is associated with lower Alzheimer's disease (AD) risk in large cohorts and lower brain MAPT levels.
2. Interpretation: We hypothesized that AD risk variants exhibit MAPT haplotype-dependent association. Through haplotype-stratified association analyses using data from the Alzheimer's Disease Genetics Consortium (ADGC) on 18,841 participants, we identified 11 loci with MAPT H1or H2-specific AD risk association. Eight genes at these loci had significant differential expression in AD versus control brains. Six genes were members of the neuronalenriched brain transcriptional co-expression network implicated in synaptic transmission.
3. Future directions: Replication of MAPT haplotypestratified associations should be sought in larger cohorts.
Candidate genes from this study should be evaluated for the presence of functional variants that may influence tau-related outcomes. Emerging larger cohorts with multiomics data and generation of more complex model systems will enable these studies.
In the current study, we sought to further elucidate the role of MAPT H1 and H2 haplotypes in AD susceptibility by leveraging the genomewide genotype data available from the sizable ADGC case-control series. Using haplotype-stratified analyses, we tested the hypothesis that AD risk variants exhibit MAPT haplotype-dependent association and may therefore potentially identify novel AD risk variants with implications for functional pathways. Analysis of a stratum with a more homogeneous AD risk profile with respect to MAPT H1/H2 haplotype may help uncover loci that have differential influence on AD risk in a MAPT context-specific manner. For example, given the association of MAPT H2 with lower brain MAPT levels, it is plausible that those loci with MAPT H2-specific associations harbor genes that influence neurodegeneration via pathways that are not dependent on elevated tau.
In contrast, AD risk associations in H2 non-carriers may enrich for loci that confer risk in a tau level-dependent fashion.
Our approach herein is akin to pursuing GWAS in an apolipoprotein E gene (APOE)-stratified fashion. 19 Although MAPT haplotypes tested to date in the literature clearly have smaller effect sizes than that of APOE genotypes for AD risk, it is nonetheless worthwhile to pursue this MAPT haplotype-stratified analysis not only because of its potential to identify novel loci but also because of the plethora of data implicating tau in AD in functional studies. 20 In this study, we evaluated known International Genomics of Alzheimer's Project (IGAP) AD 21 risk loci in a MAPT haplotype-stratified analysis, which did not reveal evidence of MAPT haplotype-specific associations. We also identified novel AD risk loci with association only in MAPT H2 carriers (six loci) or H2-non-carriers (five loci). We characterize genes near both the known and the new loci for their expression levels and co-expression networks in a brain transcriptome data set of AD and control temporal cortex. 22,23 Our findings, which require replication in larger cohorts, suggest that MAPT haplotype-stratified GWAS may identify novel loci, and that genes at these loci are expressed predominantly within neuron-enriched networks implicated in synaptic transmission.

Study populations
The ADGC data were used for this study. Subjects available through the ADGC have been described previously and are available through ftps from the UPENN server (alois.med.upenn.edu). [24][25][26][27] The data set included all the covariates required for the analysis and all actual and imputed genotypes. Post-quality control (post-QC) data for both the actual and imputed genotypes and designations for all the sub-cohorts included in the ADGC data were obtained. The demographics detailing each cohort and stratified group are described in Table S1. The cohort for the expression analysis was the Mayo Clinic RNAseq data set. 22 Detailed methods are provided in Supplementary Methods.

AD risk association analysis
Variants were evaluated for association with AD using multivariable logistic regression implemented in PLINK. 28 Both joint (full data set of 21 cohorts analyzed jointly, adjusting for cohort) and meta-(separate cohorts) analyses were performed. For the meta-analysis, a random effects method was adopted due to presence of heterogeneity, I 2 > 25. 29 An additive model for the minor alleles determined in the unstratified data set was applied with the covariates age, sex, and PC1-3 (principal components 1-3) used throughout all models. A second model using the additional APOE covariate in the joint and meta-analyses was also evaluated. Two IGAP loci variants rs4147929 and rs9331896 were filtered out of the original data set due to the QC procedures described previously. 27 They were evaluated separately for the joint analyses using the same method above.

Epistasis analysis
SNP-SNP interactions of epistasis between each of the 3,067,502 SNPs and the H2 tagging variant rs8070723-G were conducted. Two models were evaluated for the H2 tagging variant, a carrier model (H1H1 and H1H2+H2H2) and a dosage model (H1H1, H1H2, and H2H2). The analysis was performed by creating a distance matrix in PLINK between each SNP and rs8070723-G. Two general linear models (with SNPx rs8070723-G interaction and without interaction) were executed using age, sex, ADGC cohort, and PC1-3 as covariates followed by an analysis of variance (ANOVA) to assess the significance between the models using the chi-square method as implemented in R.

Gene expression analyses
Differential gene expression and co-expression network analyses were conducted as previously published. 23,32 For each gene, multiple linear regression was performed in which normalized gene expression was the dependent variable, diagnosis (AD vs control) was the independent variable of primary interest and sex, flowcell, age at death, RNA integrity number (RIN), and center from which the samples were obtained were the covariates. Weighted Gene Co-Expression Network Analysis (WGCNA) was utilized to identify brain co-expression networks and test their associations with AD as we reported previously. 23

Visuals
The figures were generated using the lattice 33 and metafor packages in R and Inkscape (www.inkscape.org).

MAPT haplotype-specific association analysis at known AD risk loci
Using genome-wide genotype data from 21 cohorts within ADGC, we tested the hypothesis that AD risk variants exhibit MAPT haplotypespecific association. Following QC measures, approximately 3 million variants with a minor allele frequency (MAF) ≥0.02, and all index variants identified by the IGAP consortium 21 were retained for analysis and evaluated for MAPT haplotype-specific association.
GWAS analyses with AD were performed using joint and metaanalyses. There was no evidence of population stratification based on the quantile-quantile plots (QQ plots) ( Figure S1) and the genomic inflation factors of 1.04, 1.04, and 1.01 for the unstratified, H2 non-carrier, and H2 carrier joint analyses, respectively. The joint and meta-analyses yielded similar results with respect to genomic inflation.
Likewise, the addition of APOE as a covariate did not significantly alter the results. We adopted the joint analysis approach without APOE covariate as the primary model.
We first evaluated the previously reported IGAP 21 AD risk variants to determine if they exhibit MAPT haplotype-specific association.
As expected, the unstratified analysis results were similar to those reported in the IGAP study, albeit with reduced significance due to the smaller cohort size ( Figure S2,

Genome-wide MAPT haplotype-specific AD risk association analysis
To identify any additional AD risk variants with MAPT haplotypespecific association, we evaluated the genome-wide results for the unstratified, H2 non-carrier and H2 carrier groups ( Figure 1). We tested for significance of MAPT haplotype-specific associations by genome-wide epistasis analysis with rs8070723 ( Table 2). We defined loci with MAPT haplotype-specific AD risk associations as being discordant. To be classified as discordant, the following criteria had to be met: Discordant locus (1) has AD risk association P value of < 1E-05 in one of the stratified analysis, but statistically insignificant in the other one (P > 5E-02); (2) has nominally significant epistasis interaction with rs8070723 (P < 0.05).
We identified five loci in the H2 non-carriers and six in the H2 carriers with discordant MAPT haplotype-specific AD risk associations ( Figure 1, Table 2). These loci (nearest genes at loci) are as follows: In the H2 non-carriers: chr4 (TBC1D9), chr4 (GAL-  Figure S3. We checked the regional association plots of the discordant loci to determine whether any of them represented known IGAP AD risk loci ( Figure 2 Unstratified contains all subjects from the 21 ADGC cohorts used for analysis. The H2 tagging variant rs8070723 was used to stratify study participants into H2 non-carriers (H1H1) and H2 carriers (H1H2+H2H2).
g Joint analysis performed with logistic regression in PLINK using as covariates age, sex, cohort, and PC1-3. Rs4147929 is missing in the GSK cohort. N, number of subjects analyzed; OR, odds ratio for AD risk for the minor allele; 95% CI, 95% confidence interval, P-value.
h Meta-analysis performed with logistic regression analysis in PLINK using the random effects model, and using the covariates age, sex, and PC1-3. Meta-analysis could not be performed for rs4147929 and rs9331896. N, number of cohorts analyzed; OR, odds ratio, P-value.
i Epistasis analyses performed in R. All cohorts were combined and stratified into two groups by the H2 tagging variant rs8070723: H2 non-carriers (H1H1) and H2 carriers (H1H2+H2H2). Results from the carrier model are shown. By definition, all discordant loci had nominally significant epistasis P values, although none reached genome-wide significance ( Table 2).
Considering the 21 IGAP and 11 discordant loci evaluated, and applying a study-wide epistasis P -value of 1.52E-3 (Bonferroni P = 0.05/33), there was one discordant SNP with MAPT H2 non-carrier-specific association and two discordant SNPs in the MAPT H2-carrier group.
The SNP with the smallest epistasis P value and MAPT H2 non-carrierspecific association is rs4354897 on chromosome 15 (Table 2)   (1.14-1. e Joint analysis performed with logistic regression in PLINK using as covariates age, sex, cohort and PC1-3. N, number of subjects analyzed; OR, odds ratio; 95% CI, 95% confidence interval, P-value.
f Meta-analysis of logistic regression analysis in PLINK using the random effects model, and using the covariates age, sex and PC1-3. N, number of cohorts analyzed; OR, odds ratio, P-value.
g Epistasis analyses performed in R. All cohorts were combined and stratified into two groups by the H2 tagging variant rs8070723: H2 non-carriers (H1H1) and H2 carriers (H1H2+H2H2). Results from the carrier model are shown.
F I G U R E 2 Regional association plots of discordant MAPT haplotype-stratified association results: The figures are shown for the 11 loci depicted in Table 2 and reflect the results of haplotype-stratified joint association analyses without the APOE covariate. Discordant loci results with significance in the (A) H2 non-carriers or (B) H2 carriers.
To determine the joint effect of the discordant SNPs and MAPT haplotypes on AD risk, we performed a bivariate analysis (Table S2).
The MAPT H2 non-carriers with the SNP major homozygote genotypes were designated as the reference. We tested the AD risk association of each SNP genotype in the MAPT H2-carrier or H2 non-carrier background against this reference. The bivariate analysis results are consistent with their corresponding MAPT haplotype-specific associations and depict the joint effect of each SNP genotype and the MAPT haplotype on AD risk.

Brain expression analyses of MAPT-stratified AD risk association loci genes
We characterized the brain expression patterns of the genes at the discordant MAPT-stratified association loci (Table 2) using the temporal cortex (TCX) RNAseq transcriptome data from Mayo Clinic. 22,23,32 Of the 17 genes at the 11 discordant loci, 12 were present in this data set (Table S3). We evaluated these genes for differential expression (DE) between neuropathologic AD and control TCX RNAseq data. In addition, we determined the brain gene co-expression networks, 34 which harbor these genes and annotated these networks for their enriched gene ontology (GO) biological processes 35 and brain cell types, as described previously. 23,32 Eight of the 12 genes evaluated had significantly different expression in AD versus control TCX (Table S3).
The genes with transcriptome-wide significant differential expression We performed the same analyses also for the known AD risk loci genes (Table S3). Eight of the 17 IGAP loci genes with brain expression data had significant differential expression, both at the gene (q < To determine whether any of the MAPT haplotype-specific AD risk SNPs influenced brain expression levels of MAPT or the "synaptic transmission" co-expression module TCX1, which also harbors MAPT, we performed expression quantitative trait loci (eQTL) and module QTL (modQTL) analyses, respectively, as described previously. 14,32,37,38 None of the MAPT haplotype-specific AD risk SNPs had significant associations with temporal cortex MAPT levels or the "synaptic transmission" module eigengene (data not shown). We conclude that these MAPT haplotype-specific loci are not likely to confer AD risk through their influence on brain gene expression of MAPT or synaptic transmission network genes.

DISCUSSION
Despite significant progress in identifying genetic risk factors and the increased understanding in Alzheimer's disease (AD) etiology, the ability to develop effective preventions or cures continues to remain elusive. Novel approaches to analyzing available multiscale genomic and phenotypic data will provide further insights into the complexity of AD and provide mechanisms to foster the development of precision medicine.
In this study we sought to evaluate available genome data by performing a stratified analytic approach. Stratified methods based on sex [39][40][41] and APOE 19  MAPT haplotype-dependent association. We tested previously identified AD risk loci 21 to determine whether they have differential associations in a MAPT haplotype context-dependent manner. We also extended this analysis genome-wide to determine if this approach may identify novel AD risk variants.
We found that the index AD risk variants reported previously had similar directions of associations in both the MAPT H2 non-carrier and H2 carrier analyses. Epistasis analysis with these and the MAPT H1/H2 haplotype tagging variants revealed no evidence of differential association (P > 0.05) for all but two AD risk loci. Even though CD2AP-rs10948363 and ZCWPW1/PILRB-rs1476679 had nominally significant epistasis (P = 0.035 and 0.022, respectively), the estimated effects of these variants were largely overlapping in the MAPT H2 carriers and non-carriers. These findings are not surprising given that the loci that rise to significance in the overall GWAS are likely to have a more consistent effect across stratified groups.
In contrast, stratified analysis may uncover novel loci with groupspecific associations that may be missed in the combined cohort.
Although we did not identify any MAPT haplotype-specific associ-  14,15,18 The preferential protection of rs11665676-T in MAPT H2 non-carriers may be due to the fact that in the presence of the protective MAPT H2 haplotype, any further protection conferred by this variant may be negligible. This may explain the lack of association of rs11665676-T with lower AD risk in MAPT H2-carriers.
The discordant rs11665676 variant resides within an intron of NECTIN2 (aka PVRL2), which is within a LD region with BCAM and in proximity to the TOMM40-APOE-APOC1 locus. 42 It has been shown previously that the LD structure of the polymorphisms across these five genes displayed heterogeneity between AD and control individuals, suggesting that the genes within this region in addition to APOE may play a role in AD risk. 42,43 Indeed, a highly polymorphic variant of TOMM40 (poly-T variant) was found to associate with AD risk and its endophenotypes independent of APOE in some studies. 44 Given the complexity of this region on chromosome 19, including LD across multiple genes, plentiful polymorphisms, and the strong APOE 2/ 3/ 4 effect on AD risk, alternative approaches focused on haplotype analysis of this region are proposed to uncover novel variants that influence AD independent of APOE. 45 Our analysis of stratifying samples according to specific genotypic/haplotypic backgrounds provides another approach in the discovery of polymorphisms that may influence AD risk under a specific genomic context. Our approach identified a polymorphism in NECTIN2 (PVRL2), which is enriched in APOE 3 carriers and which has differential protective association in MAPT H2 non-carriers.
This finding suggests a biological link between NECTIN2 and/or APOE In a previous APOE-stratified analysis, 19 a variant in the MAPT region, rs2732703-G, which is more common in H2 carriers, was found to confer greater protection from AD in APOE 4 negative individuals.
This finding is different and independent of our report, and suggests that variability at the MAPT locus influences APOE association with AD risk, whereas our results indicate that variability at the APOE locus has distinct AD risk association on different MAPT haplotype backgrounds. Both findings support the notion of heterogeneity at both 2), also known as poliovirus receptor-related 2 (and formerly as herpesvirus entry mediator B, HVEB), encodes a plasma membrane glycoprotein that has been implicated in a multitude of central nervous system (CNS) functions. 43 NECTIN2 is involved in adherens junction, which is important to maintain blood-brain barrier and to prevent the spread of viral infections. In our brain expression data 22,23,32 analyzed herein, we determined NECTIN2 to be significantly elevated in AD TCX, and to reside in a co-expression module enriched for "defense response" GO biological process and microglia-enriched genes. These findings support a role for this gene in innate immune pathways.
Our findings along with prior association of another NECTIN2 variant (rs6859) with AD risk in African Americans independent of APOE, 46 merit further evaluation of this gene as a plausible AD gene.
In addition to the NECTIN2 variant at the APOE locus, MAPTstratified analysis revealed one other discordant association in a known AD risk locus, which was PICALM intronic SNP rs140869727 that revealed increased risk in MAPT H2-carriers. The minor A allele of rs140869727 has frequency (MAF) of 0.17 and is rarer than the PICALM locus index IGAP SNP rs10792832, which has a MAF of 0.36. The latter did not have differential MAPT haplotype-stratified association, whereas rs140869727 had epistasis (P = 4.51E-03). We concluded that the discordant rarer SNP may be tagging a PICALM variant, which confers greater risk of AD in MAPT H2 carriers. PICALM was found to associate with both 3R and 4R tau inclusions in AD and primary tauopathies, and soluble PICALM levels were inversely correlated with phosphotau, 47 suggesting a biological link between this protein involved in clathrin-mediated endocytosis and tau.
We identified nine discordant loci that were not previously identified in AD risk GWAS, including the largest recent studies. 48,49 The four novel H2 non-carrier-specific associations were near TBC1D9 (chr4), ADAMTSL3 encodes a glycoprotein that localizes to the extracellular matrix, belongs to a family of metalloproteases, and is proposed to be a candidate gene for schizophrenia, with proposed function in synaptogenesis. 50 Of interest, another H2 non-carrier-specific asso- and a nectin-like protein (CADM2 = NECL3 = SYNACM2) as candidates is noteworthy. CADM2 was also identified as a locus for habitual physical activity, along with APOE, 60 and was also suggested as a gene that may link obesity with psychiatric traits. 61 The three other candidate genes at the AD risk loci identified in MAPT H2-carriers-C11orf21, STK32B, and NKX6-1-were also implicated in CNS diseases or function. C11orf21 has an intronic variant rs77007065-A, which confers AD protection in MAPT H2 carriers (0.68, CI = 0.58 to 0.81, P = 9.78E-06) and is one of the most discordant SNPs (epistasis P = 5.62E-06). This variant is also upstream of TSPAN32, which together with C11orf21 resides in a region of differential methylation in autistic brain samples. 62 STK32B is a serine/threonine kinase and resides at a locus previously identified in a GWAS for essential tremor. 63 The promoter region of this gene is differentially methylated in blood samples from adolescents with generalized anxiety disorder. 64 Finally, NKX6-1, which is a transcription factor, was found to be involved in midbrain dopaminergic neuron differentiation, 65 in addition to its role in the differentiation of pancreatic ß islet cells. 66 Whether these are the genes that harbor functional variants that influence AD risk in a MAPT haplotype-dependent manner and their biological interaction with tau-related pathways remains to be established.
In our study, we also performed a systematic evaluation of all of the candidate genes at the discordant AD risk loci for their expression in AD versus control temporal cortex (TCX), 22,23,32 their membership in brain gene co-expression networks identified in these samples, and annotation of these networks for their enriched biological processes and CNS cell types. For these analyses, we utilized the Mayo Clinic Brain RNAseq data generated by our group, and implemented approaches as previously described. 22,23,32 We also analyzed the candidate genes at the known IGAP AD risk loci 21  Nevertheless, the concurrent presence of GWAS candidate genes within networks that are enriched in processes known to be perturbed in the disease process (such as "immune response," "synaptic transmission," "axon ensheathment") provides further strength for the candidacy of these genes and information about the pathways with which they are likely to be involved. The presence of half of the discordant loci in "synaptic transmission" networks suggests that the MAPT haplotype stratified approach may be preferentially identifying neuronal genes that are involved in this crucial process in a MAPT haplotypedependent manner. This finding is congruent with known and proposed roles of tau in synaptic transmission or its disruption in AD. 67 In comparison, the un-stratified GWAS appears to uncover genes that pertain to a wider spectrum of pathways and cellular processes that may be due to the lack of the dependency on MAPT haplotype context.
Because the transcriptome data was obtained in bulk brain tissue in a region affected with AD neuropathology, the observed transcriptional differences between AD and controls may reflect cell population changes. 22 Despite this caveat, we and others have successfully utilized bulk brain transcriptome data to identify transcriptional networks that associate with neurodegenerative diseases and their endophenotypes. 23,32,42,43,45 Many of these networks are enriched in pathways and genes that have been implicated previously in these diseases through independent data including genetic associations. 32,36,42 This suggests that integrative analysis of transcriptional networks and disease association data can provide cross-validation for the genes.
This approach also provides transcriptional context for the candidate genes discovered from disease GWAS as demonstrated here.
In summary, we performed a MAPT H1/H2 haplotype-stratified association in the ADGC GWAS data and identified 11 loci with evi- There are several limitations to our study. Notwithstanding their novelty, the MAPT H1/H2 haplotype-stratified association results should be interpreted with caution due to falling short of genomewide significance (P < 5.0E-08), as they may represent false-positive findings. It will be important to apply this approach in larger available GWAS data and seek confirmation. Given that MAPT H2 haplotype is rarer, our MAPT H2 carriers were smaller in size (n = 7360) than MAPT H2 non-carriers (n = 11,481). This may explain the presence of two loci that approached genome-wide significance in the MAPT H2 noncarriers, whereas the strongest association remained at P = 1.99E-06 in the MAPT H2 carriers. We also acknowledge that our MAPT H1/H2 haplotype definition was based on the tagging variant rs8070723 and that the H1 haplotype, which has considerable variation, 10,11 can be divided into additional sub-haplotypes. Future studies utilizing whole genome sequencing (WGS) can enable more accurate assignment of haplotypes, although sub-haplotypic stratification would require even greater sample sizes. We discovered that many of the candidate genes at the discordant AD risk loci are differentially expressed in AD TCX and reside in the "synaptic transmission" co-expression network, which also harbors MAPT. Despite their intriguing biological implications, it is possible that these congruent genomic and transcriptomic findings are coincidental. Definitive determination of biological interactions between the discordant loci genes with MAPT requires detailed studies in model systems, which is beyond the scope of this work. Our findings provide testable hypotheses for such functional studies. Finally, our brain transcriptome data are driven from bulk tissue, where the gene expression findings may simply reflect cell population changes and where biologically important differential expression results in rarer cell types may be obscured. It will be important to evaluate brain cell-type specific expression patterns of the genes nominated in this study in the single-nucleus and single-cell transcriptome data from AD and control brains, once sizable data sets become available.
Our study represents an alternative approach in leveraging available GWAS data for discovery of loci and genes that may confer AD risk in a MAPT context-dependent manner. Integrative utilization of independent genomic and transcriptomic data provide cross-validation for our findings. The candidate genes that emerge from this study should be evaluated for the presence of functional variants that may influence tau-related outcomes in model systems or human cohorts. Emerging larger cohorts with multi-omics data and generation of more complex model systems should enable these future studies.

ACKNOWLEDGMENTS
We thank the patients and their families for their participation, without which these studies would not have been possible. Research.

DECLARATIONS OF INTEREST
None.

ADGC ACKNOWLEDGMENTS
The National Institutes of Health, National Institute on Aging (NIH-NIA) supported this work through the following grants: ADGC,