Variants in PPP2R2B and IGF2BP3 are associated with higher tau deposition

Abstract Tau deposition is a key biological feature of Alzheimer’s disease that is closely related to cognitive impairment. However, it remains poorly understood why certain individuals may be more susceptible to tau deposition while others are more resistant. The recent availability of in vivo assessment of tau burden through positron emission tomography provides an opportunity to test the hypothesis that common genetic variants may influence tau deposition. We performed a genome-wide association study of tau-positron emission tomography on a sample of 754 individuals over age 50 (mean age 72.4 years, 54.6% men, 87.6% cognitively unimpaired) from the population-based Mayo Clinic Study of Aging. Linear regression was performed to test nucleotide polymorphism associations with AV-1451 (18F-flortaucipir) tau-positron emission tomography burden in an Alzheimer’s-signature composite region of interest, using an additive genetic model and covarying for age, sex and genetic principal components. Genome-wide significant associations with higher tau were identified for rs76752255 (P = 9.91 × 10−9, β = 0.20) in the tau phosphorylation regulatory gene PPP2R2B (protein phosphatase 2 regulatory subunit B) and for rs117402302 (P  = 4.00 × 10−8, β = 0.19) near IGF2BP3 (insulin-like growth factor 2 mRNA-binding protein 3). The PPP2R2B association remained genome-wide significant after additionally covarying for global amyloid burden and cerebrovascular disease risk, while the IGF2BP3 association was partially attenuated after accounting for amyloid load. In addition to these discoveries, three single nucleotide polymorphisms within MAPT (microtubule-associated protein tau) displayed nominal associations with tau-positron emission tomography burden, and the association of the APOE (apolipoprotein E) ɛ4 allele with tau-positron emission tomography was marginally nonsignificant (P  = 0.06, β = 0.07). No associations with tau-positron emission tomography burden were identified for other single nucleotide polymorphisms associated with Alzheimer’s disease clinical diagnosis in prior large case–control studies. Our findings nominate PPP2R2B and IGF2BP3 as novel potential influences on tau pathology which warrant further functional characterization. Our data are also supportive of previous literature on the associations of MAPT genetic variation with tau, and more broadly supports the inference that tau accumulation may have a genetic architecture distinct from known Alzheimer’s susceptibility genes, which may have implications for improved risk stratification and therapeutic targeting.


Introduction
The Alzheimer's disease biomarker cascade is widely accepted to include neocortical amyloid and tau deposition as prominent features (Jack et al., 2013). Brain amyloidosis is thought to have substantial heritable underpinnings, with numerous studies identifying genetic variants associated with differences in amyloid burden (Thambisetty et al., 2013;Ramanan et al., 2014Ramanan et al., , 2015Apostolova et al., 2018;Yan et al., 2018). In comparison however, the genetic influences on tau deposition in Alzheimer's disease are poorly understood.
The APOE (apolipoprotein E) E4 allele, the strongest known genetic risk factor for sporadic Alzheimer's disease, is well-known to be associated with increased amyloid load (Reiman et al., 2009;Vemuri et al., 2010;Ramanan et al., 2014Ramanan et al., , 2015 but in population genetics studies appears to display only conditional associations with tau pathology (Farfel et al., 2016;Hohman et al., 2018;Ramanan et al., 2019). A few studies have investigated genetic determinants in relation to levels of cerebrospinal fluid tau (Kauwe et al., 2008;Cruchaga et al., 2013). However, it remains an open question as to why some individuals may be more susceptible to the accumulation of tau pathology in Alzheimer's disease while others are more resistant, a fundamental unanswered issue given that the burden and topography of tau deposition is closely related to cognitive symptoms in Alzheimer's disease.
With the recent validation of positron emission tomography ligands for tau (Villemagne et al., 2015;Saint-Aubert et al., 2017), we have an opportunity to noninvasively assess tau pathology in vivo in samples large enough for genetic analyses that are conducive for novel discoveries. In this work, we hypothesized that common genetic variants would be associated with susceptibility or resistance to tau deposition in older adults, and conducted a genome-wide association study (GWAS) of taupositron emission tomography burden to test this hypothesis.

Selection of participants
The Mayo Clinic Study of Aging (MCSA) is a population-based prospective study of older individuals residing in Olmsted County, Minnesota (Roberts et al., 2008;Petersen et al., 2010). Starting in 2004, individuals between 70 and 89 years old were identified for recruitment using the Rochester Epidemiology Project medical records linkage system (Rocca et al., 2012;St Sauver et al., 2012). In 2012, the study was extended to include those aged 50 and older, and starting in 2015, tau-positron emission tomography was added to the MCSA. Clinical data (through questionnaires and in-person history), neuropsychological assessment, neuroimaging and blood laboratory tests were assessed at selected visits. Clinical diagnoses were made by a consensus panel, incorporating all available information. All study protocols were approved by the Mayo Clinic and Olmsted Medical Center Institutional Review Boards and written informed consent was obtained from all participants or their surrogates. Our inclusion criteria included individuals over age 50 with tau-positron emission tomography and genomewide genotype data.

Demographic and clinical data
Age, sex, years of education, the number of active medications, and scores on the Mini Mental Status Exam, and Clinical Dementia Rating Sum of Boxes were ascertained for each patient at clinical visit. As a measure of cerebrovascular disease risk, an index score of chronic late-life cardiac, vascular and metabolic conditions was ascertained from health care records as a summation of the presence or absence of hypertension, hyperlipidemia, cardiac arrhythmias, coronary artery disease, congestive heart failure, diabetes and stroke (Vemuri et al., 2017a). The Charlson Comorbidity Index was used as a measure of overall health and risk of near-term mortality (Charlson et al., 1987).
For relatedness and population ancestry and structure analyses, SNP pruning was performed including restriction to autosomal SNPs with call rate >95%, MAF >5%, Hardy-Weinberg equilibrium P > 1 Â 10 À5 , exclusion of LCT, HLA and polymorphic inversion regions, and linkage disequilibrium (LD) pruning using a 1000 SNP sliding window to select SNPs with r 2 < 0.2. Twenty-eight subjects were removed due to non-Caucasian ancestry as determined using STRUCTURE version 2.3.4, and using 585 samples from the 1000 Genomes Project as population anchors. Twenty-five pairs of individuals exhibited significant relatedness, defined as PLINK identity by descent PI_HAT >0.25 (Zeng et al., 2015), and therefore one individual from each pair was excluded based on the following criteria, in order of priority: (i) missing tau or amyloid positron emission tomography data; (ii) clinical diagnosis as cognitively unimpaired (as compared to mild cognitive impairment or dementia); (iii) lower call rate. APOE E2/E3/E4 allele status determined through genotyping of its defining SNPs (rs429358 and rs7412) on the Illumina chip displayed 100% concordance with allele status previously determined in the MCSA sample via standard restriction digest methods (Hixson and Vernier, 1990). Principal component analysis using SNPRelate version 1.22 (Zheng et al., 2012), identified no significant effects of population stratification within the final dataset. As a conservative measure to account for any potential confounding effects of population structure, the first five principal component eigenvectors were utilized as covariates in all genetic analyses. Following these procedures, data passing quality control was available for 506 136 SNPs and 1727 MCSA participants.
Data from genome-wide imputation were used in this study to evaluate genomic regions and variants of interest which were not directly assayed by the GWAS chip. Prior to imputation, the orientation of all genotyped markers in relation to the plus strand alignment of the Haplotype Reference Consortium reference panel (McCarthy et al., 2016) was verified. Imputation was performed for all autosomal chromosomes with Minimac version 4-1.0.2 via the Michigan Imputation Server  and using default phasing and other recommended quality control parameters and the EUR reference population. Following imputation and removal of monomorphic variants, data was available for 18 873 872 SNPs across the genome. After application of quality control filters as applied to the GWAS chip dataset (SNP call rate < 95%, MAF < 1% and Hardy-Weinberg equilibrium P < 1 Â 10 À5 , and subject call rate < 98%) and removal of SNPs with ambiguous alignment or no identifying rs number, 6 417 232 SNPs were available for analysis. Imputation quality for the dataset was high based on the squared correlation coefficient between imputed allele dosages and masked genotypes (r 2 ) being 0.992. This imputed data was used for finemapping of regions (250 kb upstream and downstream) surrounding SNPs with genome-wide significant association in the GWAS and to analyse a hypothesis-driven list of SNPs previously associated with Alzheimer's disease clinical diagnosis in large case-control studies through the International Genomics of Alzheimer's Project (Lambert et al., 2013;Kunkle et al., 2019).

Neuroimaging data
The acquisition, processing and summary measure details for positron emission tomography imaging performed on MCSA study participants are described in detail elsewhere (Jack et al., 2017). All analyses utilized an inhouse fully automated image processing pipeline with 43 atlas-defined regions of interest (ROIs) propagated from an MRI template. Positron emission tomography tracer choice was based on longstanding in-house protocols to maintain consistency.
Tau-positron emission tomography was performed with AV-1451 ( 18 F-flortaucipir), synthesized on site using the precursor compound supplied by Avid Radiopharmaceuticals (Xia et al., 2017;Lowe et al., 2019). Regional tau burden was computed from median tracer uptake in each ROI divided by the cerebellar crus grey matter ROI to yield a standardized uptake value ratio (SUVR). The primary outcome measure for this study was tau burden in a composite meta-ROI, computed using median tau uptake in the entorhinal, amygdala, parahippocampal, fusiform and inferior and middle temporal ROIs, divided by the cerebellar crus grey matter ROI (Jack et al., 2017). For comparison, post hoc models additionally used as outcomes tau burden in the entorhinal and inferior temporal cortices, regions known to exhibit early and prominent tau pathology in Alzheimer's disease (Johnson et al., 2016;Vemuri et al., 2017b). Where individuals had multiple longitudinal tau positron emission tomography scans, the first scan was used for the cross-sectional analyses performed in this study. Tau positivity (versus negativity) was defined by meta-ROI SUVR !1.25 as previously described .
Amyloid-positron emission tomography imaging, completed during the same visit as tau-positron emission tomography for all but one individual in the sample, was performed using Pittsburgh compound-B (Klunk et al., 2004). Global cortical amyloid load was computed for each participant from median tracer uptake in the prefrontal, orbitofrontal, parietal, temporal, anterior cingulate and posterior cingulate/precuneus ROIs, divided by the median uptake in the cerebellar crus grey matter ROI. Amyloid positivity (versus negativity) was defined by global cortical SUVR ! 1.48 as previously described .

Statistical analyses
Distributions for all neuroimaging biomarkers were assessed using scatter plots and histograms. We elected against the use of a data transformation for the tau-and amyloid-positron emission tomography phenotypes in view of their expected non-normal distribution in a large population-based sample predominated by cognitively unimpaired older adults. Genetic analyses were performed using PLINK version 1.9. Association tests for each SNP with tau-positron emission tomography burden were conducted using linear regression under an additive genetic model and including age at tau-positron emission tomography, sex and the first five genetic principal component eigenvectors as covariates.

Hypothesis-driven genetic variant associations with tau-positron emission tomography
In the initial, hypothesis-driven phase of this study, we assessed for associations of variants previously demonstrated as relevant for Alzheimer's disease, including the APOE E2 and E4 alleles and 43 SNPs which displayed genome-wide significant associations with a clinical diagnosis of probable Alzheimer's disease in the largest available case-control studies (Lambert et al., 2013;Kunkle et al., 2019) through the International Genomics of Alzheimer's Project consortium (Supplementary Table 1). Imputation was used to determine genotypes for any SNPs not directly assayed by the chip array, with all but one SNP (HLA-DRB1 rs9271058) successfully imputed after quality control. In this phase we also tested for any associations within chip genotype data for 31 SNPs in MAPT (microtubule-associated tau protein) given the direct biological relevance of this gene to tau deposition (Pittman et al., 2006). For these hypothesis-driven analyses, a significance threshold of P < 6.58 Â 10 À4 (0.05/76 variants) was used.

GWAS of tau-positron emission tomography
Next, we performed a GWAS to discover novel SNP associations with tau-positron emission tomography burden. In this study phase, a conservative significance threshold of P < 5 Â 10 À8 was used based on a Bonferroni correction of one million independent tests (Pe'er et al., 2008). Standardized beta coefficients were generated to describe effect sizes, representing the additive change in standardized tau-positron emission tomography burden (mean ¼ 0, standard deviation ¼ 1) corresponding to each additional minor allele. Manhattan and Q-Q plots were generated using Haploview version 4.2 (Barrett et al., 2005), and regional Manhattan plots were generated using the web-based tool LocusZoom (Pruim et al., 2010). Public databases were utilized for LD calculations based on reference population data, including LDLink (Machiela and Chanock, 2015) and Ensembl version 98 (Hunt et al., 2018).

Post-GWAS: complementary models
SPSS Statistics version 22.0 (IBM Corp., Armonk, NY) and SAS version 9.4 (SAS Institute, Inc, Cary, NC) were used for other analyses to extend and complement the GWAS. To assess for potential confounding factors impacting significant SNP associations, demographic and clinical variables were analysed for genotypic group differences using ANCOVA models for continuous variables and logistic regression for categorical variables. Post-GWAS linear and logistic regression models were used to assess the robustness of significant SNP associations, including via the application of additional covariates (APOE E4 status, cardiovascular and metabolic condition, global amyloid burden) and stratification of the sample by amyloid or tau status (positive versus negative). Regression models using stepwise forward entry were used to examine the independent variance explained by individual target SNPs after accounting for age, sex, the first five genetic principal components, and global amyloid burden.

Post-GWAS: voxel-wise analyses
Voxel-wise analyses were performed using SPM version 12 (Wellcome Trust Centre for Neuroimaging) to characterize the whole-brain spatial distribution of associations for target SNPs, as previously described (Ramanan et al., 2015). For these analyses, age, sex and the first five genetic principal components were included as covariates in an ANCOVA model. A study-specific explicit grey matter mask was used, and results were displayed at a voxelwise significance threshold of family-wise error-corrected P < 0.05 with minimum cluster size (k) ¼ 100 voxels. Surface renderings of the voxel-level results were generated for visualization. To ensure that voxel-wise tau associations were not driven by partial volume effects or neurodegeneration, for comparison we tested for associations of these SNPs with tau when correcting for partial volume and separately tested for voxel-level grey matter SNP associations.

Post-GWAS: fine-mapping and functional annotation
To fine-map top association signals from the GWAS and screen for proximal multi-SNP combinations tagging those variants, we performed exploratory haplotype proxy analyses in PLINK version 1.07. Using regional imputed data flanking SNPs with genome-wide significant associations, haplotype combinations of 2-5 SNPs in strong LD with the top SNP and with MAF >1% were scanned over maximum windows of 500 SNPs and 250 kb and tested for association with tau-positron emission tomography, covarying for age, sex and the first five genetic principal components. Separately, a variety of public databases were utilized for functional annotation of variants of interest, with the goals of assessing potential SNP-gene relationships as well as LD and tissue-specific gene expression patterns. These resources included LDLink (Machiela and Chanock, 2015), Ensembl version 98 (Hunt et al., 2018), the Human Protein Atlas (Uhlen et al., 2015) and the Single Cell Atlas of the Entorhinal Cortex in Human Alzheimer's disease database (Grubman et al., 2019).

Data availability
Data from this study are available from the authors upon reasonable request.

Sample characteristics
The GWAS sample included 754 individuals over 50 years of age with tau-positron emission tomography imaging, with mean age 71.9 years (standard deviation 10.4 years) and with men comprising 55% of the sample ( Table 1). The preponderance of individuals in the sample was cognitively unimpaired (87.6%). Comparable to previous estimates of APOE allele frequency in individuals of European descent (Heffernan et al., 2016), 29% of the GWAS sample carried at least one E4 allele. A majority of individuals were classified as amyloid negative by positron emission tomography (61%).
Hypothesis-driven analyses of Alzheimer's disease risk loci and MAPT with tau-positron emission tomography burden Neither the APOE E2 nor E4 allele demonstrated association with tau burden in the primary meta ROI, with the summary statistics for the E4 allele being marginally nonsignificant prior to correction for multiple comparisons (P ¼ 0.06, b ¼ 0.07). These results did not appreciably change when using a dominant ( versus additive) genetic model. A significant association of APOE E4 with entorhinal cortex tau-positron emission tomography burden (P ¼ 3.30 Â 10 À5 , b ¼ 0.14) and a nominal association with inferior temporal cortex tau-positron emission tomography burden (P ¼ 0.03, b ¼ 0.08) were identified, both of which were completely attenuated after covarying for global amyloid burden, consistent with prior analyses of a smaller sample from the MCSA cohort (Ramanan et al., 2019). Three SNPs within MAPT displayed nominal associations with tau burden, including rs1467967 which tags a common MAPT haplotype , though none survived correction for multiple comparisons (Supplementary Table 1). No nominal associations with tau were identified for any of the SNPs reported as associated with Alzheimer's disease clinical diagnosis in prior large consortium case-control studies (Supplementary Table 1).
Confounders do not explain the associations of PPP2R2B rs76752255 and IGF2BP3 rs117402302 with higher tau-positron emission tomography burden The associations of PPP2R2B rs76752255 and IGF2BP3 rs117402302 with higher tau remained genome-wide significant when including APOE E4 (P ¼ 7.94 Â 10 À9 ; b ¼ 0.20, and P ¼ 2.81 Â 10 À8 ; b ¼ 0.19, respectively) or cardiovascular and metabolic condition (P ¼ 4.31 Â 10 À9 ; b ¼ 0.20, and P ¼ 4.63 Â 10 À8 ; b ¼ 0.19, respectively) as additional covariates to the main regression model. When global amyloid burden was included as a covariate, the association of PPP2R2B rs76752255 remained genomewide significant (P ¼ 2.17 Â 10 À9 ; b ¼ 0.19), while the association of IGF2BP3 rs117402302 was partially attenuated (P ¼ 7.81 Â 10 À6 ; b ¼ 0.14). Post hoc analyses stratified by amyloid-positron emission tomography status (positive versus negative) revealed the same direction of association for these SNPs regardless of amyloid status, though the magnitudes of tau-positron emission tomography SUVR difference were greater in amyloid positive individuals ( Fig. 3B and D). Based on logistic regression, carriers of PPP2R2B rs76752255-C were more likely than non-carriers to be tau-positive [P ¼ 0.01, OR ¼ 3.24 (95% CI 1.27-8.23)] when covarying for age, sex, genetic principal components and global amyloid burden. No allelic differences in tau status were identified for IGF2BP3 rs1174024302 after accounting for global amyloid burden. Overall, carriers of rs76752255-C were significantly more impaired than non-carriers based on the Clinical Dementia Rating scale (P < 0.001) despite having no or marginally significant differences in age, sex, education, cerebrovascular disease risk and brain amyloid burden (Supplementary Table 3). Carriers of rs117402302-A similarly displayed greater impairment than non-carriers on the Clinical Dementia Rating (P < 0.001) and had higher amyloid burden (P < 0.001) but no differences in age, sex, education or cerebrovascular disease risk. When a stronger relatedness threshold was used (PI_HAT >0.125) resulting in removal of an additional subject from analysis, the associations of PPP2R2B rs76752255 (P ¼ 1.01 Â 10 À8 ; b ¼ 0.20) and IGF2BP3 rs117402302 (P ¼ 4.09 Â 10 À8 ; b ¼ 0.19) with higher tau remained genome-wide significant with unchanged beta weights, indicating no spurious effect from cryptic relatedness in the primary sample.
Voxel-wise analyses find associations of PPP2R2B rs76752255 and IGF2BP3 rs117402302 with higher taupositron emission tomography burden We performed whole-brain voxel-wise analyses to assess the broader spatial distribution for the associations of the two target SNPs with tau-positron emission tomography burden. Compared to minor allele noncarriers, individuals with either the PPP2R2B rs76752255-C or IGF2BP3 rs117402302-A alleles displayed higher tau on a voxel level in an Alzheimer's disease-pattern distribution, with predominant clusters in the medial, lateral and inferior temporal and posterior cingulate regions (Fig. 4 and Supplementary   Figure 1 Manhattan plots for the GWAS of tau-positron emission tomography burden. Observed Àlog 10 P-values are displayed (yaxis) for all single SNPs tested in the GWAS. Genome-wide significant associations (P < 5 Â 10 À8 ; red line) with tau-positron emission tomography burden were identified on chromosomes 5 (rs76752255) in PPP2R2B and on chromosome 7 (rs117402302) near IGF2BP3. Suggestive associations (P < 1 Â 10 À6 ; blue line) were identified on additional chromosomes.
Variants associated with Tau PET BRAIN COMMUNICATIONS 2020: Page 7 of 15 | 7 Figure 2 Regional manhattan plots for the GWAS of tau-positron emission tomography burden. All genotyped and imputed SNPs within a 500 kb region surrounding PPP2R2B rs76752255 (A) and IGF2BP3 rs117402302 (B) are plotted based on their association P-values, NCBI build 37 genomic position, and recombination rates calculated from the 1000 Genomes Project reference data. The colour scale of r 2 values is used to label SNPs based on their degree of LD with the target SNP. Top SNP associations in the region are labelled. Genes in the region are labelled below the plots, with arrows denoting 5 0 -to-3 0 orientation. Plots were created using the LocusZoom software suite. Table 5). The results were unchanged when additionally correcting for partial volume, and corresponding voxel-wise grey matter analyses revealed no significant differences between minor allele carriers versus noncarriers for either SNP, indicating that the tau associations were not driven by partial volume effects related to neurodegeneration. For comparison, voxel-wise analyses for APOE E4 showed only modest focal medial temporal signal ( Supplementary  Fig. 3), in contrast to the robust Alzheimer's disease pattern of association seen for the PPP2R2B and IGF2BP3 SNPs identified through GWAS.
Proxy analyses suggest that PPP2R2B rs76752255 may tag a nearby haplotype Using regional imputed data to maximize genomic coverage, we screened for haplotypic multi-SNP combinations Figure 3 Associations of PPP2R2B rs76752255 and IGF2BP3 rs117402302 with higher tau. Tau-positron emission tomography SUVR data for a composite 'meta' ROI are shown in relation to PPP2R2B rs76752255 and IGF2BP3 rs117402302 genotype. (A) and (C) In the full study sample, the rs76752255-TC and rs117402302-GA genotypes were associated with higher tau-positron emission tomography burden. (B) and (D) In post hoc stratified analyses, these alleles demonstrated the same direction of association (protective) for tau-positron emission tomography burden regardless of amyloid status, though the magnitude of SUVR difference was relatively greater in amyloid-positive individuals than in amyloid-negative individuals. One individual from the full study sample was not included in these analyses due to not having an available amyloid-positron emission tomography scan. that may tag the top association signals from the GWAS. Within the region surrounding PPP2R2B rs76752255, despite no single SNPs displaying substantial LD with rs76752255, 6 haplotypes (combinations of 2-5 SNPs) exhibited evidence for serving as proxies (r 2 > 0.8). Among these, the haplotype consisting of rs77205756-T/ rs17105275-T displayed strong LD with rs76752255 and association with the tau-positron emission tomography phenotype at an even more robust significance level (P ¼ 5.20 Â 10 À11 , r 2 ¼ 0.83 with rs76752255, b ¼ 0.17, frequency ¼ 1.7%), suggesting that the GWAS discovery in PPP2R2B may be tagging a nearby haplotype. In contrast, no haplotypes within the region of IGF2BP3 rs117402302/ rs138919567 were found to be proxies for these SNPs.

Discussion
To our knowledge, this study of a population-based sample of older adults represents the first reported GWAS of tau-positron emission tomography burden. Using this approach, we discovered novel associations of PPP2R2B and IGF2BP3 variants with higher tau deposition in older adults. These findings advance extant knowledge about the potential heritable factors associated with tau pathology in older adults, and highlight the importance of accounting for differential susceptibility and resistance to tau deposition when considering risk of developing Alzheimer's disease dementia. Our data also support the inference that the genetic and molecular underpinnings of tau deposition in Alzheimer's disease may be distinct from established Alzheimer's disease susceptibility genes, arguing for a broadening of thought regarding models of risk stratification and therapeutic targeting.
The influences on tau deposition in aging and neurodegenerative disease are presumed to be complex and multifactorial. b-amyloidosis is a risk factor for tau accumulation within and outside of the medial temporal lobe (He et al., 2018;Jack et al., 2018Pontecorvo et al., 2019;Ramanan et al., 2019). However, a substantial proportion of older individuals who are cognitively unimpaired harbor amyloid pathology (Sperling et al., 2013;Jansen et al., 2015;Petersen et al., 2016), including some individuals who meet biomarker criteria for amyloid positivity but not for tau positivity. These different biomarker profiles result in differences in longitudinal clinical trajectories . Given that cognitive decline in Alzheimer's disease is more closely related to tau accumulation than to amyloid accumulation (Nelson et al., 2012;Ossenkoppele et al., 2016;Bejanin et al., 2017), understanding why certain individuals may have relatively more resistance (avoidance) or resilience (coping) against tau deposition may be crucial for improved risk stratification and therapeutic targeting in Alzheimer's disease (Arenaza-Urquijo and Vemuri, 2018).

Figure 4
Voxel-wise whole brain analyses of the associations of PPP2R2B rs76752255 and IGF2BP3 rs117402302 with higher tau. Glass brain surface renderings from whole-brain voxel-wise analyses are displayed. Intensity scales are shown below, with brighter colours on the primary red-orange-yellow scale indicating progressively stronger association signals, and with individual regions comprising the composite 'meta-ROI' overlayed and denoted by distinct colour scales as labelled. Significant clusters of association are shown for PPP2R2B rs76752255-C (left panel) and IGF2BP3 rs117402302 (right panel) with higher tau-positron emission tomography burden, with predominant effects observed in the medial, lateral and inferior temporal and posterior cingulate regions.
Our new findings support the concept that heritable factors can influence susceptibility and resistance to tau pathology. It is noteworthy that individuals with the PPP2R2B susceptibility allele in this study displayed significantly higher levels of tau and clinical impairment despite having no or marginally significant differences with their counterparts in age, sex, education, cerebrovascular disease risk, APOE E4 status, and brain amyloid burden, supporting the concept that tau-associated genetic variation may impact clinically-relevant Alzheimer's disease pathophysiology independently from other Alzheimer's disease dementia-associated risk factors. However, individuals with the IGF2BP3 susceptibility allele had higher amyloid load compared to their counterparts even though both groups had similar ages, suggesting that this allele may impact multiple components of the Alzheimer's disease biomarker cascade. The complementary voxel-wise analyses also demonstrated that the whole-brain topography of higher tau burden associated with the PPP2R2B and IGF2BP3 variants was of a frank Alzheimer's disease pattern, and was not driven by isolated focal signal hotspots or partial volume effects related to neurodegeneration. Collectively, these notions lend additional credence to the theory that Alzheimer's disease is etiologically complex and that a combination of multiple genetic and lifestyle/environmental factors can influence its key pathophysiologic processes on an individualized basis (Reitz, 2016).
Plausible biological mechanisms can be inferred from existing literature to account for these novel associations. PPP2R2B is highly expressed in the brain (Uhlen et al., 2015) and regulates the activity of the PP2A (protein phosphatase 2A) family of enzymes, which serve as a principal factor in controlling the phosphorylation status of tau and whose inhibition or downregulation have been shown to promote tau phosphorylation (Torrent and Ferrer, 2012;Sontag and Sontag, 2014). Triplet expansion in PPP2R2B causes spinocerebellar ataxia type 12 (Srivastava et al., 2017), but variation in this gene had not previously been associated with Alzheimer's disease in humans. Based on an atlas of single cell gene expression from entorhinal cortex tissue samples (Grubman et al., 2019), PPP2R2B has lower expression in Alzheimer's disease compared to controls in numerous cell types, with the strongest association in oligodendrocytes (FDR-corrected P ¼ 8.92 Â 10 À171 ). This background supports the potential relevance of genetic variation in PPP2R2B for impacting neurodegenerative processes and highlights the value of our GWAS discovery in this region.
IGF2BP3 has well-known roles in a number of cancers (Lederer et al., 2014), including as a glioblastoma marker (Suvasini et al., 2011). An orthologue of IGF2BP3 (IGF2BP1, also known as IMP1) is known to bind the 3 0 untranslated region of tau mRNA (Atlas et al., 2004). More broadly, insulin and insulin-like growth factor pathways represent targets of active interest in Alzheimer's disease (Arvanitakis et al., 2020). This backdrop includes Alzheimer's disease model system data which support a neuroprotective role for IGFs, and a hypothesis that resistance to this signaling may lead to disinhibition of pro-apoptotic and pro-inflammatory pathways as well as tau phosphorylation related to phosphatidylinositol 3-kinase, protein kinase B (Akt) and glycogen synthase kinase 3 pathways (Benarroch, 2012). Nevertheless, the import of these potential mechanisms in relation to tau in Alzheimer's disease is still unclear, and our findings suggest that additional study of the IGF2BP3 locus may be warranted. More broadly, since top association signals may not always be indicative of the true causal genetic change and given that both coding and non-coding variants can exert functional impacts at distance (Kapranov et al., 2007;Consortium, 2012), our findings argue for further study of the wider loci containing rs76752255 and rs117402302 to more precisely characterize the functional architectures of these genomic regions.
Nominal associations with tau burden were identified for 3 SNPs within MAPT, including the haplotype-tagging SNP rs1467967 which was previously associated with brain MAPT expression (Allen et al., 2014) and levels of cerebrospinal fluid tau (Babi c Leko et al., 2018). Mutations in MAPT are well-known causes of certain tau-related frontotemporal lobar degenerative and atypical Parkinsonian syndromes (Rademakers et al., 2004;Hö glinger et al., 2011;Strang et al., 2019), some of which include tau pathology that can be effectively assessed by AV-1451 tau-positron emission tomography (Spina et al., 2017). Our data provide some additional support for the association of MAPT genetic variation with tau burden, and more broadly reinforces prior work suggesting roles for MAPT in the pathophysiology of Alzheimer's disease and other tauopathies (Allen et al., 2014;Desikan et al., 2015;. There continues to be strong interest in whether a mechanistic relationship exists between APOE and brain parenchymal tau that is independent of amyloid, as has been suggested by analyses of primary tauopathies and selected studies using cellular and animal model systems (Shi et al., 2017;Zhao et al., 2018;Wadhwani et al., 2019). Human in vivo imaging studies along the Alzheimer's disease spectrum have also assessed this issue, which may be impacted by the population being analysed (Ossenkoppele et al., 2016;Ramanan et al., 2019). The findings from this study support the hypothesis that in the general population of older adults, the association of APOE with tau accumulation would appear to be amyloid-dependent and at minimum may be substantially weaker than its association with amyloidosis.
In addition, our study found no associations with taupositron emission tomography burden for other SNPs previously associated with clinically defined probable Alzheimer's disease risk in large case-control studies. Given that several of these genetic variants have displayed associations with amyloid-positron emission tomography burden in prior studies (Ramanan et al., 2014(Ramanan et al., , 2015Apostolova et al., 2018), our findings suggest that the genetic background influencing tau pathology may meaningfully differ from the set of genes typically proposed as top loci for Alzheimer's disease clinical diagnosis. This seeming dichotomy argues for a broadening of thought about potential targets for risk prediction and clinically meaningful interventions in Alzheimer's disease, with particular mind to the potential value of multi-pronged therapy given that factors both related and unrelated to amyloid are likely to influence the disease.
This work has limitations. Although larger than many of the published genome-wide studies of amyloid-positron emission tomography and other Alzheimer's disease endophenotypes, our GWAS of tau-positron emission tomography included a relatively modest sample size compared to non-biomarker-based studies. As such, the novel associations reported in this study warrant a degree of caution. In addition, given the novelty of this approach and its setting consisting of a large, single-site, populationbased sample with tau-positron emission tomography, no comparable replication cohort was available. As the first study on a new imaging genetics dataset, we focused on directly genotyped SNPs in the GWAS, which is an advantage relative to the robustness of our top findings. However, follow-up studies utilizing genome-wide imputed SNP data will provide higher density genomic coverage and facilitate complementary analytical approaches. The relatively low MAFs and lack of minor allele homozygotes for the top SNPs in our study also limited evaluation for potential interactions with other factors and precluded assessment for dosage effects of these alleles. In addition, the cross-sectional design of our study was not equipped to test whether longitudinal imaging and clinical trajectories differ based on genotype profiles. Although outside the scope of this study, which focused on common genetic variation, genome sequencing data would allow for association testing among rare variants, such as the APOE Christchurch variant recently proposed as protective even in the setting of the PSEN1 (presenilin 1) E280A mutation which causes autosomal dominant Alzheimer's disease (Arboleda-Velasquez et al., 2019). Furthermore, the AV-1451 tracer measures Alzheimer's disease-type mixed 3R/4R tau deposits and not isolated 3R or 4R tau aggregates, and the specificity and sensitivity of AV-1451 to Alzheimer's disease-related tauopathy are not perfect. These features could have import for interpretation of imaging genetics associations using tau-positron emission tomography (Leuzy et al., 2019).
Nevertheless, this study highlights the value of pairing advanced molecular imaging with genetics to illuminate potential mechanisms underlying Alzheimer's disease pathophysiologic features. In particular, the novel associations of PPP2R2B and IGF2BP3 with susceptibility to tau pathology warrant further investigation through validation studies and functional characterization, and may have significant implications for future risk counseling and therapeutic development.
Molecular Imaging, AVID Radiopharmaceuticals and the NIH outside the submitted work. Dr Mielke reported receiving research support from the NIH, Department of Defense and unrestricted research grants from Biogen outside the submitted work. Dr Jack reported consulting for Eli Lilly, serving on an independent data monitoring board for Roche, and serving as a speaker for Eisai but receives no personal compensation from any commercial entity; he also reported receiving research support from the NIH and the Alexander Family Alzheimer's Disease Research Professorship of the Mayo Clinic. Dr Petersen reported receiving consulting fees from Hoffman-La Roche, Inc, Merck, Inc, Genentech, Inc, Biogen, Inc, GE Healthcare and Eisai, Inc outside the submitted work. Dr Vemuri reported receiving grants from the NIH during the conduct of the study. No other disclosures were reported.