Tau Polygenic Risk Scoring: a Cost-Effective Aid for Prognostic Counseling in Alzheimer’s Disease

Abstract

Tau deposition is one of two hallmark features of biologically defined Alzheimer’s disease (AD) and is more closely related to cognitive decline than amyloidosis. Further, not all amyloid-positive individuals develop tauopathy, resulting in wide heterogeneity in clinical outcomes across the population with AD. We hypothesized that a polygenic risk score (PRS) based on tau PET (tau PRS) would capture the aggregate inherited susceptibility/resistance architecture influencing tau accumulation, beyond solely the measurement of amyloid-β burden. Leveraging rich multimodal data from a population-based sample of older adults, we found that this novel tau PRS was a strong surrogate of tau PET deposition and captured a significant proportion of the variance in tau PET levels as compared with amyloid PET burden, APOE (apolipoprotein E) ε4 (the most common risk allele for AD), and a non-APOE PRS of clinical case-control AD risk variants. In independent validation samples, the tau PRS was associated with cerebrospinal fluid phosphorylated tau levels in one cohort and with postmortem Braak neurofibrillary tangle stage in another. We also observed an association of the tau PRS with longitudinal cognitive trajectories, including a statistical interaction of the tau PRS with amyloid burden on cognitive decline. Although additional study is warranted, these findings demonstrate the potential utility of a tau PRS for capturing the collective genetic background influencing tau deposition in the general population. In the future, a tau PRS could be leveraged for cost-effective screening and risk stratification to guide trial enrollment and clinical interventions in AD.

Introduction

Tau accumulation is one of two hallmark features of biologically defined Alzheimer’s disease (AD) [22] and represents a target for risk stratification and treatment given its close relationship with neurodegeneration and cognitive impairment [23, 33]. Brain tau positivity (T+) can be assessed in vivo through cerebrospinal fluid (CSF) biomarkers [7], and more recently through validated positron emission tomography (PET) ligands [58, 71] and evolving research-based plasma assays [41, 64]. However, relatively little is known about accurate prediction of who may be at risk for abnormal tau deposition in the future, particularly at presymptomatic or early symptomatic stages where preventive strategies may be of highest yield and where appropriate trial enrollment may impact drug success or failure.

Amyloid positivity (A+) is generally considered to be necessary for the development of tauopathy in AD [31], but a nontrivial proportion of older individuals remain A+/T- during life [26]. This observation, along with the wide temporal dissociation between rising amyloid levels and later tau deposition and cognitive decline [16], suggests that other processes may impact tau accumulation and spread. In particular, there is growing evidence to support genetic factors as influencing differential susceptibility (versus resistance) to tau deposition [10, 12, 19, 51, 53], with initial data pointing to this inherited architecture as being multifactorial and likely distinct in key aspects from the gene variants associated with clinically diagnosed probable AD dementia in case-control designs [32, 34] as well as amyloidosis [53, 74].

Polygenic risk score (PRS) calculation is a well-validated approach for capturing the aggregate influence of genetic variation on complex outcomes, as a step toward precision medicine [6]. While the PRS framework has been applied widely to assess the collective genetic risk of clinically diagnosed probable AD dementia [9, 13], we sought to define a PRS for predicting tau accumulation as a potentially important factor in guiding prognostic clinical counseling. Specifically, we hypothesized that a PRS of tau PET burden would uniquely capture the aggregate genetic susceptibility/resistance to tau accumulation beyond solely the measurement of amyloidosis. In this study, we analyzed a large population-based sample of older adults to (1) calculate a tau PRS, (2) assess the utility of the tau PRS in comparison with the APOE (apolipoprotein E) ε4 allele, an “AD risk” PRS, and amyloid PET burden, and (3) validate the association of the PRS with complementary in vivo biomarkers and postmortem measures of tau pathology.

Methods

Selection of Participants

Data was drawn from the Mayo Clinic Study of Aging (MCSA), a population-based prospective study of older adults residing in Olmsted County, Minnesota [49, 55, 57, 61]. All study protocols were approved by the Mayo Clinic and Olmsted Medical Center Institutional Review Boards. Written informed consent was obtained from all participants or their surrogates. Primary inclusion criteria for discovery analyses in this study included age 50 years or older with the presence of a tau PET scan and GWAS data. Validation studies analyzed CSF biomarker and postmortem neuropathologic data in MCSA subsamples independent from the tau PET/PRS sample.

Demographic and Clinical Data

Age at the relevant assessment (neuroimaging, lumbar puncture, or death), sex, and years of education were ascertained for each patient. In the MCSA, a measure of cerebrovascular disease risk (CMC) was ascertained from health care records based on the presence or absence of conditions influencing vascular health [68]. APOE ɛ2/ɛ3/ɛ4 allele status was determined via standard digest methods [18].

Cognitive Assessment Data

Cognition was assessed longitudinally through composite z-score measures from neuropsychological test battery items [55]. A composite score for memory was derived from the delayed recall tasks of the WMS-R Logical Memory II, WMS-R Visual Reproduction II, and AVLT [25]. A global cognitive summary score was derived from the memory, executive (TMT Part B, WAIS-R Digit Symbol), language (BNT, category fluency), and visuospatial (WAIS-R Picture Completion, WAIS-R Block Design) functioning domain scores [69]. For each individual task, z-scores were calculated based on the mean and standard deviation for cognitively unimpaired MCSA enrollees aged 50 and older, and the domain scores were themselves weighted back to age and sex distributions within the Olmsted County population [56].

GWAS Data

GWAS array data was acquired for 1783 participants from the MCSA sample using the Illumina Infinium Global Screening Array-24 v2.0, which included 658,805 single nucleotide polymorphisms (SNPs) across the genome [52, 53]. Subject-level exclusion criteria included a call rate<98%, sex discordance with clinical data, heterozygosity rate, batch effects, significant relatedness (PI_HAT>0.25), or non-Caucasian ancestry (determined using STRUCTURE version 2.3.4). For directly genotyped variants, SNP-level exclusion criteria included call rate<95%, Hardy-Weinberg equilibrium p<1×10⁻⁵, and minor allele frequency (MAF)<1%. Genome-wide imputation was performed with the Michigan Imputation Server [11] using Minimac version 4–1.0.2 and the Haplotype Reference Consortium reference panel [39]. Following imputation, identical SNP-level exclusion criteria were applied, and in addition, variants with low imputation quality (r²<0.8) were removed and missing genotypes from the GWAS array were substituted with imputed data where applicable. This resulted in a total of 7,295,984 SNPs available for analysis. To account for population stratification, SNPRelate [78] was used to generate principle component eigenvectors for use as covariates.

Neuroimaging Data

In the MCSA, the acquisition and analysis of tau PET and amyloid PET scans using an in-house fully automated image processing pipeline are described in detail elsewhere [27]. Tau PET was performed with AV-1451 (¹⁸F-flortaucipir), synthesized on site with precursor supplied by Avid Radiopharmaceuticals [37]. Amyloid PET was performed with Pittsburgh compound B (PiB) [28]. Standardized uptake value ratio (SUVR) measures for tau and amyloid PET were generated by normalizing median tracer uptake in target regions of interest (ROIs) to the cerebellar crus grey matter. The target tau PET measure was a meta-ROI computed from the entorhinal, amygdala, parahippocampal, fusiform, and inferior and middle temporal ROIs [27]. As a complementary measure, tau PET positivity was defined by meta-ROI SUVR≥1.25 [25]. The target amyloid PET measure was global cortical amyloid load, computed from the prefrontal, orbitofrontal, parietal, temporal, anterior cingulate, and posterior cingulate/precuneus ROIs. Amyloid PET positivity was defined by global cortical SUVR≥1.48 (≥22 centiloids) [37].

Cerebrospinal Fluid Biomarkers

A subset of MCSA participants underwent lumbar puncture for CSF collection. For these samples, CSF hyperphosphorylated tau (p-tau) was analyzed using the Elecsys phospho-tau (181P) CSF electrochemiluminescence immunoassay (Roche Diagnostics, Basel, Switzerland) [17]. Full details on acquisition, processing, and quality control are described elsewhere [42, 67].

Postmortem Neuropathologic Analyses

A subset of MCSA participants underwent neuropathologic examination of postmortem brain tissue using standardized sampling and dissection protocols [43, 65]. Briefly, formalin-fixed and paraffin-embedded tissue sections were cut at 5 μm thickness and used for histologic and immunohistochemical studies. Braak neurofibrillary tangle stage [4] was assessed using either phospho-tau antibody (AT8; 1:1000; Endogen, Woburn, MA) or modified Bielschowsky silver stain [20, 65].

Statistical Analyses for Tau PRS Calculation

As a precursor to PRS generation, a GWAS of tau PET was performed using PLINK version 1.9 [5]. For the GWAS, linear regression under an additive genetic model was performed, utilizing meta tau PET SUVR as the outcome. Covariates for the GWAS included the first five genetic principal component eigenvectors (based on standard recommendations for most GWAS [50, 66]) along with age at neuroimaging and sex as common demographic variables.

The tau PRS was calculated per recommended guidelines [44] utilizing R Statistical Software (version 3.6.2). Variants with association p<1×10⁻⁷ in the GWAS were pruned by LD (r²>0.8) to isolate the top independent GWAS signals. As an additional conservative step, we excluded any uncommon SNPs (MAF≤10%) where the association with tau PET weakened (p>1×10⁻⁷) when using a dominant model. For the remaining 17 SNPs, variants were selected for inclusion in the final PRS based on linear regression models utilizing the best subset variable selection procedure, the best model was indicated by having the smallest Bayesian Information Criterion. To account for a potential “winner’s curse” phenomenon, we applied a correction factor to generate penalized β coefficients (β_corr= β x 0.9369) for use in the final PRS [1]. The tau PRS value for each subject was computed as the sum of (allele count x β_corr) for each SNP in the final model, where the allele count was denoted as 0/1 for uncommon (MAF≤10%) SNPs (to limit the potential for rare homozygotes to overly influence the results) or 0/1/2 for common (MAF>10%) SNPs. The adjusted R² value from the model was used to estimate proportion of variability in tau PET explained by the PRS, after accounting for covariates (age, sex, first 5 principal components). A bootstrap approach using 100 bootstrap samples was used for internal validation of the PRS [44]. In each bootstrap sample, an “optimism value” was calculated as the difference in adjusted R² between the bootstrap and original samples. The median of these optimism values was then subtracted from the adjusted R² for the original sample to obtain an internally validated adjusted R² for the PRS. In addition, we assessed the calibration of the PRS by comparing actual tau PET burden with predicted tau PET burden from linear regression analysis including the PRS as a covariate.

Additional Statistical Analyses

Other analyses to evaluate the utility of the tau PRS were performed with R Statistical Software version 3.6.2 and SPSS Statistics version 22.0. Linear regression models were used to test the association of the tau PRS with tau PET burden after inclusion of additional covariates, including global amyloid PET burden, CMC (as this was shown previously to be associated with regional tau PET burden [51]), APOE ε4 allele status, and a non-APOE AD risk PRS based on variants and effect sizes (sum of allele count x log(odds ratio)) from a large published AD case-control GWAS (Supplementary Table 1, online resource) [32]. We also separately performed a sensitivity analysis testing the association of the tau PRS with tau PET values while including only individuals with high amyloid PET burden (≥68 centiloids) [30]. To test for an interaction between amyloid PET burden and the tau PRS on tau PET levels, we assessed the following model: tau PET SUVR = age + sex + first 5 genetic principal component eigenvectors + global amyloid PET SUVR + tau PRS + (global amyloid PET SUVR × tau PRS).

For independent validation, we analyzed two distinct MCSA subsamples which did not overlap with the tau PET PRS. Linear regression was used to test for association of the tau PRS with CSF p-tau levels in 303 participants, including age, sex, CMC, and the first 5 genetic principal components as covariates, with global amyloid PET burden included as an additional covariate in secondary models. Logistic regression was used to test for association of the tau PRS with postmortem tau pathologic burden in 122 participants, specifically against the presence (versus absence) of high Braak neurofibrillary stage (≥stage IV) given that Braak stages of IV or greater are most consistently associated with at least a mild dementia syndrome [15]. Age at death, sex, and the first 5 genetic principal components were included as covariates.

Finally, we utilized linear mixed effects models to investigate the association of the tau PRS with longitudinal cognitive trajectories in 1502 MCSA participants with amyloid PET data and at least two longitudinal composite z-score measures of global cognitive and memory functioning [25, 69]. Terms of interest included time from baseline, age at baseline, sex, education/occupation score [69], cycle number at baseline, amyloid PET burden, tau PRS, and the interaction between amyloid PET burden and the tau PRS. We also included terms for time² and age² to allow for curvature as well as interaction terms for all variables with time and time² to assess whether each predictor affected the slope and curvature, respectively, of the cognitive trajectories. Nested models determined that random effects for subject-specific intercepts, slopes, and curvatures all were necessary (p<0.001). Parsimonious final models were constructed using backwards elimination of predictors while respecting nested terms from higher order interactions.

Availability of Data and Materials

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

Results

Sample Characteristics

The discovery sample was identical to Ramanan, et al. [53] and included 754 individuals with mean age 71.9 years (standard deviation 10.4 years) and comprising 45% women (Table 1). This dataset included a wide range of tau PET values across age ranges (Supplementary Figure 1, online resource). For validation, we analyzed two distinct MCSA subsamples which did not overlap with the tau PET PRS sample, including 303 individuals with CSF tau biomarker data and 122 individuals with postmortem tau neuropathology data.

Table 1:

Sample Characteristics

	Tau PET (N=754) MCSA Discovery	CSF p-tau (N=303) MCSA Validation	Postmortem (N=122) MCSA Validation
Age (years)	71.9 (10.4)	75.2 (10.4)	87.7d (6.9)
Sex	412 (55%) men 344 (45%) women	166 (55%) men 137 (45%) women	70 (57%) men 52 (43%) women
Education (years)	14.8 (2.6)	14.5 (2.8)	14.9 (3.2)
CMC	2.0 (1.6)	2.3 (1.6)	2.9 (1.7)
APOE ε4 Status a	534 (71%) negative 219 (29%) positive	210 (70%) negative 91 (30%) positive	83 (68%) negative 39 (32%) positive
Diagnosis b	659 (87.6%) CU 75 (10.0%) MCI 18 (2.4%) DEM	248 (82.4%) CU 47 (15.6%) MCI 6 (2.0%) DEM	--
Amyloid Status via PET c	460 (61%) negative 293 (39%) positive	170 (56%) negative 133 (44%) positive	--
Meta-ROI Tau PET SUVR	1.21 (0.13)	--	--
CSF p-tau	--	21.7 (10.5)	--
Braak Stage ≥ IV	--	--	54 (44%)

Values displayed as mean (standard deviation) or number (percentage)

Abbreviations: CMC = score of cardiovascular/metabolic conditions (range 0–7); CU = cognitively unimpaired; MCI = mild cognitive impairment; DEM = dementia; ROI = region of interest; SUVR = standardized uptake value ratio

^aMissing for one individual in the discovery sample and 2 individuals in the validation sample

^bConsensus clinical diagnosis at the first PET/clinical visit; missing for 4 individuals in the discovery sample and 2 individuals in the validation sample

^cMissing for one individual in the discovery sample

^dAge at death for postmortem neuropathologic sample

GWAS and PRS Calculation for Tau PET Burden

To identify candidate genetic variants for the PRS, we tested over 7 million SNPs for association with tau PET burden. There was no indication of spurious inflation of GWAS p-values due to population stratification (λ=1.008). A total of 53 variants displayed at least a suggestive association (p<1×10⁻⁷) with tau PET burden (Supplementary Figure 2, online resource). Following LD pruning (r²≤0.8), 21 SNPs remained to evaluate further for PRS inclusion, including 13 SNPs which displayed genome-wide significant association (p<5×10⁻⁸; Supplementary Table 2, online resource). Based on the 21 candidate variants, we identified the most parsimonious combination of SNPs to include in the PRS using linear regression models with a best subset variable selection procedure [76]. The final PRS model included 14 SNPs (Table 2).

Table 2:

Variants Included in the Tau Polygenic Risk Score

					GWAS Results (Dominant Model if MAF ≤ 10%)			Best Subset Selection Results		Penalizedd β for PRS
Variant	Genea	MA	MAF	βb (95% CI)		Adj.c R2	P-value	β (95% CI)	P-value	Score
rs115862481	RP11-550H2.1; RP11-550H2.2 a	C	1.1%	0.165 (0.105, 0.224)		0.008	7.69 × 10−8	0.083 (0.033, 0.133)	0.001	0.078
rs77137782	C2orf50	G	1.5%	0.146 (0.093, 0.198)		0.018	6.37 × 10−8	0.104 (0.061, 0.148)	2.94 × 10−6	0.098
rs10169939	RMDN2	T	41.5%	0.035 (0.022, 0.047)		0.028	5.76 × 10−8	0.025 (0.015, 0.036)	1.84 × 10−6	0.024
rs57978791	MPP4	C	2.2%	0.117 (0.075, 0.159)		0.018	5.98 × 10−8	0.077 (0.042, 0.112)	1.72 × 10−5	0.072
rs17867080	LINC01378	G	2.8%	0.104 (0.066, 0.142)		0.013	9.84 × 10−8	0.061 (0.029, 0.092)	0.0002	0.057
rs76752255	PPP2R2B	T	1.5%	0.147 (0.097, 0.197)		0.035	9.89 × 10−9	0.081 (0.039, 0.123)	0.0001	0.076
rs117402302	IGF2BP3; Metazoa_SRP a	G	1.5%	0.144 (0.093, 0.194)		0.013	3.88 × 10 −8	0.082 (0.040, 0.124)	0.0002	0.077
rs148653042	ERICH1-AS1	G	2.6%	0.104 (0.066, 0.142)		0.021	8.44 × 10−8	0.068 (0.037, 0.099)	2.04 × 10−5	0.064
rs77358498	PPFIBP2	C	1.1%	0.216 (0.153, 0.279)		0.021	3.91 × 10−11	0.172 (0.119, 0.225)	2.77 × 10−10	0.161
rs138168944	SLC5A102; FIBIN a	G	2.1%	0.120 (0.077, 0.163)		0.025	7.37 × 10−8	0.091 (0.055, 0.126)	7.86 × 10−7	0.085
rs12575860	RP11-683O4.1; CTD-2555I5.1 a	C	10.6%	0.055 (0.035, 0.075)		0.030	7.63 × 10−8	0.044 (0.027, 0.060)	2.40 × 10−7	0.041
rs80161412	NTM	T	1.1%	0.179 (0.118, 0.240)		0.022	1.24 × 10−8	0.116 (0.065, 0.167)	8.72 × 10−6	0.109
rs113184873	GGACT	G	1.0%	0.188 (0.125, 0.251)		0.038	8.08 × 10−9	0.096 (0.043, 0.149)	0.0004	0.090
rs76772393	CTD-2591A6.2; RP11-412H8.2 a	C	4.2%	0.090 (0.058, 0.122)		0.009	4.41 × 10−8	0.046 (0.019, 0.073)	0.0008	0.043

^aNearest gene(s) indicated if the target SNP is intergenic or overlapping multiple elements

^bβ values represent the change in tau PET SUVR corresponding to the presence of the minor allele (MAF ≤ 10%) or to each additional minor allele (MAF > 10%)

^cAdjusted R² values (as a goodness of fit measure for the regression terms) listed for each factor as part of a multiple linear regression including these 14 variants as predictors and age, sex, and the first 5 genetic principal components as covariates, utilizing a dominant model for MAF≤ 10% and an additive model otherwise

^dObtained by multiplying the β coefficient from the “best subset selection results” by the heuristic shrinkage estimate (0.9369)

Abbreviations: Chr=Chromosome; MA=minor allele; MAF=minor allele frequency; β=regression coefficient; CI=confidence interval

Tau PRS is a Strong Surrogate for Tau PET Burden

In the discovery sample, the tau PRS ranged from 0 to 0.79 (Figure 1b) and was robustly associated with tau PET burden (p=3.10×10⁻⁷⁵), including within defined strata of PRS ranges (Supplementary Table 3, online resource). A calibration plot (Figure 1b) demonstrated strong agreement between measured tau PET burden and predicted burden based on the PRS. Overall, 34% of the sample (255/754) carried minor alleles for 2 or more SNPs from the PRS, while 18% (135/754) carried no minor alleles from the PRS SNPs. Following bootstrap internal validation to account for potentially overly optimistic estimates, the proportion of variance (adjusted R²) in tau PET collectively explained by age, sex, the first 5 genetic principal components, and the tau PRS was 38.6%, including 27.8% uniquely explained by the tau PRS and 10.7% explained by age. The association between the tau PRS and tau PET burden remained strong when further covarying for cerebrovascular disease risk (p=3.79×10⁻⁷⁶), which is additionally known to influence neurodegeneration in older individuals and may have modest associations with tau deposition [68].

Figure 1: Comparison of Tau PRS Versus Measured Tau PET Burden.

(a) Meta ROI tau PET burden is plotted against tau PRS values for the discovery sample. (b) A calibration plot displays measured tau PET burden against predicted tau PET burden from the PRS. Individuals were grouped according to their measured PRS (0, 0.001 – 0.050, 0.051 – 0.100, 0.101 – 0.150, 0.151 – 0.200, or >0.200). The dashed line indicates the ideal reference line where predicted and measured values are equal. Vertical lines represent the 95% CI for the given mean observed tau PET value.

Tau PRS Uniquely Predicts Tau Burden Over and Above Amyloid Burden Alone

As expected, global amyloid PET burden was strongly associated with tau PET burden (p=1.34×10⁻³⁴). However, amyloid levels explained only 16.2% of the variance in tau burden, a more modest contribution as compared with the tau PRS (27.8%). When the tau PRS and amyloid burden were included together as independent variables in a regression, the association with tau PET burden was stronger for the tau PRS (p=1.61×10⁻⁶⁵, β_std=0.51) than for amyloid levels (p=7.54×10⁻²⁴, β_std=0.32). We also observed a statistical interaction between the tau PRS and amyloid PET burden (p=4.53×10⁻⁸), whereby the effect of the tau PRS on tau PET levels was stronger in individuals with higher amyloid levels. Collectively, these findings indicate unique added value of the tau PRS in predicting tau PET burden, beyond the sole knowledge of brain amyloid levels. In a sensitivity analysis of 123 individuals having high amyloid PET burden (≥68 centiloids), the tau PRS remained associated with tau PET burden (p=5.21×10⁻²³, β_std=0.76).

Tau PRS is a Better Predictor of Tau Burden than APOE ε4 or an AD Risk PRS

To further assess the utility of the tau PRS, we compared it against other relevant genetic markers. As reported previously [53], the association of APOE ε4 dose with meta-ROI tau PET was marginally non-significant (p=0.06, β_std=0.07), and (in contrast to findings for the tau PRS) was non-significant after covarying for amyloid load. A non-APOE “AD risk” PRS was also not associated with tau PET burden (p=0.09, β_std=0.06), and again substantially attenuated after additionally covarying for amyloid burden.

Tau PRS is Associated with CSF p-tau and Postmortem Braak Stage in Independent Samples

In an independent (i.e., non-overlapping with the tau PET discovery cohort) sample of 303 MCSA participants, higher tau PRS values were associated with higher CSF p-tau levels (p=0.04, β_std=0.11). This association was partially attenuated after additionally covarying for amyloid PET burden (p=0.065, β_std=0.09). As with the tau PET data, there was a statistical interaction between the tau PRS and amyloid PET status on CSF p-tau levels (p=5.52×10⁻³), including a larger effect size of the tau PRS in amyloid-positive (p=0.045, β_std=0.17, n=133) versus amyloid-negative (p=0.19, β_std=0.10, n=170) individuals. Among MCSA participants with both PET and CSF biomarker data (n=287), tau PET and CSF p-tau levels were significantly correlated (r=0.40, p=1.35×10⁻¹²), reinforcing the suitability of CSF p-tau data for validation. Among the combined 590 MCSA participants with GWAS and CSF biomarker data, the tau PRS was robustly associated with CSF p-tau levels (p=3.97×10⁻⁵, β_std=0.16), remaining significant after covarying for amyloid PET burden (p=4.14×10⁻³, β_std= 0.10).

For additional validation, we analyzed postmortem neuropathologic data in another independent sample of 122 MCSA participants. A higher tau PRS was associated with higher likelihood of Braak neurofibrillary tangle stage IV or greater (p=0.03, β=6.83), indicating an association of the genetic risk score with substantial AD tau pathology at autopsy.

Tau PRS is Independently and Interactively Associated with Cognitive Trajectories

We utilized linear mixed effects models to investigate the association of the tau PRS with longitudinal cognitive trajectories (n=1502; Supplementary Tables 4–7, online resource). The total models including all fixed and random effects included excellent fit (R²=0.96 and 0.92 for global cognition and memory, respectively). Using nested models to test for significance, the tau PRS was associated with improved model fit for both cognitive outcomes (p<0.001) independently from amyloid burden. Plots display the impact of higher tau PRS being associated with steeper slope (i.e., faster decline) in cognition over time (Figure 3). We also observed a significant association of the time² x tau PRS interaction term on memory (p=0.01, β=−0.04), indicating that a higher tau PRS was associated with accelerated rate of cognitive decline (i.e., with curvature of trajectory) even after accounting for amyloid burden and all nested terms. Further, we observed a statistical interaction between the tau PRS and amyloid burden, such that for individuals with high tau PRS and high amyloid levels, predicted global cognitive functioning would start lower, decline more rapidly, and accelerate that decline more quickly (Figure 2).

Figure 3: Data-Guided Model for Future Clinical Applications of a Tau PRS.

The genetic architecture influencing susceptibility and resistance to amyloid (G_A) and tau (G_T) deposition are shown as largely distinct and influencing different levels of the Alzheimer’s disease pathway. In this model supported by our analyses, G_A contributes to the development of amyloidosis, while G_T influences the accumulation and spread of tau (including a degree of synergism with amyloidosis) which is proximally related to cognitive decline amidst a multifactorial picture (including non-amyloid and non-tau pathways relevant for the disease). In the future, a high tau polygenic risk score (PRS) could provide a blood-based, cost-effective measure to stratify risk of cognitive decline to guide counseling and intervention. Figure created with BioRender.com.

Figure 2: Tau PRS Influences Cognitive Trajectory Particularly in the Setting of High Amyloid.

Longitudinal trajectories of predicted cognition from linear mixed model analyses are displayed for men (left panels) and women (right panels). The curves display predicted performance over time in global cognitive functioning (top row) and memory functioning (bottom row) for a 72-year-old individual, representing the mean age in the sample. “High” and “low” values are defined as beyond one standard deviation from the sample mean. Higher tau PRS values were associated with steeper slopes indicating faster cognitive decline, with this effect observably greater in the setting of high amyloid burden. For global cognitive functioning, we observed an interaction between the tau PRS and amyloid levels, such that for individuals with high tau PRS and high amyloid levels, predicted global cognitive functioning would start lower, decline more rapidly, and exhibit accelerated rate of decline with time.

Discussion

In this study of a population-based sample of older adults with tau PET, we generated a novel endophenotype PRS to capture the aggregate genetic susceptibility/resistance to tau accumulation. This tau PRS accounted for a robust proportion of the variance in tau PET levels over and above amyloid PET burden, APOE ε4, or a non-APOE PRS of probable AD risk variants. We also identified a statistical interaction between the tau PRS and amyloid levels, indicating that the genetic architecture influencing tau deposition may be a critical factor underlying whether amyloidosis ultimately leads to tauopathy. Using independent samples, we validated the tau PRS as associated with CSF p-tau levels and postmortem Braak neurofibrillary tangle stage. We further observed evidence of independent and interactive (with amyloid levels) effects of the tau PRS on worsened cognitive trajectories. These findings demonstrate the potential value of a tau PRS which could be leveraged for future therapeutic development and risk stratification including guiding clinical trial design and enrollment.

Abnormal amyloid accumulation is generally considered to be necessary but not sufficient for the development of cortical tau pathology in AD [31, 38]. Beyond the presence of amyloidosis, it is presumed that tau accumulation is influenced by genetic factors [10, 12, 19, 51, 53] in addition to lifestyle and environmental factors [2, 68]. More precise characterization of this architecture has high potential for predictive clinical application given the close temporal and topographic relationship between tau deposition and cognitive decline in AD [3, 21, 46, 47]. The strongest known genetic risk factor for sporadic AD is the APOE ε4 allele [8], which has been robustly associated with increased amyloid load in numerous studies [53, 54, 70, 74]. Although there is high-quality scientific evidence to support a role of APOE ε4 in tau pathophysiology [59, 72, 77], human studies suggest that this role would appear to be more nuanced and on balance more modest than the relationship between APOE and amyloidosis [47, 51, 53, 73]. Altogether, this backdrop implies that clinically meaningful risk stratification regarding relative susceptibility to tau accumulation would be anticipated to involve factors beyond solely amyloid status and APOE ε4.

Beyond APOE, case-control GWAS have identified numerous risk variants for clinically diagnosed AD dementia [32, 34]. Aggregate case-control PRS measures have been investigated for correlation with relevant imaging and fluid biomarker endophenotypes [9, 35, 62, 75], and one “AD risk” PRS has been developed for potential clinical use [14]. However, case-control GWAS/PRS approaches have limitations, including that clinically probable AD dementia is incompletely correlated with biologically defined AD [63], and that a substantial proportion of nondemented older adults who could be classified as controls have extant AD pathophysiology [24]. Our approach of generating a biomarker-defined PRS offers the advantage of specificity related to a central biological process in AD. The gene variants forming our tau PRS may also offer insights into potential mechanistic targets related to tau deposition in AD, including PPP2R2B (protein phosphatase 2 regulatory subunit B) and NTM (neurotrimin, also known as IgLON2) which encodes a neuronal cell adhesion molecule [60], among others.

We observed that the tau PRS accounted for a relatively substantial proportion of the variance in tau PET burden, an important finding given that most reported neuroimaging genetics associations have displayed modest individual effect sizes [40]. This variance explained by aggregate genetic risk for tau accumulation was larger than, and included unique contributions distinct from, the fraction explained by amyloid PET burden. We also observed a statistical interaction between the tau PRS and amyloid levels, such that the relationship between amyloid and tau burden was particularly strong among individuals with a high tau PRS. Collectively, these results support the concept that genetic heterogeneity may partly explain individual variation in the relationship between amyloidosis and the development of tauopathy in AD.

The observed relationship between the tau PRS and worsened longitudinal cognitive trajectory in our data is important given the close relationship of tau accumulation with cognitive decline in AD [3, 46]. In addition, the independent validations against CSF p-tau and postmortem Braak stage offer further support for our findings (including via the gold standard of neuropathologic burden), though it should be noted that these measures are strongly but imperfectly correlated with tau pathology assessed by PET. In exploratory sensitivity analyses within this dataset (N=511 from the tau PET/PRS sample using the same MRI scanner), we found that higher tau PRS values were strongly associated with lower cortical thickness (p = 4.05 × 10⁻⁵, β = −0.15) within a composite AD signature region of interest (as described previously [27]), additionally supporting the potential clinical utility of the tau PRS score. Future work will be focused on more broadly evaluating the prognostic utility of the tau PRS. Regardless, we anticipate that our findings will stimulate additional needed studies in other cohorts, particularly as further large samples with GWAS and tau PET data mature in the future.

We took several steps to address limitations of sample and phenotype in this study, including utilizing conservative PRS calculation methods, applying recommended internal validation techniques, and pursuing validation studies with data independent from the training set. Nevertheless, additional replication in other cohorts will be important. The size of our discovery cohort was large for a tau PET study but was still relatively modest in comparison to many GWAS/PRS works, which could influence findings for some of the relatively uncommon variants (1≤MAF≤10%) included in the tau PRS. While our analyses of the population-based MCSA offer the benefit of generalizability, the relatively small proportion of individuals in the sample with cognitive impairment could impact the results. The MCSA population also lacks racial and ethnic diversity, which could have implications for the broader clinical applicability of genetic testing and counseling related to its findings.

This work has other limitations. Although outside the scope of this study, which focused on aggregate common genetic variation influencing overall AD pattern tau PET burden, future work assessing rare genetic variation, environmental factors, gene interactions with sex and other variables, and differential topographic distributions of tau would also be informative. Further, the AV-1451 tracer used for tau PET in this study measures AD-type mixed 3R/4R tau deposits with robust but not perfect specificity and sensitivity [36]. It should also be emphasized that our tau PRS is likely specific to AD-type tauopathy and would not be anticipated to generalize to non-AD neurodegenerative tauopathies. Finally, our tau PRS was drawn from a GWAS of cross-sectional tau PET data using the static measure of DNA variation, and thus is not designed to capture dynamic changes in transcription, translation, or methylation which may contribute to shorter-term longitudinal tau accumulation and widespread involvement.

In the future, a tau PRS could be useful as part of a multi-pronged approach to clinical risk stratification in individuals considered at risk for AD dementia. Specifically, for individuals found to have evidence of amyloidosis through CSF, PET, or novel plasma [45] biomarkers, a high tau PRS could serve as a noninvasive, cost-effective screening tool to guide prognostic counseling, follow-up intervals, and decision-making on therapeutic application or enrollment in clinical trials (Figure 3), understanding that the relationship of a tau PRS with future cognitive trajectory is likely complex. It is widely anticipated that future treatment regimens for Alzheimer’s disease may be comprised of individually-tailored combination therapies, modeled on approaches used for other complex disorders such as hypertension, cancer, and HIV/AIDS [48]. In this schema, the tau PRS could inform whether a preventive or therapeutic strategy targeting tau accumulation would be of high value, particularly before the development of substantial symptoms. As emerging disease-modifying therapies targeting amyloid, tau, and other mechanisms evolve [29], the tau PRS could also be gauged as a factor in predicting likelihood of treatment response to disease-modifying drugs. Further validation of our findings in other cohorts and extension to other relevant settings is warranted. Nevertheless, this novel work serves as proof of concept that a PRS for tau accumulation may provide a powerful tool for improved risk stratification and mechanistic understanding in AD.