Proof-of-concept recall-by-genotype study of extremely low and high Alzheimer’s polygenic risk reveals autobiographical deficits and cingulate cortex correlates

Genome-wide associations studies (GWAS) demonstrate that genetic risk for Alzheimer’s disease (AD) can be partly explained by the cumulative impact of thousands of single-nucleotide polymorphisms (SNPs) [1, 2]. Downstream analysis of AD genetic architecture has now uncovered many novel mechanistic insights into the aetiology of AD, suggesting multiple, novel components of distinct molecular aetiology [3, 4]. Furthermore, polygenic risk scores (PRS) derived from these AD GWAS show predictive capacity to identify individuals at high risk for AD, which may provide clinical utility for early detection, intervention, and diagnosis [5‐8]. The AD-PRS has been further linked to an increase across a range of preclinical AD features, including peripheral, neuroimaging, and cognitive markers of brain health, suggesting common AD genetic risk may manifest before the onset of symptoms, via alterations in neurobiological process that increase susceptibility in later life.

A significant proportion of common genetic risk for AD can be explained by the APOE locus, which may manifest via alterations in brain structure/function. Multiple GWAS of neuroimaging features have linked APOE status to amyloid burden [9], white matter hyperintensities [10, 11] and functional-temporal coherence of blood oxygen level dependency (BOLD) signals [12, 13]. Ongoing population AD-PRS studies that consider APOE have demonstrated associations with feature of brain health such as hippocampal volume [14]. As APOE status accounts for a considerable proportion of AD genetic risk, it remains largely unknown how other AD genetic risk factors link to brain structure and function. However, recent studies demonstrate an association between AD-PRS which excludes the APOE region (non-APOE AD-PRS), where non-APOE AD-PRS is independently linked to cognitive trajectories [15] and neuroimaging features of brain health such as hippocampal volume [16‐18] and cerebrovasculature [19, 20].

However, the shared variance between AD-PRS and these preclinical features is small, limiting the utility of integrating biomarkers into prediction models of preclinical AD. One recent strategy is to assess risk in individuals with extremely high AD-PRS (defined as over two standard deviations from the sample mean). Exploring risk for AD in individuals with extreme AD-PRS values allows us to explore in vivo correlates of genetic susceptibility while limiting confounding and reverse causation that exist in samples of clinically ascertained participants [21].

Understanding non-APOE AD-PRS contributions to brain health remain a critical avenue of exploration when linking the common genetic architecture of AD to brain health, as to build a multiplex model of AD susceptibility and quantify downstream effects. Here, we first describe the recall-by-genotype (RbG) approach for neuroimaging non-APOE AD-PRS, based on the genotyping and AD-PRS estimation across a larger, population sample. We have previously demonstrated that RbG studies based on PRS with smaller number of participants can successfully reproduce observations made in larger studies. By targeting groups at the extremes of PRS, the number of participants required to provide sufficient statistical power to observed differences is drastically reduced. For example, in a recall-by-genotype study for schizophrenia PRS, we were able to observe increased prevalence of psychotic symptoms and striatal reward–linked brain activity in a sample of approximately two hundred asymptomatic participants [22], previously observed in samples of thousands [23, 24].

By assaying a non-APOE AD-PRS in a large, genotyped population, we can recruit a subset of individuals from the general population who have either extremely low or high AD-PRS, enriching the sample for a large amount of variation in AD-PRS. There is considerably increased AD-PRS risk (as indexed by odds ratio [OR]) in AD-RPS between the 1st and 10th decile (more than 30-fold difference in AD risk the current study, offering considerably more power (approximately fourfold increase) than an opportunistic sample (see ‘Materials and Methods’ for further details)). This study aims to understand mechanisms by which the cumulative effect of risk SNPs for AD affect the brain. We hypothesise that burden of AD risk SNPs will be related to cognitive performance and in vivo measures of brain health such as macrostructure—based on prior observations in AD, mild cognitive impairment, and at-risk populations. We expect this research will help to elucidate biological processes by which AD genetic variants may lead to reduced brain health. By identifying the preclinical features that are linked to non-APOE genetic risk, we hope to identify neurocognitive features that reflect AD susceptibility before the onset of symptoms. We anticipate this work may guide future detection, intervention and prevention strategies that mitigate or offset these components of adverse brain health.

We recruited sixteen individuals with either very low or high (> ± 2 SDs from population mean) to perform an assessment of autobiographical memory and T1-weighted structural MRI. Based on our estimation, an opportunistic sample (assuming a random selection of AD-PRS from the broader population) would have been significantly underpowered to detect associations with AD-PRS. However, as we were able to capitalise on variance in AD-PRS across the wider, larger sample, we were able to quadruple the power to detect AD-PRS effects, reducing the sample required from N = 90 to N = 16 [21]. We were further able to limit confounding with stringent exclusion criteria parameters.

The participants with a high AD-PRS showed marked reductions in (i) autobiographical memory and (ii) cingulate thickness compared to the low AD-PRS group. However, we did not observe an association between AD-PRS and hippocampal volume, suggesting the shared variance may be smaller than our study power afforded or that associations may be explained by confounding from genetically correlated traits such as intelligence/years of education. Several studies have demonstrated a negative association between AD-PRS hippocampal volume across the lifespan [14, 17, 35, 36]; however, a significant proportion of the shared variance has been attributed to an association with the APOE locus [14, 42]. Our study broadly conforms to two prior observations. First, a substantial body of evidence has previously demonstrated that increased AD genetic risk is linked to reduced cognition, with studies showing negative genetic correlations between cognition and AD common variants and AD-PRS studies supporting this observation. Second, this study supports prior observations that cortical thickness of midline / cingulate structures are reduced in individuals with high AD-PRS [37]. More specifically, reduced anterior cingulate thickness has previously been linked to AD genetic risk within a endocytosis pathway-specific AD-PRS analysis [43], across MCI groups and via associations with meta-memory in AD [44]. More recently, a negative association between AD-PRS and caudal anterior cingulate thickness has further been demonstrated in a large, pre-pubescent sample (N_ABCD > 4000), suggesting that this alteration may be a risk factor that is expressed across the lifespan [45]. While the sample age range (58–76) and autobiographical memory assessment deficit do not allow us to delineate between prodromal and early disease effects [46], we suggest that the assays we report are an initial showcase of the ability to detect AD-PRS-related differences in significantly smaller samples that could be employed at point in the lifespan, using biological readouts that are more complex and not scalable in large samples / cohorts. While we observed converging evidence linking AD-PRS with cingulate thickness, the study must be considered with the following limitations. While the sample size provided > 80% power to detect an association with AD-PRS, a larger or replication sample would have allowed us to further assess the validity of our findings. While this was not possible for the Survey of Autobiographical Memory (SAM), we did replicate the negative association between AD-PRS and cingulate thickness in a larger sample (N = 31,966). Here, we used a larger AD GWAS dataset to estimate AD-PRS [41], which became available after recruitment, an advantage of working with secondary data and updated GWAS derivates, compared to our recall-by-genotype study, which was constrained by AD-PRS estimations made before recruitment (see last section of the limitations section within this discussion). The effect size of this association was considerably smaller, this is however to be expected in a sample with additional potential sources of confounding and heterogeneity [47]. Second, we did not have a comparable group with an average AD-PRS (for example, participants with an AD-PRS in a middle decile). Therefore, any group differences we observe here may reflect higher SAM and thickness in the very low AD-PRS group, rather than preclinical alterations in the high AD-PRS group. Third, we also acknowledge that the cross-sectional design does not reveal important information such as MCI / AD conversion or trajectories which would have helped to establish the utility of the observed features in the prediction of future neurodegeneration. Fourth, while individual AD-PRS can be considered in relation to the larger sample from which they were derived, it is currently a challenge to provide an individual context about their AD-PRS as a standardised assessment. Future studies of AD-PRS working towards increased portability and generalisability, so individuals’ genetic risk can be considered independent from the sample from which their AD-PRS were estimated, may prove useful for generating AD-PRS based on existing normative samples [5, 6]. Fifth, we acknowledge that AD-PRS only represents a summated total of all known, common AD risk variants. While there are initiatives to assess AD genetic risk via the partitioning of PRS into specific biological pathways, AD-PRS may still reflect a heterogeneous, biologically unspecific estimate, making it difficult to mechanistically implicate specific causal processes. Last, we acknowledge that recall-by-genotype studies using PRS require investigators to recruit on the basis of a specific GWAS data set and with a specific PRS approach (for example, a specific P-threshold, PRS method, and GWAS training data), where variability in the process has the potential to change the position of individuals within the wider recall sample and their respective position in the AD-PRS groups, limiting their flexibility compared to compared to re-analyses of secondary data based on newer AD GWAS derivatives. While ongoing studies are working towards a standardised metric for PRS assessment [48], recall-by-genotype approaches are likely to continue being affected by ongoing GWAS and downstream methods development.

While our observations suggest that prodromal or early markers of AD pathophysiology can be observed in the high AD-PRS group, we suggest moreover that the recall-by-genotype design demonstrates that selecting specific individuals based on their PRS reflects appropriate biological features, which has several translation applications. For instance, (1) while we collected self-report memory / structural MRI features, more complex biological readouts could be assayed that are not scalable in big data such as expensive biological experiments (e.g. iPSc collection and generation) [3]; (2) genetics-first characterisation can occur across the entire lifespan, establishing precise timelines for trajectories of genetic risk, enhancing prediction, intervention opportunities, adding a layer of precision to the commonly used characteristics [49], and (3) empower clinical trials for individuals at increased genetic risk, with implications for timeliness, power, and cost [50]. Recall-by-genotype of less common, missense single-nucleotide variants (for example, functional, amino-acid change conferring SNPs within genes such as ABI3, PLCG2, TREM2) could also further provide mechanistic insight into the aetiology of preclinical AD, with known functional roles in modifying immune system physiology [51, 52], which have further been linked to features of brain health [53‐55]. In conclusion, we document the first recall-by-genotype study for AD-PRS and observe neurocognitive features with distinct profiles between participants with very low and high AD-PRS. This recall-by-genotype approach further permits the exploration of experimental preclinical methods currently not available in large neuroimaging-genetic databases such as MRI-derived measures of cerebrovascular and neurometabolic structure and function, respectively, as well as molecular characterisation via stem cell-phenotyping [3].