Independent role of Alzheimer’s disease genetics and C-reactive protein on cognitive ability in aging

Apolipoprotein E ( APOE ) ε4 , the strongest genetic risk factor for late onset Alzheimer’s Disease (LOAD), has been associated with cognitive decline independent from AD pathology, but the role for other LOAD risk genes in normal cognitive aging is less studied. We examined the effect of APOE ε4 and several different polygenic risk scores (PRS) for LOAD on cognitive level and decline in aging, using longitudinal data from the UK Biobank. While PRS-LOAD including all variants (except APOE) predicted cognitive level, APOE ε4 and PRS-LOAD based on 17 non- APOE gene variants with strong association to AD (p<5e-8) predicted age-related decline in verbal numeric reasoning. The effect on decline were partly driven by four variants involved in the immune system. Those variants also predicted serum levels of the inflammatory marker C-reactive protein (CRP), but CRP did not mediate the effect on decline. Those findings suggest genetic variations in immune functions play a role in aspects of cognitive aging, that may be independent of LOAD pathology as well as systemic inflammation measured by CRP.


Introduction
Late-onset Alzheimer's disease (LOAD) is the most prevalent neurodegenerative disease and the major cause of dementia in older people ( Scheltens et al., 2016 ;Reitz et al., 2020 ). About 79% of LOAD risk have been attributed to genetic variations ( Gatz et al., List of abbreviations: LOAD, Late-onset Alzheimer's disease; APOE, Apolipoprotein; GWAS, Genome-wide association studies; AD, Alzheimer's disease; SNP, Singlenucleotide polymorphism; ADNI study, The Alzheimer's Disease Neuroimaging Initiative; CSF, Cerebrospinal fluid; UKB, UK Biobank; APP, Amyloid precursor protein metabolism/Aß-formation; RCT, Reverse cholesterol transport; CRP, C-reactive protein; PRS-LOAD, Genome-wide polygenic risk score for Late-onset Alzheimer's disease; APP PRS-LOAD, Amyloid precursor protein metabolism/Aß-formation pathway polygenic risk score; Tau PRS-LOAD, Tau protein binding pathway polygenic risk score; Lipids PRS-LOAD, Lipids metabolism pathway polygenic risk score; RCT PRS-LOAD, Reverse cholesterol transport pathway polygenic risk score; Immune PRS-LOAD, Immune response pathway polygenic risk score; LD, Linkage Disequilibrium; ICD-10, International Statistical Classification of Diseases  2006 ). While the apolipoprotein ( APOE) ε4 allele constitutes the strongest single genetic risk variant for LOAD, genome-wide association studies (GWAS) have identified multiple other risk variants associated with Alzheimer's disease (AD) ( Kunkle et al., 2019 ). A model with three genetically defined subtypes of AD were recently suggested ( Frisoni et al., 2022 ): (1) autosomal dominant AD with early onset, caused by a few rare gene variants with high penetrance, (2) APOE ε4 -related AD, where APOE ε4 constitutes a gene variant with intermediate penetrance and high frequency in the population ( ∼28 %), and 3) APOE ε4 -unrelated AD, influenced by a larger number of low-penetrance risk gene variants with generally later disease onset than APOE ε4 -related AD. Although the third subtype is the most common form of AD and has equally high heritability ( Lo et al., 2019 ), the genetics of AD in this group is least studied. Based on GWAS results, polygenic risk scores (PRS) can be calculated to estimate cumulative genetic disease risk across the whole genome and to study the impact of overall disease genetics on underlying mechanisms of the disease. PRS-LOAD (excluding the APOE locus) have been repeatedly associated with cognitive decline ( Kauppi et al., 2020 ;Kumar et al., 2021 ), as well as AD biomarkers    size for the link between PRS-LOAD on AD risk is higher in APOE ε4 non-carriers relative to carriers .
Pathway enrichment analyses of genes identified from GWAS have provided important evidence of a role for the immune system in AD ( Sims et al., 2020 ), and confirmed a genetic link to Amyloid β (A β)-, and tau pathways, already known to be important in the LOAD pathology ( Kunkle et al., 2019 ). Two other major pathways enriched among LOAD genes are lipid transport and endocytosis ( Sims et al., 2020 ). Some researchers have started to examine genetic risk load in those and other pathways in relation to disease endophenotypes, by calculating a PRS that is restricted to SNPs located in or near genes with a functional annotation to these pathways ( Hari Dass et al., 2019 ). For example, one study in 258 healthy old adults and 451 individuals with mild cognitive impairments from the ADNI study examined PRS-LOAD based on significant gene variants from the immune-, endocytosis-, and lipid-pathways ( Femminella et al., 2021 ). The immune PRS was associated with worse mini-mental state examination, and more pronounced neurodegenerations, while the endocytosis PRS was associated with tau-pathologies in cerebrospinal fluid (CSF) ( Femminella et al., 2021 ). The effect of genetic risk load related to the immune pathway is of particular interest given the well-documented role for neuroinflammation in AD, other aging-related diseases, and normal aging ( Franceschi et al., 2018 ;Li et al., 2021 ).
In the present study, we first tested if our previous findings that both APOE ε4 and PRS-LOAD predicted decline but not level of cognitive ability in aging ( Kauppi et al., 2020 ), could be replicated in the UK Biobank (UKB), a large-scale population-based cohort on a variety of diseases and traits, including cognitive performance (n = 488,247 with baseline cognitive data, age 40-70).
We hypothesized that such effect would be most prominent in the upper age span (60-70). Next, we tested the role for genetic risk load, excluding the APOE locus , in each of five major LOAD pathways on clinically healthy cognitive aging ( Fig. 1 ). Pathways were defined from previous enrichment analyses where risk genes were found for the following pathways: amyloid precursor protein metabolism/Aß-formation (APP), tau protein binding, protein metabolism, reverse cholesterol transport (RCT) and immune response ( Kunkle et al., 2019 ). We hypothesized that risk load in the immune pathway would be linked to cognitive decline also in normal aging. As the amyloid and tau pathways are more specifically linked to the pathophysiology of AD, we did not expect genetic risk load in those pathways to influence clinically normal cognitive aging. Further, we hypothesized that any observed link between immune PRS-LOAD and cognitive aging would be mediated by levels of C-reactive protein (CRP), a clinically important biomarker of systemic inflammation ( Luan and Yao, 2018 ). As the focus was to examine the genetic underpinning of normal cognitive aging, we only included cognitively healthy individuals.

Participants
The UK Biobank is a prospective cohort study with over 50 0,0 0 0 participants ( Sudlow et al., 2015 ). Participants of the study were adults aged between 40 and 70 years old at the baseline assessment. The baseline recruitment started in early 2006 and was completed in 2010 and included 22 centers across the UK. The UK Biobank has ethical approval from the NHS National Research Ethics Service as a research tissue bank (References 16/NW/0274 and 11/NW/0382) and all necessary patient/participant consent was obtained. The assessment process included a touchscreen questionnaire, a face-to-face interview with a research nurse, and blood sampling. Participants were excluded from the present analysis based on the following criteria; missingness of selected cognitive performance tests, non-British ancestry, failure of genotyping quality control (as defined below), and relatedness to another study participant (third degree or closer, KING estimated kinship co-efficient > 0.0442) ( Fig. 2 ).

Genotyping and quality control
DNA was extracted from stored blood samples, and genotyping was performed using 2 closely related arrays with over 95% common content, the UK BiLEVE array and the UK Biobank Axiom array. Samples were analyzed in batches of approximately 4700 items, initial quality control of the genotyping data was done by Affymetrix. A more detailed information on the array and sample processing is available on the UK Biobank website ( https://www. ukbiobank.ac.uk ). Quality control and imputation (to over 90 million SNPs, indels and large structural variants) has been performed by a collaborative group headed by the Wellcome Trust Centre for Human Genetics. Post-imputation quality control (QC) was performed based on genotype call rate < 10%, minor allele frequency (MAF) < 1%, and SNP missingness < 5%. For details on genotyping see ( Bycroft et al., 2018) .

Genetic risk assessment
Before calculating PRS, SNPs with ambiguous strand alignment were removed, as were SNPs within the APOE region (44.4-47.0 Mb on chromosome 19 on the hg19 assembly). Allele dosages for APOE ε 4 and ε 2 were instead included as separate variables in all analyses, coded linearly as 0, 1, or 2 ε 4/ ε 2 alleles for each individual.
Thereafter, PRS-LOAD were calculated using the summary statistics from the AD GWAS performed by Kunkle et al. ( Kunkle et al., 2019 ) including 21,982 AD cases and 41,944 cognitively normal controls, using the clumping plus p-value thresholding approach (C + T) ( Purcell et al., 2009 ;Wray et al., 2007 ). Following the C + T approach, linkage disequilibrium (LD) clumping was first performed in PLINK version 1.9 ( Purcell et al., 2007 ), by discarding SNPs within 10 0 0 kb of, and in r 2 ≥ 0.1 with another more significant SNP. The European sample of the 10 0 0 Genomes Project phase 3 ( The 10 0 0 Genomes Project Consortium, 2016 ) was used as LD reference panel for clumping, after removal of SNPs with genotype call rate < 1% and MAF < 1%. After clumping, 2 different SNP p -value thresholds were used for inclusion of SNPs to the PRS-LOAD, p < 1, that is, including all SNPs (N = 315,585), and p ≤ 5 × 10 −8 , that is, SNPs that were associated to AD in the GWAS by N = 17). Using PLINK version 1.9 ( Purcell et al., 2007 ), PRS for AD were then calculated for each individual by summing the alleles of the included SNPs weighted by each SNPs beta value from the GWAS ( Kunkle et al., 2019 ).

Pathway-based PRS-LOAD
Pathway-based PRSs were calculated based on the results from pathway enrichment analyses performed in the AD GWAS by Kunkle et al. ( Kunkle et al., 2019 ), identifying the following pathways: amyloid precursor protein metabolism/Aß-formation (APP PRS-LOAD), tau protein binding (Tau PRS-LOAD), lipids metabolism (Lipids PRS-LOAD) (including protein-lipid complex assembly, protein-lipid complex, protein-lipid complex subunit organization and plasma lipoprotein particle assembly pathways), reverse cholesterol transport (RCT PRS-LOAD) and immune response (Immune PRS-LOAD) as enriched among AD risk genes ( Fig. 2 ). Further in the following sections, we will address each pathway by its abbreviation stated before. First, genes belonging to each pathway were identified from Kunkle et al. ( Kunkle et al., 2019 , Supplementary Table 3). Thereafter, to identify SNPs linked to each gene, we defined start and end base positions of these genes according to RefSeq ( Pruitt et al., 2007 ) in genome build hg19, and we added 35 kb upstream and 10 kb downstream to these positions. To assign SNPs in the AD GWAS summary statistics ( Kunkle et al., 2019 ) to these genomic ranges, we used the function snps.to.genes from the package PAGWAS in R version 4.0.3. We thereby created a summary statistics file only containing SNPs related to one of the 5 pathways, which was used to calculate pathway specific PRSs, after removal of SNPs with ambiguous strand alignment and SNPs within the APOE region, as described above (including LD clumping for each pathway separately). As with the whole genome PRS two different SNP p -value thresholds were used for inclusion of SNPs to the PRS-LOAD, that is, including all SNPs outside the APOE loci ( p < 1), and snps SNPs that were associated to AD in the GWAS by Kunkle et al. at GWAS significance level ( p ≤ 5 × 10 −8 ; Supplementary Table 1). Polygenic risk scores were standardized to Z scores, that is, mean = 0, standard deviation (SD) = 1, where higher scores equate to increased risk.

Dementia diagnosis
As the primary interest was to examine the genetic underpinnings of normal cognitive aging, participants that were reported to have developed dementia were excluded from the analysis (both at baseline and during follow-up period). The International Classification of Diseases ICD-10 codes (corresponding to the UK Biobank data field ID; A81.0, F00, F01, F02, F03, F05, G30, G31.0, G31.1, G31.8, and I67.3) were used to identify participants with dementia if one or more of these codes were recorded as a primary or secondary diagnosis in the health records up to the date when data has been downloaded (February, 2022).

Serum CRP level
Serum levels of the inflammatory marker C-reactive Protein (CRP) (mg/L) were measured by immunoturbidimetric highsensitivity analysis on a Beckman Coulter AU5800 ( Milton et al., 2021 ). The distribution of serum CRP levels was significantly skewed and therefore transformed with a natural log (Supplementary Fig. 1).

Cognitive assessment
To assess cognitive function, participants were asked to complete several tasks on a touchscreen. These tasks were designed specifically for UK Biobank and completed by participants without supervision. Four tests were selected for the current analysis that were administered at the baseline recruitment period; Verbal number reasoning (VNR), Numeric memory (NM), Visual memory (VM), and reaction time (RT). The NM test were excluded in the longitudinal analyses due to inconsistent follow up design. All tasks have been shown to have moderate correlations with other standard cognitive tests in the same domains (r = 0.33-0.52) and showed moderate test-retest correlations over 4 weeks (r = 0.41-0.55; Fawns-Ritchie and Deary, 2020 ). The detailed information on each cognitive test could be found in the cognitive testing section of the UK Biobank website ( https://biobank.ndph. ox.ac.uk/showcase/label.cgi?id=10 0 026 ). Briefly, during the NM test (UKB Data-Field 4282) participants were asked to recall a 2-digit number after a short pause, whereafter the number of digits were increased by one. The test continued until the participant made a mistake or reached the maximum of twelve digits. The VNR test (UKB Data-Field 20016) consisted of thirteen logic/reasoningtype questions administered during a 2-minute time limit. RT (UKB Data-Field 20023) was measured using the mean response time (ms) in a symbol matching test. RT scores were significantly skewed and transformed with a natural log (ln). During the VM test (UKB Data-Field 399) participants were asked to memorize the positions of 6 card pairs, and then had to match them from memory with minimum possible errors. The score of this test is the number of errors that each participant made. VM scores were also significantly skewed, and zero-value inflated, therefore transformed with a ln + 1 formula. ( Lyall et al., 2016 ).

Statistical analyses
Analyses were run in 2 stages ( Fig. 2 ), where the exploratory stage constituted analyses aimed to confirm results from previous studies in UKB. In this stage, we first run cross-sectional analyses to examine the association of PRS-LOAD with level (baseline) of cognitive test performance, using linear models, with the 4 tasks described above as outcome variables, including age, age 2 , sex, APOE ɛ 4, and APOE ɛ 2, and the 10 first principal components for genetic ancestry (to control for population stratification) as covariates. Thereafter, we run longitudinal linear mixed-effect models to examine the effect of PRS-LOAD on change in cognition over time, using the same covariates. Here, NM was excluded due to lack of longitudinal data. Mean centered age was used as a timescale, and interaction with mean centered age was allowed for all covariates in the model. The model allowed for random subject-specific intercept.
In the main analysis stage ( Fig. 2 ), we first examined the role for PRS-LOAD on cognitive level and decline in aging using linear mixed effect models with the same covariates and design as above, but here limited to older individuals (60-70 years at baseline), as cognitive decline is limited before this age (e.g., Kauppi et al, 2020 ). Next, we used the same model design to examine the effect of pathway-based PRS-LOAD on cognitive level and slope, with separate models for each of the 5 pathways. Based on results from the exploratory stage, those analyses were restricted to the VNR test only. A Bonferroni corrected threshold of p < 0.01 (controlling for five pathway PRS-LOAD) was considered to be significant, and p < 0.05 was considered as significant at the uncorrected level. Thereafter, we used linear regression models to examine if each of the 5 pathway-based PRS-LOAD were associated with log CRP level, and if so, whether including logCRP levels at baseline as a covariate to the linear mixed models would modify any observed effect of pathway-based PRS-LOAD on cognitive slope. STATA (STATA 17.0) were used for all analyses.

Descriptive characteristics and group-level analyses
Selected baseline descriptive characteristics of the UK Biobank population stratified by APOE ɛ 4 carriers are presented in Supplementary Table 4, including cognitive task performance, CRP, BMI and smoking status. Among the 361,971 participants (46.2% male), the mean age at baseline recruitment was 56.85 years (SD = 7.92). Among the whole group of participants (40-70 years old), the mean predicted value of the VNR test was 6.5 at the age 55. Increased age predicted poorer performance on the VM and RT tasks, while age 2 predicted VM, RT and VNR, that is, poorer performance with accelerated older age (Supplementary Table 5). Around 80% of individuals with longitudinal data had a non-missing value of each cognitive performance test at 2 occasions only (Supplementary Table 6).

Exploratory analyses of PRS-LOAD on cognitive level and slope
In the exploratory stage of the project ( Figure 2 ), we examined the effect of PRS-LOAD on cognitive level and slope in individu-als from the full age range 40-70 at baseline. To maximize sample size, we first run cross-sectional analyses to examine the effect of PRS-LOAD (without the APOE locus) on cognitive test performance, calculated with 2 different PRS p -value thresholds; p < 1 (i.e., including all SNPs) and p ≤ 5 × 10 −8 (GWAS significance level), on each cognitive performance test at baseline (Supplementary Table 7). Genome-wide PRS-LOAD was associated with several cognitive performance tests, with the strongest effect seen on VNR at p -value thresholds < 1 (t = -8,15, p < 0.0 0 01), while no significant effect was seen for PRS-LOAD on RT, VM or the NM tests (Supplementary Table 7). Also, there was no significant effect of APOE ɛ 4 or APOE ɛ 2 carrier status on baseline level of any cognitive test ( p 's > 0.05, data not shown), except for an effect of APOE ɛ 4 on the VM test (ß = 0.0051, p = 0.018). Those results confirm previous analyses in UKB with similar design ( Hagenaars et al., 2016 ).
Thereafter, we run longitudinal analyses to examine the effect of APOE ɛ 4, APOE ɛ 2, and PRS-LOAD on cognitive slope among individuals age 40-70 years ( Fig. 2 ). APOE ɛ 4 was associated with increased age-related decline in VNR (ß = -0.012, p < 0.0 0 01), but no other associations between any of the genetic predictors and cognitive slope was seen (i.e., for any of the 3 cognitive tasks, Supplementary Table 5).

Effect of PRS-LOAD and pathway-based PRS on cognitive aging
In the main analysis stage ( Fig. 2 ), we tested whether AD genetics would impact cognitive aging, by fitting longitudinal models restricted to individuals aged 60-70 years at baseline ( Table 1 ). Based on cross-sectional analyses showing the strongest genetic effect for the VNR test, we limited analyses of this stage to VNR only. Here, we examined PRS-LOAD and our 5 pathway-based PRS-LOAD ( Fig. 1 ), based both on all SNPs (PRS p < 1) and GWAS-significant SNPs (PRS p < 5e −8 ). In this age group, PRS-LOAD significantly predicted decline when using GWAS-significant SNPs, but not when including all SNPs ( Table 2 ). There was no significant effect of APOE ɛ 4 or APOE ɛ 2 carrier status on VNR cognitive test level, but APOE ɛ 4 was again associated with increased age-related decline on VNR (ß = -0.0166, p < 0.003; Table 1 ).
When including all SNPs to the pathway PRS-LOAD, there was a suggestive effect of the Lipids-, APP-, and RCT-pathways on VNR slope, but no effects remained after controlling for multiple comparisons ( Table 2 ). When including only GWAS-significant SNPs (outside the APOE loci) to the pathway PRSs, only 1-5 SNPs remained in each pathway, and several SNPs were present in more than one pathway, and several pathway-based PRSs-LOAD were therefor moderately correlated (r = 0.51-0.59 for Immune and RCT, Lipids and Tau, APP and RCT pathways) and Lipids and APP pathways were strongly correlated (r = 0.77) (Supplementary Table 8). Controlling for multiple comparisons, only the he immune PRS-LOAD. Based on 4 SNPs, were significantly associated with steeper cognitive slope (immune: ß = -0.0069, p = 0.009). The one (1) SNP that was present in the RCT pathway at this p -threshold was also significantly associated with decline (ß = -0.0087, p < 0.001), but this SNP was also part of the immune pathway. Of note, the immune pathway contained the largest number of genes among the pathways (Supplementary Table 2), and the correlation between the immune PRS-LOAD that contained only GWAS significant SNPs and the immune PRS-LOAD containing all SNPs was low (r = 0.08). The other pathways contained fewer genes, and consequently showed moderate to high correlation between the score including all SNPs and the score including only GWAS significant SNPs (r = 0.4-0.7). Mean centered age (65 years) was used as a timescale (age). * * p = < 0.01, * * * p = 0.001. RT, reaction time; VM, visual memory (measured as number of mistakes); VNR, verbal numeric reasoning.
Next, we aimed to examine if the genetic effect of Immune PRS-LOAD on cognitive slope was modified by logCRP ( Table 3 ). Including logCRP in the model, the association between the immune PRS-LOAD and slope of the VNR task performance was slightly attenuated but remained statistically significant (ß = -0.0063, p = 0.020). There were no interactions between logCRP and immune PRS-LOAD on cognitive performance ( p > 0.05, data not shown).

Discussion
In this study, participants from the UK Biobank were used to examine the role of genetic risk factors of AD, including both APOE ɛ 4 and risk load in disease-relevant pathways, on cognitive ability and change across 10 years. To investigate the influence of genetic risk factors of AD on clinically healthy cognitive aging we focused on individuals that did not develop dementia during the study period, that is, to avoid secondary effect on cognition due to pathological changes related to clinical dementia. In our longitudinal models, APOE ɛ 4 predicted decline but not level of verbal numeric reasoning (VNR). No genetic effects were seen for the 2 other cognitive measures included, reaction time and visual memory, but due to differences in quality of tests in UKB, we here avoid interpretations of the results in relation to specific cognitive domains. Those results provide large-scale support for previous findings indicating a function of APOE ɛ 4 not only in AD , but in brain aging in a broader sense ( Honea et al., 2009 ). For example, breakdown of the blood-brain barrier in APOE ɛ 4 carriers was recently linked to age-related cognitive decline independently from LOAD pathologies ( Montagne et al., 2020 ). However, as most individuals who develop LOAD do not carry the APOE ɛ 4 allele , the primary purpose of this study was to investigate the role of non-APOE genetic risk factors of LOAD in cognitive aging. We found that PRS based on non-APOE gene variants were associated to cognitive level when including all SNPs, while age-related cognitive decline was predicted by PRS only when limiting the PRS to SNPs that were previously associated to AD at GWAS-significant level ( p < 5e-8). By categorizing risk variants into 5 pathways with documented genetic enrichment in AD, we aimed to gain more biologically informative results than from only studying genetic risk across the whole genome. In healthy old adults (aged 60-70 at baseline), the effect on cognitive decline, measured by VNR, was again strongest for genetic risk load based on GWAS-significant SNPs only for all pathways except APP. Risk load based on variants annotated to the immune pathway (including one SNP also annotated to RCT) were significantly associated with steeper cognitive decline, while risk load in APP-and lipid pathways showed trend-level associations with cognitive performance slope. Notably, those effects were observed when risk load was assessed using only the few GWAS significant variants that were present in each pathway. Thus, those results suggest that mainly gene variants with a strong effect on AD, but with diverse biological functions, are linked to cognitive decline also in clinically healthy aging, and that this effect may be somewhat more prominent for variants with a role in the immune system.
In the first stage of this study, we tested for large-scale replication of our previous findings of a link between both APOE ɛ 4 and PRS-LOAD on decline, but not level of cognition in clinically healthy old adults ( Kauppi et al., 2020 ). Similar results were also reported by Gustavson et al., associating AD PRS with memory decline ( Gustavson et al., 2023 ). Confirmatory analyses of the whole sample, aged 40-70 at baseline, showed replication of the effect of APOE on decline, while higher PRS-LOAD was instead linked to lower level, but not decline, of the VNR task. As age-related decline in most cognitive domains (e.g., episodic memory and processing speed) has been found to start at about age 60 at group level (e.g., Kauppi et al., 2020 ), we restricted the main analyses to participants aged 60 and above at baseline. In the older subgroup, we found a significant but weak effect of PRS-LOAD on VNR decline, but in contrast to our previous study, this was only seen when restricting the PRS to GWAS-significant SNPs ( p < 5e-8), and not for a PRS calculated based on all SNPs ( p < 1). However, there might not be direct comparison, as the work by Kauppi et al. had an extensive test battery with reliable tests representing several cognitive domains, which allowed for calculations of a cognitive composite score reflecting cognitive aging in a broader sense.
Other studies have shown that optimal discrimination of AD case-control status based on PRS can be achieved using a more stringent p -threshold ( Leonenko et al., 2021 ;Ritchie et al., 2020 ), while cognitive ability is typically best predicted by PRS models using all SNPs ( Koch et al., 2021 ;Coombes et al., 2020 ). This is likely related to a higher polygenicity of cognitive performance than of AD. Thus, it is possible that our results reflect that cognitive decline in the UKB population is more strongly linked to preclinical AD than in our previous study, where longer followup time allowed to restrict analyses to participants that remained non-demented for 6 years after the last time point. Lack of replication could also be related to cohort differences including the older age span and longer follow-up time in the previous study, and differences in cognitive test paradigms. In the well-established Lothian birth cohort ( Harris et al., 2014 ;Ritchie et al., 2020 ), PRS-LOAD were associated with baseline cognitive level, but had only marginal and nonsignificant relations with the gradient of cognitive performance slope.
APOE ɛ 4 was linked to decline in the VNR task both in the whole group (40-70) and when restricting analyses to only individuals in the age span of cognitive aging (60-70), but unrelated to both level and slope of the VM and RT tasks. Importantly, UK Biobank cognitive tests were less stable across time than well-validated tests administered under standardized conditions; VM test had the lowest test-retest reliability, while VNR and RT tests showed relatively good stability ( Fawns-Ritchie and Deary, 2020 ).
In the next step, we examined the role of genetic risk load in biological pathways in relation to cognitive aging, by focusing on decline of the VNR test in aging (from age 60 and above). A previous smaller study also investigated the role for immune-and other pathway PRS-LOAD in cognitive ability and other AD endophenotypes, and found that pathway-based PRSs-LOAD (including APOE ) were generally less predictive than an overall PRS or APOE alone ( Darst et al., 2017 ). Here, we did not include SNPs from the APOE locus, but were still able to detect an effect of the immune-PRS LOAD, using four GWAS significant ( p < 5e-8) SNPs only, on more pronounced cognitive decline with a somewhat stronger effect than the genome wide PRS-LOAD (17 SNPs above GWAS significance, excluding APOE ), and somewhat weaker than the effect of APOE ɛ 4 . However, this effect was not specific to the immune pathways, but was also seen for the RCT pathway, where only one (1) SNP was included at GWAS significance level. It should be noted though that there is genetic overlap between pathways, and the single RCT SNP in the clusterin gene (CLU) was annotated to both the RCT and the immune pathway. Nevertheless, results point to an effect of strong AD risk genes with diverse biological functions on cognitive decline also in clinically healthy old adults. The results justify further investigations of the role for genetic variations in immune-related genes, and how those may relate to neuroinflammation as a risk factor for more pronounced cognitive decline in aging and dementia. Results also warrant further investigations including reliable measures of other cognitive domains to investigate if the effect seen is domain specific or reflect cognition in a broader sense.
To further follow-up on those findings, we also investigated if the link between higher immune PRS-LOAD and more pronounced cognitive decline, measured using VNR, could be mediated by levels of the serum inflammatory marker CRP. Levels of chronic inflammation increase with age, supported by a concept often referred to as "inflammaging" ( Santoro et al., 2021 ). Individual variation in inflammatory markers in aging has been linked to various diseases, including AD, and dementia ( Giunta et al., 2008 ), as well as cognitive aging ( Fard and Stough, 2019 ). The nonspecific marker of inflammation CRP plays an important role in the monitoring of inflammation, neurodegeneration, tissue injury, and recovery ( Luan and Yao, 2018 ), and has been associated with cognitive performance in both UK Biobank and other samples ( Milton et al., 2021 ;Rentería et al., 2020 ). As there are no fully identified mechanistic links between inflammation and cognitive aging, it is possible that inflammation could rather constitute a biomarker of rate of aging than being causally linked to cognitive functioning ( Mooijaart et al., 2011 ). GWAS on CRP levels has found inflammation to be a somewhat heritable trait ( Said et al., 2022 ), but to the interplay between genetic factors, environmental events and lifestyle factors on individual variations in chronic inflammation need to be better revealed. Here, we first demonstrated that higher Immune PRS-LOAD predicted lower CRP levels. Interestingly, higher CRP is associated with lower cognitive performance, while high genetic risk for AD based on both Immune PRS-LOAD and APOE ɛ 4 were associated with lower blood CRP levels. It was reported previously that APOE ε4 has been associated with lower levels of blood CRP levels in cognitively healthy middle-aged and older populations ( Martiskainen et al., 2018 ), those relations may reflect a genetically determined suppression of the immune system, that is, chronically lower levels of systemic inflammation, that may make the individual more vulnerable to pathogens. The cross-sectional association between CRP on cognition may instead reflect ongoing infections with momentary negative influence on cognitive ability ( Wang et al., 2022 ). This was however not directly tested within the current study. Our results confirm that AD risk load in Immune PRS-LOAD are linked to levels of inflammation similarly to APOE ɛ 4 . We also found that the effect of genetic risk load in the immune PRS-LOAD on cognitive aging was not mediated by CRP. Our results imply that AD risk load in immune-related genes may have parallel effects on chronic systemic inflammation and cognitive decline, or that CRP levels do not capture all aspects of lifelong inflammation. However, it is possible that other inflammatory markers would better capture the role for immune-related AD risk genes in cognitive aging. CRP is primarily a marker of acute inflammation and not an ideal marker for chronic inflammation. The association of immune PRS-LOAD with cognitive decline in clinically healthy participants may be mediated by neuroinflammatory processes that are genetically influenced, but unrelated to systemic inflammation measured in serum. As aging itself is the strongest risk factor for AD, normal cognitive or biological aging is difficult to discriminate from prodromal AD without biomarkers of AD pathologies. Thus, those results need to be followed up in cohorts with CSF or blood-based biomarkers of amyloid and tau.
Despite the observed link between immune PRS and cognitive decline, we found that baseline CRP levels did not predict future decline in cognition in the UKB. Those results contrast previous reports of a link between CRP and cognitive decline ( Rentería et al., 2020 ;Zheng and Xie, 2018 ), although some studies did also fail to detect a link to cognitive decline ( Mooijaart et al., 2011 ). Increased baseline levels of CRP ( Engelhart et al., 2004 ;Schmidt et al., 2002 ) and other inflammatory proteins in plasma ( Bettcher et al., 2021 ) in healthy adults have also been associated with subsequent development of all-cause dementia or AD. A recent metaanalysis of 170 studies demonstrated an increased peripheral level of high-sensitivity CRP in individuals with AD diagnosis compared to healthy controls ( Shen et al., 2019 ), although some studies found opposite results with lower levels of CRP in individuals with AD ( O'Bryant et al., 2010 ). Also, sex specific characteristics should be accounted for, as several reports found that CRP level had statistically significant negative association with cognitive functioning only in females ( Canon and Crimmins, 2011 ;Watanabe et al., 2016 ), while others report an effect in males only ( West et al., 2020 ).

Strengths and limitations
The major strengths of this study include the large sample size, with participants from the well-established UK Biobank. However, several limitations should also be considered. By design, UK Biobank is not representative to the general population, as no random sampling was performed. Although the cognitive tests in the UK Biobank are novel and standalone versions of cognitive tests, these tests were self-administered from a computer, within a limited time period, which supposedly could introduce some bias. While quality of the cognitive tests might not meet the standard of clinically used neuropsychological assessment, several tests showed substantial concurrent validity ( Lyall et al., 2016 ), also cognitive scores in participants with repeat measurements as previously reported have a reasonable stability and showed testretest reliability. Nevertheless, the generalization to cognitive aging based on the used cognitive assessments is fragmentary, due to the lack of long-term follow up, absence of biomarkers that can discriminate between preclinical AD and nonpathological aging and incomplete list of available cognitive domains. As the study population was a middle-aged cohort, survival bias should be acknowledged.
In conclusion, our findings suggest that LOAD risk gene variants with strong effect on AD risk related to several major disease pathways, including immune function, influence decline in verbal numeric reasoning in clinically healthy aging. To what extent results may be generalized to other cognitive domains remains to be revealed. Yet, genetic risk load in immune-response pathways and actual level of inflammation were independent predictors of cognitive aging. Future studies need to investigate other potential mechanisms for the link between immune-related LOAD risk genes and cognitive aging than CRP, as well as to what extent those effects reflect preclinical onset of AD pathologies.

Disclosure statement
The authors declare that they have no competing interests.