Associations of circulating C-reactive proteins, APOE ε4, and brain markers for Alzheimer’s disease in healthy samples across the lifespan

The apolipoprotein E gene ε4 allele (APOE ε4) and higher circulating level of C-reactive protein (CRP) have been extensively investigated as risk factors for Alzheimer's disease (AD). Paradoxically, APOE ε4 has been associated with lower levels of blood CRP in middle-aged and older populations. However, few studies have investigated this intriguing relation and its impact on neurological markers for AD in younger ages, nor across the whole lifespan. Here, we examine associations of blood CRP levels, APOE ε4, and biomarkers for AD in a cognitively healthy lifespan cohort (N up to 749; 20-81 years of age) and replicate the findings in UK Biobank (N=304 322; 37-72 years of age), the developmental ABCD study (N=10 283; 9-11 years of age), and a middle-aged sample (N=339; 40-65 years of age). Hippocampal volume, brain amyloid-β (Aβ) plaque levels, cerebrospinal fluid (CSF) levels of Aβ and tau species, and neurofilament protein light protein (NFL) were used as AD biomarkers in subsamples. In addition, we examined the genetic contribution to the variation of CRP levels over different CRP ranges using polygenic scores for CRP (PGS-CRP). Our results show APOE ε4 consistently associates with low blood CRP levels across all age groups (p <0.05). Strikingly, both ε4 and PGS-CRP associated mainly with blood CRP levels within the low range (<5mg/L). We then show both APOE ε4 and high CRP levels associate with smaller hippocampus volumes across the lifespan (p <0.025). APOE ε4 was associated with high Aβ plaque levels in the brain (FDR-corrected p =8.69x10-4), low levels of CSF Aβ42 (FDR-corrected p =6.9x10-2), and lower ratios of Aβ42 to Aβ40 (FDR-corrected p =5.08x10-5). Blood CRP levels were weakly correlated with higher ratio of CSF Aβ42 to Aβ40 (p =0.03, FDR-corrected p =0.4). APOE ε4 did not correlate with blood concentrations of another 9 inflammatory cytokines, and none of these cytokines correlated with AD biomarkers. CONCLUSION: The inverse correlation between APOEε 4 and blood CRP levels existed before any pathological AD biomarker was observed, and only in the low CRP level range. Thus, we suggest to investigate whether APOEε 4 can confer risk by being associated with a lower inflammatory response to daily exposures, possibly leading to greater accumulation of low-grade inflammatory stress throughout life. A lifespan perspective is needed to understand this relationship concerning risk of developing AD.


Introduction
Late onset Alzheimer's disease (AD) is the major cause of dementia, but our understanding of its aetiology is incomplete (Scheltens et al., 2021). Clinical manifestation of AD typically occurs after 60-65 years of age, and includes declined cognitive performance and brain atrophy (Burns and Iliffe, 2009), whereas the pre-symptomatic phase of AD could start 15-20 years before the clinical diagnosis. The most popular hypothesis for the development of ADthe amyloid beta (Aβ) cascade hypothesis (Beyreuther and Masters, 1991;Selkoe and Hardy, 2016)suggests the disease is initiated by extracellular aggregation of Aβ42 plaques, which in turn causes accumulation of intracellular neurofibrillary tangles composed of hyperphosphorylated tau proteins, and subsequently induces neuroinflammation in the brain. Ultimately, the cascade leads to widespread neuronal and synaptic dysfunction, brain atrophy, and cognitive impairment. Whereas the Aβ-cascade hypothesis explains early-onset AD to a certain extent, it appears to play a smaller role in the more common late-onset form of AD. Recent genome-wide association studies (GWAS) for late-onset AD have reported up to 30 genomic risk loci (Jansen et al., 2019;Kunkle et al., 2019) that implicate involvement of Aβ processing, immune response and lipid metabolism in AD development. However, most of these variants show very weak associations with AD (Scheltens et al., 2021).
The apolipoprotein E gene (APOE) ε4 allele is the strongest genetic risk factor established for late onset AD (Corder et al., 1993;Strittmatter and Roses, 1995). APOE has three major variants: ε2, ε3, and ε4. Carriers of one APOE ε4 copy have 3-4-fold increased risk of developing AD compared to non-carriers, whereas the risk associated with carrying two ε4 copies has been estimated to be up to 15-fold (Farrer et al., 1997;Neu et al., 2017). APOE ε4 carriers have a younger age of onset for AD (Bonham et al., 2016;Sando et al., 2008), and several studies have shown that APOE ε4 affects AD in an age-dependent manner, with highest risk for 60-70 years old persons, but smaller effect for those above 75 years (Bonham et al., 2016;Corbin et al., 2018). Nevertheless, the molecular mechanisms for how APOE ε4 increases AD risk are unclear (Belloy et al., 2019). Proposed mechanisms of APOE ε4 are largely linked to the Aβ-cascade hypothesis, including increased production and decreased clearance of Aβ in the brain, increased formation of intracellular tau tangles, perturbed synaptic function and altered microglial response (Yamazaki et al., 2019). However, APOE ε4 has also been associated with reduced cognitive performance and smaller hippocampi in younger ages well before brain Aβ plaques are detectable (Filippini et al., 2009;Luiza K. Axelrud et al., 2018;O'Dwyer et al., 2012). Thus, other physiological pathways linking APOE ε4 to AD risk remain to be explored.
Among several postulated mechanisms by which APOE ε4 could impact AD risk, immune-related pathways have gained increased attention (Yamazaki et al., 2019). In particular, APOE ε4 has been associated with low concentration of CRP (Eiriksdottir et al., 2006;Haan et al., 2008;Martiskainen et al., 2018;März et al., 2004). CRP is an acute-phase serum protein, produced mainly by the liver but also other organs, including the brain (Du Clos, 2000;Dumitrescu et al., 2020). High CRP levels typically indicate a recent infection and sustained medium concentration may suggest chronic inflammation (Dumitrescu et al., 2020). For AD, although the general trend is that high CRP levels correlate with high risk of developing AD, such a trend may not be causal (Prins et al., 2016). Furthermore, a recent study suggested the effect of blood CRP levels on AD may depend on the age of the sample (Gabin et al., 2018). Despite that both APOE ε4 and CRP derived from blood measures (Desikan et al., 2015;Fernandes et al., 2020;Morgan et al., 2019;Silverman et al., 2012) have been extensively studied in terms of their association with AD, less efforts have been made to understand the intriguing relation between the three. Although the association between APOE ε4 and CRP levels and AD development may depend on age, previous studies were rarely performed on samples younger than the currently accepted age of pre-symptomatic AD (Eiriksdottir et al., 2006;Haan et al., 2008;Martiskainen et al., 2018;März et al., 2004). Thus, interpretations of the APOE ε4 vs. CRP relation in terms of AD risk are incomplete.
Here, we investigate associations of APOE ε4, blood CRP levels and established biomarkers for AD in a cognitively healthy adult lifespan cohort (N up to 749; 20-81 years of age), and in three replication datasets comprising different age-ranges: a UK Biobank subsample (N = 304,322; 37-72 years of age), the developmental ABCD study (N = 10,283; 9-11 years of age), and a middle-aged sample (N = 399; 40-65 years of age). Total hippocampal volume (HippV), brain amyloid-β (Aβ) plaque levels measured by PET-imaging, and CSF levels of Aβ, tau and NFL were examined as AD biomarkers. We also studied genetic contributions to the variation in CRP levels in different ranges using a polygenic score for CRP (PGS-CRP). We hypothesized that the previously observed negative APOE ε4-CRP relations could also exist in young people well before the accepted AD pathological biomarkers are detectable. If so, that may indicate that any effects of lower inflammatory response in APOE ε4 carriers could accumulate throughout the lifespan.

Methods
Fig. 1 shows our analysis steps, datasets and sample age distributions, and analytical models used in the present study. In brief, we used our local LCBC sample for discovery and three independent external datasets (UKBB, UB-BBHI and ABCD) for replications.

Participants
The study was approved by the Regional Committee for Ethic in Medical Research in Norway. All participants provided written informed consent.
COGNORM sample. Participants with CSF biomarkers, cytokine levels, and DNA genotypes from the COGNORM-study (Idland et al., 2017) 2012 to 2013, Oslo University Hospital and Diakonhjemmet Hospital, Oslo) were included. The COGNORM-study included 172 patients of at least 65 years of age undergoing elective gynecological, urological and orthopedic surgeries in spinal anesthesia in the two hospitals. Patients with dementia, previous stroke with sequelae, Parkinson's disease and other conditions that probably affect cognition, were excluded. In the present study, 99 participants were included (64-93.0 years of age; 49 women; see Figs. S8 and S9 for CSF and blood marker distributions). Body mass index (BMI) was also collected for these participants.

UK biobank population (UKBB).
Participants in the UK Biobank (UKBB) (Sugden et al., 2019) with HippV, CRP levels and DNA genotype information, were included. Ethical approval was obtained from the National Health Service National Research Ethics Service (Ref. 11/NW/ 0382) and all participants provided written informed consent. The dataset released February 2020 was used, including 502,507 participants, of whom 40,682 had undergone MRI scanning, 487,409 have DNA genotypes, and 468,569 had CRP measured. Participants with diagnosis of any neurodegenerative disease (N = 2,030; ICD10, G30-32), any injury to the head (N = 12,722; ICD10 S00-09), were excluded from our study. Participants with ambiguous APOE genotypes (N = 12,252), and that were genetically related with at least one other participant, were also excluded. In total, 304,322 participants (37-72 years of age; 164,303 women; 12,123 having CRP > 10 mg/L) were included to study the association of APOE ε4 with CRP levels; and 26,573 participants (45-82 years of age; 13,894 women; 1,581 having 2 MRI scans) were used to study the association of CRP and APOE ε4 with HippV.
ABCD population. The ABCD study aims to discover factors affecting human brain development from childhood to adolescence (Casey et al., 2018), and has included > 10,000 children 9-10 years old. In the present study, we analyzed the full release 3.0, which also included brain scans for the second timepoint for a subset. There was no blood CRP measure available. In total, 10,283 unique participants (4,865 girls; 9-11 years of age), among which 3,519 have two MRI scans, were included here.
UB-BBHI population. Samples were obtained from participants in the Barcelona Brain Health Initiative (BBHI) (Cattaneo et al., 2018). BBHI is a population based prospective study including cognitively preserved individuals without a diagnosis of a medical neuropsychiatric condition. Inclusion in our study was based on the availability of blood CRP measures, genotypes, and MRI scans. In total, 339 individuals were included (179 women; 41-67 years of age).

LCBC, COGNORM and UB-BBHI Sample.
Genotyping and imputation for participants of these three samples were performed previously using identical protocols (Walhovd et al., 2020). Briefly, buccal swabs were collected and genotyped using the Global Screening Array (GSA; Illumina, Inc.) with shared custom content. Pre-imputation quality check were performed using the GenomeStudio and PLINK (Chang et al., 2015) software. Samples with low call-rate (<0.95), with genetically determined non-European ancestry, genetically related to other participants, or with abnormal heterozygosity rates, were excluded. Single nucleotide polymorphisms (SNP) with low frequencies (minor allele frequency (MAF) < 0.01), or with violation of the Hardy-Weinberg equilibrium (HWE, p < 5x10 -6 ), were excluded. Sample imputation to the HRC data (McCarthy et al., 2016) were performed using Minimac3 (Das et al., 2016) with pre-phased haplotypes estimated by SHAPEIT2 (Delaneau et al., 2011). After imputation, SNPs with MAF < 0.01, imputation R 2 < 0.6, or HWE p < 10 -6 , were excluded from subsequent analysis.
UK biobank population (UKBB). Imputed genotypes for the 487,409 participants were obtained from UKBB (Category 100314). The UKBB Axiom Array from Affymetrix (about 90% of participants) and the UK BiLEVE Axiom Array (5% participants) were used for genotyping. Samples with autosome missing call-rate > 0.02, or mismatched genetically inferred and self-reported sex, were removed. Genotype imputation was performed by the UKBB center based on haplotypes from the HRC (McCarthy et al., 2016) and UK10K (Walter et al., 2015) data (further details see (Sugden et al., 2019)).
ABCD population. Imputed genotypes from the ABCD release 3.0 were obtained (https://abcdstudy.org/). Briefly, the Affymetrix NIDA SmokeScreen Array was used for genotyping, which resulted in 733,293 SNPs. After quality control of these genotyped SNPs, the TOPMED reference data was used for imputation. Detailed information can be obtained from https://nda.nih.gov/study.html?id = 901. In the present study, we filtered out rare variants (MAF < 0.05), and variants showing strong deviation from Hardy-Weinberg equilibrium (p < 10 − 6 ).

APOE genotype determination
APOE genotypes in the four datasets were determined by the two SNPs, rs429358 and rs7412. Four haplotypes (https://www.snpedia. com/index.php/APOE), TCCT, CTCT, TCTC and CTTC, of the two SNPs could not be resolved to a unique APOE haplotype, and such cases were excluded from subsequent analysis. Our present study focused on the effect of the APOE ε4 allele on CRP levels and HippV. APOE ε4 was coded as binary, i.e. carrier vs. non-carrier (low proportions of ε4 homozygotes were observed in all four samples; see Table S1 for ε4 frequencies).

Population structure
To correct for subtle population structure effects in statistical analyses, the first 10 principal components (PC) were computed using PLINK (Chang et al., 2015) for the LCBC, UB-BBHI and ABCD sample. Imputed genotypes were first quality checked by removing SNPs that have MAF < 0.05, HWE p < 10 -6 , missing-rate > 0.05. Then, correlated SNPs were removed by the PLINK command -indep-pairwise 100 50 0.1. The top 10 PCs were obtained by the -pca command from PLINK. This procedure was independently performed for the LCBC, UB-BBHI and ABCD population. PCs for the UKBB population were computed by the UK biobank team.

Polygenic scores for CRP levels (PGS-CRP)
Effect sizes of genome-wide significant SNPs for CRP levels reported by Dehghan et al (Dehghan et al., 2011) and Ligthart et al (MacBean et al., 2020) were extracted from their published papers. Among 60 reported SNPs, 56 also existed in our sample. The SNP rs4420638 close to the APOE gene showed the largest effect on CRP levels and thus, was excluded in PGS computation. PGS for all participants were computed using the weighted sums methods implemented by the -score function from PLINK.

LCBC, COGNORM and UB-BBHI sample
Circulating levels of CRP were measured by the AM-438 Quantification of CRP in Dried Blood Spots (DBS) using MSD Mesoscale ECL platform. One 3.1 mm punch from dried human whole blood samples (DBS) was eluted in 60 µL kit diluent and left standing at 4 • C overnight. After bringing the eluate to room temperature, 10 µL of the eluate were diluted in 500 µL kit diluent and analyses were performed on a MESO® QuickPlex SQ 120 Multiplex Imager using the V-PLEX Human CRP kit (K151STD-2; MSD, Rockville, Maryland, USA) as described in the manual. The lower limit of detection (LLD) was 0.08 mg/L. Samples with levels below this value were set to half of this value, 0.04 mg/L. Among the 533 samples used to discover the APOE-CRP relation, 42 had a CRP level below LLD. The number of samples with a below LLD CRP levels analyzed for the CRP-HippV relation is 83.
UK biobank population (UKBB) The serum CRP level for all participants were measured by the UK Biobank biomarker panel (http://www.ukbiobank.ac.uk/uk-biobank-biomarker-panel/). Levels of CRP were determined by the Beckman Coulter AU5800 platform (Beckman Coulter (UK), Ltd) using immune-turbidimetry. The manufacture's detection limits were 0.08-80 mg/L. No sample had a level < 0.08 mg/L.

LCBC, COGNORM and UB-BBHI population
Participants from the LCBC and COGNORM sample were scanned at a total of four Siemens scanners at two sites; 1) Oslo University Hospital; 2) Curato, currently Aleris, Oslo: A 1.5 T Avanto equipped with a 12 channels head coil (site 1 and 2), a 3 T Skyra equipped with a 24 channels Siemens head coil (Site 1) or a 3 T Prisma equipped with a 32 channels head coil (site 1) (all Siemens Medical Systems, Erlangen, Germany). The pulse sequence used for morphometric analyses were one to two 3D sagittal T1-weighted MPRAGE sequences (Supplementary Information). All scans were reviewed for quality and automatically corrected for spatial distortion due to gradient nonlinearity  and B1 field inhomogeneity (Sled et al., 1998). Images were first automatically processed cross-sectionally for each time point with the FreeSurfer software package (version 6.0; http://surfer.nmr.mgh.harvard.edu/). The segmentation procedure automatically labels each voxel as one of 40 structures (Fischl et al., 2002), using a probabilistic brain atlas specific for the current image acquisition protocol (Han et al., 2006). Detailed description for image processing is presented in Supplementary Information. Brain scans for the UB-BBHI participants were acquired using a 3 Tesla Siemens PRISMA scanner with a 32-channel head coil and detailed protocols have been reported previously (Cattaneo et al., 2018).

UK Biobank population (UKBB).
MRIs were collected and processed by the UK Biobank (https://www.ukbiobank.ac.uk) (Alfaro-Almagro et al., 2018). Imaging data were collected using 3.0 T Siemens Skyra scaners (32 channels head coil). Anatomical T1-weighted magnetization-prepared rapid gradient echo (MPRAGE) images were obtained in the sagittal plane at 1 mm isotropic resolution, and T2 weighted FLAIR images were acquired at 1.05x1x1 mm resolution in the sagittal plane. Images were processed by the UK Biobank using the FreeSurfer 6.0 software package.

PET-Aβ status
A total of 126 participants (44.4-80.8 years of age) from the LCBC sample underwent a 18F-flutemetamol-PET scan, sensitive to Aβ accumulation (Wolk et al., 2011). Images were acquired on a General Electric Discovery PET/CT 690 scanner at Aleris Hospital and Radiology, Oslo, Norway. The PET-amyloid beta status, i.e. positive or negative, was estimated by applying principal component analysis on the standardized uptake value ratios (SUVR) from a set of 68 FreeSurfer-derived cortical regions (Mormino et al., 2014). The cut-off between groups was determined using Gaussian mixture modeling (R package mclust v5.2) for the first two principal components. We fitted 18 models, ranging from 1 to 9 mixtures, allowing for either equal or unequal variance, and selected the model with the lowest Bayesian information criterion value. As previously reported in healthy older participants (Mormino et al., 2014), the optimal model consisted of a two-distribution model with unequal variance. Participants with a > 0.5 probability of belonging to the high Aβ distribution were classified as Aβ positive, and the remaining as Aβ negative. Detailed description for PET image processing is presented in Supplementary Information.

CSF AD biomarkers for the COGNORM sample
CSF specimens were collected in polypropylene tubes at the onset of anesthesia prior to administration of the anesthetic agent. The specimens were centrifuged, aliquoted and stored at − 80 • C (Idland et al., 2017). CSF AD biomarkers, Aβ-42 and Aβ40, were measured using the Meso Scale Discovery Aβ Triplex Assay (Meso Scale Discovery, Rockville, Maryland), and the ratio of Aβ-42 to Aβ40 was computed. Levels of total tau (CSF-t-tau) and phosphorylated 181 total (CSF-p-tau) levels were determined using INNOTEST enzyme-linked immunosorbent assay (ELISA; Fujirebio) at the Sahlgrenska University Hospital, Sweden. CSF levels of neurofilament light protein (NFL) were determined by using a commercial ELISA (UmanDiagnostics, Umea, Sweden). Identical protocols for measuring cytokine levels as in LCBC were applied to the COGNORM sample. No data for PET-Aβ was available for this sample.

Effects of APOE ε4 and CRP levels on HippV
To take advantage of the longitudinal brain scans available for subsets of participants, generalized additive mixed models (GAMM; R package 'mgcv') were used to model the nonlinear relation between HippV and age, taking into account intraindividual correlations. APOE ε4 status, estimated intracranial volume (ICV), sex, and the top 10 PCs were coded as fixed effects, and participant unique identifiers were set as the random effect. In addition, we allowed different smooth age functions for the two APOE ε4 groups. Standardized HippV (mean zero and standard deviation one) was set as the dependent variable. To estimate the effect of CRP levels on HippV, instead of APOE ε4 status, the log10transformed CRP levels were set as the main predictor. Moreover, both CRP levels and APOE ε4 status and the interaction term between the two were used as main predictors to test whether the two variables contribute to HippV variation independently. All statistics reported are on the log10 scale of CRP levels. Bonferroni correction was applied to account for multiple testing.

Effects of APOE ε4 and CRP levels on PET-Aβ status and CSF marker levels
Logistic regression analyses were performed to estimate the effect of APOE ε4 and CRP levels on PET-Aβ status. PET-Aβ status was used as dependent variable; APOE ε4 status and CRP levels were set as the main predictor, separately. Age at measurement, sex, BMI, and the top 10 PCs were used as covariates. Linear regression models were then used to estimate the effect of APOE ε4 and CRP levels on CSF marker levels. Each marker was analyzed separately. In each model, the CSF marker level was inverse-normal transformed to have zero mean and one standard deviation and was used as dependent variables. APOE ε4 and CRP levels were used as the main predictor, separately. Age at measurement, sex, BMI were used as covariates. Here, False Discovery Rate (FDR) was used for multiple testing correction (Benjamini and Hochberg, 1995).

Replication analysis
Corresponding models for the relation between APOE ε4 and HippV were applied in the UKBB and ABCD replication samples. For the UB-BBHI sample, which had no longitudinal brain scans, corresponding GAM models were used.

Results
The LCBC sample had a wider age-range than the other samples (Fig. 1B), but had fewer participants aged 35 to 65 years. The UKBB sample had higher CRP levels than those of the LCBC and UB-BBHI samples (Fig. S1, two-sided t test, p < 2x10 -16 ). ABCD participants had smaller HippV than those in the other three samples, and the UB-BBHI participants had the largest HippV among the four (pairwise two-sided t tests, Bonferroni-corrected p < 0.05; Fig. S2). The LCBC sample has a higher frequency of APOE ε4 than the other three samples; but, the ε4 frequencies in all four samples are consistent with a recent report of geographical frequency differences (Table S1) (Kern et al., 2015).

Stronger genetic control over lower CRP levels
To investigate if the genetic contributions to CRP levels vary with blood CRP levels, we stratified CRP levels into 5 bins and regressed log10 transformed CRP levels in each bin on APOE ε4 status, including age and sex as covariates in the models. The false discovery rate procedure was used for multiple testing correction (along test with the full CRP range in each dataset, 18 tests performed). We found that whereas ε4 was associated with CRP levels overall (bin: All, FDR-corrected p < 0.05), only in the bin with lowest CRP level ranges (<1mg/L) was ε4 nominally associated with lower CRP level in the LCBC and the UB-BBHI samples (FDR-corrected p = 0.09 and 0.02, respectively; Fig. 2A and Table S2). To test whether this was due to too few participants having high CRP levels in the two healthy samples, we performed the same analysis for the UKBB dataset (N = 304,322). The results clearly showed that the higher the CRP levels, the less significant was the association between APOE ε4 and CRP levels( Fig. 2A-B).
We next regressed the log10 transformed CRP levels in each bin on the PGS-CRP (constructed without APOE-region SNPs). The same covariates as used above for APOE ε4 were also included in these models. PGS-CRP showed a weaker effect on CRP levels than for APOE ε4 (Fig. 2B, Table S3), and PGS-CRP values were only positively associated with overall CRP levels (bin: [0-100] mg/L) in both the LCBC and UB-BBHI sample (FDR-corrected p < 0.05, Table S3). Whereas, due to large samples in each bin, the UKBB showed significant association for the three bins with lower CRP levels (FDR-corrected p < 0.05, Fig. 2B).
Thus, the same trend as observed for ε4 was also apparent for PGS-CRPthe higher the CRP levels, the less their variations were attributable to genetic variation.

APOE ε4 is associated with lower CRP levels across lifespan
Across the whole lifespan, we observed a strong and consistent negative association between APOE ε4 and CRP levels (Fig. 2C). Specifically, in the LCBC samplewhere 54% of participants were below 40 years of age -APOE ε4 carriers were found to have lower CRP levels on average across the lifespan (N = 533; beta = -0.15, p = 7.94x10 -4 ) as compared to non-carriers. When the analysis was restricted to those < 40 years of age, similar effect sizes were obtained but not statistically significant (N = 289; beta = -0.10, p = 0.12), possibly due to reduced sample size. In the first replication dataset (UB-BBHI, N = 339), we also found a strong negative effect of ε4 on CRP levels that existed across the whole age-range (41-67 years; beta = -0.21, p = 2.22x10 -3 ). The concave curve in Fig. 2C for UB-BBHI may be due to too few individuals having the ε4 allele (N = 69). In the very large UKBB dataset, the effect size of ε4 became smaller but more significant (Fig. 2C; N = 304,322; APOE ε4 carrier vs non-carrier: beta = -0.13, p < 2x10 -16 ). Although we always modelled the age function by regression splines, in this large UKBB sample, CRP levels were almost linearly increasing with age (approximated p for age < 2x10 -16 ). The positive associations between CRP levels and age were also significant in the LCBC and UB-BBHI sample (p < 0.05) (Fig. 2C).

High CRP levels associate with small hippocampal volumes
CRP levels showed stronger association with smaller HippV than did APOE ε4 status (Fig. 4). In the LCBC sample (N = 507, N-MRI scans = 1,344, 20-81 years of age), increased CRP levels associated with small hippocampi (beta = -0.17, p = 1.21x10 -2 ). Including APOE ε4 status as an additional covariate did not qualitatively change this association. In UKBB, higher CRP levels were also associated with smaller HippV (beta into six bins based on measured CRP levels(mg/L): <1, <1 mg/L; [1-3), equal to or higher than 1 mg/L and lower than 3 mg/L; [3-5), equal to or higher than 3 mg/L and lower than 5 mg/L; >=7, equal to or>7 mg/L. Colors indicate if uncorrected p < 0.05; C) Association between APOE e4 and CRP levels across age range and independent samples. Fitted CRP levels were obtained by fitting generalized additive model (GAM), which model age by regression splines, and plot against age at measurement. Colors indicate e4 status. Names of datasets and their sizes are shown on top of each panel. 1.69x10 -2 ). Moreover, in this big dataset, we did not find significant interactions between CRP levels and APOE status on HippV (p = 0.9). Possibly due to the small sample size for the UB-BBHI sample (N = 339), we did not observe significant associations between CRP levels and HippV (p = 0.95). We also investigated CRP by age interactions using the largest UKBB sample (Supplementary Information). We found that individuals with high CRP levels had a faster hippocampal atrophy (p = 1.77x10 -4 , Fig. S13).

PGS-CRP was not associated with AD markers
Because the APOE gene has been associated with CRP levels in several previous studies, we tested if genetically predicted CRP levels (excluding APOE-region SNPs) could predict AD biomarkers in the three samples. There were no associations between PGS-CRP and any AD biomarkers studied. In addition, there were no associations between PGS-CRP and APOE ε4 status (Fig. S3). Whereas these results may suggest the effect of CRP on AD biomarkers was largely driven by APOE ε4, it should be noted that both CRP levels and APOE e4 were significantly associated with HippV in the multivariate model.

Discussion
By using four complementary populations comprising an age-span of 9-90 years (total N = 315 753) we showed that APOE ε4 was continuously and consistently associated lower circulating CRP levels. This association is mainly for CRP levels in the range of normal physiological conditions (CRP < 3 mg/L). Our findings extend previous knowledge by indicating that the association exists before the typical age at which accepted AD biomarkers can be observed (~40 years) (Selkoe and Hardy, 2016) and continues beyond the typical age of onset for AD (~60 s). Moreover, we demonstrated both APOE ε4 and higher CRP levels are associated with smaller hippocampal volumes, a marker that has frequently been linked to low cognitive performance and high risk of AD. Whereas APOE ε4 was clearly associated with PET-Aβ and CSF markers for AD, CRP did not show convincing associations with any of

CRPcut
High Low   Fig. 4. Associations between APOE ε4 and hippocampal volume. Fitted total hippocampal volumes by GAMM (LCBC sample (Disc_LCBC), UK Biobank subsample (Rep2_UKBB)) and by GAM (Rep1_UB-BBHI) were plotted against baseline age (x axis). For demonstration purpose, hippocampal volumes were stratified by high CRP level (above the sample median) and low CRP level (below the sample median). Log10-transformed CRP level were input as continual variable in model fitting including sex, and age at measurement (smoothing variable in GAM(M), x axis), BMI and the top 10 principal components as covariates. For GAMM, random intercepts were modelled by taking subject identifiers as random term. Sample sizes were shown on the top of each panel.

Fig. 5.
Associations of APOE ε4 with CSF markers. Distributions of CSF Ab42 to Ab40 ratios (A) and level of CSF Ab42 (B) were stratified into APOE e4 carriers and non-carriers (x axis). In total 59 individuals from the COGNORM sample had both CSF marker and APOE data and were analyzed.
these AD biomarkers. Our study is to our knowledge the largest examining the CRP-APOE ε4 association, and covers the entire adult lifespan as well as relations to other well-established biomarkers for AD also in early adult life. By taking advantage of big data, the results suggest genetic variants contribute to the variations of CRP levels only within the normal physiological range (Dumitrescu et al., 2020). In the relatively smaller CRP samples (i.e. LCBC and UB-BBHI), <30% participants had CRP levels above 1 mg/L (Table S2 and S3), possibly suggesting the null association between genetic markers (both APOE ε4 and PGS-CRP) and higher CRP ranges could be due to low statistical power. However, also in the large UKBB subsample -where>20 000 participants had CRP > 7 mg/L and 12 123 had CRP levels > 10 mg/Lthe results clearly showed that the effect of genetic variants on CRP levels decreased with increased CRP level. This means the inherited genetic variations, here APOE ε4 status and PGS, do not influence CRP's ability to respond to acute infections or other stimuli, where high CRP levels (>10 mg/L) are expected (Dumitrescu et al., 2020). Instead, chronic and low-grade inflammation as indexed by CRP is under strong genetic control. Martiskainen et al. (2018) have recently reported a similar inverse APOE ε4-CRP relation in a large population (N ~ 6,100; 50-75 years of age). Their study showed an association between high CRP levels and low plasma Aβ42, but no association between APOE ε4 and plasma Aβ42 levels. In a small postmortem AD sample with neurofibrillary pathology (N = 71), the authors reported null associations between CRP expression levels and APOE ε4 in brain tissue. Here, we confirmed the APOE ε4-CRP relation in three independent samples, with a sample size five times larger and a broader age-range. In our LCBC sample, we also observed significant associations between APOE ε4 and PET-Aβ, CSF Aβ42 as well as the ratio of CSF-Aβ42 to Aβ40. It should also be noted that the origins of blood or plasma Aβ arise not only in brain but also the liver, kidney, skeletal muscle, skin, and other organs (Roher et al., 2009;Selkoe et al., 1988), the relations between blood Aβ, CSF-Aβ and PET-Aβ have been inconclusive (Le Bastard et al., 2009;Schindler et al., 2019;Toledo et al., 2011), and thus that differences in physiological measurements likely in part explains some inconsistencies between reports. We observed that whereas high blood CRP level was nominally associated with a low CSF-Aβ42 to Aβ40 ratio in our sample, this association was completely driven by APOE ε4 status. Importantly, our sample included participants as young as 20 years where there are hardly any detectable neuropathological antecedents to AD in the brain. Thus, the interpretation by Martiskainen et al. that the relation between APOE ε4 and CRP may be driven by an anti-inflammatory effect of plasma Aβ, is questionable. Notably, CRP did not show relations to Aβ brain biomarkers in our sample, suggesting another, possibly independent pathway.
Several other explanations for the APOE ε4 and CRP relation have been proposed. Because APOE ε4 and high circulating CRP levels both have been linked to increased risk of cardiovascular diseases, März et al. have suggested that the reduced CRP levels in APOE ε4 carriers might be caused by reduced activity of the hepatic mevalonate pathway, i.e. reduced cholesterol synthesis (März et al., 2004). This hypothesis was supported by the clinical observation that inhibiting mevalonate pathway by statin treatment promoted lower CRP levels (Balk et al., 2003). However, a later report from a large Icelandic cohort implicated that the effect of APOE ε4 on lowering CRP levels may be independent of statin-related pathways (Eiriksdottir et al., 2006). Moreover, in a population of nonagenarians, Rontu et al. interpreted the effect of APOE ε4 on extended longevity due to reduced inflammatory response as indicated by low CRP levels (Rontu et al., 2006). At the other end of life, Berrahmoune et al. showed that APOE ε4 was associated with low CRP levels among children (10-20 years of age). Whereas the association between APOE ε4 and low circulating CRP levels has been reported before, most previous studies have been based on clinical case-control samples with a relatively narrower age range, and the interpretations of the associations have been related to the mechanism of the focal diseases in the studies.
Our study fills a gap based on four cognitively healthy populations, with an age span of 9 to 90 years. In the LCBC sample, the levels of both inflammatory (CRP, IL8) and anti-inflammatory (IL1RA) markers monotonically increase with age (Fig. S14). This observation cannot simply be explained by the 'inflamm-aging' hypothesis (Franceschi et al., 2000), i.e., increased chronic inflammation with age. Instead, both inflammatory and anti-inflammatory parameters are constantly changing with age across the lifespan (Riveros-Mckay Aguilera et al., 2020). Whereas CRP levels were highly correlated with other cytokine levels ( Fig. S6), no associations were observed between APOE ε4 and these other cytokines, which is in accordance with previous findings (Martiskainen et al., 2018;März et al., 2004;Rontu et al., 2006). Thus, the interpretation of the APOE ε4 and CRP association in terms of immune response remains elusive. However, some possible directions for further research may be suggested: Since the inverse correlation between APOE ε 4 and blood CRP levels exists before any pathological AD biomarker is observed and exists only in the low CRP level range, one should consider what could be the effects of a lifetime of lowered inflammatory response in terms of increased AD risk. Based on our findings, a novel hypothesis is that APOE ε 4 can confer risk by being associated with a lower inflammatory response to daily exposures, possibly leading to greater accumulation of low-grade inflammatory stress though the lifespan. Since the currently measured CRP did not show convincing associations with any of the classical AD biomarkers, nor the currently investigated cytokines, further research should aim to delineate a possible pathway.
Although we found that being an APOE ε4 carriers and having higher CRP levels were associated with small hippocampal volumes across the lifespan in the three populations, these effects were relatively weaker than that of APOE ε4 on the CRP levels. On the one hand, the association between APOE ε4 and hippocampal volume may suggest an offset effect, i.e., APOE ε4 carriers start with smaller hippocampi than non-carriers, as shown in our previous work (Walhovd et al., 2020). On the other hand, low CRP levels were associated with large hippocampal volumes. Still, whether there is also an offset effect of CRP levels on hippocampal volume is unclear. The effects of APOE ε4 (carriers vs. non-carriers) and CRP levels PET-amyloid beta load in the brain and CSF marker levels were evaluated by general linear models, including age at measurement, BMI, sex and the top 10 genetic principal components as covariates (APOE related analysis only). Logistic regression models were used for PET-Aβ positiveness; and linear regression models were used for other markers, which have been normalized before modelling. Sample size (N), t-score (regression coefficient divided by standard error), and p (FDR corrected p in parentheses) are also shown. The following abbreviations were used: CRP, C-reactive protein; PET-Aβ, total amyloid beta load in the brain; CSF-Aβ40, amyloid beta with 40 amino acids in cerebral spinal fluid; CSF-Aβ42, amyloid beta with 42 amino acids in cerebral spinal fluid; CSF-Aβ42/Aβ40, ratio of CSF-Aβ42 to CSF-Aβ40 in cerebral spinal fluid; CSF-t-tau, level of the total tau proteins in cerebral spinal fluid; CSF-p-tau, level of the phosphorylated tau proteins in cerebral spinal fluid; CSF-NFL, level of the neurofilament light proteins in cerebral spinal fluid.
We replicated our findings in three independent datasets with varying age ranges, sizes, and data acquisition platforms, indicating the results are robust. We also accounted for the effect of BMI by including it as a covariate in the models. A limitation is that, due to the high cost of measuring AD biomarkers, only a small number of participants had PET-Aβ and CSF markers in our sample. That we did not find associations between APOE ε4 and CSF-Aβ40, CSF-NFL and CSF tau levels might therefore be due to low statistical power, and it remains to be seen whether future larger scale studies may identify such associations. In addition, for a subset of participants (COGNORM study), cytokine levels were measured three years later than the brain scan. However, such a complication does not affect our conclusions.

Conclusion and perspectives
By synthesizing four cognitively healthy populations with wide age spans, we observed that APOE ε4 was associated with low whole blood/ serum CRP levels across the lifespan. While a relationship between acute inflammation and dementia risk has been established (Komaroff, 2020), we think in a lifespan perspective, further investigations of the role of very low-grade inflammatory responses may also be worthwhile. Based on our findings, it may be fruitful in further research to focus on whether APOE ε 4 can confer risk by being associated with a lower inflammatory response to daily exposures, possibly leading to greater accumulation of low-grade inflammatory stress though the lifespan. Our study shows previous interpretations of these associations are incomplete, and future mechanistic studies are needed to resolve this intriguing relation.

Author contributions
YW, AMF and KBW conceived and designed the study. JMR, MP, FM, AP, JMT, MC, DB, JMY, IKA, AMF, KBW, LOW and AVI collected sample and performed MRI and cytokine measurements. YW performed statistical analysis. All authors contributed to the interpretation of the results. YW and KBW wrote the draft. All authors edited and approved the manuscript.

Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Norway, funded the first years of the COGNORM study. Part of the research was conducted using the UK Biobank resource under application number 32048. We are grateful to all participants for their time and commitment. The funding sources had no role in the study design. The ABCD data were obtained from the Adolescent Brain Cognitive Development (ABCD) Study (https://abcdstudy.org), held in the NIMH Data Archive (NDA). ABCD is a multisite, longitudinal study designed to recruit more than 10,000 children aged 9-10 years into early adulthood. The ABCD Study is supported by the National Institutes of Health and additional federal partners under award numbers U01DA041022, U01DA041028, U01DA041048, U01DA041089, U01DA041106, U01DA041117, U01DA041120, U01DA041134, U01DA041148, U01DA041156, U01DA041174, U24DA041123, U24DA041147, U01DA041093, and U01DA041025. A full list of supporters is available at https://abcdstudy.org/federal-partners.html. A listing of participating sites and a complete listing of the study investigators can be found at https://abcdstudy.org/scientists/workgroups/. ABCD consortium investigators designed and implemented the study and/or provided data but did not necessarily participate in analysis or writing of this report. This manuscript reflects the views of the authors and may not reflect the opinions or views of the NIH or ABCD consortium investigators. Part of the computations in this work were performed on resources provided by UNINETT Sigma2-the National Infrastructure for High Performance Computing and Data Storage in Norwaywith project no. (nn9769k/ns9769k).