Circulating C-reactive protein influences polygenic risk of inflammatory genes expressed in brain endothelia for Alzheimer’s disease

Abstract

Background:

C-reactive protein (CRP) is a key marker of systemic inflammation that affects blood vessel endothelial function, including in the brain. Since endothelial dysfunction is linked to Alzheimer’s disease (AD), we investigated whether elevated CRP level interacts with genetic pathways in brain endothelial cells to influence AD risk.

Methods:

Using AD genome-wide association study (GWAS) data, we developed multiple polygenic risk scores (PRSs) including single nucleotide polymorphisms (SNPs) in genes expressed in brain endothelial cells, excluding the APOE region, that are involved in inflammation, synaptic transmission, and other pathways.

Results:

Analysis across three independent cohorts revealed that individuals with low inflammatory PRSs (<50%) and elevated blood CRP level were associated with an increased risk of AD; in contrast, those with high inflammatory PRSs (≥50%) did not exhibit this CRP-related AD risk increase. Further examination of individuals with a low inflammatory PRS showed that elevated CRP was associated with lower cerebrospinal fluid (CSF) Aβ42 level and temporal lobe atrophy. Among individuals with a high inflammatory PRS, elevated CRP level was negatively correlated with CSF pTau181 and brain tauopathy, suggesting a potential protective mechanism against tau pathology. Key inflammatory PRS genes, which were impacted by circulating CRP for AD, included APP, IL6ST, and FN1, are involved in amyloid pathology, wound healing, and coagulation.

Conclusion:

Our findings highlight two distinct genetic-dose dependent backgrounds: “vulnerable” (<50% inflammatory PRS) and “resilient” (≥50% inflammatory PRS), and support a Genome-Internal Environment (G × IE) interaction model, linking peripheral inflammation to AD risk.

1. Introduction

Circulating blood creates an internal environment (IE) that influences various tissues and cells, particularly endothelial cells lining blood vessels. Elevated C-reactive protein (CRP) is a marker of inflammation severity in the IE. Previous research showed that peripheral CRP affects Alzheimer’s disease (AD) risk in APOE ε4 carriers. CRP promotes tau pathology in APOE ε4 knock-in mice, but not in APOE ε2 or APOE ε3 knock-in mice (Zhang et al., 2021). It is also found that AD risk is influenced by the interaction of CRP and monocyte chemoattractant protein-1 (MCP-1) with variants in APOE, SPI1, CD33, HLA-DRB1, and NAV3, likely through effects on brain endothelial cells (Tao et al., 2018; Huang et al., 2022; Huang et al., 2023; Huang et al., 2025). Since immune and endothelial cells share inflammatory gene expression patterns and pathways (Raivich et al., 1999), both brain and peripheral cell types may contribute to AD pathology (Zhang et al., 2023). Based on these findings, we hypothesized that the influence of variants in inflammatory genes expressed in brain endothelial cells on AD-related processes might be modified by CRP levels in blood.

Genome-wide association studies (GWAS) have identified over 75 AD risk loci (Bellenguez et al., 2022), with several linked to inflammatory pathways (Hikami et al., 2011; Walker et al., 2015; Foster et al., 2019; Puigdellivol et al., 2020; Pimenova et al., 2021; Torvell et al., 2021; Wissfeld et al., 2021; Patrick et al., 2022). A polygenic risk score (PRS), which represents the combined effects of multiple genetic variants on a trait and serves as a continuous variable for disease risk, has become a common tool in AD genetic research (Clark et al., 2022). Leonenko et al. demonstrated that the best approach for predicting AD uses APOE ε4 carrier status and a PRS that excludes the APOE region as separate predictors (Leonenko et al., 2021). Recent advances have also developed pathway-specific and cell-type-specific PRSs (Kumar et al., 2022; Yang et al., 2022; Harrison et al., 2023; Yang et al., 2023) to better understand genetic influences on brain structure and disease processes like AD. Gene-environment interactions (G × E) have been studied extensively with the genetic component represented by single variants, gene-based variant sets, or PRS (Marderstein et al., 2021; Boye et al., 2024; Herrera-Luis et al., 2024). However, relatively few studies (Marderstein et al., 2021) have focused on the impact of biological and physiological conditions (i.e., the internal environment (IE)) such as circulating inflammatory status/factors, and their interaction with genetic factors.

Systemic inflammation induces reactive and proinflammatory microglia and astrocytic phenotypes that promote AD pathology (Walker et al., 2019). Mapping the genes and identifying their modifying factors are important for AD prevention and intervention (Herrera-Luis et al., 2024). In this study, we applied a hypothesis-driven strategy that relies on functional pathways to prioritize gene and variant sets (i.e., PRS), and thus enhances power for detecting G × E interactions. Accordingly, we constructed pathway-specific AD PRSs to investigate the association of the interaction between specific aggregated genetic effects and peripheral CRP level (i.e., G x IE signals) with AD. Various PRS methods have been developed (Purcell et al., 2007; Choi and O’Reilly, 2019; Prive et al., 2021), and one popular tool is the PRS-CS (Ge et al., 2019), which infers posterior effect sizes of single nucleotide polymorphisms (SNPs) using the GWAS summary statistics and a linkage disequilibrium (LD) reference panel. We generated PRSs for genes involved in (Zhang et al., 2021) inflammation, (Tao et al., 2018) synaptic transmission, and (Huang et al., 2022) other pathways using Hallmark and Gene Ontology (GO) gene sets, brain endothelial cell expression, and AD GWAS summary statistics (Bellenguez et al., 2022). Analyses of these pathway PRSs and blood CRP levels were conducted using data from the Framingham Heart Study (FHS), UK Biobank (UKB), and Alzheimer’s Disease Neuroimaging Initiative (ADNI) (Fig. 1).

Fig. 1.

Flowchart for the study

The flowchart and details of the endothelial AD-related PRSs generation, statistical analyses, and in-depth analyses are shown.

*doi:https://doi.org/10.1038/s41593-023-01334-3

**doi:https://doi.org/10.1038/s41588-022-01024-z

2. Methods

2.1. Participants

Discovery analyses were conducted using data from the Framingham Heart Study (FHS) which is a single-site, multigeneration, community-based, prospective cohort study in Framingham, Massachusetts (Tao et al., 2018). Surveillance for incident AD/dementia was initiated in 1975, and diagnoses of AD and other causes of dementia were made through a consensus panel. (Kannel et al., 1979) We excluded those with baseline and incident ‘other causes’ of dementia. Scores for executive function, language, and memory domains in FHS were derived based on neuropsychological tests as previously described (Mukherjee et al., 2020; Mukherjee et al., 2023; Scollard et al., 2023). This study focused on the Offspring (Generation 2) cohort participants who had genome-wide genotyping data and serum CRP measurement data derived from blood samples obtained at Exam 7 (baseline) and who were evaluated for cognitive decline and AD/dementia until 2019. Prevalent AD cases that were diagnosed before Exam 7 (baseline) were excluded. Characteristics of the sample included in this study are shown in Table 1. Informed consent was obtained from all study participants, and the study protocol was approved by the Institutional Review Board (IRB) of Boston University Medical Campus.

Table 1.

Basic characteristics and different pathway polygenetic risk scores (PRS) between cognitive controls and incident AD in the FHS cohort.

Characteristics		All subjects	Cognitive controls	Incident AD	P-valued
N subjects, No (%).		3,069	2,883 (93.94)	186 (6.06)	-
Age when measuring CRP levels, mean (SD)		60.81 (9.38)	60.13 (9.14)	71.30 (6.44)	<0.001 a
Follow-up-to-AD, years, median (Q1-Q3)		18.41 (14.49-19.38)	18.50 (15.83-19.43)	9.68 (6.33-13.59)	<0.001 b
Male, No. (%)		1,399 (45.58)	1,333 (46.24)	66 (35.48)	0.005 c
Years of education, mean (SD)		13.95 (2.42)	13.99 (2.41)	13.37 (2.59)	0.002 a
APOE ε4e, No. (%)		613 (19.97)	554 (19.22)	59 (31.72)	<0.001 c
C-reactive protein (mg/L), median (Q1-Q3)		2.15 (1.01-5.13)	2.14 (1.00-5.10)	2.20 (1.15-5.62)	<0.001 b
PRS-50KB	Inflammatory PRS (nSNPs = 2,054), mean (SD)	0.19 (0.92)	0.19 (0.92)	0.22 (0.91)	0.69a
	Synaptic transmission PRS (nSNPs = 1,555), mean (SD)	−9.00 (2.24)	−9.00 (2.23)	−8.49 (2.25)	0.003 a
	Other PRS (nSNPs = 78,057), mean (SD)	−0.48 (0.13)	−0.48 (0.13)	−0.45 (0.14)	<0.001 a

3,069 FHS participants were divided into two groups based on those who had not (Cognitive controls) versus those who developed AD (Incident AD) until 2024. Demographics, APOE ε4 carriers and CRP levels were illustrated and compared between the two groups. 3 pathways PRS, 1) inflammatory, 2) synaptic transmission, and 3) others, surround 50KB gene region extension were generated and compared between the two groups. Means (SD) with T-tests for normal distributions and medians (Q1-Q3) with Wilcox rank sum tests for skewed distributions were used to analyze continuous variables, while n (%) with the χ² tests were used for categorical variables for the AD-control comparisons. P values indicating statistical significance are shown.

Abbreviations: AD: Alzheimer’s disease; CRP: C-reactive protein; APOE: apolipoprotein E; PRS: polygenic risk score.

^a.P value from t-tests.

^b.P value from Wilcox rank sum tests.

^c.χ² test p value.

^d.P value for the comparison between the two groups.

^e.APOE ε4 = ε34 + ε44

^f.PRSs were scaled by 10,000.

Replication and validation analyses were conducted using data from the UK Biobank (UKB) and the Alzheimer’s Disease Neuroimaging Initiative (ADNI). The UKB collected genetic, biomarker, medical record and self-reported demographic and clinical information from more than 500,000 persons living in the UK (Allen et al., 2012). Included in this study were 361,005 self-reported white participants after excluding subjects who did not have baseline CRP measurement, lacked genetic information, had prevalent AD/dementia at baseline, or were nonwhite (self-reported). Controls who reported cognitive impairment or had a family history of AD or dementia in one or both parents (i.e., “proxy” AD cases (Marioni et al., 2018; Jansen et al., 2019)) were also excluded. The AD status for each individual was determined on the basis of medical or hospital inpatient records. Individuals who had ICD-10 codes of F00 (AD dementia) or G30 (AD) were classified as AD. AD diagnoses were recorded until March 30, 2017, which was used as the end date for survival analysis of incident AD.

ADNI, a longitudinal multicenter study launched in 2003, collects neuroimaging, AD biomarker in CSF and blood, and clinical and neuropsychological test data which can be used for developing accurate models to predict the diagnosis and progression of AD (Petersen et al., 2010). Participants underwent longitudinal in-depth neuropsychological evaluations (Aisen et al., 2010), and consensus diagnoses of cognitive normal (CN), mild cognitive impairment (MCI), and AD were assigned based on established research diagnostic criteria (Huang et al., 2020). After filtering out self-reported non-European ancestry subjects and those without CRP and genotype information, 444 ADNI-1 cohort participants were included in the analysis. MCI to AD conversion was determined by comparing the baseline Clinical Dementia Rating (CDR=0.5) with the most recent CDR score. MCI subjects whose most recent CDR scores were ≥ 1.0 were classified as ‘converters’.

2.2. Genotyping and genotype imputation

SNP genotype data for the FHS cohort that were previously filtered and imputed were obtained from the Trans-Omics for Precision Medicine (TOPMed) Imputation Server (https://imputation.biodatacatalyst.nhlbi.nih.gov/#!). Genotype calling and imputation in the UKB dataset were performed as previously described. (Bycroft et al., 2018) We imputed genotypes for ADNI participants using the TOPMed reference panel. The imputation quality (r²) of all SNPs was > 0.4. After checking, all the imputed SNPs we used in constructing PRSs had high imputation quality (r2>0.8).

2.3. Blood CRP measurements

CRP levels (Exam 7) in FHS participants were measured using enzyme-linked immunosorbent assay (ELISA) with a Dade Behring BN100 nephelometer (Wilson et al., 2005) from fasting blood samples that were collected at Exam 7 from the antecubital vein (details have been previously described) (McDermott et al., 2005). In UKB, CRP (high sensitivity, hs-CRP) was measured in the period of 2006-2010 and a second time in 2012-2013 for UKB subjects by immunoturbidimetric-high sensitivity analysis on a Beckman Coulter AU5800. Measurements obtained during the first period were used as the baseline level for this study. Details have been previously described (http://biobank.ctsu.ox.ac.uk/crystal/label.cgi?id=17518; http://biobank.ctsu.ox.ac.uk/crystal/field.cgi?id=30710; https://biobank.ctsu.ox.ac.uk/showcase/showcase/docs/biomarker_issues.pdf). Plasma CRP was measured in ADNI participants using the Human Discovery MAP Panel and measurement platform. (Ray et al., 2007; Trojanowski et al., 2010) CRP level was log-transformed or evaluated according to cutoffs of 3 mg/L or 8 mg/L to define low-grade inflammation in subsequent analyses.

CRP was measured in each cohort using standard clinical assays. Although absolute values can vary across assay platforms, all methods are validated for clinical use and demonstrate strong concordance in ranking individuals by inflammatory status. To account for potential assay-related heterogeneity, CRP levels were harmonized by applying the established clinical threshold of ≥3 mg/L to define elevated CRP. Primary analyses used the harmonized CRP values, and sensitivity analyses stratified by cohort and assay type were performed to confirm robustness.

2.4. Plasma AD biomarkers

Plasma Aβ42 and phosphorylated Tau181 (pTau181) levels obtained for the FHS Offspring cohort participants at exam 7 were log-transformed to conform to a normal distribution. Details of blood sample collection, processing, and quality control for measurements of Aβ42 and pTau181 in the FHS have been previously reported. (Romero et al., 2020) (McGrath et al., 2022),

2.5. Brain imaging

A subset of the FHS Offspring cohort participants underwent brain magnetic resonance imaging (MRI) scanning between 3/1999 and 12/2017, as previously described¹². Specifically, participants were imaged using a 1.5 T MRI (Siemens Medical, Erlangen, Germany) with a 3-dimensional T1-weighted coronal spoiled gradient-recalled echo sequence. All images were transferred to and processed by the University of California Davis Medical Center without knowledge of the clinical information. Segmentation and quantification of the total cerebral cranial volume (TCV), frontal lobe (FBV), parietal lobe (PBV), temporal lobe (TBV), and hippocampal (HPV) were performed using semi-automated procedures as previously described. (DeCarli et al., 2005) TCV was determined using a convolutional neural network method (Fletcher et al., 2021). Nonlinear co-registration of images to the Desikan-Killiany-Tourville atlas enabled the calculation of regional gray matter volumes (Aljabar et al., 2009). MRI measures were corrected for head size by calculating the percent of these volumes relative to the TCV. The normalized brain volume variables were log-transformed and their normality was confirmed. Each image set underwent rigorous quality control, including assessments of the original acquisition and image processing quality.

2.6. Cerebral spinal fluid AD biomarkers

Aβ42, total tau (t-Tau) and p-Tau181 levels in cerebrospinal fluid (CSF) were measured in ADNI participants using the multiplex xMAP Luminex platform (Luminex, Austin, TX, USA) with INNOBIA AlzBio3 (Innogenetics, Ghent, Belgium) immunoassay kit-based reagents (Olsson et al., 2005; Shaw et al., 2009). Further details of ADNI methods for CSF acquisition and CSF measurement can be found at https://adni.loni.usc.edu/methods/ . The most recent biomarker measurements for each individual were used for analyses.

2.7. AD-related neuropathological traits

Eleven AD-related proteins were measured by antibody-specific immunostaining in brain tissue obtained from 79 FHS donors including the amyloid species Aβ40 and Aβ42; tau-related species p-Tau231, p-Tau202, p-Tau396, p-Tau181, t-Tau and AT8 expression; Iba1 density, microglial activation marker CD68, and PSD-95 (Cherry et al., 2020). These variables were normalized using the rank-based inverse normal transformation. Other AD-related neuropathological traits were assessed, including Braak stage for neurofibrillary degeneration, CERAD score for the density of neocortical neuritic plaques, CERAD semi-quantitative score for diffuse plaques, microinfarcts, arteriolosclerosis, atherosclerosis, cerebral amyloid angiopathy (CAA), macrophage infiltration and hippocampal sclerosis. Neuropathological evaluations were performed by neuropathologists blinded to all demographic and clinical information at the BU-ADRC (Au et al., 2012).

2.8. Pathway AD PRS generation

To create pathway-specific AD PRSs, we first selected pathway genes from Hallmark & Gene Ontology (GO) gene sets (Liberzon et al., 2015) (Fig. 1). Inflammatory-related genes were identified from Hallmark gene sets using the terms COMPLEMENT, ANGIOGENESIS, INFLAMMATORY RESPONSE, IL6-JAK-STAT3 SIGNALING and TNFA SIGNALING VIA NFKB. Synaptic transmission-related genes were identified from GO terms with keyword ‘synaptic transmission’. In the primary analysis, these pathway gene sets were pruned to include those expressed in ≥ 5% of brain vasculature endothelial cells (i.e., count > 0). Endothelial-cell-expressed genes were identified using publicly available single-nucleus RNA-seq data from human brain vasculature (Sun et al., 2023). Pathway gene counts and the gene selections at each step are presented in Supplementary material 2.

Pathway SNPs were selected using summary statistics from a large AD GWAS based on GRCh38 (Bellenguez et al., 2022). SNPs assigned to the APOE gene (i.e., within±20/50/100/200KB based on our strategy in Fig. 1), without LD information in 1000G (phase 3 EUR) LD reference panel or with minor allele frequency (MAF) < 5% were excluded. Pathway SNPs were chosen based on criteria (Fig. 1), including AD GWAS P-value and region surrounding the pathway gene coding region. To avoid a high overlap of pathway SNPs between inflammatory and synaptic transmission pathways, the overlap rates across different P-value thresholds and gene region extensions were calculated and plotted (Figure S1). Based on this analysis, a significance threshold of P < 0.05 and gene region extension of 50 KB (other extension sizes were evaluated as sensitivity analysis) were used for selecting pathway SNPs and generating PRS for main analyses since they had acceptable overlap rates between pathways of <25%. The remaining SNPs that satisfied the requirements were assigned to ‘other pathways’.

We estimated the posterior effect size of each SNP using PRS-CS (Ge et al., 2019), a powerful Bayesian method that incorporates linkage disequilibrium (LD) patterns across the genome using a Bayesian continuous shrinkage prior that shrinks small/noisy effects toward zero more gently than hard thresholds, which helps to estimate SNP effect sizes more accurately. Output from PRS-CS are posterior SNP effect sizes adjusted for LD and shrinkage, which were used to compute individual PRS in test samples (Fig. 1). All parameters were set to default. Pathway AD PRSs for each individual in the FHS, UKB and ADNI-1 cohorts were calculated by PLINK2 ‘score’ based on the posterior effect size of specific pathway SNPs, as previously suggested (Ge et al., 2019).

Sensitivity analyses for constructing the PRS were conducted in the FHS cohort by (Zhang et al., 2021) using all inflammatory genes regardless of endothelial expression (add noise, more genes/SNPs included); (Tao et al., 2018) setting GWAS P-value thresholds to 1 (i.e., add noise) or 0.005 & 5e-5 & 5e-8 (i.e., reduce noise but may also reduce the true signal); (Huang et al., 2022) a variant-based strategy that assign each variant to their nearest gene (within a fixed window size). These analyses did not change our conclusions (Table S1–7, except for smaller P-value cutoffs 5e-5 and 5e-8 that resulted in much fewer variants (and thus much smaller signals and less significant)), which suggested that the main signals in PRSs were driven by SNPs with high weight (i.e., posterior effect size) and would not be masked by including more or less ‘noise’ SNPs. Additionally, other gene region extensions of 20/100/200 KB were also evaluated for less or more SNPs compared to ±50KB of gene region (Table S1), and our conclusions still held. We applied a more stringent LD r² threshold of 0.01 (i.e., excluded SNPs with LD r² > 0.01) for the APOE variants (rs7412 and rs429358) as a sensitivity analysis. Since APOE gene were in the synaptic transmission pathway but not in the others, we conducted the analysis for this pathway and placed the results in Table S6. No significant difference was observed.

2.9. Statistical analysis

Statistical analyses were performed using the R statistical environment (R 4.2.1). Group differences were assessed using t-tests for normally distributed continuous variables, the Wilcoxon rank sum test for continuous variables with skewed distributions, and the chi-square test of independence for categorical variables. Statistical significance thresholds were determined by Bonferroni correction to correct for multiple testing.

The association of AD with pathway PRSs in each dataset was evaluated using logistic regression models including covariates for age, sex, years of education, APOE ε4 carrier status, and the first 5 principal components (PCs) of population structures. AD incidence and MCI to AD conversion were assessed using Cox proportional hazards models which included a pathway AD PRS as a continuous variable, blood CRP level (either log-transformed as a continuous variable or a dichotomous variable based on < vs. ≥ 3 mg/L), a term for the interaction between the pathway AD PRS and CRP level, and the same covariates. We used Cox proportional hazards models, adjusting for the same covariates as above, to evaluate how CRP levels were associated with AD incidence across groups of individuals stratified by inflammatory PRS, increasing in 10% intervals from below the 10th percentile to above the 90th percentile. Results from each dataset were combined by meta-analysis using the inverse-variance weighted method with the R package ‘meta’.

The association of the interactions between blood CRP and pathway AD PRSs with MRI traits, as well as with log-transformed plasma AD biomarkers, was evaluated in the FHS dataset using linear regression models including the same covariates described above. MRI brain volumes were divided by the total cerebrum cranial volume and multiplied by 100% to adjust for brain size.

Since FHS did not measure the CSF AD biomarkers, we used the ADNI-1 dataset with the concentrations of CSF Aβ42 and ptau181. Linear regression models were used to study the relationship between plasma CRP and the CSF AD biomarkers with adjustments for age at baseline, sex, years of education, and APOE ε4 within different inflammatory PRS stratification groups. The biomarker variables were log-transformed for normal distribution, and the last measurement was used.

We further examined brain pathologies in relation to inflammatory AD PRS. Due to the limited number of autopsied samples FHS brain bank, subjects were divided into six groups based on inflammatory PRS percentiles: <30%, <40%, <50%, >50%, >60%, and >70%. For continuous outcome variables including brain Aβ40, Aβ42, p-Tau231, p-Tau202, p-Tau396, p-Tau181, t-Tau, AT8, Iba1 density, CD68, and PSD-95, we applied a rank-based inverse-normal transformation. Linear regression models were then used to assess the association between plasma CRP levels and these neuropathological brain measures within each PRS group, adjusting for baseline age, sex, years of education, and APOE ε4 status. For ordinal neuropathological outcomes including Braak stage, CERAD scores (neuritic and diffuse plaques), microinfarcts, arteriolosclerosis, atherosclerosis, cerebral amyloid angiopathy (CAA), macrophage infiltration, and hippocampal sclerosis, we used ordinal regression models (logistic regression models for binary outcomes); covariates were not included in these models due to the small sample sizes within each PRS group.

2.10. Bioinformatic analysis methods

We performed additional analyses involving genes containing SNPs that were used for generating the inflammatory-related AD PRS (Fig. 1), an approach referred to as variant-level interaction analysis. After adjusting for age, sex, years of education, and APOE ε4 status, we examined the interaction between each inflammatory SNP and blood CRP levels in relation to AD incidence using Cox proportional hazards models; in relation to AD-related continuous traits using linear regression models. Results were visualized through bubble plots. The gene set overrepresentation analysis for the selected inflammatory genes (n=206 for 50Kb extended window) was conducted using the R package ‘clusterProfiler’ (Yu et al., 2012) with GO terms. The top 50 enrichment results were collected (Supplementary material 3). The differential expression (DE) patterns of the selected inflammatory genes between AD and controls were examined for endothelial cells in single-nucleus RNA-seq data from human brain vasculature with MAST hurdle model (Finak et al., 2015), adjusted for the number of expressed genes, age, sex, PMI, ethnic group, batch, brain region and other dementia-related pathology (Lewy body dementia, Parkinson’s disease and vascular contributions to cognitive impairment and dementia (VCID)), as previously showed (Sun et al., 2023) (Supplementary material 4). Next, from the top 20 enriched Gene Ontology (GO) terms, we selected those that included at least one gene with a SNP showing a significant interaction with blood CRP (p < 0.01). These terms and their associated genes were visualized using a heatmap and network graph. We also summarized the expression patterns of the top genes based on both average expression and the percentage of cells expressing each gene across five brain cell types (endothelial, ependymal, fibroblast, pericyte, and smooth muscle cells) in a bubble plot.

The STRING protein-protein interaction analysis for three selected key interactive genes (APP, FN1 and IL6ST) was performed using the web tool https://string-db.org/. The minimum required interaction score was set to ‘highest confidence (0.90)’ with other parameters set to default.

3. Results

3.1. Characteristics and comparisons of the FHS participants without and with AD development

The 3,069 participants from FHS generation 2 (Gen 2) Exam 7 had an average follow-up period of 18.4 years (Table 1) and were used as the discovery cohort. 186 (6.06%) of the participants developed AD during the follow-up. Subjects were divided into two groups: those who did not develop AD (cognitive controls) and those who developed AD (incident AD). As expected, participants who developed AD were older (P<0.001), had fewer years of education (P=0.002), and had higher blood CRP levels (P<0.001). More female participants (P=0.005) and APOE ε4 carriers (P<0.001) were found in those who developed AD.

3.2. Characterizations of PRSs and the influence of circulating CRP on them for AD risk

We first analyzed each polygenic risk score (PRS) set, using a 50KB gene extension, for their association with AD risk using logistic regression models and t-tests. Individuals who developed AD had higher PRS for synaptic transmission (P=0.003) and other pathways (P<0.001) compared to cognitively normal controls. In contrast, inflammatory PRSs showed no significant difference between groups (P=0.69) (Table 1). After adjusting for age, sex, education, APOE ε4 status, and the first five principal components (PCs), only the other PRSs remained significantly associated with AD risk across all gene region extensions (20KB, 50KB, 100KB, 200KB) (P<0.004) (Table S8).

As the inflammatory PRSs did not contain CRP, we examined the interaction between blood CRP levels and each PRS set in relation to AD risk using Cox regression models adjusted for covariates in the FHS cohort. In contrast to PRS sets alone, the inflammatory PRSs, but not synaptic transmission and other PRSs, showed significant interaction with blood CRP levels—both as a continuous log-transformed measure (P=0.004) and as a binary variable using a 3 mg/L cutoff (P=0.02), for AD risk (Table 2). Similar results were observed with gene region extensions of 20KB and 100KB (Table S8).

Table 2.

Interactive effects of different pathway PRSs and blood CRP levels on incident AD.

	Incident ADb FHS		Incident ADb UKB		Incident AD from MCIb ADNI-1		Meta-analysis: FHS + UKB + ADNI-1
Pathway PRSa	HR (95% CI)	P value	HR (95% CI)	P value	HR (95% CI)	P value	Z-score	P value	Directions
Log (CRP)
Inflammatory	0.80 (0.68-0.93)	0.004	0.92 (0.85-0.98)	0.02	0.88 (0.72-1.08)	0.21	−3.59	3.3×10−4	− − −
Synaptic transmission	0.99 (0.94-1.05)	0.80	0.98 (0.95-1.00)	0.07	0.99 (0.91-1.09)	0.91	−1.73	0.08	− − −
Others	0.77 (0.29-2.07)	0.60	0.87 (0.54-1.42)	0.59	0.43 (0.10-1.81)	0.25	−1.03	0.30	− − −
CRP (≥ 3 vs. < 3 mg/L)
Inflammatory	0.68 (0.49-0.94)	0.02	0.79 (0.66-0.94)	0.008	0.49 (0.28-0.84)	0.01	−4.01	6.0×10−5	− − −
Synaptic transmission	0.98 (0.86-1.12)	0.78	1.00 (0.94-1.07)	0.97	0.99 (0.78-1.25)	0.91	−0.18	0.85	− + −
Others	0.93 (0.11-7.68)	0.95	0.37 (0.11-1.23)	0.11	0.28 (0.01-12.50)	0.51	−1.56	0.12	− − −

Three cohorts, Framingham Heart Study (FHS), UK Biobank (UKB) and ADNI-1, were used independently and combined for meta-analysis for AD incidence. Three endothelial AD PRSs (i.e., inflammatory, synaptic transmission, and others pathways) within the gene region extension of 50KB were generated. Blood CRP concentration [logarithm (2A-upper portion)] and 3 mg/L cutoff (2B-lower portion) were applied to the models. Using cox proportional hazards regression (HR) with 95% CI, the interactions between the different PRSs and blood CRP levels for AD risk (FHS and UKB) / MCI-to-AD conversion (ADNI) were studied in each cohort and meta-analysis. P values for interaction terms are shown.

Abbreviations: AD: Alzheimer’s disease; CRP: C-reactive protein; HR: hazard ratio; PRS: polygenic risk score.

^a.PRSs were scaled by 10,000.

^b.Cox proportional hazards regression models for the interaction: incident AD / MCI-to-AD conversion ~ PRSs * blood CRP + PRSs + blood CRP + age at baseline + sex + years of education + APOE ε4 status + top 5 PCs. Raw P values for interaction terms are shown.

We validated the association between different PRSs and AD by using the UKB and ADNI-1 cohorts (Table S9) using logistic regression models and Cox regression models. In UKB (n=361,005), again only inflammatory PRSs, but not synaptic transmission and others PRSs, interacted with blood CRP levels to influence AD risk (P=0.008), although inflammatory PRSs alone was also associated with AD risk (Table 2 and Table S9). With the data on MCI-to-AD conversion in the ADNI-1, we found that the inflammatory PRSs interacted with the blood CRP cutoff 3 mg/L to influence MCI-to-AD conversion (P=0.01) (Table 2).

We further conducted the meta-analysis using the interaction results from the combined FHS, UKB and ADNI-1 cohorts. Again, significant interactions between the inflammatory PRSs and blood CRP levels for increased AD risk were observed, i.e., for continuous log CRP level (P=3.3×10⁻⁴), and for the binary CRP with cutoff 3 mg/L (P=6.0×10⁻⁵) (Table 2). No interactions of the synaptic transmission or others PRSs with blood CRP were found for AD risk.

3.3. Stratification of the inflammatory PRSs and circulating CRP for AD risk and cognitive domains

We divided the inflammatory PRSs into percentiles, increasing by 10% increments from <10% to >90%, and examined how blood CRP levels related to AD risk using Cox regression models (Fig. 2). In the FHS cohort, we observed a genetic dose-dependent pattern. Among individuals with lower inflammatory PRS (<50%), whom we labeled the “vulnerable PRS” group, higher blood CRP levels were associated with increased AD risk (Fig. 2A). In contrast, for those with higher inflammatory PRS (≥50%), labeled the “resilient PRS” group, this association disappeared, suggesting a reduced impact of CRP on AD risk in this group. These findings were generally replicated in the UKB cohort (Fig. 2B). Kaplan–Meier analysis further supported this pattern. In the FHS cohort, high blood CRP levels (>3 mg/L) significantly increased AD risk only in the vulnerable PRS group (≤50%), but not in the resilient PRS group (>50%) (Fig. 2C). Similar results were seen with the 20KB gene extension inflammatory PRS (Figure S2).

Fig. 2.

The influence of blood CRP levels on various percentile cutoffs of endothelial inflammatory PRS for AD incidence

The interactive effect between genetic vulnerability and resilience (pathway PRSs) in the context of peripheral inflammation (blood CRP levels) was examined for AD risk. To investigate this, Cox proportional hazard regression models (A and B) and Kaplan-Meier Survival Curves (C) were employed to analyze the joint effects of endothelial inflammatory PRS and blood CRP levels on AD risk in the FHS (A and C) and UKB (B) cohorts. In A and B, participants were stratified into groups based on endothelial inflammatory PRS percentiles (increasing by 10% increments from <10% to >90%). The CRP cutoff of 3 mg/L was used in the analysis. Cox proportional hazards regression models were used to examine the association between blood CRP levels and incident AD in each PRS group, adjusting for age, sex, education, APOE ε4, and the top 5 principal components. The results are presented as hazard ratios (HRs) and p-values. Figure 2C shows Kaplan-Meier survival curves with p-values, illustrating the time to AD onset based on the PRS (<50% and ≥50%) and the CRP cutoffs (3 mg/L). In the ADNI-1 cohort, the outcome was conversion from MCI to AD. A meta-analysis combining data from FHS, UKB, and ADNI-1 was performed using Cox proportional hazard regression models (D). Participants were grouped by endothelial inflammatory PRS percentiles (increasing by 10% increments from <10% to >90%), and the interaction between PRS and blood CRP levels was examined for AD risk. The results are presented as meta beta estimates and p-values in bar plots.

Significant associations are indicated by *=P<0.05, **=P<0.01, and ***=P<0.005.

We then performed a stratified meta-analysis combining FHS, UKB, and ADNI cohorts (Fig. 2D). The results consistently showed that elevated CRP levels (≥3 mg/L) increased AD risk only in individuals with lower inflammatory PRS (<50%). This CRP-AD relationship was not observed in those with higher PRS (≥50%).

To explore this interaction related to cognitive function, we examined cross-sectional cognitive data in the FHS cohort using linear regression models. Consistent with the AD findings, individuals with low inflammatory PRS (<50%) and high CRP (>3 mg/L) showed worse performance or trends toward worse performance in language (Fig. 3A) and memory (Fig. 3B) domains. Conversely, individuals with high inflammatory PRS (≥50%) and high CRP levels showed neutral or even better cognitive scores in these domains. This suggests that a high inflammatory PRS may offer resilience against CRP-related cognitive decline, while a low inflammatory PRS increases vulnerability to the effects of peripheral inflammation in aging. No significant relationships were found between inflammatory PRS, CRP, and the executive function domain (Figure S3).

Fig. 3.

Cross-sectional associations between blood CRP concentrations and cognitive domains across different endothelial inflammatory PRSs percentiles The study investigated genetic vulnerability versus resilience (low vs high inflammatory PRSs) in relation to cognitive function amid peripheral inflammation (blood CRP levels) using linear regression models. Participants from FHS were categorized into groups based on a 10% increase of the inflammatory PRS: increasing by 10% increments from <10% to >90%. Two genetic patterns were identified: <10% to <50% as vulnerable PRS and >50%, to >90% as resilient PRS. Linear regression models were utilized to examine the associations between blood CRP levels (with a cutoff of 3 mg/L) and cognitive outcomes: (A) language domain score, (B) memory domain score. Adjustments were made for age, sex, years of education, APOE ε4 genotype, and the top 5 principal components. The results are presented as beta estimates and p-values, with significance indicated as *=P<0.05, **=P<0.01, and #=P<0.1.

3.4. Characterizations of the inflammatory PRS and circulating CRP for the brain atrophy and AD biomarkers

To further validate the interaction between inflammatory PRS and blood CRP in AD, we examined their relationships with brain MRI volumes and CSF AD biomarkers using linear regression models. For MRI brain volumes, elevated CRP levels were significantly associated with reduced temporal lobe volume (TBV), again in a genetic dose-dependent pattern—the strongest effect seen in the vulnerable PRS group (Fig. 4A). This CRP-brain volume relationship was not observed for total cerebral volume (TCV), hippocampus (HPV), parietal (PBV), or frontal brain volumes (FBV) (data not shown).

Fig. 4.

Associations between blood CRP concentrations, temporal lobe brain volume and the CSF AD biomarkers across varying percentiles of endothelial inflammatory PRSs.

We examined the genetic vulnerability versus resilience (low vs high inflammatory PRSs) in relation to temporal lobe brain volume and AD biomarkers in the presence of peripheral inflammation (blood CRP) using linear regression models. Participants were divided into groups according to percentiles of the inflammatory PRS: increasing by 10% increments from <10% to >90%. Linear regression models were employed to investigate the relationships between blood CRP levels and the (A) temporal brain volume (TBV) (divided by TCV), after adjusting for age, sex, years of education, APOE ε4 status, and the top 5 principal components.

The ADNI-1 dataset, which included genetic data, blood CRP levels, and the cerebrospinal fluid (CSF) AD biomarkers, was utilized based on different levels of inflammatory PRSs. The participants were divided into groups according to the inflammatory PRS percentiles: increasing by 10% increments from <20% to >80%. Blood CRP concentration was divided by two cutoffs, < vs. ≥ 3 and 8 mg/L as the determining factors. Linear regression models were applied to examine the associations of these CRP cutoffs and the CSF AD biomarkers, CSF Aβ42 (B) and CSF pTau181 (C), as outcomes, in different PRS groups, while adjusting for age, sex, years of education, and APOE ε4 status.

Results are presented as beta estimates and p-values, with significance indicated as *=P<0.05, **=P<0.01, and #=P<0.1.

Next, we analyzed CSF AD biomarkers in the ADNI cohort, adjusting for age, sex, education, and APOE ε4 to validate the findings. Among individuals with low inflammatory PRS (<50%, vulnerable PRS), decreasing PRS percentiles combined with higher CRP levels were associated with lower CSF Aβ42 levels, consistent with greater amyloid burden (Fig. 4B). This relationship was absent in the resilient PRS group (≥50%). For CSF pTau181, the pattern reversed. Among those with high inflammatory PRS (≥50%, resilient PRS), increasing PRS combined with higher CRP was associated with lower CSF pTau181 levels, suggesting less tau pathology (Fig. 4C). No such relationship for CSF pTau was observed in the vulnerable PRS group.

In contrast to the AD biomarkers in central nervous system, higher blood CRP levels were nonlinearly associated with plasma Aβ42, especially in individuals with low inflammatory PRS (<50%, vulnerable PRS); but this association was weak or absent in those with high inflammatory PRS (≥50%, resilient PRS) (Figure S4). No relationship was observed between CRP levels and plasma pTau181 in either genetic group.

3.5. Inflammatory PRS, Blood CRP, and AD neuropathology

We also examined the interaction between inflammatory PRS, blood CRP, and AD-related neuropathology in a small FHS autopsy sample (n=82) using linear regression models and ordinal regression models. Inflammatory PRS was split into low (<50%, <40%, <30% vulnerable PRS) and high (≥50%, ≥60%, ≥70% resilient PRS) groups, and CRP levels were classified using 3 mg/L and 8 mg/L cutoffs. Among individuals with low inflammatory PRS, higher plasma CRP levels (especially ≥8 mg/L) were positively associated with microglial marker CD68 (Fig. 5A) and showed a trend toward higher AT8 tau pathology (Fig. 5B). In contrast, in the high PRS group (resilient PRS), elevated CRP (≥8 mg/L) was negatively associated with AT8 staining (Fig. 5B) and lower Braak stage (Fig. 5C), indicating less tau pathology. Additionally, elevated CRP was negatively associated with cerebral amyloid angiopathy (CAA) in this resilient group (Fig. 5D). Consistent with Fig. 4, we observed a genetic dose-dependent pattern: 1) Individuals with low inflammatory PRS (vulnerable PRS) were more sensitive to high circulating CRP levels, showing increased AD risk and brain neuroinflammation, lower CSF Aβ42, and more TBV atrophy. 2) In contrast, those with high inflammatory PRS (resilient PRS) showed protective effects—high circulating CRP levels were linked to lower CSF pTau181 and low level of tauopathy, and no increased AD risk.

Fig. 5.

Associations between blood CRP levels and brain neuropathologies under different endothelial inflammatory PRS background

Linear regression and ordinal regression models were applied to study the associations of blood CRP levels with the various AD brain neuropathologies in different PRS groups. The data of FHS participants who donated their brains and had blood CRP measurements when they were alive were used. Due to a small sample size (n<80), participants were divided into two groups based on the endothelial inflammatory PRS median values: increasing by 10% increments from <30% to 70%, and blood CRP levels were divided into the cutoffs 3 and 8 mg/L. Immunostaining levels of CD68 (A) and AT8 (B); the neuropathology, Braak stage (0, 1, 2, 3, 4, 5, 6) (C) and cerebral amyloid angiopathy (0, 1, 2, 3) (D) were used as the outcomes. CD68 and AT8 were normalized using the Rank-Based Inverse Normal Transformation. Beta estimates and p values are illustrated in the bar plots. *=P<0.05, **=P<0.01 and #=P<0.1.

3.6. In-depth characterization of the variants and genes in the inflammatory PRS influenced by circulating CRP for AD risk

We aimed to identify specific endothelial-expressed AD risk genes that may interact with peripheral CRP levels to influence AD risk. Within the inflammatory PRS (206 genes, Fig. 1), we first performed variant-level interaction analysis in the FHS cohort using Cox regression models. Thirteen key gene loci—ABCA1, ACVRL1/ACVR1B, APP, COL4A2, FN1, HRH1, IL1R1, IL6ST, LIRA, MCL1, SLC4A4, TNIP1 and ZBTB10 —showed significant interaction with elevated CRP levels (≥3 mg/L) in predicting AD incidence (p < 0.01) (Table S10, Figure S5).

Next, we examined how these genes relate to cognitive function, AD pathology, and biomarkers using linear regression models and ordinal regression models (Fig. 6A). Most of these genes not only interacted with CRP to influence AD risk but were also linked to amyloid biomarkers. A subset was additionally associated with tau pathology. We then conducted Gene Ontology (GO) pathway enrichment analysis on the 206 inflammatory PRS genes. The top 18 enriched pathways included processes like wound healing, cytokine production, proteolysis regulation, coagulation, and tyrosine phosphorylation (Fig. 6B). Nine of the 13 key interactive genes TNIP1, MCL1, IL1R1, IL6ST, FN1, ABCA1, ACAVR1B, ACVRL1 and APP, were involved in these pathways. Gene expression analysis showed that MCL1, ACVRL1, and APP were upregulated in brain endothelial cells of AD patients, while the other seven genes were downregulated, with only TNIP1 reaching statistical significance (Fig. 6B). Notably, APP was expressed in over 60% of brain endothelial cells, and IL6ST and FN1 were expressed in over 40%, though both were downregulated in AD (Fig. 6C). STRING analysis revealed that these endothelial-expressed genes are connected to AD-related biological pathways (Fig. 6D): 1) APP interacts with PSEN1 and CLU (Aβ production); 2) FN1 interacts with STAT3 (inflammation) and vascular genes TGFB1, ICAM1, ITGA5; and 3) IL6ST interacts with STAT1, STAT3, and JAK2 (inflammation).

Fig. 6.

In-depth analysis of the individual genes of the endothelial inflammatory PRS and circulating CRP to AD risk.

We investigated the key genes of the SNPs in the inflammatory PRS impacted by elevated blood CRP for AD risk. Using Cox proportional hazard regression models, top gene variants with interaction of the CRP cutoff (3 mg/L) for AD and AD-related traits and the P-values for AD < 0.01 are shown in bubble plots (A). The 206 inflammatory genes in the generated inflammatory PRSs were used as input for gene overrepresentation analysis for GO terms. 18 significant GO terms among the top 20 (B) were selected and shown based on the criteria that at least one interactive gene in (A) was involved. Using the endothelial cell single-nucleus transcriptome data from human brain vasculature, the differential expressions of the inflammatory genes between AD and control were studied with MAST hurdle model, adjusted for the number of expressed genes, age, sex, PMI, ethnic group, batch, brain region and other dementia-related pathology. The top interactive genes that involved in the selected GO pathways were plotted in heatmap, where the selected genes belonged to the corresponding GO pathways were colored in blue or red based on the log fold change (logFC). TNIP1 on the left part of the dash line indicates differentially expressed between AD and control endothelia with statistical significance. The expression levels (i.e., mean expression and proportion of cells expressed) of the genes in A were studied across five cell types (endothelial, ependymal, fibroblast, pericyte and smooth muscle cells) (C). Three key interactive genes (APP, FN1 and IL6ST) expressed in brain endothelia and their interactions with other related inflammatory genes in STRING analysis were highlighted (D). Since EDN1 is an endothelial specific gene, 4 of the GO pathways that included EDN1 gene were selected for network graph visualizations (E). It shows the connections between the selected four GO terms and the genes involved in each pathway. The labels of the significant DEGs were highlighted in yellow (up-regulated) or blue (down-regulated) in the AD enodothelia. The size of the gene ‘circle’ represents significance levels in the DE analysis. The gene ‘circle’ color represents the logFC levels in the DE analysis (orange to blue=high to low). The green circle highlighted the interactive genes in (A). The red rectangle highlighted the endothelial-specific gene EDN1.

We also examined EDN1, an endothelial-specific gene downregulated in AD brain. EDN1 participates in four key pathways—cytokine signaling, response to external stimuli, wound healing, and coagulation—which overlap with those involving APP, IL6ST, and FN1 (Fig. 6E). Further, several inflammatory genes (CX3CL1, JAK2, CLU, IFNGR1, PDGFB, STAT1, STAT2) were downregulated in AD brain endothelia, particularly within cytokine signaling pathways, while STAT3 was upregulated. Angiogenesis-related genes (PLAUR, TGFB1, NOTCH4, CDKN1A) were also significantly upregulated in the wound healing pathway. In summary, our analysis suggests that brain endothelial inflammation and angiogenesis are key mechanisms through which peripheral inflammation (CRP elevation) may drive or worsen AD pathogenesis.

4. Discussion

Growing evidence suggests that brain endothelial dysfunction plays a crucial role in AD progression (Fang et al., 2023; Yue et al., 2024), with inflammation in endothelial cells acting as a key driver of disease pathology (Kinney et al., 2018). However, the interplay between genetic predisposition, peripheral inflammation, and AD risk remains poorly understood. While all the risk genes for late-onset AD are not fully penetrated, the interaction between AD genetic vulnerabilities and peripheral circulating inflammatory proteins for AD risk has been less studied than the risk genes alone. We found two distinct genetic profiles (Table 2, Fig. 2): 1) Vulnerable group (inflammatory PRS <50%)-High CRP levels significantly increased AD risk in a genetic dose-dependent pattern. 2) Resilient PRS group (inflammatory PRS ≥50%)-CRP levels did not impact AD risk. This distinction may be biologically and clinically significant, as it suggests that individuals with a low endothelial inflammatory genetic profile are more susceptible to AD when systemic inflammation is high, whereas those with a higher genetic inflammatory load may be more resistant. Previous studies showed that the APOE ε4, but not APOE ε2 or APOE ε3, allele is influenced by circulating CRP to increase AD risk (Tao et al., 2018; Tao et al., 2021). Although the mechanism is not clear, this study extends this by showing a broader genome-internal environment (G × IE) interaction in the blood-brain axis for AD risk.

In response to infections and injuries, circulating inflammatory proteins like CRP release into blood and generate an IE that significantly impacts various cells, particularly endothelial cells lining blood vessels. Our study provides evidence for genome (G)-by-IE (G × IE) interactions, showing that peripheral inflammatory factors can modulate genetic effects on AD risk in a genetically dose-dependent manner. While some studies have linked elevated CRP levels to increased AD risk, others have not found a clear association (Sundelof et al., 2009; O’Bryant et al., 2010; Song et al., 2015). Our findings and others suggest that this inconsistency may stem from genetic differences related to inflammation and AD, which determine an individual’s vulnerability or resilience via brain endothelia responding to IE inflammation (Grammas, 2011). Specifically, the interaction between genetic factors, such as APOE or inflammatory PRSs identified in this study, and CRP may influence AD risk by altering gene regulation in certain cell types like brain endothelia in response to inflammation. One possible mechanism for the G × IE interaction is that CRP or other proinflammatory proteins directly or indirectly impact individual genotypes or combinations of genetic variants, as shown with the inflammatory PRS in this study. For instance, CRP-induced transcription factors (TFs) may bind differently to TF binding sites depending on genotype, leading to variations in gene expression associated with AD risk. Alternatively, genetic variants and circulating inflammatory factors might drive different epigenetic modifications or produce distinct gene products, which could interact its downstream pathways in brain endothelia for AD susceptibility. Future research could investigate these interactions at the functional genomic and proteomic levels to better understand how peripheral inflammation and genetic factors converge to influence AD risk.

The inflammatory PRS used in this study was derived specifically from genes expressed in brain endothelial cells, rather than from peripheral endothelial/inflammatory pathways more broadly. This tissue-specific design likely contributes to the stronger and more consistent PRS×CRP associations observed for CNS-derived biomarkers (CSF Aβ42, CSF pTau181, temporal lobe atrophy, and neuropathological markers) (Fig. 4 and 5), which directly reflect processes occurring at the neurovascular interface. In contrast, plasma AD biomarkers integrate a mixture of central and peripheral signals and are influenced by hepatic metabolism, renal clearance, and peripheral inflammatory states that are not captured by a brain-endothelium–focused PRS (Figure S4). Thus, the attenuated or absent PRS×CRP effects in plasma biomarkers may reflect this biological specificity: a genetic score anchored in brain endothelial transcriptional programs is more likely to modulate CNS pathology than peripheral biomarker levels. This interpretation aligns with our overall model of CRP-modulated endothelial vulnerability underlying AD risk.

The vulnerable endothelial inflammatory profile (PRSs<50%) showed genetically dose-dependent responses to peripheral inflammatory factor, CRP, for AD risk (Fig. 2). Notably, the top interactive genotypes were primarily associated with amyloid pathology in the brain (Figs. 3–4, 6A). Two of the top genes we discovered in the Inflammatory PRS of brain endothelia, APP (Selkoe and Hardy, 2016) and FN1 (Pedrero-Prieto et al., 2019; Bhattarai et al., 2024) (Fig. 6C–D) are involved in Aβ generation/aggregation. Other amyloid-related AD risk genes in the pathways of inflammatory PRSs included TNIP1, CLU and PSEN1 (Figs. 6B–E). These genes are directly involved in APP processing, leading to increased Aβ production and amyloid pathology (Selkoe and Hardy, 2016; Reitz, 2013; Panyard et al., 2024), highlight a potential link between peripheral inflammatory responses at the brain endothelia and Aβ pathology in the AD brain. Other top pathways are wound healing and angiogenesis, including several angiogenesis-related genes, IL6ST, FN1, PLAUR, TGFB1, NOTCH4 and CDKN1A (Fig. 6E). Since Aβ damages cerebrovasculature, angiogenesis and vascular repair may play some role in the interplay between peripheral inflammation, brain endothelial inflammation, and brain AD pathology (Jefferies et al., 2013).

Another genetic profile we identified was the resilient inflammatory PRSs (≥50%) which exhibited resilience against AD risk in the presence of peripheral inflammation (Fig. 2). These resilient inflammatory PRSs carriers showed negative associations with tauopathy in CSF and in the brain (Figs. 3–4). Basic research studies demonstrate that Aβ induces tau phosphorylation and aggregation (Lewis et al., 2001; Gamblin et al., 2003), and tauopathy, rather than Aβ, is the later event and the primary driver of cognitive decline in AD (Mungas et al., 2014). It has been shown that tau can enter brain vascular endothelial cells, promoting cellular senescence, while tauopathy in AD mouse models leads to the accumulation of phosphorylated tau (pTau) in endothelial cells (Hussong et al., 2023). Human brain studies have further confirmed the presence of tau and pTau species, including pTau181, pTau217, and pTau231, within blood vessels in AD brains (Hoglund et al., 2024; Zhang et al., 2024). Thus, we hypothesized that specific genotypes associated with high inflammatory PRSs (≥50%) may confer resilience to peripheral inflammation in AD risk by not only reducing Aβ production/aggregation but also attenuating tau phosphorylation and accelerating the proteolysis of pTau.

This study has several limitations. CRP was measured using different clinical assays across cohorts, which may affect absolute values, but all assays are validated and consistent results across cohorts support robustness. PRS were generated using a single tool (PRS-CS), and future analyses with alternative methods may provide additional insights. The gene-centric strategy for generating pathway PRSs can miss some functional variants far away from the gene body and would be one limitation. However, we believe it is still acceptable since the closer variants to main gene body (e.g., <50KB) contributed to most of the gene function and regulation, while variants far away from the gene body has trivial effect on that gene (such as those >1MB distance) (Zhang et al., 2014). Gene and variant selection were based on endothelial expression, but many of these genes are also expressed in immune cells, so contributions from other cell types such as microglia cannot be excluded. The sample size for neuropathological analyses was relatively small, and participants were limited to Caucasian descent, highlighting the need for validation in larger and multiethnic cohorts. Finally, while our findings suggest potential mechanisms linking CRP, endothelial pathways, and AD risk, they remain observational, and preclinical studies are needed to establish causality.

Nevertheless, one hallmark of aging is the increased susceptibility of elderly individuals to infections, injuries, and chronic peripheral conditions like cardiovascular diseases, diabetes and metabolic syndrome, which contribute to a chronically pro-inflammatory IE in circulating blood. Our study provides evidence that G × IE interaction plays a critical role in maintaining healthy brain endothelial function, which may contribute to cognitive resilience during aging. Despite that genetic background is a key element in this process, antagonizing some inflammatory factors for certain genetic carriers could be a practical approach to modify the genetic vulnerabilities for AD prevention. Future drug development and clinical trials using anti-inflammatory medications for AD should account for genetic background and IE factors, such as CRP levels, to adopt a personalized medicine approach.

Appendix A. Supplementary data

Supplementary data to this article can be found online at https://doi.org/10.1016/j.nbd.2026.107257.

Data availability

The FHS data are available at dbGaP (https://www.ncbi.nlm.nih.gov/projects/gap/cgi-bin/study.cgi?study_id=phs000007.v33.p14) upon reasonable request/application. The UKB data are available at https://www.ukbiobank.ac.uk/ upon reasonable request/application. The ADNI data are available at https://adni.loni.usc.edu/ upon reasonable request/application.