FHS is a single-site, multi-generation, community-based, prospective cohort study of health in Framingham, Massachusetts. The current study included 1,566 participants (Mean_age = 65.93 ± 8.91; 46.36% female) from the Offspring cohort (Gen 2) with data on genome wide single nucleotide polymorphisms (SNPs) and circulating CD34+CD133+ cells measured at examination 8 (2005–2007) who were followed subsequently for incident AD dementia. Relevant details about this cohort have been previously described [15]. Fifty-nine of them developed incident AD through the period ending in 2019 (Figure S1). Informed consent was obtained from all participants and the study protocol was approved by the Institutional Review Board of Boston University.

We used the existing data on circulating EPCs and EMCs in the FHS cohort, and the data has been published in several papers [16–18]. Briefly, fasting blood samples were collected from participants in order to count circulating progenitor cells and were stored at 2–8°C [19] and peripheral blood mononuclear cells (PBMCs) were used for cell phenotyping [17]. Then, PBMCs were isolated via Ficoll (General Electric) gradient centrifugation, counted, incubated with FcR blocking antibodies (Miltenyi Biotec), labeled with anti-human KDR PE antibody (R&D Systems), anti-human CD34 FITC antibody (BD Biosciences), anti-human CD133 APC antibody (BD Biosciences). Samples were washed and fixed with 2% paraformaldehyde. Cells were analyzed on BD FACSCalibur cytometer running CellQuest software (BD Biosciences). Data were analyzed using FlowJo 8 (FlowJo LLC) and positive populations were identified by comparisons of stained samples with samples incubated with isotype-matching IgG antibodies conjugated to FITC and PE. Compensation was performed using single-stain controls. Percentages of CD34+, CD34+CD133+, CD34+CD133–, CD34–CD133+, CD34–CD133–, CD34+, and CD34+KDR+ (VEGFR2) cells were quantified in the data. Similarly, anti-human CD31 PE (BD Biosciences) and anti-human CD45 PerCP antibody (BD Biosciences) were used to analyze CD31+, CD31–, CD31+CD45–, CD31+DIM, and CD31+lymphoid cells. We used the data of both the proportions of each cell type among the total number of PBMCs (%) and the concentrations of each cell type (cell number per mL) shown in Table 1.

The laboratory researchers who conducted the measurements of different circulating cells were not involved in the process of AD dementia diagnoses and were blind for the cognitive status of the participants. All of the Gen 2 participants were followed for incident AD and dementia. The Mini-Mental State Examination (MMSE) was administered beginning at the time of the 5th health exam (i.e., 1991–1995) to monitor changes in cognitive status. A decrease in MMSE performance of 3 or more points from the immediately preceding exam or 5 or more points across all the exams would indicate a change in cognitive status that warranted review by a dementia diagnostic panel consisting of at least one neurologist and one neuropsychologist. Furthermore, from 1999–2005, all the surviving Gen 2 participants were invited for an in-depth cognitive examination, which also screened for incident cognitive impairment that warranted review by the dementia diagnostic panel. Consensus diagnostic procedures according to the criteria of Diagnostic and Statistical Manual of Mental Disorders (DSM-IV) and National Institute of Neurological and Communicative Diseases and Stroke/Alzheimer’s Disease and Related Disorders Association (NINCDS-ADRDA) for dementia and AD were conducted and have been previously described [20]. Incidences of dementia and AD after the CD34+CD133+ measurement at exam 8 were used for the analyses.

We used the existing data on brain volumes from FHS for this study. The brain magnetic resonance imaging (MRI) protocol has been reported in detail elsewhere [21, 22]. A Siemens 1-T MR machine with a T2-weighted double spin-echo coronal imaging sequence was used (Siemens Healthineers, Erlangen, Germany). A central laboratory blinded to demographic and clinical information processed digital information on brain images and quantified the brain data with a custom-written computer program operating on a UNIX, Oracle Solaris platform (Sun Microsystems, Santa Clara, California, USA). The semiautomated segmentation protocol for quantifying gray matter vs. white matter volumes in total cerebral brain volume (TCV), frontal lobar, parietal lobe, occipital lobe, and temporal lobe brain volume as well as hippocampal volume (HPV) has been described elsewhere [23], as has the interrater reliabilities for these methods. TCV was determined using a convolutional neural network method [15]. Non-linear co-registration of images to the Desikan-Killiany-Tourville atlas [16] enabled calculation of regional gray matter volumes [17, 18]. MRI volumes were corrected for head size by calculating the percentage of TCV. Each image set underwent rigorous quality control including assessments of the original acquisition and image processing quality.

To study the genotypes that were potentially impacted by circulating CD34+CD133+ progenitor cells for AD, GWAS were performed using the Framingham analytical pipeline. We conducted an interaction GWAS with log-transformed CD34+CD133+ cell frequency by using a logistic regression model for AD. The relationship between AD and interaction terms (CD34+CD133+ frequency and genome-wide SNP dosage) was tested using GEEPACK (logistic regression utilizing generalized estimating equations) while adjusting for age, sex, years of education, and the first 10 population principal components (PCs) for controlling ancestry/population substructures. Additive models were assumed. Only autosomal SNPs on chromosomes 1–22 were considered. A statically filtered set of dosage files for MACH imputation via the 1,000 genomes reference files from March 2012 of only samples of European ancestry was used. The filtering criterion was r² ≥ 0.1. Moreover, only SNPs with minor allele frequency (MAF) ≥ 10% were chosen for the analysis. In addition, samples with a genotyping rate less than 97% and with an excess number of heterozygote observations or Mendelian errors were removed. Manhattan plot, Q-Q plot, and genomic control [31] were used for visualization and quality control, and LocusZoom [32] was used to present the regional information. Only P values < 5.0 × 10^-8 were considered to be statistically significant for the SNP-CD34+CD133+ progenitor cell interactive effects for AD.

ROSMAP data were used to investigate the functions of the variants at the DNA methylation levels and gene expression levels and to explore their relationships with AD (https://dss.niagads.org/cohorts/religious-orders-study-memory-and-aging-project-rosmap/). ROSMAP participants have annual blood draws that provide a repository of stored serum, plasma, and cells. Clinical evaluation, self-report, and medication review are used to document medical conditions. Moreover, subjects are neurologically evaluated every year, and a review of all of the ante-mortem data at the time of death leads to a final clinical diagnosis for each participant; specifically, each individual receives a diagnosis of AD, mild cognitive impairment (MCI), or no cognitive impairment (NCI). The details of this methodology have been described on the website.

We additionally explored the gene expression patterns across different AD brain regions by using Agora (https://agora.adknowledgeportal.org/genes), which included 9 brain regions, e.g., the anterior cingulate cortex (ACC), cerebellum (CBE), dorsolateral prefrontal cortex (DLPFC), frontal pole (FP), inferior frontal gyrus (IFG), posterior cingulate cortex (PCC), parahippocampal gyrus (PHG), superior temporal gyrus (STG), and temporal cortex (TCX). Two AD-related traits, e.g., BRAAK (neurofibrillary tangles) and CERAD (neuritic plaques), were also measured. In addition, GTEx Portal (https://gtexportal.org/home/) was used to study the associations between the selected SNPs and gene expression (eQTLs). The details of these procedures have been previously described [33].

We used Brain xQTLServe from ROSMAP (https://mostafavilab.stat.ubc.ca/xQTLServe/) to explore the associations between SNPs and DNA methylation (mQTLs). Genotype data were generated from 2,093 individuals of European descent. Of these individuals, DNA methylation (450K Illumina array, n = 468) was derived from postmortem frozen samples of a single cortical region, DLPFC. Gene expression data were generated by using RNA-sequencing (Illumina HiSeq) from the DLPFC of 494 individuals at an average sequence depth of 90 M reads. The details of this methodology have been previously described [34].

Data were analyzed by using R version 4.2.1. The Cox models were applied to circulating cells including CD34+CD133+, CD34+, CD34+CD133–, CD34–CD133+, CD34–CD133–, or CD34+KDR+ (VEGFR2), CD31+, CD31–, CD31+DIM, and CD31+lymphoid cells to examine the relationship between CD34+CD133+ or quartiles and dementia/AD diagnosis risk with adjustments for age, sex, years of education, APOE ε4, and different vascular diseases (Table S1). Only CD34+CD133+ progenitor cells were found to be associated with AD dementia.

Circulating CD34+CD133+ progenitor cells were first divided into four quartiles based on cell numbers. Quartiles of cell numbers were compared on the number of subjects, age at baseline, sex, years of education, APOE ε4, incident AD/dementia status, CHD, HTN, and different cerebrovascular diseases using analysis of variance (ANOVA), Chi-squared tests, and Kaplan-Meier survival analysis. We further stratified the subjects into the absence and the presence of different vascular pathologies and examined the relationship between CD34+CD133+ quartiles and dementia/AD diagnosis risk by using the Chi-squared test in each subgroup. Using ANOVA and Cuzick trend test, we also studied the relationship between CD34+CD133+ quartiles and brain volumes in the absence and presence of vascular diseases.

Furthermore, the Cox model was applied to the stratification of 1) no vascular diseases, 2) peripheral vascular diseases, and 3) cerebrovascular diseases with adjustments for age, sex, years of education, and APOE ε4. Additionally, the main effects and interaction effects (with CD34+CD133+) of GWAS-selected SNPs were obtained by using logistic regression and Cox proportional hazard regression adjusting for age, sex, years of education, APOE ε4, and PCs. These interaction effects were further assessed by applying stratification analyses of genotypes with Cox proportional hazard regression models and a Kaplan-Meier survival analysis.

We used the existing data from the FHS Gen2 cohort, which included 1,566 participants with an average age (mean ± SD) of 65.93 ± 8.91 years at exam 8 when circulating endothelial cells were measured at baseline, of which 46.36% were females. The concentrations (cell number per mL) and proportions of total PBMCs (%) of different EPCs and EMCs in this cohort were shown (Table 1, 2nd column).

There were 59 cases of incident AD after exam 8. Using Cox proportional hazards regression models, we examined the relationships between the proportions of different circulating cell populations (log-transformed) with following EPC marker combinations, CD34+CD133+, CD34+CD133–, CD34+, CD34–CD133+, CD34–CD133–, or CD34+KDR+ (VEGFR2), and AD or all-cause dementia risk. In addition, we studied EMCs including CD31+, CD31–, CD31+DIM, and CD31+lymphoid cells for AD risk using the same analyses. Among them, only CD34+CD133+ EPCs, but none of the other cell types, significantly impacted the risk of AD (HR = 0.64, P = 0.02) and all-cause dementia (HR = 0.63, P = 0.006) after adjusting for age, sex, years of education, APOE ε4, and vascular diseases (Table 1).

We further studied CD34+CD133+ EPCs in detail. The median concentration of CD34+CD133+ EPC counts was 2.61 × 10^-4/mL; the median proportion of circulating CD34+CD133+ EPCs was 0.032% mononuclear cells measured at baseline. The relationship between CD34+CD133+ cell proportion (%) and AD risk or all-cause dementia risk was studied after adjusting for different covariates (Table 2). CD34+CD133+ EPC proportion was associated with reduced risk of AD or all-cause dementia without adding covariates (Model 1), and remained significant after adjusting for covariates of age, sex, years of education (Model 2), and the addition of APOE ε4 and vascular diseases (Model 3), showing that a higher frequency of circulating CD34+CD133+ EPCs decreased AD risk (HR = 0.64, 95% CI = 0.44–0.94, P = 0.02) and all-cause dementia risk (HR = 0.63, 95% CI = 0.45–0.87, P = 0.006). We also used the concentration (per mL) of CD34+CD133+ EPCs and found similar associations with reduced risk of AD dementia (Table S1).

The participants were then divided into four quartiles based on the circulating CD34+CD133+ proportions (%) (Q1 = 0.002–0.020; Q2 = 0.020–0.032; Q3 = 0.032–0.049 and Q4 = 0.049–0.609). We found that younger (P = 0.007) and female (P = 0.002) participants were more likely to be in the higher CD34+CD133+ quartile (Table 3), with no significant differences in education and APOE ε4 carriers across the four quartiles. Vascular disease rates did not show statistical difference across the four CD34+CD133+ quartiles (Table 3). However, after adjusting for the confounders, CD34+CD133+ EPC cells were positively associated with HTN and body mass index (BMI) with statistical significance (Table S2). In contrast to positive associations with some vascular diseases, with an increase in the CD34+CD133+ EPC quartiles, the incident rates of AD (5.85% vs. 3.94% vs. 2.99% vs. 2.09%, P = 0.04) and all-cause dementia (7.66% vs. 5.56% vs. 4.41% vs. 2.86%, P = 0.02) decreased (Table 3). In parallel, Kaplan-Meier survival curves consistently revealed a reverse and a dose-dependent relationship between four increased CD34+CD133+ quartiles and decreased AD probability (P = 0.03, Figure 1).

We hypothesized that CD34+CD133+ EPCs would show a significant association with AD dementia when vascular damage exists, in which the repair by the EPCs is needed. To approach this, the participants were stratified into those without any vascular disease vs. peripheral vascular diseases vs. cerebrovascular diseases. We found that increased CD34+CD133+ EPC frequency was significantly associated with decreased AD risk among those who had peripheral vascular diseases (HR = 0.62, 95% CI: 0.40–0.94, P = 0.02) or cerebrovascular disease (HR = 0.58, 95% CI: 0.35–0.96, P = 0.03, Table 2). Whereas this trend was not observed in individuals without any vascular diseases although the number of participants was small. Similar trends were observed for all-cause dementia risk among those individuals with peripheral vascular diseases (HR = 0.61, 95% CI: 0.43–0.88, P = 0.008) or with cerebrovascular disease (HR = 0.53, 95% CI: 0.35–0.81, P = 0.003). Again, similar to the proportion (%), the concentration of CD34+CD133+ EPCs per mL was found to have similar associations with AD dementia in the presence of peripheral or cerebrovascular diseases (Table S1).

We then examined this relationship in each specific vascular disease (Figure 2A–N). We found that increased CD34+CD133+ quartiles had a significantly dose-dependent association with decreased AD risk among participants who had HTN (P = 4.6 × 10^–4, Figure 2E) and CMB (P = 0.01, Figure 2K). Increasing CD34+CD133+ EPC quartiles also had a reverse relationship with the risk of all-cause dementia among those who had CHD (P = 0.05, Figure 2D), HTN (P = 1.8 × 10^–4, Figure 2F), stroke (P = 0.02, Figure 2H), silent infarct (P = 0.01, Figure 2J), CMB (P = 0.02, Figure 2L), or had a high WMHI level (P = 0.03, Figure 2N).

In this FHS cohort, compared to those participants without any vascular diseases (0.78%), participants with peripheral vascular diseases, including CHD (8.27%, P = 4.1 × 10^–5) and HTN (7.42%, P = 7.5 × 10^–5), and participants with cerebrovascular diseases, including clinical diagnoses of stroke (10.83%, P = 3.3 × 10^–6), silent infarct (10.97%, P = 1.8 × 10^–6), CMB (15.12%, P = 1.8 × 10^–8), and high volume of WMHI (7.59%, P = 7.0 × 10^–5), had significantly higher rates of incident AD. A similar but stronger trend was observed for the incident rates of all-cause dementia in participants with these vascular diseases. However, we did not find the interactive effects between CD34+CD133+ EPCs or the other cell types and the combined vascular diseases for the risk of AD dementia (Table S3), probably due to that the proportion of those without any vascular diseases was small (n = 256) with only 3 cases of AD onset so that we did not have power to detect the interactive effects.

To study the relationship between CD34+CD133+ EPCs and brain volumes, we used a subsample (n = 1,172) with available brain MRI data. Since white matter pathologies in the brain are linked with vascular diseases [35], we examined brain gray versus white volumes of different brain lobes across four CD34+CD133+ proportion quartiles in the absence and the presence of vascular diseases by using ANOVA and Cuzick trend test (Table 4). We examined 11 brain regions and found that an increase in the CD34+CD133+ EPC quartiles was associated with bigger cerebrum white matter, but not grey matter, volumes in the presence of vascular diseases (P = 0.004), especially related to the white matter volumes in frontal lobe (P = 0.007) and temporal lobe (P = 0.046). In addition, the higher CD34+CD133+ EPC quartiles, the bigger volume of hippocampus in the presence of vascular diseases (P = 0.023). We also studied these associations in the presence of particular vascular diseases. In the presence of HTN, CD34+CD133+ EPCs were significantly associated with the brain volumes of cerebrum white matter (P = 0.0006), frontal white matter (P = 0.004), temporal white matter (P = 0.014), and hippocampus (P = 0.0016, Table 4). This finding was consistent with the relationship between CD34+CD133+ EPCs and reduced AD risk in HTN (Figure 2E). In the presence of CMB, CD34+CD133+ EPCs were significantly associated with frontal gray matter and hippocampus, but the relationships were not linear.

Studies have shown that brain endothelial damage increases AD pathogenesis in the brain in certain genotype carriers [6], therefore, we hypothesized that peripheral circulating CD34+CD133+ EPCs may reduce AD risk in particular genotypes. To test this hypothesis, we first conducted a GWAS for AD with the interaction term between CD34+CD133+ cell frequency and SNP as predictor using a logistic regression model. For SNPs passing a genome-wide significance (P < 5.0 × 10^–8), a Cox regression model was also applied to validate the interaction effects on AD incidence. As depicted in the Manhattan plot, 8 SNPs showed significant interactions with CD34+CD133+ frequency for AD risk (Figure 3A). Seven of the 8 SNPs with a P value < 5.0 × 10^–8 were located in gene KIRREL3 on chromosome 11 (two of the 7 SNPs selected for further analysis: rs580382 and rs4144611) and 1 of the 8 SNPs was in gene EXOC6B (chromosome 2, rs61619102) (Figure 3A, Figure S2). These associations are supported by evidence with surrounding SNPs (Figures 3B and 3C, Table S4). Notably, for these 3 SNPs, although their main SNP effects were not associated with AD risk (P > 0.40), their interactive effects with CD34+CD133 cells were significantly associated with AD after adjusting for age, sex, years of education, APOE ε4, and PCs using both logistic and Cox regression analyses (Table 5).

Consistently, in genotype stratification analysis adjusting for age, sex, years of education, APOE ε4, and PCs, we found that CD34+CD133+ frequency was negatively associated with AD incidences among the homozygotes of KIRREL3 rs4144611 TT carriers (HR = 0.29, 95% CI: 0.15–0.57, P < 0.001), KIRREL3 rs580382 CC carriers (HR = 0.31, 95% CI: 0.17–0.57, P < 0.001), and EXOC6B rs61619102 CC carriers (HR = 0.49, 95% CI: 0.31–0.75, P < 0.001) (Table S5). Further examination of these findings revealed a dose-dependent effect of increased CD34+CD133+ cell quartiles on the reduction in AD risk among KIRREL3 rs4144611 TT or rs580382 CC homozygotes as well as in EXOC6B rs61619102 CC homozygotes (Table 6, Table S6, and Table S7). In contrast, these impacts from CD34+CD133+ cells on AD risk were not present in the counterpart genotypes of these 3 SNPs, e.g., KIRREL3 rs4144611 either TG or GG; rs580382 either CT or TT; as well as EXOC6B rs61619102 either GC or GG carriers (Tables S5, S6, and S7).

These relationships were also observed in Kaplan-Meier survival curves. Among the homozygotes of KIRREL3 rs4144611 TT carriers (P = 0.003, Figure 1B), KIRREL3 rs580382 CC carriers (P = 1.2 × 10^–4, Figure 1C), and EXOC6B rs61619102 CC carriers (P = 0.001, Figure 1D), higher CD34+CD133+ quartiles corresponded to a lower incidence of AD development and difference in AD-free probability with statistical significance. Again, among their counterpart genotype carriers (KIRREL3 rs4144611 GG + TG or rs580382 TT + CT carriers and EXOC6B rs61619102 GG + GC carriers), there were no such relationships between CD34+CD133+ quartiles and AD or all-cause dementia risk. Additionally, unlike for AD dementia risk, we did not find associations of these genotypes or the interactive effects of these genotypes and CD34+CD133+ cells for any vascular disease (data not shown).

To address why these two genes could interact with CD34+CD133+ EPCs for AD risk, we explored if the mRNA expressions level of gene KIRREL3 and EXOC6B are associated with AD pathology. We first used the RNA-seq dataset generated from the DLPFC region and monocytes in the ROSMAP study. As shown in Figure S3, brain KIRREL3 expression was associated with AD (P = 0.02), but peripheral KIRREL3 expression was not (Figure S3A). Compared to controls, brain expression level of KIRREL3 was lower in AD brains specifically in the TCX (P = 2.7 × 10^–5) regions (Figure S3B). To investigate the relationship between KIRREL3 SNPs and lower mRNA levels in the AD brain, we investigated the epigenetics of KIRREL3 SNPs and found that the methylation site cg11751545 was consistently negatively correlated with KIRREL3 gene expression in DLPFC (Figure S4A). Two KIRREL3 genotypes (rs4144611 GG + GT or rs580382 TT + CT) were negatively associated with DNA methylation level of two CpG sites within KIRREL3 (cg11751545; cg04445570; P < 0.05) (Figure 3B, Figure S4B). Based on these results, we hypothesized that the rescue effect of CD34+CD133+ EPCs may reduce AD risk only in vulnerable KIRREL3 rs4144611 TT carriers (Table 5 and Figure 1B and Figure S3C) and rs580382 CC carriers (Tables 5 and 6) because the low brain expression of KIRREL3 is regulated by the high brain methylation of cg11751545 in these carriers.

While peripheral monocyte expression of EXOC6B was significantly and positively associated with AD (P < 0.001), the brain expression of EXOC6B was not associated with AD after Bonferroni correction (Figure S5A). Consistently, EXOC6B rs61619102 G allele was positively associated with EXOC6B gene expression as an eQTL compared to C allele in the following two types of peripheral tissues: visceral and subcutaneous adipose tissues (Figure S5B). Thus, we reasoned that the impact of EXOC6B polymorphisms on AD was mainly elicited through the peripheral system and that the EXOC6B rs61619102 homozygous genotype (CC) itself (which was associated with a low level of expression in the peripheral system) interacted with circulating CD34+CD133+ EPCs to reduce AD risk (Table 5; Figure 1D and Figure S5C).