Genetic and regulatory architecture of Alzheimer's disease in the APOE region.

INTRODUCTION: Apolipoprotein E (APOE) ε2 and ε4 alleles encoded by rs7412 and rs429358 polymorphisms, respectively, are landmark contra and pro "risk" factors for Alzheimer's disease (AD).
METHODS: We examined differences in linkage disequilibrium (LD) structures between (1) AD-affected and unaffected subjects and (2) older AD-unaffected and younger subjects in the 19q13.3 region harboring rs7412 and rs429358.
RESULTS: AD is associated with sex-nonspecific heterogeneous patterns of decreased and increased LD of rs7412 and rs429358, respectively, with other polymorphisms from five genes in this region in AD-affected subjects. The LD patterns in older AD-unaffected subjects resembled those in younger individuals. Polarization of the ε4- and ε2 allele-related heterogeneous LD clusters differentiated cell types and implicated specific tissues in AD pathogenesis. DISCUSSION: Protection and predisposition to AD is characterized by an interplay of rs7412 and rs429358, with multiple polymorphisms in the 19q13.3 region in a tissue-specific manner, which is not driven by common evolutionary forces.

BACKGROUND

The strongest evidence for genetic predisposition to Alzheimer's disease (AD) was reported for the apolipoprotein E (APOE)/ translocase of outer mitochondrial membrane 40 (TOMM40) region 19q13.3 with the APOE ε4 allele as the strongest genetic risk factor for AD development in various populations1 and the APOE ε2 allele as a protective factor against AD.2, 3 However, even the pathogenic role of the ε4 allele in AD remains poorly understood, consistent with the inefficiency of AD clinical trials4 and finding of cognitively normal homozygous ε4 carriers among centenarians.5 Understanding the protective role of the ε2 allele has lagged behind the ε4 research because of, in part, seemingly smaller effects of this allele on AD.3

Mainstream research considers the effects of risk alleles in genetics of such complex traits as AD as a result of incomplete penetrance.6 We emphasize inherent heterogeneity in the effects of the same alleles on AD. This view is supported by evolutionary biology, which argues that the conceptual problem in the genetics of traits that make bodies vulnerable to disease(s) in post‐reproductive life, called age‐related traits, is an uncertain role of evolution in establishing their molecular mechanisms.7 Increased human life expectancy8 and changes in the environment9, 10, 11, 12 contribute to this problem. Accordingly, in the framework of evolutionary biology, age‐related traits are viewed as the results of indirect mechanisms such as co‐evolution with fast‐evolving pathogens, mismatch with environments, reproductive success at the expense of health, and so on,7 that increase heterogeneity.

Following the framework of evolutionary biology, we examined the molecular signatures of AD in the APOE region, represented by 32 single nucleotide polymorphisms (SNPs) from five genes (BCAM, NECTIN2, TOMM40, APOE, and APOC1), as differences in linkage disequilibrium (LD) patterns in mega‐samples of 2673 AD‐affected and 16,246 AD‐unaffected subjects of European ancestry. We emphasized protective and detrimental heterogeneous signatures involving the APOE ε2 and ε4 alleles, encoded by rs7412 and rs429358, respectively. We show that susceptibility to AD is the result of a complex interplay of these SNPs with SNPs from other genes in the APOE region, which is not driven by common evolutionary forces characteristic for the general (AD‐unaffected) population.

METHODS

Data availability

This article was prepared using limited access data sets obtained though dbGaP (accession numbers phs000007.v28.p10, phs000287.v5.p1, phs000428.v1.p1, and phs000168.v2.p2) and the University of Michigan. Phenotypic Health and Retirement Study (HRS) data are available publicly and through restricted access from http://hrsonline.isr.umich.edu/index.php?p=data.

RESEARCH IN CONTEXT

Systematic review: Recently, we reported significant molecular signatures of Alzheimer's disease (AD) in the apolipoprotein E (APOE) region, which excluded the ε2 and ε4 alleles. A literature review (PubMed and Google Scholar) identified few other publications, which reported significant associations of linkage disequilibrium (LD) structures with AD. These relevant publications are appropriately cited.

Interpretation: Susceptibility to AD is the result of a complex interplay of the ε2 and ε4 alleles with other alleles from different genes in the APOE region, which is not driven by common evolutionary forces. Accordingly, this interplay is the result of AD‐specific exposures, which, therefore, can be amendable to AD preventive interventions even with natural, for example, lifestyle, factors.

Future directions: This work suggests an approach to examine the potential role of complex genotypes/haplotypes in the AD etiology in loci with complex LD structures. Further work should be focused on elucidating personalized, that is, more homogeneous, group‐specific, polygenic profiles of AD risk and protection.

Study cohorts and phenotypes

We used data from five studies. Data for older individuals were drawn from the Framingham Heart Study (FHS) original (FHS_C1) and offspring (FHS_C2) cohorts,13 Cardiovascular Health Study (CHS),14 Health and Retirement Study (HRS),15 and the National Institute on Aging (NIA) Late‐Onset Alzheimer's Disease Family Based Study (LOADFS)16 for individuals of Caucasian ancestry. In LOADFS, FHS, and CHS, AD was defined based on diagnoses made according to National Institute of Neurological Disorders and Stroke and the Alzheimer's Disease and Related Disorders Association. A diagnosis of AD in HRS was defined based on ICD‐9:331.0x codes in Medicare service use files. Individuals with AD constituted the case group, n = 2673, and those without AD constituted the non‐case group, n = 16,246 (Table 1). Data from the FHS third‐generation cohort (FHS_C3) and Coronary Artery Risk Development in Young Adults (CARDIA) cohort (Table 1) were used in comparative analyses of LD patterns in younger and older individuals.

Table 1

Basic characteristics of the genotyped participants in the selected studies

Cohort	N	AD cases (%)	Men (%)	Birth year mean (SD)	Age at baseline mean (SD), years	Age at DNA mean (SD), years	Age at the end of follow‐up mean (SD), years	Follow‐up through
LOADFS	3715	1850 (49.8)	1395 (37.6)	1928.5 (12.5)	73.5 (12.5)	73.5 (12.5)	77.3 (10.9)	2015a
HRS	7226	263 (3.6)	3129 (43.3)	1934.2 (8.4)	60.6 (8.7)	73.2 (8.4)	79.1 (8.1)	2012
CHS	4326	252 (5.8)	1884 (43.6)	1914.1 (5.7)	72.8 (5.6)	73.5 (5.7)	83.5 (5.4)	2002
FHS_C1	631	205 (32.5)	210 (33.3)	1911.8 (4.2)	35.7 (4.2)	84.1 (4.3)	91.4 (4.8)	2012
FHS_C2	3021	103 (3.4)	1383 (45.8)	1935.8 (9.6)	34.7 (9.7)	60.3 (9.7)	72.2 (9.2)	2012
FHS_C3	3980	NA	1862 (46.8)	1960.5 (8.9)	40.2 (8.8)	40.2 (8.7)	47.8 (9.0)	2012
CARDIA	1941	NA	909 (46.8)	1957.5 (3.5)	25.0 (3.6)	25.0 (3.6)	40.4 (3.8)	2011

AD denotes Alzheimer's disease and related dementias.

N denotes genotyped sample after excluding individuals with missingness for SNPs >5% and missing information on AD.

Large proportion of AD cases in LOADFS is due to case‐control design.

Large proportion of AD cases in FHS is due to older age of participants of this cohort at the end of follow‐up (mean age for total sample is 91.4 years) and larger proportion of women (66.7%) who are at higher risk of AD.

CHS, Cardiovascular Health Study; FHS_C1, Framingham Heart Study (FHS) original cohort; FHS_C2, FHS offspring cohort; FHS_C3, FHS third generation cohort; HRS, Health and Retirement Study; LOADFS, NIA Late‐Onset Alzheimer's Disease Family Study; CARDIA, Coronary Artery Risk Development in Young Adults cohort; NA, not applicable; SD, standard deviation.

Information on age at onset of AD in LOADFS was not known for all cases.

Genotypes

Genotyping was performed using the same customized Illumina iSelect array (the IBC‐chip, ≈50 K SNPs) in the FHS and CHS cohorts, Affymetrix 500 K in the FHS, Illumina HumanCNV370v1 chip (370 K SNPs) in the CHS, Illumina HumanOmni 2.5 Quad chip (≈2.5 M SNPs) in the HRS, and Illumina Human 610Quadv1_B Beadchip (≈610 K SNPs) in the LOADFS.

Thirty‐two SNPs representing the BCAM‐NECTIN2‐TOMM40‐APOE‐APOC1 locus (Table S1) were not in perfect LD (r < 0.8) and directly genotyped in at least two cohorts.

We excluded individuals with >5% missingness. For cross‐platform comparisons, we selected directly genotyped target SNPs or their proxies (r > 0.8 in the 1000 Genomes Project, CEU population) using all available arrays for each study. Non‐genotyped SNPs were imputed (IMPUTE217) according to the 1000 Genomes Project Phase 3 integrated variant set release (SHAPEIT2) in the NCBI build 37 (hg19) coordinate. Retaining SNPs with high imputation quality (info > 0.8), rs11668536 in FHS/FHSO (info < 0.66) was excluded (details in Table S1).

Statistical analysis

Associations between AD and each selected SNP were evaluated using an additive genetic model, with the minor allele as an effect allele. Given limited information on AD age at onset in the LOADFS, the associations in this study were characterized using a logistic model with AD as a binary outcome and random effects to adjust for potential familial clustering (gee package in R). Associations in the other studies were evaluated using the Cox proportional hazard mixed‐effects regression model (coxme package in R) to adjust for familial clustering. The time variable in the Cox model was the age at onset of AD or the age at right censoring in 2002 (CHS) and 2012 (FHS and HRS). All statistical tests were adjusted for (all studies) age, sex; (CHS) field center; (FHS) whether the DNA samples had been subject to whole‐genome amplification; and (HRS) HRS cohorts. Meta‐statistics were evaluated using METAL.18

Linkage disequilibrium analysis

We have used methods detailed in Ref. 19. In brief, LD was characterized by the correlation coefficient r using haplotype‐based and genotype‐based methods. Differences in their LD estimates indicate deviation from Hardy‐Weinberg equilibrium (HWE). This information is important because HWE in the entire sample does not guarantee HWE in subsamples and/or at the haplotype level (see below), and thus, the observed deviation from HWE may be biologically plausible. Significance of the LD estimates was characterized using chi‐square statistics, defined as χ2 = rN, where N = 2n is the number of gametes and n is the sample size. Given the potential loss of power because of inferring haplotypes from genotypes, we used a more conservative estimate, with n instead of N. We employed an LD contrast test20 to compare the LD estimates between the AD‐affected and unaffected groups. This test was used to characterize the significance of the differences in pairwise estimates of LD between these two groups. Significance of the r estimates and the differences in the pairwise estimates of LD were corrected for multiple testing. For the 32 SNPs examined, this represented 496 (=32 × 31/2) tests. We adopted a conservative Bonferroni correction for significance, P ≤ 10−4, despite some correlation between these SNPs. Asymptotically valid confidence intervals were constructed using asymptotic variance adapted from.21

Functional annotation

Potential regulatory functions of the selected SNPs were annotated using the Ensembl genome browser (https://www.ensembl.org/), RegulomeDB (http://www.regulomedb.org/), and HaploReg (http://archive.broadinstitute.org) databases. Information on expression quantitative trait loci (eQTLs) was obtained from the GTEx (v7 release) portal (https://www.gtexportal.org/).

RESULTS

Study overview

Molecular signatures of AD were examined as the difference of LD patterns in mega‐samples of AD‐affected and unaffected subjects of Caucasian ancestry, with men and women combined and separately pooled from four independent studies comprising five cohorts: LOADFS, HRS, CHS, FHS_C1, and FHS_C2 (Table 1). LD patterns were characterized by 32 non‐proxy SNPs (defined as LD with r < 0.8), representing the BCAM, NECTIN2, TOMM40, APOE, and APOC1 genes in the 19q13.3 region (Table S1) including two SNPs, rs429358 and rs7412 SNPs, whose minor alleles encode the APOE ε4 and ε2 alleles, respectively. We examined the potential role of survival selection in the AD signatures by contrasting LD patterns between older AD‐unaffected individuals from those five cohorts (who were at exponentially increased mortality risk) and younger individuals (who were at negligible mortality risk), enriched by subjects from two additional cohorts, FHS_C3 and CARDIA (Table 1).

Unless explicitly stated, the results of LD analyses are presented using a haplotype‐based method (details in Materials and Methods).

Of the examined 32 SNPs, the minor allele of rs429358 was associated with the highest risks of AD development, β = 1.26, P = 8.05 × 10−130, whereas the minor allele of rs7412 showed the strongest protective effect, β = −0.59, P = 1.02 × 10−9 (Table S1). The effect directions were consistent in all studies for rs429358, but not for rs7412. The largest magnitude of effects for these SNPs was observed in LOADFS (β = 1.78, P = 9.33 × 10−97 for rs429358 and β = −1.50, P = 2.77 × 10−14 for rs7412) and the smallest in FHS_C1 (β = 0.51, P = 1.87 × 10−3 for rs429358 and β = 0.10, P = 6.48 × 10−1 for rs7412) (Figure 1).

Figure 1

Forest plots for the associations of (A) rs429358 (ε4‐coding SNP) and (B) rs7412 (ε2‐coding SNP) with Alzheimer's disease (AD). LOADFS, NIA Late‐Onset Alzheimer's Disease Family Study; HRS, Health and Retirement Study; CHS, Cardiovascular Health Study; FHS_C1, Framingham Heart Study (FHS) original cohort; FHS_C2, FHS offspring cohort; SE, standard error; N, sample size. Meta indicates the results from the meta‐analysis. Horizontal bars show 95% confidence intervals

Molecular signature of Alzheimer's disease

We contrasted LD patterns of the entire APOE region between AD‐affected and unaffected individuals (Table S2) and found that they differed significantly (P < 2 × 10−4). The pattern of the difference represents a molecular signature of AD illustrated by a heat map for Δr = r (Figure 2). Figure 2 shows the complex rearrangement of LD in AD cases compared with non‐cases spanning the entire region. Our analysis identified 193 of 496 (=32 × 31/2) SNP pairs (38.9%) with Δr values significant at the Bonferroni‐adjusted level: P ≤ P = 10−4. For 33 additional SNP pairs, we observed suggestive significances: P < P < 10−3.

Figure 2

Molecular signature of Alzheimer's disease (AD). Upper‐left triangle: Linkage disequilibrium (LD) pattern (r, %) in the pooled sample from all studies, non‐cases, for 32 single nucleotide polymorphisms (SNPs). Lower‐right triangle: Heat map for Δr = r−r representing the molecular signature of AD. Red denotes r > r and blue denotes r < r. Purple and yellow show the estimates with opposite signs of r and r. For convenience, positive sign of r has been selected. Legend on the right shows color‐coded P‐values. The heat map shows that LD changes for the vast majority of SNPs in the entire region spanning all five genes. Numerical estimates are shown in Table S2

Molecular signatures of AD estimated using the genotype‐based method (Table S3) were qualitatively the same as those estimated using the haplotype‐based method, with significant differences observed between cases and non‐cases (P < 2 × 10−4). The genotype‐based method provided 153 SNP pairs significant at P < P and 33 additional SNP pairs with suggestive significance (P < P < 10−3).

For 149 SNP pairs, the estimates of Δr were significant at P ≤ P in both the haplotype‐ and genotype‐based methods. Given that all SNPs in the large sample of non‐cases were in HWE at P HW > 10−3, the discordant estimates of Δr for 44 SNP pairs between these two methods indicated SNPs with a plausible biological role because the deviation from HWE occurs in cases (Table S1) and/or at the haplotype level, that is, when (see Materials and Methods). Accordingly, important biologically plausible information can be missed using the genotype‐based method alone.

The APOE ε2 (rs7412) and ε4 (rs429358) coding SNPs are parts of the molecular signature of AD

In non‐cases, rs7412 and rs429358 SNPs were in significant LD between each other, r = 11.6%, P = 7.95 × 10−94, and with most of the other SNPs (Table S2). The strongest LD for rs429358 was observed with rs2075650 (r = 70%, TOMM40) and rs12721046 (r = 69%, APOC1) SNPs. For rs7412, the strongest LD of r = 37% was with rs283813 (NECTIN2).

Rearrangement of LD between AD cases and non‐cases was characterized by a significant increase in LD of rs429358 with 13 SNPs, including rs7412 (Figure 3B), and decrease in LD of rs7412 with 8 SNPs (Figure 3A). Although the change in LD was somewhat larger for rs429358 with nearby SNPs from the TOMM40‐APOE‐APOC1 locus (Figure 3), LD changed regardless of genomic distance between the other SNP pairs. LD of rs429358 and rs7412 SNPs changed in opposite directions with the same four SNPs (rs8106922, rs405509, rs440446, and rs439401) from the TOMM40‐APOE‐APOC1 locus. LD for rs429358 and rs7412 with SNPs from the BCAM‐NECTIN2 locus also changed in opposite directions but for non‐overlapping SNPs. Significant changes in LD between rs7412 and 8 SNPs as well as between rs429358 and 13 SNPs were not explained by LD between those 8 or 13 SNPs. This is because LD between these 8 or 13 SNPs can be very small (Figure 3, brackets) and, therefore, it cannot be explained by clustering of specific alleles from different SNPs in the same subjects. The latter implies genetic heterogeneity. The changes in LD between AD cases and non‐cases observed for rs7412 and rs429358 in the mega sample of pooled studies were consistent in independent studies (Table S4). Consistency of changes in LD for other SNPs in independent studies was reported in Ref. 22.

Figure 3

Significant ε2‐ and ε4‐related molecular signatures of Alzheimer's disease (AD). (A) The ε2‐related signature is characterized by a significant decrease (blue) in linkage disequilibrium (LD) for rs7412 with eight single nucleotide polymorphisms (SNPs) in AD cases compared with non‐cases. (B) The ε4‐related signature is characterized by a significant increase (red) for LD of rs429358 with 13 SNPs, including rs7412, in AD cases compared with non‐cases. Insets show examples of small LD between SNPs indicated by brackets. Vertical lines show 95% confidence intervals. Numerical estimates are shown in Table S2

Molecular signatures of AD in men and women

We evaluated LD structure for the selected 32 SNPs in AD‐affected and unaffected men and women separately (Table S5). The 95% confidence intervals for Δr in men and women well overlapped for all SNP pairs, implying no significant difference in Δr between these sexes.

LD patterns in younger and older individuals

We examined the role of survival selection in the molecular signature of AD by contrasting LD patterns in older subjects with no AD who were 55 years and older at biospecimen collection (N = 14,803), and younger subjects who were <55 years at biospecimen collection (N = 6565). We excluded four SNPs from this analysis (rs7026, rs4803760, rs440277, and rs11667640) because they were imputed for most subjects (95.4%) from the young group. The 55‐year cutoff was used to separate younger individuals who were under negligible mortality risk in modern developed countries from those who were under exponentially increasing mortality risk. This choice allowed consideration of LD patterns in the younger group as a proxy for the evolutionary selected LD structure in the APOE genomic region. This analysis did not identify significant differences in LD patterns between these two groups. At the level of individual SNP pairs, only two pairs in the BCAM‐NECTIN2 locus exhibited significant differences (Δr at P ≤ P = 10−4) in these large samples (Table S6). No significant differences were identified in the TOMM40‐APOE‐APOC1 locus (Figure 4).

Figure 4

Linkage disequilibrium (LD) patterns in younger and older individuals. Upper‐left triangle: LD pattern (r, %) in younger subjects who were <55 years at biospecimens collection for nine single nucleotide polymorphisms (SNPs) from the TOMM40‐APOE‐APOC1 locus. Lower‐right triangle: Heat map showing Δr as the difference in LD estimates in younger and older samples. Older sample included subjects with no Alzheimer's disease (AD) who were 55 years and older at biospecimens collection. Numerical estimates are shown in Table S6

Regulatory architecture in the APOE region across cell types and tissues

Using data from Ensembl, 10 of 32 SNPs were identified as regulatory variants in active expression states in a variety of tissues ranging from one to 63 of the 68 cell types (Table 2). For seven of them, RegulomeDB assigned functionality scores of 1b to 2a corresponding to strong regulatory potential (Table 2). Most SNPs may affect transcription factor (TF) binding ability. Altered motifs for TFs were identified for 28 SNPs in HaploReg (Table S7). The protein motifs at these sites are for known TFs that could contribute to the complex regulation of genes in this region. HaploReg showed that 10 SNPs could affect the binding of various proteins (from one to seven), suggesting that they could be in actively transcribed regions. Five more SNPs affected protein binding according to RegulomeDB (Table S7). Twenty‐six SNPs acted as eQTLs for the nearby protein‐coding genes, according to GTEx, affecting expression in a number of tissues (Table 2 and Table S7).

Table 2

Functional annotation of 32 SNPs in the APOE region

ID	SNP ID	LD cluster	Function	Regulatory feature	Gene	Active	Poised	Score	M0&M1 macrophage	CD14+ monocytes	NHA	NHLF	Selected eQTL
1	rs7026		3'UTR	PFR	BCAM	6	2	5					NECTIN2
2	rs1871045		Downstream		BCAM			4					NECTIN2
3	rs11668536		Downstream		BCAM								NECTIN2
4	rs10402271	ε4	Downstream		BCAM			5					NECTIN2
5	rs4803760	ε2	Intergenic					6					BCAM
6	rs1871046		Intron	Promoter, TFBS	NECTIN2	46	20	2a					No
7	rs4803763		Intron		NECTIN2			5					NECTIN2
8	rs440277	ε4	Intron	PFR	NECTIN2	5	1	1f	Yesa	Yesa	Yes		NECTIN2, FOSB
9	rs3852856		Intron		NECTIN2								NECTIN2
10	rs377702		Intron		NECTIN2			2b					NECTIN2, FOSB, CLASRP
11	rs8105340		Intron		NECTIN2			6					NECTIN2
12	rs12610605	ε4	Intron		NECTIN2			5
13	rs4803766		Intron		NECTIN2			5					NECTIN2
14	rs17561351		Intron	PFR	NECTIN2	5	1	1b					NECTIN2a
15	rs8104483		Intron	PFR	NECTIN2	5	1	1b					NECTIN2a
16	rs4081918	ε4	Intron	PFR	NECTIN2	5	1	1f	Yesa	Yesa	Yes		NECTIN2a
17	rs519113	ε2	Intron		NECTIN2			1f					BCAM, NECTIN2a
18	rs2075642		Intron		NECTIN2			5					NECTIN2
19	rs387976	ε2	Intron	OCR	NECTIN2	1	1	5				Yes	NECTIN2
20	rs11667640	ε4	Intron		NECTIN2			4					NECTIN2
21	rs6859	ε4	3'UTR		NECTIN2			4					NECTIN2
22	rs11673139	ε4	Intron		NECTIN2			4					NECTIN2, MARK4
23	rs283813	ε2	Intron		NECTIN2			5					No
24	rs157580	ε4	Intron	Promoter	TOMM40	63	5	1f	Yesa	Yesa	Yes		APOE, APOC1, DMPK
25	rs2075650		Intron		TOMM40			1f					No
26	rs8106922	ε2, ε4	Intron		TOMM40			5					DMPKa
27	rs405509	ε2, ε4	Upstream	Promoter	APOE			1f					APOE
28	rs440446	ε2, ε4	Missense intron	Promoter	APOE	18	32	4	Yesa	Yes	Yes	Yes	APOE, APOC1
29	rs429358		Missense	Coding region, exon 4				5					No
30	rs7412		Missense	Coding region, exon 4				4					APOE
31	rs439401	ε2, ε4	Non coding transcript exon	PFR	APOE‐APOC1	13	3	1b	Yesa		Yesa	Yesa	APOE, APOC1
32	rs12721046		Intron		APOC1			6					No

Linkage disequilibrium (LD) cluster indicates SNPs in LD with rs429358 and rs7412 SNPs, whose minor alleles code the APOE ε4 and ε2 alleles, respectively.

Table includes activity levels (active/poised) for 68 cell types (epigenoms).

Column “Score” shows RegulomeDB score based on the integration of multiple high‐throughput datasets with 1a being the highest score and 6 being the lowest score. Note that because RegulomeDB focuses on noncoding SNPs, missense SNPs may not have large scores.

NHA denotes epigenetic signature in normal human astrocytes cells.

NHFL denotes human lung fibroblasts.

eQTLs denote expression quantitative trait loci selected for affected protein‐coding gene in specific cell types.

OCR, open chromatin region; PFR, promoter flanking region; TF, transcription factor; TFBS, TF binding site.

Active state.

The APOE ε4 allele–related LD cluster (Figure 3B) includes five SNPs located in promoter regions of the associated genes, which were active in 5 to 63 of 68 cell types (Table 2). We found that all five regulatory variants shared the same feature, exhibiting the active state in M0 and M1 macrophages from venous blood. Four of them were active (rs440277, rs4081918, and rs157580) or poised (rs440446) in CD14+ monocytes (Table 2). One of five promoter variants, rs439401, was active in normal human astrocytes (NHAs) and four variants (rs440277, rs4081918, rs157580, and rs440446) exhibited a poised epigenetic signature in NHAs. The APOE ε2 allele–related LD cluster (Figure 3A) included two SNPs in promoter regions and rs387976 SNP in open chromatin, which were active in one to 18 of 68 cell types (Table 2). All three variants shared an active (rs439401) or poised (rs440446 and rs387976) expression state in normal human lung fibroblasts (NHLF). Variants in poised expression states can be epigenetically activated at a later stage in development or in response to exogenous stimuli.23, 24

DISCUSSION

Unlike few small‐scale prior studies examining associations of LD patterns with AD,25, 26 we found that AD was associated with a highly heterogeneous molecular signature in the APOE region, which included rs7412 and rs429358 encoding the ε2 and ε4 alleles, respectively, and SNPs from all five genes in the BCAM‐APOC1 locus, regardless of genomic distance between them. This signature is represented by the pattern of differences in LD structures between AD‐affected and unaffected subjects (Figure 2). The AD signature is consistent with a haplotype rather than a single allele origin of AD.27, 28, 29 Significant changes in LD indicate complex genetic architecture of AD in this region that is consistent with the view on AD as a continuum, rather than distinct clinically defined entities, driven by multimodal cognitive decline.30 No significant differences between the AD signatures in men and women were identified. Our results show that rs429358 and rs7412 are an inherent part of this signature. This finding indicates that the role of the ε4 and ε2 alleles in AD is dependent on the other SNPs in this locus. Indeed, decreased LD of rs7412 with eight SNPs in this locus in AD‐affected subjects compared with unaffected subjects shows that the larger LD strengthens the protective effect because the large LD is observed in unaffected subjects. Likewise, increased LD of rs429358 with 13 SNPs in AD‐affected subjects shows that the larger LD strengthens the detrimental effect because the larger LD is observed in AD‐affected subjects. Complexity of the molecular signature of AD implies that other SNPs in this locus can indirectly modify the effects of the ε4 and ε2 alleles in AD pathogenesis. Changes in the LD of the ε4‐ or ε2 allele–coding SNPs with the other SNPs in a heterogeneous manner (Figure 3) indicate more homogeneous carrier groups of detrimental or protective polygenic variants. This finding naturally strengthens a gene‐based precision‐medicine approach31 to AD treatment and prevention. The lack of the role of survival selection (Figure 4) in the AD signature implies that the LD pattern for the 32 SNPs in AD‐unaffected subjects was likely evolutionary selected, whereas that in AD‐affected subjects was not driven by the same evolutionary forces. This result offsets potential age‐related bias and is consistent with the uniquely human origin of AD, which is sensitive to the modern environment.32 More detailed analyses are required to better understand driving force of the AD signatures, for example, whether they are the result of AD‐related selection within a given human generation, AD‐related selection across recent generations within families or communities, or AD‐related divergence of ancestral groups.

Our bioinformatics analysis identified regulatory variants from the APOE ε4‐ and ε2 allele–related LD clusters (Figure 3), which shared the same features within each cluster. A hallmark for regulatory variants from the ε4‐allele LD cluster was an active state in primary macrophages (M0) and pro‐inflammatory M1 macrophages and active or poised expression states in CD14+ monocytes and NHAs. Monocytes that originate in the bone marrow can differentiate into specific tissue macrophages and dendritic cells in response to inflammation/infection. Blood monocyte–derived macrophages, representing innate immunity, can contribute to the immune response in the central nervous system (CNS) along with brain‐resident macrophages (microglia).33 A pro‐inflammatory (M1) macrophage response causes neurotoxicity.34 Enrichment in these specific immune cells is consistent with the role of peripheral monocytes/macrophages, along with microglia, in Aβ clearance and a potential role in AD.33, 35 It is important to note that our results are in line with recent advances implicating monocyte‐specific eQTLs in AD36 and the AD susceptibility alleles as significant eQTLs in CD14+ monocytes.37 Given crosstalk between macrophages/microglia and astrocytes, they show neurotoxic or neuroprotective phenotypes. M1 macrophages particularly induce astrocyte proliferation and a reactive phenotype. The interaction between macrophages and astrocytes plays an important role in the increasing inflammatory response leading to neurodegeneration.38 Astrocytes are implicated in the induction of neuroinflammation and AD, and apoE‐mediated Aβ clearance, which may be impaired by the reactive phenotype.39 Stressed, dysfunctional astrocytes are connected with ε4‐associated AD.40 Thus, the shared features of regulatory variants from the ε4 allele LD cluster highlight its connection with changes in immune response and inflammation in the CNS and the APOE ε4–dependent crosstalk of astrocytes with macrophages in neuroinflammation in AD. This suggests that the ε4 allele LD cluster is the result of rebalancing of neuroinflammatory tolerance mediated by astrocytes and macrophages in an exposure‐dependent manner.

A common feature of regulatory variants in the ε2 allele–related LD cluster is having an active or poised state in NHLFs. Lung fibroblasts play a role in airway inflammation and remodeling. Pulmonary health is important in risk prevention of cognitive decline and dementia.41 In addition, rs4803760 (intergenic NECTIN2‐BCAM) and rs519113 (NECTIN2) are eQTLs for BCAM in lung. The ε4‐ and ε2‐allele LD clusters have two common promoter variants (rs440446 and rs439401). Of interest, rs439401 is located in the APOE‐APOC1 intergenic region, which includes a specific macrophage, adipocyte, and astrocyte enhancer for the APOE gene,42 and the peroxisome proliferator–activated receptor γ (PPARγ) regulatory region,42 which may simultaneously affect transcriptional regulation. PPARγ is implicated in the regional transcriptional regulation of chr19q13.32 with the highest increase in expression observed for APOE messenger RNA (mRNA).42, 43 It plays a role in determining anti‐inflammatory macrophage (M2) phenotype,44 astrocyte inflammatory brain pathology,45 and airway and lung inflammation.46

Thus, polarization of the ε4‐ and ε2‐allele–related heterogeneous LD clusters differentiates cell types and implicates specific tissues in AD pathogenesis. These clusters can be a result of alteration in functional properties of complex regulatory networks in specific cell/tissue types linked with activation and function of immune cells (ie, pro‐[M1] and anti‐inflammatory [M2] macrophages) directed by the tissue‐specific micro‐environmental effects and other factors.47 Specifically, the detrimental ε4 allele LD cluster highlights the simultaneous effects of macrophage and astrocytes, whereas the protective ε2 allele LD cluster is implicated in non‐brain tissue. Our results support the idea that the effect of even the strongest genetic risk factor of AD, the APOE ε4 allele, can be naturally altered by changing the epigenetic landscape earlier in life by lifestyle and environmental interventions to decrease negative epigenetic changes in the APOE region and macrophage‐driven “inflamm‐aging.”48, 49 However, they indicate the critical role of heterogeneity and show that it can be informatively dissected as directed by molecular signatures of AD in the APOE region.

CONFLICTS OF INTEREST

The authors have no conflicts of interest related to this study.

Table S1. Basic information on SNPs and the associations of these SNPs with AD in the combined sample and each cohort separately.

Click here for additional data file.

Table S2. Linkage disequilibrium estimates using haplotype‐based method in the pooled sample of all cohorts.

Click here for additional data file.

Table S3. Linkage disequilibrium estimates using genotype‐based method in the pooled sample of all cohorts.

Click here for additional data file.

Table S4. Linkage disequilibrium estimates using haplotype‐based method in the pooled sample of all cohorts and each cohort separately for rs7412 and rs429358.

Click here for additional data file.

Table S5A. Linkage disequilibrium estimates using haplotype‐based method in the pooled sample of all cohorts for women.

Table S5B. Linkage disequilibrium estimates using haplotype‐based method in the pooled sample of all cohorts for men.

Click here for additional data file.

Table S6. Linkage disequilibrium estimates using haplotype‐based method in younger and older individuals.

Click here for additional data file.

Table S7. Functional annotation of potential regulatory SNPs in the APOE region.

Click here for additional data file.