The role of ABCA7 in Alzheimer’s disease: evidence from genomics, transcriptomics and methylomics

The ABCA7 gene consists of 47 exons and encodes a complete transmembrane transporter protein of 2146 amino acids with a molecular weight of 220 kDa [11]. As a typical ABC transporter, ABCA7 contains two intracellular nucleotide-binding domains (NBD) with conserved Walker A and B motifs for ATP hydrolysis, two transmembrane domains, and two extracellular loops (Fig. 1) [12]. A detailed ABCA7 structure (e.g. obtained through X-ray crystallography or cryogenic electron microscopy), however, is currently not available [RCSB PDB id: Q8IZY2, rcsb.org, accessed January 2019] [13]. Under physiological conditions, ABCA7 is expressed in brain tissue at a generally low abundance, which is comparable to other human tissues and cell types [Genotype-Tissue Expression (GTEx) project, accessed January 2019]. Two studies examining ABCA7 expression in mice brain, or human primary neurons from fetal brain, respectively, denoted the principal ABCA7 expressing cell type as neurons [14] or microglia [15]. More recently though, with the advent of large-scale single-cell sequencing projects, ABCA7 expression was observed to be relatively comparable across all different cell types in mouse brain [16]. A similar experimental set-up with human brain, unfortunately, did not detect sufficient ABCA7 RNA levels to determine cell type-specific expression [17].

In analogy to ABCA1—the closest homolog to ABCA7 and a more extensively studied protein—ABCA7 is predicted to regulate lipid metabolism. In vitro studies show that both ABCA1 and ABCA7 conduct apolipoprotein A (apoA)-I mediated export of cholesterol and phospholipids, but with different preferred substrates and efficiencies [19‐22]. Additionally, ABCA7 expression is upregulated when cholesterol is depleted via the sterol regulatory element-binding protein 2 (SREBP2) pathway, which has an inverse regulatory effect on ABCA1 [23]. Furthermore, in vivo Abca7 knockout mice have shown lower serum levels of cholesterol and high-density lipoprotein (HDL) [14], an altered phospholipid profile in the brain [24], and disruption of lipid rafts in thymocytes and antigen-presenting cells [25]. When analyzing Abca7 deficiency in mouse primary macrophages, however, no effect on cholesterol and phospholipid efflux was observed [14]. Hence, while a role of ABCA7 in lipid metabolism seems very likely, more research is needed to understand the specific effects across different tissues and cell types, and in relation to pathological effects.

In addition to lipid metabolism, ABCA7 has been linked to the regulation of phagocytosis. This connection was hypothesized based on the homology between ABC transporters (in particular the ABCA sub-family) and C. elegans cell corpse engulfment gene ced-7, which is involved in the engulfment of cell corpses during programmed cell death [26]. While both ABCA1 and ABCA7 have similarity to CED-7, only depletion of Abca7, and not Abca1, affected phagocytosis in mouse macrophages in vitro and in vivo [27]. Upon Abca7 deletion in mouse fibroblasts, phagocytosis was impaired [23]. Furthermore, Abca7 knockout in AD mouse models showed increased amyloid deposition in the brain and reduction of oligomeric uptake of both Aβ_1-40 and Aβ_1-42 in macrophages and microglia [28, 29]. ABCA7 presumably affects phagocytosis of Aβ aggregates rather than soluble Aβ, since microdialysis studies showed no difference in clearance rate of soluble Aβ [24]. In regard to the expression of ABCA7 across different cell types in the brain, as described above, it is important to note that in addition to microglia, also neurons, glial cells, and cerebrovascular cells are involved in Aβ uptake [30], and ABCA7 may play a role in them as well. Recently, Lamartinière et al. reported on a putative new role of ABCA7 as a regulator of cholesterol homeostasis and Aβ efflux at the blood–brain barrier [31]. Besides Aβ clearance, ABCA7 was also shown to directly affect Aβ production. Both in vitro and in vivo findings show an increase of Aβ upon suppression of ABCA7, presumably through β-secretase cleavage [32].

Finally, two in vivo studies assessed the role of ABCA7 on cognitive and behavioral aspects of AD. Logge et al. examined Abca7 knockout mice and observed impaired novel recognition memory in males and impaired spatial reference memory in females [33]. Impaired spatial memory was further corroborated in a cohort of male and female mice [24].

In the last decade, several GWAS [34‐40] were conducted on large cohorts of AD patients and controls, which introduced novel genetic risk loci for AD, including ABCA7. These studies provide AD-associated common single-nucleotide polymorphisms (SNPs), referred to as “sentinel” SNPs, which are presumed to be in linkage disequilibrium (LD) with a functional genetic variant. Because LD patterns and allele frequencies differ between populations, genetic association signals may differ between GWAS depending on the ethnicity of the cohorts. In Caucasians, three associated SNPs were pinpointed (Table 1, Fig. 1): intronic SNP rs3764650 (odds ratio (OR) = 1.2, Table 1) was the first sentinel SNP in the ABCA7 locus to obtain genome-wide significance [35] and is the most frequently studied variant in ABCA7 regarding the effect of common variants on (endophenotypes of) AD, as described below. A second Caucasian-based GWAS identified common missense variant rs3752246 (p.Gly1527Ala, OR = 1.2, Table 1) [38]. In silico analysis of rs3752246 suggests a benign effect (e.g. a Phred-scaled CADD score of 5.3, which ranks its deleteriousness at 30% of all genomic SNPs [41]). Finally, the largest peer-reviewed meta-analysis to date established intronic ABCA7 SNP rs4147929 (OR = 1.1, Table 1) as AD-associated sentinel SNP. SNPs rs3752246 and rs4147929 are in complete linkage disequilibrium (LD) in European populations (D′ = 1.0, R² = 0.97) [42]; hence, both tag the same functional variant. The LD between rs3764650 and rs4147929 (or rs3752246) is modest (D′ = 0.64, R² = 0.21) [42], suggesting either variable haplotype structure of the same functional genetic variant between Caucasian populations, or tagging of distinct functional genetic variants. In African Americans, ABCA7 intronic common variant rs115550680 was identified as AD-associated sentinel SNP. In contrast to Caucasian populations, rs115550680 (OR = 1.8, Table 1) had a much stronger effect size approaching that of APOE ε4 in African American populations [39]. Further research into the contribution of common ABCA7 variants to AD in other ethnicities has mostly been limited to these sentinel SNPs. Meta-analysis of rs3764650 in a mixed population comprising five Asian cohorts and a Turkish cohort failed to show association [43]. These findings are most likely explained by different haplotype structures between these populations and/or the relatively small cohorts used in these studies (1871 patients and 3406 controls). Larger studies and hypothesis-free analysis of common variants are needed to understand the role of ABCA7 in non-Caucasian and non-African American populations.

Ensuing identification of common AD-associated ABCA7 variants, several studies were performed to further understand the effects of these variants on biomarkers of AD, brain morphology, and symptomatic consequences (Table 2). Focusing on endophenotypes instead of clinical diagnosis results in more homogeneous phenotype definitions, which may lead to more accurate results. In comparison with the abovementioned large-scale GWAS, however, these follow-up endophenotype studies are conducted on smaller study populations at different stages of the disease (e.g. cognitively healthy elderly, individuals with mild cognitive impairment (MCI), AD patients, and post-mortem brain analyses), and results should be interpreted accordingly.

One of the major efforts to study the role of AD-associated sentinel SNPs on amyloid pathology was conducted in the frame of the Alzheimer’s Disease Neuroimaging Initiative (ADNI). Within ADNI, a continuously growing number of cognitively normal individuals, individuals with MCI, and AD patients are recruited and subjected to ¹⁸F-florbetapir positron-emission tomography (PET), which measures amyloid deposition in the brain of living subjects. In the first two ADNI studies, no significant association between ABCA7 GWAS SNPs and amyloid deposition was found [61, 9]. However, in the latest ADNI study comprising the largest cohort, the risk allele of ABCA7 sentinel SNP rs3752246 was significantly associated with increased amyloid deposition. Specifically, of all 20 AD risk SNPs under study, the effect size of ABCA7 was the highest after APOE. Further analysis revealed that the effect of ABCA7 on amyloidosis was significant in the cognitively healthy and MCI individuals, but not in the AD group, suggesting an early effect of ABCA7 [46]. These findings are in line with a Pittsburgh compound B PET study on cognitively normal individuals which reports association of rs3752246 and amyloid deposition [45]. In addition, post-mortem brain silver staining showed increased neuritic plaque formation for carriers of the rs3764650 risk allele [44]. Finally, one study found decreased amyloid β (Aβ) levels in CSF of non-demented Aβ positive subjects with an rs3764650 risk allele; no association was observed with tau CSF levels [47]. Hence, overall, carrying an ABCA7 risk allele increases amyloid deposition in the brain (Table 2). Currently, no evidence is available for a role of ABCA7 sentinel SNPs in (phosphorylated) tau homeostasis, although this was studied to a lesser extent than amyloid pathology. No studies have yet investigated molecular markers of other pathomechanisms through which ABCA7 could be involved in AD (e.g. lipid metabolism and immune-related endophenotypes) in the brain, CSF, and/or plasma.

Brain atrophy, which can be measured by magnetic resonance imaging (MRI), correlates with neuronal loss and as such has value in diagnosis of AD and monitoring of disease progression [62]. The correlation between AD GWAS SNPs and MRI measurements has been studied in several studies across different brain regions. Within the frame of ImaGene, 4071 non-demented individuals were subjected to hypothesis-free whole brain MRI analysis, and the volumetric results were correlated to genetic AD risk loci. Of all 19 genetic AD risk loci under study (including APOE), the strongest association—albeit not passing conservative multiple-testing thresholds—was obtained for the ABCA7 rs4147929 risk allele and decreased gray matter in the left postcentral gyrus [49]. This suggests an effect on brain morphology by ABCA7 in the asymptomatic stages of AD, yet further research reaching statistical significance is required. Wachinger et al. used MRI to study asymmetry in four pre-selected brain regions (hippocampus, amygdala, putamen, and caudate) and found significant association between rs4147929 and hippocampal asymmetry, particularly within the AD subgroup [52]. Additionally, studies with relatively smaller study populations reported associations between ABCA7 GWAS SNPs and decreased gray-matter density in dementia patients [50], increased cortical and hippocampal atrophy in controls and individuals with MCI [48], and dissociation of entorhinal cortex resting-state functional connectivity based on functional MRI in cognitively healthy elderly [51]. In summary, evidence is presently linking AD-associated ABCA7 common variants to increased brain atrophy already at preclinical stages of AD (Table 2). Further research to understand the consequences on the entire brain, however, is warranted.

Evidence for association between ABCA7 and cognitive decline has been inconsistent; which is in line with other genetic risk factors of AD apart from APOE. On the one hand, correlations between ABCA7 SNPs and clinical symptoms have been observed, yet based on relatively small cohorts, or with associations not passing multiple-testing correction. These include: association with baseline cognitive performance and linear rate of change [59], increased likelihood of clinical symptoms in individuals with AD neuropathology [60], later age at onset and shorter disease duration [53], and behavioral generalization deficits [51]. On the other hand, large-scale studies found no, or complex associations between ABCA7 sentinel SNPs and cognitive decline. For instance, the Three-City Dijon study, a longitudinal study comprising 4931 individuals older than 65 years of age found no association between ABCA7 common variants and global cognition, verbal fluency, visual memory, information processing, or literacy [63]. A second longitudinal study with 3267 female and 3026 male community-dwelling individuals found association between rs3764650 and cognitive decline, but only in females [57]. Third, in a study of 1153 middle-aged adults enriched for parental history of AD, again no association was found between ABCA7 sentinel SNPs and cognitive decline. However, a multiple-testing-corrected significant interaction was observed between APOE ε4 and both rs3764650 and rs3752246, which are not in strong LD. Strikingly, this interaction was consistent with a ‘flip-flop’ effect [64]: the effect on memory of ABCA7 genotypes is in opposite directions, depending on the APOE genotype [54]. Fourth, Carrasquillo et al. reported a significant association between the risk allele of rs3764650 and increased rates of memory decline in individuals with a final diagnosis of MCI or AD [56]. Hence, while overall the effects of ABCA7 GWAS SNPs on cognition seem to be minimal, these SNPs may alter cognition in subgroups stratified on gender, APOE-status, or disease progression (Table 2).

Two studies examined the effect of common risk loci of AD on the risk of developing posterior cortical atrophy (PCA), an atypical form AD, in which patients manifest with progressive decline in visuospatial, visuoperceptual, literacy, and praxis skills [65]. Both studies observed a significant association of the ABCA7 rs3764650 risk allele and increased risk for PCA (Table 2) [55, 58]. This risk effect did not differ from the risk to develop typical AD [58]. Additionally, Logue et al. assessed the shared common variant etiology between AD and age-related macular degeneration (AMD), the most common form of severe blindness and vision loss among the elderly. Within AMD, they found a weak association for ABCA7 AD sentinel SNP rs3752246, albeit with a different direction of effect. Stronger association with AMD was observed for another ABCA7 SNP (rs3752228), which was not in LD with any of the ABCA7 AD sentinel SNPs, hence suggesting a possible role of ABCA7 in AMD, but most likely through a different mechanism than AD [66]. Finally, a recent study examined the effect of ABCA7 sentinel SNP rs4147929 on vascular dementia, ischemic heart disease, ischemic cerebrovascular disease, and lipid levels in 104,258 individuals from the Danish general population, but did not observe an association with these endophenotypes [67].

GWAS associations were paramount in establishing ABCA7 as a risk gene for AD. Yet to understand which function of ABCA7 is compromised in AD, and to translate this into potential therapeutic targets, the downstream pathophysiological mechanisms should be elucidated. Sentinel SNP rs3752246 corresponds to a missense variant, but as described above, this amino acid change is not expected to affect the function of ABCA7 based on in silico prediction. Instead, a potential functional mechanism of common variants in ABCA7 could be the alteration of ABCA7 expression. In particular, a decrease of ABCA7 would be in line with the previously described AD-related effects observed in vitro and in vivo, as well as the effect of rare ABCA7 variants, which will be discussed below. According to the Genotype-Tissue Expression (GTEx) project [accessed December 2018], the minor alleles of rs3752246 and rs4147929 reduce the expression of ABCA7 in multiple tissues, including brain regions such as the cerebellum and the cerebral hemisphere. In addition, Vasquez et al. reported reduced ABCA7 expression in the anterior cingulate brain region for carriers of the rs3764650 risk allele in both AD patients and controls [68]. Allen et al., on the other hand, found no significant effect of rs3764650 on the expression in temporal cortex and cerebellum; assessed in a larger cohort of autopsied AD brains [69]. Hence, at the molecular level these sentinel SNPs do not provide sufficient insight into how genetic variation in ABCA7 contributes to AD. To resolve this issue, it is important to identify the underlying functional genetic variants, which are tagged by the sentinel SNPs. Furthermore, since these functional variants may not be in perfect LD with the sentinel SNPs, and this LD structure may differ between populations, it may be beneficial to use the functional variants rather than the indirect sentinel SNPs for future prognostic and diagnostic purposes.

First, while unlikely since each GWAS SNP tags many other variants in LD, the sentinel SNPs themselves could exert a pathological effect. One study modelled the rs3764650 (intron 13) and rs3752246 (exon 33) risk alleles in vitro. They identified possible promoter-enhancing capabilities for the rs3764650 major non-risk allele, and the rs3752246 risk allele led to increases in secreted Aβ40 and Aβ42, and β-secretase activity [70]. Nevertheless, since this study is based on the overexpression of ABCA7 and APP in human embryonic kidney cells (HEK293) or Chinese hamster ovary (CHO) cells, there is no guaranteed relevance of these findings in human brain tissue. In contrast to the functional sentinel SNP hypothesis, a study performed second-generation sequencing of the entire genomic ABCA7 locus, including introns and regulatory regions, to examine all genetic variations across the locus for association with AD. This study, performed on a Caucasian cohort, identified rs78117248 as the common SNP with the strongest AD association, comprising a twofold risk-increasing effect (Table 1). Furthermore, this variant was in very strong LD (D′ > 0.91) with all three Caucasian GWAS SNPs (rs3764650, rs4147929, and rs3752246), and conditional regression analysis showed that the sentinel SNP associations could be explained by rs78117248 [71]. This association was further replicated in an independent non-Hispanic White cohort [72]. While the genetic evidence is supportive of a role for this SNP in AD susceptibility, it is located deep intronic, and inspection of the local genomic context suggests that substitution of a single nucleotide at this position will have a minimal effect on protein function and expression. This suggested that another non-genotyped variation in ABCA7 in LD with this SNP would be the true functional variant.

Recent work points towards a repetitive DNA sequence in ABCA7 that was poorly covered in sequencing studies [73]. In intron 18, a variable number of tandem repeats (VNTR) was identified with highly variable lengths ranging from 300 bp to more than 10 kb. Carrying a risk allele of rs3764650 and rs78117248 was significantly correlated with longer VNTR lengths. Alleles longer than 5.6 kb were defined as expanded VNTRs, which had a 4.5-fold enrichment in AD (Table 1). VNTR alleles shorter than this cutoff did not seem to differ in length distribution between patients and controls. More than 7% of the AD patient cohort was observed to carry an expanded ABCA7 VNTR. Furthermore, based on ABCA7 transcript analysis in Epstein–Barr virus transformed lymphoblastoid cell lines of patients and controls, increasing VNTR length was associated with decreasing ABCA7 expression. In addition, alternative splicing of the VNTR-flanking exons (exon 18 and exon 19) was examined and increasing VNTR length was significantly correlated with an exon 19 skipping isoform, which results in loss of an ATP-binding domain of ABCA7. These ABCA7 transcription and splicing effects are potentially regulated through respective binding of transcription factors and splice factors to the VNTR core motif [73]. VNTR length showed an inverse correlation with Aβ_1-42 levels, corroborating the evidence from sentinel SNPs that common genetic variation in ABCA7 has an effect on Aβ. No significant associations were found in relation to total tau and phosphorylated tau in CSF of AD patients [73]. Further research is needed to delineate this VNTR’s contribution to AD in other populations. This is not trivial, due to the difficult detection of expanded tandem repeats. Yet, recent technological advances such as long-read sequencing can overcome these difficulties as demonstrated for the ABCA7 VNTR with the use of whole-genome long-read sequencing [74].

The strong ABCA7 GWAS signal observed in African Americans can be explained by another biological variant. By sequencing carriers of sentinel SNP rs115550680, a 44-bp exonic deletion (rs142076058, p.Arg578 fs) was identified which causes a frameshift in the ABCA7-coding sequence resulting in the formation of a premature termination codon (PTC) [75]. While the functional consequences of this deletion remain to be explored, the mutant transcript was detectable in blood of carriers, and a resultant truncated protein would lack both NBD domains, as well as nine predicted transmembrane domains. Follow-up of p.Arg578 fs in an African American case–control cohort showed a 1.8-fold enrichment (Table 1) in AD patients, with 15.2% of the patients carrying the deletion compared to 9.7% in controls [75]. The deletion was observed in all affected individuals of a large early-onset AD sibship of Caribbean Hispanic descent. An independent African American case–control study, however, could not replicate the association, with a carrier frequency of 9.2% in patients and 7.4% in controls. Differences in cohort ancestry between studies could be a potential explanation for this discrepancy [76].

A discovery that firmly confirmed ABCA7 as a risk gene for AD was the enrichment of rare variants in the gene in AD patients. Numerous independent studies reported this observation in 2015, including hypothesis-free studies such as a large-scale imputation study [77] and a small exome-array association study [78], as well as targeted gene resequencing studies [71, 79]. Particularly rare PTC variants, comprising frameshift variants, nonsense variants, and variants causing out-of-frame splicing, were highly enriched in AD [80, 81, 75, 71, 82, 83, 72, 76, 84‐86, 77, 79].

Thirteen published studies have examined the associations of ABCA7 PTC variants in AD–control cohorts [80, 81, 75, 71, 82, 83, 72, 76, 84‐86, 77, 79]. All, but one, observed a higher frequency of heterozygous PTC variants in patients than controls; yet a large range of ORs is reported varying from approximately 1.4 [76] to 5 [85] in studies reporting an enrichment in patients. This relatively low 1.4 OR was reported in a study conducted on an African American cohort [76], and was mostly driven by a substantially weaker risk effect for common African American PTC p.Arg578 fs than its originally reported effect size (Table 1). Meta-analysis of six Caucasian studies that conducted a comprehensive NGS-based analysis of the entire ABCA7 coding sequence on cohorts including > 900 individuals indicates a 2.6-fold PTC enrichment in AD patients compared to controls (Fig. 2, Table 1).

Cautious interpretation of these Caucasian risk estimates is warranted though since c.5570 + 5G > C (rs200538373), the most frequent Caucasian PTC variant, was not always included in calculations of study-specific effect estimates. This variant does not directly perturb the canonical sequence of a splice site, but causes the preferential usage of a cryptic splice donor site with a frameshift as a consequence [85, 77]. It is unclear whether this alternative splicing always occurs on the mutated allele, or if splicing is only partially affected. An in vitro assessment of c.5570 + 5G > C supports a consistent splicing effect of c.5570 + 5G > C [89]; whether this extrapolates to different human cell types needs to be verified.

According to the Genome Aggregation Database (gnomAD, accessed January 2019) [90], the carrier frequency of ABCA7 PTC variants (including c.5570 + 5G > C) in the general European population is 2.1%, while most hypothesis-free studies report between 3.0 and 4.4% of AD patients carrying a PTC variant, with differences most likely driven by founder effects [81, 71, 83, 85, 77]. In African American individuals, a staggering 14.5% carries a PTC variant. This cumulative frequency further increases to 21.7% in AD patients [76]. This high PTC mutational frequency is primarily due to the common PTC variant p.Arg578 fs which is proposed to explain the African American GWAS signal at ABCA7 (Table 1). With the exception of variants in canonical splice sites, c.5570 + 5G > C, and the ABCA7 VNTR, no other ABCA7 variants have been reported to affect splicing, but we note a lack of comprehensive studies, suggesting these carrier frequencies may be underestimated. Importantly, the observation of ABCA7 PTC variants in control individuals and the high frequency of certain variants among African Americans indicates that ABCA7 PTC variants per se are not sufficient to cause disease, but rather risk variants of intermediate to high penetrance. Identification of factors that modify the penetrance of these variants will be pivotal in understanding how ABCA7 PTC variants affect AD risk.

PTC variants can lead to the formation of truncated ABCA7 proteins, which, in theory, can underlie the AD risk-increasing effect by several mechanisms. First of all, truncated proteins may cause toxicity in the cell through a gain of function (GOF) effect. In this case, pathogenic mutations are expected to cluster in parts of the gene, affecting particular protein domains. Second, truncated proteins may present a dominant negative effect: an abolition of the function of proteins binding to the mutated protein; often observed in multimeric protein complexes. Third, the truncated proteins may no longer exert an effect, or can be subject to quick degradation, which leads to a loss-of-function (LOF). To avoid the first two deleterious scenarios, cells have a mechanism termed nonsense-mediated mRNA decay (NMD) to track down transcripts containing a PTC which are subsequently degraded prior to the formation of truncated proteins. Intervention through NMD results in LOF. In general, loss of one gene copy is tolerated; however, for some dosage-sensitive genes, losing half of the expression may cause disease, which is termed haploinsufficiency [91]. In the case of ABCA7 in AD, heterozygous PTC variants are observed across the entire transcript without an apparent pathological enrichment to certain protein domains (Fig. 1), making a GOF effect to be an unlikely mechanism for all PTC variants. Furthermore, the ABCA7 protein channel is formed as a monomer [11], reducing the likelihood of a dominant negative effect. As such, ABCA7 haploinsufficiency is the most plausible mechanism leading to increased AD risk. This mechanism would be in line with the previously discussed in vitro and in vivo evidence, and the effect of common ABCA7 variants (in particular the ABCA7 VNTR). Further research into this hypothesis, by studying ABCA7 expression, however, has been limited and has proven to be difficult. Currently, we do not know which isoforms of the ABCA7 protein form a functional entity, nor do we know what the physiological levels of functional ABCA7 are in different cell types and/or tissues. Large variance is observed in brain ABCA7 mRNA and protein expression, both between PTC variant carriers and non-carriers, and within the group of PTC variant carriers [80, 71]. Further allele-specific analyses of these transcripts reveal that incomplete NMD (or NMD escape) of ABCA7 transcripts harboring a PTC is common for most PTC variants and the abundance of these PTC transcripts can vary from 5% (strong mRNA degradation through NMD) to approximately 50% (almost no NMD degradation) [80, 75, 85]. Interestingly, long-read phasing of ABCA7 cDNA revealed that for several PTC variants, the induced PTC could be removed from the transcript by alternative splicing (i.e. restoration of the reading frame by usage of a cryptic splice site, or removal of a nonsense variant through in-frame exon skipping). These “PTC rescue mechanisms” have the potential to increase the abundance of functional ABCA7 protein [85]. Varying NMD efficiency and PTC rescue splicing, however, explains only a fraction of ABCA7 expression differences, and the effect on penetrance and clinical presentation remains to be explored. Further research into different ABCA7 isoforms, the upstream factors regulating ABCA7 expression, and correction for brain tissue composition (i.e. correction for the influx and activation of microglia) is necessary to understand when ABCA7 haploinsufficiency kicks in and how this increases the risk for AD.

Carrying an ABCA7 PTC variant, in general, more than doubles the risk for developing AD, with local variation due to founder effects. Two reports show an increased proportion of positive familial history for AD patients with an ABCA7 PTC compared to the overall patient cohort from which they originated [92, 85]. While familial clustering is in line with an increased penetrance of these variants, observation of segregation patterns mimicking autosomal dominant inheritance is exception rather than rule. Only four pedigrees have been reported so far; moreover, co-segregation in these pedigrees was neither significant nor clear for all carriers (Fig. 3) [75, 71, 72, 93].

Two studies examined whether ABCA7 PTC variants could also affect the risk of other neurodegenerative conditions by comparing healthy controls to, respectively, 840 Parkinson disease (PD) patients [94] and a cohort of 381 brain autopsied non-AD dementia patients (enriched for vascular dementia, frontotemporal dementia, dementia with Lewy bodies, and progressive supranuclear palsy) [80]. A strong enrichment (OR = 4.9) of ABCA7 PTC variants was observed in PD, albeit not reaching significance (p = 0.07) [94]. For the non-AD dementia cohort, a significant enrichment was observed [OR = 3.1 (1.2–7.7)] and this effect size was similar when compared to the enrichment of AD patients and controls [80]. Hence, these reports suggest that ABCA7 PTC variants may play a general role in neurodegeneration, but more research with meticulously diagnosed patient cohorts is necessary.

Since AD GWAS were mainly based on LOAD patients, ABCA7 was originally considered a LOAD risk gene. Studies investigating ABCA7 PTC variants on the other hand have included a substantial number of EOAD patients and found an enrichment of PTC variants in these individuals as well [81, 92, 75, 83, 85]. While Bellenguez et al. observed slightly increased enrichment of ABCA7 PTC variants in EOAD, compared to LOAD [81], other studies did not observe an effect of ABCA7 PTC variants on onset age [75, 71]. Overall, ABCA7 PTC variant carriers have a very wide onset age range, from 46 [83] to 90 years old [71], even within carriers of the same variant. While an onset age of 46 is very young for AD, onset ages for carriers of pathogenic mutations in causal genes, such as PSEN1, are generally lower [85]. The wide variation in onset ages is to be expected for (rare) risk variants.

Only few studies so far have investigated clinical characteristics of ABCA7 PTC variant carriers other than onset age. In a series of 22 AD patients carrying an ABCA7 PTC variant, all presented with a classical AD phenotype, with progressive memory impairment [92]. Neuropathological assessment of autopsied brains of PTC variant carriers was in accordance with typical AD [92, 85]. Additional research into the downstream phenotypic effects of PTC variants has been scarce. Only one study assessed a possible relation with amyloid deposition: Martiskainen et al. examined the effect of the c.5570 + 5G > C (rs200538373) risk allele on Aβ₄₀ and Aβ₄₂ levels in blood plasma of healthy individuals, but no associations were observed [95]. Nevertheless, the correlation between plasma amyloid levels and AD has been contradictory, and in general, only marginal differences are observed between patients and controls [96]. Further investigation of relevant biomarkers is, therefore, necessary.

Analogous to PTC variants in ABCA7, other variant categories most likely influence the risk to develop AD. In particular, missense variants have the potential to destroy important protein domains, or they can interfere with protein stability and degradation; both of which can lead to a LOF effect. One study reported the co-segregation of missense variant p.Arg880Gln (rs143718918) with autosomal dominant AD (Fig. 3) [93]. Based on in silico prediction, this variant has a high chance to disrupt ABCA7 function. With a Phred-scaled CADD score of 28.6, p.Arg880Gln is considered to be among the 0.14% most deleterious variants of the genome [41]. Further analysis of ABCA7 missense variants has been limited, and no meta-analysis exists to date. Nevertheless, several NGS studies have reported the observed frequencies of missense variants in AD patients and controls [81, 71, 83, 85, 86, 77], albeit using different criteria to determine which variants are to be reported, potentially biasing downstream analyses. While all studies focus on rare variants, some studies only report variants with in silico-predicted deleterious effects on the ABCA7 protein (Fig. 2). Meta-analysis of all published reports on missense variants in ABCA7 suggests this class of variants also confers a 1.8-fold enrichment in AD patients (Table 1); however, the validity of these effect sizes has to be confirmed with additional research. On the one hand, they could have been artificially augmented due to different missense selection criteria across the studies. On the other hand, however, we may be underestimating the effect of missense variants on AD, since the direction of effect for all missense variants under study is considered to be equal. In reality, it is most likely that while some missense variants may be deleterious, others will have no effect, or may even protect from disease. Rare variant association testing can take these different effect directions into account (e.g. Sequence Kernel Association Test (SKAT) [97]), though they should be conducted on variants which are not selected based on predicted pathogenicity, and the underlying variants need to be reported with carrier frequencies for interpretation. Better understanding of the function and structure of ABCA7, as well as experimental assays to evaluate the effect of variants, are needed to enable classification of missense variants. A bidirectional effect is supported by common ABCA7 missense variants: two studies observed significant protective effects for p.Gly215Ser (rs72973581, Table 1) [85, 86]. This residue is situated in the extracellular domain of ABCA7. Interestingly, the minor and protective serine amino acid is evolutionary conserved in several mammals, as well as the homologous human ABCA1 protein, which favors a functional effect of this missense variant.

In addition to genetic variations, epigenetic modifications of DNA and DNA-bound histones can alter gene expression. One of the best-known and most studied modifications is methylation of cytosine nucleotides at position 5 of the pyrimidine ring. This 5-methylcytosine occurs most prominently in a cytosine:guanine (CpG) sequence. These motifs often cluster together in so-called CpG islands, and regions flanking these islands are termed CpG island shores. Historically, DNA methylation was mostly associated with gene expression silencing, yet increasing research shows that, depending on the genomic context, the effects of methylation are highly variable [98].

The role of methylation in AD was assessed at genome-wide scale within the frame of the Religious Order Study (ROS) and the Memory and Aging Project (MAP). DNA CpG methylation was examined in more than 700 gray-matter brain tissues from the dorsolateral prefrontal cortex. Of the 415,848 examined CpG dinucleotides, 71 reached genome-wide significance; two of which corresponded to AD GWAS loci: BIN1 and ABCA7 [99]. In-depth analysis of the ABCA7 locus showed that three CpG sites (cg02308560, cg24402332, and cg04587220) associated with AD after correction for multiple testing (Fig. 1, Table 1). Methylation of each of these CpG dinucleotides correlated with a higher risk of developing AD [99, 100]. Additionally, 12 CpG sites in ABCA7 correlated with Aβ load and 18 CpG sites correlated with tau tangle density [100]. Across all of these endophenotypes, cg02308560 and cg04587220 generally had the strongest associations, which were independent from ABCA7 AD GWAS sentinel SNP genotypes [101, 99]. They are only 30 base pairs apart from each other and located approximately 5.6 kb downstream of the ABCA7 gene, embedded in a CpG island which is part of a Polycomb repressed region situated in an intron of the HMHA1 gene [100]. DNA methylation is an important regulator of Polycomb-mediated repression, which can further regulate the expression of nearby genes, such as ABCA7 [102]. Overall, CpG methylation in ABCA7, however, did not seem to have strong effects on brain ABCA7 mRNA expression [100].

In addition, Humphries et al. examined methylation in a smaller series of brain tissue samples from the temporal pole. For all 20 AD GWAS loci under study, the strongest effects were observed in the ABCA7 locus. More specifically, a CpG island shore, located in exon 12 and intron 12 of ABCA7 (chr19:1045074–1045679, Fig. 1), was hypomethylated in LOAD when compared to healthy control brains, and brains of patients with dementia with Lewy bodies (Table 1) [103]. Due to usage of different methylation detection techniques, this CpG island shore was not assessed in the previously discussed methylation studies.

Not all AD-related methylation studies, however, identified significant changes in the ABCA7 locus. Lunnon et al. performed a (smaller sample size) genome-wide methylation assessment using the same methylation array as described above, but in different brain regions (entorhinal cortex, cerebellum, superior temporal gyrus, and prefrontal cortex) and blood. None of the top 100 associated methylation markers across these tissues was positioned in the ABCA7 locus [104]. Additionally, Yamazaki et al. analyzed methylation of five CpG sites in the promotor region of ABCA7 in peripheral blood samples, but found no differential methylation between patients and controls [105]. Hence, while several independent studies have reported a role for CpG methylation of ABCA7 in AD risk, the mode of action remains to be elucidated. More research, accounting for factors such as brain regions and tissue composition, could provide more clarity into the downstream effects of ABCA7 methylation.

In addition to previously discussed DNA-based findings, alterations on the ABCA7 mRNA level have been linked to AD as well. Based on cDNA cloning and northern blotting, early literature on ABCA7 described two isoforms of ABCA7: full-length ABCA7 (“type I”, RefSeq ID: NM_019112, GENCODE ID: ENST00000263094.10_1) and a splice variant (“type II”). The type II isoform originates from alternative splicing of the ABCA7 pre-mRNA which does not remove intron 6 from the coding sequence. This introduces a PTC, which is then followed by a novel in-frame start codon. The coding sequence initiating at this novel start codon and containing the 41 downstream exons corresponds to type II ABCA7 (resembles GENCODE ID ENST00000435683.6_1) [106]. The abundance of both isoforms was highly variable between human tissues. In vitro experiments further showed that both isoforms translate into proteins with different cellular localization. Type I resided both on the plasma membrane and in intracellular compartments. Type II on the other hand was predominantly localized in the endoplasmic reticulum. Finally, in contrast to type I, type II did not show efficient apoA-I-mediated lipid release [106].

As a consequence of large-scale RNA-sequencing studies and public transcript repositories such as GENCODE [107], we now know that the transcriptional landscape of ABCA7 is far more complex. According to GENCODE version 28 [Accessed January 2019], 19 different transcripts are allocated to the ABCA7 gene. Nevertheless, it is unknown which of these are translated into proteins and what their functional relevance is. These known transcripts, however, are only the tip of the iceberg. Two reports analyzed splicing patterns of a dozen of ABCA7 exons in depth and identified alternative splicing correlated with ABCA7 VNTR lengths, and found splicing events which could potentially rescue a deleterious PTC variant. Incidentally, naturally occurring alternative splicing was observed in virtually all exons [85, 73], which gives rise to a myriad of potential transcripts. Large deviations from the full-length ABCA7 transcript are presumably detrimental to its functions, since the bulk of the ABCA7 transcript is essential to form an ABCA7 protein with a transmembrane pore and ATP-hydrolyzing NBD domains (Fig. 1). Yet, more research is needed to support this hypothesis and to understand the consequences of smaller transcript alterations.

Due to the relatively low expression of ABCA7, and consequently low sequencing read counts in RNA-sequencing experiments, it is not trivial to study alternative splicing in ABCA7. Nevertheless, in the largest AD meta-analyzed transcriptome-wide association study (TWAS) to date, significant association was found between ABCA7 splicing—measured in the dorsolateral prefrontal cortex—and AD [108]. These findings were further corroborated by another study analyzing RNA from the temporal lobe brain region, in which differences in ABCA7 splicing between LOAD and controls were observed [103]. The strongest TWAS alternative splicing signal mapped to the 3′ end of the ABCA7 transcript, which is most likely the result of technical limitations. The algorithms used for characterization of alternative splicing require a minimum of supporting sequencing reads, and RNA-sequencing coverage is often enriched in the 3′ transcript regions. Hence, due to the low expression of ABCA7, this alternative splicing analysis was probably most powerful in the 3′ end [109]. Further research with higher sequencing depth, or enrichment of ABCA7 transcripts could provide more clarity.

Interestingly, while evidence is limited compared to neurodegenerative diseases, ABCA7 has also been linked to cancer. ABCA7 expression is increased in ovarian cancer tissues [110, 111], and this upregulation has been shown to accelerate the epithelial-to-mesenchymal transition, a process involved in the initiation of metastasis. In vitro knockdown of ABCA7 in ovarian cancer cells reversed markers of epithelial-to-mesenchymal transition, suggesting that ABCA7 is actively involved in this process. Furthermore, ovarian cancer patients with high ABCA7 mRNA expression levels, have a poorer survival rate than patients with low ABCA7 expression [111]. In addition to ovarian cancer, increased expression of ABCA7 was observed in pancreatic ductal adenocarcinoma tumors [112]. Melanoma cell lines on the other hand showed lower expression levels in comparison with melanocytes [113].

Hence, while more research is needed, it seems that an increase of ABCA7 expression may lead to a higher risk of specific cancer-type development, which is opposite to ABCA7 loss-of-function as a pathomechanism of AD. This observation is in line with several epidemiological reports observing an inverse comorbidity between central nervous system (CNS) disorders and cancer [114]. Biological evidence further supports this finding; hence, this is most likely a true association rather than an indirect effect caused by a survival bias. Notably, inverse up- and downregulation of genes in CNS disorders and cancer have been observed [115], and several genetic loci have shown opposite effects in AD and cancer GWAS studies [116]. Further research into ABCA7 inverse comorbidities is warranted to understand the implications of ABCA7 expression as a possible therapeutic target for dementia.

In the past 8 years, human-based studies have demonstrated numerous ABCA7-related associations with AD; including four GWAS SNPs, 52 PTC variants, rare missense variants, a protective common missense variant, a tandem repeat, CpG methylation, and alternative splicing. As such, compelling evidence supports the involvement of ABCA7 in AD pathophysiology (Fig. 4). Combined, these ABCA7 alterations contribute substantially to the occurrence of both EOAD and LOAD patients (e.g. the heritability of liability attributable to expanded ABCA7 VNTRs and PTC variants in the Caucasian cohort is approximately 4.5% [117]) Furthermore, in specific cases, PTC variants have been observed to achieve high disease penetrance, leading to inheritance patterns resembling autosomal dominancy. Nevertheless, cautious interpretation is needed regarding the effects of ABCA7 alterations on the phenotype. Overall, variable disease penetrance and onset ages are observed, and it is currently not known whether ABCA7 variants are AD specific or if they contribute to other dementia phenotypes as well. Hence, a better understanding of ABCA7 modifiers is warranted. An important avenue of investigation in this regard, and potentially the biggest caveat in ABCA7 research to date, is the expression of (different isoforms of) ABCA7. AD-associated markers in ABCA7, as well as in vitro and in vivo experiments, point towards a decrease of ABCA7 as the main pathological mechanism. As a consequence, varying ABCA7 expression levels could modify the phenotype. Moreover, targeting ABCA7 transcription or splicing to increase functional ABCA7 expression levels could be a valuable therapeutic approach. However, based on the Accelerating Medicines Partnership—Alzheimer’s Disease (AMP-AD) consortium which includes the currently largest collaborative post-mortem brain RNA-sequencing project, no significant differential expression of ABCA7 is observed between patients and controls [https://​agora.​ampadportal.​org; accessed January 2019]. This is further corroborated by a recent meta-analysis of AD brain transcriptomics data [118].

Several hypotheses could underlie this situation. First, perturbation of ABCA7 expression is potentially not a canonical pathological pathway of AD, and therefore, not observed in a (underpowered) heterogeneous population of AD patients. In this scenario, ABCA7 haploinsufficiency could give rise to a molecular subtype of AD, with potentially different clinical and therapeutic characteristics. Hence, patient stratification based on ABCA7 variant status could be advantageous, both to detect decreased ABCA7 expression and for downstream investigations. Second, ABCA7 downregulation due to deleterious missense variants (e.g. by causing misfolding and/or increased protein degradation) are not observed on a transcript level. ABCA7 mRNA downregulation is also partially masked for PTC variants because of NMD escape and PTC rescue splicing. Additionally, differences in splicing (e.g. as observed with different ABCA7 VNTR lengths) are generally not detected in conventional ABCA7 quantification. Third, decreased ABCA7 may contribute to AD pathology in the early stages of disease; however, it is possible that at later stages (which are measured in post-mortem brain), ABCA7 is upregulated as a compensatory mechanism in AD brains; e.g. to increase phagocytic clearance of amyloid plaques. Last, ABCA7 expression is likely to differ between cell (sub-) types and cellular states (e.g. ramified versus activated microglia), and this cell type composition in turn varies between patient and control brains which biases overall expression analyses. Additionally, it is not completely understood in which cell type(s) and brain regions ABCA7 exerts its (AD-related) functions.

Hence, to provide an answer to these hypotheses, there is a need for: (1) a better understanding of the function of ABCA7 and in which cell types, brain regions, and disease stages it plays a role; (2) functional studies and assays which can differentiate deleterious variants and splicing events from functional ABCA7 forms; (3) more sensitive assays for both the mRNA and protein level which are capable of detecting and quantifying different isoforms of the lowly expressed ABCA7 gene. Recent developments in the sequencing field can provide solutions. On the one hand, single-cell sequencing can give novel insights into the cell- and tissue-specificity of ABCA7. While not always trivial for (post-mortem) brain tissue, adaptations to this technique such as single nuclei sequencing [119] and spatial transcriptomics [120] can provide an alternative. On the other hand, long-read sequencing techniques can sequence entire mRNA molecules from start to end and, therefore, provide a better understanding and quantification of all possible ABCA7 transcripts [121, 122]. For both single-cell and long-read sequencing, high sequencing depth is necessary to detect the relatively lowly expressed ABCA7 transcripts, or specific capture of ABCA7 mRNA prior to sequencing needs to occur. Other developments which may benefit our understanding of ABCA7 include better in vitro models based on induced pluripotent stem cells (iPSCs), which can be differentiated into neuronal and glial cell types of interest [123]. These models can be extended to three-dimensional organoids to encompass the interplay between different brain cells [124]. Furthermore, improvements in the field of cryogenic electron microscopy have recently allowed to determine the protein structure of ABCA1 [125]. Hence, a similar approach should be feasible for ABCA7, which would enable improved in silico predictions of ABCA7 functions and effects of variants and splicing.

Currently, to determine methylation, or genotype sentinel SNPs, rare variants, and tandem repeats, different technologies are used (e.g. bisulfite sequencing or methylation arrays, SNP arrays, NGS and Southern blotting). Consequently, these markers are often studied independently, leading to a missed understanding of their combined effects on AD. One way to resolve this issue would be the usage of long-read sequencing, which enables simultaneous detection of SNPs, structural variants such as tandem repeats, and DNA methylation. Recently, we showed that robust whole-genome long-read sequencing is possible on PromethION (Oxford Nanopore Technologies), and accurate detection of ABCA7 VNTR alleles, including expansions, was achieved [74]. With continuous improvements to single-nucleotide accuracy, these whole-genome long-read sequencing techniques will hopefully soon enable simultaneous detection of all DNA-related markers in ABCA7 and other AD loci, with clinical-grade quality [126]. In addition, this would enable discovery of other AD-associated structural variants and epigenetic markers, and would facilitate easier transition from research to ABCA7 genetic screening in the clinic.