Fine-mapping of the human leukocyte antigen locus as a risk factor for Alzheimer disease: A case–control study

Summary

In a case–control study, Jennifer Yokoyama and colleagues present fine-mapping of the human leukocyte antigen genetic region to identify haplotypes associated with Alzheimer disease.

Introduction

Alzheimer disease (AD) is a progressive neurodegenerative disorder and has a global burden of approximately 46 million people worldwide, with prevalence projected to double over the next 20 y [1]. The hallmark features of the disease include the accumulation of amyloid plaques, tau neurofibrillary tangles, and neuronal destruction, leading to brain atrophy and loss of cognitive function. The etiology of these processes stems from synergistic interactions of environmental and genetic factors, many of which remain obscure and therefore complicate research efforts aimed at identifying efficacious therapies.

The three largest genetic contributors identified thus far are rare variants in amyloid precursor protein (APP) and presenilin 1 and 2 (PSEN1, PSEN2) [2]. These variants are uncommon, cause an early onset form of the disease, and typically segregate in an autosomal dominant fashion. Studies of late onset AD (typically defined as onset age >65 y) have demonstrated the risk of sequence variants such as the common ε4 allele of apolipoprotein E (APOE), rare variation in TREM2 [3–7], and MAPT [8], as well as numerous common variants contributing modest AD risk [9], including single nucleotide polymorphisms (SNPs) in the following loci: CR1, BIN1, INPP5D, MEF2C, CD2AP, ZCWPW1, NME8, EPHA1, CLU, PICALM, MS4A4, CELF1, FERMT2, ABCA7, CD33, CASS4, PTK2B, SORL1, SLC24A4-RIN3, DSG2, and HLA-DRB5/HLA-DRB1 [9,10]. However, there remain additional unexplained genetic contributions to nonfamilial forms of AD, suggesting polygenic contributors as well as the potential for epistatic and epigenetic interactions [11].

There is increasing support for the role of neuroinflammation in the etiology of AD as well as evidence that inflammatory processes are an early event in the brains of patients with AD [12]. Several studies have provided biochemical and histological evidence of classic immune components, including active microglia [13–15], complement factors [16,17], inflammatory cytokines [18], and C-reactive protein [19] within the parenchyma of AD brains. This is further supported by work in mouse models providing strong evidence for the role of complement-dependent destruction of synapses by phagocytic microglia prior to plaque deposition; similar mechanisms may even contribute to age-related cognitive decline [20,21]. Given these findings, there is a great deal of interest in identifying genetic determinants of inflammation related to AD susceptibility.

Located on chromosome 6p21, the major histocompatibility complex (MHC) is a dense region of approximately 150 genes that encode the human leukocyte antigen (HLA) immunoregulatory proteins [22]. Because of their proximity to each other, many of the MHC genes exist in linkage disequilibrium (LD) and are inherited as haplotypes with varying frequencies in global populations. MHC genes encode cell surface receptors and are classified based on their ability to present endogenous or exogenous antigens to T cells. MHC class I proteins exist on the surface of all nucleated cells and present fragments of antigens generated intracellularly to CD8+ T cells to induce a cytokine-mediated immune response. MHC class II molecules are only expressed by professional antigen-presenting cells, including B cells, macrophages, and microglia, and present exogenous material taken into the cell via endocytic vesicles to CD4+ T cells. Together, the diverse repertoire of the human immune system partly stems from the extremely polymorphic nature of the MHC class I and II regions.

Many associations are established between neurodegenerative and autoimmune diseases, specific class I and II alleles, and combinations of alleles (haplotypes) in the HLA region. Previous genome-wide association studies (GWASs), pleiotropic analyses, and meta-analyses by our group and others have investigated MHC susceptibility loci in a wide range of diseases, including AD [9,23,24]. However, because of the complex genetic organization of the HLA region and differences in the haplotype substructure of different ethnic populations, as well as differences in sequencing and allelic imputation methods, studies have yet to definitively elucidate which genes and specific alleles contribute to the observed association signals.

As mentioned, HLA-DRB5/HLA-DRB1 has been implicated in numerous GWASs as a significant contributor to AD risk [9]. This prior work has established a significant association of the HLA locus to AD risk in over 75,000 individuals, yet the specific allele or alleles contributing to this association remain elusive. We thus used a robust HLA imputation method and case–control approach to fine-map the contributions of HLA polymorphisms and haplotypes to AD in over 11,500 patients and controls from independent cohorts from the University of California, San Francisco (UCSF) Memory and Aging Center (MAC) and the Alzheimer’s Disease Genetics Consortium (ADGC) (S1 File). We also examined longitudinal neuropsychological measures of cognitive function and cross-sectional biomarker data from cerebrospinal fluid (CSF) from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) to assess the clinical relevance of identified risk haplotypes.

Materials and methods

Participants were consented (as described below) for research in accordance with the Institutional Review Board at the University of California, San Francisco, and institutional review boards at each site for multicenter study data approved all aspects of this study as they fall under the purview of the respective research groups (ADNI and ADGC).

Participants

UCSF MAC cohort

The participants included in this study were 309 white individuals over the age of 50 y, including 191 controls and 118 individuals with AD seen at the UCSF MAC between 1999–2012 who were genotyped as part of their participation in longitudinal research on neurodegenerative disease and healthy cognitive aging. DNA from the UCSF MAC cohort was collected from 2000–2012, and genotyping was performed in 2012. Because individuals are followed up longitudinally, we verified clinical diagnosis at the beginning of this study (May 2015). A multidisciplinary team of neurologists, neuropsychologists, and nurses performed a detailed evaluation on individuals with AD and established a diagnosis according to consensus criteria for AD [25]. Individuals included as controls underwent a similar assessment and were diagnosed as having normal cognition for their age. Participants who carried a known genetic risk variant in APP, PSEN1, or PSEN2 were excluded from this study. Participants or surrogates completed written informed consent for all genetic research related to neurodegenerative disease and healthy cognitive aging during their initial visit in accordance with the Institutional Review Board at the University of California, San Francisco.

ADGC

The ADGC is an NIH-funded collection of GWAS data created for the goal of identifying genetic contributions to late-onset AD. Participants included in this study were from 30 merged datasets combined by Boehme, Mukherjee, Crane, and Kauwe and included 28,730 individuals carrying either an AD or cognitively normal control clinical diagnosis [26] A list of the datasets and basic information is included in S1 Table; full details on the datasets and the merging process are available at http://kauwelab.byu.edu/Portals/22/adgc_combined_1000G_12032014.pdf [26]. Analyses were limited to white individuals for maximum statistical power to reduce potential for confounding due to the known population-based contribution to diversity in the HLA region. Participants were recruited and seen between 1984–2012. Written informed consent for genetic studies falling under the purview of the ADGC was obtained from all study participants, and institutional review boards at each site approved all aspects of this study. Specific consent for this study was obtained from the ADGC based on an application describing the proposed work.

ADNI

We also utilized data from 346 individuals recruited for participation in the ADNI study with data from SNP genotyping and longitudinal cognitive scores. All individuals included in this study had a minimum of two clinic assessments. At baseline, 120 individuals were cognitively normal (CN) older adults, 113 individuals were diagnosed with mild cognitive impairment (MCI), and 113 with AD. Of these, 163 individuals also had CSF measurements of plasma biomarkers available (S2 Table). The ADNI cohort is well characterized and has been used in previously published studies [27–29]. The clinical severity of symptoms in the MCI and AD groupings was measured using the Clinical Dementia Rating sum of boxes (CDR-SB) [30]. A clinician diagnosed each participant using a structured protocol that utilized clinical judgment and neuropsychological tests that are provided in S1 Methods. The mean follow-up time was 3.15 ± 2.04 y for control participants (n = 91), 2.39 ± 1.71 y for participants with MCI (n = 148), and 1.37 ± 0.75 y for patients with AD (n = 69). Written informed consent was obtained from all study participants for research studies falling under the purview of ADNI, and the University of California, San Francisco Institutional Review Board approved all aspects of this study.

Genotype acquisition

UCSF MAC cohort

Patient and control genotypes were obtained via genotyping on the Illumina Omni1-Quad array (Illumina, San Diego, California) using manufacturer’s instructions. APOE genotype was determined with a TaqMan Allelic Discrimination Assay for the two SNPs, rs429358 and rs7412, on an ABI 7900HT Fast Real-Time PCR system (Applied Biosystems, Foster City, California) using the manufacturer's instructions.

ADGC

Details of genotyping in the 30 datasets that comprise the combined ADGC dataset are available online [26] and partially described in previously published papers [10].

ADNI

Haplotypes were determined using genotypes from the Human610-Quad BeadChip (Illumina, San Diego, California) as previously described [31]. APOE genotypes were determined by Cogenics (now Beckman Coulter; Pasadena, California).

CSF biomarker measurements

ADNI

Baseline CSF biomarkers levels were measured using the Human DiscoveryMAP panel developed by Rules Based Medicine (Myriad RBM; Austin, Texas). The Human DiscoveryMAP panel is commercially available and measures a collection of metabolic, lipid, inflammatory, and other AD-relevant indicators. We limited our analyses to 28 immune proteins in the panel that were associated with inflammatory or immune processes (S1 List). The samples were processed and analyzed by Myriad RBM and checked for quality by the ADNI Biomarker core. CSF amyloid β 1–42 was measured using the AlzBio3 Luminex xMAP immunoassay (Innogenetics, Ghent, Belgium) according to previously described methods [32]. This method utilizes a monoclonal antibody specific for amyloid β 1–42 that is chemically bonded to color-coded beads along with analyte-specific detector antibodies. Additional details are available in S1 Methods.

Clinical assessments

ADNI

In this study, we analyzed two neuropsychological measures of cognitive function and one measure of clinical severity in ADNI participants. The Rey Auditory Verbal Learning Test (RAVLT) [33] is a test of verbal memory. It begins with the administrator reading a list of 15 unrelated words to the participant, who is then asked to verbally repeat as many of the words as they can. This happens for a total of five learning trials, and the administrator records the number of words correctly recalled after each trial. The test administrator then reads a set of 15 new words to the participant (interference word list), and, immediately following this, the participant is asked to recall as many of the first list of words as possible (immediate recall score). After a 30-min delay during which unrelated tests are administered, the participant is asked to recall as many words as possible from the initial list (delayed recall score). The RAVLT “forgetting score” is calculated as the difference between immediate recall versus delayed recall scores [33]. The forgetting score remains relatively stable over time in individuals with consistent memory function; the forgetting score tends to get smaller as the number of recalled items decreases. The 11-item Alzheimer’s Disease Assessment Scale (ADAS) cognitive subscale assesses learning and memory, orientation, and several aspects of language including production, comprehension, and constructional and ideational praxis [34,35]. Higher scores indicate more impairment. Finally, the Clinical Dementia Rating (CDR) scale is a measure of three cognitive domains (memory, problem solving, and orientation) and three functional domains (self-care, community engagement, and hobbies). Information is collected directly from the study participant, as well as from a study informant. The scores for the six domains are combined into the CDR sum of boxes (CDR-SB) score [36].

Statistical analysis

Cohort demographic summary statistics

Summary statistics for participants’ age, sex, age of onset, and APOE ε4 carrier status were calculated using R.

Imputation of HLA alleles

HLA genotypes were derived from chromosome 6 SNP data using an imputation program, HLA Genotype Imputation with Attribute Bagging (HIBAG) v1.3, which calculates predictions of genotype by averaging HLA-type posterior probabilities over an ensemble of classifiers built on bootstrap samples [37]. It relies on a training set of known HLA and SNP genotypes. We imputed the following HLA genes: A, B, DRB1, DQA1, and DQB1. For the UCSF MAC cohort and the ADGC merged dataset, a training set for four-digit resolution using ethnic-specific models for Europeans based on Omni1_Quad_v1_0_H was used. For the ADNI cohort clinical biomarker analyses, we used four-digit resolution ethnic-specific models for Europeans derived from Illumina Human610-Quad v1.0.

Quality control of HLA imputation

Based on the distribution of posterior probabilities for each of the five imputed alleles (see S1 Fig), we chose a call threshold (CT) of 0.75. As previous studies have shown that a CT of 0.5 leads to HIBAG prediction accuracies of 94.8%–99.2% for individuals of European ancestry [38], we expect our more stringent CT will correspond to similar or higher HIBAG prediction accuracies based on assumed accuracy of imputed ADGC SNPs. After excluding samples with any imputation probability below this cutoff at any locus, our final ADGC cohort size was 11,381.

Calculating locus and haplotype odds ratios (ORs)

OR estimates for patients with AD and cognitively normal controls were calculated using a statistical package designed to specifically probe associations with the HLA (BIGDAWG), including tests of Hardy-Weinberg equilibrium and case–control association analyses for haplotypes as previously described [39]. Analyses were performed for each cohort (UCSF + ADGC) separately and in combination. As this was a fine-mapping study based on a previous genome-wide significant, and replicated, finding at HLA-DRB5; and considering that this study represents a first analysis of the highly polymorphic HLA region in the context of AD, a complex disease, we did not require a multiple testing correction. To strike a balance between reducing Type I error while also allowing for full exploration of the loci underlying this association with the MHC region, we implemented a stepwise assessment of HLA gene contributions to AD: using allelic information, we established a priori significance at p < 0.05 at the haplotype level based on the prior GWAS-significant results. We then examined the contingency table from which the haplotype result was derived to identify the specific allele(s) contributing to the association signal. We accepted allele-level significance at p < 0.05 given the haplotype-level significance [40,41]. Based on a sample size of 11,690 in our combined UCSF+ADGC cohort, with 326 degrees of freedom and an alpha of 0.05, we had 64.1% power to detect an OR of 1.21 based on the haplotype frequencies of AD versus cognitively normal controls for the top associated five-allele haplotype.

Biomarker and cognitive data

Discrete and continuous demographic variables were compared across the ADNI cohort using chi-squared and ANOVA analyses, respectively. Linear mixed effects models were used to assess the relationship between the risk haplotype of interest and changes in the longitudinal cognitive measurements, ADAS and RAVLT, while controlling for baseline and time interactions of age, sex, education, baseline CDR-SB score (to account for baseline differences in clinical severity/diagnosis), and APOE ε4 carrier status. Use of linear mixed effects models allowed us to account for variable data missingness across participants by estimating subject-specific slopes. This enabled us to estimate cognitive changes for each individual despite varying numbers of visits. Missing data were omitted from the analyses, and all participants were required to have at least two time points to be included in the analysis. All interactions and main effects were modeled as fixed effects with random slopes and intercepts across individuals. The main effects of all variables were included in all longitudinal analyses but have been omitted from the definitions below to improve their clarity.

The linear mixed effects model for ADAS11 scores was defined as follows:

The linear mixed effects model for the RAVLT forgetting score was defined as follows:

Cross-sectional CSF biomarker analyses

Linear models were used to test for an association between baseline CSF biomarker levels and the haplotype of interest. We controlled for age, sex, education, baseline CDR-SB score (to account for baseline differences in clinical severity/diagnosis), and APOE ε4 dosage.

Results

Five-allele haplotype analysis implicated DR15 in AD risk

The discovery UCSF cohort consisted of 309 individuals with clinically diagnosed AD and cognitively normal older adult controls (Table 1). Because of the small sample size, all imputed alleles were included in the haplotype analysis (HLA A, B, DRB1, DQA1, and DQB1). We performed association analysis on the four haplotypes with sufficient frequency in this small cohort. Of these four, one showed a significant association with AD risk: HLA A*02:01~B*07:02~DRB1*15:01~DQA1*01:02~DQB1*06:02 (OR = 3.69; 95% confidence interval [CI] 1.16–13.69; p = 0.01) (Table 2).

After quality control, 11,381 individuals were available for analysis in the validation ADGC cohort (Table 1). Of the 318 haplotypes available for analysis, 12 five-allele haplotypes were significantly associated with AD risk (p < 0.05, Table 2). The strongest association was HLA A*03:01~B*07:02~DRB1*15:01~DQA1*01:02~DQB1*06:02 (OR = 1.22 [1.08–1.38], p = 8.5 x 10⁻⁴). This haplotype differed from the UCSF finding by one allele, at HLA-A. The third most significant haplotype association in the ADGC cohort was A*02:01~B*13:02~DRB1*07:01~DQA1*02:01~DQB1*02:02, which showed a protective effect, (OR = 0.66 [0.50–0.89], p = 4.2 x 10⁻³). This haplotype shared the HLA-A allele associated with AD risk in the UCSF discovery analysis. The full A*02:01~B*07:02~DRB1*15:01~DQA1*01:02~DQB1*06:02 haplotype associated with AD in the UCSF cohort was not significant in the ADGC cohort (p = 0.30).

In combined analysis of both the UCSF and ADGC cohorts, 326 haplotypes were available for analysis (additional haplotypes beyond the 4 + 318 haplotypes analyzed in the separate UCSF and ADGC cohorts resulted when sufficient numbers of AD and CN controls for rare haplotypes became available in the combined UCSF + ADGC dataset). HLA A*03:01~B*07:02~DRB1*15:01~DQA1*01:02~DQB1*06:02 (OR = 1.21 [1.08–1.37), p = 9.6 x 10⁻⁴) and A*02:01~B*13:02~DRB1*07:01~DQA1*02:01~DQB1*02:02, [OR = 0.66 [0.49–0.88], p = 3.8 x 10⁻³] remained as two of the three most significant associations with AD (Table 2). Locus-level analyses of the combined cohort showed independent AD associations of B*07:02, DRB1*15:01, DQA1*01:02, and DQB1*06:02 (Table 3).

Five-allele haplotype contributed to AD risk independently of APOE ɛ4 and may be driven by ɛ4-negative individuals

We next assessed whether the strong genetic AD risk factor APOE ɛ4 can account for the most significant five-allele haplotype association we identified in the combined UCSF+ADGC cohort. We recoded individuals as carriers or noncarriers of the A*03:01~B*07:02~DRB1*15:01~DQA1*01:02~DQB1*06:02 haplotype and assessed in a logistic regression framework the independent contributions of the risk haplotype, APOE ɛ4 carrier status, and whether there was an interaction between the two. As expected, APOE ɛ4 was strongly associated with AD risk (p < 2 x 10⁻¹⁶). The five-allele risk haplotype remained a significant contributor to AD risk (p = 0.036), but there was no statistically significant interaction between haplotype and APOE ɛ4 (p = 0.19). However, dividing the cohort by APOE ɛ4 carrier status showed that the frequency of the A*03:01~B*07:02~DRB1*15:01~DQA1*01:02~DQB1*06:02 haplotype was higher only in individuals with AD who are negative for ɛ4 (Table 4). Analysis of variance performed separately in ɛ4 carriers and noncarriers resulted in a significant association of the five-allele haplotype with AD only in ɛ4-negative individuals (p = 0.036 in ɛ4 noncarriers; p = 0.90 in ɛ4 carriers).

Separate class I and class II haplotype analyses corroborated A*03:01~B*07:02 and DR15 in AD risk

Given the different roles of HLA receptors in recognizing endogenous (class I) or exogenous (class II) ligands, we also assessed class I (HLA A~B) and class II (HLA DRB1~DQA1~DQB1) haplotypes separately for their role in AD risk in the combined UCSF+ADGC cohort. Of 202 analyzed class I haplotypes, ten two-allele haplotypes were significantly associated with AD (p < 0.05), including A*03:01~B*07:02 (p = 0.03, OR = 1.1 [1.0–1.2]) (Table 5). Only one three-allele class II haplotype (out of 30 analyzed) was associated with AD risk, DR15 (p = 0.025, OR = 1.1 [1.0–1.2]). Together, these two separate haplotypes represent the most strongly associated five-allele haplotype identified in the combined analysis.

Class I haplotype A*03:01~B*07:02 was associated with baseline CSF amyloid levels

We utilized a subset of the ADNI cohort with genetic and cognitive data available to assess the disease-specific relevance of the class I A*03:01~B*07:02 and class II DR15 haplotypes across the AD spectrum, including cognitively normal controls, individuals with MCI, and those with AD. We analyzed the two haplotypes separately to assess whether class I and class II risk-associated haplotypes were correlated with similar or different clinical measures of AD. The cohort was balanced with respect to age, sex, and haplotype distributions (Table 6). The cohort was significantly different with respect to education and number of time points and showed expected differences in CDR-SB baseline score, APOE ɛ4 carrier status, ADAS baseline score, and RAVLT forgetting baseline score.

Carrying A*03:01~B*07:02 was associated with higher baseline levels of amyloid β as measured in CSF (Fig 1, p = 0.01). Traditionally, CSF amyloid levels are inversely correlated with amyloid burden in the brain; our results suggest that carrying A*03:01~B*07:02 is correlated with lower amyloid levels in the brain [43]. This is observed despite the fact that there were no statistically significant differences in baseline clinical or biomarker measures in patients with versus without the risk haplotype (S3 Table, S2 Fig). A*03:01~B*07:02 was not associated with any other baseline measures and was not associated with change in longitudinal measures over time.

DR15 risk haplotype correlated with worse cognitive decline and greater baseline inflammation across the AD spectrum

Longitudinal analysis of cognitive data identified a statistically significant association between the number of alleles of the DR15 risk haplotype and ADAS cognitive scores (p = 0.03), as well as with RAVLT forgetting scores (p = 0.02, S4 Table). The DR15 haplotype was associated with worse decline over time on both measures, corresponding to increasing longitudinal ADAS cognitive scores and decreasing longitudinal RAVLT forgetting scores over time (shown relative to noncarriers in Figs 2 and 3). DR15-associated changes in cognitive trajectory occurred despite the fact that there were no baseline differences in clinical severity or cognitive function in patients with AD based on DR15 carrier status (S5 Table). In addition, baseline biomarker measures most relevant to AD were similar in both patients with AD who are DR15 carriers and those who are noncarriers (S5 Table, S3 Fig), indicating that all patients had equivalent baseline disease severity.

In a subset of individuals who also had baseline CSF data available (S2 Table), we tested whether the DR15 risk haplotype altered any biomarker measures of immunological function and inflammation. We tested 28 analyte levels related to immune function and inflammation (S1 List). At baseline, there was an association between chemokine CC-4 (CC4) and age (p = 0.02, S4 Fig), as well as CC4 with dose of DR15 risk haplotype (p = 5.18 x 10⁻³, Fig 4, S6 Table). Although not reaching strict statistical significance after adjustment for the 28 biomarkers tested (at Bonferroni adjusted p < 1.79 x 10⁻³), this analysis provides suggestive biomarker evidence of heightened baseline inflammation in individuals carrying the DR15 risk haplotype.

HLA haplotype risk effects differed by sex

Given previous reports of greater risk effects of DR15 in female patients with multiple sclerosis (MS) [44] and the stronger effect of APOE ɛ4 in females [45], we assessed whether men versus women showed similar or different HLA haplotype associations with AD risk. When split by sex, two of the three most significant five-allele haplotypes from the combined sex analysis were significant in an individual sex. The five-allele haplotype A*03:01~B*07:02~DRB1*15:01~DQA1*01:02~DQB1*06:02 was significant only in men (OR 1.31 [1.09–1.58], p = 0.0035) (S7 Table). However, A*02:01~B*13:02~DRB1*07:01~DQA1*02:01~DQB1*02:02 was significant only in women (OR 0.68 [0.46–0.99], p = 0.034) (S7 Table). Similar findings appeared in separate class I and class II haplotype analyses. Only men showed significant associations with class I haplotype A*03:01~B*07:02 (p = 0.027), and only women showed significant associations with A*02:01~B*13:02 (p = 0.0049) (S8 Table). Finally, class II haplotype DR15 was only significantly associated with AD risk in men (p = 0.01) (S8 Table). Locus-level analyses were consistent, with only men showing significant associations with ten alleles, including B*07:02 (p = 0.013), DRB1*15:01 (p = 0.0096), DQA1*01:02 (p = 0.029), and DQB1*06:02 (p = 0.01) (S9 Table). There were four individual alleles associated with AD risk in women, none of which were components of any of the top three significant five-allele haplotypes in the combined sex analysis (S9 Table).

Iterative subanalyses corroborate role of HLA-A*03:01~B*07:02~DRB1*15:01~DQA1*01:02~DQB1*06:02 in AD

To attempt to alleviate concern over possible Type I error in this analysis, we randomly split the combined ADGC+UCSF cohort ten times (maintaining the same proportion of AD:controls) and reran the five-allele haplotype analysis in the 20 resulting (smaller) cohorts. Two of the top-associated five-allele haplotypes showed p-values < 0.05 in over half of the randomly split analyses (S10 Table), which was more than any of the other “top” haplotypes from the original analysis. This included the one we focused on in this study (A*03:01~B*07:02~DRB1*15:01~DQA1*01:02~DQB1*06:02), which showed significance in 11 iterations of the randomly split analysis, with p-values from 0.026–0.0001 and ORs of 1.21–1.40, further corroborating the contributions of this haplotype and its subcomponents to AD risk.

Discussion

In a total of over 11,000 individuals, we found evidence suggesting that the five-allele HLA haplotype A*03:01~B*07:02~DRB1*15:01~DQA1*01:02~DQB1*06:02 is a risk factor for AD and that this effect may be driven by men who do not carry the major AD risk factor, APOE ɛ4. Locus-level analysis further confirmed AD associations of the individual alleles B*07:02, DRB1*15:01, DQA1*01:02, and DQB1*06:02. In separate class I and class II haplotype analyses, the class I A*03:01~B*07:02 haplotype and the class II DRB1*15:01~DQA1*01:02~DQB1*06:02 (DR15) haplotype were both significantly associated with risk for AD. We assessed the clinical relevance of each of these haplotypes separately in a smaller cohort representing the spectrum of cognitively normal controls and individuals with MCI and AD. Carrying the MHC class I haplotype A*03:01~B*07:02 was associated with higher CSF amyloid levels, suggesting lower levels of amyloid in the brains of haplotype carriers across the AD spectrum. The class II haplotype DR15 was associated with greater rate of decline on two different measures of cognitive function relevant to AD in a dose-dependent manner. In a subset of the same cohort, carrying the DR15 risk haplotype was also associated with higher baseline levels of CC4, a biomarker of AD-related inflammation [46]. Taking these findings together, this study provides evidence for the contribution of the A*03:01~B*07:02~DRB1*15:01~DQA1*01:02~DQB1*06:02 haplotype and its components, A*03:01~B*07:02 and DR15, to risk of AD.

Over 30 y of research into HLA alleles and risk of AD has yielded mixed conclusions due in part to limitations in mapping alleles within this complicated genomic region. Early studies mapped risk of AD to the HLA region of chromosome 6 [47], and the studies that followed differed significantly in their methodological approach, the identities and resolution of the alleles studied, the ethnicity of the study cohorts, and the inferences drawn from the data. MHC class I molecule HLA-A*02 has been shown to either be associated with increased risk of AD or to have no effect in nearly 15 different studies [48–61]. Given that only B*07:02, DRB1*15:01, DQA1*01:02, and DQB1*06:02 showed significant locus-level associations with AD, our findings are consistent with an ambiguous role of HLA-A*02 in AD. In terms of class II alleles, one study by Mansouri and colleagues demonstrated a link between DRB1*15:01~DQB1*06:02 and AD in a small cohort of Tunisians [62], consistent with our findings. Previous GWASs have found that AD risk is associated with a SNP in DRB5 [63]. As there is strong LD between DRB5*01 and DRB1*15:01~DQB1*06:02 [64], it is possible that the AD association we have detected with DR15 is due in part or wholly to DRB5. Finally, our finding that HLA associations with AD are stronger in APOE ɛ4-negative individuals is consistent with prior work for different HLA alleles [65,66].

The HLA region has been studied to a varying extent for its contributions to neurological disease, and many of the risk alleles implicated in the present studies have also been linked to other disorders. Most notably, the class II DR15 haplotype is the most consistently replicated genetic finding in MS [67–69]. DR15 also correlates with worse clinical progression in women with relapsing-onset MS (e.g., younger age at onset and more subcortical atrophy) [44]. Class I allele B*07 has also been associated with MS risk, particularly in those also carrying DRB1*15 [70]. In one Parkinson disease (PD) study, four alleles identified in a risk haplotype overlapped with our top five-allele haplotype association [71]. Similar to AD, other studies have also implicated the HLA-DRB5 region in PD risk [72]. In one small autism study, the class I allele B*07 and class II allele DQB1*06:02 were both associated with disease risk [73]. Finally, class II allele DQB1*06:02 has been associated with marked increased risk for [74,75], and worsened severity of [76,77], narcolepsy. These findings are consistent with our study, in which we identified a dose-dependent association between DR15 and greater cognitive decline in individuals representing the AD spectrum.

Participants who carried at least one copy of the class I haplotype A*03:01~B*07:02 on average had higher baseline CSF amyloid levels, suggesting lower amyloid burden in the brains of these participants. Similar findings have been observed in APOE ɛ4-negative individuals when compared to APOE ɛ4-positive individuals across the phenotypic spectrum of cognitively normal to early MCI, suggesting higher brain amyloid β in carriers [78]. This finding raises the possibility that there could be a tau-mediated effect on AD clinical symptoms, as the AD group did not differ in clinical measures by haplotype carrier status. On the other hand, DR15 haplotype carriers demonstrated subtle differences in baseline inflammatory biomarker levels, as well as a worse cognitive trajectory over time, suggesting a disease-modifying effect that could be mediated by changes in immune function.

Our study benefited from several strengths. The primary discovery cohort was a well-characterized sample of patients who received extensive clinical evaluations at the UCSF Memory and Aging Center. The replication dataset from the ADGC of over 11,000 AD and cognitively normal control individuals is the largest dataset to date used to explore immunogenetic contributions to AD risk. Lastly, longitudinal data from ADNI allowed us to probe the potential clinical relevance of haplotype findings across the AD spectrum. Our study also has caveats that are important to consider. Our imputation program predicted accuracy for individuals with European ancestry that was likely higher than 94.8%–99.2% based on our more stringent call threshold in comparison to other studies [38]. Imputed HLA alleles have been shown to be reliable classification tools in studies with similar methodologies [38,79,80]; however, future studies would benefit from direct sequencing of HLA alleles to avoid potential imputation inaccuracies. Because of limitations in the imputation package selected for HLA allele calling, we were only able to impute genotypes for a subset of MHC class I and II genes. For example, the imputed genes available did not include DRB5, which was indicated in previous studies to be associated with AD risk or pathological processes [9,24,81]. DRB5 is on the DR15 haplotype, so it is likely that the association we identified reflects these previous results. However, we can neither directly confirm nor refute this possibility in the present study. The DR15 risk haplotype is most common in Europeans, and to minimize genetic heterogeneity in population substructure, we limited the present analysis to white individuals of non-Hispanic descent. Additional studies are required to assess the identified HLA risk haplotypes and component alleles for their contribution to AD in more diverse populations where patterns of LD differ and may uncouple alleles that were tightly linked in our study population, though the initial study identifying class II associations with AD in Tunisians suggests this may be a generalized risk phenomenon. Although p < 0.05 may be considered lenient based on the number of total alleles tested, it is also true that all of these alleles represent only five genes within one genomic region that has been previously linked to AD risk. Despite reduced statistical power due to low frequency of HLA haplotypes imparted by the extraordinary diversity of this region, we feel that this study is an important first step in elucidating the underlying contribution of the HLA to AD risk given the medical implications of ultimately identifying immune-related therapies as a means of modifying a complex, common disease. We have greater confidence in our findings due to corroborating clinical validity as identified in the ADNI cohort. Iterative subanalyses of the combined study cohort further support a role of our top five-allele haplotypes in AD risk. In addition, two of the main alleles of interest we identified, DRB1*15:01 and DQB1*06:02, have been linked to AD risk in two prior studies, further supporting our results. We also identified several other risk haplotypes in our analyses beyond the ones we focused on in this study; the clinical relevance of these additional haplotypes and alleles requires further investigation. Future work is also required to test whether these findings extend to early-onset and atypical clinical syndromes with underlying AD pathology.

In summary, we present evidence for a role of the HLA class I A*03:01~B*07:02 haplotype and the HLA class II DR15 haplotype in AD risk. Our study also suggests that these risk haplotypes may be associated with CSF AD biomarker levels (class I) and greater decline in cognition over time, as well as higher levels of inflammation across aging (class II). The results of our study indicate that the broad A*03:01~B*07:02~DRB1*15:01~DQA1*01:02~DQB1*06:02 haplotype may contribute genetic risk to AD beyond that contributed by the established risk factor APOE ɛ4, particularly in men. As components of this haplotype are well-established risk factors in MS, PD, autism, and narcolepsy, we propose that they may contribute to underlying biological risk mechanisms in multiple neurological diseases. Future work is required to establish the precise molecular processes underlying this risk association, as well as to expand this finding to broader, diverse populations of AD and potentially even other neurodegenerative conditions.

Supporting Information

Attachment 1

Attachment 2

Attachment 3

Attachment 4

Attachment 5

Attachment 6

Attachment 7

Attachment 8

Attachment 9

Attachment 10

Attachment 11

Attachment 12

Attachment 13

Attachment 14

Attachment 15

Attachment 16

Attachment 17

Attachment 18