The MUC6/AP2A2 Locus and Its Relevance to Alzheimer’s Disease: A Review

Abstract

We recently reported evidence of Alzheimer’s disease (AD)-linked genetic variation within the mucin 6 (MUC6) gene on chromosome 11p, nearby the adaptor-related protein complex 2 subunit alpha 2 (AP2A2) gene. This locus has interesting features related to human genomics and clinical research. MUC6 gene variants have been reported to potentially influence viral—including herpesvirus—immunity and the gut microbiome. Within the MUC6 gene is a unique variable number of tandem repeat (VNTR) region. We discovered an association between MUC6 VNTR repeat expansion and AD pathologic severity, particularly tau proteinopathy. Here, we review the relevant literature. The AD-linked VNTR polymorphism may also influence AP2A2 gene expression. AP2A2 encodes a polypeptide component of the adaptor protein complex, AP-2, which is involved in clathrin-coated vesicle function and was previously implicated in AD pathogenesis. To provide background information, we describe some key knowledge gaps in AD genetics research. The “missing/hidden heritability problem” of AD is highlighted. Extensive portions of the human genome, including the MUC6 VNTR, have not been thoroughly evaluated due to limitations of existing high-throughput sequencing technology. We present and discuss additional data, along with cautionary considerations, relevant to the hypothesis that MUC6 repeat expansion influences AD pathogenesis.

INTRODUCTION

This review article is about a putative Alzheimer’s disease (AD)-associated genetic polymorphism within the mucin 6 (MUC6) gene, and <4000 base pairs (bp) away from the gene that encodes adaptor-related protein complex 2 subunit alpha 2 (AP2A2), on human chromosome 11p15.5 (1). The genomic locus of interest is a poorly annotated variable number of tandem repeat (VNTR) region, largely unreadable by conventional sequencing methods. Here, we review the topical literature on MUC6, the risk-associated polymorphic VNTR region, and AP2A2. We examine whether it is credible that a novel high-impact risk allele could be discovered, given that so many AD-linked genetic loci have already been identified. Evidence is provided in support of the hypothesis that MUC6 VNTR repeat expansion is associated with AD risk. We also discuss reasons to be skeptical of the hypothesis. Relevant data are presented from the University of Kentucky AD Center (UK-ADC) autopsy cohort. Finally, critical knowledge gaps are highlighted.

Mucin 6 (MUC6)

The MUC6 gene is located in a recombination prone region approximately 1 million bp from the telomere of human chromosome 11p (2, 3). MUC6 is clustered along with 3 other mucin family genes (MUC2, MUC5A, and MUC5B) and AP2A2. MUC6 and AP2A2 are in close proximity, oriented in opposite directions—their 3′-untranslated regions nearly overlap (Fig. 1). The juxtaposition of AP2A2 and MUC6, with possible sharing of 3′ regulatory sequence elements, is conserved on chromosomes of other vertebrates including frogs (Xenopus tropicalis).

FIGURE 1.

The region on human chromosome 11 which contains the MUC6 variable number of tandem repeat (VNTR) region. This locus is ∼1 million bp from the chromosome 11p telomere. The window of human chromosome 11p15.5 shown here (A) is relatively small (<24 kb) and contains only the 3′ portions of both AP2A2 and MUC6 genes. The yellow arrow indicates the MUC6 exon 31 which contains the VNTR region. Note that, in this annotated assembly, the MUC6 Exon 31 is <4 kb in size, although the actual MUC6 exon 31 is usually >10 kb (often >15 kb). This underscores the extremely incomplete annotation for this VNTR-containing exon. Source: https://genome.ucsc.edu/. Panel (B) provides a selective overview of previously reported findings related to MUC6 exon 31. Panel (C) depicts, in cartoon form, the distinction between single-nucleotide variants (SNVs) and variation in a VNTR region. Individual SNVs are by definition small (usually 1 bp), whereas VNTR regions’ tandem repeat expansion can lead to a much larger change in the local genomic architecture. For the MUC6 VNTR region, each repeat unit is ∼507 bp in length.

Mucin genes encode proteins that become glycosylated and serve as gel-like substances to protect and lubricate epithelial surfaces (4). The MUC6 protein is expressed preferentially in the gastrointestinal (GI) tract, and also in genitourinary and pulmonary epithelia (5–9). Here, we will focus on the repetitive domains and gene expression regulation of MUC6, as may be relevant to the hypothesis that the MUC6 VNTR is an AD risk allele.

A conspicuous feature of the mucin gene family is that long repetitive genetic sequences reside in exons, and are transcribed and translated into proteins (10). The repetitive DNA regions encode polypeptide motifs enriched with residues for O-linked glycosylation—proline, threonine, and serine amino acids, thus, they are termed PTS domains. These glycosylated polypeptides help constitute the “gel forming” components of mucus (10).

Another interesting characteristic of the mucin gene family is that the different mucin genes’ expression may be subject to orchestrated regulation by inflammation-related transcription factors (11). In cells transcribing the mucin genes clustered together on chromosome 11p15.5, the local chromatin forms three-dimensional structures for coordinated exposure of multiple mucin genes’ transcriptional promoter sequences (12). These interactions are complicated (13) and context-specific: In various inflammatory diseases, there is differential regulation of chromosome 11p15.5 mucin genes’ expression (14–17).

MUC6 transcripts are expressed at very low levels in human brain (Fig. 2). However, MUC6 protein function may have an impact on the brain. There is increasing appreciation of brain-gut interactions, which may theoretically be affected by the microbiome—that is, intestinal bacteria. Mucins, including MUC6, constitute a barrier for intestinal organisms: MUC6 has been reported to block Helicobacter pylori infections (18, 19), and mucin proteins play similar functions in altering risk for other bacterial and viral infections (4, 20–22). The intestinal microbiome may play mechanistic roles in neurodegeneration, including AD (23–25); therefore, a microbiome-modulating gene may affect AD indirectly.

FIGURE 2.

Screenshots from the Genotype-Tissue Expression (GTEx) data set depicting human tissue-specific gene expression patterns for MUC6 (A) and AP2A2 (B). Note that MUC6 is expressed strongly in stomach tissue but not appreciably in the CNS (yellow bars). By contrast, AP2A2 is expressed at high levels in CNS tissue. The source of the data for these figures was the GTEx Portal, https://gtexportal.org. This figure was adapted from Katsumata et al (1), with permission.

MUC6 may also influence susceptibility to non-GI viral infection, which in theory could be relevant to AD risk. There is a body of literature supporting the hypothesis that AD is caused or exacerbated by viral, particularly herpesvirus, infection (26–29). In a recent study by Readhead et al (30), herpesviruses’ genetic imprints were increased in AD brains. Readhead et al also performed a genome-wide association study (GWAS) to assess the associative impact of human genetic variants on vulnerability to brain infection by herpesviruses (30). Remarkably, the human genetic variant “associated with the most viruses… (rs71454075) falls within the glycoprotein mucin 6” (30)—and, this virus-associated polymorphism resides in the MUC6 VNTR region. MUC6 gene variation was also associated with vulnerability to respiratory syncytial virus infection (31). The association(s) between AD and infectious disease are hypothetical and controversial (see, e.g., [32]). However, the prior published findings suggest that complex processes related to the gut microbiome and/or viral immunity may be modulated by MUC6. It is unknown whether the regulation of MUC6 gene expression affects or is affected by expression of the nearby gene, AP2A2.

Adaptor-Related Protein Complex 2 Subunit Alpha 2 (AP2A2)

In contrast to MUC6, AP2A2 is transcribed robustly in many human tissues including the brain (Fig. 2). The AP2A2 polypeptide is a component of the adaptor protein complex, AP-2 (33, 34), a 4-protein multimer that is located on a subset of endocytic vesicles (35, 36). This is an evolutionarily ancient protein complex and the AP2A2 gene has conserved orthologs in plants, fungi, and invertebrates (37). Functionally, AP-2 participates in assembling the components of early clathrin-coated vesicles (CCVs) at the plasma membrane (33, 35, 38) (Fig. 3). The AP-2 complex interacts directly with lipids, clathrin, “cargos” that are undergoing CCV-mediated endocytosis, and accessory CCV proteins such as PICALM and BIN1 (39–41).

FIGURE 3.

Clathrin-coated vesicle (CCV) formation at the plasma membrane involves late-onset Alzheimer’s disease (AD) risk-associated gene products including BIN1 and PICALM proteins. The CCV is here shown in early stage of formation at the plasma membrane. Clathrin is depicted with a dashed line. The AP2A2 protein is a component of the AP-2 adapter complex which is located on the plasma membrane and then on CCVs as the vesicle is internalized. Protein cargos that may be related to AD pathogenesis, including Tau, APP, BACE1, APOE, and various viruses have been reported to be processed via CCV-mediated endocytosis. Other (non-CCV) endocytosis mechanisms also occur for some of these proteins. For this cartoon, some relevant structures were minimized for the sake of clarity.

Cargos for CCV-mediated endocytic internalization via AP-2 have been reported to include the amyloid precursor protein (APP), beta-site APP cleaving enzyme 1 (BACE1), Tau, synaptic vesicle proteins, and many others (33, 35, 41–47). It also has been proposed that the apolipoprotein E (APOE) protein and lipoprotein receptors are internalized via CCV endocytosis (43, 48–50). There is an extensive literature documenting the roles of CCV endocytosis and AP-2 in viral infection (51–54). In sum, AP2A2 function may tie in with multiple mechanisms linked to AD pathogenesis.

AP-2 is a stable protein complex comprising 4 domains, which are termed alpha, beta, mu, and sigma (39). AP2A1 and AP2A2 are homologous human genes that encode the alpha subunit of the AP-2 complex. Relatively little is known about the molecular neurobiology of AP2A2 per se. Publicly accessible databases indicate that the AP2A2 transcript is expressed in a broad range of CNS cell lineages (data not shown). At the tissue level, both AP2A1 and AP2A2 are highly expressed in human cerebellum, but AP2A2 is expressed at lower levels in the cerebral cortex according to the GTEx Portal (https://gtexportal.org/ [55]); Figure 4. These findings are intriguing because the cerebellum is relatively resistant to AD-type pathology (56, 57).

FIGURE 4.

Screenshots from the Genotype-Tissue Expression (GTEx) data set depicting human CNS gene expression for the adapter complex (AP-2) alpha subunit-encoding gene transcripts, AP2A2 (A) and AP2A1 (B). These 2 genes encode proteins homologous in sequence—80.2% identity and 90.3% similar according to https://embnet.vital-it.ch/software/LALIGN_form.html. Both genes’ transcripts are expressed at high levels in the cerebellum (red asterisks), but the AP2A1 (compared to the AP2A2 transcripts) are expressed at higher levels in the cerebral cortex (green asterisks). Implications of this difference are not known; there have been relatively few studies of AP2A2 expression regulation in the human brain. BA, Brodmann area. The source for these figures was the GTEx Portal, https://gtexportal.org.

We performed immunohistochemical staining for AP2A2 in human brains. In nondemented control brains, AP2A2 immunoreactivity showed a punctate pattern in neurons, consistent with CCV/vesicular structures (1). In AD brains, AP2A2 colocalized with phospho-tau-immunoreactive neurofibrillary tangles (NFTs); by contrast, AP2A2 colocalization with phospho-tau was not seen in progressive supranuclear palsy (PSP)-tufted astrocytes (Fig. 5). Intriguingly, a prior study reported a similar immunostaining pattern using an antibody against CCV-associated protein PICALM—the PICALM protein also was colocalized with phospho-tau in AD NFTs but not with tufted astrocytes in PSP (58).

FIGURE 5.

Photomicrographs of human brain (aged control [A–C], Alzheimer’s disease [D–F], and progressive supranuclear palsy [PSP; G–I]) show the colocalization of AP2A2 protein with phospho-tau-immunoreactive neurofibrillary tangles of Alzheimer’s disease. Panels (C), (F), and (I) have merged staining results combined with blue DAPI counterstain to visualize cell nuclei. AP2A2-immunoreactive cells are seen in aged control brain but are not colocalized with phospho-tau (A–C). In Alzheimer’s disease cerebral neocortex, there was a strong tendency for colocalization of an accentuated AP2A2 signal (D) that colocalized with phospho-tau (E); see yellow arrows. By contrast, cells that were immunonegative for phospho-tau tangles (blue arrow) had less AP2A2 immunoreactivity. In PSP basal ganglia, the phospho-tau immunoreactivity (yellow arrowheads) in tufted astrocytes did not colocalize with AP2A2 immunoreactivity. The colocalization was quantified and methods published in detail in (1). Thanks to Dr Peter Davies for the anti-phospho-tau (PHF-1) monoclonal antibody. Scale bars = 100 μM.

As to the impact of AP2A2 genetic variation, prior GWAS have linked AP2A2 gene variants with AD risk—yet not conclusively (Tables 1 and 2). (By convention, we refer to single nucleotide polymorphisms as SNPs if there is common variability at that locus in human populations and, more generally, as single nucleotide variants [SNVs] if the minor allele is <1% prevalent [59]). A common AP2A2 SNP (rs10751667) was listed as a “suggestive” AD-linked allele in a large, meta-analytic GWAS: nominal p∼6 × 10⁻⁷ association with the AD phenotype, not statistically significant after correcting for multiple comparisons (60) (Table 1). In a separate study, the same SNP (rs10751667) was associated with a subtype of mild cognitive impairment (61). Another recent GWAS reported a different AP2A2 SNP, rs10794342, again with suggestive nominal (p∼4 × 10⁻⁶) association with AD risk (62) (Table 1).

TABLE 1.

AP2A2 Single Nucleotide Polymorphisms (SNPs) That Showed “Suggestive” Nominal Association With Alzheimer’s Disease (AD) Phenotype in Genome-Wide Association Studies (GWAS), Along With Expression Quantitative Trait Locus (eQTL) Status^*

AP2A2 SNP With Nominal GWAS Association With the AD Phenotype	Reference	Nominal p Value for Association With AD Phenotype	AD Association Statistically Significant After Correction for Multiple Comparisons?	eQTL for AP2A2?*	eQTL for MUC6?*
rs10751667	Lambert et al (60)	p = 6.3 × 10−7	No	Yes; p = 5.7 × 10−30*	Yes; p = 2.9 × 10−35*
rs10794342	Nazarian et al (62)	p = 4.4 × 10−6	No	Yes; p = 5.9 × 10−40*	No*

^*eQTL status and p values indicated were according to “Multiple Tissue eQTL Comparison Meta Analysis RE2” at GTEx Portal, https://gtexportal.org/. For details on the methods and numbers of samples involved in the GTEx work, see https://gtexportal.org/home/documentationPage#staticTextAnalysisMethods.

TABLE 2.

AP2A2 Single Nucleotide Polymorphisms (SNPs) From Table 1 in More Recent Alzheimer’s Disease GWAS With Larger (but still overlapping) Samples

Reference	Kunkle et al (64)		Jansen et al (63)
	Stage 1	Stage 2
Sample size, n	63 926*	18 845*	455 258*,†
rs10751667	1.4 × 10−4‡	2.2 × 10−6‡	7.5 × 10−3‡
rs10794342	3.6 × 10−4‡	9.2 × 10−5‡	7.2 × 10−3‡

^*Overlapping samples with Lambert et al (60).

^†The Jansen et al study included 47 493 Alzheimer’s disease “proxy” cases and 328 320 “proxy” controls as determined by reported parental family history.

^‡Nominal p values from summary data are shown; these were not statistically significant after correcting for multiple comparisons.

Larger, meta-analytic AD GWAS have been published in the past few years. We highlight the results of studies by Jansen et al (63) and Kunkle et al (64), for which we obtained summary statistics (Table 2). Again, the associations between AP2A2 SNPs and the AD phenotype were nominally statistically significant, and with the same risk alleles as in prior reports. A similar finding was reported in the study by Marioni et al (65) (data not shown), but this study sample (the included cases and controls) was highly similar to the later Jansen et al (63) study, for which results are shown in Table 2. More generally, evaluating the abovementioned GWAS independently of each other is not possible via summary statistics because all of these GWAS samples had some degree of overlap.

The 2 marginal AD-associated SNPs (rs10751667 and rs10794342) are in moderate linkage disequilibrium (LD) with each other (D’ = 0.99, r² = 0.58 in “all populations” according to ldlink.nci.nih.gov), which means that their genetic variants tend to be co-inherited. In the Nazarian et al (62) and Lambert et al (60) papers, the less common alleles were protective (OR = 0.82 and OR = 0.93, respectively). Variation of both rs10751667 and rs10794342 SNPs is associated with altered expression for AP2A2, i.e., both are AP2A2 expression quantitative trait loci (eQTL), whereas rs10751667 is also an eQTL for MUC6. Table 1 shows results from the GTEx Portal website describing meta-analyses of data derived from CNS and non-CNS tissues. As an example of results from brain tissue, according to the GTEx Portal, the rs10794342_T allele is associated with lower AP2A2 expression in human cerebellum (p < 1 x 10⁻⁷ based on the analysis of 209 brain samples). It is an intriguing possibility that a particular polymorphism or haplotype is a risk allele for AD and also could alter expression levels for both AP2A2 and MUC6. The GTEx Portal indicates that rs10794342 and rs10751667 SNPs are also associated with alternative AP2A2 splicing—thus, are sQTLs for AP2A2—in specific human tissues (data not shown), another phenomenon that may merit further study. The SNPs shown in Table 1 are <75 kbp away from the MUC6 VNTR region, and are in LD with SNPs in the MUC6 VNTR region (data not shown). Thus, it is possible that the AD-linked AP2A2 SNPs are a “flag” (proxy feature) for the status of the MUC6 VNTR.

Additional published data have provided suggestive links between AP2A2 gene variants and the AD phenotype. A recent GWAS of brain transcriptomic and splicing data, correlating with AD phenotype data, reported that transcript expression variability for both AP2A2 and the homologous AP2A1 gene was associated with the AD phenotype (66). A separate study analyzed AD GWAS and transcriptomic data and found that AP2A2 was a “potential disease-causal” gene (67). Further, in a large, inbred Amish family with genetic risk for AD that was independent of APOE, a genomic screen showed that the genetic marker most strongly associated with the AD phenotype (highest multipoint logarithmic of the odds, or LOD, score) was located <500 kb from AP2A2 on chromosome 11p (68).

Variants of other genes that encode CCV-related proteins have been repeatedly associated with late-onset AD risk (58, 69–72) (Fig. 3). In a recent large GWAS that incorporated many different study cohorts, PICALM and BIN1 gene variants were the non-APOE risk alleles with the lowest p values (73). PICALM and BIN1 proteins have also been implicated in Tau proteinopathy and in APP/Aβ processing (40–42, 58, 74–77). Additional endocytic genes—for example, SORL1 and CD2AP—were also associated with APP processing, autophagy, and clinical dementia (71, 72, 78). Taken together, these results provide theoretical support for the hypothesis that a gene variant which affects AP-2/CCV function could have an impact on AD risk.

The MUC6 VNTR

The term VNTR describes repetitive end-to-end iterations of genomic sequences ≥6 bp in length (79, 80). The number of repeat units is inherited in the germline and often differs from individual to individual (79, 81). There is a wide variety of VNTR sequences in the human genome (79). We note that the genomic segment we are referring to as a VNTR region could also be referred to as a copy number variant (82, 83), structural variant (84, 85), satellite DNA (86), insertions/deletions (87), simple sequence repeats (88), or tandem repeat expansion, according to different sources. The larger (with more repeats) versions of many different tandemly repeated sequences in the human genome have been associated with neurologic diseases (89), and VNTRs can be transcriptional regulators for nearby genes (90). In the case of the C9ORF72 tandem repeat expansion, now known to be a strong neurodegenerative disease genetic risk factor (91), the existence of that tandem repeat expansion was signaled by GWAS that identified nearby coinherited SNPs (92–94). Those SNPs are not themselves pathogenetic but they served as proxies to flag the disease-causing tandem repeat locus.

Other investigators have studied the MUC6 VNTR region (95–101). No other known human tandem repeat sequence is highly similar to the MUC6 VNTR region. Most human genome VNTR regions contain repeated sequences that are shorter (usually 10–60 bp) than those in the MUC6 gene (80). By contrast, the tandemly repeated sequences in MUC6 are ∼507 bp each, and these unit repeats are composed of PTS-encoding subunits (79, 95, 101). The number of ∼507-bp repeat sequences in the human MUC6 VNTR region is usually between 15 and 36 repeats (Fig. 1): In ∼95% of cases, the MUC6 VNTR region’s size ranges between 7.5 and 18 kb (19, 98). MUC6 PTS domain-encoding DNA sequences are highly polymorphic (sequences differ between human individuals), causing one group to hypothesize that “the actual amino acid sequence is less important as long as it contains a sufficient number of anchor sites for O-glycans…” (95).

Because of the known GI tissue expression of MUC6, a major focus in prior work has been on how the MUC6 tandem repeat polymorphisms are associated with risk for GI diseases—mainly cancer and infection (19, 96, 102, 103). Findings from these previous studies indicated that the short version(s) of the MUC6 VNTR was the risk alleles, in contrast to what we reported in AD (19, 99). Interestingly, minisatellite polymorphisms outside of exon 31 were also located in and near the MUC6 gene, and were associated with MUC6 expression (96, 103). Many other studies have evaluated MUC6 as a pathologic biomarker in human diseases, but most did not address tandem repeat polymorphism per se.

Unfortunately, the MUC6 VNTR region is extremely difficult to sequence. This difficulty stems from technical factors, related to the size, number, and interindividual differences of the tandem repeats, as well as the incomplete annotation of the gene (95). There has not been a full-length MUC6 transcript sequence reported to date (partial cDNA sequences have been described) (7, 10, 97, 100, 104). The technical difficulty in sequencing also applies to DNA haplotypes (79, 95). The annotated GRCh37/GRCh38 reference panel sequences of the MUC6 VNTR-related haplotypes were imperfectly stitched together (95), and both the reference (Fig. 1) and the alternate contigs (i.e., the standard annotation sequences) are too small, not necessarily representative of any individual, much less a population’s repertoire.

In the Alzheimer’s Disease Sequencing Project (ADSP) whole-exome sequencing (WES) data set (105, 106), the analytic results reflected the difficulty in sequencing. More specifically, MUC6 exon 31 produced low mapping quality scores and a commensurate high percentage of SNV calls that failed to pass bioinformatics pipeline quality control filters. For chromosome 11 overall, 89% of SNV calls passed Baylor University (BAYLOR) and Broad Institute (BROAD) filters as defined previously (1), and only 11% failed to pass one or other filter. In sharp contrast, within MUC6 exon 31, almost two-thirds of SNVs failed both filters (Fig. 6A). Correspondingly, many sequencing reads that were aligned to MUC6 exon 31 had low mapping quality (Fig. 6B). These data help explain why the MUC6 VNTR region was largely excluded from many prior studies.

FIGURE 6.

The MUC6 VNTR is a region with poor mapping quality in high-throughput Alzheimer’s Disease Sequencing Project (ADSP) whole-exome sequencing (WES) data, resulting in undependable sequencing results. Shown are the percentage of single nucleotide variants (SNVs) that passed quality control (QC) filters in Baylor University (BAYLOR) and Broad Institute (BROAD) analyses (A) and mapping quality (MAPQ) score in the AP2A2/MUC6 genomic region (B). The percentages were calculated from variant level-filtering results for the BROAD and BAYLOR pipelines contained in consensus VCF file of ADSP WES data (A). The MAPQ scores were obtained from exome alignment BAM file for subject ID = NA06984 (Utah Residents with Northern and Western European Ancestry) in 1000 Genomes Project Phase 3 (B). Note the poor MUC6 MAPQ scores in the Exon 31 region. This figure was adapted from Katsumata et al (1), with permission.

Among other notable features of the mucin genes is their reported potential to vary, within individuals over time, and between human populations. (It is a challenge to determine whether the detected variability signals were due partly to sequencing mistakes.) For example, mucin genes including MUC6 were on a short list of genes reported to accumulate exomic mutations within individuals during aging (107). In a mucin gene (MUC2) that is near MUC6 on chromosome 11p15.5, the PTS-encoding repeat sequences differed markedly between “healthy Caucasians” and “healthy Asians” (108). Systematic comparisons among diverse groups have not been reported for MUC6, although a high degree of MUC6 VNTR polymorphism has been demonstrated within different populations (19, 99, 109). Variation between diverse populations is a relevant concern because prior studies found that the impacts of specific late-onset AD-risk alleles can vary according to ethnoracial factors (110–114). However, there are abundant scientific and historical reasons to be cautious with research at the nexus of genetics and race, including in dementia research (115).

The Missing/Hidden Heritability Problem in AD Research

A salient question in the field of AD research is: What proportion of risk for developing the AD phenotype is attributable to currently unknown genetic factors? Addressing this involves 2 related questions: (1) What proportion of AD risk is attributable to genetic factors overall? and (2) What proportion of AD risk is explained by known genetic risk factors? There is uncertainty about the answers to all of these questions. Multiple levels of complexity are involved, related to the fields of genomics, neuropathology, epidemiology, and large multicenter studies. Prior reviews and meta-analyses have remarked that the disease-related operationalizations (diagnostic criteria and thresholds/cut-points applied) and also the results of prior AD studies “vary greatly” (116, 117). For all the unanswered questions, there are grounds to hypothesize that a substantial proportion of risk for developing the AD phenotype is due to currently unknown genetic factors—this is the basis for positing a “missing/hidden heritability problem” in AD (118–120).

As to the degree to which AD is explained by genetic variation, there is not a clear consensus from the literature. In a large Swedish twin study, heritability was estimated to explain 58%–79% of AD cases (119). Another group reported that genetic differences explained 53% of phenotypic variance of AD (121). It is also possible that heritability varies across AD subtypes. It was reported that attributable risk due to genetics for episodic memory loss and working memory loss in elderly subjects was approximately 60% and 70%, respectively (122). A study of dysexecutive AD symptoms estimated the heritability of dysexecutive clinical spectrum symptoms (a quantitative measure) to be 68% (123). Different approaches have been used to estimate the impact of specific genetic factors on AD neuropathologic subtypes (124, 125). We emphasize that the heritability of AD is challenging to estimate because data are required from diverse populations (each estimate of heritability may be unique to a particular population, perhaps related to that group’s environment [126]), with neuropathologic corroboration, and there are many potential confounders.

The known genetic risk factor with the largest impact on public health is the APOE ε4 allele. The AD-associated APOE ε4 allele is present in approximately one-fourth of persons, with gene frequency varying among different human populations and greatly enriched in many AD study cohorts (127–129). The APOE ε4 allele was initially demonstrated to be associated with the AD phenotype by testing relatively small groups: APOE status was compared between 95 AD cases and 139 “unaffected” controls (130).

Extensive GWAS were performed subsequently. Recent AD GWAS sample sizes were large, and imputation approaches are constantly improving (64, 73, 131, 132). These advances enabled the identification of dozens of individual non-APOE risk alleles, including common variants with subtle effect sizes and rare alleles with larger effects (60, 64, 73, 133–136). Collectively, and combinatorially, these gene variants may have large impact.

Despite progress in the field, major uncertainties persist. For example, it is not known what proportion of the non-APOE-related AD risk is conferred via genetic mechanisms. Determining the scope of this residual genetic risk requires community-based research cohorts because hospital- and clinic-based samples have strong recruitment biases (137–140). In a high-quality community-based study (Rush University Medical Center cohorts), among n = 1017 autopsied subjects, AD pathology was observed in 62.9%, whereas among the same 1017 individuals, only 25.6% had the APOE ε4 allele (125). In an analysis of data from the population-based Rotterdam Study, the proportion of incident dementia deemed attributable to the APOE ε4 allele was estimated to be 20% (141). These data indicate that a large amount of AD risk is not explained by APOE.

Assessing cumulative genetic risk for AD in an individual or a population requires integration of complex genotype data (64, 131, 132). AD polygenic risk scores (PRS) incorporate information about many gene variants to predict an individual’s risk of expressing the AD phenotype. PRS explained over 80% of AD risk in selected samples (133, 142–147). Although these studies provided valuable insights into AD genetics, few of these studies had population-representative samples. Some of the prior published papers incorporated data on the variation of thousands of putative risk alleles in the PRS using an unfiltered study design, which may theoretically have included proxies from the MUC6 region along with many other polymorphisms. Studies that generated a PRS using a predetermined set of well-validated AD-linked SNPs (148–151) found that predicting AD risk using known risk alleles may be useful in various settings (152), but these studies also indicated that there are yet-uncharacterized AD risk factors because a substantial proportion of AD risk was not explained by the known gene variants. According to multiple sources, if one assumes that ∼25% of the population harbors the APOE ε4 genotype, then the known common risk variants identified by GWAS explained only 30%–50% of total phenotypic variance of late-onset AD (118, 119, 121, 153–158).

To summarize the current authors’ perspective regarding genetic influence on late-onset AD risk: (1) It is currently not known how much of AD risk overall is attributable to genetic factors, but it is often estimated that genetic factors are responsible for ∼50%–80% of AD risk and (2) It also is not known how much of AD risk is attributable to known AD-linked risk alleles, but it may be as low as 30%–50%. With this degree of uncertainty, there are many possibilities, including that there may be currently unknown environmental factors that strongly influence AD risk. Yet as is relevant to the current review article, prior studies also indicate that high-impact AD-associated risk factors may remain to be characterized.

Where Might the Missing Heritability of AD Be Found?

Although increasing GWAS sample sizes and identifying rare disease-associated alleles are potential strategies for filling in the missing/hidden heritability (159–163), the challenges go deeper. Surveying the literature provide indications of the need to develop new approaches for genotype/phenotype correlations. Choosing one example from among many: A relatively easily measured human trait is height. Twin studies indicate that height is highly heritable; ∼80% of human height is due to genetic factors (164–166). A recent large GWAS (n = 253 288 included individuals) identified 697 gene variants associated with human height at genome-wide statistical significance (167). These data increased the explanation of the genetic influences on human height from prior studies (168), yet, collectively, these 697 gene variants still explained only one-fifth of the phenotypic (height) variance in the sample (167). Even taking into account “all common variants” only “captured 60% of heritability” (167). Thus, despite large sample sizes and improving imputation methods, some phenotypes may not be fully explained by readily characterized small genetic variants.

One reason for the persistent knowledge gaps is that much of the human genome has remained practically off-limits to existing high-throughput genome-wide sequencing analyses. Related challenges include coping with large structural variants, genomic regions with substantial interindividual variation, and incomplete annotation of the human genome. Further, genomic regions with repetitive sequences were trimmed off by the bioinformatics pipelines used to identify variants in both GWAS and “next-generation sequencing” (NGS) data (169–171).

A recent, directly relevant study by Ebbert et al analyzed ADSP NGS data systematically and reported that there are “many regions with few mappable reads that we call dark by depth, and others that have ambiguous alignment, called camouflaged. … The number of genes affected by dark and camouflaged regions was surprisingly high. We identified 36 794 total dark regions across 6054 gene bodies, 3804 of which were protein coding genes…” (171).

The discovery of extremely widespread “dark or camouflaged” regions in human genomes has profound implications. This may be particularly true for studies that correlate NGS data with neurologic disease endophenotypes. Even if one hypothesized that genomic regions which can be reliably sequenced are more likely to be disease-driving than the many regions that confound current sequencing technologies, it would still be quite credible that the extensive “dark or camouflaged” regions contain genetic variants which influence human phenotypes. However, for dozens of different human neurologic diseases, a specific tandem repeat-rich genomic region was the established cause of the disease phenotype (172–174). This tendency makes it all the more likely that the missing/hidden heritability for AD could be partly explained by analyzing the widespread, and hitherto poorly characterized, repeat-rich segments of the human genome (83).

The MUC6/AP2A2 Locus and the AD Endophenotype: Data and Cautionary Considerations

The MUC6 locus was initially identified to contain a candidate AD-associated polymorphism through analyses of ADSP WES data (n = 5142 AD cases and 4889 control subjects of European ancestry were included) (1). We employed an optimal sequence kernel association test (SKAT-O) (175) to survey genes for SNVs associated with the AD phenotype (1). The results of that gene-based analysis are shown in Figure 7, using data charted in “Manhattan plot” format. Several aspects of these findings can be appreciated: 1> The reported association between MUC6 gene variation and the AD phenotype was strong (p < 10⁻²⁵); 2> The other genes associated significantly with the AD phenotype in this analysis agreed with prior published studies (TREM2 [176, 177] on chromosome 6 and BCAM [178], which is located near APOE on chromosome 19); and, 3> There was not a haphazard pattern of nominally positive “hits” that would indicate a systematic tendency toward spurious false-positivity.

FIGURE 7.

Manhattan plot of gene-based analysis of Alzheimer’s Disease Sequencing Project (ADSP) whole-exome sequencing (WES) data, showing association with Alzheimer’s disease phenotype in that data set. Optimal-unified sequence kernel association test (SKAT-O) association results are depicted for n = 10 031 (5142 AD cases and 4889 controls) subjects. The gene with lowest p value in this genome-wide scan was mucin 6 (MUC6), with p = 1.38 × 10⁻²⁶, and the second lowest p value was triggering receptor expressed on myeloid cells 2 (TREM2). Relatively rare alleles in both TREM2 and basal cell adhesion molecule (BCAM), which is located near APOE on chromosome 19, were both previously associated with the AD phenotype (106, 176, 178). This figure was adapted from Katsumata et al (1), with permission.

With regard to the hypothesis that MUC6 contains an AD-associated polymorphism, there were reasons to be skeptical. Each of the MUC6 SNVs that was initially found to be associated with the AD phenotype was located within the same unique, poorly annotated, and highly polymorphic VNTR region (1). It was previously demonstrated that the MUC6 VNTR region is prone to producing false-positive results in WES data because of the extremely high local polymorphism (109, 179). One group suggested that MUC6 belongs on a “blacklist” of genes that should be viewed with extraordinary skepticism when predicting pathogenicity based on WES results, because of this “excess heterogeneity” (180). The very low nominal p value for the MUC6/AD association (Fig. 7) could be interpreted as another indicator that technical bias influenced the results. As stated above, it is possible that the MUC6 polymorphism could be variable between populations of different ancestry. This needs to be evaluated carefully, because population structure could also introduce false-positive associations. A consideration of these factors makes it seem like a credible hypothesis that MUC6 genotype/phenotype associations reported by outside authors (e.g., the association between a MUC6 VNTR gene variant and brain herpesvirus infection [30]) could be false-positives.

Despite these warning signals, we assert that MUC6 should not be universally or compulsorily ignored. Merely because a gene contains many sequence elements that are highly variable in a population does not mean that all of those genetic variants are equally benign. As specific examples, other genes that are hypothesized to be pathogenetic (e.g., HLA-DRB1 in multiple sclerosis and AD [181–184] and FIG4 in amyotrophic lateral sclerosis [185–187]), also were “blacklisted” for extreme heterozygosity, as MUC6 was (180). It may be argued that MUC6 deserves added attention because the gene contains a highly polymorphic VNTR region, which has been credibly linked to human diseases (see above). Moreover, thousands of published WES studies did not produce a signal at the MUC6 locus. The prior literature does, however, indicate that an abundance of caution is merited when studying the association between MUC6 polymorphism and human disease.

Concerns have also been raised about potential sources of systematic bias in the ADSP WES data set that we used to identify the MUC6 genotype/AD phenotype association. The MUC6 SNV calls were derived using bioinformatics pipelines affected by many potential confounders (106, 188). In 2016, the National Institute on Aging Genetics of Alzheimer’s Disease Data Storage Site (NIAGADS) posted a topical document titled “Review and Proposed Actions for False-Positive Association Results in ADSP Case-Control Data”; the entire document can be read online (189). Although not mentioning MUC6 specifically, this posting did refer to the ADSP WES data set version that we had analyzed (phs000572.v6.p4), and stated that “differential effects between cases and controls [led] to strong but spurious associations”. Among the notes of caution were: “Different capture protocols were used at the different sequencing centers, and these … could result in biases.” And: “Association analysis of case-control genotypes with AD with covariate adjustment for time period of sample sequence processing at the Broad Institute appeared to eliminate most potentially spurious associations.” This analysis was complemented by “Sanger sequencing … . to validate heterozygous calls” (189). (The reported Sanger sequencing experiments focused on SNV calls but not on tandem repeat expansion.) Ensuing published work that evaluated AD-linked rare variants in the ADSP WES data set did not mention MUC6 (171, 176–178).

Consistent with the NIAGADS cautionary posting, we reported experimental evidence that some of the individual MUC6 SNV calls that had initially focused our interest in the MUC6 VNTR region were not correct (1). For a small sample of cases, we generated SNV calls via a direct polymerase chain reaction (PCR)-based assay, and the results were different from the SNV calls for the same cases according to the initial ADSP data (1). Based on these findings, if there was a true positive signal of pathogenicity at the MUC6 locus, it likely was not a scenario where individual disease-causing gene variants were identified.

We conclude that the phenomenon of the MUC6 gene variants’ status being associated with the AD phenotype (Fig. 7) may have been influenced by many factors. As described above, there are reasons to be skeptical: The MUC6 region’s extreme heterozygosity raises the likelihood of a false-positive signal; specific SNV calls in the early ADSP data set were incorrect; and, the initial ADSP analyses were systematically biased. However, the full impact of that bias was not determined, and we underscore the critical distinction between the SNVs which were shown to be problematic and/or mis-called, versus the tandem repeat expansion of MUC6 that was not addressed by any other workers in the dementia field. The known data are compatible with a testable hypothesis: There was a meaningful underlying biological association between the number of MUC6 tandem repeats and the AD phenotype. Put another way, the early analyses of WES data may have provided a proxy flag, signaling that MUC6 tandem repeat expansion is associated with AD risk.

Additional Studies, Preliminary Assessment of Public Health Impact, and Knowledge Gaps

Following the results of ADSP WES data analyses, we used experimental tools tailored specifically for studying the MUC6 VNTR region: The focus was on the association between AD pathology and MUC6 tandem repeat expansion, rather than on SNV-type genetic variation. The study sample was the UK-ADC autopsy cohort (190) as described in detail (1). Consistent with prior studies (19, 98), analyzing the autopsied subjects’ genomic DNA with PCR revealed a wide range of MUC6 VNTR sizes. In a Discovery subsample (n = 119) of autopsied subjects, the size of the MUC6 VNTR region was associated with the severity of phospho-tau pathology in neocortical brain regions (Fig. 8). For the risk allele, we used as a threshold criterion the largest 33%ile MUC6 VNTR region size (top tertile) as measured by PCR, because this was the cut-point in the Discovery cohort with the strongest association between the MUC6 VNTR region size and phospho-tau pathologic severity (1). That main finding was reproduced in a separate Replication cohort (n = 173).

FIGURE 8.

Experiments were performed to test the association(s) between Alzheimer’s disease phospho-Tau (pTau) pathology, detected with immunohistochemistry (A–C), and the size of the MUC6 VNTR as measured using a PCR-based assay (D). Digital quantification of pTau pathologic burden was used for correlation with the size of the MUC6 VNTR polymorphism. Panels (A–C) show representative results of pTau immunohistochemistry (IHC) from the temporal neocortex (Brodmann areas 21/22) of a subject with dementia. The actual pTau immunostain is shown in panel (A), and at higher magnification in panel (B). The false-colored data (C) conveys the quantified pTau pathology (dark red). Parameters from different brain regions were analyzed to compare the amount of pTau pathology in subjects stratified by the size of the largest detected MUC6 VNTR region according to PCR (D). PCR results are shown for 5 representative cases; those with one amplicon band on the gel are hypothesized to be homozygous, 2 bands indicate a heterozygote. The slower running (higher on the gel) samples represent the larger alleles. Results are summarized (E). The reason we compared pTau pathology in temporal neocortex in the Replication cohort cases with the largest 33%ile amplicons as measured by PCR (top tertile), to the remaining two-thirds of cases, is that this is the region and threshold in the Discovery cohort with the strongest association between the MUC6 VNTR region size and pTau pathology (1). This figure was adapted from Katsumata et al (1), with permission.

None of our prior published analyses combined the abovementioned Discovery and Replication cohorts from the UK-ADC data set. We here show results of analyses factoring in all of these subjects together. The criterion for operationalizing the MUC6 risk allele was again that the size of the VNTR region was in the top tertile in that cohort according to the PCR assay. The summary data were analyzed to compare parameters related to AD neuropathologic changes using conventional stage-based AD pathology severity metrics.

Tables 3 and 4 show the results for comparisons of AD neuropathology in the UK-ADC combined cohort, stratifying by APOE ε4 allele status and MUC6 VNTR allele status. Table 3 shows how the MUC6 VNTR and APOE ε4 allele genotypes were correlated with the number of individuals in specific Braak NFT stage (191) groups. There were a total of n = 101 cases with Braak NFT stage VI (which indicates the most severe AD pathology) in this analysis. The assessment of MUC6 VNTR risk allele status reduced the number of severe AD cases unexplained by a risk allele—the number of MUC6−/APOE− cases with Braak NFT stage VI was only n = 15. A test using analysis of deviance for a logistic regression model of Braak NFT stage VI, regressed on MUC6 VNTR status, APOE ε4 presence, and their interaction, exhibits high statistical significance (LRT-p = 1.08 × 10⁻⁷). A reciprocal tendency for enrichment of MUC6−/APOE− cases can be observed among subjects with relatively low neurofibrillary pathology, Braak NFT stages 0–III. Table 4 shows the results for neuritic amyloid plaques, graded according to the Consortium to Establish a Registry for Alzheimer’s Disease (CERAD) neuritic plaque density scores (192). Among the subjects who lacked the APOE ε4 allele, carrying the MUC6 VNTR risk allele correlated with having highest grade density of neuritic amyloid plaques. The same general trend was present among those with the APOE ε4 allele. This is the first demonstration of amyloid plaque-associated pathology correlated with the MUC6 VNTR genotype.

TABLE 3.

Braak Neurofibrillary Tangle (NFT) Stage Groups in the UK-ADC Autopsy Cohort (1) (n = 292^*), Stratifying by APOE ε4 Status (APOE4+/−) and by Whether or Not the MUC6 VNTR Region Is in the Top Size Tertile (MUC6+/−)

Allele Status†	Total n, Each Group	n, Braak NFT Stage VI Cases	% Braak NFT Stage VI Cases	% Braak NFT Stages 0–III Cases
APOE4−/MUC6−	n=104	n=15	14.4	55.8
APOE4−/MUC6+	n=51	n=22	43.1	31.4
APOE4+/MUC6−	n=71	n=32	45.1	23.9
APOE4+/MUC6+	n=60	n=32	53.3	16.7

^*n = 6 had missing Braak NFT stage.

^†“+” indicates presence of at least 1 putative risk allele.

TABLE 4.

Association Between CERAD Neuritic Amyloid Plaque Density Scores and PCR-Measured MUC6 VNTR Region Size Stratified by APOE ε4 Allele Status in the UK-ADC Cohort (1); n = 292

	Numbers of Subjects by Category		p Value*
	Bottom 2 Tertiles VNTR Size	Top Tertile VNTR Size
APOE ε4 (−): n = 159
CERAD none, possible, or probable	n = 72	n = 22	0.0025
CERAD definite (NP density highest)	n = 34	n = 31
APOE ε4 (+): n = 133
CERAD none, possible, or probable	n = 26	n = 13	0.12
CERAD definite (NP density highest)	n = 47	n = 47

^*Chi-squared test result, two-tailed test.

CERAD, Consortium to Establish a Registry for Alzheimer’s Disease (192).

Published data (1) and a literature review provide support for hypotheses that link the MUC6 VNTR region polymorphism with AD pathogenesis (Fig. 9). The expanded version of the MUC6 VNTR region was associated with decreased AP2A2 expression (1). MUC6 VNTR polymorphism thus may affect CCV function, which is related to autophagy and functions in processing neurodegenerative disease-associated proteins such as APP and Tau (41, 42, 75). Alternatively, or in parallel, the combination of impacts on MUC6 and AP2A2 could influence the gut microbiome and/or viral immunity, therefore affecting pathogenetic cascades relevant to AD. The long-term effects of the variation of the MUC6 VNTR region may cause or accentuate pathologies, including tau neurofibrillary degeneration and neuritic amyloid plaques, in aged persons’ brains.

FIGURE 9.

Hypotheses related to pathogenetic mechanisms: a variable number of tandem repeat (VNTR) region within the MUC6 gene on Chr. 11p constitutes a common driver of the Alzheimer’s disease (AD) phenotype. The large-sized VNTR (i.e., the risk allele, which is ∼30% prevalent in our preliminary studies) is associated with the severity of AD-type pathology, particularly tau neurofibrillary pathology. The MUC6 VNTR region is immediately downstream of the AP2A2 gene. The AP2A2 gene product is an “adaptor” protein involved in clathrin-coated vesicle (CCV) formation at the plasma membrane (33, 35, 38). We found that AP2A2 expression is dysregulated in persons with the large-sized VNTR region (1). There are also lines of evidence connecting MUC6 with the gut microbiome and viral immunity as discussed in the text of the present review; these inflammatory influences may contribute to altered risk of AD.

It is premature to make generalizations about the influence of the MUC6 VNTR expansion on public health. For now, we can only analyze the relatively small UK-ADC autopsy cohort sample. We performed preliminary calculations of the AD odds ratio (OR), using the combined sample from the UK-ADC autopsy cohort (n = 292); see Figure 10. The MUC6 VNTR risk allele is ∼35% prevalent (top tertile) in this cohort, and the MUC6 VNTR risk allele OR for AD is 2.10 (95% CI 1.07–4.18) using the severity of temporal neocortex NFT pathology as the endophenotype, compared to APOE ε4 OR = 2.87 (95% CI 1.55–5.58) for the same endophenotype and in the same cohort.

FIGURE 10.

Preliminary estimations of public health impact of the MUC6 VNTR polymorphism: risk allele prevalence and Alzheimer’s disease (AD) “odds ratio” (OR). According to our preliminary data in the University of Kentucky AD Center cohort, the AD OR for the top-tertile risk allele is 2.10 (95% CI 1.07–4.18) using Tau pathology in the temporal neocortex (Brodmann areas 21/22) as the endophenotype, compared to APOE ε4 OR = 2.87 (95% CI 1.55–5.58). In prior work, other non-APOE common late-onset AD-risk alleles’ OR were <1.5 (60). The UK-ADC cohort is a relatively small sample that lacks socioeconomic or ethnoracial diversity. Additional studies are required to better understand the public health impact of the MUC6 VNTR polymorphism.

Among tested subjects in the UK-ADC sample, persons with the APOE ε4 allele were relatively likely to also have the MUC6 VNTR risk allele (Table 3): 32.8% of APOE ε4(−) subjects had the MUC6 VNTR risk allele, whereas 45.8% of the APOE ε4(+) subjects had the MUC6 VNTR risk allele. Although this was only a marginal trend (p < 0.03) in this small cohort, this phenomenon is notable. Primarily, to evaluate the impact of the MUC6 VNTR risk allele appropriately, we factored each subject’s APOE ε4 allele status into our analytic model (1), so as not to overestimate the associative impact of the MUC6 VNTR risk allele.

The UK-ADC sample’s enrichment of subjects with both the MUC6 VNTR risk allele and (in the same persons) the APOE ε4 allele may have other implications. As noted previously, AD-related study samples are often influenced by recruitment bias to include a disproportionately high number of individuals at genetic risk for AD-type dementia. (The allele frequency for the APOE ε4 allele was >45% in this dementia-enriched UK-ADC convenience sample, although the population frequency is ∼25%.) If the MUC6 VNTR risk allele was indeed influencing AD risk, one would hypothesize that there should be a corresponding tendency to recruit high-risk subjects with both the large MUC6 VNTR and the APOE ε4 allele; this bias may not be restricted to the UK-ADC cohort. If this were true, then, an additional hypothesis would follow: Just as an estimate of the impact of the MUC6 VNTR must be adjusted to account for the presence of the APOE ε4 allele in carriers of both risk alleles, so should an estimate of the impact of the APOE ε4 allele be adjusted to correct for the presence of the MUC6 VNTR risk allele. In other words, a portion of AD risk which hitherto has been attributed to APOE, may actually have been due to the MUC6 VNTR allele, because dementia research study subjects may tend to harbor both risk alleles simultaneously due to recruitment bias. Further, persons with the APOE ε4 allele but with a risk mitigating (short) MUC6 VNTR allele may theoretically be less likely to be represented in dementia research cohorts.

Although the data gathered related to the MUC6 VNTR and AD risk are suggestive so far, there are also reasons to be cautious, in addition to those mentioned above. The sample that we assessed was relatively small by genomics study standards. These findings require further corroboration. The UK-ADC autopsy cohort is a relatively homogenous demographic sample, so there is a need for assessment of more diverse and population-representative cohorts. There are other key knowledge gaps that may help guide future studies. For example, we do not know the actual MUC6 VNTR sequences for any of the cases we have analyzed. To characterize the MUC6 VNTR polymorphism, a PCR-based assay was developed which indicated the approximate size of the MUC6 VNTR (Fig. 8D), because we lacked access to a quantitative tandem repeat assay such as is available for C9ORF72 repeat number characterization (193, 194). An improved MUC6 VNTR assay or sequencing platform is required for both research and possible future clinical purposes. At this locus and others, there remains much to be learned about the “dark and camouflaged” (171) regions of the human genome.

Funding included grants P30 AG028383, R01 AG057187, R01 AG042475, R56 AG057191, and U01 AG016976 from the National Institute on Aging/National Institutes of Health (NIH). The Genotype-Tissue Expression (GTEx) project was supported by the Common Fund of the Office of the Director of the NIH, and by NCI, NHGRI, NHLBI, NIDA, NIMH, and NINDS. We are deeply grateful to all of the study participants, clinical workers, and researchers who helped make this article possible. See Supplementary Data for additional acknowledgments.

The authors have no duality or conflicts of interest to declare.