Exploratory Data from Complete Genomes of Familial Alzheimer Disease Age-at-Onset Outliers

Abstract

Identifying genes that modify the age-at-onset (AAO) of Alzheimer disease and targeting them pharmacologically represent a potential treatment strategy. In this exploratory study, we sequenced the complete genomes of six individuals with familial Alzheimer disease due to the autosomal dominant mutation p.Glu280Ala in PSEN1 (MIM# 104311; NM_000021.3:c.839A>C). The disease and its age-at-onset are highly heritable, motivating our search for genetic variants that modulate AAO. The median AAO of dementia in carriers of the mutant allele is 49 years. Extreme phenotypic outliers for AAO in this genetically isolated population with limited environmental variance are likely to harbor onset-modifying genetic variants. A narrow distribution of AAO in this kindred suggests large effect sizes of genetic determinants of AAO in these outliers. Identity by Descent (IBD) analysis and a combination of bioinformatics filters have suggested several candidate variants for AAO modifiers. Future work and replication studies on these variants may provide mechanistic insights into the etiopathology of Alzheimer disease.

Here we describe exploratory data on the complete genomic sequences of six individuals from a well-characterized kindred with familial Alzheimer disease caused by a fully penetrant autosomal dominant missense mutation encoding p.Glu280Ala in PSEN1 on chromosome 14q24.3 [Lopera et al., 1997]. The gene product of PSEN1, presenilin 1, is a component of the gamma-secretase complex that cleaves amyloid beta (A4) precursor protein (APP; MIM# 104760) into amyloid beta (Aβ), the main component of Alzheimer disease (AD) senile plaques [Glenner and Wong, 1984]. The median age at onset (AAO) of dementia in mutation carriers is 49 years [95% CI 49–50] [Acosta-Baena et al., 2011]. AAO in sporadic AD has a high heritability with a predicted genetic contribution of 42% [Li et al., 2002]. A study in a monogenic early-onset AD population predicted a heritability of AAO of 93% [Marchani et al., 2010]. APOE (MIM# 107741), the major known modifier gene of AAO, can also shift the AAO in carriers of the p.E280A mutation [Pastor et al., 2003], suggesting shared mechanisms in the monogenetic and sporadic forms of the disease.

Extreme outliers in AAO provide an opportunity to identify disease modifying factors. In this exploratory study, four extreme outliers for AAO (two a decade on either side of the mean) were selected for whole genome re-sequencing. Two individuals with the average AAO were also sequenced as controls. This study was approved by Institutional Review Boards at the Universidad de Antioquia where the samples were obtained and informed consent was obtained from all subjects. Additional information about the study participants is provided in Supp. Table S1. The narrow distribution of AAO in carriers of the p.E280A allele suggests the presence of determinants of AAO with large effect sizes in the outliers. Due to common founders and a high occurrence of endogamy and consanguinity, this kindred contains limited genomic variability and consequently a narrowed scope of search for variants that affect phenotype. Similarities in education, socio-economic status, and lifestyle among population members reduce environmental variance thereby increasing the likelihood of modifier genes among the factors that control AAO.

Although insightful, genome-wide association studies (GWAS) remain underpowered to detect association with variants of low frequency or small effect sizes. Many statistically associated variants reside in noncoding regions and lack an obvious biological mechanism of action. Associated single nucleotide polymorphisms (SNPs) may also serve as proxies for the causal SNP due to strong linkage disequilibrium (LD) and thus themselves make no contribution to phenotype. Arrays used in GWAS are typically unable to detect structural variants, which can contribute significantly to disease phenotype. Using whole-genome sequencing, we obtained a comprehensive assessment of genomic variation in each individual and address some of these shortcomings. By combining a variety of bioinformatics analyses, our exploratory investigation highlights a variety of biologically plausible mechanisms of how variants might modulate the Alzheimer AAO. While this approach can capture the scope of genetic variation, it lacks statistical significance and therefore serves as an initial step toward identifying candidate loci.

Sequencing was performed by Complete Genomics using unchained combinatorial probe anchor (cPAL) chemistry [Drmanac et al., 2010]. An average of 192 Gb of sequencing reads was generated for each genome covering over 96% of both the genome and exome at an average depth of >60X for all samples. Reads were aligned to the reference genome (NCBI Build 37) and variant calling was performed by Complete Genomics. Each genome averaged 3.46±.07 million SNPs with a genome-wide average transition to transversion ratio of 2.13, consistent with current estimates from the 1000 Genomes Project Consortium. The union of all genomic variants in the 6 genomes comprised 8.2 million variants including 6.93 million SNPs, 505,000 insertions, 567,000 deletions, and 198,000 substitutions. 1.05 million novel variants absent from dbSNP 132 were observed. 1.16 million of the variants were shared across all 6 genomes and 450±90 thousand variants were private to each genome.

The intersection of variants shared in the early or late onset outlier phenotype pairs will contain the AAO modifier genes common to the outliers. A filter was established to find variants that were 1) homozygous (two copies of the same variant) in both genomes of the extreme phenotype pair and absent or heterozygous in the remaining 4 genomes, or 2) present in the phenotype pair and absent from the other 4 genomes. Variants falling into these categories are hereafter classified as tracking with the disease AAO. To determine which variants are likely to contribute to disease phenotype, we focus primarily on non-synonymous variants (NSVs) that track with the disease AAO. Future work will assess contributions from non protein coding variants such as the rare homozygous variant NR_030326.1: n.29T>C located in the mature sequence of MIR596 and present only in the late-onset phenotype pair. The results of our screening process are summarized in Supp. Table S2. 275 and 151 distinct genes contained NSVs tracking with early and late onset respectively. Seven genes contained both potential risk and protective variants, similar to APOE in which the ε4 allele is a risk factor and the ε2 allele is protective [Corder et al., 1993; Corder et al., 1994], potentially implicating these genes as molecular switches for the onset of the disease (Supp. Table S3).

We used Identity by Descent (IBD) mapping to identify regions of these genomes inherited from a recent common ancestor, which are likely to contain causal alleles shared between individuals with the same phenotype. We expect casual alleles to be rare because common variants tend to have smaller effect sizes [Kryukov et al., 2007] which likely cannot account for our extreme AAO phenotypes. Some casual variants are also likely to have arisen recently or else they would have been removed by purifying selection [Browning and Thompson, 2012]. Variants may have arisen through drift during a recent bottleneck in the founder population. Rare, recent, or founder introduced causal variants are readily detected through IBD mapping.

Using the GERMLINE algorithm [Gusev et al., 2008], we identified IBD segments between all pairs of individuals and visualized these regions using Circos [Krzywinski et al., 2009] (Fig. 1). The parameter used for IBD determination was a minimum region of 3 MB containing fewer than 4 mismatches between two genomes. With about 1 SNP/KB, in the human genome, we expect around 3000 SNPs in the minimum region. The length and number of IBD segments decline exponentially after every generation and are largely undetectable by current algorithms after ~25 generations. Interestingly, the only region of IBD (excluding centromeres) shared in all 6 genomes was located on chromosome 14 and contained the causal PSEN1 p.E280A mutation (Fig. 1A). This result demonstrates the utility of this method in identifying regions likely to contain causal variants. We next identified regions of IBD between the pairs of extreme outliers and the NSVs within these regions that track with AAO. Relationship inference analysis, although underestimating kinship coefficients, suggests that cryptic relatedness did not contribute to IBD sharing between our extreme phenotype pairs. (Supp. Table S4).

Figure 1.

Identity By Descent (IBD) Mapping. IBD segments were determined for each pair of genomes in our study and plotted as tracks within the center of the circle. Labels indicate which two genomes are being compared in each track. The six individuals in the study are labeled Early (E) 1 and 2, Mean (M) 1 and 2, and Late (L) 1 and 2, depending on the age at onset. IBD segments between the two early outliers (Early 1 and Early 2) are shown red. IBD segments between the two late outliers (Late 1 and Late 2) are shown in green. Numbered chromosomes are shown in the outermost track, with centromeres drawn in black. Within the chromosomes, nonsynonymous variants (NSVs) tracking with AAO are plotted as colored circles. A: The only region of IBD shared across all six genomes contains the causal PSEN1 p.Glu280Ala mutation. B: A stretch of IBD unique to the late-onset pair contains 3 rare variants. C: A stretch of IBD unique to the early-onset pair contains 6 nonsynonymous variants in 5 genes.

In the delayed onset pair, a unique stretch of IBD on chromosome 16 spanned 5.3 MB and contained 3 NSVs that track with the disease onset (Fig 1B). One of the variants in CCL22 (Supp. Table S5) was very rare. The other two variants were rare, homozygous and affect the genes GPR56 (MIM# 604110; NM_001145770.1: c.918A>C, p.Gln306His) and CCDC135 (NM_032269.5: c.1298C>T, p.Pro433Leu). The protein encoded by GPR56 is involved in brain cortical patterning. The gene product of CCDC135 is involved in motile cilia and could possibly affect cerebrospinal fluid flow [Yang et al., 2011]. Another region of IBD shared in the late-onset pair fell on chromosome 6 and contained a rare variant in MYO6 (MIM# 600970; NM_004999.3:c.1120T>C, p.Tyr374His) encoding a protein which is highly expressed in the brain and enriched at the postsynaptic density [Osterweil et al., 2005].

In the early-onset phenotype pair, a stretch of IBD on chromosome 1 contained 6 distinct NSVs including several rare variants (Fig. 1C). The six variants affected 5 genes: EPHA10 (Table 1), INPP5B (MIM# 147264; NM_005540.2: c.136G>A, p.Gly46Ser), RLF (MIM# 180610; NM_012421.3: c.5053C>G, p.Gln1685Glu), DEM1 (NM_022774.1: c.515G>T, p.Gly172Val and NM_022774.1: c.343G>A, p.Asp115Asn), and KCNQ4 (MIM# 603537; NM_004700.3: c.1365T>G, p.His455Gln). Many of these genes are functional in neurons. INPP5B is a type 2 5-phosphatase that localizes to the Golgi and is co-localized with the gene product of RABEP1 (MIM# 603616) [Shin et al., 2005]. RABEP1-positive endosomes are one of the earliest pathological hallmarks of AD [Cataldo et al., 2000]. KCNQ4 encodes a potassium voltage gated channel critical in regulating neuronal excitability. Another region of IBD on chromosome 1 contained 3 NSVs in the genes SWT1 (NM_017673.6: c.442A>G, p.Ile148Val), PRG4 (Supp. Table S5), and HMCN1 (MIM# 608548; NM_031935.2: c.7253T>C, p.Ile2418Thr). An IBD segment on chromosome 6 contained two nonsynonymous variants in the early-onset pair. These variants reside within the carnitine transporter SLC22A16 (MIM# 608276; NM_033125.2: c.755T>C. p.Val252Ala) and glutaminyl-tRNA synthase QRSL1 (NM_018292.4: c.32C>T, p.Ala11Val). Carnitine may protect against the toxicity of Aβ [Virmani et al., 2001] thus defects in this transporter could accelerate the onset of the disease.

Table 1.

Nonsynonymous variants potentially implicated in modifying age at onset

Onset	Gene	Variant	Filters	Putative Function*
Early	ACE: Angiotensin I converting enzyme MIM# 106180	NM_000789.3 c.731A>G, p.Tyr244Cys chr17:61557773	SIFT score = 0 MAF = 0.008	Several variants in this gene are associated with Alzheimer disease. ACE degrades Aβ [Hu et al., 2001]
Early	EPHA10: EPH receptor A10 MIM# 611123	NM_173641.2 c.205C>T, p.Arg69Cys chr1:38227722	100% conserved across orthologs IBD in early onset	Neuronal cell communication [Aasheim et al., 2005]
Early	LIF: Leukemia Inhibitory Factor MIM# 159540	NM_002309.3 c.280G>A, p.Ala94Thr chr22:30639969	Conserved across domains/species MAF = 0.010	Cytokine with roles in neuronal differentiation and signaling
Early	MTFP1: Mitochondrial fission process 1 MIM# 610235	NM_001003704.2 c.368A>T, p.Asp123Val chr22:30824621	MAF = 0.023	Loss of function results in caspase cascade activation, apoptosis
Early	NEFM: Neurofilament, medium polypeptide MIM# 162250	NM_005382.2 c.1423G>A, p.Ala475Thr chr8:24774791	MAF = 0.004	Neuron intracellular transport and axoskeletal component
Early	SEC14L2: MIM# 607558	NM_012429.3 c.410A>G, p.His137Arg chr22:30803579	MAF = 0.008	Cholesterol biosynthesis and phospholipid transport
Early	SEC14L4: MIM# 612825	NM_174977.3 c.907C>T, p.Leu303Phe chr22:30887825 c.973G>C, p.Gly325Arg chr22:30887668	90% conserved across species MAF = 0.0082	Biogenesis of Golgi-derived transport vesicles
Late	NPC1: Niemann-Pick disease C1 MIM# 607623	NM_000271.3 c.709C>T, p.Pro237Ser chr18:21140367	MAF = 0.006	Intracellular trafficking of cholesterol, Niemann-Pick disease
Late	RIC8A: Resistance to inhibitors of cholinesterase MIM# 609146	NM_021932.4 c.624_626delCCC, p.Pro208_Pro209? chr11:209898	Amino Acid Deletion	C.elegans RIC8A mutants have defective synaptic transmission [Miller et al., 1996]
Late	SEMA4C: sema domain, Ig domain TM domain, and short cytoplasmic domain 4C MIM# 604462	NM_017789.4 c.2077G>T, c.2078A>T, p.Glu693Leu chr2:97526787	Substitution MAF = 0.018 MAF = 0.018	Axon guidance Cortical neurons expression Interacts with post-synaptic density protein 95 (DLG4, MIM# 602887) [Wu et al., 2009]

After all bioinformatics analyses, a promising list of candidate factors was compiled.

^*Unless otherwise stated, functions are gathered from UniProtKB/Swiss-Prot Function Entries. Please refer to these for relevant citations.

Mutations are reported using standard nomenclature (http://www.hgvs.org/mutnomen/) with reference transcript listed, complementary DNA (cDNA) sequence position numbered such that +1 corresponds to the A of the ATG initiation codon, and a protein sequence derived from and numbered based on the same reference. Chromosome and starting positions are given in GRCH37 coordinates. MAF = minor allele frequency. IBD = identity by descent.

Although we lack the statistical power to detect association between genomic variants and AAO, several bioinformatic tools were implemented to predict their effects on gene function. The most promising results of these combined analyses are presented in Table 1. Nonsense and frameshift variants are likely to cause a loss of protein function (Supp. Table S6). For all NSVs, loss of function in the resulting protein was predicted using SIFT [Ng and Henikoff, 2006] and PolyPhen2 [Adzhubei et al., 2010] which account for evolutionary conservation and structural features. Novel NSVs are likely to have deleterious effects. Only one novel NSV segregated exclusively in the early-onset phenotype pair. Located in C5AR1 (MIM# 113995; NM_001736.3: c.524G>A, p.Arg175Gln), this variant was given the maximum deleterious prediction by PolyPhen2. This position is highly conserved across species and the gene resides in region of linkage with AD. This gene is expressed in hippocampal and cortical neurons and a role in Alzheimer neurodegeneration has been proposed [Farkas et al., 2003], making this variant an extremely promising candidate for follow up studies. Rare variants are more likely to have functional impact than common variants because deleterious variants will be eliminated through negative selection [Ford, 1965]. Variants were sorted by frequency observed in the 1000 Genomes Project Consortium data and rare variants were evaluated as candidate modifiers of AAO (Table 1, Supp. Table S5).

For each NSV tracking with AAO, we determined whether it was previously tested in association with AD. The AlzGene database contains the results from over a thousand candidate gene studies and GWAS related to AD [Bertram et al., 2007]. By identifying variants whose statistical association is already shown, we can see if the phenotypic outliers are enriched for risk genes. Linkage analysis has identified chromosomal regions likely to harbor the causal variants in AD, but the underlying variants often remain unknown. Having already identified almost all the genetic variation within our sequenced individuals, we can then determine the variants within known linkage regions. Putative AD linkage regions were obtained from genome-wide linkage meta-analyses [Butler et al., 2009] and allow us to take advantage of prior studies to evaluate the potential effects of our variants. Through this analysis, we found that both early-onset individuals are homozygous for a variant in MTHFR (MIM# 607093; NM_005957.4: c.665C>T) which encodes a p.Ala222V substitution in methylene tetrahydrofolate reductase that reduces the enzyme's activity by 70% and confers an Alzgene meta-analysis odds ratio of 1.13. This variant results in elevated homocysteine levels, which is a reported risk factor for AD [Frosst et al., 1995; Seshadri et al., 2002]. The delayed onset pair shared a variant in APH1B (MIM# 607630; NM_031301.3: c.651T>G) encoding a p.Phe217Leu substitution. The gene product of APH1B is a member of the gamma-secretase complex. Alzgene meta-analysis provides an OR of 1.52 for carriers of this allele. In our sample, this allele seems to work in the opposite direction, possibly by mitigating the effects of the deleterious PSEN1 mutation.

We next sought to identify whether individual genes contained multiple NSVs, which should increase the likelihood of a loss of function. Roughly half of the genes whose variants co-segregate with AAO contained multiple NSVs, with a third of these co-segregating with AAO. The number of NSVs that track with onset per gene are shown in Supp. Figure S1. From this analysis, we observed two missense mutations in the phospolipid-transporting ATPase encoded by ATP10A (MIM# 605855; NM_024490.3: c.2350G>A and c.3516G>C) that co-segregate exclusively in the early-onset genomes and have MAF < 0.10. The resulting substitutions p.Ala784Thr and p.Trp1172Cys both affect highly conserved amino acids.

A problem with studying the genetics of complex diseases is the difficulty in identifying gene-gene interactions. Current statistical methods are severely underpowered to detect all but the strongest epistatic interactions. We employed a simple but novel paralog analysis to begin to address this issue. Through gene duplication, functionally similar paralogs may compensate for reduced expression of one another. By examining variation within paralogous gene sets, we may be able to enhance our predictions of the effects of a given variant. A list of paralogs for each of the genes containing NSVs that track with disease onset was generated from HomoloGene, Ensembl, and Pseudogene annotations. We then identified all the NSVs contained in these paralogs and in which genomes these variants were found. Our initial list of genes whose variants track with AAO contained 15 sets (2 or more genes per set) of paralogous genes. About half of the genes whose variants track with AAO have paralogs that also contain NSVs. Many genes located in regions in linkage with AD have paralogs that are also in regions of linkage or have variants that are extremely rare. For example, in the late onset genomes, UNC13D (MIM# 608897) only has only 1 paralog BAIAP3 (MIM# 604009) encoding a brain-specific angiogenesis inhibitor. UNC13D lies in a region statistically linked with AD and BAIAP3 contains a rare (MAF = 0.0544) homozygous variant (NM_001199096.1: c.1532A>C, p.Asp511Ala). In the early-onset pair, there were homozygous NSVs in BICC1 (MIM# 614295; NM_001080512.1: c.2827T>C, p.Ser943Pro) and its only paralog ANKS6 (NM_173551.3: c.1930G>A, p.Val644Ile). Both of these genes are in regions with high LOD scores for AD linkage. These observations suggest the hypothesis that variants in both paralogs might be required to have a functional effect, with neither variant significantly associated with AD on its own. Furthermore, deleterious variants in genes that lack redundant paralogs may be more likely to have a cellular phenotype (Supp. Table S7).

In conclusion, by combining several bioinformatic approaches and incorporating a priori biological information into our exploratory analysis of relatively few complete genomes, we can identify the scope of possible variants putatively related to the disease phenotype. Unlike GWAS, our analysis can highlight novel and rare variants and, by primarily focusing on NSVs, provide readily interpretable biological mechanisms as shown in the IBD analysis. Future work and replication studies on these variants may provide mechanistic insights into the etiopathology of Alzheimer disease. As the number of sequenced genomes increases, the regions of shared IBD between phenotypic groups will shrink, further narrowing our search scope for causal variants. Identification of modifier loci for phenotypic variation in an extended genetically isolated pedigree affected with a monogenic form of Alzheimer disease could apply to the more common sporadic form of the disease and point the way toward much needed novel pharmacologic targets.