Alzheimer’s disease (AD), the most common form of dementia, is a devastating and incurable condition affecting over 5.2 million individuals in the US alone (Alzheimer’s Association Statistics, 2014). As populations across the globe age, this figure is expected to increase vastly, posing a huge threat to public health (Brookmeyer et al., 2007). With no cure or effective treatments currently available, the burden of the disease will vastly increase, yet it is thought that just delaying the onset of the condition by a few years could significantly decrease the number of people developing the disorder (Brookmeyer et al., 2007). However, with our current limited knowledge of the etiology of the disease, the appropriate targets and interventions remain unclear. Uncovering genetic loci associated with the condition gives clues to the important pathways and mechanisms involved in the disease process, which can translate to novel targets for diagnoses and treatments.

The rare, familial, early onset form of the condition is comparatively well understood, known to be caused by rare damaging mutations inherited in a Mendelian fashion in three genes (APP, PSEN1, and PSEN2) related to amyloid processing, which cause a buildup of amyloid in the brain or shift the ratios of the types of Aβ produced. The late onset form, however, is a complex disorder with multiple genetic and environmental risk factors. Although the heritability of the disease is estimated to be around 60–80% (Gatz et al., 2006), our understanding of the genes and loci involved has been limited until relatively recently.

The APOE locus was identified as a genetic risk factor for late onset AD (LAOD) in the early 1990s via linkage studies in large AD affected families (Pericak-Vance et al., 1991). As a genetic risk factor for a late onset, complex disorder, APOE has an unusually large effect size, which facilitated its discovery by these methods. The risk associated ε4 allele conveys an increase in risk of around 2–3x in heterozygous form, and around 15x in homozygous form, while the ε2 allele is associated with decreased risk (Farrer et al., 1997). Even with such an effect size, the population attributable fraction of APOE’s ε4 allele is estimated to be only 27.3% (Lambert et al., 2013), so a large proportion of the heritability remained unexplained. Despite a wealth of research over the following years, further linkage studies as well as numerous candidate gene investigations failed to find any further genetic risk loci for AD that were robust and replicable (Ertekin-Taner, 2010).

The next major leap forward came about via the advent of the genome wide association study (GWAS). GWAS enable researchers to test for association between a given disease or trait and virtually all loci within the human genome in a single pass. A caveat of this is that with such a large number of tests being conducted, the sample sizes required to give sufficient power are very large. This technique was facilitated by the coming together of a number of crucial factors. Firstly, international collaborative efforts such as the 1000 genomes project, which set out to catalog all common human variation across multiple populations (Abecasis et al., 2010). This meant in turn that genotyping arrays could be developed that captured the vast majority of human variation in a single array. Finally, and again through large collaborations, sample cohorts of a large enough size to give sufficient power to these studies became available.

The results of the first two truly large scale GWAS in AD were published in October 2009, and brought about the first major advance in the AD genetics field in over a decade. The two studies collectively identified three new loci associated with AD risk, in the genes CLU, PICALM, and CR1 (Harold et al., 2009; Lambert et al., 2009). Subsequently, these loci have been extensively replicated and shown to be linked to a variety of AD associated endophenotypes, giving extremely strong evidence for their genuine involvement in the etiology of the condition (Carrasquillo et al., 2010; Corneveaux et al., 2010; Jun et al., 2010; Seshadri et al., 2010). Since then, through ever increasing sample sizes and meta-analyses, a further 17 genes have been reported to be associated with AD via GWAS, most recently in a “mega”-meta-analyses featuring almost 75000 samples (Seshadri et al., 2010; Hollingworth et al., 2011; Naj et al., 2011; Lambert et al., 2013).

While the successes of GWAS in AD are irrefutable, there remain some crucial limitations. Firstly, GWAS can only really detect trait associated common variants – these are the variants included in the widely available genotyping chips used for GWAS to date. Rare variants are poorly targeted by the design. Secondly, although a plethora of loci associated with AD have been identified, the variants underlying these associations remain largely unclear. Often the strongest associated variants within an identified locus fall in the intronic regions of genes, or even in intergenic regions, making it unclear what the actual causative factors in the area are. The early theory that rare, causative mutations were the underlying source of the association being tagged by common GWAS variants have fallen out of favor since in AD as well as numerous other fields of research, extensive resequencing efforts at GWAS loci have proved largely unfruitful (Guerreiro et al., 2010; Hunt et al., 2013). Indeed, where rare, associated variants have been detected within GWAS genes, their effects seem to be independent of the GWAS hit, indicating that different forms of variation within a single locus (e.g., rare coding and common regulatory) can exist and have separate effects on the trait(Bettens et al., 2012).

Aside from not knowing the specific causative factors, the GWAS identified loci individually do not explain a large proportion of disease risk. Each of the loci identified by GWAS affects disease risk only a small amount, so even with APOE and all 20 GWAS genes combined, there still remains a large amount of missing heritability (Manolio et al., 2009). Another limitation is the increasingly large sample sizes required to detect the increasingly subtle effects on disease risk that GWAS loci have. With these ever diminishing returns, the remaining unknown heritability is unlikely to be entirely explained by this approach.

The remaining missing heritability of AD is likely to be complex, and unlikely to be resolvable using a single methodology. While other common variants of low risk undoubtedly remain to be discovered, the likelihood of these explaining the full heritability of AD is low. Non-coding, regulatory variations, rare coding variants of moderate to high effect sizes and epigenetics all likely contribute to AD heritability. Indeed, two articles recently published have presented the findings of the first major epigenome wide association studies in the AD field, and together identified and replicated four loci where epigenetic alterations were associated with AD risk (De Jager et al., 2014; Lunnon et al., 2014), and it is likely more sites where epigenetic modifications relate to AD risk exist (Lord and Cruchaga, 2014).

Although AD is a condition with typical onset far beyond reproductive age, in general, deleterious alleles are more likely to be rare due to the laws of purifying selection (Gibson, 2011), so high impact alleles are unlikely to be detected by GWAS, which target common variants. Furthermore, as rare causative alleles are discovered, they are likely to have a greater explanatory power over the etiology of complex traits than common risk variants do.

Sequencing technologies have seen huge advances in accuracy, throughput, and cost effectiveness in recent years (Cirulli and Goldstein, 2010). Similar to the way in which the development of genotyping arrays facilitated the discovery of common disease associated variants via GWAS, the rapid improvements in sequencing technologies and enrichment strategies over recent years have paved the way for the detection of rare, disease associated variants via exome or whole genome sequencing (WGS).

Despite the huge decrease in sequencing costs, it is still prohibitively expensive for most studies to conduct WGS on large cohorts. WGS is also problematic in terms of data handling, storage, and interpretation. A number of possible strategies are available to circumvent this issue, such as targeted sequencing, exome arrays, exome sequencing, and selection of highly informative subjects, such as members of multiply affected families, specific populations with low heterogeneity, or extremes of phenotypes.

Targeted sequencing, e.g., for loci identified by GWAS, is a cheaper alternative to whole genome or whole exome sequencing (WES), and has had some successes (Rivas et al., 2011). Although it has been attempted for several of the early AD GWAS hits, it proved largely unfruitful, as has been seen to be the case in other complex disorders (Guerreiro et al., 2010; Ferrari et al., 2012; Hunt et al., 2013).

Exome arrays offer a platform for genotyping exonic low frequency and rare variants in a chip based assay, bridging the gap between the sparse genotyping of GWAS and sequencing based study designs. A number of commercially available exome arrays are available, often allowing for the incorporation of custom content to tailor the platform to specific diseases or traits. The heterozygous concordance rate between exome sequencing data and exome array data has been estimated to be 98.14% (Wang et al., 2013). Exome array based studies have been shown to be successful in identifying several loci associated with insulin levels in a cohort of 8229 Finnish non-diabetic individuals (Huyghe et al., 2013). However, there are issues with the technology, and to date, no variants associated with AD have been published using the exome array method. Chung et al. (2014) used the Affymetrix Axiom Exome Genotyping Array in their study of 1005 Korean subjects, but the only significant association signals were found to be due to APOE. One of the reasons for this may be that the selection of variants for exome arrays was mainly based on individuals of European origin, so their utility in other populations may not be great. The sample size utilized here was also significantly smaller than Huyghe et al.’s (2013) design, making it unclear whether the lack of significant associations were due to genetic differences in the sample population compared to the chip design, or a simple lack of power. The inclusion of variants in exome arrays is also dependent on their genomic context, not all are compatible with the array type genotyping design (Do et al., 2012). Exome arrays also do not offer the ability to detect previously unknown variants, which targeted or exome sequencing designs do. Exome sequencing, although only targeting 1–2% of the genome, targets the most commonly disease associated regions (Ng et al., 2009). Additionally, it lessens the issues associated with data handling, processing and storage, as well as being easier to interpret (as bioinformatics programs for prediction of consequences of exonic mutations are generally more advanced than those for non-coding changes), and is also cheaper, enabling larger sample sizes to be utilized than would be possible with WGS.

The power of association tests between single variants and traits decreases as the minor allele frequency (MAF) of the variant does. This means very large samples will be needed for rare variant association studies unless the effect size of the variant is particularly large. It is also not clear at present what the most appropriate way to correct for multiple tests in rare variant studies is. The typical GWAS significance threshold of 5×10^-8 is based on approximately one million independent tests being conducted simultaneously. Although rare variants are unlikely to be completely independent, there are far more rare variants than GWAS genotyped common variants, so the required significance threshold could be even more stringent, making attaining significance in single variant analyses problematic. Evaluating multiple rare variants within a given region (e.g., a gene) can help combat this. A number of methods have emerged over the last few years to enable this. The simplest of these are burden tests, but these do not account for variants in the same test unit having opposing directions of effects, which can cancel out any meaningful signals (Morris and Zeggini, 2010; Lee et al., 2014). Indeed, recent evidence from the Alzheimer’s field suggests many loci do have variants which both increase and decrease disease risk (e.g., the protective alleles recently reported in APP and APOE, both long standing AD risk loci harboring deleterious variants’ Jonsson et al., 2012; Medway et al., 2014). Combined tests, such as SKAT-O can take in to account both risky and protective variants within the same locus, aggregating the individual association signals in to a single combined association score, which can be much more powerful than the single variant approach (Lee et al., 2012, 2014).

Selecting extremes of the phenotypic spectrum has promise for rare variant discovery since such individuals are likely to be enriched for rare disease causing variants. In terms of AD, this could be particularly severe cases or those with early onset vs. cognitively healthy extremely old individuals. It is also possible to use levels of biomarkers, which typically follow normal distributions, and conduct sequencing in the highest and lowest measuring individuals. Using quantitative traits as endophenotypes has been shown to have increased power compared to standard designs, reducing the number of individuals to be sequenced substantially (Li et al., 2011; Benitez et al., 2013b). Another potential way to enrich subjects for genetic risk factors is to utilize individuals with a strong family history of the condition, since pedigrees with multiple affected individuals are more likely to have pre-disposing genetic variants than typical sporadic cases. Indeed, the way in which traditional genetic linkage approaches can be effectively used in conjunction with new technological and analytical methods for the identification of rare, disease causing mutations is currently being explored (Santorico and Edwards, 2014).

It is likely that, as with GWAS, combining several individually smaller studies in to one large meta-analysis will prove an effective way of increasing sample sizes, and may produce findings beyond those of the separate studies alone. At present, however, there is little standardization in the analysis of next generation sequencing data across studies, which could make combining them problematic. Methods will need to be developed to adequately deal with heterogeneity between studies, particularly given the high variance in rare variant detection and quantification across different sequencing platforms and analysis strategies (O’Rawe et al., 2013).

Another strategy for increasing power without affecting costs is to utilize imputation, inferring genotypes in unsequenced individuals to increase effective sample size. However, the accuracy of imputation decreases as MAF decreases, so imputation of rare variants is unreliable at present for unrelated individuals. As more sequencing data, both whole genome and exome, is generated, the reference panels upon which imputations are based will become more extensive, making an improvement in imputation accuracy for rare variants likely in the coming years. Additionally, several methods are being developed to infer the genotypes of rare variants in related individuals, using sequencing data, GWAS data, and pedigree information (Cobat et al., 2014; Saad and Wijsman, 2014).

A further complication is that rare variant studies are more susceptible to cross population differences in allele frequency. Rare variants, which are more likely to be recent in origin, can show vast differences in frequencies across populations (Raska and Zhu, 2011). Whilst this can be leveraged to be advantageous (Hatzikotoulas et al., 2014), it also provides difficulties, since subtle underlying population substructure can inflate false positive rates (Keen-Kim et al., 2006). Although robust methods exist for correcting for such stratification in GWAS data, the same cannot currently be said for rare variant association studies (Liu et al., 2013). These matters will be discussed in the context of rare variant association studies in AD below.

Despite being faced with this plethora of issues, in recent years several research groups have had success in identifying low frequency and rare variants associated with AD, with some variants and genes yielding protective effects, and others increasing AD risk.

One of the major lessons that can be learned from these rare variant association studies in AD so far is that population stratification can have a profound affect both on findings in initial reports and replication. Large sample sizes are required for replication in order to distinguish whether a failure to replicate is due to a lack of power or a lack of association. A lack of association may indicate the initial finding to be a false positive result, or highlight differences in the etiology of the disease between the discovery population and the replication cohort. Replication in both the same and alternative populations is crucial, particularly at the gene rather than variant level, since associated variants can differ between populations.

The overriding aim of all of these studies is to better understand the etiology of AD. Identification of new genes and variants associated with AD will bring new targets for disease diagnostics and treatments. A number of major resequencing efforts are being undertaken in AD, with significant insights into disease biology likely from their results.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.