GWA study can be of a case-control design of unrelated individuals with the disease phenotype of interest (cases) and unaffected individuals (controls) or of a quantitative phenotype [26]. The key feature of these studies is that they are based on association analysis and therefore require unrelated individuals (as opposed to genetic linkage studies in which families are required).

Dense genotyping arrays containing hundreds of thousands to millions of markers have been developed for GWA studies. Two major commercial sources of genotyping arrays are Illumina and Affymetrix and they offer a variety of specific array platforms that should be chosen based on study population and hypothesis of the project. Selection of markers that are included on the genotyping arrays has relied largely of HapMap [27] and more recently 1000 Genome [28] projects. Specific content has been developed to provide a variety of array platforms that range from basic coverage to make large-scale projects affordable (OmniExpress Bead Chip, for example) to providing the most comprehensive coverage across multiple populations (Omni2.5 Bead Chip for example that was designed using 1000 Genome data).

Analysis of common genetic variants is performed in three steps: [1] testing for departures from Hardy–Weinberg Equilibrium (HWE) proportions; [2] estimate of variant–variant LD to assess the genetic structure of the cohort [24]; and [3] association analysis under a specific genetic model (additive, recessive, or dominant) via logistic regression. Logistic regression models usually include demographic covariates such as age and gender. Instead of using basic race/ethnicity categories from self-reports, principal components (PCs) from the genome-wide genetic variant data are also included in the regression model. This is done because even among a sample of a single racial/ethnic group such as nonhispanic white individuals, consideration of the effects of population stratification is very important. Significance level in GWA studies is most commonly assessed using the concept of genome-wide significance, which uses an estimated number of independent genomic regions based on the distribution of LD in the genome for a specific population; for European populations, this is approximately p < 5 × 10⁻⁸ [17].

One of the fundamental concepts in the analysis of GWA studies is that of meta-analysis. In this approach, findings from multiple studies can be combined to increase the power of identification of significant genetic variants. An essential principle in meta-analysis is that all studies included examined the same hypothesis and that the original analysis was performed in an almost identical fashion [29]; when this is not possible, statistical approaches to assess heterogeneity of analysis methods are used [26]. Software package METAL is often used for meta-analysis of GWA studies [30] and detailed protocols for this type of study have been published [31].

Imputation analysis of genotype data is often performed in GWA studies for two reasons: [1] to identify additional genetic risk loci, and [2] to provide information on the same set of variants and thus facilitate meta-analysis. In this approach, a reference panel such as Hapmap or 1000 Genome data are used in conjunction with genotype data to infer genotypes that were not directly measured on the genotyping array using haplotypes (groups of variants that are inherited together, or are in high LD, in a specific population). The most important consideration in imputation is to use the reference population that is best matched to the study cohort. In practice, study sample haplotypes may match multiple reference haplotype and imputation assigns probability of the presence of specific allele(s). This analysis is performed using software such as IMPUTE [32] or MaCH [33].

While tools for functional annotation of genetic variants can be applied to most significant hits from the GWAS analysis, additional fine mapping by denser genotyping or sequencing is generally required to identify variants that are likely to have functional consequences. Functional annotation strategies are discussed in the next section.

Whole genome, whole exome, and targeted sequencing projects are mostly designed to identify rare genetic variants but they will also capture information on common variation. Whole genome sequencing (WGS) studies capture variation in both coding and noncoding variants while whole exome sequencing (WES) studies are focused on coding variants (although newer target enrichment strategies capture UTRs and some promoter and intronic sequence). WES is ~10 fold less expensive than WGS thus allowing for a larger number of samples to be profiled, and the data produced are less complex to analyze. Targeted resequencing studies are well suited for following up on GWAS or linkage loci and are generally designed to capture the entire locus, including both coding and noncoding regions. Loci with previously associated common variants through GWAS can be fine mapped and are more likely to contain functional rare variants and therefore may be the most fruitful application of next generation sequencing to complex disease. It was initially proposed that two affordable strategies for identification of disease-causing variants were [1] sequencing of affected individuals in a pedigree followed by genotyping of candidate variants to demonstrate co-segregation with disease in the family and [2] extreme-trait sequencing of a small number of individuals at the tails of the trait distribution followed by targeted sequencing or genotyping in a larger cohort [35, 36]. As the cost of sequencing has decreased, other study designs have been adopted. Statistical considerations for the design of rare variant association studies are discussed in detail elsewhere [37].

The first sequences of the human genome published a decade ago [38, 39] were accomplished using automated Sanger sequencing with dideoxy chain termination [40] at a cost of approximately $2.7 billion to produce a draft sequence of the human genome (http://​www.​genome.​gov/​11006943). In contrast, today a human genome can be sequenced at 20× coverage for roughly $2500 using next generation sequencing (NGS) technology. NGS refers to a group of strategies that rely on a combination of template preparation, sequencing and imaging, and genome alignment and assembly methods, producing gigabases (GB) of sequence data per run in the form of short reads [41]. The Illumina sequencing platform is the most commonly used at the present time and its read length is currently capped at 2 × 125 bp. Single molecule technologies such as from Pacific Biosciences do not require the clonal amplification of molecules to be sequenced, but rather a single DNA molecule is sequenced by synthesis using a DNA polymerase. Single molecule sequencing technologies promise a much simpler sample preparation and offer longer read lengths but generally have lower accuracy than short-read sequencers, which makes them less desirable for rare variant studies. However, their long read lengths are useful in the regions of the genome that are difficult to assemble using short reads, with highly repetitive genomic regions being the best example. A combination of short and long read length technologies provides the best solution in some cases. Single molecule technology is reviewed elsewhere [42].

Another important aspect of sequencing exomes or genomic regions of interest, such as those identified by GWA studies, is target capture. Large genomic regions are most efficiently and cost-effectively captured using hybridization-based approaches such as Agilent SureSelect and Nimblegen SeqCap while PCR-based approaches offered by Life Technologies AmpliSeq and Agilent Haloplex are more suited for capture of smaller regions [43].

Four main steps in the analysis of sequencing data are [1] base calling and sequence assembly [2], variant detection [3], statistical analysis to identify significant associations, and [4] functional analysis of the identified variants. Several recent reviews have detailed discussion of these steps [36, 44, 45] with the most important points briefly summarized here. Data analysis workflow begins with the base calling and alignment of sequence data to the reference genome [46]. Base-calling procedures generate per-base quality values (QVs) that are typically converted to Phred-like quality score [47]. Most alignment software provides run metrics that allow the user to assess quality of sequence data; these include number of raw reads, number of mapped reads, number of unique reads, sequence coverage, and coverage for and percent of “on target” reads for targeted approached [44]. The quality controls metrics allow the researchers to determine potential experimental and alignment biases and remove low quality and poorly mapping reads from further analysis.

The high quality, aligned reads are then analyzed to identify DNA sequence variants, most commonly single nucleotide variants (SNVs); information on structural variants (SVs) and copy number variants (CNVs) can also be obtained. Multiple software options exist for variant calling with both those that perform variant calls in individual samples and those that use multiple samples to call variants [45, 47]; commonly used pipelines include GATK [48] and SAMTools [49], among others. Single marker statistical analysis of common variants (>5 %MAF) from sequence data is identical to that used for GWA studies. On the other hand, most study populations will be underpowered to conduct single marker tests of association for rare variants (0.1–1 % range of MAF) despite the expectation of high effect sizes for rare risk alleles (relative risks ~2.5–5.0). The general approach taken is therefore to test for association between disease status and the accumulation of rare variants across the risk locus or gene units rather than with any single variant. Three main classes of collapsing tests are [1] tests that use group summary statistics on variant frequencies in cases and controls; [2] those that test for similarity in unique DNA sequences in different individuals; and [3] regression models that test collapsed sets of variants (and other variables) as predictors of the phenotype [36]. Many of the features of the different approaches are combined in a single software SKAT-O, where O refers to optimal [50] that is commonly used in these analyses.

Following association testing, identified variants are analyzed for functional consequences. Algorithms that predict deleteriousness of protein-coding variants such as SIFT or PolyPhen are used to prioritize nonsense and frameshift mutations because they result in loss of protein function [51, 52]. SIFT has been incorporated into ANNOVAR pipeline [53] while PolyPhen is a part of the SeattleSeq pipeline (http://​snp.​gs.​washington.​edu/​SeattleSeqAnnota​tion) as well as the PLINK/SEQ suite (http://​atgu.​mgh.​harvard.​edu/​plinkseq/​) for comprehensive functional annotation of variants. Methods for prioritizing noncoding variants based on nucleotide sequence conservation are also being utilized in large-scale sequencing studies; one commonly used of the many available algorithms (listed in Ref. [54]) is the Genomic Evolutionary Rate Profiling (GERP) algorithm [55]. These algorithms use comparative genomics, generally limited to mammalian species as nucleotide sequence is less conserved than protein sequence, to estimate nucleotide-level evolutionary constraints in genomic sequence alignments and assign conservation scores. Higher conservation scores are indicative of a more likely regulatory function and can be used to prioritize noncoding variants for further studies. A recently developed Combined Annotation–Dependent Depletion (CADD) method integrates many diverse measures of functional relevance such as deleteriousness, conservation, and other scored into a single measure (C score) for each variant that can be used to objectively prioritize variants [56]. An alternative to post hoc analysis of variants in associated genes/loci is to incorporate functional information into the test and stratify or weigh rare alleles by functional significance. A number of tests allow for inclusion of prediction scores in test statistics; PLINK/SEQ, for example, includes previously computed PolyPhen scores [57].