Ion channels and schizophrenia: a gene set-based analytic approach to GWAS data for biological hypothesis testing

Abstract

Schizophrenia is a complex genetic disorder. Gene set-based analytic (GSA) methods have been widely applied for exploratory analyses of large, high-throughput datasets, but less commonly employed for biological hypothesis testing. Our primary hypothesis is that variation in ion channel genes contribute to the genetic susceptibility to schizophrenia. We applied Exploratory Visual Analysis (EVA), one GSA application, to analyze European-American (EA) and African-American (AA) schizophrenia genome-wide association study datasets for statistical enrichment of ion channel gene sets, comparing GSA results derived under three SNP-to-gene mapping strategies: (1) GENIC; (2) 500-Kb; (3) 2.5-Mb and three complimentary SNP-to-gene statistical reduction methods: (1) minimum p value (pMIN); (2) a novel method, proportion of SNPs per Gene with p-values below a pre-defined α-threshold (PROP); and (3) the truncated product method (TPM). In the EA analyses, ion channel gene set(s) were enriched under all mapping and statistical approaches. In the AA analysis, ion channel gene set(s) were significantly enriched under pMIN for all mapping strategies and under PROP for broader mapping strategies. Less extensive enrichment in the AA sample may reflect true ethnic differences in susceptibility, sampling or case ascertainment differences, or higher dimensionality relative to sample size of the AA data. More consistent findings under broader mapping strategies may reflect enhanced power due to increased SNP inclusion, enhanced capture of effects over extended haplotypes or significant contributions from regulatory regions. While extensive pMIN findings may reflect gene size bias, the extent and significance of PROP and TPM findings suggest that common variation at ion channel genes may capture some of the heritability of schizophrenia.

Introduction

General trends in neuropsychiatric genetics

For most neuropsychiatric disorders, strong association or linkage findings have been uncommon; replication has been sparse. This, in combination with generally inconclusive results from the recent swell of GWAS work (Alaerts and Del-Favero 2009; Medina et al. 2009; Moskvina et al. 2009), has led investigators to conclude that most neuro-psychiatric disorders have complex genetic architectures. Whether this complexity is mediated by genetic interactions (expressed at the individual level) or genetic heterogeneity (expressed at the population level) is a matter of ongoing debate. Nonetheless, given this complexity, it is unlikely that further traditional genetic investigations will yield conclusive results.

Sequencing technology is rapidly becoming logistically and financially feasible for many institutions. However, the rate-limiting factor in genetic discovery under these methods will be the development of computational and statistical approaches that can extract meaning from the vast sequencing data generated. Given the 3 billion potential sites of variation across the genome, and the likely contribution of low-frequency and rare variants to disease risk, even comprehensive whole-genome sequencing of a massive target population may not produce readily obvious patterns of variation with clear causative implications. Alternatively, we suggest that a fruitful approach to genetic discovery using data at the whole genome level, including that derived from GWAS or sequencing, will be twofold. First, it will be hypothesis driven. Second, it will entail a methodology capable of measuring the collective contribution to overall risk of disease by biologically defined sets of variants.

Schizophrenia

A growing body of research within different psychiatric disorders, including autism, obsessive–compulsive disorder (OCD), attention deficit hyperactivity disorder (ADHD), and schizophrenia, supports dimensional symptom structures. Symptoms of schizophrenia may be roughly defined by dysfunction along three, at least partially distinct, phenotypic dimensions: executive functioning (symptoms of cognitive disorganization), social responsiveness (‘negative’ symptoms, or asociality) and sensory-perceptual integration (classical ‘positive’ symptoms, including hallucinations and delusions). Replicable findings localize executive functional capacities to prefrontal cortex (Camchong et al. 2006; Huffaker et al. 2009; MacDonald et al. 2005; Nigg and Casey 2005; Perlstein et al. 2001; Russell et al. 2003), social responsiveness to medial prefrontal circuitry and cingulate (Ashwin et al. 2007; Fujiwara et al. 2007; Pavuluri et al. 2007; Singer 2007), and sensory-perceptual integration to higher-order sensory/heteromodal association cortices (Booth et al. 2002; Buchanan et al. 2004; Ho et al. 2009; Keshavan et al. 2003; Silbersweig and Stern 1998). Despite advances in our understanding of schizophrenia at the level of both phenotype and brain system, however, untangling the molecular and genetic substrates of schizophrenia has been especially challenging.

Most generally, the accumulated results of genetic studies to date strongly suggest that schizophrenia is a complex genetic disorder whose genetic architecture is likely characterized by locus, allelic and trait heterogeneity at the population level with genetic interactions and epigenetic mechanisms mediating disease expression at the individual level (see Moore et al. 2006; Moore and Williams 2005; Thornton-Wells et al. 2004 for reviews of these concepts). Additionally, evidence is mounting that disease risk in the population is, in part, mediated by inherited or de novo rare variants (Alaerts and Del-Favero 2009; Gorlov et al. 2008; McClellan et al. 2007) or copy number variations (CNVs) (Alaerts and Del-Favero 2009; Glessner et al. 2010; ISC 2008; McCarroll 2008; Need et al. 2009; Stefansson et al. 2008; Walsh et al. 2008).

Specific genetic and pathophysiologic hypotheses in schizophrenia have evolved over many decades; the most longstanding centers on dopamanergic dysfunction (Glatt et al. 2003; Miyamoto et al. 2003; Verhoeff 1999). Alternate hypotheses tend to focus on the role of other neuro-transmitter systems including GABAergic (Lang et al. 2007; Petryshen et al. 2005; Reynolds and Harte 2007), glutamatergic (Corti et al. 2011; Dekeyne and Millan 2003; Lang et al. 2007; Nudmamud-Thanoi et al. 2007; Reynolds and Harte 2007), serotonergic (Gaddum and Hameed 1954; Schildkraut 1965; Woolley and Shaw 1954) and noradrenergic (Breier et al. 1998; Farley et al. 1978; Siuta et al. 2010; Sternberg et al. 1982; Wise and Stein 1973) systems and interactions thereof (see (Lang et al. 2007) and (Meltzer and Roth 1995) for recent reviews).

Importantly, the functional specificity of each of the putatively relevant CNS neurotransmitter systems is sub-served not only by the neurotransmitter-specific proteins (i.e., receptors, transporters, metabolic enzymes) expressed therein, but also by the unique subset of ion channels expressed in each system (Beaulieu and Gainetdinov 2011; Krause et al. 2002; Lachowicz and Sibley 1997; Missale et al. 1998; Nicoll 1988; Wolfart et al. 2001). However, despite a growing body of genetic, pharmacologic, functional and brain expression analyses implicating ion channels as primary molecular mediators of numerous physiological and pathophysiological processes likely involved in schizophrenia (see references discussed below), we could identify no published reports explicitly testing an ion channel hypothesis in schizophrenia.

Hypothesis and aims

The primary working hypothesis of this study is that genetic variation in a host of ion channel genes (including voltage- and ligand-gated subtypes) contribute to susceptibility to schizophrenia across affected populations. Moreover, we suspect the genetic architecture of these disorders, and of the contributions of ion channels in particular, is characterized by extensive locus, allelic and trait heterogeneity. We hypothesize that across the entire population of individuals affected by schizophrenia, there are many distinct ion channel gene variants and that most such variants are uncommon to rare. That said, we expect that some proportion of disease risk may be mediated by variants in linkage disequilibrium with the common tagSNPs representing ion channel genes in GWAS arrays.

Because our hypothesis entails that substantial locus and allelic heterogeneity characterize the risk architecture of the ion channel gene contribution to disease risk, we submit that gene set-based analytic approaches (GSAs) are well suited to the detection of such genetic contributions in case–control data. Our primary aim, then, was to test our biological hypothesis using GSA. Secondary aims were to compare results derived using (1) three SNP-to-gene mapping approaches and (2) three analytic approaches to SNP-to-gene reduction for GSA.

Background and rationale for our hypothesis, aims and methods

Ion channels and schizophrenia

Systematic examination of the role of ion channels in mediating risk of schizophrenia has not yet been undertaken. However, pathophysiologic, etiologic and pharmacologic roles for particular ion channel genes or proteins in schizophrenia have been explored.

Most extensively studied among the ion channel candidates is the calcium-activated potassium channel, KCNN3. Localized to chromosomal band 1q21.3 (Wittekindt et al. 1998), this gene was proposed as a functional candidate for schizophrenia and bipolar disorder because of its role in mediating neuronal excitability through its regulation of the slow component of hyperpolarization (Ritsner et al. 2002). As reviewed by Ritsner et al. (2002), several early association studies found evidence for association between schizophrenia and KCNN3 CAG repeat length (Bowen et al. 1998; Cardno et al. 1999; Chandy et al. 1998; Dror et al. 1999; Stober et al. 1998), though several follow-up analyses did not replicate the initial findings (Antonarakis et al. 1999; Austin et al. 1999; Chowdari et al. 2000; Joober et al. 1999; Li et al. 1998; Rohrmeier et al. 1999; Stober et al. 2000; Ujike et al. 2001). Despite the lack of consistent replication, several lines of evidence provide additional support (Ritsner et al. 2002), including: a large genomewide linkage scan in 22 extended schizophrenia pedigrees which revealed highly significant linkage evidence at 1q21-q22 (Brzustowicz et al. 2000), the finding of a KCNN3 mutation in a schizophrenic patient (Bowen et al. 2001) that trafficked to the nucleus and suppressed endogenous KCNN3 in a dominant negative manner (Miller et al. 2001), and an independent finding that schizophrenic patients had significantly different allele sizes than controls (Saleem et al. 2000). The complicated trajectory of investigations of KCNN3 in schizophrenia demonstrates a relatively typical pattern in candidate gene investigations across many neuropsychiatric disorders. However, under an architecture characterized by extensive genetic heterogeneity, such a trajectory of across samples and methodologic approaches is to be expected.

More recent literature from various disciplines implicates a host of other ion channel genes in schizophrenia. Candidate gene association (Bigos et al. 2010; Green et al. 2010; Nyegaard et al. 2010; Wei and Hemmings 2006), genome-wide (Curtis et al. 2011; Tam et al. 2010), gene-wide (Moskvina et al. 2009), gene set enrichment (Glessner et al. 2010) and meta-analytic (Huffaker et al. 2009) studies have yielded evidence for association between schizophrenia and several voltage-gated ion channels (KCNH2, KCNE1, KCNE2CACNA1C, CACNA1B, CACNA1F, CACNA5). In addition, Papassotiropoulos et al. (2011) were able to replicate (in three independent samples) evidence of association of a voltage-gated sodium channel (SCN1A) polymorphism with human short-term memory performance, known to be impaired in schizophrenia. Moreover, a further fMRI investigation demonstrated SCN1A allele-dependent activation differences in brain regions typically involved in working memory processes and implicated in schizophrenia.

Numerous functional, pharmacologic and translational studies have demonstrated functional links between schizophrenia, related endophenotypes or phenotypic dimensions and physiological and/or pathophysiological functioning in particular ion channel classes, including voltage-gated potassium channels [KCNH2(Huffaker et al. 2009); KCNQ channels (Fedorenko et al. 2008; Kapfhamer et al. 2010; Sotty et al. 2009); CHIP3, KCNA1, KCNAB1(Duncan et al. 2008)], outward-rectifying potassium channels [KCNK10 (Xiao et al. 2009); KCNK2/10, TREK channels (Thummler et al. 2007)], large conductance, voltage- and calcium-sensitive potassium channels [KCNMB1(Wong et al. 2005)], and voltage-gated calcium channels (CAC-NA1C (Bigos et al. 2010)), as well as broader classes of ion channels (Butler-Munro et al. 2010). Notably, in their recent review of mRNA expression profiling studies in animal models of schizophrenia, Van Schijndel and Martens (2010) note that while the majority of the genes implicated across the 26 profiling studies they reviewed were implicated only in a single study, genes belonging to sodium and potassium voltage-gated channel families were implicated in seven. Finally, reviews have highlighted the promise of ion channels as therapeutic targets in pharmacologic development due, in particular, to their membrane localization, structural heterogeneity, highly differentiated and specialized CNS distribution, essential roles in diverse physiological processes and likely relevance to several neuropathological diseases including schizophrenia (Ashcroft 2006; Camerino et al. 2008; Hansen et al. 2008; Judge et al. 2007).

Alongside the specific evidence supporting an etiologic or pathophysiologic role for ion channels in schizophrenia, evidence from molecular genetics and biology supports the candidacy of ion channels in mediating risk for human neuropsychiatric disorders more generally. Approximately 410 human genes encode known ion channel proteins (eEnsembl, Ensembl release 62, accessed 04/29/11). Ion channels have important roles in diverse processes including nerve excitation, cell proliferation, sensory transduction, and learning and memory (Ashcroft 2006). At least 60 ion channel genes have been associated with human disease (i.e., channelopathies); many channelopathies are genetically heterogeneous with the same phenotype being caused by mutations in different ion channel genes (locus heterogeneity) or affecting different alleles within the same gene (allelic heterogeneity). In addition, there a several mechanisms by which genetic mutation can disrupt channel function, including alteration of channel activity (e.g., disruption of ligand binding or open probability), alteration in channel synthesis or membrane targeting or, less commonly, alteration of single channel conductance. Furthermore, the form and severity of the clinical phenotype conferred by a particular channel mutation depends on the relative contribution the mutated channel makes to the electrical activity of the cells relevant to the phenotype of interest. In other words, there may be minimal disruption to electrical activity in some neurons and significant disruption in others, depending on the complement of ion channels expressed in each. That said, in some cases very subtle change in channel activity conferred by mutation can produce severe disease (Ashcroft 2006; Du et al. 2005).

The dimensional nature of human neuropsychiatric disorders necessitates that their molecular substrates have tightly regulated and specific expression patterns in order to preferentially disrupt specialized cognitive functions, while preserving basic brain functions and leaving other higher-order processes largely intact. There are several mechanisms by which the expression of ion channel proteins in the human CNS is regulated. First, there is a multitude of discrete isoforms and molecular subunit combinations within each ion channel subclass, each of which has distributional specificity that is controlled developmentally, temporally and adaptively (Abernethy and Soldatov 2002; Ader et al. 2008; Angelino and Brenner 2007; Anselmi et al. 2007; Catterall and Few 2008; Cerda et al. 2011; Chahine et al. 2008; Dai et al. 2009; Gaudet 2007; Goetz et al. 2007; Gotti et al. 2009; Hashimoto and Panchenko 2010; Jegla et al. 2009; Jensen et al. 2008; Kurokawa et al. 2009; Levitan 2006; Marionneau et al. 2011; Milstein et al. 2007; Molina et al. 2006; Olsen and Sieghart 2009; Paoletti 2011; Perrais et al. 2010; Picton and Fisher 2007; Tseng et al. 2007; Vanoye et al. 2010; Waxman et al. 2000; Yang et al. 2009). Second, most ion channels have known alternatively spliced isoforms, offering additional modulatory capacity (Chahine et al. 2008; Daniel and Ohman 2009; Honore 2008; Jenkinson 2006; Kurokawa et al. 2009; Lin et al. 2009; Mohapatra et al. 2007; Vazquez and Valverde 2006; Yan et al. 2008). Third, there are likely an infinite number of isoform- or complex-specific interactions with other CNS genes that further extend the capacity for developmental, temporal, and adaptive modulation (Chahine et al. 2008; Dib-Hajj and Waxman 2010; Fernandez-Alacid et al. 2009; Thompson and Begenisich 2006; Waxman 2007). Thus, ion channels are quite likely primary molecular mediators of the functional specificity of brain subsystems and circuitry and, therefore, are rational candidates in the search for molecular substrates of complex neuropsychiatric disorders.

From an evolutionary perspective, ion channels, relative to most other neuronally expressed proteins, demonstrate a remarkable, and relatively unique, simultaneity of phylogenetic conservation (Anderson and Greenberg 2001; Strong and Gutman 1993; Trimmer and Rhodes 2004), species differentiation (Abernethy and Soldatov 2002; Anderson and Greenberg 2001; Strong and Gutman 1993), structural diversity (Anderson and Greenberg 2001; Ashcroft 2006; Strong and Gutman 1993; Waxman 2000), and developmental and distributional specificity (Abernethy and Soldatov 2002; Anderson and Greenberg 2001; Freudenberg et al. 2007; Mechaly et al. 2005; Sailer et al. 2004; Stocker 2004; Stocker et al. 2004; Waxman 2000; Wolfart et al. 2001). Together, these qualities support a critical role for ion channels in the functioning and functional specialization of nervous systems, from the most primitive eukaryotes to human beings. In a very compelling bioinformatics analysis of human CNS-expressed ion channels, Freudenberg et al. (2007) examined the phylogenetic attributes of human ion channels relative to other CNS-expressed genes (Freudenberg et al. 2007). First, they observed that relative to invertebrate, vertebrate genomes had an increased percentage of ion channel genes. This pattern was shared by other important CNS-expressed genes. Second, they found that, in contrast to the other CNS-expressed genes, ion channels have longer intron and protein sequences, features typical of genes with more specialized expression patterns. Third, they found that ion channels, in contrast to non-channel genes, have increased human-rodent transcription start site conservation, indicating the functional relevance of mutations affecting ion channel transcriptional regulation. Taken together, the molecular genetic and phylogenetic characteristics of ion channels strongly support their candidacy in the genetic risk of neuropsychiatric disorders.

Thus, we propose that ion channels, as a class, represent ideal etiologic candidates underlying the genetic susceptibility to schizophrenia. Moreover, given the diversity of the implicated genes within this class and the likely heterogeneous nature of the underlying genetic architecture of schizophrenia, we submit that GSA represents an optimal and currently accessible approach to test such a hypothesis.

Rationale for gene set-based approach

Gene set- and pathways-based analytic approaches have emerged over the last decade, following the advent of genome-wide expression and GWAS microarray platforms. One of the earliest GSA approaches, gene set enrichment analysis (GSEA), was introduced by Subramanian et al. (2005) for the analysis of expression microarray data. Wang et al. (2007) published a study applying a modified GSEA algorithm to analyze individual-level genotype data from a GWAS while many additional investigations have elaborated on the relative strengths of gene set-based approaches in genetic discovery (Al-Shahrour et al. 2007; Chen et al. 2009; Medina et al. 2009). As a result, GSA methods and applications designed to apply similar algorithms to marker-level summary results (e.g., marker-level p values) of GWAS have emerged [e.g., gene set-based analysis of polymorphisms (GeSBAP) (Medina et al. 2009), Exploratory visual analysis (EVA) (Reif et al. 2005, 2007), improved gene set enrichment analysis (iGSEA) (Zhang et al. 2010), GSA-SNP (Nam et al. 2010), exploratory gene association network (EGAN) (Paquette and Tokuyasu 2010)].

Most simply defined, GSA is a set of methods that examine the extent to which more significant test statistics tend to aggregate or cluster within gene groups (or sets) that share some biological characteristic(s), or modules of functionally related genes (Medina et al. 2009). Thus, GSA may employ any type of annotations to classify genes into biologically relevant gene sets, such as shared molecular function, tissue expression patterns or shared biological pathways involvement. In addition to often being more biologically informative, GSAs, like pathways-based analyses, can overcome the limitations imposed by single-marker analysis of such high-throughput data, such as their extensive multiple testing corrections requirements through gene- or gene set-based data reduction and analysis.

Rationale for hypothesis-driven investigations

Thus far, GSA methods have been largely applied for exploratory analyses of large, high-throughput datasets, including the results of GWAS. Though apparently uncommon, these methods can also be employed to conduct explicit biological hypothesis testing. Notable strengths of exploratory analyses include the potential for illuminating latent patterns in the data that were not detected on initial (marker-level) analysis and that may not have been considered if inconsistent with extant biological perspectives. That said, hypothesis-driven analyses remain critical to scientific progress and have several advantages that we sought to exploit in our investigation. First, the scientific method entails that scientific investigations begin with a biological hypothesis that, in turn, should inform the selection of both the biological material and the methodologic approach to its analysis. Second, without an a priori hypothesis, interpretation of results relies on speculation rather than scientific deduction. Finally, the testing of explicit hypotheses frees the investigator from the burden of extensive hypothesis testing correction necessary under ‘hypothesis-free’ or exploratory analysis. Depending on the gene set or pathways annotations used, exploratory GSA necessarily entails the correction for hundreds to several thousand hypothesis tests.

Rationale for secondary aims: mapping strategies and GSA methods

The current debate in application of GSA methods, and the focus of our present comparative analyses, centers on two challenges: (1) the SNP-to-gene mapping strategy employed and (2) the analytic approach for SNP-to-gene statistical reduction. SNP-to-gene mapping is a pre-processing step in which the investigator must determine how the SNP-level data will be mapped to known genes in the human genome. The investigator may choose to map (and therefore include in the analysis) only those SNPs that lie within genic regions (i.e., coding sequences (CDS), exonic, intronic, 5′UTR, 3′UTR) or may opt to map all SNPs falling within some predetermined distance from the start/end of a gene’s boundaries (e.g., 500 Kb upstream and downstream of a genes start and end location.) While 500 Kb up/downstream should be expected to capture most enhancer/promoter regions (Wang et al. 2007), several recent studies have suggested that measured associations at common SNPs may capture information about regional rare variants that may be as far as 2.5 Mb away from the tagSNP (Dickson et al. 2010; Wang et al. 2010). Thus, it remains unclear whether extensive up/downstream mapping approaches may capture additional information about association within extended haplotypes and, of course, whether such information will ultimately enhance the accuracy or utility of GSA. To address the question of whether SNP-to-gene mapping parameters influence GSA results, we employed three mapping strategies in each of our analyses: GENIC (SNPs within gene boundaries), 500 Kb (GENIC plus 500 Kb up/downstream) and 2.5 Mb (GENIC plus 2.5 Mb up/downstream) mapping.

The second, and perhaps more challenging, problem facing the application of GSA strategies to GWAS data is the analytic method employed for SNP-to-gene reduction or the derivation of gene-level summary statistics from the original SNP-level summary data (p values, in particular). Until quite recently, the most commonly employed SNP-to-gene reduction method was to use the minimum SNP-level p value to represent each gene. The rationale, which is still valid, is that in a gene-centric analysis one wants the ‘best representative’ statistic for each gene. The use of the minimum p value also enables the analysis to capture effects of genes that may contribute a single susceptibility locus to the overall risk of disease. (Any method that combines effects across SNPs within a gene will necessarily reduce the magnitude of the single, best locus contribution and this may not be warranted.)

The now well-recognized disadvantage of this approach is that genes represented by a large number of SNPs on genotyping arrays (usually, but not always, larger genes) are more likely to have lower minimum p values based on chance alone. Thus, the use of minimum p value in GSA may be more likely to produce evidence of enrichment in gene sets that disproportionately contain larger genes. Given this, it is essential to address this potential bias by employing complimentary methods that account for gene size. We believe it is worthwhile to note, however, the mere fact is that a statistical result may be due to chance does not mean that the finding is due to chance. While it is true that ion channel genes, as a class, have larger-than-average gene size (and thus, that many ion channel genes have more than average probes on microarrays and GWAS platforms), it is also true that they represent excellent biological candidates and that this candidacy may, in part, be due to the size of their genes. Thus, we elected to conduct a set of GSAs to compare the results obtained using minimum p value (pMIN method) for each gene to those obtained using two additional SNP-to-gene reduction strategies that account for variation in gene size: proportion of SNPs per Gene with p value < α-threshold (PROPα), and the truncated product method (TPM) (see “Methods” for further description).

Finally, our choices of SNP-to-gene reduction methods were made not only to address the issue of gene size bias, but also for their statistical and biological complimentarity. We believe that the GSA results obtained using each method may have unique implications for the genetic architecture of schizophrenia. If real (i.e., not a statistical artifact), gene sets implicated by pMIN analysis alone may be those in which genes have a single locus, or finite set of proximal loci, contributing to disease susceptibility. The PROPα method, by contrast, will be more robust to the identification of gene sets in which its gene members have disproportionately larger numbers of disease-associated SNPs, suggesting the potential for substantial allelic heterogeneity. Finally, the TPM method will implicate gene sets in which a disproportionate number of its genes had very strong evidence of association and/or a large number of disease-associated SNPs and would most strongly implicate those genes manifesting both characteristics. Thus, the confluence of findings across pMIN, PROP and TPM methods may suggest the relative likelihood of contributions from allelic and locus heterogeneity to risk architecture. In summary, marker-level GWAS data used in the GSAs cannot, however, address the nature and extent of genetic epistasis relevant at the individual level.

Methods

Overview of data preparation and analytic approach

From each original GWAS dataset, we created three experimental datasets based on three distinct SNP-to-gene mapping strategies (GENIC, 500 Kb, 2.5 Mb) to be examined. On each experimental dataset (e.g., EA-GENIC), we derived a gene-level p value using three distinct analytic strategies (pMIN, PROP, TPM). For ease of reference, we refer to each analysis by original GWAS dataset, experimental mapping dataset, analytic strategy and α-threshold as in EA-500 Kb-TPM01 for the European-American GWAS, 500 Kb mapping strategy, Truncated Product Method strategy for SNP-to-gene reduction and TPMα < 0.01. We conducted separate GSAs on each set of experimental values, testing for enrichment of 11 nested ion channel activity gene sets in each (see Fig. 1).

Fig. 1.

Flow Diagram Illustrating SNP-to-gene mapping and analytic processing steps for gene set analyses. Flow diagram depicting the preprocessing procedures used to transform the original EA and AA GWAS data prior to performing the gene set analyses. The center column (Methods) depicts the methods used to transform the original and mapping datasets into the final EA and AA GSA datasets used in EVA to generate final GSA results. The far left (EA Data) and right (AA Data) columns each depict the starting (original GWAS), intermediate (Mapping) and final (GSA) datasets resulting from each methodological step

Gene set annotations

Arguably, the most comprehensive biological classification scheme for genes based on their molecular function is the molecular function gene ontology (GO-MF). According to the AmiGO project, the Molecular Function ontology captures the elemental activities describing the actions of a gene product at the molecular level. Relative to other gene classification schemes (e.g., biological pathways), such elemental activities are most likely to be comprehensive, well-characterized and reliable. With regard to ion channel gene classes, in particular, it is important to note that several commonly used biological pathways annotations (e.g., Panther) do not include most known ion channel genes within any pathways (Askland et al. 2009). (For example, the KCNN3 gene has ‘No pathway information available’, Panther Database, accessed 04/28/11).

The molecular function ontology term, ‘Ion channel activity’ is defined as ‘catalysis of facilitated diffusion of an ion (by an energy-independent process) by passage through a transmembrane aqueous pore or channel without evidence for a carrier-mediated mechanism.’ [AmiGO, accessed 02/08/11]. There are 394 human genes that fall within this molecular function gene set. This gene set is a 6th generation term of the molecular function ontology under the following lineage (from least to most specific): molecular function (15,540 gene products), transporter activity (1,189), transmembrane transporter activity (926), substrate-specific transporter activity (855), substrate-specific channel activity (403), ion channel activity (394). The human ion channel activity set contains 4 daughter terms: anion channel activity, cation channel activity, ligand-gated ion channel activity and voltage-gated ion channel activity. There is substantial overlap among the genes included among the daughter subsets (e.g., many cation channel activity genes are also ligand- and voltage-gated channel activity genes).

Our primary hypothesis is that, over the entire affected population, a diverse set of ion channel gene variants contributes to the genetic susceptibility to schizophrenia. However, due to random sampling effects, any particular case–control population may, by chance, carry disproportionate disease-associated variation in a specific ion channel subset. Thus, if under this condition, we test only for the most general ion channel activity set, we could miss strong enrichment that might be expected, by chance, within a relevant subset. Thus, we elected to simultaneously test for enrichment of the most general set, ion channel activity, as well as 10 specific ion channel activity subsets. Though any one of the subsets are potentially relevant, we selected the 10 most basic subsets in terms of their molecular function because of strong reliability of the functional definitions as well as their reasonable size: potassium channel activity (131 genes), sodium channel activity (32), calcium channel activity (81), chloride channel activity (73), calcium release channel activity (9), voltage-gated ion channel activity (192), voltage-gated potassium channel activity (98), voltage-gated sodium channel activity (15), voltage-gated calcium channel activity (27), and voltage-gated chloride channel activity (16).

SNP-to-gene mapping

The original GWAS datasets, obtained via dbGaP permission, used in the current analysis contained 729,454 (EA) and 845,814 (AA) SNP-level summary statistics (a SNP-level Odds ratio and corresponding p value). Our analytic approach is gene-centric, in that it is measuring the extent to which particular sets of genes, based on shared molecular functions, are enriched for more significant test statistics. As such, the approach entails deriving gene-level summary statistics for all genes represented in the original SNP-level datasets. To do this, each SNP was first mapped to its nearest gene, using the Affymetrix 6.0 Annotation dataset, Release 30 (Release date November 15, 2009). Affymetrix annotations provide gene mapping information (including gene symbol, distance from gene, SNP-Gene relational information) for 98.8% of genotyped EA and 98.7% of genotyped AA SNPs. When more than one gene mapped to a single SNP, a single gene was assigned using the following prioritization method: CDS > UTR > exon > intron; and for intergenic SNPs, the gene with the closest up/downstream distance was assigned. The resulting, fully mapped SNP EA and AA datasets were then subject to further reduction into separate mapping sets, as below.

Based on prior biological rationale, we employed three mapping strategies to derive three mapping datasets from each primary (EA and AA) GWAS dataset. The first strategy retained only those SNPs that mapped within the boundaries of a gene, including CDS, exonic, intronic, 5′UTR, 3′UTR (we refer to this as the GENIC dataset). The GENIC mapping strategy is considered the most conservative as it does not include SNPs that may lie within up/downstream promoter/enhancer regions. The second strategy retained only SNPs residing within and up to 500 Kb up- or downstream of a gene’s boundaries (500 Kb dataset). The 500 Kb approach, consistent with that employed by Wang et al. (2007), should capture most effects mediated (or represented) by SNPs within the promoter/enhancer elements in addition to the genic SNPs. The third and broadest strategy retained all SNPs within 2.5 Mb of its nearest gene. This approach was included in order to investigate the potential for effects of long-range haplotypes that may be captured by very distant SNPs (Dickson et al. 2010; Wang et al. 2010).

Analytic approaches to SNP-to-gene reduction for GSA

To test our biological hypothesis, we applied three analytic approaches to derive gene-level summary statistics from the SNP-level GWAS data for each mapping set. The first, and thus far most commonly employed method, was to use the minimum p value to represent each gene (pMIN). For each of the three mapping datasets, the lowest SNP-level p value was selected to represent each gene in the GSA. Since pMIN does not control for the number of mapped SNPs, we also employed two additional analytic approaches, both of which control for the number of SNPs mapped to each gene in their derivation.

In the second approach, PROPα, we used a novel procedure to derive a summary statistic (p value) to each gene based on the gene’s proportion of mapped SNPs with p values less than specified α-thresholds. For each mapping set, we used EVA to calculate a proportion of SNPs/gene with p < α and derived a corresponding Fisher’s exact probability of finding such a proportion by chance given the size of the gene. The resulting p value was then used as the experimental statistic for each gene to conduct the GSA in EVA. As PROP is a novel strategy, we employed three different α-thresholds (standard = 0.05, conservative = 0.01 and very conservative = 0.001) to assess sensitivity.

Our third approach was the truncated product method (TPM), originally developed for combining p values by Zaykin et al. (2002) (Zaykin et al. 2002) and recently employed by Moskvina et al. (2009) (Moskvina et al. 2009) in their analysis of bipolar and schizophrenia GWAS data. This method is intermediate between Fisher’s product method (Province 2001) and Sidak’s correction (Abdi 2007; Sidak 1967). For each gene, the test statistic is calculated as the product of all p values <α. The null distribution for this statistic depends on the total number of SNPs for that gene, and is calculated with a formula described by Zaykin et al. (Zaykin et al. 2002). The resulting p value is then used as the experimental statistic in the GSA. Following the example of Zaykin et al., for genes containing over 1,000 SNPs we used a Monte Carlo algorithm to approximate the null distribution. This was done to avoid arithmetic overXow. For genes with no SNPs that had p value <α, Sidak’s correction (Sidak 1967) was applied. For TPM, we employed two α-thresholds for comparability to other published and future work (0.05 and 0.01). All TPM calculations were performed in R and the resulting p values were exported and uploaded into EVA to conduct the GSAs.

GSA by exploratory visual analysis (EVA)

To conduct our GSA, we employed the Exploratory Visual Analysis (EVA) application (Reif et al. 2005; Reif et al. 2007). EVA enables the use of either Fisher’s Exact or permutation-based significance testing among gene sets. For each experimental dataset (i.e., EA-GENIC, EA-500 Kb, EA-2.5 Mb, AA-GENIC, AA-500 Kb, AA-2.5 Mb), the gene-level p values derived under each analytic method (i.e., pMIN, PROPα and TPMα) were used to perform all GSAs in EVA using an enrichment threshold of p < 0.05. Thus, for each analysis, EVA calculates the proportion of genes in each GO-MF-defined gene set whose experimental p value is <0.05 then derives a corresponding Fisher’s exact probability of Finding such a proportion by chance given the size of the gene set.

EVA parameter settings

In most gene set-based approaches, the investigator must choose a threshold criteria that allows calculation of the enrichment statistic from a comparison of the ‘high’ to the ‘low’ scores in the experiment of interest. For example, when using the pMIN method, the investigator must choose a pMIN cutoff value for the calculation of the enrichment statistic. This choice is essentially an arbitrary one, so convention is often followed. We chose a more conservative pMIN cutoff of p = 0.01 because minimum p values do not follow a uniform distribution and an expectedly large proportion of genes had pMIN <0.05. In the case of the PROP and TPM methods, the derived gene-level p values follow a standard uniform distribution. Thus, for our primary analyses, we choose a standard threshold of p = 0.05 for those experiments.

Multiple hypothesis testing correction

In the analysis of each experimental dataset, we simultaneously tested 11 ion channel hypotheses. Under the gene ontology structure employed (GO-MF), these gene sets (and therefore, the respective hypotheses) are not independent. In fact, ion channel activity is a parent term under which the remaining ten ion channel subsets are subsumed. Thus, we employed an a priori threshold of 0.025 for our primary analyses and also applied the extremely stringent Bonferroni correction, which assumes complete independence (Bonferroni-p < 4.55E–03).

Results

Results of our primary European-American (EA) and African-American (AA) GSAs are shown in Tables 1 and 2, respectively. All tested ion channel gene sets reaching nominal (p < 0.05) or a priori significance (p < 0.025) levels, under our primary EVA enrichment threshold of p = 0.05, are shown in the tables.

Table 1.

GSA results (p values) for all tested ion channel gene sets exceeding nominal significance in European-American sample

Analytic parameters			Significant GSA results for tested gene sets
Statistical reduction method	Mapping strategy	EVA enrichment cut-off	Ion channel activity	Voltage-gated ion channel activity	Potassium channel activity	Calcium channel activity	Chloride channel activity	Voltage-gated chloride channel activity	Calcium-release channel activity
pMIN	Genic	0.01	8.12E–05**	0.0412	NS	9.67E–05**	5.91E–03*	NS	4.69E–04**
pMIN	500 Kb	0.01	8.55E–06**	5.54E–03*	8.87E–03*	5.05E–04**	NS	NS	1.60E–04**
pMIN	2.5 Mb	0.01	1.55E–05**	7.12E–03*	0.0106*	6.35E–04**	NS	NS	1.73E–04**
PROP001	Genic	0.05	NS	NS	NS	NS	NS	NS	9.20E–04**
PROP001	500 Kb	0.05	2.86E–03**	NS	1.01E–03**	NS	NS	NS	3.29E–05**
PROP001	2.5 Mb	0.05	8.62E–03*	NS	6.08E–03*	NS	NS	NS	3.41E–05**
PROP01	Genic	0.05	0.02945	NS	NS	0.0309	NS	NS	0.0113*
PROP01	500 Kb	0.05	0.03288	NS	0.0160*	NS	NS	NS	1.26E–03**
PROP01	2.5 Mb	0.05	0.03212	NS	0.0158*	NS	NS	NS	1.25E–03**
PROP05	Genic	0.05	NS	NS	NS	NS	NS	NS	0.0462
PROP05	500 Kb	0.05	NS	NS	NS	0.0225*	NS	NS	8.88E–03*
PROP05	2.5 Mb	0.05	NS	NS	NS	0.0220*	NS	NS	8.82E–03*
TPM01	Genic	0.05	0.01926*	NS	NS	NS	0.0265	0.0393	0.0319
TPM01	500 Kb	0.05	3.76 E–03**	0.0462	5.40E–03*	NS	NS	NS	4.13E–04**
TPM01	2.5 Mb	0.05	7.91E–03*	NS	0.0152*	NS	NS	NS	4.30E–04**
TPM05	Genic	0.05	NS	NS	NS	NS	NS	0.0407	5.68E–04**
TPM05	500 Kb	0.05	NS	NS	NS	0.0428	NS	NS	1.45E–03**
TPM05	2.5 Mb	0.05	NS	NS	NS	0.0448	NS	NS	1.48E–03**

There were no significant GSA findings for sodium channel, voltage-gated potassium, voltage-gated calcium or voltage-gated sodium channel gene sets

pMIN (minimum p value method), PROP05, PROP01, PROP001 (proportion method using α = 0.05, 0.01 and 0.001, respectively), TPM05, TPM01 (truncated product method using α = 0.05 and 0.01, respectively)

^*Indicates gene set reached a priori significance threshold of <0.025

^**Indicates gene set significant after Bonferroni correction

Table 2.

GSA results (p values) for all tested ion channel gene sets exceeding nominal significance in African-American sample

Analytic parameters			Significant GSA results for tested gene sets
Statistical reduction method	Mapping strategy	EVA enrichment cut-off	Ion channel activity	Voltage-gated ion channel activity	Potassium channel activity	Calcium channel activity	Chloride channel activity	Voltage-gated potassium channel activity	Voltage-gated calcium channel activity	Calcium-release channel activity
pMIN	Genic	0.01	2.00E–08**	1.68E–05**	5.19E–04**	1.19E–03**	8.63E–04**	7.59E–03*	1.50E–03**	0.0155*
pMIN	500 Kb	0.01	4.13E–07**	1.90E–04**	1.44E–05**	8.39E–04**	0.0234*	5.78E–03*	NS	NS
pMIN	2.5 Mb	0.01	3.53E–07**	2.42E–04**	1.73E–05**	4.00E–04**	0.0258	6.59E–03*	NS	NS
PROP001	Genic	0.05	NS	NS	NS	NS	NS	NS	NS	NS
PROP001	500 Kb	0.05	NS	NS	NS	NS	NS	NS	NS	NS
PROP001	2.5 Mb	0.05	NS	NS	NS	NS	NS	NS	NS	NS
PROP01	Genic	0.05	NS	NS	NS	NS	NS	NS	NS	NS
PROP01	500 Kb	0.05	NS	NS	0.0499	NS	NS	NS	NS	NS
PROP01	2.5 Mb	0.05	NS	NS	NS	NS	NS	NS	NS	NS
PROP05	Genic	0.05	NS	NS	NS	NS	NS	NS	NS	NS
PROP05	500 Kb	0.05	NS	NS	9.33E–04**	NS	NS	0.0366	NS	NS
PROP05	2.5 Mb	0.05	NS	NS	2.92E–03**	NS	NS	NS	NS	NS
TPM01	Genic	0.05	NS	NS	NS	NS	NS	NS	NS	NS
TPM01	500 Kb	0.05	NS	NS	NS	NS	NS	NS	NS	NS
TPM01	2.5 Mb	0.05	NS	NS	NS	NS	NS	NS	NS	NS
TPM05	Genic	0.05	NS	NS	NS	NS	NS	NS	NS	NS
TPM05	500 Kb	0.05	NS	NS	0.0423	NS	NS	NS	NS	NS
TPM05	2.5 Mb	0.05	NS	NS	NS	NS	NS	NS	NS	NS

There were no significant GSA findings for sodium channel, voltage-gated chloride channel or voltage-gated sodium channel gene sets

pMIN (minimum p value method), PROP05, PROP01, PROP001 (proportion method using α = 0.05, 0.01 and 0.001, respectively), TPM05, TPM01 (truncated product method using α = 0.05 and 0.01, respectively)

^*Indicates gene set reached a priori significance threshold of <0.025

^**Indicates gene set significant after Bonferroni correction

Minimum p value (pMIN) method

Multiple ion channel gene sets were significantly enriched under all three mapping strategies using the pMIN approach in both the EA and AA analyses. Several sets retained significance under conservative multiple testing correction across all mapping strategies, including the parent set, ion channel activity.

Proportional (PROP) method

Under all mapping strategies and all α-thresholds (0.05, 0.01, 0.001), all EA PROPα GSAs produced statistically significant enrichment (p < 0.025) of at least one tested ion channel gene set. The most robust enrichment was detected under the 500 Kb and 2.5 Mb mapping strategies. Of particular note in the EA PROPα analyses is the fact that for both the 500 Kb and 2.5 Mb datasets, the significance of the implicated ion channel sets increased with decreasing (i.e., more conservative) α-thresholds. Thus, in the 500 Kb set for example, employing an α-threshold of 0.001 (i.e., PROP001) to calculate the proportion of subthreshold SNPs for each gene produced more statistically significant gene set enrichment for the implicated sets than did the use of α = 0.01 (PROP01) and α = 0.05 (PROP05). As shown in Table 1, several of the implicated gene sets also retained significance after extremely conservative Bonferroni corrections across one or more mapping/analytic approaches including: ion channel activity, potassium channel activity and calcium-release channel activity.

By contrast, the AA GSAs produced more circumscribed findings. The potassium channel activity (KCN) set was significantly enriched under the two broadest mapping strategies (500 Kb and 2.5 Mb) and for PROP05 analyses. Of note, the KCN findings in the AA 500 Kb-PROP05 (p = 9.33E–04) and 2.5 Mb-PROP05 (p = 2.92E–03) GSAs also retained significance after Bonferroni correction.

Truncated product method (TPM)

Under all mapping strategies and both α-thresholds (0.05, 0.01), each EA TPM analysis produced statistically significant enrichment (p < 0.025) of at least one tested ion channel gene set. Similar to the EA PROP analyses, we found more significant enrichment (except for the calcium-release channel gene set) under the more conservative α-threshold (TPM01). The ion channel and potassium channel activity gene sets were enriched only in the conservative TPM01 analyses. The most robust enrichment was detected under the 500 Kb-TPM01 analysis, wherein both ion channel activity (p = 3.72E–03) and calcium-release channel activity (p = 4.13E–03) retained significance after Bonferroni correction.

No tested ion channel gene set reached a priori significance threshold in any AA TPM analysis. In the AA 500 Kb-TPM05 analysis, the potassium channel activity gene set did reach nominal significance (p = 0.04228), consistent with the significance of this gene set in the AA 500 Kb-pMIN and -PROP05 analyses and its nominal significance in AA 500 Kb-PROP01. In addition, when the gene set enrichment criterion was relaxed to p < 0.075 in a post-hoc TPM01 analyses, two tested ion channel gene sets each reached nominal significance in the 2.5 Mb and 500 Kb mapping sets: voltage-gated sodium channel (2.5 Mb: p = 0.02911; 500 Kb: p = 0.02826), and calcium channel (2.5 Mb: p = 0.04125; 500 Kb: p = 0.03902) activity.

Discussion

Differential implications for pMIN, PROP and TPM

Overall, the pMIN approach found significant enrichment of several ion channel gene sets in all three mapping strategies for EA and AA datasets. The extent of the findings under pMIN, relative to the other methods, does support the possibility that pMIN may be biased by gene size. However, we would again emphasize that while large genes are more likely to have disproportionately low minimum p values by chance, such statistical probabilities cannot address the potential that such differences in distribution of more significant minimum test statistics may have a biological basis. Until more definitive investigations, such as deep resequencing and functional analyses, have been completed, the results produced under pMIN methods may yet provide valid biological insights for future investigations.

The fact that the PROP and TPM analytic methods, each of which do account for gene size, found significant enrichment for several tested ion channel gene sets across all three mapping strategies in the European-American sample provides very strong statistical evidence that variation at common polymorphic markers within these genes may explain some proportion of the heritability of schizophrenia in these populations. Interestingly, however, the fact that our GSAs, like most, utilize measures of association at common polymorphic markers does not necessarily suggest that the relevant underlying functional variants are, themselves, common in the population. As alluded to in the introduction, recent investigations have demonstrated the plausibility, perhaps likelihood, that tagSNPs may be capturing effects of rare variants over extended genomic haplotypes. While it remains unclear exactly what type of functional variants (i.e., common, low-frequency or rare) may be creating the effects measurable in GSAs, our findings support the pathophysiological relevance of this class of genes in mediating susceptibility to schizophrenia. Thus, we would submit that, in combination with previous findings implicating specific ion channel genes, our findings suggest that ion channel genes, as a class, are ideal candidates for deep resequencing. In addition, our finding of strong GSA evidence across all three methods for the EA sample suggests that both locus and allelic heterogeneity are likely to factor strongly into the genetic architecture of schizophrenia, and the contribution of variation in ion channels in particular. For the AA sample, significant findings were limited to the pMIN and PROP05 analyses, which may suggest a greater role for allelic heterogeneity within this population (see “Comparison of EA and AA results”, below, for further discussion of EA-AA discrepancies).

Results generated under genic versus broader mapping strategies

A closer examination of some specific results may be revealing. Under the most stringent thresholds (i.e., PROP001 and TPM01) in the EA analyses, the 500 Kb and 2.5 Mb mapping strategies yielded very similar results (see Table 1). Ion channel activity and the potassium and calcium-release channel activity subsets each were strongly implicated, though their rank orders varied somewhat across mapping and analytic strategies. By contrast, under the GENIC mapping strategy, the TPM01 analysis found significant enrichment for the general ion channel activity set only, while the PROP001 found significant enrichment only in the calcium-release channel set. The more limited findings under the GENIC, as opposed to the broader, mapping sets may reflect true discrepancies in the extent of disease-related genomic variation captured by intra- versus proximal intergenic SNPs. In other words, these results could suggest that the disease susceptibility contributed by these genes is primarily via regulatory alterations (e.g., regulation of temporal or regional expression patterns), rather than via structural or splicing alterations. An alternative interpretation of the discrepant findings between GENIC and broader mapping strategies in the EA sample is that the GENIC results are simply less robust because they were derived from the analysis of a much more limited set of SNPs (i.e., 278424 EA-Genic SNPs vs. 675986 EA-500 Kb SNPs vs. 720477 EA-2.5 Mb SNPs), or because intergenic SNPs are likely to harbor more false positives.

Comparison of EA and AA results

Comparing the particular significant results yielded in the EA and AA analyses, we note that there are far fewer significantly enriched ion channel gene sets in the AA analyses beyond the AA-pMIN analyses. One interpretation of the less substantial AA findings is that there are true ethnic differences in the susceptibility genes for schizophrenia and that ion channels are not prominent molecular substrates of schizophrenia phenotypes in AA populations. This is certainly possible given their rather distinct evolutionary histories and genomic architectures (Tishkoff et al. 2009; Tishkoff and Williams 2002), relative to EA populations. There are also several alternative explanations for the ethnic discrepancies in our findings. First, it may simply be a random sampling effect. Namely, due to the extensive genetic heterogeneity of schizophrenia, it’s possible that the particular sample of 1,241 AA cases represented in the GWAS analyzed did not harbor a substantial amount of disease-associated variation in these genes (but that they do exist in the larger affected population). A second alternative is that case ascertainment strategies inadvertently selected for different phenotypic characteristics in the two ethnic groups. In fact, it has been shown through previous epidemiologic work that there are differences in diagnostic rates and patterns between European and African-American samples in schizophrenia (Neighbors et al. 2003; Trierweiler et al. 2000; Trierweiler et al. 2006). Third, the AA results may have been less robust because of higher dimensionality relative to sample size, as compared to the EA sample. More specifically, the EA GWAS genotyped at 729,454 SNPs in 1,404 cases and 1,442 controls while the AA GWAS genotyped 835,143 SNPs in 1,241 cases and 979 controls. This could have resulted in reduced power to detect gene set enrichment in the AA relative to the EA datasets. This may be particularly true if, as the AA-pMIN and highly significant AA-PROP05 results suggest, there is a substantial contribution to disease risk of variation in the KCN gene set. As potassium channels are the largest specific subclass of ion channels, the issues of higher dimensionality are especially pertinent. Finally, given the known reductions in LD in populations of African ancestry, it is possible (perhaps likely) that the AA SNP panel employed is insufficient to capture the extent of genomic variation present in African-American genomes. Thus, although the AA panel is comprised of a greater number of SNPs relative to the EA panel, it captures proportionally less than does the EA panel.

Comparison of genes conferring enrichment in the potassium channel activity gene set: EA versus AA

Our findings of significant to highly significant enrichment in the potassium channel (KCN) activity set in both populations (EA-500 Kb-pMIN, EA-2.5 Mb-pMIN, EA-500 Kb-PRO P01/PROP001, EA-500 Kb-TPM01, EA-2.5 Mb-PROP01/PROP001, EA-2.5 Mb-TPM01; AA-Genic-pMIN, AA-500 Kb-pMIN, AA-2.5 Mb-pMIN, AA-500 Kb-PROP05, AA-2.5 Mb-PROP05) are noteworthy for several reasons. First, aside from the early work in the KCNN3 gene, the smattering of independent findings for other potassium channel genes described in the “Introduction”, and one recent study investigating the mechanism of lithium in bipolar disorder (Butler-Munro et al. 2010), the general class of potassium channel genes has been explored much less extensively than other ion channel classes in neuropsychiatric disorders. Following the significant bipolar GWAS findings for CACNA1C (Sklar et al. 2008), much attention [e.g., follow-up GWAS meta-analyses (Ferreira et al. 2008)] has been paid to the potential role of this gene and related voltage-gated calcium channels and calcium-channel interactants [e.g., ANK3 allelic heterogeneity study (Schulze et al. 2009)] in both bipolar disorder and schizophrenia. In addition, a role for sodium channel genes has long been speculated due to their role as pharmacologic targets of many of the standard, and most effective, mood stabilizing agents used in the treatment of bipolar disorder. Thus, it is interesting that it was the more neglected, yet largest and most functionally diverse, ion channel subset that produced the most consistent findings of all subsets tested.

Finally, one of our most interesting observations derives from closer examination of the specific genes conferring enrichment to the KCN gene set across the samples and analytic methods. As shown in Table 3, across the 500 Kb analyses in which the KCN gene sets reached at least nominal significance (EA-pMIN, EA-PROP01, EA-PROP001, EA-TPM01, AA-pMIN, AA-PROP05, AA-PROP01, AA-TPM05), the KCN gene set enrichment is conferred by a total of 45 KCN genes. Of those 23 genes (51%) contribute to enrichment of the KCN gene set only in AA analysis(es), 8 (18%) contribute only in EA analysis(es) and 14 (31%) contribute in both EA and AA analyses. Thus, the majority of genes (69%) mediating the enrichment of this gene set across samples and methods are unique to the results of only one ethnic sample. In addition, from the list of genes mediating enrichment of the KCN gene set, we can observe that only a very small number (~4, <10%, based on our literature review) have been previously implicated in genetic or other investigations (Borsotto et al. 2007; Duncan et al. 2008; Tam et al. 2010; Xiao et al. 2009). Interestingly, three of the four contributed to enrichment in GSA(s) of one ethnic group only (see Table 3).

Table 3.

Comparison of specific potassium channel genes contributing to enrichment of the potassium channel activity gene set in the 500 Kb analyses across samples and statistical reduction methods

GENE	Gene size (bp)	EA-pMIN	EA-PROP01	EA-PROP001	EA-TPM01	AA-pMIN	AA-PROP05	AA-PROP01	AA-TPM05	SIG analysis count	SIG sample(S)
KCNQ1	614,039	8.06E–03	4.54E–02	NS	2.82E–02	1.59E–03	6.90E–03	3.62E–02	1.25E–03	7	EA & AA
KCNA4	6,722	3.89E–04	NS	9.11E–04	3.95E–03	2.25E–03	2.42E–02	NS	3.15E–02	6	EA & AA
KCND2	476,600	2.05E–03	7.12E–03	NS	1.27E–03	2.02E–03	2.92E–02	NS	2.47E–02	6	EA & AA
KCNMA1	767,531	7.77E–04	5.43E–03	3.49E–03	1.22E–04	1.15E–03	NS	NS	NS	5	EA & AA
KCNT2	382,586	8.28E–05	NS	1.17E–02	6.54E–03	1.84E–03	NS	1.31E–03	NS	5	EA & AA
TMEM38B	80,619	3.44E–05	4.52E–03	2.54E–02	1.47E–04	5.51E–04	NS	NS	NS	5	EA & AA
KCND3	411,800	2.01E–03	2.02E–03	NS	4.82E–04	7.04E–03	NS	NS	NS	4	EA & AA
KCNE2*	7,117	NS	NS	NS	NS	2.67E–04	3.39E–03	1.88E–03	1.30E–04	4	AA
KCNH7	467,716	NS	NS	NS	NS	2.21E–03	9.05E–04	4.27E–02	8.15E–04	4	AA
KCNJ15	45,084	NS	NS	NS	NS	1.80E–03	3.40E–03	7.14E–03	1.90E–04	4	AA
KCNK10*	142,011	6.60E–03	NS	NS	NS	8.24E–04	2.63E–03	NS	8.47E–04	4	EA & AA
KCNK13	124,087	8.68E–04	4.13E–04	NS	4.03E–05	4.17E–03	NS	NS	NS	4	EA & AA
KCNS3	54,237	2.63E–04	6.29E–03	2.84E–02	2.35E–04	NS	NS	NS	NS	4	EA
TMEM38A	27,894	NS	NS	NS	NS	1.35E–03	4.38E–02	1.19E–02	3.67E–03	4	AA
KCNA2	2,569	8.40E–03	NS	NS	NS	NS	1.53E–02	NS	1.92E–02	3	EA & AA
KCNB1	110,676	NS	NS	NS	NS	8.82E–03	3.29E–02	NS	3.49E–02	3	AA
KCNC2	169,615	NS	NS	NS	NS	8.44E–04	2.68E–03	NS	7.78E–04	3	AA
KCNH6	25,936	1.88E–03	2.38E–03	NS	6.08E–04	NS	NS	NS	NS	3	EA
KCNK3	40,900	5.49E–04	NS	2.60E–02	2.02E–02	NS	NS	NS	NS	3	EA
KCNQ2*	68,723	3.30E–03	1.35E–02	NS	5.97E–03	NS	NS	NS	NS	3	EA
PKD2	70,109	NS	NS	NS	NS	1.30E–03	1.83E–02	NS	5.40E–03	3	AA
HCN2	27,266	NS	NS	NS	NS	NS	2.60E–02	NS	1.37E–02	2	AA
KCNB2	533,168	2.37E–03	NS	NS	NS	9.42E–03	NS	NS	NS	2	EA & AA
KCNC1	36,665	3.63E–03	NS	NS	4.13E–02	NS	NS	NS	NS	2	EA & AA
KCNG3	52,080	NS	NS	NS	NS	2.91E–03	NS	3.63E–02	NS	2	AA
KCNIP1	450,177	2.48E–03	NS	NS	NS	2.08E–03	NS	NS	NS	2	EA & AA
KCNIP4	1,610,967	5.78E–03	NS	NS	NS	5.88E–03	NS	NS	NS	2	EA & AA
KCNK17	15,459	NS	NS	NS	NS	NS	2.10E–02	NS	4.20E–02	2	AA
KCNQ5	630,781	NS	NS	NS	NS	3.81E–03	4.38E–02	NS	NS	2	AA
ABCC8	84,002	9.22E–03	NS	NS	NS	NS	NS	NS	NS	1	EA
KCNA1*	8,348	NS	NS	NS	NS	4.49E–03	NS	NS	NS	1	AA
KCNA3	3,346	NS	NS	NS	NS	1.13E–03	NS	NS	NS	1	AA
KCNA5	2,764	NS	NS	NS	NS	3.07E–03	NS	NS	NS	1	AA
KCNE4	3,493	NS	NS	NS	NS	5.30E–03	NS	NS	NS	1	AA
KCNF1	2,288	NS	NS	NS	NS	3.03E–03	NS	NS	NS	1	AA
KCNG1	19,482	4.39E–03	NS	NS	NS	NS	NS	NS	NS	1	EA
KCNG2	36,148	7.96E–03	NS	NS	NS	NS	NS	NS	NS	1	EA
KCNH1	450,902	NS	NS	NS	NS	9.45E–03	NS	NS	NS	1	AA
KCNH5	394,639	NS	NS	NS	NS	2.20E–03	NS	NS	NS	1	AA
KCNH8	412,117	NS	NS	NS	NS	1.31E–03	NS	NS	NS	1	AA
KCNK5	40,504	NS	NS	NS	NS	2.58E–03	NS	NS	NS	1	AA
KCNK9	159,677	7.10E–03	NS	NS	NS	NS	NS	NS	NS	1	EA
KCNQ3*	422,993	NS	NS	NS	NS	4.52E–03	NS	NS	NS	1	AA
KCNS2	3,775	NS	NS	NS	NS	2.24E–03	NS	NS	NS	1	AA
KCNU1	270,959	NS	NS	NS	NS	2.28E–03	NS	NS	NS	1	AA

^*Of note, the genes indicated by an asterix are those that have previous evidence of association or functional significance in schizophrenia. See text for relevant references

To summarize, our review of previous literature found significant independent findings supporting various ion channel candidates in genetic, pharmacologic and functional neuroscience investigations and minimal replication of specific genes across diverse genetic investigations while the present analyses found strong replication of KCN gene set enrichment across methods and ethnic samples. Thus, we submit that this confluence of results provides especially compelling evidence, not only for a role of ion channel genes in mediating risk of schizophrenia, but also for an architecture characterized by substantial genetic heterogeneity. By extension, these findings suggests that the particular KCN genes that have been idiosyncratically subject to independent investigation may represent only a small proportion of the KCN genes likely to be relevant to the genetic architecture of schizophrenia.

Future directions

Though more computationally demanding, some emergent GSA applications enable the incorporation of information on LD structure into the SNP-to-gene mapping procedures. Such approaches represent an alternative to the more conventional use of arbitrary distances based on gene boundaries and may offer some advantages in cross-sample comparisons.

Findings from ours and future GSAs in schizophrenia may be most useful to direct targeted resequencing efforts. While whole genome and whole exome sequencing are becoming more feasible, the capacity to store, manipulate and analyze such large scale data are major rate-limiting steps in their use for genetic discovery. In addition to being computationally more manageable, the use of targeted resequencing around biologically rational candidates will be more immediately interpretable. In addition, the development of gene set based analytic approaches to whole-genome sequencing data will enable both exploratory and hypothesis-driven analysis on a larger scale.

Ultimately, the confirmation of disease-related functional variants in ion channel genes will have profound implications for biomarker and pharmacologic development. As common markers and rare functional variants are identified across the population of affected individuals, catalogs can be developed from which biomarker arrays can be constructed for diagnosis and risk profiling. The ultimate application of such genetic information will be in their use for personalized treatment plans. Though still far short of their potential diversity and scale, existing ion channel modulating drugs comprise an extremely successful and highly profitable drug class (Xie et al. 2004). This pharmaceutically tractable class of molecules holds the potential for highly specialized molecular targeting. Building on currently available targeting mechanisms, further developments may enable more specialized ion channel class (e.g., calcium channel), subclass (e.g., inwardly rectifying potassium channels), protein domain (e.g., ionic pore) or conformation state (e.g., open, closed or inactivated) targeting. Finally, the evolution in gene-based therapies may offer the possibility of altering expression or functioning of specific ion channels.