Genomewide association study of opioid dependence: multiple associations mapped to calcium and potassium pathways

Abstract

Background

We report a GWAS of two populations, African- and European-American (AA, EA) for opioid dependence (OD) in three sets of subjects, to identify pathways, genes, and alleles important in OD risk.

Methods

Design employed three phases (based on separate sample collections). Phase 1 included our discovery GWAS dataset consisting of 5,697 subjects (58% AA) diagnosed with opioid and/or other substance dependence (SD), and controls. Subjects were genotyped using the Illumina OmniQuad microarray, yielding 890,000 SNPs suitable for analysis. Additional genotypes were imputed using the 1000 Genomes reference panel. Top-ranked findings were further evaluated in Phase 2 by incorporating information from the publicly available SAGE dataset, with GWAS data from 4,063 subjects (32% AA). In Phase 3, the most significant SNPs from Phase 2 were genotyped in 2,549 independent subjects (32% AA). Analyses were performed using case-control and ordinal trait designs.

Results

Most significant results emerged from the AA subgroup. Genomewide-significant associations (p<5.0×10⁻⁸) were observed with SNPs from multiple loci - KCNC1*rs60349741 most significant after combining results from datasets in every phase of the study. The most compelling results were obtained with genes involved in potassium signaling pathways (e.g., KCNC1, KCNG2, and KCNA4). Pathway analysis also implicated genes involved in calcium signaling and long-term potentiation.

Conclusions

This is the first study to identify risk variants for OD using GWAS. Our results strongly implicate risk pathways, provide insights into novel therapeutic and prevention strategies, and may provide biologically bridge OD and other non-SD psychiatric traits where similar pathways have been implicated.

INTRODUCTION

Opioid dependence (OD) is associated with serious medical, legal, and social problems, and co-occurring psychiatric disorders. The cost of OD to society in 2002 was ~$181 billion (1). Although genomewide association study (GWAS) is the method of choice to identify risk genes for complex traits, none has been published for OD, despite an estimated heritability of >0.60 (2). The strongest GWAS-derived noteworthy and replicable genomewide significant (GWS) results so far for drug dependence (DD) traits identified a set of loci mapping to a chromosome 15 nicotinic receptor gene cluster (e.g., ref. (3)) for nicotine dependence (ND) and related traits; we also reported a GWS association of the FAM53B locus to cocaine dependence (CD) (4).

Few DD GWAS studies have been attempted, and those that have been published are underpowered by modern standards, partly because they used dichotomous traits (i.e., DD diagnoses). Here, we used a relatively large sample and augmented power with an ordinal trait analytic design that allowed us to take into account both the presence or absence of OD and the severity of affection (including the ability to distinguish between subjects with zero and those with one or two symptoms). This increased power by enabling us to use more of the available phenotypic information than standard diagnosis-based analyses. Some of these strategies have been used previously in successful efforts to map ND risk alleles, most notably the use of large clinical samples (3). We further increased our analytic power by including, for some analyses, data from the Study of Addiction: Genetics and Environment (SAGE) sample (5, 6), which includes SD trait information. This dataset is available to the scientific community through an application process and will henceforth be referred to as “public domain.”

Our GWAS discovery sample consisted of 2,379 European Americans (EA), including 1,383 subjects with OD; and 3,318 African Americans (AA), including 683 subjects with OD. A second phase sample of 4,603 EAs and AAs from the SAGE study, and a third phase sample including 2,549 EAs and AAs ascertained in a manner identical to that of the discovery sample, were used to replicate and extend our findings.

Thus, our study took place in three “phases” that differed with respect to samples and genotyping. “Phase 1” designates our own GWAS sample. “Phase 2” designates the addition of SNP data from SAGE (which used a very different recruitment strategy, but similarly ascertained subjects, and was genotyped on a different microarray) combined with our sample by meta-analysis; this is the core of the GWAS discovery and replication strategy. “Phase 3” designates our own smaller replication sample, where individual SNPs rather than GWAS arrays were genotyped. With these strategies, we identified genetic variants that increase risk for OD and related heritable traits.

MATERIALS AND METHODS

Subjects and Diagnostic Procedures

The (Phase 1) GWAS discovery sample included 5,697 subjects. A second identically ascertained sample comprising 2,549 subjects was used for replication (Phase 3). All of these subjects were recruited for studies of the genetics of drug (opioid or cocaine) or alcohol dependence (AD). The sample consisted of small nuclear families (SNFs) originally collected for linkage studies, and unrelated individuals. Subjects were recruited at five eastern US sites (Table 2). Our previous OD linkage study (7) included a subset of the SNFs included in this study. Subjects gave written informed consent as approved by the institutional review board at each site, and certificates of confidentiality were obtained from NIDA and NIAAA.

Table 2.

(a) Demographic and diagnostic information for the entire sample. Yale University School of Medicine (APT Foundation), New Haven, CT; University of Connecticut Health Center, Farmington, CT; the University of Pennsylvania School of Medicine, Philadelphia, PA; the Medical University of South Carolina, Charleston, SC; and McLean Hospital (Harvard Medical School; Belmont, MA.; (b) Subject characteristics

A. Recruiting sites	Phase 1 (SNFs)				Phase 1 (Unrelateds)				Phase 3				Site Total
	AA		EA		AA		EA		AA		EA		AA	EA	Total
Recruiting site	Male	Female	Male	Female	Male	Female	Male	Female	Male	Female	Male	Female
Yale (APT Foundation)	199	257	141	108	453	370	485	290	223	198	474	477	1,700	1,975	3,675
University of CT	174	227	155	161	455	355	451	296	127	93	315	299	1,431	1,677	3,108
MUSC	42	84	52	47	53	109	33	29	21	24	47	47	333	255	588
McLean Hospital	44	36	42	30	10	6	18	11	0	2	2	3	98	106	204
Univ. Pennsylvania	9	11	0	0	288	136	20	10	51	64	43	39	559	112	671
										Phase 2
SAGE (Phase 2)	1083		736		2235		1643		643	668	1222	1530	1,311	2,752	4,063
Total Phases 1+2+3													5,432	6,877	12,309

B. Subject Characteristics
	African Americans				European Americans
Phase/OD Status	#	Male (%)	Mean age	Mean symp. count	#	Male (%)	Mean age	Mean symp. count
Phase 1/OD cases	683	62	44.2	6.1	1383	60	37.2	6.5
Phase 1/OD Abusers	49	76	44.8	1.5	57	72	36.9	1.6
Phase 1/Controls	2550	49	40.6	0.04	854	55	39.6	0.1
Phase 1/Exposed Controls	639	61	42.2	0.14	366	64	37.1	0.21
Phase 2/OD cases	105	61	39.9	6	193	60	35.6	5.9
Phase 2/OD Abusers	2	100	40.5	0	5	100	41	0
Phase 2/Controls	1204	48	40	0.03	2554	43	38.8	0.03
Phase 2/Exposed Controls	158	67	40.8	0.2	383	69	35.9	0.2
Phase 3/OD cases	142	67	45.2	5.9	407	60	35.6	6.4
Phase 3/OD Abusers	10	90	44.8	2.2	22	55	36.9	1.1
Phase 3/Controls	601	47	39.9	0.02	1255	46	42.3	0.02
Phase 3/Exposed Controls	85	78	42.3	0.13	137	61	36.7	0.16

All subjects were interviewed using an electronic version of the Semi-Structured Assessment for Drug Dependence and Alcoholism (SSADDA) (8, 9) to derive DSM-IV diagnoses (10) of lifetime OD and other major psychiatric traits. The test-retest (κ=0.94) and interrater (κ=0.91) reliability of the OD diagnosis was excellent (9).

Genotyping and Quality Control

Samples for Phase 1 were genotyped on the Illumina HumanOmni1-Quad v1.0 microarray containing 988,306 autosomal SNPs, at the Center for Inherited Disease Research (CIDR) and the Yale Center for Genome Analysis. Genotypes were called using GenomeStudio software V2011.1 and genotyping module V1.8.4 (Illumina, San Diego, CA, USA).

Follow-up genotyping (Phase 3 sample) was performed using a custom Illumina GoldenGate Genotyping Universal-32, 1536-plex microarray. Most SNPs included in the custom array were selected for studies of other phenotypes. Additional SNPs were genotyped individually using the TaqMan method (11).

44,644 SNPs on the microarray and 135 individuals with call rates < 98% were excluded, and 62,076 additional SNPs were removed due to minor allele frequencies (MAF) <1%. After data cleaning and quality control, 5,697 individuals and 889,659 SNPs remained for imputation. We identified several instances where identical DNA marker profiles were linked to two different interview forms. When demographic information (sex, date of birth, number of reported children) was consistent across interviews, one sample was randomly removed from analysis; when demographic information was inconsistent, both were removed. Genetic relationships were examined in the family-based sample by calculating pairwise identity by descent (IBD) proportion estimates using PLINK (12). Pairs of individuals whose IBD proportions did not match their reported genetic relationship were assigned to two different families and pairs of individuals who shared more than 25% of their alleles IBD were assigned to the same family. Self-reported males with X chromosome heterozygosity > 20% and self-reported females with heterozygosity < 20% were excluded unless their true identity could be determined.

To verify and correct the misclassification of self-reported race, we compared GWAS data from all subjects with genotypes from the HapMap 3 reference CEU, YRI, and CHB populations. Principal components (PC) analysis was conducted in the entire GWAS sample using Eigensoft (13, 14) and 145,472 SNPs that were common to the GWAS dataset and HapMap panel (after pruning the GWAS SNPs for linkage disequilibrium (LD) (r²) >80%) to characterize the underlying genetic architecture of the samples. The first PC score distinguished AAs and EAs; these groups were subsequently analyzed separately. We then conducted PC analyses within the two groups and the first three PCs were used in all subsequent analyses to correct for residual population stratification.

Public Domain GWAS Sample: SAGE

In the Phase 2 analyses described below, we included publically available GWAS data from SAGE (http://www.ncbi.nlm.nih.gov/projects/gap/cgi-bin/study.cgi?study_id=phs000092.v1.p1) (15). The SAGE dataset contained 1,311 AA and 2,750 EA unrelated individuals (Table 2). SAGE includes individuals from the Collaborative Study on the Genetics of Alcoholism (COGA) (5), the Family Study of Cocaine Dependence (FSCD) (6), and the Collaborative Genetic Study of Nicotine Dependence (COGEND) (15). The COGA sample is a set of unrelated individuals recruited in Indiana, New York, St. Louis, Connecticut, Iowa, and San Diego selected for genotyping from a larger set of 8,000 subjects. Cases met criteria for DSM-IV AD. FSCD contained cases and controls from the greater St. Louis metropolitan area. All cases met criteria for DSM-IV AD and most also met criteria for DSM-IV CD. Controls were from the same communities and had consumed alcohol, but had no lifetime history of substance dependence. A subgroup of FSCD subjects was not alcohol dependent, but had a lifetime diagnosis of DSM-IV cannabis dependence or dependence on another illicit drug. COGEND subjects were recruited in MO and MI. Cases met criteria for DSM-IV alcohol and/or ND. SAGE samples were genotyped on the Illumina Human 1M array containing 1,069,796 total SNPs. Control subjects were selected from the non-dependent population and did not meet criteria for AD, ND, or illicit DD.

Genotypes were imputed with IMPUTE2 (16) using the genotyped SNPs and the 1000 Genomes reference panel released in June 2011 (http://www.1000genomes.org/), which contains phased haplotypes for 1,094 individuals of various ancestries, including 379 samples of European descent (CEU, FIN, GBR, IBS, and TCI), 286 of Asian descent (CHB, JPT, and CHS), 181 admixed American samples (PUR, CLM, and MXL), and 246 samples of African descent (ASW, LWK, YR) (17). EA and AA samples were imputed separately. Genotyped SNPs with Hardy-Weinberg equilibrium P-values <10⁻⁵ were set to missing and imputed genotypes were used instead. We retained 18,564,419 SNPs with derived information content > 0.8 in at least one of the population groups.

Analytical Models

1. We employed a model with the imputed minor allele dosage as the dependent variable used DSM-IV symptom count (Sympcount_adj) for OD and each of three other major SD diagnoses (CD, AD, and ND) as ordinal predictors (with adjustment for sex, age, and the first three PCs of ancestry [Figure S1 in Supplement 1]). The mutual adjustment of OD for measures of dependence on other substances facilitated the identification of genetic risk factors unique to each substance and limited confounding due to comorbid dependencies. All individuals contributed to this analysis, including those meeting DSM-IV criteria for OD (who met 3–7 criteria) and those having no symptoms of OD. The ordinal trait model has greater power to detect genetic associations than a univariate model based on disease status because of greater information content and improved specificity of the dependence measure. The β coefficient and p-value for the OD symptom count (adjusted for the symptom counts for CD, AD, and ND) were used to assess the magnitude and significance of the association, respectively. Ordinal trait data were derived for all samples included in the study.

2. Case-control status was the outcome in models that included as controls only individuals who had used opioids at least once without becoming dependent. This accounts for the fact that subjects who were never exposed might be better classified as phenotype “unknown” than as “unaffected.”

Association tests in the Phase 1 and Phase 2 GWAS datasets were performed using linear or logistic association models embedded in generalized estimating equations (GEE) to correct for correlations among related individuals (18). The Phase 3 sample containing unrelated individuals only was evaluated using linear and logistic models. All models were adjusted for age, sex, and the first three PCs of ancestry. The data were analyzed separately within population groups and the results were combined by meta-analysis using the inverse variance method implemented in the computer program METAL (19). In Phase 2, Phase 1 SNPs with p-values<10⁻⁴ in either population were tested for association in the SAGE dataset in the specific population group(s) and statistical model(s) yielding the specified result in Phase 1. Tests were restricted in this manner to minimize their number and because it is unlikely that a SNP not meeting the cut-off for follow-up would attain a noteworthy significance level when combining results from the discovery and replication samples. Phase 1 and 2 results were combined within population groups by meta-analysis. In Phase 3, we selected for replication ⁷⁶ SNPs (Sympcount_adj, N=44 and case-control, N=28, including 58 unique SNPs) with p<10⁻⁴ in either population in the Phase 2 analyses.

Multiple Testing Considerations

A p-value of 5.0×10⁻⁸ was used as a threshold for GWS. Results were not adjusted for testing in two populations because we tested three distinct a priori hypotheses: (1) SNPs are associated with OD and related traits in AAs, (2) SNPs are associated with OD and related traits in EAs, and (3) associations are evident with the same SNPs in AAs and EAs (in meta-analysis). Results were not adjusted for testing two phenotypes, as the rank order correlation between the symptom count and case-control variables is 1.0 (that is, no affected subject can have a lower symptom count than a control), and the point-biserial correlation is 0.95.

Pathway Analysis

Meta-analyzed GWAS results (separately by population) from the discovery and SAGE datasets were used to identify biological pathways related to OD. First, the number of independent SNP association tests for each gene in the genome was computed according to the method of Li and Ji (20). Next, the smallest p-value for an individual SNP within each gene was multiplied by the number of independent tests in that gene to create a list of genes significantly associated with OD after correcting for the number of tests within that gene (P_adj<0.05). The significant genes were evaluated by pathway analysis, performed using the Ingenuity Pathway Analysis (IPA) software suite (http://www.ingenuity.com) to identify an overrepresentation of selected genes within canonical pathways that were defined using information culled from multiple sources (Kyoto Encyclopedia of Genes and Genomes, interactome studies, manual curation, etc). A Fisher’s exact p-value for each pathway indicated whether the pathway contained more significantly associated genes than expected by chance. To validate the findings, we uploaded the same gene list into the Database for Annotation, Visualization and Integrated Discovery (DAVID; http://www.david.niaid.nih.gov) and performed a gene enrichment analysis using its functional annotation tools.

RESULTS

As noted, three independent datasets were employed in these OD-phenotype analyses, in three phases. EAs and AAs were analyzed separately in each phase. In Phase 1, GWAS was conducted in a discovery dataset containing 5,697 subjects. In Phase 2, the first replication phase, SNPs with p-values <10⁻⁴ in either or both populations combined were tested for association in the SAGE dataset including 4,063 subjects. In Phase 3, the second replication phase, SNPs with p-values <10⁻⁴ in either population in the Phase 2 analyses were evaluated in a set of 2,549 unrelated subjects ascertained in the same manner as the Phase 1 sample. Meta-analyses were conducted in a combined sample of 12,309 subjects. Analyses of symptom count accounted for comorbidity with other SD symptoms (sympcount_adj).

In both Phase 1 analyses, we identified population-specific variants and variants associated with risk in both EAs and AAs and (Figures S2 and S3 in Supplement 1]), including GWS associations in AAs with PITPNM3 (rs9913974, p=8.4×10⁻⁹) for the case-control model and KCNC1 (rs60349741, p=1.3×10⁻⁸) and HHLA2 (rs73204138, p=3.7×10⁻⁸) for the Sympcount_adj model. The p-values for these findings are negligibly inflated (Figures S4 and S5 in Supplement 1]) and the SNPs were well imputed (Table S1 in Supplement 1]). We also identified a possible association with a DISC1 polymorphism (rs2738888, AA p=8.47×10⁻⁴; EA p=1.74×10⁻³; meta p=5.22×10⁻⁶ (discussed below)).

In Phase 2, we repeated the primary OD analyses for SNPs with p<10⁻⁴ (403 SNPs for the case-control model, 1,522 SNPs for the Sympcount_adj model in AAs) in the independent SAGE sample. Four distinct regions contained GWS SNP sympcount_adj p-values in AAs after meta-analysis of the results from Phases 1 and 2 (Table 1). These included the regions containing the potassium voltage-gated channel genes KCNG2 (rs62103177, p=6.7×10⁻¹⁰; Figure 1) and KCNC1 (rs60349741, p=7.5×10⁻⁹; Figure S6 in Supplement 1), and APBB2 (rs115368721, p=1.8×10⁻⁹; Figure S7 in Supplement 1), and PARVA (rs73411566, p=1.8×10⁻⁸; Figure S8 in Supplement 1). Several GWS results from Phase 1 did not remain so in Phase 2, including SNPs mapped to HHLA2 and PITPNM3.

Table 1.

Genomewide significant association results for measures of opioid dependence – adjusted symptom count model. Imputation qualities ranged from 0.91–0.99 (Supplementary Table 3).

Chr	BP	SNP	SNP Type	Gene	Phase 1		Phase 2		Phase 1+2	Phase 3		Phase 1+2+3
					MAF	P-value	MAF	P-value	P-value	MAF	P-value	P-value
4	41056716	rs114070671*	intron	APBB2	0.08	4.58E-06	0.08	8.27E-05	2.26E-09	0.09	0.55	1.04E-07
4	41056839	rs115368721	intron	APBB2	0.08	3.74E-06	0.08	1.16E-04	1.92E-08		--not tested --
11	12454744	rs73411566	intron	PARVA	0.03	5.48E-05	0.03	3.06E-05	1.75E-08		--not tested --
11	17796111	rs60349741	UTR_3	KCNC1	0.06	1.33E-08	0.07	7.79E-02	9.26E-09	0.05	0.45	4.81E-07
18	77624479	rs62103177	intron	KCNG2	0.06	8.00E-05	0.06	1.10E-07	6.68E-10	0.06	0.15	3.60E-10

^*LD proxy for rs115368721

Figure 1.

Regional association (Manhattan) plot showing Phase 1+2 meta-analyzed results for association of Sympcount_adj with SNPs mapped to the KCNG2 region on chromosome 18 in the AA population. The SNPs are color coded according to r² with the most significant SNP shown in purple. The SNP with the strongest evidence for association is shown twice: once with the Phase 1+2 meta-analysis p-value (purple with “crosshairs”) and once after the inclusion of the Phase 3 samples (purple without “crosshairs”). (They are nearly superimposed in this figure.) The light blue line and right Y-axis show the observed recombination rate in the HapMap YRI samples.

Although no GWS association signals identified in the meta-analysis of the results from Phases1+2 were significantly associated with the corresponding trait in the much smaller Phase 3 dataset at p<0.05, three of the results improved or remained significant after meta-analysis that included the Phase 3 data (Table 1). The associations of rs62103177 in KCNG2 with Sympcount_adj improved to p=3.6×10⁻¹⁰ (Figure 1). Two of the GWS SNPs (rs115368721 in APBB2, rs73411566 in PARVA) in Phase 2 were imputed and assays could not be designed to genotype them in Phase 3. LD-based proxies were used for APBB2 SNPs, but no suitable proxies (i.e., r²>0.4) were available for the PARVA SNP.

Pathway Analysis

Pathway analysis was performed including genes with at least one SNP with a gene-based corrected p-value <0.05. Separate analyses were performed with gene sets obtained from the case-control (62 genes) and Sympcount_adj (198 genes) models. The most remarkable pathways were identified from the case-control model. The first and third most significant pathways resulting from this analysis were calcium signaling (p=0.002, false discovery rate (FDR)=0.15, Figure 2) and synaptic long-term potentiation (LTP) (p=0.004, FDR=0.17, Figure 3), respectively. (The second most-significant canonical pathway (p=0.0024) captured genes that influence cardiac hypertrophy.) The Ca²⁺ signaling pathway was identified by significant associations in genes encoding calcineurin A (PPP3CA), calcium/calmodulin-dependent protein kinase II beta (CAMK2B), histone deacetyltransferase 9 (HDAC9), and cAMP responsive element binding protein 5 (CREB5). Ca²⁺ signaling is also critical to LTP; thus, this pathway emerged as significant in part due to associations with PPP3CA, CAMK2B, and CREB5 (Table 3). Reanalysis using the DAVID software revealed that the top gene ontology functional category was “calcium ion binding” (p=0.0089) and the top KEGG pathway was “LTP” (p=0.034). These pathways (complete list, Table S2 in Supplement 2) are identical to the pathways identified by IPA.

Figure 2.

The role of genes identified in GWAS of Phase 1+2 AA subjects using a case-control model in the canonical pathway “calcium signaling.” The gene or gene families with a significant association result are shown in orange. The calcineurin family was identified by an association in PPP3CA, the CREB family by CREB5, the HDAC family by HDAC9, and the CAMK2 family by CAMK2B (see table “pathway SNPs”). For a full description of the functional relationships represented by the various lines and arrows, see:http://www.biolreprod.org/content/suppl/2010/09/29/biolreprod.110.085910.DC1/biolreprod.110.085910-3.pdf.

Figure 3.

The role of genes identified in GWAS of Phase 1+2 AA subjects using a case-control model in the canonical pathway “synaptic long-term potentiation.” The gene or gene families with a significant association result are shown in orange. The calcineurin family was identified by an association in PPP3CA, the CREB family by CREB5, and the CAMK2 family by CAMK2B (see table “pathway SNPs”). For a full description of the functional relationships represented by the various lines and arrows, see:http://www.biolreprod.org/content/suppl/2010/09/29/biolreprod.110.085910.DC1/biolreprod.110.085910-3.pdf.

Table 3.

Pathway SNPs

chr	SNP	Gene	P	BP	Total SNPs	Independent SNPs	Adjusted P
7	rs2075069	CAMK2B	1.27E-06	44284564	61	20	2.55E-05
4	4-102128801	PPP3CA	3.15E-04	1.02E+08	67	18	5.66E-03
7	rs9639575	CREB5	1.19E-04	28417214	135	52	6.16E-03
7	7-18189274	HDAC9	2.40E-05	18189274	219	73	1.75E-03

DISCUSSION

We report here results for a GWAS for OD in two different population groups in the United States. We made use of GWAS and replication samples collected by us and available data from the SAGE project, which were collected using a similar psychiatric interview, were used for replication. Several of the top-ranked genes encode proteins that participate in potassium and calcium signaling pathways. Although calcium signaling genes have been studied in addiction biology, they were not previously considered key genetic candidates. These findings therefore add substantially to our knowledge of the biology of OD. Our most compelling results were obtained in the AA population.

KCNC1 and KCNG2, loci containing some of the most significantly associated SNPs, encode potassium voltage-gated channel subunits. Another locus, PITPNM3, encodes a protein that is involved in phosphatidylinositol transport, but that also binds calcium. Variants located ~196 kb upstream from PPP3CA were associated with OD in the case-control model (although not GWS) in both populations (rs6419156, P_AA=1.4×10⁻⁶, P_EA=5.6×¹⁰⁻⁷); the effect directions were opposite, suggesting that population-specific causal variants or a single causal variant occurring on distinct haplotypes may be responsible. PPP3CA encodes a calcium-dependent, calmodulin-stimulated protein phosphatase involved in calcium signaling. (A trinucleotide repeat at this locus was reported to be nominally associated to SD in AAs but not EAs (21)).

Calcium signaling and LTP (which relies on calcium signaling for neurotransmitter release) were two of the best-supported pathways identified in pathway analysis (although these results did not meet FDR-defined significance after correction for the number of pathways evaluated). Thus, potassium and calcium transport and signaling mechanisms appear to play essential roles in OD risk.

Possibly relevant to these findings, CACNA1C (which encodes a voltage-dependent calcium channel) is one of the best-supported GWAS-identified risk genes for bipolar affective disorder (22) and schizophrenia (23). Further, a large mega-analysis of psychiatric traits also landed on several calcium-system genes as contributing to risk for a set of psychiatric illnesses: two of the four GWS SNPs mapped to voltage-gated calcium channel subunit genes, and calcium signaling pathways were identified as important for the five disorders studied (24). Potassium-calcium signaling may couple neuronal signaling to vasodilation in the brain (25) and opioids can regulate calcium conductance via increasing potassium conductance in μ-opioid receptors (26). One of the genes in the calcium signaling pathway that is strongly associated with OD risk, CAMK2B, modulates activation of ionotropic (including AMPA) glutamate receptors (27). AMPA-mediated glutamatergic signaling in the hippocampus plays an important role in context-dependent sensitization to morphine. Behavioral sensitization in animal models is of interest as a neural basis of addiction (28, 29). LTP (which underlies learning and memory) was also implicated directly in pathway analysis. These pathways may interact to produce biologically important effects both on behavior following an initial exposure to opioids and on impaired control over use following chronic exposure (i.e., relapse risk, a key issue in the natural history and treatment of OD).

There were additional GWS associations (albeit not-yet-replicated) of biological interest. PARVA, encoding alpha parvin, which is reported to play roles in cell adhesion and cytoskeletal organization, was associated in the GWAS for OD (Phase 2 meta p=1.75×10⁻⁸). APBB2 (Phase 2 meta p=2.26×10⁻⁹) encodes a member of a family of proteins that bind the amyloid precursor protein that may be important for signal transduction and is associated with Alzheimer disease risk (30). A closely related gene, APBB1, was shown previously to be associated to ND (31).

An interesting possible OD risk gene identified in Phase 1 was DISC1, which (among other functions) regulates development of the cerebral cortex. DISC1 was originally identified as a schizophrenia risk gene in a cytogenetic study (32) and considerable evidence since has tied it to cognitive phenotypes in human (33) and animal (34) models. We observed nominally significant associations with DISC1 in both EAs and AAs (meta p=5.2×10⁻⁶). Subsequent to this common-variant association finding, we used deep sequencing at this locus to identify a set of rare variants in DISC1 that contribute to risk for OD (35). Thus, this study ties OD to other cognitive phenotypes via a common gene, and, as schizophrenia was one of the five disorders identified as associated to calcium pathways (24), also through gene pathways. Previously, we reported that another schizophrenia-associated locus, NRG1, is associated with cannabis dependence in AAs (36).

Strengths of this study are a relatively large sample size, including SAGE data to increase power. All subjects were assessed using comprehensive, standardized instruments. Our primary sample has unusually good representation of AA subjects. Most of the strongest findings emerged from this understudied population; some of these were supported by findings from the EAs. We employed several novel analytic strategies, including models defining OD symptoms as a quantitative trait adjusted for addiction to other substances. Such approaches minimize concern about the misclassification of controls because SD requires exposure to the drug. An overall picture emerges of genes and pathways involved in regulating calcium and potassium signaling as major contributors to OD risk.

Our study has several limitations. The GWS and other top-ranked findings were identified with imputed SNPs; however, empirically, our results show that the top-ranked SNPs in our study were imputed with excellent quality. Although our strongest results were with imputed SNPs, they are not rare. (Two had MAF=0.08; two, 0.06; and one, 0.03.) The primary concern about rare (MAF<1%) SNPs is the greater rate of false positive results possible due to large effect sizes (i.e., greatly increased risk) derived from small allele frequency differences between cases and controls or genotyping errors. However, studies examining this issue have concluded that nominally significant results occur less frequently than expected, even for low MAF SNPs, provided there is no genotyping or imputation quality bias between cases and controls (37, 38). Moreover, the concern about false positives is substantially less when the results are replicated or supported by evidence from multiple samples, which is the case for the associations that we reported. Also, although our sample size is reasonable, experience has shown that many true associations are detected only in samples or meta-analyses that are much larger. The relatively small size of the sample used in Phase 3 for the primary trait analyses and as replication for the OD subtype findings is in our view the most likely cause of few replication observations in that sample. Specifically, there were proportionally fewer AAs available for Phase 3 (n= 805, ~24% as large as the discovery sample), which is the part of the sample in which most positive results were initially observed. Ideally, the replication sample would be larger than the discovery sample. It is also noteworthy that there was little evidence of association in EAs for any of our top findings in AAs, however, all but one of the SNPs reported in Table 1 were monomorphic in EAs. Finally, our findings are not adjusted for testing association in two populations and with two trait models (case-control and ordinal). Bonferroni correction is too conservative given the high correlation between the traits and distinct hypotheses for EAs and AAs, populations that often have distinctive risk loci and/or distinctive risk alleles at the same locus. Future studies in large independent populations are necessary to address some of these concerns.

In conclusion, we identified numerous novel GWS associations with OD. We found compelling evidence for association in several brain systems, most consistently for calcium and potassium signaling pathways, which should provide the impetus for further research on these brain systems in OD. Further, these findings are largely consonant with a set of SNP and pathway-based findings that are emerging for a broader range of neuropsychiatric traits, and as such, may bridge OD and other neuropsychiatric phenotypes.