The opioid epidemic continues to cause human and economic devastation in the United States and elsewhere [1]. U.S. deaths due to the misuse of opioids reached a record high of more than 42,000 in 2016, which prompted the U.S. Department of Health and Human Services to declare opioid abuse to be a public health emergency in 2017. Several factors led to the epidemic, including the over-prescription of opioid analgesics, increased availability of heroin and illicit synthetic opioids [2], and possibly progressively worsening labor market opportunities, especially for those with low levels of education [3]. Despite early interventions to curb over-prescription, the epidemic continued to worsen [4]. Cocaine dependence (CD) has received less attention recently but its U.S. prevalence (likely significantly underreported [5]) remains at about 1%, resulting in a significant public health burden [6]. Cocaine users experience mortality rates four to eight times higher than the general population [7] and have an increased risk of suicide [8]. In contrast to opioid dependence (OD), pharmacological treatments for CD do not exist [9].

Moderate to high heritability estimates for both OD (43–50%) [10] and CD (65–78%) [11, 12] strongly suggest a genetic component to risk. Several OD risk genes have been identified through modestly-sized genome-wide association studies (GWAS) [13–16], but they have not been replicated and explain only a tiny fraction of the total trait heritability. Similarly, CD risk genes have been identified and there is evidence of genetic overlap between CD and other psychiatric disorders [17, 18], but collectively these results explain little trait heritability. Two relatively large OD GWAS papers have been published recently [Polimanti et al. [16] (4,503 OD cases) and Zhou et al. [19] (28,317 OD cases)]. These studies identified a single genome-wide significant (GWS) association for OD: a coding variant in the μ-opioid receptor gene OPRM1 (opioid receptor mu 1 gene). The sample sizes available for these studies still lag far behind those available for other complex diseases with a similar public health impact, and no large scale GWAS for CD has been performed. This problem is exacerbated by the fact that controls exposed to illicit drugs (a prerequisite for developing OD or CD) represent a small subset of the total control population. They may also be limited by the lack of differentiation between dependence upon prescription opioids and heroin, which may have a different set of risk factors.

In light of the deficits in our understanding of the genetic factors contributing to OD and CD risk, we have sought to derive phenotypes that are under more direct genetic control that could facilitate understanding the disorders and identification of potential drug targets and risk pathways. Our group recently identified variants near OPRM1 associated with usual methadone dosage in a sample of individuals with OD [20]. Here, we report findings from a GWAS of two related phenotypes: the time between first reported use and first DSM-IV diagnosis of OD and CD. We analyzed each substance independently. Both cocaine and opioids are highly addictive [21]. Approximately 50% of individuals who ever use heroin develop OD [22], while about four percent of people who try cocaine become addicted within two years, with another 16% in a prodromal stage of addiction [23]. A few studies (one conducted in the primary cohort analyzed here) have examined risk factors associated with rapid progression to dependence. Conduct disorder and childhood physical abuse predicted rapid more development of OD and CD, while and alcohol and nicotine dependence diagnoses were associated with slower progression, and African Americans (AAs) progressed to OD more rapidly than European Americans (EAs) [24].

There is also an interplay between use of medically indicated or recreationally used prescription opioids or heroin in determining who develops OD. Prescription opioids used for pain relief are generally safe when taken for a short time and as prescribed by a doctor. In a U.S. veteran population, individuals who used prescription opioids recreationally had 19 times the odds of initiating heroin use, fewer than four percent of people who abuse prescription opioids initiate using heroin within five years [25]. Despite our inability to distinguish between disease course with regard to specific opioids and their pattern of initiation, these related phenotypes have two properties that make them attractive for variant discovery: namely (1) they do not include individuals who were never exposed to opioids or cocaine and (2) they may be better indicators of an individual’s genetic risk for OD/CD than case-control status because they distinguish between cases who develop the disorder slowly (i.e. are more genetically resistant) and those who develop it rapidly (i.e. are highly genetically susceptible), assuming these differences are partially under genetic control.

Across all cohorts, the mean TD was shorter for opioids than cocaine (Tables 1 and 2). In Yale-Penn, older age at interview was strongly associated with longer TD for both CD [hazard ratio (HR) = 0.97 per year, P = 3.51 × 10⁻⁹⁷] and OD (HR = 0.98 per year, P = 3.02 × 10⁻²⁷). EA ancestry was associated with a longer TD for CD (HR = 0.79, P = 3.70 × 10⁻¹⁸) and a shorter TD for OD (HR = 1.43, P = 1.20 × 10⁻²¹). Female sex was associated with longer TD for CD (HR = 0.97, P = 0.046) and shorter TD for OD (HR = 1.12, P = 0.001). Shorter TD was significantly correlated with a higher number DSM-IV dependence criteria endorsed (i.e. more severe dependence), with Pearson correlation coefficients of 0.54 and 0.53 in AAs and EAs, respectively, for OD and 0.44 and 0.48 in AAs and EAs, respectively, for CD. The distribution of DSM-4 CD and OD criteria counts in Yale-Penn are shown in Supplemental Figure 1.

Boxplots for the distributions of age at onset and the TD for each substance in Yale-Penn are shown in Supplemental Figure 2. There were 4,989 individuals exposed to both cocaine and opioids that contributed to both analyses. There were 2,908 individuals with a lifetime DSM-4 dependence diagnosis for both drugs.

In the discovery sample, we identified tone region containing variants associated with TD at a GWS level (P < 5 × 10⁻⁸) in the gene family with sequence similarity 78-member B (FAM78B) on chromosome 1 (rs61835088, P = 2.96 × 10⁻⁸) for CD in the combined EA and AA sample. Each copy of minor (C) alleles at rs61835088 was associated with 0.57 fewer years between first use and CD diagnosis among those who converted. The Manhattan plots for CD and OD in Yale-Penn AAs and EAs are shown in Supplemental Figure 3. There was no evidence for inflation of the test statistics for either substance in either EAs or AAs (Supplemental Figure 4).

No significant gene-based tests were observed after correcting for the number of valid gene-based test results, which varied by population and dependence outcome. The top ranked gene (P = 4.5 × 10⁻⁶) in gene based tests was calcium voltage-gated channel subunit alpha1 B (CACNA1B) in the EA CD analysis. One gene set (membrane depolarization) showed a significant enrichment, with six of 83 genes, including CACNA1B, showing nominally significant gene based test results after Bonferroni correction (P = 0.04) for CD in EAs.

Three hundred and fifty one variants with suggestive P-values < 1.0 × 10⁻⁵ in the discovery sample (149 for CD, and 202 for OD) were tested in the replication samples. Several of these did not yield a valid result due to low MAF or imputation quality in their respective ancestry group or substance. Valid results were obtained for 142 CD SNPs and 200 OD SNPs in COGA and 170 OD SNPs in CATS. After correcting for the number of tests performed, none of the GWS hits in the discovery met criteria for replication. Several variants were at or near nominal significance with the same effect direction as Yale-Penn in COGA/CATS and improved the association signal after meta-analysis. The strongest signal in any of the replication samples was in the OD analysis with rs8063946 (P = 3.70 × 10⁻⁴ in COGA AAs) in the gene FTO alpha-ketoglutarate-dependent dioxygenase (FTO).

After combining all results, the association of rs61835088 with CD TD strengthened (HR = 0.87, P = 3.71 × 10⁻⁹, Table 3, Figure 1). For the same outcome, the association with intergenic chromosome 21 SNP rs2825295 in AAs surpassed the GWS threshold in the combined Yale-Penn/COGA meta-analysis (P = 2.19 × 10⁻⁸, Table 3, Figure 1). Each minor (T) allele at rs2825295 was associated with 0.43 fewer years between first cocaine use and dependence among those who converted. The top variant in regions with P-values < 1.0 × 10⁻⁶ after combining all data are shown in Tables 3 (CD) and 4 (OD in AAs). For OD TD, all but one of the top associations was in the AA portion of the sample. There was one OD TD signal in the combined EA and AA sample, rs8063946 (P = 1.87 × 10⁻⁷) in the gene FTO, which for clarity is not shown in Table 4. Near-GWS evidence of association was obtained in the combined samples with SNPs in three other genes previously associated with addiction phenotypes: rs114341823 in GRIN2B [44, 45], which encodes the glutamate ionotropic receptor NMDA type subunit 2B was associated with OD TD in AAs (P = 1.45 × 10⁻⁷); the variant in FTO [46–48] described above; and rs73721103, which is between alpha-aminoadipic semialdehyde synthase (AASS) and LOC102724527 but is ~100 kilobases from PTPRZ1 [49, 50], which encodes the protein tyrosine phosphatase receptor type Z1 with CD TD in AAs (P = 2.81 × 10⁻⁷).

None of the variants we identified result in amino acid changes and any function they might have is likely regulatory, or as a result of linkage to a functional variant. Several of the top CD TD SNPs showed evidence for regulatory potential in various databases. According to the Genotype Tissue Expression database (https://gtexportal.org/home/) rs61835088 is a significant eQTL for FAM78B in tibial nerve tissue. This variant was also an eQTL for that gene in prefrontal cortex and blood according to QTLbase (http://mulinlab.org/qtlbase) [51]. Several variants, including rs61835088, rs73404786 were significant methylation QTLs (mQTL) in prefrontal cortex. Rs73721103 was a significant eQTL for the gene ring finger protein 133 (RNF133) in blood. Only one OD TD SNP showed evidence for regulatory potential: rs11728570 as a significant mQTL in blood. Neither SNP Function Prediction (https://snpinfo.niehs.nih.gov/) nor Braineac (http://braineacv2.inf.um.es/) provided any additional information about potential functionality of the top results.

Here we present the results from GWASs of CD and OD related phenotypes measuring the time interval between first use of cocaine or opioids and a DSM-IV dependence diagnosis on the respective drug. These analyses, made possible by the extensive phenotyping performed on three independent cohorts, identified two GWS associations, one specific to AAs and one in the full sample, between relatively common variants and enhanced susceptibility to CD that were not detected at GWS in GWAS that used the bivariate case-control status for dependence as outcomes. Both of these were strengthened after combining meta-analysis with the replication sample. To our knowledge, this is the first time these trait definitions have been studied using GWAS. Although none of the GWS associations in the discovery sample were replicated after correcting for the number of variants tested in the COGA and CATS cohorts, 13 (9%) of the CD TD SNPs and 11 (5.5%) of the OD TD SNPs were nominally significant in one of those samples with the same effect direction.

The function of FAM78B, associated with progression from cocaine use to dependence, is not well understood but it is highly expressed in the brain [52] and interacts with several other proteins with diverse functions [53]. According to the STRING protein-protein interaction database (as https://string-db.org) [54], the proteins with which it is predicted to interact include ones related to mitochondrial function (ATP5F1), neurofilament network integrity (SNCG), and NOTCH signaling (XXYLT1).

Other genes that were among the top hits, but not GWS, have functions with potential links to SUDs. The genes GRIN2B and FTO, associated with progression from opioid use to dependence; and PTPRZ1, associated with cocaine TD have all previously been linked to SUD traits. Different variants than the one identified in this study in GRIN2B have been associated with OD [45] and chronic ketamine use [44]. Its protein product directly binds calcium/calmodulin-dependent protein kinase 2-alpha (encoded by CAMK2A), which can lead to synaptic long-term potentiation by facilitated CAMK2 response to synaptic calcium [55]. We previously identified this pathway as an important modulator of OD risk [13] and this result strengthens the evidence that an individual’s propensity to establish and reinforce substance-specific neural circuits may be an important factor driving their genetic predisposition to SUDs, and that calcium homeostasis may at least partially drive that propensity. FTO has also been associated with addictive behaviors. Although initially identified as a regulator of eating and obesity traits [56, 57], though probably not through an effect of FTO itself but rather as a consequence of variants in the gene that affect expression of IRX3 and IRX5 [56, 58]. FTO has also been associated with alcohol dependence [46–48] and connectivity in a dopamine-dependent reward circuit of meso-striato-prefrontal regions of the human brain [59]. PTPRZ1 binds to two neurotrophic cytokines [pleiotrophin (PTN) and midkine (MK)] that, along with other functions, contribute to the extinction of cocaine and amphetamine-seeking behaviors [49, 50] and limit morphine withdrawal syndrome [60]. The variants we identified, however, were ~100 KB away from the gene and may not strongly suggest a role for this gene in SUD risk. The top gene identified through gene based tests, CACNA1B in EAs for CD TD, encodes a pore-forming subunit of the presynaptic neuronal voltage-dependent calcium channel. Genes from this family of neuronal regulators of calcium and potassium levels were associated with their respective traits in our previous OD and CD GWAS papers [13, 18]. This, along with the fact that the membrane polarization gene set was significantly enriched among our results provides further evidence for the role of genetically determined differences in synaptic membrane potential acting as mediators of addiction risk.

We also looked up the results for the top TD SNPs in our previously published CD [18] and OD [13] GWAS of case-control status where controls were exposed to cocaine or opioids. There was modest evidence for association of each TD GWS variant identified with risk of dependence (not TD) on the respective substances, although not at the GWS level. Rs61835088 (P = 0.007) and rs2825295 (P = 0.001) were associated with binary CD diagnosis, and rs4860439 (P = 1.11 × 10⁻⁶) and rs7032521 (4.10 × 10⁻⁷) with OD diagnosis. The OPRM1 OD variant (rs1799971) identified by Polimanti et al. [16] and Zhou et al. [19] was nominally associated with OD TD in Yale-Penn EAs (P = 0.05).