Association genome scanning can help determine which chromosomal regions and genes contain allelic variants that predispose to substance abuse. This approach does not require family member participation, gains power as genomic marker densities increase (17-20), identifies smaller chromosomal regions than linkage-based approaches, fosters pooling strategies that preserve confidentiality and reduce costs (21-24), and provides ample genomic controls that can minimize the chances of unintended ethnic mismatches between disease and control samples.

Our initial pooled association genome scans compared allelic frequencies at 1,494 SNPs in unrelated polysubstance abusers and controls (24). Forty-one SNPs displayed nominally “reproducibly positive” allele frequency differences between abuser and controls in both European- and African-American samples (24). These reproducibly positive markers cluster closer to each other and to positive markers from linkage studies of addictions than anticipated by chance (25, 26). However, this density of SNP markers provided information about possible associations with addiction for only a modest number of the blocks of restricted haplotype diversity found in these subjects' genomes.

“10k” SNP microarrays (Affymetrix) use size-selected PCR products of genomic restriction fragments that have been ligated to universal linker sequences and amplified by using single PCR primer pairs. These arrays allow studies of more SNP markers. This information adds to previous data sets. Only a single SNP is assessed in both 1,494 and 10k arrays. This approach thus improves the power of pooled association genome scanning by seeking associations between addiction vulnerability and allelic variants located on more of the blocks of restricted haplotype diversity found in these subjects' genomes. However, this approach also requires careful validation before use with pooled samples, because it differs from the method previously validated for pooled samples and 1,494 SNP arrays (24).

We thus now report validation and use of pooled association genome scanning by using 10k arrays hybridized with size-selected amplicons from end-ligated XbaI genomic DNA restriction fragments of pooled genomic DNAs. DNAs come from European- and African-American polysubstance abusers who report dependence on at least one illegal substance vs. matched controls free from any reported significant lifetime use of any addictive substance. We use this approach to generate >57 million person genotype equivalents from quadruplicate determinations. We discuss the convergent data that these results provide, the genetic architecture for polysubstance abuse that the results support, the classes of candidate genes that they nominate for roles in human addiction, and the implications of these findings for pooled association genome scanning approaches to complex genetic disorders.

Research volunteers provided informed consents, self-reported their ethnicities, and provided drug use histories and Diagnostic and Statistical Manual of Mental Disorders (DSM)IIIR or IV diagnoses, as described (24, 27, 28). Three hundred ninety-two unrelated European-American “abusers” averaged 35 years of age and total drug use scores of 2.87 and demonstrated DSMIII-R or DSMIV dependence on at least one illegal abused substance. 240 “control” European-Americans averaged age 32, reported no significant lifetime histories of use of any addictive substance, and averaged 0.61 total drug use scores. Four hundred forty abusers of self-reported African-American descent averaged 34 years of age and total drug use scores of 2.98. One hundred eighty-one controls averaged 36 years of age and total drug use scores of 0.53. We excluded subjects with DSM = 2 scores who did not reach DSM dependence criteria for an illegal substance. DNAs from 316 of these African-American and 559 of these European-American subjects were assessed for 1,494 SNPs in different blocks of restricted haplotype diversity as described (24). No results for DNAs from 304 African-American and 73 European-American subjects have been previously reported.

Genomic DNA was prepared from blood as described (24, 27, 28), quantitated by spectrophotometry, picogreen and Hoechst dye fluorescence, and diluted to 10 ng/μl. Validation studies compared: (i) allelic determinations from individual Center for the Study of Human Polymorphisms (CEPH) DNAs and (ii) results from pools of the same DNAs. Test-retest variability was also studied.

Allelic frequencies in polysubstance abusers and controls were compared by using pools made by combining equal amounts of DNA from 20 individuals of the same ethnicities and phenotypes. We used hybridization probes prepared from genomic DNA as described (Affymetrix GeneChip Mapping Assay Manual) with precautions to avoid contamination. Fifty nanograms of pooled genomic DNA was digested by XbaI, ligated to adaptor, amplified by PCR by using 3-min 95°C hot start; 35 cycles of 20 sec, 95°C; 15 sec, 59°C; 15 sec, 72°C; and a final 7-min 72°C extension. PCR products were purified (MinElute 96 UF kits, Qiagen, Valencia, CA), digested for 30 min with 0.04 unit/μl DNase I to produce 30- to 200-bp fragments, end-labeled by using terminal deoxynucleotidyl transferase and biotinylated dideoxynucleotides, and hybridized to 10k arrays (Affymetrix), which were stained and washed as described (Affymetrix Gene-Chip Mapping Assay Manual) by using immunopure streptavidin (Pierce), biotinylated antistreptavidin antibody (Vector Laboratories), and R-phycoerythrin streptavidin (Molecular Probes). Arrays were scanned and fluorescence intensities quantitated by using an Affymetrix array scanner, as described (24).

Allele frequencies for each SNP in each DNA pool were assessed based on hybridization intensity signals from four arrays, allowing assessment of hybridization to the 10 sense and 10 antisense “perfect match” cells on each array that are complementary to the PCR products from alleles “A” and “B” for each diallelic SNP. Each array was analyzed as follows: (i) “Background” values, the average fluorescence intensity from the 5% of cells with the lowest values, were subtracted from the fluorescence intensity of every cell. (ii) Background-subtracted values were normalized by division by the average value obtained from the 5% of cells with the highest values. (iii) Normalized hybridization intensities from the 20 array cells that corresponded to the perfect match “A” and “B” cells for each SNP were averaged. (iv) “A/B ratios” were determined by dividing average normalized A values by average normalized B values. (v) Arctangent transformations were applied to each ratio to aid combination of data from arrays hybridized and scanned on different days. (vi) Average arctan values from the four replicates of each experiment were determined. (vii) Mean and standard deviations of average arctan values for each diagnostic and ethnic group were calculated. (viii) Mean arctan A/B ratios for abusers were divided by mean arctan A/B ratios for controls to form abuser/control ratios for European- and African-American samples. (ix) A t statistic for the differences between abusers and controls of each racial/ethnic group was generated by using the formula:

Alleles for some SNPs were genotyped in individual samples from those subjects whose DNAs had not been analyzed before the current report (Supporting Text, which is published as supporting information on the PNAS web site).

Chromosomal positions of 11,482 SNPs as well as previously linked and associated markers were determined by using National Center for Biotechnology Information and NetAffx (Affymetrix) data. Forty SNPs could not be accurately positioned.

Although there is no universally accepted method for analyzing association genome-scanning data, we used preplanned analyses based on those that have been favorably reviewed (29). We identified SNPs with abuser/control ratios in the top or bottom 2.5% of all abuser vs. control comparisons in European- and African-American samples and assessed t statistics. We focus on the 38 “nominally reproducibly positive” SNPs that met at least three of the four criteria: top or bottom 2.5% outlier status for abuser/control differences in each of two ethnic/racial samples and 5% outlier t values scores in each of the two ethnic/racial samples (Table 1).

Table 1. Thirty-eight SNPs that meet criteria for “reproducibly positive” differences between substance abusers and controls

To seek chromosomal clustering of SNPs that displayed abuser/control allelic differences in the current data, we identified the 51 SNPs from the current arrays that lay within 0.1 Mb of one of the 38 nominally reproducibly positive SNPs. This distance is arbitrary but allows comparisons with previous analyses (30). We then assessed the extent to which these nearby SNPs met these same four criteria. To seek convergence between current and other association data, we repeated these analyses with data sets consisting of (i) the current nominally reproducibly positive SNPs, (ii) SNPs that met the same criteria in association analyses of samples from unrelated alcohol-dependent and control individuals sampled from the Collaborative Study on the Genetics of Alcoholism (COGA) as part of the Genetics Analysis Workshop (C.J. and G.R.U., unpublished work), selecting the alcohol-dependent proband who displayed maximal use of dependence on illegal substance from each COGA pedigree for analyses, and (iii) SNPs associated with addiction vulnerability in prior studies (24). We also sought convergence between the current data and the simple sequence length polymorphism markers that had been linked to addiction vulnerability in prior studies by using ± 5-Mb intervals to assess convergence with these prior linkage results (Supporting Text).

To provide a control for the possibility that the abuser/control differences observed at the nominally positive SNPs were due solely to occult ethnic/racial differences in the frequencies of alleles at these same SNPs between abusers and controls, we subtracted the arctan A/B ratios for African-American abusers from those for European-American abusers and the values for African-American controls from those for European-American controls. These differences were averaged for each SNP to form an averaged ethnicity difference score. Kolmogorov-Smirnov testing (graphpad prism v. 3.00, GraphPad, San Diego) assessed whether the distribution of averaged ethnicity difference scores for SNPs that displayed abuser/control differences with the largest t values differed significantly from the distribution of these ethnicity difference scores for all SNPs.

To assess the power of our current approach, we used the observed standard deviations and mean abuser/control differences for the SNPs that provided the largest differences between control and abuser population means for European- and African-American samples. We used the program ps v. 2.1.31 (http://biostat.mc.vanderbilt.edu/twiki/bin/view/main/powersamplesize), within-group standard deviations corresponding to 0.15, 0.1, 0.05, 0.025, and 0.01 differences in means, α = 0.05, and sample sizes and abuser/control ratios from the current data set.

Observed results were compared with those expected by chance by using 100,000 Monte Carlo simulation trials, as described (24).

To provide insights into some of the genes likely to harbor variants that contribute to individual differences in addiction vulnerability, we selected a candidate gene for each where this was possible. We identified candidate genes when the reproducibly positive SNP lay (i) within the gene or (ii) in 3′ or 5′ flanking sequences that were contained on a block of high restricted haplotype diversity along with exon sequences from that gene, as defined by haploviewer (http://www.hapmap.org) using data from U.S. Center for the Study of Human Polymorphisms (CEPH) individuals. Markers that could not be clearly localized or did not meet these criteria are annotated as n/a in Table 1.

Pooled genotyping displayed features that supported the validity of results. Regression analyses that examined the relationships between (i) “observed allele ratios,” background-subtracted, normalized, arc-tan transformed hybridization intensity ratio values obtained from pools of Center for the Study of Human Polymorphisms (CEPH) DNAs, and (ii) “expected allele ratios,” the fraction of A and B alleles obtained from individually genotyping these same individuals yielded average slopes of 1.01. Ninety-seven percent and 80% of SNPs displayed observed/expected correlation coefficients >0.9 and >0.97, respectively. Ninety percent and 55% of SNPs displayed standard deviations of replicate results of <10% and <5% of mean values, respectively.

Abuser/control hybridization ratios for the 11,522 SNPs examined here fell into nearly Gaussian distributions with mean values close to one for both European- and African-American samples (Fig. 1). Hybridization intensities for sense vs. antisense oligonucleotides displayed 0.78 r² values. Mean arctan A/B ratios ± average SEM for pool-to-pool differences for the 11,522 SNPs were 0.79 ± 0.021 for European-American abusers and 0.79+/-0.025 for controls and 0.79 ± 0.0019 for African-American abusers and 0.79 ± 0.028 for controls. Mean SEMs for the differences among the four replicate studies of each DNA pool were 0.033.

Fig. 1. Main axes: Chromosomal distributions (labels 1-22 and X) that map abuser/control ratios to the chromosomal position of each corresponding SNP for European- (blue) and African-American (red) samples. The positions of the SNPs whose data yield outlier abuser/control (more ...)

For analyses, we selected 582 candidate positive markers that represented the 2.5% of SNPs with greatest and the 2.5% of SNPs with the smallest abuser/control ratios in European-Americans (values >1.12 or <0.88) and assessed which of these SNPs also displayed the greatest 2.5% or smallest 2.5% differences between allelic frequencies in African-American abuser and control samples (values >1.13 or <0.89) (Fig. 1; yellow symbols; 30 were expected by chance; Monte Carlo P < 0.001).

SNPs for which the t values for abuser/control differences lie in the upper 5% of all t values (nominal P < 0.05) are also highlighted in yellow on the supplemental axis to the right of each chromosomal depiction.

Thirty-eight “nominally reproducibly positive” SNPs display outlier abuser/control differences and t values (Table 1). We identified a candidate gene for many of the 38 nominally reproducibly positive SNPs based on information from mapviewer (Golden Software, Golden, CO) and Unigene (National Center for Biotechnology Information), the proximity between the gene and the nominally reproducibly positive SNP, and the quality of data supporting the gene (Table 1).

To provide a control for the possibility that observed abuser/control differences might reflect occult ethnic/racial stratification and correspondingly different allelic frequencies at the SNPs that display these abuser/control differences, we compared the distributions of differences between European- and African-American allele frequencies for (i) all SNPs and (ii) SNPs for which abuser/control differences achieved nominally significant t values. Kolmogorov-Smirnov testing revealed no differences in the distribution of the ethnic differences in the frequencies of alleles in these two groups (P = 0.88 and = 0.89 for European- and African-American comparisons, respectively).

The power to detect abuser/control allele frequency differences of 0.01, 0.025, 0.05, 0.1, and 0.15, based on the number of pools and the observed pool-to-pool variability was 0.3, 0.65, 0.85, 0.95, and 0.95 and 0.25, 0.5, 0.7, 0.85, and 0.85 for European- and African-American samples, respectively.

We identified evidence for clustering of results from assessments of data from the 51 SNPs that were positioned within ± 0.1 Mb of each of the 38 nominally reproducibly positive SNPs. Thirteen of these 51 nearby SNPs displayed either abuser/control ratios in the upper or lower 2.5% of all abuser/control ratios and/or nominal t values in the upper 5% of all t values (Monte Carlo P = 0.015).

The reproducibly positive results from this data set can be compared with results from other association and linkage results for addictions. There are 84 SNPs from the current data set that lie within ± 0.1 Mb of the SNPs that were “reproducibly positive” in parallel association analyses of unrelated alcohol-dependent and control individuals sampled from the Collaborative Study on the Genetics of Alcoholism/Genetic Analysis Workshop data set (C.J. and G.R.U., unpublished work). Ten of these nearby SNPs display outlier abuser/control values in the current data set (Monte Carlo P = 0.011). Thirty-nine SNPs from the current data set lie within ± 0.1 Mb of the 42 SNPs that were “reproducibly positive” in our prior association study (24). There were six outlier abuser/control values at these SNPs (Monte Carlo P = 0.07). Individual genotyping for these SNPs, using samples from the individuals in the current study who were not included in any previous report, correlated with data obtained from pools. Pearson correlation coefficients were 0.64, 0.69, 0.60, and 0.82 for African-American abusers, African-American controls, European-American abusers, and European-American controls, respectively.

There are 136 markers that were nominally positive in a prior addiction linkage study that could be assigned reasonably accurate chromosomal positions (references in Table 2, which is published as supporting information on the PNAS web site). Twenty of these previously linked markers fell within 5 Mb of one of the 38 nominally reproducibly positive SNPs from the present study (Monte Carlo P = 0.3).

The results of this study support the efficacy of array-based pooled association genome scanning approaches and identify 38 candidate polysubstance abuse vulnerability loci. These loci are identified in both European- and African-American samples and display at least one outlier t value for the differences between abuser and control allelic frequencies. We discuss the strengths and weaknesses of these results, the ways in which they converge with results of previous association- and linkage-based studies, and the classes of genes they nominate to play roles in human substance abuse vulnerability.

The reliability and validity of the current approach are supported by data that document the reliability and validity of clinical assessments (24), test-retest correlations, correlations between sense- and antisense-hybridization patterns, correlations between pooled and individually assessed genotypes, the extent to which markers are positive in both European- and African-American samples, the extent to which some reproducibly positive markers lie near each other, and the extent to which they lie near markers defined in other studies of addiction molecular genetics. The validity of a related pooling approach using the same arrays has also been recently documented by Butcher et al. (31). Many of these results also support the idea that common allelic variants that are relatively frequent in several populations provide substantial contributions to addiction (24-26, 30).

Modeling studies support substantial power for the current methods. They also support the likelihood of both false-positive and -negative results. We document 0.25-0.95 power to detect 0.01-0.10 allele frequency differences at the SNPs used here. Modeling the current approach with the new program gene detective (D. Naiman and G.R.U., unpublished work) suggests 0.55 power to detect functional allelic variants that increase risk to a sibling 1.5-fold (λs = 1.5) (data not shown). Both of these approaches suggest that the current study can detect many of the allelic variants that predispose to addiction vulnerability in these samples, but neither indicates that the current data set is free from false negatives. It is also likely that some of the positive results listed here represent false positives. We make many comparisons in this study. In initial results of secondary analyses using multiple analysis of variance and permutation tests, we have obtained evidence for interactions between array-to-array and SNP-to-SNP differences, as well as correlations between the pool-to-pool standard deviations and abuser/control ratios (r² = 0.58) (C.J., D. Naiman, and G.R.U., unpublished work). We have evaluated the potential impact of these correlations and interactions on our choice of the 38 “reproducibly positive” SNPs in several ways. Standard deviations of pool-to-pool or array-to-array differences for the 38 SNPs were not appreciably larger than those for all SNPs (0.092 vs. 0.090 and 0.060 vs. 0.066, respectively). The mean ethnic differences in allele frequencies for the 38 SNPs were no greater than those for all SNPs (0.14 vs. 0.15). There were only modest differences between the minor allele frequencies, as reflected in the mean arctan A/B hybridization ratios for the 38 reproducibly positive SNPs and those observed for all SNPs (0.48 vs. 0.31). None of these secondary analyses thus suggests that the positive results obtained at the 38 SNPs identified here were due to the modest correlations or interactions noted in the overall data sets.

Interesting candidate genes lie near many of the reproducibly positive markers identified in this work. Reproducibly positive SNP rs1395475 lies on chromosome 4 acetaldehyde dehydrogenase/alcohol dehydrogenase gene cluster, adjacent to a number of markers linked to vulnerability to alcoholism (24, 32) and polysubstance abuse (33). Taken together, these findings underscore the likely contributions of variants at this broad locus to addiction vulnerability.

Cell signaling molecule genes that lie near reproducibly positive SNPs (Table 1) include those that signal within and between cells; G protein-, phosphorylation-, and calcium-based signaling is especially implicated by these observations. Reproducibly positive SNPs lie 5′ to the OR5AK2/3 cluster of G protein coupled “olfactory” receptor genes, within the RGS5 regulator of G-protein-mediated signaling gene, within the RYR3 channel for calcium release from intracellular stores, and in the 5′ flank of the PDE11A phosphodiesterase gene. They lie within the PPP3CA gene encodes a catalytic subunit for the calcium-dependent protein phosphatase calcineurin and in the 3′ flank of the PPP1R3A gene that encodes the 3A regulatory inhibitory subunit of protein phosphatase 1. They lie within the tetraspanin gene TM4SF13 and in the 5′ flank of the SLC35B4 transporter gene. Several of these genes display interesting features. RYR3 is expressed in brain (34) and RyR3 knockout mice display learning and locomotor changes (35). Calcineurin activities are centrally implicated in memory, including mnemonic features that may be important for addictions. The PDE11A phosphodiesterase cleaves cAMP and cGMP (36). The synaptotagmin 14 gene may also be included in this group; other family members bind calcium and alter vesicle exocytosis/endocytosis (37). The gene that encodes the SLC35B4 transporter is expressed in brain regions that include amygdala and hippocampus (38). The tetraspanin TM4SF13 l can alter lipid raft association for membrane-spanning proteins (39) and is expressed in whole brain, cerebral cortex, hypothalamus, hippocampus, sympathetic trunk, and dorsal root ganglia (40).

Gene regulatory and/or developmental genes lie near reproducibly positive SNPs. The cordon-bleu-like gene (41) is a brain-patterning gene that could modulate development of midline neural structures, including those implicated in reward and addiction (42). The PHF15/PHD finger protein 15 gene's homology to the JADE1 developmental patterning regulator suggests roles in developmental modulation (43). The GATA-binding protein 3 and LSM11 are other interesting candidate genes near reproducibly positive SNPs.

Disease-related genes also lie near nominally reproducibly positive SNPs. Atrophin-interacting protein 1 (44) is expressed largely in brain, where it interacts with atrophin, the protein in which polyglutamate expansions cause dentatorubral and pallidoluysian atrophy. Utrophin is a homolog of the dystrophin gene in which mutations can cause muscular dystrophy. Utrophin isoforms are expressed in adult brain regions (45), including frontal lobe, hypothalamus, and medulla, as well as dorsal root ganglia and the developing peripheral nervous system. OFCC1 is near chromosomal breakpoints found in cleft-palate syndromes and in a candidate locus for schizophrenia. TSSC1 is expressed in cerebral cortex, hypothalamus, medulla, and cerebellum, displays homologies with the Rb-associated protein p48 (46), and could thus be implicated in brain transcriptional repression (47).

We have identified a reproducibly positive SNP within the gene for a cell adhesion molecule neurexin 3. Normal neuron-specific neurexin 3 expression is essential for the functional organization of Ca²⁺ channels, sites for vesicular release, and NMDA glutamate receptors (48). These results, and possibly even data for the minor histocompatibility antigen HB-1, also add to previous data that nominate variants in other genes that regulate cell-cell recognition for roles in human addiction vulnerabilities (49).

The current data provide support for loci nominated in prior SNP association and linkage-based studies and produce previously undescribed “reproducible substance abuse vulnerability” regions (3, 26, 30) (Table 2). As we identify more and more of the allelic variants that contribute to substance abuse vulnerability, we will be better able to understand addictions themselves. As pooled association genome scanning is used more frequently, we believe it is likely to aid in identification of vulnerability loci for increasing numbers of complex disorders.

We acknowledge financial support from National Institute on Drug Abuse Intramural Research Program, dedicated help from Dr. Fely Carillo and other Johns Hopkins-Bayview support staff, and passionate statistical discussions and help from Dr. Daniel Naiman, Department of Mathematical Sciences, Johns Hopkins University.