Most of the SNPs assessed by these commercial microarrays were chosen in ways that are not based on hypotheses about the underlying biology of any particular disorder. However, to the extent that there is a body of knowledge concerning the biology of a disorder, not all genes may be as likely, a priori, to contain disease associated variants. Thus, if a commercial microarray is used for a GWAS, we might ask – how well does the microarray cover biologically relevant genes for which there is a priori reason to believe their products are involved in the disease of interest [3]? For example, because genes that encode nicotinic cholinergic receptors have a clear biochemical connection to nicotine dependence, as do alcohol dehydrogenases for alcoholism, we should be especially vigilant in testing the hypothesis that variants in these genes might influence addiction vulnerability.

These results suggest that a significant amount of common genetic variation in these addiction related genes is unaccounted for by these commercial SNP microarrays. The deficiency is particularly high in the African sample. This is likely due to the lower LD in this older population, which means more SNPs are required for tagging. While the Illumina 1M clearly provides the best coverage, we would still need to add 23,441 SNPs to tag all common SNPs in the HapMap African sample for these 910 genes. The best-case scenario is when the Illumina 1M is used for European-Americans. But even in this case, there are still 5,117 SNPs that are not well represented.

Table 2 shows some examples of coverage by the Illumina 610 Quad microarray for ten genes that are of particular interest. We chose this array because it offers a median level of coverage among the seven arrays we studied. These genes were among the most highly prioritized by the addiction researchers with whom we consulted. The selection process involved a number of criteria, including pharmacogenetic pathways, gene expression data, and mouse models. For example, CDH13 (Cadherin 13) is known to be expressed in neurons in the human adult cerebral cortex, midbrain, thalamus and medulla regions [5]. Because it is also known to inhibit neurite extension [6] and activate a number of signaling pathways [7]–[10], it is a strong candidate for the genetic study of addiction phenotypes [11]. CDH13 contains 2,414 SNPs that are common in the African population, and only 50% of these are tagged by the by the Illumina 610 Quad microarray. Figure S1 shows the complete distribution of individual gene coverage percentages using our primary set of 910 genes for the Illumina 610 Quad microarray in each population.

We have designed a SNP database (available at http://zork.wustl.edu/nida/neurosnp.html) to systematically determine how to supplement these commercial microarrays for addiction. Our database includes a SNP prioritization score based on the genomic information network (GIN) method introduced by S. Saccone and colleagues [4]. This method was originally designed to systematically incorporate a priori biological hypotheses into the prioritization of SNPs after a genome-wide association study. The method begins with a set of SNPs that are ranked by their association p-values, and then increases the rank of a SNP when it is determined to be biologically relevant to the phenotype according to an a priori set of conditions, such as being in a biologically relevant gene, and additionally, perhaps, being a missense mutation. The score is a measure of biological relevance to addiction, and can be used independently of association p-values to prioritize which SNPs are selected to supplement commercial microarrays. The score incorporates SNP/gene functional properties (such as coding and promoter regions), human/mouse evolutionary conservation, and a quantitative trait locus (QTL) mapping method that utilizes mouse models to identify genes associated with addiction phenotypes (Chesler and colleagues, submitted). Figure S2 shows the distribution of prioritization scores for our genome-wide SNP database, and Figure S3 shows the GIN network model we used to model addiction, which was adapted from the nicotine dependence model used by Saccone and colleagues [4].

In addition to our primary set of 910 genes, the mouse systems genetics method (Chesler and colleagues, submitted) that identified 7,842 additional genes with potential biological relevance to addiction through mouse QTL and gene expression correlation analysis and the GIN prioritization scores reflect this quantitative assessment of biological relevance. Genes with a large number of mouse associations are prioritized more highly, and those with a relatively low number receive little increase in the prioritization score relative to arbitrary genes (see the methods section for details). These additional data provide a broader measure of biological relevance to addiction which may be useful for prioritizing SNPs for further study after a GWAS [4] or fine mapping a region of genetic linkage. This method has the effect of combining information from the expert nomination process and the mouse systems genetics data. SNPs in the 486 expert nominated genes, the determination of which did not involve the mouse data, receive a uniform increase in priority. If there is additional evidence from the mouse data of relevance of the gene to addiction, the priority is increased further depending on the extent of the evidence, which is measured by the number of mouse phenotypes that link to the gene.

Table S2 shows the distribution of phenotypes that map to mouse genes, both for the entire set of mouse genes considered and for the top 424 genes that were used for our primary analysis of SNP microarrays (these were mapped to human genes via NCBI Homologene). More detailed information on this latter set of genes can be found in supporting file S1 which is discussed in more detail below. Complete details on the data and experiments for this mouse systems genetics project are described in Chesler and colleagues (submitted).

In order to determine the coverage of regions inferred to be undergoing recent adaptive selection [12], [13], all SNPs detected by the LD decay (LDD) test in the Perlegen and HapMap datasets were compared to the Illumina 1M and Affymetrix 6.0 SNPs. Uncovering evidence for recent selection is an additional approach to defining functional human genomic variation. The LDD test identifies alleles undergoing selection by searching for an expected increase with distance in the fraction of inferred recombinant chromosomes surrounding a selected variant. This method is insensitive to local recombination rate because it relies on LD differences between the two alleles at a site, while the local rate influences the extent of LD surrounding both alleles. While over 99.9% of the selected regions defined by the LDD test fall within +/−10 kb of a SNP present on these microarrays, there are some important exceptions. For example, the extensive LD surrounding the selected DRD4 7R allele [14] is not captured by these arrays, which contain very few SNPs in the region (only 1 in 100 kb). In general, however, the extensive long-range LD exhibited by these recently selected alleles (up to 1 Mb), and the current density of microarray SNPs, indicates that most of these evolutionarily important alleles can be “tagged” by an adjacent SNP surrogate.

The combined set of 910 genes used for our analysis of SNP microarrays is available as a spreadsheet in supporting file S1. The spreadsheet contains detailed annotation, including the logical category used by the NeuroSNP project, such as “Nicotine System” and “Dopamine System” (further documentation of these categories and other columns is contained in the spreadsheet – see the sheet labeled “Column Descriptions”). Other columns include the Entrez Gene ID and gene symbol, the full name of the gene as well as all known symbol aliases and alternative descriptions, build 36.2 physical mapping data and mouse homologs. Some columns contain links to external databases, such as GenoPedia (http://www.hugenavigator.net/HuGENavigator/startPagePedia.do), which contains a list of all human diseases that have been linked to the gene, including links to publications. The spreadsheet also contains links to the Knowledgebase of Addiction Related Genes (KARG, http://karg.cbi.pku.edu.cn) [15], and also GeneNetwork (http://genenetwork.org) for additional information on mouse systems genetics data. We have also created a web site (http://zork.wustl.edu/nida/neurosnp.html) that contains a searchable database of this set of genes, as well as downloadable files for the gene and SNP databases. These resources will allow investigators to both gather new biologically relevant targets for genetic association studies of addiction, and also to discover new information on well-known targets, such as the extent of tagged coverage in various population by commercial SNP microarrays.

These methods and resources were developed with the intention that they would be useful to researchers who wish to test a priori biological hypotheses, either within the context of a GWAS, or for a more targeted study such as studying specific addiction-related genes or fine mapping a region of genetic linkage for an addiction-related phenotype. The need for this kind of approach has been discussed in the literature [3], and to this end we have used multiple domains to develop a collection of genes with evidence of biological relevance to addiction. The biological principles guiding the selection criteria, such as biochemical pathways and expression data, do not necessarily imply the existence of genetic variants within these genes that influence addiction phenotypes. For example, metabolic pathways related to nicotine are an obvious source for cataloging genes that are biologically relevant to genetic studies of nicotine dependence, but these normal biological systems do not necessarily involve genes with variants that influence abnormal phenotypes. Therefore, the utility of these methods and resources depend on the subjective preferences of investigators on the genetics of addiction and their specific a priori biological hypotheses [3].

While the primary utility of the resource we have developed is the supplementation of commercial SNP microarrays, it has several other useful applications. For example, the GIN prioritization scoring method is useful for interpreting the results of a GWAS, and can be used to prioritize SNPs for further study after a GWAS [4], as well as prioritize tests of gene-gene interaction. As the genomic coverage of commercial SNP microarrays improves, and subsequently whole-genome sequencing becomes the new standard, the problem of multiple testing will continue to hinder progress in understanding complex interactions underlying the genetics of addiction and other complex diseases. Therefore tests of gene-gene interaction will in most cases require a mechanism of prioritization, and our database will be a useful resource for this approach. By limiting tests of gene-gene interaction to genes that have a biological connection to the phenotype, the issues of multiple testing and computational tractability are substantially reduced. And even within a set of biologically relevant genes, using the GIN prioritization scores to further refine interactions tests, such as testing only within the top 100 SNPs ranked by these scores, will further reduce the problem.

Table 1 shows that a substantial amount of variation in high priority regions for addiction is currently unaccounted for by these commercial microarrays. Ultimately, the actual number of SNPs used to supplement a microarray will depend on the genotyping platform being used for supplementation. This platform could involve a different technology than that of the original microarray being supplemented. For example, the Affymetrix 6.0 array could be supplemented with custom genotyping on the Illumina GoldenGate platform (http://www.illumina.com), and because some of the SNPs used for Table 1 may not perform well on the GoldenGate platform, the numbers reported in the table may need to be reduced.

An additional 424 genes were nominated based on behavioral genetic analysis of 255 measures of addiction related phenotypes obtained by the Tennessee Mouse Genome Consortium in the recently expanded panel of over 60 BXD recombinant inbred mouse lines (Chesler and colleagues, submitted). Because we are interested in identifying addiction associated pathway members that are polymorphic in humans, but not necessarily polymorphic in mice, our approach identified both candidate genes for mouse phenotypic variation, and biomolecular correlates of the mouse phenotypes. Each gene was chosen because it either resided in a significant or suggestive QTL interval or was a gene expression correlate of multiple addiction-related phenotypes. The criteria used to place genes on the list included correlations with p<.0001, and QTLs, with genome-wide permutation p<.05 or p<.33, the conventional thresholds for significant and suggestive loci. Convergence of evidence of these effects in multiple addiction assays was the criteria used for list membership. Mouse genes were then mapped to human genes via the HomoloGene database (http://www.ncbi.nlm.nih.gov/sites/entrez?db=homologene).

To explore genetic variation in this combined set of 910 human genes, we created a general purpose genome-wide SNP relational database. The foundation for this database was Build 128 of dbSNP (http://www.ncbi.nlm.nih.gov/projects/SNP), which was our source of physical mapping data and SNP/gene functional properties. We then examined the SNP coverage of these genes as provided by seven commercial microarrays: the Illumina HumanHap300 Duo, HumanHap550, HumanHap650Y and Human1M (http://www.illumina.com), and the Affymetrix Genome-Wide Human SNP Array 5.0 and 6.0 (http://www.affymetrix.com). To assess genomic coverage of common SNPs by these microarrays, we used genotype data for four populations from the International HapMap Project, Public Release 23a (http://www.hapmap.org): African (Yoruba people of Ibidan, Nigeria – YRI), Chinese (Beijing – CHB), European-Americans (CEPH – CEU), and Japanese (Japan – JPT). To estimate LD, we used the program HaploView (version 4.0, http://www.broad.mit.edu/mpg/haploview) [18] to estimate r² for all SNPs within 500 kb of each other. The commonly used condition r²≥0.8 was used to assess whether a SNP is tagged through LD in a given population by a given SNP microarray. General database management was done with a combination of SAS [19] and Perl [20].

To provide a mechanism for prioritizing SNPs when supplementing SNP microarrays for addiction, we used the genomic information network (GIN) technique introduced by Saccone and colleagues [4]. The GIN method assigns each SNP a numeric prioritization score indicating the biological relevance for addiction: the higher the score, the greater the priority. Figure S3 shows the network model we used for addiction, which is a modification of the nicotine dependence model used by Saccone and colleagues. The score incorporates a number of factors, including SNP/gene functional properties (such as coding and promoter regions), and evolutionary conserved regions (ECRs, provided by ECRbase [21], http://ecrbase.dcode.org). The original GIN method introduced by Saccone and colleagues incorporated LD into the prioritization score through the use of LD proxies. This is more appropriate when prioritizing SNPs for replication after an initial GWAS. In our current implementation, where we are selecting SNPs to supplement arrays for the discovery phase of a GWAS, we have eliminated the LD component in order to avoid redundancy among the selected SNPs.

Table S1

Table S2

Figure S1

Figure S2

Figure S3

File S1

File S2