In 1997 Dr. Kenneth Olden, director of the National Institute of Environmental Health Sciences (NIEHS), convened a historic conference titled “The Environmental Genome Project” held 17-18 October 1997 in Bethesda, Maryland. This symposium explored the feasibility of the Environmental Genome Project (EGP) and generated significant discussion. The EGP is designed to explore the relationship between common genetic polymorphisms and environmentally induced disease in human populations (Olden and Wilson 2000). This key symposium focused on themes that even today remain of great importance: a) the known interactions between genetic variants, environmental agents, and disease risk; b) the current and emerging technologies to identify and type DNA polymorphisms in the human genome; c) sequence diversity and human population genetics; and d) the available functional tools to analyze DNA polymorphisms.

The discussion and follow-up to this 1997 conference have had a profound impact on human genetic analysis and have formed the foundation of the EGP as well as many other large-scale projects aimed at defining the variability of the human genome (Collins et al. 1997; Olden and Wilson 2000) and studying gene- environment interactions (Collins et al. 2003). In this overview, we discuss progress in the EGP and prospects for the future in terms of the themes envisioned by Dr. Olden.

More than a century ago the link between environ-mental exposure and disease susceptibility was first recognized with the discovery of the association between exposure to coal soot and cancer in young chimney sweeps (Doll 1975). Other early examples developed from typing protein polymorphisms in human populations. These include hemolysis in individuals with glucose 6-phosphate dehydrogenase deficiency after exposure to antimalarial drugs and other oxidants (Motulsky 1972); increased risk of emphysema from cigarette smoking in individuals with ₁ antitrypsin deficiency (Eriksson 1965; Lieberman et al. 1969) and, lactose intolerance in individuals with lactase deficiency (Dahlqvist et al. 1963; Haemmerli et al. 1965; Klotz 1964). Over the past three decades, numerous links between DNA variations in the enzymes that metabolize and/or detoxify carcinogens and susceptibility to specific cancers with exposure to environmental agents have been reported (Kelada et al. 2003). In the future new technologies for directly quantifying environmental exposures will be required to improve the accuracy of gene-environment associations (Rothman et al. 1999). New system-based approaches such as proteomic and metabolic profiling will likely play a central role in these analyses and will provide quantitative data on environment exposures. These are now being implemented into studies of toxigenomics and the EGP (Waters and Fostel 2004).

The links between environmental exposures and polymorphism analysis have a long history, and association studies are clearly suited to approaching these analyses (Bell and Taylor 1997). Association studies compare the frequency of a polymorphic marker, or a set of markers, in affected and unaffected individuals. Because recombination along the chromosomes is averaged over the genetic history of the population at large, this should randomize any association between a given polymorphism and phenotype, or environmental influence, unless it is closely linked with specific alleles in the genome. Although association studies have greater power to detect variants with low relative risk, whole-genome association studies will require extremely dense sets of polymorphic markers--on the order of hundreds of thousands to more than a million markers that can rapidly be typed on large numbers of samples (Kruglyak 1999).

To develop such high-density genetic maps, studies have focused on the identification of single-nucleotide substitutions because they are the most abundant form of sequence variation in the human genome (Cooper et al. 1985; Kruglyak 1997; Wang et al. 1998). If one considers the size of the human population (~ 6 billion), with a mutation rate of approximately 2 10^-8 per base pair per generation, then every site in the genome compatible with survival has mutated an average of 240 times in just the most recent generations. However, most of these base substitutions are extremely rare in human populations. Only a fraction of the variation that exists has minor allele frequencies (MAF) exceeding 1%, and these are referred to as single nucleotide polymorphisms (SNPs). Recent estimates predict that > 15 million SNPs with MAFs exceeding 1%, and > 7 million markers with MAFs exceeding 5%, will be found in the human genome (Kruglyak and Nickerson 2001).

The discovery of SNPs in the human genome has been aided by the development of a panel of samples known as the polymorphism discovery resource (PDR) panel proposed by the National Human Genome Research Institute (Collins et al. 1998). The PDR panel was designed to discover human genetic variation while being sensitive to the ethical, legal, and social issues of population definition and not to assess the frequency of variations in specific ethnic subpopulations. Therefore, all identifying demographic information was removed from the individual samples. However, samples in this panel are representative of individuals drawn from the U.S. population, including Americans of European, African, Mexican, and Asian descent and Native Americans. Until recently, variation discovery in the EGP has focused on the PDR panel of 90 samples. This sample size is sufficient to detect polymorphic sites occurring at > 5% MAF in any one of the ethnic subpopulations (Kruglyak and Nickerson 2001). This focused the discovery efforts on the identification of common polymorphisms for association studies.

A number of strategies have been used to identify SNPs, and of these, DNA sequencing has become the dominant technology. To date, > 10 million SNPs have been uniquely mapped on the human genome (build 123; http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?db=snp). Most variants in the current database have been identified as single-base mismatches by comparing sequences from overlapping BAC (bacterial artificial chromosome) clones that were sequenced for the human genome or by comparing the reference genome sequence with sequences obtained by shotgun sequencing (Altshuler et al. 2000b; Sachidanandam et al. 2001; Venter et al. 2001). Frequency information is available for only a subset of these SNPs, although this is rapidly changing with the emergence of the HapMap data set (http://www.hapmap.org/) and the Perlegen data sets
(Patil et al. 2001; http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?db=snp).

A surrogate strategy that has emerged to identify common SNPs in the absence of real frequency data is to rely on SNPs identified by two independent discoveries for each of the two alleles (Gabriel et al. 2002; Reich et al. 2003). These are being referred to as “double-hit” SNPs, and several recent analyses have shown these variants from the database are likely to have MAFs sufficient to be detected again in another population survey (Carlson et al. 2003; Reich et al. 2003). However, as many studies have shown, the patterns of variation in the genome are influenced by a number of factors, and the analysis of these patterns will require genotype data for each site as well as its relationship to its surrounding sites (Carlson et al. 2003; Wang and Todd 2003). The availability of comprehensive genotype information greatly aids in the selection of the most useful SNP markers for large-scale genotyping. Since its inception, the EGP has focused on generating nearly complete genotype information using targeted DNA sequencing of genes across 90 PDR samples and has provided substantial insights in the variability of the human genome (Livingston et al. 2004).

Figure 1. A schematic overview of the base-excision repair pathway adapted from Matsumoto (2001). This pathway is composed of more than 30 interacting genes. The polymorphism data across this pathways allow investigators to probe disease risk not only for individual members of the pathway but also for the entire pathway of interacting genes, thereby permitting analysis of multigenic contributions to disease risk.

After the initial EGP symposium, NIEHS investigators provided substantial input into the development of a list of 550 candidate environmental response genes for targeted variation discovery. These candidates include genes involved in DNA repair, apoptosis, cell cycle control, and drug metabolism (for the complete list, see GeneSNPs at http://genome.utah.edu/genesnps). These candidates are distributed across all the human chromosomes except for the Y-chromosome, and altogether represent > 2% of all known human genes (International Human Genome Sequencing Consortium 2004). Efforts for the EGP have also focused on completing SNP discovery across entire pathways of interacting genes, such as the base-excision repair pathway illustrated in Figure 1. Other pathways include the nucleotide-excision repair, mismatch repair, double-stranded break repair, and transcription-coupled repair pathways. Many members of these pathways have been implicated in cancer susceptibility (Han et al. 2004; Ide and Kotera 2004; Mohrenweiser et al. 2002).

To date, the discovery efforts for the EGP represent the largest resequencing effort ever attempted across the human genome. A total of 371 genes have been scanned, and on average, approximately 53% of the genomic sequence for each gene has been examined for variation across the 90 PDR samples. Notably, approximately 20% of these candidate genes have already been implicated in Mendelian diseases, including disease genes for rare forms of cancer susceptibility, such as the breast cancer susceptibility loci BRCA1 and BRCA2, neurofibromin 1 (NF1), retinoblastoma locus RB1, Wilms tumor locus WT1, and ataxia telangiectasia mutated (ATM). In total, > 8.6 Mb of baseline human reference sequence has been scanned across the 90 PDR samples, generating > 770 Mb of sequence (the equivalent of resequencing human chromosome 3 four times). Sixty percent of the candidate genes have been scanned for variation across > 75% of the entire reference gene sequence, and for many, nearly complete sequences are available. For each candidate, all exons, 1.5 kb upstream of the cDNA sequence, 1.5 kb downstream of the last exon, and a significant amount of intronic sequence have been examined. These efforts have uncovered > 50,000 SNPs and produced 4.5 million genotypes, which are cataloged at the GeneSNPs database (http://genome.utah.edu/genesnps) and the dbSNP database
(http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?db=snp).

The overall nucleotide diversity () in the EGP genes is 6.7 10^-4 (equivalent to one SNP every 1,500 bp between any two chromosomes), and the SNP frequency across the 180 chromosomes averaged one SNP every 173 bp. These figures are consistent with previous genomewide estimates of nucleotide diversity and SNP frequency (Carlson et al. 2004a; Halushka et al. 1999; Li and Sadler 1991; Nickerson et al. 1998; Sachidanandam et al. 2001; Stephens et al. 2001). However, significant variance around this mean is observed, and nucleotide diversity varied more than 63-fold, from 0.72 10^-4 (equivalent to one SNP every 13,800 bp between any two chromosomes or a frequency of one SNP every 324 bp) for MARCKS-like protein (MLP) to 45.6 10^-4 (one SNP every 221 bp between any two chromosomes, or a frequency of one SNP every 63 bp) for small proline-rich protein 1B(cornifin, SPRRIB). To contrast these genes, none of the 16 variable sites in MLP had an MAF > 5% in the sequenced population (n = 90 samples), and therefore, none can be considered common in the population. In

Figure 2. A GeneSNPs (http://www.genome.utah.edu/genesnps/) view of E2F2. E2F2 represents an average of the genes scanned for polymorphism discovery based on its size and nucleotide diversity. The seven exons encoding this gene are depicted by blue rectangles for coding and green for untranslated (UTR) sequences in the mRNA. For this gene, 24 kb was scanned for polymorphisms, which includes sequences 5’ to the first exon (1.7 kb) and 3’ of the last exon (1 kb). Vertical descending lines indicate the position of the SNPs identified in this sequence. The length of the vertical lines represents the frequency of the minor allele, and the color indicates whether the SNP location is in flanking (black), intronic (brown), synonymous (yellow), nonsynonymous (red), or UTR (green) sequences.

comparison, 76 polymorphisms were identified in SPRRIB, and 43 of these variable sites (56%) were common in the population, having an MAF > 5%. The variability between these genes reveals the importance of detailed candidate gene studies. Although the average of all genes is consistent with genomewide levels of sequence diversity (Carlson et al. 2004b; Halushka et al. 1999; Li and Sadler 1991; Nickerson et al. 1998; Sachidanandam et al. 2001; Stephens et al. 2001), there is significant gene-to-gene and region-to-region variability that makes it difficult to predict the genetic structure of any given candidate or region in the genome (Clark et al. 2003). Because of this variability, there is only one gene sequenced to date, the cell cycle gene E2F transcription factor 2 (E2F2), that reflects the overall average diversity and size of the candidate genes sequenced by the EGP. A representation of the polymorphism distribution and gene structure of E2F2 is shown in Figure 2, and each of the candidate genes being examined by the EGP is available in a similar format via the GeneSNPs database
(http://genome.utah.edu/genesnps). The GeneSNPs database was developed specifically for the EGP and is so highly regarded that many of its features have been emulated by other databases, including dbSNP (http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?db=snp) and PharmGKB (http://www.pharmgkb.org/; Klein and Altman 2004).

For E2F2, 21.3 kb was scanned across the 90 PDR samples, and 112 single-nucleotide substitutions and six small insertion/deletion polymorphisms were identified. These polymorphisms are depicted by position in the gene in Figure 2 by vertical descending bars whose length is proportional to the allele frequency in the PDR. The nucleotide diversity across E2F2 is 6.9 10^-4, or one SNP every 1.4 kb between two random chromosomes. The number of common polymorphisms is similar to that of other average genes, with 41% of the total (46 of 112 SNPs) having an MAF > 5% in the PDR 90 panel. Also typical of the average gene, E2F2 has four coding SNPs (cSNPs), with two that are predicted to change the amino acid sequence (nonsynonymous) indicated by the red vertical bars in Figure 2.

As described previously (Livingston et al. 2004), the average candidate gene contains approximately 34 common SNPs. Because functional analysis via genotype-phenotype studies or by animal models is costly, reducing the number of sites (from an average of 34) is a major consideration in effective study design. Two computational approaches have been taken by the EGP project to directly identify phenotypically relevant SNPs.

One of these approaches has focused on testing the nonsynonymous (potentially functional) variations in coding sequences (Botstein and Risch 2003; Collins et al. 1997; Kruglyak and Nickerson 2001) for direct association studies and to target specific cSNPs for the development of new animal models. Of the nearly 50,000 SNPs found in the 371 candidate genes, 1,085 nonsynonymous cSNPs (ns-cSNPs) have been identified. Therefore, on average only 2% of the variability in a gene sequence is the result of amino acid substitutions.

Figure 3. The number of ns-cSNPs per gene, for EGP candidate genes containing these substitutions. Of the 371 genes scanned across the 90 PDR samples, only 221 genes (60%) had one or more ns-cSNPs.

Interestingly, despite an average of a little more than two ns-cSNPs per gene, there is substantial variability among the candidate genes, as shown in Figure 3. Of the 371 candidate genes sequenced to date, only 221 genes (60%) contained at least one ns-cSNP. Among genes with ns-cSNPs, there is substantial variation in the number of ns-cSNPs per gene, which ranges from 1 to 21. More than 15 ns-cSNPs were detected in five genes: insulin-like growth factor receptor 2 (IGF2R) with 21 ns-cSNPs, REV3-like, catalytic subunit of DNA polymerase zeta (REV3L) with 21 ns-cSNPs, protein kinase,DNA-activated, catalytic polypeptide (PRKDC) with 20 ns-cSNPs, exonuclease 1 (EXO1) with 17 ns-cSNPs, and the excision repair cross-complementing rodent repair deficiency, complementation group 6 (ERCC6) with 16 ns-cSNPs. Further analysis of these highly variable outlier genes could produce new functional insights and should be pursued in more detailed analyses.

In individual candidate genes, ns-cSNPs were further analyzed by applying two computational approaches that have been developed to detect functionally significant amino acid changes, SIFT (Ng and Henikoff 2003) and PolyPhen (Ramensky et al. 2002; Sunyaev et al. 2001). To date, 119 ns-cSNPs (~ 11% of total ns-cSNPs) have been identified as potentially deleterious by both of these approaches. Of these, only 11 potentially deleterious ns-cSNPs (dSNPs) had MAFs exceeding 5%. This important category of ns-cSNPs, the ones that commonly occur in the population, represents only a minor fraction of the total ns-cSNPs identified in the EGP (1% of the total ns-cSNPs and < 0.03% of the total SNPs identified). It is worth noting that the candidate genes with the highest number of ns-cSNPs also had multiple ns-cSNPs with predicted functional consequences based on both SIFT and PolyPhen predictions, including REV3L, PRKDC, and EXO1. IGF2R and ERCC6 did not have sufficient comparative data for accurate prediction with these two programs but are also likely to be highly polymorphic. A recent study exploring cSNPs in human genes associated with high-density cholesterol levels revealed larger numbers of rare ns-cSNPs with predicted functional significance when individuals at the extremes of the phenotypes were sequenced and compared (Cohen et al. 2004). Therefore, it is possible that perusing EGP genes with highly polymorphic coding regions will also be productive in terms of phenotype and direct functional analysis.

Figure 4. Conserved noncoding regions identified by TraFaC for CDC25A. The regulogram depicts shared cis-elements (vertical bars) between human (left) and mouse (right) sequences in the context of their sequence similarity. Identifying conserved mouse-human regions with consensus cis-regulatory elements and mapping non-cSNPs, TraFaC can be used to predict the potential adverse effects of polymorphisms on the regulation of gene and expression. The promoter region of human CDC25A and mouse Cdc25a reveals strong conservation of consensus TFBSs in relatively the same order of occurrence. The TFBSs occurring in both genes are highlighted as variously colored bars drawn across the two genes. An SNP identified in the promoter (-503 C > T) is highlighted.

The vast majority of SNPs in the human genome are in noncoding sequences (> 91%). However, our ability to predict function in noncoding sequences is limited. Several new approaches are developing to predict functional regions in noncoding sequences through the application of comparative genomics and the mining of sequences that have been highly conserved through evolutionary history (Ahituv et al. 2004; Boffelli et al. 2004a, 2004b; Dieterich et al. 2003; Frazer et al. 2004; Sandelin et al. 2004). For the EGP, TraFaC (transcription factor binding site comparison; http://trafac.chmcc.org), a web-accessible tool for identifying transcription regulatory regions using a comparative sequence analysis approach, has been applied (Jegga et al. 2002). TraFaC generates a graphical output from BLASTZ alignments by comparing sequences of human and mouse orthologs or other sequences of interest. Potential transcription factor binding sites (TFBSs) are identified as conserved blocks in the two compared sequences. An example of the TraFaC output for cell division cycle 25a (CDC25A) is shown in Figure 4. This alignment illustrates SNPs in conserved consensus TFBSs, such as the variation (-503 C > T) in a putative PAX1 (paired box gene 1) binding site in a gene with transcriptional activating properties. This potential PAX1 binding site is located in the 5´-flanking sequence of CDC25A (Figure 4) and could adversely affect the regulation of gene expression.

Insights from comparative genomics are rapidly developing, and a number of approaches are being applied to search for functionally important non-cSNPs (Berjerano et al. 2004; Pennacchio and Rubin 2001, 2003; Thomas et al. 2003). It is likely that this area will continue to be a major focus in future studies of the gene function. In this regard, ongoing efforts from The ENCODE Project (ENCODE Consortium 2004), which is focused on rapidly determining the function of genomic sequences beyond the coding sequences, will likely be expanded to genes of interest for the EGP in the near future and could provide new candidate SNPs for functional analyses.

Although new computational approaches to predict functional SNPs in the human genome are emerging, indirect association studies will ultimately be applied to identify SNPs with function that cannot be predicted a priori using approaches similar to those described above for SNPs in coding and noncoding sequences. Indirect association studies rely on linkage disequilibrium (LD) between genetic markers to measure the association between the SNPs genotyped, as well as the SNPs in LD with the assayed site and the disease phenotype (Collins et al. 1997). The number of sites required for genotyping any gene or region of the genome will greatly depend on the strength and extent of LD. For regions with strong LD and few haplotypes, only a few sites are required to represent or “tag” the gene or region. However, if the genomic region contains many haplotypes indicating low levels of LD, many more sites will be required for an association study of sufficient power. In the genes sequenced to date, there is much variability in the patterns of LD and common haplotype diversity (Figures 5 and 6). This is true across the human genome (Clark et al. 2003; Patil et al. 2001; Phillips et al. 2003; Reich et al. 2001; Stephens et al. 2001). Based on this, it is imperative to characterize the sequence and haplotype diversity of specific genomic regions of interest in human populations to rationally select SNPs for genotyping in an association study. In this respect, the EGP data set is providing new insights into these important questions in population genetics (Livingston et al. 2004; Wall and Pritchard 2003).

Figure 5. Examples of site association or LD, as measured by r2 for common SNPs (MAF > 5%), in candidate environmental response genes. (A) BNIP1 exhibits average LD with just a few blocks of correlated SNPs for an average-sized gene. (B) CCND2 exhibits low LD and less correlation among the sites. (C) NF1 exhibits strong LD for a large gene. The top portion of each graphic illustrates the visual genotypes for each gene, in which each column represents a site (blue, common homozygote; yellow, rare homozygote; red, heterozygote; gray, missing data), and each row represents a genotype for a specific individual from the PDR sorted by site and sample correlations. The bottom portion of each graph shows the LD plot for each gene, measured by r2, and depicted on a rainbow scale (white = weak or no LD; blue to red = increasing site association strength or LD).

Figure 6. Haplotypes diversity across EGP candidate genes. Haplotypes were inferred using PHASE software, version 2.1 (Stephens and Donnelly 2003) for all SNPs with an MAF ≥ 5% in the EGP samples of 90 individuals representing 180 chromosomes.

The gene-to-gene variability observed in the EGP for nucleotide diversity is also evident in site correlations or LD. Figure 5 illustrates some of the extremes observed in LD, as measured by the metric r², across the genes involved in environmental responses. For genes with average or high LD, such as BCL2/adenovirus E1B 19kDa interacting protein 1 (BNIP1) (Figure 5A) and NF1 (Figure 5C), respectively, few sites are required for genotyping in association studies. However, for genes with very weak LD, such as cyclin D2 (CCND2; Figure 5B), many more sites will be required for a genetic association study because few sites within this gene are correlated. It is important to note that the extent of LD across a gene is independent of gene size. For example, LD extends across the 283-kb NF1, whereas fewer correlated sites are present in the smaller CCND2. For these genes with weak LD, attempts to choose sites with either LD-based (Carlson et al. 2004b) or haplotype-based (Johnson et al. 2001) selection will require typing a larger fraction of the common sites in the candidate gene. Although the stratified nature of the PDR can produce artifactual LD, the patterns of LD described here represent the range of observed patterns within the EGP data set. Particularly for genes that exhibit strong LD (e.g., NF1), these patterns appear to be consistent among the ethnic subpopulations in the PDR.

Although associations between individual sites and phenotype have proven useful in uncovering associations in the human genome (Meirhaeghe and Amouyel 2004; Tempfer et al. 2004), it is clear that the interactions between multiple sites within a region or gene may also be important and can be detected via haplotype associations. The best example of this is the association between a haplotype in the apolipoprotein E gene (APOE4) and Alzheimer’s disease (Corder et al. 1993). There are many factors that influence haplotype structure, including the mutation, gene conversion, and recombination rates. Figure 6 shows the distribution in the haplotypes per gene for 330 genes from the EGP. Haplotypes were only inferred from common SNPs with > 5% MAF and ranged from 2 for FAU [Finkel-Biskis-Reilly murine sarcoma virus (FBR-MuSV) ubiquitously expressed (fox derived); ribosomal protein S30] to 175 for IGF1R (insulin growth factor 1 receptor). It is worth noting that the average number of haplotypes per gene is 38 for this data set. However, haplotype diversity will be greatly influenced by recombination, and recent reports suggest that genes with extreme haplotype diversity may contain one or more hotspots of recombination that would greatly increase the overall number of haplotypes for any given candidate (Crawford et al. 2004a, 2004b), and several approaches have emerged to identify hotspots of recombination and will aid in developing new approaches to haplotype tagging for association analysis (Crawford et al. 2004a; McVean et al. 2004).

As envisioned by Dr. Kenneth Olden in 1997, the EGP has generated significant new insights into the diversity and genetics of environmental response genes. The initial set of target genes, 550 altogether, will be completed over the next 6 months, and targets for the next phase are already under development. It is likely that DNA sequencing will play an increasingly important role in the EGP and in all future genetic analysis with its decreasing costs and rapidly expanding scale. Over the next decade, new approaches in in situ sequencing will be tested. If successful, it is likely that such genome-based resequencing will emerge as a dominant genotyping strategy as well (Collins et al. 2003; Shendure et al. 2004). Indeed, if the $1,000 per resequenced human genome becomes a reality, sequence analysis will play an increasingly important role in whole-genome association studies in human populations because the entire spectrum of variation (common and rare) can be uncovered in a single pass.

Until complete resequencing becomes a reality, only more limited subsets of the variation identified across the human genome will be applied in association studies. Several approaches are being taken to reduce the number of sites to be tested in human studies. The first approach is to explore the sites associated with coding and conserved noncoding sequences that can be functionally assessed via computational methods.

cSNPs represents only a small subset of the variation across the genome, but if common diseases arise from mechanisms similar to rare Mendelian diseases, this subset of SNPs should be tested. However, arguments for and against this hypothesis have been raised, and numerous reviews of the issues involved are available (Cardon and Bell 2001; Pritchard and Cox 2002; Reich and Lander 2001). By extrapolating current findings for the EGP across the human genome for an estimated 24,000 genes (Ewing and Green 2000; International Human Genome Sequencing Consortium 2004; Lander et al. 2001), 150,000 cSNPs (with an MAF > 1%) may be identified. This figure is similar to prior predictions (Kruglyak and Nickerson 2001). On average, 50% of the cSNPs identified will lead to amino acid substitutions. Therefore, 75,000 amino acid-altering SNPs may be identified across the human genome. Using new computational approaches such as SIFT and PolyPhen to score potential functionally ns-cSNPs, the subset of cSNPs could drop to approximately 8,250 cSNPs with predicted functional importance. However, only a small fraction of these cSNPs, approximately 1,000, will have MAFs > 5% and could be easily tested with newly developed low-cost genotyping strategies on adequately sized human population cohorts. Although this estimate could potentially reflect the relatively high conservation bias of the EGP genes, the gene-to-gene variation observed for the EGP is consistent with previous observations of sets of genes involved in inflammation, lipid metabolism, and endocrine function (Cargill et al. 1999; Carlson et al. 2004a; Crawford et al. 2004b; Halushka et al. 1999; Stephens et al. 2001).

It is also possible to predict SNPs in noncoding sequences with functional significance using comparative genomic approaches (Dieterich et al. 2003; Frazer et al. 2004; Jegga et al. 2002; Sandelin et al. 2004; Schwartz et al. 2003). Furthermore, data from the ENCODE Project (ENCODE Consortium 2004) should generate new paradigms for these analyses. The development of new comparative sequence data from divergent species in the evolutionary tree (Wallis et al. 2004) will greatly enhance studies to identify SNP subsets in noncoding regions with predicted function (Bejerano et al. 2004). Even when more refined approaches develop for noncoding regions of biological interest, the sizes of the emerging SNP panels are likely to be comparable with cSNP sets.

Several of the predicted functional SNPs from the EGP are being developed by the Comparative Mouse Genomics Centers Consortium (CMGCC; NIEHS 2005) to generate transgenic and knockout mouse models based on human DNA sequence variants identified in environmentally responsive genes. These mouse models will become tools to improve our understanding of the biological significance of human DNA polymorphism, and many of the computationally mined SNPs from coding sequences are being translated into appropriate animal models for further studies (Ladiges et al. 2004; NIEHS 2005). Initially, the CMGCC is focusing on variation in genes involved in DNA repair or cell cycle control, because many of these are well-characterized, environmentally responsive genes that can be translated into many current studies (Angus et al. 2003; Bahassi el et al. 2002).

According to the common disease/common variant hypothesis, the genetic risk factors underlying common diseases are likely common, modest risk alleles in the human population (Altshuler et al. 2000a; Halushka et al. 1999; Reich and Lander 2001). These alleles may not be adequately predicted by current computational tools, which are better suited for identifying the highly penetrant alleles associated with rare Mendelian diseases. These denser SNP sets, like those now available for the EGP project, will engender a second, indirect approach to association mapping. This approach will develop SNP sets to exploit LD (or SNP associations) to capture the sites that cannot be predicted by computational tools to be functionally important.

It is worth noting that our limited ability to predict SNPs with functional consequences has led to the development of several other large-scale projects to discover and type common polymorphisms across the human genome (International HapMap Consortium 2003; Patil et al. 2001). These are far less comprehensive then the current EGP project but more global in their approach to the genome. An initial set of 600,000 SNPs is being developed, but even this set is under expansion because of the wide variation in LD or SNP association known to exist across the human genome. In fact, the complete data sets available through the EGP have been one driver of the need to increase the density of the current HapMap. However, the global views provided by these genomewide data sets are important, and there is an effort under way to type the common variation identified through the EGP on the HapMap samples so that the data from the EGP genes are integrated with resources emerging for the entire human genome.

From its inception, the EGP has provided the genotype data needed to drive the next-generation association mapping of genotype-phenotype-environmental interactions. This project has revealed the broad gene-to-gene variation in sequence diversity, LD, and haplotype diversity present in the human genome. Additionally, the EGP is one of the first projects to explore large noncoding genic sequences in the human genome and, as such, is one of the few projects that can fully inform association studies of candidate genes or association analysis of entire gene pathways. With the application of even a small pathway of genes, such as the base-excision repair pathway, it is clear that a new generation of software tools for association mapping will be required to fully explore even a simple candidate gene pathway for SNPs, haplotypes, and interactions between genes and with environmental modifiers. This is a key frontier for the EGP and is likely to be addressed during the next phase of the project.

To explore the relationship between environmental exposure and genetic susceptibility in the etiology of common diseases, the Environmental Genome Project is scanning environmental response genes involved in DNA repair, cell cycle control, apoptosis, drug metabolism, and other pathways for single nucleotide polymorphisms (SNPs). To date, more than 370 candidate environmental response genes have been examined and 50,000 SNPs identified across 8.6 Mb of baseline sequence, providing valuable resources for association mapping of genotype-phenotype-environment interactions. Additionally, these data are stimulating ongoing efforts to develop mouse models of potentially functional polymorphisms; explore the common variant/common disease hypothesis; address the ethical, legal, and social implications of the genetics of environmental disease susceptibility; and develop new tools, views, and strategies to improve the discovery of genetic variations responsible for sensitivity to environmental agents.

doi:10.1289/ehp.7922 available via http://dx.doi.org/