5.1. Linkage disequilibrium

LD is the phenomenon describing that two or more loci are inherited together more often than would be expected by chance (Ardlie et al. 2002a). This inheritance of chromosomal regions from shared ancestors without recombination, known as haplotypes, maintains particular combinations of alleles in the population (Reich et al. 2002). LD which is established in a population at some point in time is, however, gradually diminished by the actions of recombination and novel random mutations (Kruglyak 1999; Reich et al. 2002). A region characterized by no (or little) recombination is collected in a haplotype block, whereas regions with high recombination usually are described as divided into separate haplotype blocks. LD thus describes the non-random correlation between alleles at a pair of SNPs, usually defined by D' or r 2 values. An international joint effort to create a genome-wide map of LD and haplotype blocks has been established through the HapMap project (www.hapmap.org). The hope is that by knowing the haplotype block structure of the entire genome, one could capture the genetic variability of the genome by genotyping a much smaller number of SNPs that describe each haplotype block (Johnson et al. 2001). These haplotype tagging SNPs (htSNPs) are defined by the fact that information of their allelic variation provides relatively unambiguous information about variation at another locus (Johnson et al. 2001). Haplotypes thereby allow mapping of disease-susceptibility alleles without the need to identify and test every SNP across each chromosomal region (Kruglyak 1999; Risch 2000; Johnson et al. 2001) and many studies of genetic variants and association with a disease are based on the concept of LD. It may be assumed that a variant allele itself is partly responsible for the phenotype or it may be assumed that it is a marker allele in tight LD with the disease allele. An association study in the latter case is known as an indirect association approach and is the one used in whole-genome association mapping (Kruglyak 1999), Figure 2 . In other words, LD can be determined between individual SNPs, e.g. for constructing haplotype maps across a genomic region, as well as between an examined SNP and another SNP that is the one responsible for the phenotype causing mechanism, in which case the examined SNP is regarded as a phenotype marker (Kruglyak 1999).

Association studies make use of both categories. In the first case the primary goal is to reduce the amount of genotyping needed in order to cover a particular region so that almost complete information about variation in the region is obtained (Johnson et al. 2001). One study suggested that LD might exist only within distances of 3 kb (Kruglyak 1999), whereas several others have detected LD/significant association within 50 kb, 1 Mb or even much much further, of a susceptibility locus (Collins et al. 1999; Moffatt et al. 2000; Abecasis et al. 2001; Reich et al. 2001) An allele may show strong LD with the functional variant in one population but exhibit weak LD in other populations (de Bakker et al. 2006), or may show strong LD with one marker but exhibit weak LD with other nearby markers, even in the same population (Carlson et al. 2004). Such phenomena may make it difficult to replicate genetic associations among populations and to obtain consistent results within a genomic region in the same population. A way to improve the power of association studies and increase the probability of replications, may be to construct haplotype blocks by studying LD patterns across the genome and then optimally select a set of robust htSNPs such that all common variants are either directly genotyped or in strong LD with the genotyped tagSNPs (Goldstein 2001; Johnson et al. 2001; Stephens et al. 2001; Gabriel et al. 2002; Zhang et al. 2002; Ke and Cardon 2003; Xiong et al. 2003; Carlson et al. 2004). However, it is unclear whether haplotype block patterns and tagSNPs are consistent between populations and among repeated sampling from within a population. Recently, several large genomic regions (of ~500 kb) have been comprehensively examined as part of the ENCODE project (http://www.genome.gov/ENCODE/) (ENCODE Project Consortium 2004), which involved re-sequencing of 96 chromosomes in the selected regions to ascertain all common variants, as well as genotyping of all SNPs that are either in dbSNP or were identified by the re-sequencing. This strongly confirmed the patterns of long segments of strong LD that were seen in earlier studies (Collins et al. 1999; Moffatt et al. 2000; Abecasis et al. 2001; Reich et al. 2001).

It has been shown that also copy number polymorphisms are an important class of genetic variation in the human genome (Iafrate et al. 2004; Sebat et al. 2004; Sharp et al. 2005; Tuzun et al. 2005). Special emphasis has recently been on deletion variants and methods to detect such polymorphisms (Conrad et al. 2006; Hinds et al. 2006; McCarroll et al. 2006). It was demonstrated that most common deletions (minor allele frequency > 10%) could be assayed by neighboring SNPs, due to high degree of LD with nearby SNPs (Conrad et al. 2006; Hinds et al. 2006; McCarroll et al. 2006). Deletions causing severe diseases are rare, whereas more benign deletions are widespread (Iafrate et al. 2004; Sebat et al. 2004; Sharp et al. 2005; Tuzun et al. 2005). Methods have recently been developed by which deletions can be evaluated in SNP-based whole genome association studies; they are not directly observed in ordinary gene mapping studies. The underlying principle is that segregating deletions will leave a footprint in SNP genotype data, seen by i.e. mendelian inconsistencies, Hardy-Weinberg disequilibrium or null genotypes (Bellanne-Chantelot et al. 2005; Conrad et al. 2006; Hinds et al. 2006; McCarroll et al. 2006). A recent study, which re-sequenced the MHC haplotypes, e.g. different DR3 haplotypes, demonstrated that copy number variants were quite common and accounted for 13% of the variants identified in the region (Traherne et al. 2006). Copy number variants have not yet been evaluated for association to diseases, but offer a potentially interesting and additional tool. Further studies are needed to clarify their role and usefulness as markers of disease.

For mapping genes in complex diseases different approaches can be used. The principal distinction lies between candidate gene approaches, in which biological information plays the primary role in selecting genes for detailed analysis, and positional cloning strategies, in which the major effort is in localizing susceptibility variants through examination of observed or inferred patterns of co-segregation between the disease phenotype and marker genotypes within families or populations.

5.2. Candidate gene approaches

The traditional candidate gene approach has a focused strategy, it is a classical hypothesis-driven approach and usually focuses on one or a few genes, proteins or biological pathways of high information content and pathophysiological importance. Candidate gene studies have until recently been the only practical alternative to classical linkage analysis. Genes can either be functionally/biologically selected candidates or positional candidates, selected by e.g. their location in a region of linkage. In both cases studies are hypothesis-based, relying on some kind of evidence or prior knowledge that the chosen genes might affect risk of disease. Additionally, the use of new tools of genomics (e.g. transcriptional profiling and proteomics) from human or animal studies may also identify candidate genes. Candidate gene approaches are based on testing for association, either of known allelic variants in selected genes or of newly identified variants after re-sequencing parts of, or the entire selected gene(s). In its simplest form allele or genotype frequencies are compared between cases and controls, alternatively family-based controls can by used, with the advantage of avoiding a potential problem of population stratification in the control samples. Family-based designs have the advantages that they are robust against population admixture and stratification, and allow both linkage and association to be tested for (Laird and Lange 2006), but do require more genotyping. For family-based studies, the TDT test (Spielman et al. 1993) can be used, which implies that under the null hypothesis, Mendel's laws determine which marker alleles are transmitted to the affected offspring from heterozygous parents. The test compares the observed number of transmitted alleles with what was expected for Mendelian transmission. An excess of alleles of one type among affected offspring indicates that a disease-susceptibility locus for a trait of interest is linked and associated with this marker locus. Originally TDT was used to test for linkage between a disease and a genetic marker in the presence of (already detected) disease association (LD) (Spielman et al. 1993; Spielman and Ewens 1996; Laird and Lange 2006). Since both linkage and association between the trait and the marker have to be present for the TDT to reject the null-hypothesis (Spielman et al. 1993; Spielman and Ewens 1996; Laird and Lange 2006), the TDT is now typically used as a test for association in family materials (Hirschhorn and Daly 2005). Association is due to lack of recombination between the marker locus and the disease-susceptibility locus. Even in the absence of prior evidence for association, TDT has been shown to be valid and can be used to test a marker, or a set of markers, for association to disease, using a data set comprising parents and one or more affected sibling(s) (Spielman and Ewens 1996). TDT is a non-parametric test, and does not require specification of disease model, or assumptions about the distribution of the disease in the population, to be valid. An extension to the classical TDT test is the Sib-ship-TDT, which also uses information from unaffected siblings (Spielman and Ewens 1998).

Case-control studies involve groups of affected and unaffected subjects all being unrelated and sampled from the same population. This approach is used for identifying statistically significant differences in allele frequency and/or genotype distributions between a genetic variant and the disease. Recruiting probands and their relatives may be more difficult than recruiting cases and controls, except probably for childhood diseases, where suitable controls are difficult to obtain. A disadvantage of the family based approach is that for late-onset diseases it may be difficult to obtain affected subjects with living parents. Furthermore, the requirement for living parents may introduce an age-of-onset bias towards younger patients for diseases that usually arise late in life. Subjects may tend to be characterized by a less severe phenotype, if they have living parents, or may be affected by subtypes of disease, e.g. a more severe phenotype related to early onset of a disease. Both situations could contribute to population stratification bias. However, family designs are generally more robust against population substructures, and use of family-based samples is a way to avoid severe stratification biases, since only within-family comparisons are made and data are thereby not affected by population structure (Spielman and Ewens 1998).

Furthermore family-based approaches have other theoretical advantages. Samples can be used to determine whether an allele has different effects on disease, dependent on whether inherited maternally or paternally. Parent-of-origin effects describes two phenomenas; genomic imprinting, which results from epigenetic modification of the genome leading to unequal transcription of parental alleles, i.e. expression of alleles depend upon the sex of the parent, from which they were inherited (Reik and Walter 2001; Glaser et al. 2006), and secondly, parent-of-origin effects on mutation rate, the preferential occurrence of some new spontaneous mutations in either the father's or mother's germ line, e.g. base substitutions or chromosomal abnormalities (Glaser et al. 2006). Discordant sib designs can control for the effects of shared environment. However, pure sibship-based association studies may be underpowered relative to case-control studies, although differences in statistical power between population and family-based designs are generally small. A study compared use of trios to case-control studies and found that the number of simplex families (trios) required for TDT was virtually equal to the number of cases needed for similar power in case-control studies with an equal number of cases and controls (McGinnis et al. 2002). This implies that less genotyping (one case and one control versus at least parents and one affected child) may be necessary in case-control studies to gain the same power and for many diseases they may even be easier to collect than families (McGinnis et al. 2002). Another design uses a strategy of collecting a much larger number of controls than cases (McGinnis et al. 2002), which might be advantageous if controls are easier to ascertain than cases for some diseases, in order to obtain sufficient power. However, cases and controls should always be genotyped the same way, in the same laboratory and derive from the same ethnically homogeneous source to minimize risk of false positive associations, due to population stratification. Laird et al. showed by simulations, that for rarer diseases (prevalence e.g. 0.1%) use of 200 trios was more powerful than 200 cases and 200 controls, whereas for more common diseases (prevalence e.g. 14%) the opposite was the case, although differences were small (Laird and Lange 2006).

Candidate gene studies often suffer from a significant lack of information concerning the underlying complex regulatory pathways, and therefore are limited by their need to focus on only few already recognized and selected proteins or genes. Studies are dependent on having predicted the correct gene or genes, usually on the basis of biological hypotheses or the location of the candidate within a previously determined region of linkage. Even if hypotheses are broad, e.g. involving testing all genes in a certain (known) functional pathway or cascade, the approach is limited by lack of complete information regarding pathways and cascades, and will probably not identify all genetic risk factors for the disease in question. The most important limiting factor is lack of complete knowledge of the pathogenesis for most complex diseases, as well as too limited sample sizes.

Association studies using known and common variants are cheaper and simpler than complete re-sequencing of selected candidate genes. To capture all variation in and around a gene, re-sequencing of the entire gene: exons, introns and regulatory regions (i.e. 5' and 3'UTR and further up- and downstream of the gene) is necessary in ideally a large number of cases and controls. Re-sequencing is, however, expensive and laborious, and in many studies so far, only coding regions of one or a few genes were sequenced. Questioning this approach, is the fact that there are several examples of disease-variants located in introns and regulatory regions (Bergholdt et al. 1995; Bergholdt et al. 2000; Horikawa et al. 2000; Prokunina et al. 2002; Helms et al. 2003; Tokuhiro et al. 2003; Ueda et al. 2003; Bergholdt et al. 2004a). SNPs in coding regions are often implicitly accepted as "the answer" when an association with a missense variant is detected. This, often without the same functional evidence, as required for a SNP in a non-coding region, and often despite the presence of many nearby variants that might be equally or more strongly associated with disease. As an example, the well known codon 17 polymorphism ( Thr17Ala ) in the CTLA4 gene, which has been shown to be associated with several autoimmune diseases, turned out to show association, only because it was in strong LD with a regulatory polymorphism in the non-coding region ( CT60, G6230A, rs3087243 in 3'UTR), which is more strongly associated with disease and therefore more likely to be causal (Ueda et al. 2003). Proper interpretation of results from association studies might be challenging, particularly when considering rare non-coding variants. A recent study observed that re-sequencing high-priority candidate genes, in which severe loss of function variants were known to cause Mendelian disorders of lipid metabolism, demonstrated that some of these genes additionally harbored less severe, relatively rare missense variants, which were associated with high, but not extreme, levels of high-density lipoprotein (Cohen et al. 2004). This highlights the complexity and importance of correctly classifying phenotypes and sub-phenotypes.

Candidate gene association studies have been used for many years and have identified many of the genes known to contribute to susceptibility to common disease. They have been powerful in identifying common variants underlying complex traits like T1D, e.g. HLA (Nerup et al. 1974; Cucca et al. 2001), INS (Bell et al. 1984; Barratt et al. 2004), CTLA4 (Nistico et al. 1996; Ueda et al. 2003), PTPN22 (Bottini 2004; Smyth et al. 2004) and IL2RA/CD25 (Vella et al. 2005).

Despite that many positive results have come from association studies, often initial positive reports of association have not been replicated. In a study of 166 putative associations (studied at least three times) only six appeared to be consistently replicated (Hirschhorn et al. 2002). Possible reasons for this lack of replication include differences in ethnicity (i.e. susceptibility allele frequencies, penetrance levels etc.), lack of statistical power, multiple-hypothesis testing, variability in study designs or in phenotype definition, population substructure, different statistical models and other biological or genetic heterogeneities between studies or populations (Hirschhorn et al. 2002; Laird and Lange 2006). Moreover the contribution of publication bias may be considerable (Lohmueller 2003), since initial positive associations are more likely to be published than negative findings. Furthermore, reasons for inconsistencies could also be false positive findings due to type 1 errors or population stratification. Prior odds for any biological candidate to contain variants capable of influencing disease susceptibility is low, given the modest effect sizes anticipated and combined with often small sample sizes assayed. Together with a probable use of too liberal thresholds for declaring significance, it is perhaps not surprising that many false positive associations have been reported. False negative findings could be due to lack of power in replication studies, and of course true differences between study populations or between different environmental or genetic modifiers can exist (Hirschhorn et al. 2002). With genome-wide association studies, involving hundreds of thousands of markers, such issues become even more important (Laird and Lange 2006).

Recently, guidelines have been proposed for more robust association studies in multi-factorial diseases (Editorial 1999; Cardon and Bell 2001; McCarthy et al. 2005). These include replication of significant findings in additional study groups of similar ethnic origin and exploration of data sets from dissimilar ethnicities, incorporation of diverse study designs, functional assessment of the presumed aetiological variants in order to provide biological substantiation of statistical findings, as well as sufficiently large sample sizes. Furthermore, genotyping a haplotype tagged set of variants across a gene may provide a systematically capture of variation in the gene. A strategy of using meta-analyses may also be used to overcome some of the problems.

With the improvements in genotyping technology, focus in gene mapping has shifted towards association studies, using large numbers of SNPs or other markers, genotyped in known linkage regions, candidate gene regions or on a genome-wide basis.

In our own studies we used the classical candidate gene approach in evaluation of the human interleukin-1 type I receptor gene, IL1RI (Bergholdt et al. 1995; Bergholdt et al. 2000), the Casitas-B-lineage lymphoma b gene, CBLB (Bergholdt et al. 2005c) and the interleukin-12-subunit B gene, IL12B gene (Bergholdt et al. 2004a). Methods used were considered appropriate at the time these studies were performed. In studies of the IL1RI and CBLB gene we performed mutational scanning by Southern blotting and/or sequencing by different methods (Bergholdt et al. 1995; Bergholdt et al. 2000; Bergholdt et al. 2005c). After identification of variants, these were genotyped in relevant cohorts and data were evaluated for association to T1D.

IL1RI was selected due to its obvious nature as a T1D candidate gene, being the membrane-bound receptor for IL-1, and thereby essential in IL-1 signaling, known to be crucial in beta cell death (Nerup et al. 1994; Mandrup-Poulsen 1996). In our studies of the IL1RI gene (Bergholdt et al. 1995; Bergholdt et al. 2000), we de novo sequenced the complex promoter region, comprising different alternative promoters, and searched for mutations herein. In the first report one variant was detected, initially by Southern Blotting, after which, the SNP and the region surrounding it was characterized by sequencing ~400 bp (Bergholdt et al. 1995). These methods were state-of-the-art methods for detection of mutations at that time and automated sequencing was a newly developed tool. We demonstrated T1D association of the identified SNP in a case-control material, genotyped by PCR-RFLP, comprising 262 cases and 189 controls (Bergholdt et al. 1995), which is a small material according to standards of today, but considered appropriate at that time. T1D association of this SNP in the IL1RI gene was subsequently confirmed in the UK population and in non-high risk HLA individuals in a Finnish cohort (Metcalfe et al. 1996), but could not be confirmed in a larger Danish T1D family material of 245 families (Kristiansen et al. 2000b). In the second report (Bergholdt et al. 2000), we de novo sequenced the remaining and larger part of the IL1RI promoter region. Four novel polymorphisms were identified. Three were genotyped (the fourth were in complete LD with one of the others and were not analyzed further), in a collection of 253 Danish T1D families (1097 individuals). Of these, 103 families were simplex families (one affected offspring), 150 were multiplex families, i.e. families with two T1D affected offspring. Genotyping was performed by PCR-RFLP assays designed for each variant and analyses of genotyping data were performed by Sib-TDT analysis (Spielman et al. 1993; Spielman and Ewens 1996; Spielman and Ewens 1998); methods still widely used for evaluation of candidate genes. One SNP in the promoter region demonstrated T1D association (Bergholdt et al. 2000). We evaluated IL-1 receptor type 1 plasma levels and stratified according to genotypes of this SNP, and demonstrated statistically significant difference among genotypes, suggesting an allele dosage effect, which, however, was not disease dependent (Bergholdt et al. 2000). The major concern regarding these studies are the limited sample sizes used, e.g. for mutation scanning by sequencing only five to eight individuals, implying that we could have missed even common variants. Variants in the IL1RI gene have also been demonstrated associated to other diseases, e.g. the HinfI variant we detected, was associated with increased risk of helicobacter pylori infection, but not gastric cancer (Hartland et al. 2004), to our knowledge it has not been tested in other autoimmune diseases (Hollegaard and Bidwell 2006). Another SNP in the gene was associated with the metabolic syndrome (McCarthy et al. 2003).

Studies of the Komeda diabetes-prone rat, a spontaneous animal model of T1D, characterized by autoimmune destruction of pancreatic beta cells, as well as signs of autoimmunity, identified the Cblb gene (Casitas-B-lineage lymphoma b), as a major susceptibility locus (Yokoi et al. 2002). Cbl-b is an ubiquitin-protein ligase and functional studies have indicated an important inhibitory role in T-cell co-stimulation. Furthermore, in human T1D, an increased frequency of other autoimmune diseases is often seen (Payami et al. 1989; Chikuba et al. 1992; Tait et al. 2004), resembling the phenotype of the KDP rat, suggesting the human CBLB gene to be a candidate for conferring susceptibility to human T1D and autoimmunity in general. We therefore screened the coding regions of the human CBLB gene for mutations, using automated sequencing, in a panel of 24 individuals affected with several autoimmune diseases, as well as five controls (Bergholdt et al. 2005c). One variant in exon 12 was in a large family material of 480 families significantly associated to T1D (Bergholdt et al. 2005c). Human CBLB gene variants have also been tested for association to T1D by two other research groups (Kosoy et al. 2004; Payne et al. 2004). The exon 12 SNP was in both these studies identified by sequencing, but not genotyped further due to a low minor allele frequency (MAF of 0.031 and 0.043), based on frequencies in sequencing panels of 16 and 32 subjects, respectively (Kosoy et al. 2004; Payne et al. 2004). The minor allele frequencies in these studies were, however, comparable with the one observed by us (0.024) (Bergholdt et al. 2005c), but their strategy may have missed important information regarding this SNP.

Furthermore, these studies were performed in heterogeneous populations, which may bias results due to population stratification. A recent minor case-control report could not demonstrate association of a SNP in exon 10 of CBLB to Graves disease in children (Chen et al. 2005). We performed pair-wise LD analyses, by use of Lewontin's standardized disequilibrium coefficient D'(Daly 2001; Gabriel et al. 2002) of the SNPs detected in CBLB . With a threshold of D' > 0.8 for reporting LD, which is generally accepted (Carlson et al. 2004), we detected one large LD block and two smaller LD blocks, covering most of the gene. This indicates that genotyping just few SNPs covering these blocks, should capture most of the variation in the gene, Figure 3 .

LD was, however, not absolute (D' < 1), which might explain lack of association of neighboring SNPs belonging to the same block. The LD-map constructed, using GOLD (Abecasis and Cookson 2000) for visualization, for the CBLB SNPs we detected, is shown in Figure 3.

Our studies of the CBLB gene were performed before the HapMap project (www.hapmap.org) was completed. Today, information from HapMap could have provided information about haplotype blocks and which SNPs to genotype directly. The HapMap project is, by no doubt, a valuable resource for candidate gene studies, reducing or ideally removing the need for re-sequencing a chosen candidate gene.

Functional studies supporting a candidate gene is of obvious importance, and may be viewed as a supplement or even an alternative to replication studies, especially when only small sample sizes are available. Also for the CBLB gene, we performed functional evaluation, and implication for functionality was demonstrated (Bergholdt et al. 2005c). An isoform of CBLB was demonstrated to be increasingly expressed in lymphocytes after cytokine exposure, suggesting a possible role of this isoform in T1D (Bergholdt et al. 2005c). We also evaluated potential gene-gene interaction between the CBLB and CTLA4 genes involved in the same functional pathway of T-cell activation. This was done by stratifying CBLB data according to high risk versus lower risk genotypes of the established CTLA4 marker, CT60 , which is associated to T1D (Ueda et al. 2003). Since CBLB association to T1D was strengthened in the high risk group, observed by an increased distorted transmission in the subgroup of families with the high risk CTLA4 genotype, interaction among the two genes were suggested (Bergholdt et al. 2005c).

We selected the IL1RI and CBLB genes as candidate genes based on their functional implication in pathways of high relevance in T1D pathogenesis. The cytokine gene, Interleukin-12 subunit B, IL12B , was selected for further examination based on its implication as a candidate gene for IDDM18 , a region suggested by linkage analysis to be potentially important for T1D (Warrington and Bengtsson 1994; Huang et al. 2000; Morahan et al. 2001). IL-12 influences T-cell responses (Adorini 2001) and drives the differentiation of T-lymphocytes towards the Th1 subset, characterized by production of cytokines leading to cell mediated immunity (Manetti et al. 1993; Liblau et al. 1995; Tian et al. 1998). Furthermore in the non-obese diabetic (NOD) mouse, IL-12 was shown to play a primary role in T1D induction (Trembleau et al. 1995; Lamont and Adorini 1996; Adorini 2001). We sought to replicate a strong association of a 3'UTR SNP in the IL12B gene with T1D, as well as an observation of allele-dosage effects on the mRNA level of the expressed gene, reported by others (Morahan et al. 2001). We could not, in two large family collections (337 + 795 T1D families) replicate T1D association. We also typed adjacent microsatellites and performed haplotype analyses, but could still not confirm association. Neither functional significance nor a significant allele dosage effect on either mRNA or protein levels could be demonstrated in our studies (Bergholdt et al. 2004a). In conclusion, association of IL12B has been demonstrated in some T1D populations (Morahan et al. 2001; Yang et al. 2006), but not in several others (Johansson et al. 2001; Dahlman et al. 2002; Davoodi-Semiromi et al. 2002; McCormack et al. 2002; Nistico et al. 2002; Holm et al. 2003; Bergholdt et al. 2004a), suggesting genetic heterogeneity. One study only observed association in late onset T1D (Windsor et al. 2004). Functional studies on IL12B variants have been conflicting as well (Morahan et al. 2001; Dahlman et al. 2002; Seegers et al. 2002; Bergholdt et al. 2004a; Stanilova and Miteva 2005). Further, our data demonstrated inter-individual differences in stimulated IL-12 expression, however, not related to genotype of the 3'UTR SNP, underlining the complexity of the IL-12 system. Large inter-individual differences have been observed by others as well (Peng et al. 2006). The significance of the IL12B gene in T1D is therefore not yet clear. Association of the IL12B gene has been shown to other diseases, i.e. asthma (Morahan et al. 2002b; Randolph et al. 2004), chronic hepatitis C (Houldsworth et al. 2005) and cerebral malaria (Morahan et al. 2002a), however, neither in rheumatoid arthritis (Orozco et al. 2005a), nor with the presence or severity of coronary artery disease (Momiyama et al. 2005).

5.3. Classical genome-wide linkage analyses

Linkage analysis is the study of co-segregation of DNA variants with a phenotype or disease. No assumptions of biological mechanisms are made, and linkage studies are thereby not limited by specific hypotheses of disease pathogenesis. Since the function of the majority of genes characterized to date is unknown, and many unidentified genes exist, linkage analysis provides an objective tool for identifying yet unknown genes or chromosomal regions contributing to multi-factorial diseases. Genome-wide scans use affected sib-pairs or multiplex families, studied with a large set of polymorphic markers covering the entire human genome. Allele sharing is compared with disease-state and a statistical estimate of excess allele sharing can be calculated, leading preferentially to identification of the genomic position of loci, that influences the disease. Classical linkage analyses are usually perfomed using microsatellites, i.e. DNA sequences made up of repeats 2-8 nucleotides in length, randomly distributed throughout the genome. These are more polymorphic and thereby more informative, than a similar number of bi-allelic SNPs. Genome wide linkage analyses normally include at least 300-400 microsatellite markers, resulting in an average marker density of approximately 10 cM or less. The number of microsatellites may subsequently be extended in order to fine-map a specific area. The quite low resolution implies that signals from linkage scans can be broad. Data from families are required to detect linkage, as co-segregation of chromosomal loci can be observed only in relatives. Statistical support for linkage is presented in the form of a LOD score (logarithm (log10)) of the likelihood ratio for linkage). A LOD score of 3, taken as strong evidence of linkage, corresponds to odds of 1000:1 in favor of linkage. In studies of affected sib-pairs usually a non-parametric linkage score is used. Thresholds developed for linkage mapping of monogenic disorders with complete information of genotype and phenotype, may, however, not be optimally suited for mapping more complex polygenic disorders. Lander and Kruglyak (Lander and Kruglyak 1995) therefore proposed that the LOD threshold for "significant" genome-wide linkage in sib-pair studies should be raised to 3.6 (p < × 10 -5 ), while that for "suggestive" genome-wide linkage (random single occurrence in a genome-wide scan) should be set at 2.2 (p < 7 × 10 -4 ). In addition they suggested reporting all nominal p-values < 0.05 without any claim for linkage.

A typical positive first-stage genome scan usually provides linkage to one or more genetic intervals of approximately 10-30 cM in genetic width, which may include up to hundreds or thousands of genes. This is a huge distance at the molecular level, quite often with few biological clues as to the exact molecular basis of the linkage signal.

Linkage studies have been most successful in mapping genes underlying monogenic diseases, based on co-segregation of markers with the disease in affected large pedigrees. Markers are typically microsatellite markers, flanking the disease gene. Genome scans have also been used in the analysis of many complex diseases (where linkage peaks are generally wider), i.e. autoimmune and infectious diseases, phsychiatric diseases, hypertension, as well as animal models of complex human diseases. However, linkage studies have had only limited success in identifying genes for multi-factorial diseases, and the genes can in most cases only explain a small fraction of the overall heritability of the diseases. Several factors may be implicated in this lack of success, i.e. low heritability, inability of a standard set of microsatellite markers, spaced 10 cM apart, to extract complete information about inheritance and inadequately powered study designs (Hirschhorn and Daly 2005). Limitations of genome scans when applied to multi-factorial diseases may furthermore include heterogeneity in disease phenotypes, population and ethnic differences, imperfect statistical and analytical methods, and the use of different polymorphic marker panels and genetic maps in different studies. This has in several cases resulted in lack of replication and identification of different loci by different research groups studying the same disease. Possible solutions may imply using denser marker sets, larger sample sizes or larger pedigrees (Hirschhorn and Daly 2005).

Even when convincing evidence of linkage is obtained, extensive studies of putative candidate genes are still required. Systematic fine mapping of identified regions is needed to characterize variation(s) of the gene or genes responsible for the observed linkage signals, and to move from a broad region of linkage, usually exceeding 10 cM (~10 million basepairs), to the causal gene or genes within a region. This should be followed by functional characterization (in vitro and in vivo) of the biological effect of the identified genetic variants. Systematic fine mapping is laborious and complex. Linkage peaks may cover many genes (up to 1000) and may compromise several susceptibility loci, as suggested from animal models (Todd and Wicker 2001; Nguyen et al. 2002). Furthermore, a linkage signal may be observed due to chance clustering of several disease loci, each with relatively weak locus-specific effects (Concannon et al. 2005). Such complexity probably also, at least partly, explains some of the difficulties in obtaining reproducible association results. Fine mapping of identified linkage regions, in the same populations, is crucial in order to characterize the disease associated variants. Recently, the first identifications of disease-associated genes in complex diseases (Mb. Crohn, Asthma, Schizophrenia and myocardial infarction), using this strategy, were published (Hugot et al. 2001; Ogura et al. 2001; Stefansson et al. 2002; Van-Eerdewegh et al. 2002; Helgadottir et al. 2004). Another approach for increasing power in fine mapping linkage regions, may be the use of stratification/enrichment for cases from i.e. multiplex families, or from families carrying the strongest evidence for linkage.

Genome-scale linkage analyses in T1D have resulted in significant linkage for the MHC region and a number of non-MHC loci, however, with varying levels of statistical significance. In no case, so far, has a specific gene variant been proven to be the cause of genetic linkage to T1D.

In our own studies we fine mapped a region on chromosome 21, which showed linkage to T1D in the Scandinavian T1D genome scan (Nerup et al. 2001). The observed linkage was novel, and mapped to the distal part of chromosome 21 (21q21.3-qTel). The initial evidence for linkage was indicated by a maximum LOD score of 1.23 (p = 0.009) (Nerup et al. 2001), and linkage was most evident in the Danish population (maximum LOD score of 2.33 (p = 0.009)), Figure 4 . The linked region on chromosome 21 comprised approximately 20 Mb, which equals ~35 cM (www.ncbi.nlm.nih.gov/genemap).

Furthermore interaction analyses had recently given additional support to the chromosome 21 region, as being important in T1D, since in neural network and decision tree analyses of Danish and Swedish T1D genome scan data, the same region was identified (Pociot et al. 2004).

Almost the complete sequence of chromosome 21 was known when these studies were performed and 225 genes were identified on the chromosome (Hattori et al. 2000). That the chromosome was almost fully sequenced meant that new strategies for identification of genes could be applied instead of classical positional cloning, and that the position of genes, microsatellites and sequence variations were identifiable, though some uncertainty still existed about the exact map order of some of the markers.

The strategy used was to narrow the linked region, initially by linkage analysis including an increased number of microsatellites covering the region of interest in a multipoint NPL analysis. For the current study we initially used the Marshfield map (http://research. marshfieldclinic.org) for selection of microsatellite markers, but after advances in the completion of the human genome sequence, we used marker order and inter-marker distances based on the physical map (NCBI, National Center for Biotechnology Information; http://www.ncbi.nlm.nih.gov) of chromosome 21 for further analyses. A recent study has demonstrated the importance and impact of correct marker order for linkage analyses (Gretarsdottir et al. 2002).

A study of asthma susceptibility demonstrated that a one-LOD decrease in the support interval of the linkage peak, is likely to contain the susceptibility locus if the region has been saturated with markers (Van-Eerdewegh et al. 2002). In the present study the information content of the region, with the additional markers, was 95-100%, and we obtained a maximum NPL score of 3.61, Figure 5 .

We identified the corresponding support interval using NPL statistics (referred to as the "one LOD drop" interval) and found that it covered approximately 6.3 Mb, which was further studied by LD mapping with gene-based SNPs. We identified 32 candidate genes in the narrowed region, Figure 6 .

Several strategies could be chosen from here. The ideal one includes re-sequencing of all candidate genes in order to identify all mutations. However, due to the still quite significant number of 32 genes, we chose to examine the genes by identifying SNPs in coding regions (cSNPs) of the genes and initially screened the identified 74 cSNPs in a panel of 96 diabetic subjects. SNPs with a minor allele frequency above 3%, were further examined. These SNPs were genotyped and examined for T1D association in a larger T1D family material (253 families). None of the SNPs showed association to T1D, whereas one SNP in the TTC3 gene did show distorted transmission to unaffected offspring, which might support a possible protective effect.

The TTC3 gene is expressed in most human tissues, including pancreas. It has been speculated that overexpression of this gene may be involved in some of the morphological anomalies observed in Down's Syndrome (Tsukahara et al. 1996). We also performed stratification of SNP data, by dividing genotype data into subgroups, according to simplex or sib-pair family status, and according to families carrying the linkage signal in the region and families not consistent with linkage, respectively. This approach is based on a theory predicting that genetically loaded individuals of families presumably provide a stronger signal, thereby facilitating identification of disease susceptibility variants (Fingerlin et al. 2004). The strategy of selecting for further studies families that show evidence for linkage to a region, has been used previously (Horikawa et al. 2000; Van-Eerdewegh et al. 2002). Interestingly, we demonstrated significant T1D association of a TTC3 SNP in the sib-pair family subgroup, and in addition three other SNPs in OLIG2, CBR1 and KCNE1 showed association in the subgroup of families responsible for the observed T1D linkage of the region.

Additionally, the degree of linkage disequilibrium in the linked region of 6.3 Mb was evaluated by pair-wise LD measurements. We were, however, only able to identify two blocks, extending more than two markers with D'-values > 0.5. Thus, there does not appear to be large blocks with strong LD in this region in the Danish population, in agreement with previous studies on chromosome 21 (Olivier et al. 2001; Patil et al. 2001), although one block containing five markers was demonstrated, spanning approximately 100 kb, Figure 7 . These data support that one or more T1D linked genes exists in the region, even though no evidence for association was found for the tested SNPs in the complete family material. Despite the evidence for linkage in the Danish population, linkage to T1D on chromosome 21 has not been reported in other genome scans. However, it cannot be ruled out that it was missed in earlier studies due to inclusion of relatively few markers from this region on chromosome 21 (Davies et al. 1994; Hashimoto et al. 1994; Concannon et al. 1998; Mein et al. 1998; Cox et al. 2001). It can be estimated that the information content of this particular region in previous genome scans did not exceed ~0.66. In the most recent combined genome scan of Scandinavian and UK/US data, chromosome 21 was not identified either (Concannon et al. 2005). For this genome scan 22 markers for the UK/US data set and 8 markers for the Scandinavian data set were used on chromosome 21 (7 markers overlapped) (Concannon et al. 2005). If population heterogeneity, regarding this chromosome 21 region, do exist, the "Scandinavian" contribution may not be strong enough compared to the large number of UK/US families, to show up in the combined results.

Certainly, further genotyping and refinement of the genetic map corresponding to this region on chromosome 21 is important. Attempts to replicate our observation have so far not been published. LD mapping with SNPs coupled with genome scans are increasingly being used and are considered valuable for fine mapping and genetic association studies (Ardlie et al. 2002a; Tabor et al. 2002). Genotyping a large amount of SNPs is still time-consuming and costly, but important in terms of fully characterizing and fine mapping a linkage region. Our study reduced the region linked to T1D from more than 20 Mb to 6.3 Mb and increased the NPL score to 3.61. Furthermore, the number of possible susceptibility genes was reduced to 32. Although we have not yet been able to identify a single gene variant responsible for the observed linkage, we demonstrated significant support for several combinations of markers within this region. Furthermore, some evidence for association to T1D of variants of the TTC3, OLIG2, KCNE1 and CBR1 genes in conditioned analyses was observed (Bergholdt et al. 2005a). As our data suggest the existence of T1D linked gene(s) in the region, further characterization of the candidate genes on chromosome 21 is important, including analyses of regulatory and promoter regions, in which variants of potential importance for regulation of gene expression and thereby functional significance, might be located. Possible significance of non-coding SNPs in disease predisposition of e.g. T1D and T2D, have been demonstrated several times, (e.g. CALP10, CTLA4, RUNX1 ) (Horikawa et al. 2000; Prokunina et al. 2002; Helms et al. 2003; Tokuhiro et al. 2003; Ueda et al. 2003). Our subsequent analyses of chromosome 21 are discussed later in this thesis.

5.4. Comparison of linkage vs. association studies

The capability of association and linkage studies to detect allelic association is influenced by several factors, e.g. power of the study, sample size, effect size and the frequencies of the alleles (Zondervan 2004). The typical frequencies of variants, that underlie common diseases are largely unknown, but several examples of common variants (e.g. frequencies > 1%) have been shown to influence disease susceptibility, most of which, however, only increase the risk of disease by two-fold or less, when examined in large populations (Lohmueller 2003). For most common diseases the phenotype presumably is determined by a combination of susceptibility and/or protective genes and interactions between multiple genetic and environmental factors.

Hence, any individual genetic variant will probably have a small effect on disease risk. It is reasonable to suppose that common alleles, as well as rare alleles, will contribute to common disease. It has been demonstrated that linkage analysis often is more powerful than association analysis for identifying rare high-risk disease alleles, whereas association studies are expected to be more powerful for the detection of common disease alleles that confer small disease risk (Risch and Merikangas 1996).

Furthermore it has been pointed out that genetic association studies can provide not only greater power, but also better resolution of location than linkage studies (Risch and Merikangas 1996). Method of choice for obtaining maximum power for evaluation of different variants depends on allele frequency and severity/function of the variant in question, as indicated in Figure 8 .

Based upon the existing sample sizes, used in the recent combined T1D genome scan (Concannon et al. 2005), it is unlikely that non-HLA loci exist with odds ratios > 3 ( λ S > 1.3).

Recent power calculations of number of affected sib-pairs needed in genome-wide linkage analyses to detect linkage for minor risk loci, according to the criteria, set forward by Lander and Kruglyak (Lander and Kruglyak 1995) for suggestive and significant linkage are shown in Table 2 . Collecting more than 4,400 T1D affected sib-pairs are one of the goals of the Type 1 Diabetes Genetics Consortium (www.t1dgc.org). Association studies of genes with a plausible biological mechanism in T1D as well as genome-wide linkage studies both continue to be valuable methods. Increasingly, however, the focus is moving to genome-wide association studies, using SNPs, which may have the power to permit detection of the additional genes with effects in the appropriate range, i.e. odds ratios < 2, if robustly powered and performed. This approach is now technically possible and will probably be broadly applied in the future.

5.5. Genome-wide association studies

Genome-wide association studies might be advantageous for identifying variants associated with common diseases (Hirschhorn and Daly 2005; Wang et al. 2005). The definition of a genome-wide association study is an association study that surveys most of the genome for causal genetic variants. Since no assumptions are made about the genomic location of the causal variants, this approach could use the strengths of association studies without having to guess the identity of the causal genes (Hirschhorn and Daly 2005). Genome-wide association studies may include up to hundreds of thousands of SNPs (Hirschhorn and Daly 2005; Wang et al. 2005), much larger numbers than in classical linkage studies with less than thousand microsatellites. Using this amount of markers and corresponding statistical tests may, however, result in numerous hypothesis-testing issues (Hirschhorn and Daly 2005; Wang et al. 2005). Genome-wide association studies are unbiased, quite comprehensive and can be used in the absence of convincing evidence regarding the function or location of the causal genes leading to a specific disease, and can obtain greater power than classical linkage analysis. Selecting an appropriate sample size is important and because variants that contribute to complex traits are likely to have modest effects, large sample sizes are necessary. Required sample sizes are further increased by the large number of hypotheses that are tested in a genome-wide association study, p-values obtained should be corrected for multiple testing. It has been proposed that p = 5 × 10 -8 (equivalent to a p = 0.05 after a Bonferroni correction for 1 million independent tests) would be a conservative threshold for declaring a significant association in a genome-wide association study (Hirschhorn and Daly 2005). Alternatively, more liberal p-value thresholds could be used, but too liberal thresholds will require extensive follow-up and replication studies to distinguish false-positives from true associations. It has, however, been argued that a classical Bonferroni correction is too conservative, requiring too low p-values and too large sample sizes. With the high density of markers, significant LD between many markers and redundancy between single markers and multimarker haplotypes, which might also be tested for association, means that independency among tests is not the case. Alternative strategies for defining thresholds have been proposed, i.e. permutation testing, to assess the probability of having observed a particular result by chance (Hirschhorn and Daly 2005). An approach using two or three stages has also been proposed and may be advantageous in selecting a minimum amount of markers, while still obtaining reasonable power (Hirschhorn and Daly 2005). Thus, an initial scan is performed with a modest threshold for passing markers as positive, however, still with a threshold ensuring power for detecting loci explaining small fractions of the phenotypic variance (Hirschhorn and Daly 2005). Using this approach a limited number of false-positives will also pass, but in a second, independent population sample, similar in size or larger, the passed markers are tested again in stage 2. SNPs typed in stage 2 could even be genotyped by a different genotyping method, minimizing the risk of reporting false associations. A third step may then be used for even greater stringency (Hirschhorn and Daly 2005). To deal with the problem of multiple hypothesis testing such multi-stage designs may also be advantageous (Hirschhorn and Daly 2005), since the number of genotyped SNPs can be reduced in each stage, allowing genome-wide significance to be achieved step by step. (Laird and Lange 2006).

As it has been estimated that the genome contains approximately 12 million common SNPs (Kruglyak and Nickerson 2001), and that groups of neighbors correlated with each other exist, from which the genotype of one SNP perfectly predicts those of correlated neighboring SNPs (htSNPs) (Johnson et al. 2001), one SNP can serve as a proxy for many others, which are then indirectly assayed in an association screen. Once the patterns of LD are known for a given region, a few tag SNPs can be chosen such that they individually or in multimarker combinations (haplotypes) capture most of the common variation within the region (Johnson et al. 2001; Gabriel et al. 2002). The precise number of tagSNPs needed is currently not known, and depends on methods used to select the SNPs, the degree of long-range LD between blocks and the efficiency with which SNPs in regions of low LD can be tagged. Various algorithms for selecting tagSNPs have been proposed (Wang et al. 2005) It is obvious that a proportionally higher density of variants must be typed from regions showing low LD as compared to regions showing a high degree of LD to capture all information.

The amount of data produced will raise issues of analysis and interpretation. To be useful, markers tested for association must either be the causal allele or highly correlated (in LD) with the causal allele (Kruglyak 1999). Knowledge regarding LD patterns on a genome-wide scale is therefore essential. Attempts to determine LD patterns on a genome-wide scale are ongoing through the HapMap project [Consortium, 2003 #649] (www.hapmap.org), which provides a haplotype-based map of informative SNPs covering the entire genome, particularly useful for methods that use markers selected on the basis of LD. Phase I of the HapMap project has been completed (2005) and includes information on approximately 300 million genotypes (Altshuler et al. 2005). It has been demonstrated that large parts of the genome falls into segments of strong LD, within which variants are strongly correlated with each other (Daly 2001; Patil et al. 2001; Gabriel et al. 2002; Altshuler et al. 2005).

Another approach proposed for selecting SNPs for genome-wide association studies is the so called missense approach, in which only missense SNPs (SNPs changing the amino acid sequence encoded) are used (Botstein and Risch 2003). A high proportion of missense mutations are among the alleles underlying diseases of Mendelian inheritance, since they often cause severe changes in protein function. However, even for such diseases, an appreciable fraction of mutations are outside the coding regions, questioning the validity of the approach. Since alleles underlying complex traits often confer more subtle effects on disease risk, they are more likely to include non-coding regulatory variants with a modest impact on expression (Hirschhorn and Daly 2005). Missense variants have been associated with multi-factorial diseases, and obviously they are more likely to have functional consequences. However, ascertainment bias may be substantial, since until recently, missense variants have been preferentially discovered and tested for disease association. In summary, such an approach is likely to be productive, whereas it remains unclear how well common risk alleles will be captured. It has been estimated that a typical gene contains one or two missense SNPs, hence the strategy would require genotyping of only 30,000 - 60,000 SNPs (Hirschhorn and Daly 2005). On the other hand identification of all common missense SNPs would require a substantial effort in advance. It has been suggested that inclusion of SNPs from conserved non-coding regions might be advantageous. By comparative genetics, i.e. between human and mouse and rat genomes, it has been demonstrated that large parts of non-coding DNA is highly conserved among species, suggesting functional significance of such regions (Loots et al. 2000; Pennacchio and Rubin 2001; Thomas et al. 2003; Boffelli et al. 2004). This, however, will require a substantially increased number of SNPs to be included, diminishing the advantage of the missense approach as compared to selecting markers based on LD. Alternative approaches, such as typing only few SNPs in or near coding regions of each gene, are limited in coverage. Such studies only survey the variants chosen and variants in LD with them, and for sure will miss regulatory variants (Buetow et al. 2001; De La Vega et al. 2002). So, unless the LD patterns of each gene are empirically determined, such an approach might have little success and can not be called a genome-wide association study. Also attempts to convert linkage studies into association studies by looking for association between disease and the microsatellites typed in a linkage study, will survey only a small fraction of the human genome, and cannot truly be considered genome-wide association studies.

To perform genome-wide association studies, the ability to genotype large amounts of SNPs in large samples, is crucial. Genotyping technology has considerably improved and become cheaper in recent years, and prices are expected to drop further. Low costs and high quality of genotyping are essential and required to make genome-wide association studies feasible. In recent years, chips (mapping arrays) have been designed (by Affymetrix, Illumina and Perlegen companies) containing panels of genome wide SNPs, offering the most cost-effective way of genotyping the large amount of SNPs necessary. SNPs were selected on different bases, e.g. chips comprising 10,000, 100,000 and 500,000 SNPs from Affymetrix (www. affymetrix.com), on which SNPs were not selected based on LD. Another chip comprising 317,000 SNPs, termed Hap300K from Illumina (www.illumina.com) were selected mainly based on HapMap phase 1 data, however, with increased amounts of SNPs close to genes and in the MHC region, as well as enriched in non-synonymous SNPs. Also Perlegen (www.perlegen.com) perform genome-wide SNP genotypings in collaboration with researchers and companies. Chips containing 1 million SNPs are underway. Using these platforms and chips, the SNPs can be genotyped at significant lower cost per SNP. The high degree of coverage for these SNP sets implies that they are likely to cover a significant fraction of the genome, and even if they are less efficient per marker than a purely LD based set, the amount of SNPs should be high enough to ensure acceptable coverage of the genome. It is important to note that none of these marker sets will be optimal for detecting the effects of rare variants with frequencies of 1% or less.

The impact of the problem of population stratification is unknown and controversial (Ardlie et al. 2002b; Freedman et al. 2004). If a population is of mixed ethnicity, population stratification may imply different subgroups within a population with different disease prevalences. Techniques have been developed to detect and correct for the phenomenon by typing dozens of unlinked markers (Devlin and Roeder 1999; Pritchard and Rosenberg 1999; Pritchard et al. 2000; Reich and Goldstein 2001). Recently it was demonstrated that in well-matched populations, population stratification was not a major problem, whereas mild stratification was difficult to rule out (Ardlie et al. 2002b; Freedman et al. 2004). Even mild stratification may, however, confer bias when searching for alleles with modest effects. The problem of false-positive associations is essential in studies using hundreds of thousands of markers, and should be considered when designing a study. Strategies to increase the efficiency of genome-wide association studies may imply use of more homogeneous populations, e.g. like the Scandinavian or use of enriched cases, e.g. by family history or low age at onset, which is expected to increase power.

Until very recently no truly genome-wide association study had been published, although a few previous studies have used part of the approach. A strong potential association between acute myocardial infarction and the lymphotoxin-alpha gene ( LTA ) was suggested in a study with relatively few individuals (initially less than 100) genotyped for approximately 65,000 SNPs (Ozaki et al. 2002). They also followed up their initial results in a replication panel, in which association of an intronic SNP in LTA was replicated. Although this was a large project involving millions of genotypes, the low power of the initial screening sample means that the rates of both false-negative and false-positives in this study were probably high. Additionally, it is not clear what fraction of the genome was surveyed by their set of SNPs, selected from a Japanese SNP database (Ozaki et al. 2002). The lymphotoxin-alpha gene was identified against the odds of what would have been expected, taking the small initial sample size into account. Only a few other reports of genome-wide screens using SNPs had been published until very recently: Macula degeneration (Klein et al. 2005), Parkinson's disease (Maraganore et al. 2005) and obesity (Herbert et al. 2006). Relatively small sample sizes were included and the Affymetrix 100K SNP chip or Perlegen Sciences SNP chip, which only covers a fraction of the human genome were used (Klein et al. 2005; Maraganore et al. 2005; DeWan et al. 2006; Herbert et al. 2006). In addition there are several reports from different diseases where only gene-based SNPs were used, only non-synonymous SNPs, or just very few SNPs in smaller cohorts, exemplified in (Namkung et al. 2005; Yamazaki et al. 2005; Abel et al. 2006; Rudd et al. 2006; Smyth et al. 2006a; Spector et al. 2006) - these are considered far from truly genome-wide, and are not discussed further, even though some have identified candidate genes.

One genome wide association study of inflammatory bowel disease was conducted and published in 2006 using a LD based set of approximately 300,000 markers (Illumina HumanHap 300 chip) in 500 cases and 500 controls (Duerr et al. 2006). The CARD15 gene was confirmed and a variant in IL23R was identified (Duerr et al. 2006).

Since then, in 2007, the field has moved extremely fast forward, significantly demonstrating the feasibility and success of genome wide association studies in complex diseases.

The first genome wide association study of T2D has been published (Sladek et al. 2007). Five loci with variants conferring T2D risk were reported in a stage 1 and preliminary stage 2 analysis (Sladek et al. 2007). The follow-up case-control cohort used for stage 2 was selected based on less stringent inclusion criteria than the original cohort used for the genome wide association study, the significance of which is unknown. The complete two-stage analysis is underway and all loci detected do need replication in other samples. Sladek et al. (Sladek et al. 2007) used two marker sets, one based on LD data from the HapMap project supplemented by a gene-centered SNP set, in total approximately 400,000 markers. However, despite this high resolution, the often replicated T2D association of the TCF7L2 gene was only detected by positive association of one single SNP, indicating that coverage of studies like this may still be insufficient to detect all important loci.

In T2D the genome side association study of Sladek et al. has rapidly been followed by several other T2D genome scans (Saxena et al. 2007; Scott et al. 2007; Steinthorsdottir et al. 2007; Wellcome Trust Case Control Consortium 2007) and a follow-up study (Zeggini et al. 2007). Together, these reports have provided convincing evidence for several new T2D regions, the HHEX-IDE region on chromosome 10, the CDKAL1 region on chromosome 6, the CDKN2A-2B region on chromosome 9, the IGF2BP2 -region on chromosome 3 and the SLC30A8 -region on chromosome 8 (Saxena et al. 2007; Scott et al. 2007; Sladek et al. 2007; Steinthorsdottir et al. 2007; Wellcome Trust Case Control Consortium 2007; Zeggini et al. 2007). All studies were performed with large sample sizes, using individuals of Northern European ancestry, and there were only slight differences in phenotypic disease characterizations of cases among the studies. The new regions identified were all found in three or more of the scans, however for some, different SNPs were used to identify the same locus (Saxena et al. 2007; Scott et al. 2007; Sladek et al. 2007; Steinthorsdottir et al. 2007; Wellcome Trust Case Control Consortium 2007; Zeggini et al. 2007). In addition in a recent genome wide association scan for obesity the FTO gene were convincingly identified, also of presumed importance in T2D (Frayling et al. 2007). These new regions for which neither the causal gene or genes, nor the causal variant is known, adds to the established T2D genes, i.e. KJNJ11, PPARG, TCF7 and WFS1 (identified through candidate-gene studies) and the TCF7L2 gene identified through fine-mapping, reviewed in (Frayling 2007). The association of TCF7L2 is the top finding in all the genome wide association scans for T2D, it has an odds ratio of 1.37, whereas all other ten T2D loci or genes only has odds ratios between 1.10 and 1.20 (Frayling 2007). However, the studies published does not rule out other yet unknown signals or even larger signals since the chips used does not cover all common, and very little of rare variation, in the genome.

Regarding T1D, in 2007 the Wellcome Trust Case Control Consortium published their large genome wide association study of 2,000 cases from each of seven common diseases, and 3,000 shared controls from two separate cohorts. The seven diseases were bipolar disorder, coronary heart disease, Chrohn's disease, hypertension, rheumatoid arthritis, T1D and T2D (Wellcome Trust Case Control Consortium 2007). For genotyping the 500 K GeneChip Affymetrix set was used.

For T1D, this report demonstrated seven regions that showed strong evidence of association at p < 5 × 10 -7 , i.e. 12q13, 12q24 and 16p13 demonstrating the highest signals in themselves, as well as 4q27 and 12p13 identified by multi-locus analyses and 18p11 and 10p15 identified by combined analysis of autoimmune cases (T1D, Rheumatoid arthritis and Crohn's disease) (Wellcome Trust Case Control Consortium 2007). Except the chromosome 10p15 region, containing the CD25 gene, which encodes an IL-2 receptor (Vella et al. 2005), these strongest associated regions have not previously been demonstrated associated to T1D. The two chromosome 12 regions, 12q13 and 12q24 both map to regions of high LD containing many positional candidate genes, including several genes involved in immune signaling. In the 16p13 region only two genes of unknown function are located. The 4q27 region contains the IL-2 and IL-21 genes, which together with the 10p15 association signal may highlight IL-2 signaling as highly interesting in T1D. Also observations in the NOD mouse support IL-2 as important (Yamanouchi et al. 2007). The associated region on 18p11 only contains a single candidate gene, the PTPN2 (protein tyrosine phosphatase, non-receptor type 2), member of the same family as PTPN22 . Regions corresponding to the already established T1D candidate genes MHC, CTLA4, PTPN22, IL2RA/CD25 and IFIH1 were also identified in this genome wide association scan with p-values < 0.001 (Wellcome Trust Case Control Consortium 2007). Significance levels corresponding to these loci were much lower than the strongest associated loci, and the INS gene region was not detected at all (Wellcome Trust Case Control Consortium 2007). No markers close to the INS gene were present on the chip used. Replication studies are required to confirm associations from genome wide association studies. John Todd et al. therefore followed up this primary genome wide association scan in another paper, aiming at validating the initial findings regarding T1D. Eleven regions from the initial study were sought replicated, as well as a few regions identified in their other previous scan, comprising only non-synonymous SNPs, in which they originally identified the IFIH1 gene (Smyth et al. 2006a). The MHC region was excluded from the follow up analysis (Todd et al. 2007). In this second report 4,000 new individuals with T1D, 5,000 controls and approximately 3,000 family trios were used (Todd et al. 2007). Robust association to T1D of only four of the newly identified chromosomal regions, 12q24, 12q13, 16p13 and 18p11, could be replicated (Todd et al. 2007). Loci that could only be identified in the primary report, but not reproduced in the second, are probably either of small effect or false-positives, underlining the risk of reporting false-positives even in such large material. Evidence for the four replicated regions is very strong, and should be followed by extensive fine mapping and functional studies in order to identify the causal genes in each region.

Another independent two-stage T1D genome wide association scan (comprising 563 cases, 1146 controls and 483 T1D family trios of European ancestry) has just been published (Hakonarson et al. 2007). In stage 1 550,000 SNPs were genotyped on the Illumina platform. This report confirmed known T1D loci as the HLA region, the INS gene region and the PTPN22 region and identified a new locus on chromosome 16p13, corresponding to a LD block only comprising the KIAA0350 gene (a sugar-binding C-type lectin) (Hakonarson et al. 2007). This locus was also identified in the Welcome Trust Case Control Consortium genome wide association scan (Todd et al. 2007; Wellcome Trust Case Control Consortium 2007).

Failure to detect known association signals does not provide conclusive exclusion of corresponding genes, since coverage in such scans of neither common nor rare variants are complete. Despite large sample sizes, power is apparently still limited to detect significance levels appropriate for genome wide studies. The recent genome wide studies do, however, represent a validation of the genome wide approach. An interesting feature as the use of shared controls for several disease phenotypes is furthermore elucidated (Wellcome Trust Case Control Consortium 2007). Additionally, it was demonstrated that the extent of population stratification in the British population was modest, when all individuals with non-European ancestry was excluded (Wellcome Trust Case Control Consortium 2007).

Several other true genome wide association studies are expected to appear in the near future. There will be a need for standardized criteria for establishing significance (perhaps based on permutation testing and replication in other populations) and for quality control. If this latter threshold is low, a high amount of incorrect genotypes, but a higher call rate will be the case, whereas if the threshold is more stringent the opposite is the case. It should be kept in mind, that common SNPs only capture common variation; rare variations and copy number variants are not captured. Many false positives will be generated, and replication of interesting findings will be essential. Extensive re-sequencing, genotyping and further functional studies are therefore mandatory in order to identify the causal gene/genes. But the field of genome wide association scans has, with these new reports, proven to be feasible, applicable and succesfull in complex diseases like T1D and T2D.

6.1. Other autoimmune diseases

Comparative analysis of the position of candidate loci for T1D with candidate loci from other autoimmune and/or inflammatory diseases shows considerable overlap (Becker et al. 1998; Becker 1999; Becker 2004). This supports a hypothesis that at least part of the underlying genetic susceptibility to T1D may be shared with other clinically distinct autoimmune diseases, as rheumatoid arthritis, systemic lupus erythematosus, multiple sclerosis, Crohn's disease etc. Such diseases may share susceptibility loci and thereby common biological pathogenetic pathways. Several of the diseases are like T1D characterized by antibody and/or T-cell reactivity, often preceding clinical onset (Roep 1996; Gottlieb and Eisenbarth 1998; Rabinovitch and Suarez-Pinzon 1998; Schranz and Lernmark 1998).

The MHC region has been demonstrated associated or linked to disease susceptibility in almost all autoimmune diseases. The protective effect of some HLA haplotypes in T1D, on the other hand, seem specific to T1D. The relationship of the MHC with immune/inflammatory diseases is thought to be due to the essential requirement of HLA molecules in normal antigen recognition, processing and presentation, ultimately leading to activation and progression of an immune system response.

This process, central to normal immune function, is disturbed in some individuals affected by disease, however, probably only in combination with other genetic and environmental modifiers causing autoimmune disease. In T1D as well as other diseases none of the genetic variants of the HLA region are disease specific but found in affected as well as un-affected individuals. A fraction of autoimmune T1D does not posses the otherwise characteristic HLA genotypes, which are thereby neither necessary nor sufficient for T1D. Additionally, the genetic location of many proposed T1D-loci overlap with loci from other autoimmune or inflammatory diseases. The genome scan approach has apart from in T1D (Davies et al. 1994; Hashimoto et al. 1994; Concannon et al. 1998; Mein et al. 1998; Cox et al. 2001; Nerup et al. 2001; Concannon et al. 2005), been used in other complex autoimmune/inflammatory diseases, e.g. multiple sclerosis (Ebers et al. 1996; Haines et al. 1996; Kuokkanen et al. 1996; Sawcer et al. 1996), Crohns (Hugot et al. 1996; Satsangi et al. 1996), rheumatoid arthritis (Remmers et al. 1996; Cornelis et al. 1998), psoriasis (Tomfohrde et al. 1994; Matthews et al. 1996), systemic lupus erythematosus (Kono et al. 1994; Morel et al. 1994; Gaffney et al. 1998; Moser et al. 1998) and asthma (Daniels et al. 1996). Familial association of different autoimmune diseases in the same pedigree (Bias et al. 1986; Grennan et al. 1986; Ginn et al. 1998), co-incidence of different autoimmune diseases in the same individual (Lorini et al. 1996; Yamato et al. 1997), and some shared clinical parameters of different autoimmune diseases (Martin et al. 1995), suggest a common biological basis for several autoimmune/inflammatory diseases. This does not, however, explain the tissue specificity for destruction and different clinical symptoms, observed for several of the autoimmune diseases. Additionally, a recent study of variants in 16 different candidate genes, previously associated with at least one autoimmune disease, did also not support the hypothesis, since a pronounced lack of association with other autoimmune diseases, of the variants tested, was observed (Smyth et al. 2006b). The genes tested were CRP, FCER1B, FCRL3, CFH, SLC9A3R1, PAD14, RUNX1, SPINK5, IL1RN, IL1RA CARD15, IBD5 -locus (including SLC22A4 ), LAG3, ADAM33 and NFKB1 (Smyth et al. 2006b). Variation within the CTLA4 gene has consistently and convincingly, however, been demonstrated associated to several autoimmune diseases, i.e. T1D, Graves disease and autoimmune hypothyroidism (Kristiansen et al. 2000a; Ueda et al. 2003; Pociot 2004). Also variation in PTPN22 has recently been demonstrated associated to several autoimmune diseases apart from T1D, e.g. rheumatoid arthritis, Grave's disease, systemic lupus erythematosus, systemic sclerosis and Wegener's granulomatosis (Bottini 2004; Onengut-Gumuscu et al. 2004; Smyth et al. 2004; Carlton et al. 2005; Jagiello et al. 2005; Lee et al. 2005; Orozco et al. 2005b; Skorka et al. 2005; Gourh et al. 2006; Onengut-Gumuscu et al. 2006; Wellcome Trust Case Control Consortium 2007).

Pathways affecting pro-inflammatory and anti-inflammatory cytokines, T-cells, antibody regulation, apoptosis and hormonal control of the immune system are the most likely to be involved in related autoimmune diseases (Rose 1997). The eventual genetic basis underlying co-localization of loci between multiple diseases could for each locus be due to either the same allele of the same gene, to different alleles of the same gene or different genes in a complex of functionally related genes or it may be coincidental. A general pattern of locus co-localization is not found in human non-autoimmune disease (Becker et al. 1998).

Even though the end-stage phenotype of a specific autoimmune disease may be clinically distinct and/or organ specific, the pathogenesis of many autoimmune diseases may still, however, involve shared processes of immune regulation. Overlapping loci between T1D and other autoimmune/inflammatory diseases suggests that in some cases, common pathways may be involved in the etiology of clinically, distinct autoimmune diseases. Environmental factors, other genetic loci, as well as epigenetic mechanisms may contribute to disease or target-tissue specification. There has not yet been a specific effort to study different autoimmune/inflammatory diseases using a standard panel of polymorphic markers to allow direct comparisons.

6.2. Type 2 diabetes

T1D and T2D [MIM #125853] has been considered genetically and pathophysiologically distinct entities, although clinically, they may sometimes be difficult to distinguish. Increasing evidence links T1D and T2D (Pietropaolo et al. 2000; Mathis et al. 2001; Wilkin 2001; Maedler et al. 2002), and there is accumulating evidence that different inflammatory mediators are of importance in the pathogenesis of both diseases. T2D is the result of insufficient insulin production to maintain normoglycemia in the face of insulin resistance and it may develop gradually over years. T2D comprises 85-90% of all diabetes, and an increasing incidence of T2D is seen in children and adolescents. Pancreatic beta cells are sensitive to a number of pro-apoptotic stimuli believed not only to be involved in the development of T1D but also T2D (Mathis et al. 2001; Donath et al. 2003). In T2D, beta cell deficit and increased beta cell apoptosis have also been reported (Butler et al. 2003) and may contribute to the relative insulin deficiency which characterizes T2D. Furthermore, cytokines induce insulin resistance in peripheral tissues (Starr et al. 1997; Kim et al. 2001a; Grimble 2002), and elevated circulating levels of pro-inflammatory cytokines in recent onset T2D subjects have been reported (Esposito et al. 2002; Kubaszek et al. 2003; Spranger et al. 2003). Hyperglycemia in T2D patients is accompanied by increased circulatory levels of pro-inflammatory cytokines (Esposito et al. 2002). Thus, cytokines produced locally as part of the autoimmune infiltrate of the pancreatic islets in T1D or expressed in beta cells and/or at high levels in circulation in T2D may be a common pathogenetic denominator in beta cell failure of the two diseases. However, in T1D cytokines produce beta cell destruction and hyperglycemia, i.e. they are causative, whereas in T2D hyperglycemia induces islet production of cytokines possibly leading to beta cell destruction, i.e. they are a consequence.

Approximately 10-15% of T2D patients have measurable circulating GAD antibodies indicating that they have a slowly progressing beta cell failure, so-called LADA (latent autoimmune diabetes of the adult) (Dahlquist et al. 1989; Li et al. 2001; Mathis et al. 2001; Naik and Palmer 2003). Clustering of T1D and T2D in families has been reported in several studies (Dahlquist et al. 1989; Carel et al. 1993; Li et al. 2001). Such observations could suggest overlap in the pathogenesis of T1D and T2D in some cases.

Use of comparative genetics between T1D and T2D is therefore obvious. Studies of "classical" T2D markers have been attempted in T1D materials. A consistent, but of modest effect, association of the PPARG (a nuclear receptor involved in glucose homeostasis, lipid storage and adipocyte differentiation) pro12ala polymorphism to T1D has been demonstrated in a large T1D family material of 2355 European families, RR 1.15 (1.04-1.28) (Eftychi et al. 2004), as well as in our own Danish family material (490 families), a report in which also a meta-analysis (containing 1691 transmissions) demonstrated association (Johansen et al. 2006). T1D association of the KCNJ11 (encoding Kir6.2, an inwardly rectifying potassium channel, forming the ATP-sensitive K + -channel in complex with the sulfonylurea receptor) glu23lys variant, and the TCF (encoding the transcription factor, HNF-1 α ) ala98val variant could not be demonstrated (Eftychi et al. 2004; Johansen et al. 2006). The IRS1 (insulin receptor substrate 1, crucial in insulin signaling) gly972arg variant demonstrated association in some populations (Federici et al. 2003; Morrison et al. 2004), but could not be confirmed in other (Eftychi et al. 2004; Johansen et al. 2006).

Another approach for gaining knowledge about T1D genetics through T2D genetics, may be by studying the genetic basis of pathways and key molecules that may contribute to both T1D and T2D risk, e.g. it is well documented that cytokine production and action are under genetic influence.

Since it was first proposed that inflammation and activated innate immunity might play a role in the pathogenesis of T2D (Pickup et al. 1997), several studies have addressed the role of markers of inflammation, e.g. cytokines, in the development of T2D (Pickup 2004). In several studies, IL-6 and TNF- α are positively correlated with measures of insulin resistance and/or plasma insulin concentrations, body mass index (BMI), and circulating triglycerides (Muller et al. 2002; Temelkova-Kurktschiev et al. 2002a; Temelkova-Kurktschiev et al. 2002b). Cytokines may play a role also in the pathogenesis of T2D by direct effects on the beta cell. Combinations of TNF- α , IL-1 β , IL-6 and IFN- γ can induce beta cell apoptosis through induction of signaling pathways that activate e.g. NF κ B (Donath et al. 2003). Variation in genes encoding all these molecules may provide insight into the pathogenesis of T2D and T1D. Specific cytokine gene variants and their role in T1D and T2D are reviewed in (Bergholdt and Pociot 2006), demonstrating still limited overlap in genetic variants evaluated in both diseases.

A recently identified and consistently replicated T2D candidate gene is TCF7L2 (Cauchi et al. 2006; Damcott et al. 2006; Florez et al. 2006; Grant et al. 2006; Groves et al. 2006; Scott et al. 2006; Zhang et al. 2006), a transcription factor involved in glucose homeostasis, capable of regulating pro-glucagon gene expression in entero-endocrine cells via Wnt signaling (Grant et al. 2006), and possibly early beta cell development (Cauchi et al. 2006). Hence, this gene may also be an interesting gene in T1D. The causal variant is not yet known. In T2D a population attributable risk of 21% was calculated (Grant et al. 2006), whereas no studies of association of TCF7L2 with T1D has yet been published. Very recently the first genome-wide association study of T2D was published (Sladek et al. 2007). The TCF7L2 gene on chromosome 10 was confirmed, and four novel loci identified, a variant in a zink transporter gene, SLC30A8 on chromosome 8, expressed exclusively in beta cells, as well as three variants corresponding to two LD blocks containing genes potentially involved in beta cell development and function ( IDE-KIF11-HHEX (chromosome 10) and EXT2-ALX4 (chromosome 11)). The PAR calculated for these loci in common is approximately 70%. The subjects used in stage 1 of this genome-wide association study were chosen based on a positive family history of T2D and lack of obesity (Sladek et al. 2007). This may enrich for individuals carrying a smaller number of risk alleles with stronger effects and for variants in genes implicated in pancreatic development and beta cell function, thereby also of potential importance in T1D, which needs to be explored.

6.3. Knowledge from syndromes

Another feature of comparative genetics is the use of knowledge from syndromes, in which T1D is one of the characteristics, and from rare forms of diabetes. Such studies may provide possibilities of examining disruption of individual components of glucose homeostasis. It should be taken into account that diabetes, resembling T1D, in some syndromes, is not associated with autoimmunity and may not have the same characteristics and underlying pathophysiological mechanisms as classical T1D. Some syndromes and rare forms of diabetes are characterized by a reduced or complete lack of beta cell mass, some manifest at birth, others are characterized by insulin resistance, reviewed in (Barrett 2001; Polak and Shield 2004). Diabetes may also result from mitochondrial disorders, of which the underlying mechanism of diabetes is unknown, but thought to be due to reduced insulin secretion (Barrett 2001). Also rare monogenic forms of diabetes can provide information (Porter and Barrett 2005). When a causative gene is identified, its contribution to genetic heterogeneity of more common types of diabetes, as well as functional effects of the relevant proteins, can be evaluated.

A few syndromes of relevance in T1D are discussed further in the following section.

6.3.1. Down's syndrome

In our own studies we have attempted to use comparative genetics by several means. Down's syndrome may serve as an example of how knowledge and studies regarding a quite common syndrome, only occasionally including T1D, may be used to extract information about T1D in general.

Down's syndrome (trisomy 21) [MIM 190685] is characterized by mental retardation and characteristic dysmorphic features, believed to be attributable to genes on chromosome 21, likely in a complex and interactive manner (Tsukahara et al. 1996; Gitton et al. 2002; Reymond et al. 2002; Kahlem et al. 2004; Patterson and Costa 2005). Down syndrome is characterized by an increased prevalence of a number of autoimmune diseases (Shield et al. 1999), and an increased prevalence of diabetes in Down's syndrome was a dogma, described in textbooks (Pickup and Williams 1991), based on a few old, biased and not population-based studies (Milunsky and Neurath 1968; Jeremiah et al. 1973; Van-Goor et al.; Anwar et al. 1998). Additionally, an earlier peak age of onset of diabetes had been suggested previously (Burch and Milunsky 1969). In none of the previous studies T1D was, however, distinguished from T2D. Pathogenesis and mechanisms behind diabetes in Down's syndrome are unclear, but signs of premature aging as well as life-style factors may bias the use of a simple age criteria for diagnosis, as was used in some previous studies. Increased autoimmunity in general in Down's syndrome has been proposed, and especially thyroid and coeliac disease have increased prevalences, including demonstration of auto-antibodies (Ivarsson et al. 1997; Karlsson et al. 1998; Agardh et al. 2002). In a study of nine Down's syndrome patients (age of patients unknown) with slightly impaired glucose tolerance, it was, however, not possible to detect insulin auto-antibodies, as an indication of autoimmune damage of pancreatic beta cells (Serrano-Rios et al. 1973). In a more recent study co-occurrence of more than two islet auto-antibodies was, however, demonstrated to be significantly more common in Down's syndrome individuals as compared to controls (Gillespie et al. 2006).

In Denmark, due to complete registries, unique opportunities exist for population based studies, and we therefore undertook an epidemiological study aiming at defining the prevalence of Down's syndrome in T1D. In this study, we demonstrated a prevalence of Down's syndrome in T1D of 3.8, which corresponds to a 4.2 times increased prevalence of Down's syndrome in T1D in Denmark, compared to the background population (Bergholdt et al. 2006). In line with some of the previous studies addressing diabetes in Down's syndrome, we also demonstrated a more than four times increased prevalence of T1D in Down's syndrome. Our study was the first proper demonstration of an increased prevalence of Down's syndrome in verified T1D, it was nation-wide population based, used validated and complete registers, as well as confirmed diagnoses of T1D. Median age at diagnosis of T1D in Down's syndrome was 6 years in our study, lower than the general median age at onset for childhood T1D of 8 years, which might suggest a greater "genetic load", perhaps attributable to genes on chromosome 21.

To determine whether T1D in Down's syndrome is immune-mediated, also immuno-genetic factors should be evaluated, and first of all, estimation of the significance of the HLA region, in Down's syndrome individuals with T1D, is necessary. This has recently been addressed in two reports, both, however, limited by the small number of individuals with Down's syndrome and T1D (Bergholdt et al. 2006; Gillespie et al. 2006). We attempted to evaluate the significance of the HLA region, in Down's syndrome individuals with T1D. HLA - DQB1 genotypes were evaluated, where this was possible. Although numbers were small (eight cases of T1D and Down's syndrome, of which HLA genotyping was possible in six), we demonstrated that one carried a dominantly protective allele, one only neutral alleles, three carried just one classical risk allele, and just one carried two classical T1D risk alleles (Bergholdt et al. 2006). In line with our findings, a newer report, comprising 40 individuals with Down's syndrome and T1D, also suggest that individuals with Down's syndrome generally do carry more high risk HLA class 2 genotypes than controls, but also that individuals with Down's syndrome and T1D carry less of the highest risk HLA class 2 genotypes when compared to T1D individuals in general (Gillespie et al. 2006). This might suggest that other genetic risk loci than the HLA region, as genes on chromosome 21, could be important in this group. It is tempting to speculate that the explanation for the increased prevalence, or at least part of the explanation, might be due to the trisomy 21, supporting the possible existence of T1D associated genes on this chromosome. The simplest mechanism would be that the extra chromosome 21 results in a linearly proportional increase of all gene products controlled by genes on chromosome 21. This might then produce an imbalance between gene products determined by chromosome 21 and those determined by genes on other chromosomes. Unfortunately, often no direct correlation between gene dosage and the amount or activity of gene products exists (Olson et al. 2004). Mechanisms by which increased dosage causes any specific Down's syndrome feature have not been established, despite extensive studies of Down's syndrome mouse models (Olson et al. 2004). However, a recent report indicated a ~1.5 fold over-expression of most of the genes in the trisomic region of the Ts1Cje Down's syndrome mouse model (Amano et al. 2004). A new promising aneuploid mouse model, Tc1, which carries an almost complete copy of human chromosome 21, have just been reported (O'Doherty et al. 2005). This Tc1 mouse has several characteristics of Down's syndrome, and is very promising for studies of Down's syndrome. A major concern, however, is that it is a mosaic, implying that not all cells are trisomic (O'Doherty et al. 2005). It is believed that some abnormalities of Down's syndrome are due to over-expression at specific loci, resulting from the presence of two identical copies of a susceptibility allele inherited from the parent of origin of trisomy (disomic homozygosity) (Feingold et al. 1995). So by inheriting two identical copies of a rare susceptibility allele, the suggestion is that an individual with Down's syndrome exceeds a liability threshold and manifests the disorder (Feingold et al. 1995). In Down's syndrome there is also an increased prevalence of a number of other autoimmune diseases, but to date no gene or chromosomal region have been able to explain this predilection for autoimmune disease (Shield et al. 1999). Our observation of a more than four times increased prevalence of Down's syndrome in T1D, and a lower median age at onset of T1D in Down's syndrome compared to other T1D individuals, substantiate the possible significance of genes on chromosome 21. Such genes may be important for the quite rare combination of T1D and Down's syndrome, but may also be involved in T1D pathogenesis in general. This represented the first population based study of Down syndrome in T1D. Further studies of auto-antibodies and the contribution of HLA and other genetic regions in T1D in Down's syndrome are clearly needed, including a larger number of well-defined T1D cases with Down's syndrome, to clarify diabetes in Down's syndrome. Also other molecular approaches, as global assessment of the transcriptome and proteome in such individuals, might be helpful new approaches. In addition to gene and protein expression patterns also temporal and spatial expression patterns of such may be important. When using a global approach also alterations in gene expression of genes on other chromosomes than chromosome 21, as well as of biological systems and pathways could be elucidated.

The distal part of chromosome 21 is known to harbor a few established candidate genes for autoimmunity, including T1D. One is the AIRE gene at 21q22.3, encoding a regulator of transcription, in which mutations have been shown to give rise to the disease APECED (Autoimmune Poly-Endocrinopathy-Candidiasis-Ectodermal Dystrophy syndrome) [MIM #240300] (Perheentupa 1996; Scott et al. 1998; Bjorses et al. 2000). APECED is a rare systemic autoimmune disorder of monogenic and autosomal-recessive inheritance (Meyer and Badenhoop 2002), where T1D is part of the syndrome in 20% of cases (Meyer and Badenhoop 2002). Almost 30 APECED causing mutations have been identified, of which two, prevails in Europe (Pearce et al. 1998; Meyer et al. 2001). The AIRE gene was originally considered a candidate gene of the linked region, and we tested these two prevailing mutations causing the APECED syndrome in Caucasoid populations, but no mutation carriers were identified. In addition, following our fine mapping, the AIRE gene was no longer localized in the linked region but distal to this, and we did therefore not continue with re-sequencing and further examination of the gene. Others have equally not been able to identify any of these two mutations in T1D populations (Nithiyananthan et al. 2000; Meyer et al. 2001; Meyer and Badenhoop 2002). Furthermore, disomic homozygocity at the APECED locus has been proven not to be able to explain autoimmunity in Down's Syndrome (Shield et al. 1999).

Another example of our use of comparative genetics in relation to chromosome 21 genes concerns the CBR1 gene. By proteome-analysis of interleukin-1 β (IL-1 β ) exposed diabetes-prone Bio Breeding (BB-DP) rat islets of Langerhans in vitro, the protein carbonyl reductase 1 (CBR1), encoded by the CBR1 gene, was demonstrated by our group to be down-regulated five-fold by IL-1 β exposure (Sparre et al. 2002). Since IL-1 β is believed to be an important cytokine involved in destruction of the beta cells (Nerup et al. 1994), this finding suggested a role for CBR1 in cytokine-mediated beta cell destruction and thereby in T1D pathogenesis. Analysis of seven cSNPs and two SNPs in the 3'UTR of the CBR1 gene did not demonstrate any significant association (Bergholdt et al. 2005a). However, in the "linked" family subgroup analysis, a significant association to T1D was demonstrated.

Another chromosome 21 gene, the Cu/Zn superoxide dismutase 1 ( SOD1 ) is a potent antioxidant, and has been implicated in genetic predisposition to alloxan (AL) -induced diabetes in mice (Mathews and Leiter 1999; Mathews et al. 2002). Alloxan (AL) and streptozotocin (STZ) are pancreatic beta cell selective toxins that have been used to probe the mechanisms underlying oxygen-mediated damage to rodent beta cells. Pretreatment with or concomittant administration of SOD1 confer protection against AL or STZ induced diabetes (Robbins et al. 1980; Oberley 1988), and cultured rodent islets can be protected from AL or STZ induced impairment (Fischer and Hamburger 1980; Oberley 1988). Specific activity of pancreatic SOD1 have been shown to be significantly increased in AL-resistant (ALR/Lt) compared with AL-susceptible (ALS/Lt) mice (Mathews and Leiter 1999), and SOD1 transgenic mice are protected against oxidative stress (Eizirik and Sandler 1989; Kubisch et al. 1997). We therefore not only genotyped cSNPs, but also known SNPs in the 3'UTR, of which none demonstrated association to T1D (Bergholdt et al. 2005a).

The Runt-related transcription factor 1 ( RUNX1 ), also called AML1, is another candidate gene for autoimmunity, including T1D, located in the linked region on chromosome 21. Indications that RUNX1 might be a common factor in autoimmunity have recently come from reports in three other autoimmune diseases, systemic lupus erythematosus, psoriasis and rheumatoid arthritis, in which disease-associated mutations in RUNX1 binding sites of three different genes was demonstrated (Prokunina et al. 2002; Helms et al. 2003; Tokuhiro et al. 2003). Further, the association of the regulatory PDCD1 variation (in a RUNX1 binding site), originally found with SLE (Prokunina et al. 2002), has been replicated in T1D (Nielsen et al. 2003). In rheumatoid arthritis an intron 6 SNP in the RUNX1 gene itself furthermore showed disease association in a case-control study (Tokuhiro et al. 2003). These reports make the RUNX1 gene interesting in T1D, supported by the location of the linkage peak we reported, in exactly this region. RUNX1 protein is a DNA-binding transcription factor with the context-specific capability of activating or repressing gene expression (Hug et al. 2004). We were, however, not able to demonstrate association to T1D neither of the known coding SNPs in this gene, nor of the intronic SNP in intron 6 (Bergholdt et al. 2005a), which was found associated to rheumatoid arthritis (Tokuhiro et al. 2003). However, T1D association of borderline significance was observed in the "linked" family subgroup analysis. In addition, data from our group have demonstrated that the mRNA expression of AML1 ( RUNX1 ) was significantly down-regulated in a rat insulin producing cell line after exposure to IL-1 using Affymetrix chip analysis (Nielsen et al. 2004), also suggesting that RUNX1 regulated gene expression may be relevant in T1D pathogenesis.

The observation of a more than four times increased prevalence of Down's syndrome in T1D, and a lower median age at onset of T1D in Down's syndrome compared to other T1D individuals, substantiate the possible significance of genes on chromosome 21. Such genes may not only be important for the quite rare combination of T1D and Down's syndrome, but may also be involved in T1D pathogenesis in general. Further studies of auto-antibodies and the contribution of HLA and other genetic regions in T1D in Down's syndrome are clearly needed. Larger numbers of well-defined T1D cases with Down's syndrome is necessary to clarify T1D in Down's syndrome, and these should preferentially be collected in a collaborative effort by several countries in order to obtain sufficient numbers. Also other new molecular approaches, as global assessment of the transcriptome and proteome in such individuals, might be helpful in elucidating underlying biological mechanisms. In addition to gene and protein expression patterns also temporal and spatial expression patterns of such may be important. The increased frequency of T1D among Down's syndrome individuals should be kept in mind when caring for people with Down's syndrome. As the life expectancy of a 1 year old person with Down's syndrome is now more than 60 years and likely to improve (Bittles and Glasson 2004), T2D is expected, just for this reason, to become an increased "problem". Additionally, the prevalence of verified T2D in Down's syndrome should therefore be evaluated, as should environmental and genetic factors underlying T2D in Down's syndrome individuals. Focus should be on life-style intervention, diet and the importance of physical activity, also as potential methods for prevention of T2D in this group. Clinical care of Down's syndrome individuals with diabetes is a specialist job, with obvious difficulties. Prolonged and maybe even continued intensive education in self-management is necessary. In only a subgroup of people with Down's syndrome diabetes self-management are obtainable, in other cases collaboration with parents or people caring for these individuals are essential. Education and level of self-management, furthermore has to be adjusted in an individualized way. Little is known about the demanding specialized treatment and care of diabetes in Down's syndrome, and further research into this field is warranted.

In conclusion, T1D has an increased frequency in Down's syndrome individuals, which should be kept in mind. The frequency of T2D is unknown, but presumed to be a problem as well. Research dealing with underlying biological mechanisms, genes on chromosome 21 as well as other genes, prevention, clinical care and management/potential self-management of diabetes in people with Down's syndrome is clearly needed.

6.3.2. Wolfram syndrome

Wolfram syndrome (MIM 222300) is a rare autosomal recessive disorder (Wolfram and Wagener 1938; Barrett et al. 1995). The disorder is also referred to as DIDMOAD (diabetes insipidus, diabetes mellitus, optic atrophy and deafness). Insulin-dependent diabetes and bilateral progressive optic atrophy are necessary for diagnosing the syndrome (Fuqua 2000). The clinical picture is, however, variable, and may additionally include neurological symptoms (Rando et al. 1992) and psychiatric illness (Strom et al. 1998). Linkage with Wolfram syndrome has been reported for a chromosome 4p16.1 region (Polymeropoulos et al. 1994; Collier et al. 1996), containing the WFS1 gene, in which mutations were identified in Wolfram syndrome families (Inoue et al. 1998; Strom et al. 1998; Hardy et al. 1999; Smith et al. 2004; Hansen et al. 2005). Diabetes seen in Wolfram syndrome is distinctive from T1D. It results from a selective islet beta cell loss, but with no evidence of autoimmunity (Karasik et al. 1989; Ishihara et al. 2004), suggesting that the WFS1 gene could be generally important in formation or survival of beta cells. It has been suggested that mutations in WFS1 may play a minor role in more common forms of diabetes (Inoue et al. 1998). WFS1 mutations have been tested for association with different diseases and specific genotype/phenotype correlations. No association of mutations in coding regions of the WFS1 gene could be demonstrated with T1D in our Danish family material (Larsen et al. 2004), whereas a Japanese study demonstrated association of one coding SNP in a small T1D case-control design (Awata et al. 2000). Also in T2D association of a two SNP-haplotype was demonstrated with WFS1 in a case-control study (Minton et al. 2002), but further studies in T1D and T2D are needed to clarify the eventual contribution of the WFS1 gene. Wolfram syndrome-associated diabetes is generally described as non-immune-mediated, however, one recent Japanese case report, demonstrated a Wolfram patient with a novel mutation in WFS1 , including diabetes-related (GAD (glutamic acid decarboxylase) and IA-2 (insulinoma-associated antigen-2)) auto-antibody positivity (Nakamura et al. 2006). The WFS1 region on chromosome 4p16 corresponds to a region previously linked to T1D (Nerup et al. 2001; Larsen et al. 2004), however not confirmed in the recent combined genome scan (Concannon et al. 2005). Identification of subgroups of diabetic patients with specific genetic causes may be important in shedding light on genetics and mechanisms, also of putative importance for the pathogenesis of T1D in general. The exact pathogenesis of Wolfram syndrome-associated diabetes, however, remains unknown (Fuqua 2000).

6.3.3. Wollcot-Rallison syndrome

Wollcot-Rallison syndrome (OMIM 226980) is a rare autosomal recessive condition characterized by neonatal or very early-onset diabetes, epiphyseal dysplasia, renal impairment, acute hepatic failure and developmental delay (Wolcott and Rallison 1972). Diabetes associated with this syndrome is characterized by beta cell loss, leading to insulin deficiency, without autoimmune pathology (Wolcott and Rallison 1972; Delepine et al. 2000). The condition was in 2000 mapped to the chromosome 2p12 locus (Delepine et al. 2000), containing the eukaryotic initiation factor 2 α kinase 3 gene, EIF2AK3 (also known as pancreatic endoplasmic reticulum kinase, PERK ). Mutations in EIF2AK3 have been demonstrated associated with the syndrome (Delepine et al. 2000; Iyer et al. 2004; Senee et al. 2004; Porter and Barrett 2005; Durocher et al. 2006). EIF2AK3 regulates protein translation and is highly expressed in pancreatic islets and plays a role in maintaining integrity of the pancreatic islets. The syndrome is extremely rare with only approximately 30 cases described world-wide (Delepine et al. 2000; Senee et al. 2004; Durocher et al. 2006). Findings have not indicated EIF2AK3 as a common diabetes susceptibility gene in diabetes in general, it was not associated to T2D (Vaxillaire et al. 2001), whereas in T1D suggestive evidence of linkage to the region was observed in Scandinavian families (Nerup et al. 2001). Furthermore a South Indian study demonstrated association to T1D of two microsatellites in the gene region, but did not demonstrate any common polymorphisms in the EIF2AK3 gene (Allotey et al. 2004). The role of this gene in T1D is therefore not clear either, and the potential use of genes underlying rare syndromes involving diabetes remains to be further elucidated.

6.4. Animal models

In animal models of autoimmune disease, the general disease phenotype can be broken down into individual loci having discrete sub-phenotypes, which is not possible in human studies. Such sub-phenotyping at specific loci can potentially provide clues to the functional significance of candidate genes at each locus, and may provide useful biological information in the search for candidate genes at specific human T1D loci. Transgene and knock-out studies in animal models allow studies of specific effects of genes either missing or over-expressed.

In T1D especially two animal models of human T1D have been successfully used, the non-obese diabetic mouse (NOD) and the bio-breeding rat (BB). Much insight into T1D pathogenesis and genetics, from syntenic regions, has come from studying these two animal models. A thorough review of the models and results obtained by use of them, is, however, considered outside the scope of this thesis. Some recent papers reviewing their use in T1D are (Todd and Wicker 2001; Lang and Bellgrau 2004; Leiter and von-Herrath 2004; Mordes et al. 2004; Mathews 2005; Melanitou 2005; Wicker et al. 2005; Yang and Santamaria 2006).

Another spontaneous rat model of T1D, the Komeda diabetes-prone (KDP) rat, is characterized by autoimmune destruction of beta cells, rapid onset of diabetes with no sex difference and no significant T-cell lymphopenia (Yokoi et al. 2002). Lymphocyte infiltration is also seen in other tissues as the thyroid gland, kidney, adrenal gland and pituitary, indicating general autoimmunity (Yokoi et al. 2002). Knowledge on genetics in the KDP rat was used in our own studies of the CBLB gene (Bergholdt et al. 2005c). In this rat model the major part of its genetic susceptibility for autoimmunity was shown to be accounted for by two loci, the MHC region on rat chromosome 20 and another locus on rat chromosome 11. This latter locus was demonstrated to contain the Casitas-B-lineage lymphoma b ( Cblb ) gene (Yokoi et al. 1997; Yokoi et al. 2002). A nonsense mutation in this gene in the KDP rat truncates the protein and transgenic rescue studies confirmed that that this mutation was pathogenetic and specific to the rat (Yokoi et al. 2002). Furthermore, functional studies of the protein, Cbl-b (an ubiquitin-protein ligase) demonstrated by use of knock-out mice, an important role in T cell co-stimulation, by acting as a negative regulator of autoimmunity (Bachmaier et al. 2000; Chiang et al. 2000). In Cblb-deficient mice, infiltration of lymphocytes in different endocrine tissue, was shown to be caused by enhanced T cell activation (Bachmaier et al. 2000; Yokoi et al. 2002). Based on this role of Cbl-b in regulation of autoimmunity, it was hypothesized that Cbl-b also could be dysregulated in humans and may contribute to autoimmune diseases in humans. In T1D an increased frequency of other autoimmune diseases is also often seen (Payami et al. 1989; Chikuba et al. 1992; Tait et al. 2004), and the CBLB gene was therefore chosen as a candidate gene for T1D and autoimmunity in general. We re-sequenced coding regions of the gene and demonstrated association to T1D of an exon 12 polymorphism, as described in chapter 5. The variant was rare (MAF = 0.024) and transmissions therefore few, underlining the importance of using large materials and the need for replication in other samples/populations.

In our study of the CBLB gene we also performed functional evaluation of the two known isoforms of the molecule. Expression levels of the two isoforms (one lacking exon 16) on the mRNA level, was evaluated constitutively and after cytokine stimulation. We demonstrated significant up-regulation, although not disease-specific, of the isoform lacking exon 16 upon cytokine exposure (Bergholdt et al. 2005c), which may suggest a possible role for this isoform in T1D pathogenesis (and/or other autoimmune diseases). No SNPs were detected in exon 16 or splice sites flanking this exon and no known functional elements of the molecule are encoded by exon 16. A mechanism for the suggested effect of cytokine stimulation is not known, and the involvement of several isoforms of CBLB also remains unclear. Due to the KDP phenotype, including several autoimmune diseases, we hypothesized that individuals with more than one autoimmune disease, may be more likely to harbor polymorphisms in CBLB . We therefore used individuals with more than one autoimmune disease for re-sequencing. Additionally, another comparative approach was used, since gene-gene interaction of CBLB with the CTLA4 gene was evaluated. The CTLA4 gene, involved in the same biological pathway of T cell regulation as CBLB , is an established candidate gene for many autoimmune diseases, e.g. Graves' disease, autoimmune hypothyroidism and T1D. Interaction between the two genes was evaluated by stratifying data from the CBLB SNP showing T1D association according to high risk genotype vs. lower risk genotypes of the established CTLA4 marker, CT60 (Ueda et al. 2003). An increase in transmission distortion was observed (Bergholdt et al. 2005c), and although based on few transmission, this may suggest interaction and supports genetic control of T cell regulation in autoimmunity.

7.1. Epistasis

Individual effects of minor susceptibility genes is, even in very large samples, difficult to demonstrate and may depend also on genetic variance at other loci (epistasis). Epistasis or gene-gene interaction is the concept of two or more loci or allele pairs, which combine interdependently to increase the susceptibility to a disease. Interactions between genes may contribute to the variation in complex traits, e.g. disease susceptibility (Cordell 2002; Carlborg and Haley 2004). Studies in model organisms suggest that epistasis may play an important role in the etiology of multi-factorial diseases and complex traits in humans. There is no consensus as to the best strategy for detecting epistatic interactions in humans (Cordell 2002; Carlborg and Haley 2004). A number of recent studies in humans and animals have identified loci that interact significantly but contribute little or no effect individually (Kuida and Beier 2000; Kim et al. 2001b; Carlborg et al. 2003). In T1D, attempts to show epistasis between classical T1D loci in humans, as well as animal models, have only provided very few examples (Cordell et al. 1995; Cordell et al. 2000; Cordell et al. 2001). Hence, new methods are in demand for detecting and characterizing epistasis, as well as for elucidating underlying biological interactions (Cordell 2002; Carlborg and Haley 2004).

Most gene products mediate their function within complex networks of interconnected proteins (Cusick et al. 2005). Therefore understanding effects of genetic interaction either by classical linkage approaches (Buhler et al. 1997; Farrall 1997; Cox et al. 1999; Cordell et al. 2000; van Heel et al. 2003) and/or novel analytical approaches is not simple. The presence of epistasis is important when investigating the genetic mechanisms underlying a complex disease because it implies that the effect of one locus on the development of a disease can be increased or decreased by effects at another locus, which is likely to result in a reduction in the power to detect the first locus (Cordell 2002). Interaction between different genetic loci, and probably also gene-environment interactions, may contribute to the complexity of the genetics of T1D. Interactions between susceptibility loci for T1D have been demonstrated in several studies (Davies et al. 1994; Cordell et al. 1995; Mein et al. 1998; Nerup et al. 2001; Concannon et al. 2005). In a large study of disease susceptibility and gene-gene interactions in a multiple sclerosis case-control data set, variation in 34 selected genes were evaluated (Brassat et al. 2006). A multifactor dimensionality reduction test (Ritchie et al. 2003b) was used to demonstrate one single variant, as well as a three locus combination showing association to MS, emphasizing the importance of looking at main effects as well as interactive effects in complex diseases (Brassat et al. 2006). Establishing the biological importance of interactions identified statistically, is, however, difficult, but important.

A conditional approach may be applied to genome scan data. A scan for epistasis could involve searching for modest individual effects, and then querying for interactions among such set of positive markers. Scans conditional on known positive results may be more powerful, as the variance explained by the major loci has been controlled for, thereby enhancing the signal from minor contributors. This indicates that initial screens for main effects also in genome-wide association studies are likely to be successful in many cases, without considering epistasis in the initial analysis, but only in subsequent analyses.

7.2. Data mining

Data mining technologies (based on pattern recognition) applied to genome wide linkage data might be a valuable tool, since these methods are able to detect complex interactions, not easily detectable by other methods. Data mining technologies can be used to identify marginal markers and combinations of marker-marker interactions predictive for disease susceptibility or protection. Importantly, looking at several interacting markers at a time may improve the ability to differentiate between patients and controls. Data mining and neural network-based methods have recently proven valuable in T1D (Pociot et al. 2004).

Genetic factors influence both susceptibility and resistance to disease, and data mining methods do allow evaluation of potentially protective effects (Pociot et al. 2004). This is not possible in classical linkage analyses, implying that for elucidation of such effects, novel analysis tools are necessary.

In our studies we have applied a data mining approach, using decision trees, for the analysis of genome scan data (Bergholdt et al. 2007b). Originally, the developed methods were based on decision tree construction and artificial neural networks (Pociot et al. 2004) applied to T1D genome scan data (genotyping data) from the Scandinavian genome scan (Nerup et al. 2001). We wanted to test whether the analytical methods previously developed, could use IBD sharing data, implying a reduced amount of information, as compared to real genotyping data used in (Pociot et al. 2004). We were given access to genome scan data from the Type 1 Diabetes Genetics Consortium (www.t1dgc.org) on 392 Scandinavian T1D multiplex families (411 affected sib-pairs) and 763 US/UK multiplex families (910 affected sib-pairs), which were part of the recent combined and largest T1D genome scan (Concannon et al. 2005). Data were recoded to IBD status and all microsatellite markers genotyped in both materials were analyzed.

We were able to detect and confirm the major loci also identified by the genome scan (Concannon et al. 2005): 6p21 (the HLA region), 2q31-33 (the CTLA4 region), 16p12-q11.1, 11p15.5 (the Insulin gene region), 16q22-q24 and 10p14-q11 (Bergholdt et al. 2007b), confirming the validity of the approach. Markers corresponding to the HLA region showed predictive signals for T1D with significance levels several orders of magnitude higher than the other markers. Interestingly, also two markers corresponding to the previously identified IDDM15 region (Delepine et al. 1997) on chromosome 6q21, were identified (Bergholdt et al. 2007b). In linkage studies identification of markers for this region generally requires special analyses, in order to separate their effect from the strong effect of the MHC region (Delepine et al. 1997; Concannon et al. 2005). A few new regions were also suggested, D17S798, D2S125, D9S175, and D8S261 (ref 8). In the complete data set, also the D4S403 was identified as of predictive value. This marker and the region surrounding it was previously demonstrated linked to T1D (Nerup et al. 2001; Larsen et al. 2004) and contains the WFS1 gene associated with Wolfram syndrome, chapter 6.3.2. In the Scandinavian data sub-set a marker corresponding to the linkage region on chromosome 21 (Bergholdt et al. 2005a), was furthermore identified providing significant predictive value for T1D (Bergholdt et al. 2007b).

Further we evaluated potential interactions between markers by use of the Enterprise Miner tree model in SAS. We used the highest ranked marginal markers to search for interaction with other markers (on other chromosomes) and the resulting combinations were tested for association to T1D. Combinations of markers, with significant predictive or protective value for T1D, were identified for the TNFA marker (a marker for the HLA region) and the D17S798 marker, interacting with markers on other chromosomes (ref. 8).

Interestingly, we demonstrated several rules of statistically significant interaction. One marker combination conferred protective value towards T1D (statistically more frequent in non-affected sib-pairs than affected sib-pairs), i.e. TNFA - D11S910 . All other combinations ( TNFA-D4S403, TNFA-D2S177, TNFA-D1S229, TNFA-D13S170, TNFA-D16S287, D17S798-D2P25, D17S798-D5S429 and D17S798-D1S197 ) conferred increased susceptibility towards T1D as compared to each of the markers alone.

To evaluate a potential biological significance of the predicted rules of genetic interaction, we constructed protein-protein interaction networks, using proteins encoded by genetic regions 5 Mb on each side of the interacting micro-satellite markers. We used a high confidence, quality controlled system integrating and validating interactions from multiple databases and species and were able to demonstrate support for functionality for all predicted rules of genetic interaction (Bergholdt et al. 2007b). These networks predict, with high stringency and confidence criteria, the most likely candidate genes responsible for the interactions predicted, as well as the biological context in which they are involved in (Bergholdt et al. 2007b). The protein-protein networks and their value in selecting functional candidate genes are further discussed in chapter 8.2.

In conclusion, novel analytical methods, based on data mining and decision trees can confirm observations known from "classical" analyses and provide new information about genetic interactions. Genetic interaction as the basis for pathogenesis is supported for T1D, and our findings certainly suggest these novel methods as valuable in searching for yet unidentified genetic and functional interactions involved in pathogenetic processes. Strong functional support to the value of the predicted interactions has come from our demonstration of protein interactions corresponding to the genetic interactions, as described in chapter 8.2.

8.1. Gene expression - Transcriptional profiling

Gene expression involves the presence of one or more transcription factor that, by binding to a promoter, leads to production/increased production of a protein or its suppression, and gene expression is thoroughly regulated. Knowledge regarding differences in gene expression, under different conditions or in different tissues, may be useful in mapping genes. Expression levels of genes may be seen as intermediate phenotypes between genomic DNA sequence variations and more complex disease phenotypes (Cheung et al. 2003); and variation in gene expression has been shown in large part to be due to polymorphisms in DNA sequence (Cheung et al. 2003; Morley et al. 2004). Genes responsible for variation in gene expression have been suggested to have an impact on phenotype (Morley et al. 2004). High heritability of gene expression has previously been demonstrated (Cheung et al. 2003) and suggests that identification of genetic determinants of expression may provide insights into the molecular basis of complex traits, like T1D. Common expression patterns of genes may furthermore reflect shared regulatory mechanisms.

Efforts have been made to identify and prioritize positional candidate genes based on their expression (Franke et al. 2004; Franke et al. 2006). It was expected that differences in expression e.g. among cases and controls, would be due to cis-acting variants in genes, but surprisingly and importantly, it was demonstrated that for most genes, differences in gene expression were determined by variants in genes located elsewhere (trans-acting) (Schadt et al. 2003; Morley et al. 2004).

Recently, a number of studies have used approaches in which linkage mapping was combined with the identification of co-regulated genes using microarrays. These studies have enabled the discovery of trans-acting expression QTLs (Morley et al. 2004; Deutsch et al. 2005; Li and Burmeister 2005; Stranger et al. 2005). This may be a promising approach also for identifying interacting susceptibility genes in multi-factorial diseases. In expression QTL analyses in general, large numbers of individuals are genotyped for a panel of polymorphic markers and simultaneously phenotyped with microarray expression profiling, followed by statistical evaluation of linkage between markers and gene expression levels (Beyer et al. 2007). Data for such genetical genomics studies in T1D are, however, still limited. Use of microarray technology is instrumental for these studies and the use of microarrays has changed molecular biology from studies of individual biological functions of a few related genes or proteins towards a more global investigation of cellular processes, shown to be complex and interrelated in biological systems (Allison et al. 2006; Hoheisel 2006). Microarrays are mainly used for transcriptional profiling and analysis of transcripts of genes, but can also be used for genotyping, which are now also done routinely. Other approaches are also possible, e.g. genome-wide epigenetic analysis (Hoheisel 2006) and approaches evaluating dynamic traits, developmental issues, including a time perspective (time-dependent genetic effects), and more functional mapping of QTLs (Wu and Lin 2006).

In our own studies we suggested, that by looking for genes that do change expression level after cytokine exposure in human islets and in a rat beta cell line, it would be possible to select or prioritize genes with functional significance in T1D pathogenesis. Gene expression may reflect an intermediate phenotype. This can be used in detection of disease gene variants, if the presence of DNA variation leads to some functional consequence, e.g. represented by altered gene expression. Ultimately, DNA variation in one of the candidate genes may be shown to be associated with a disease associated phenotype, such as the intermediate RNA expression level. Differences in expression levels may have other causes than genetic variation, which should also be explored. However, the current sample collection size of 8 human pancreatic islet preparations makes such analysis impossible. It will not be possible to study enough human islet preparations to establish significant correlations between gene expression and variations. We focused on an approach to pin-point potential functional candidate genes for further studies and have previously genotyped all coding SNPs in all the examined chromosome 21 genes (Bergholdt et al. 2005a). In the future, it would probably be more relevant and important to search for potential regulatory SNPs in these genes.

We used array-based approaches (Bergholdt et al. 2007a) for evaluation of the 32 positional T1D candidate genes identified on chromosome 21 (Bergholdt et al. 2005a). Eight human pancreatic islet preparations were initially used to compare expression levels of the human chromosome 21 genes in un-stimulated versus cytokine stimulated conditions. For this we used a so called Low Density Array (TaqMan) platform, containing expression assays for the selected genes as well as control genes (Bergholdt et al. 2007a). We searched for genes showing differential expression between cytokine stimulated and unstimulated conditions. Human genes down regulated after cytokine exposure of the islets were ITSN1, TTC3, CHAF1B, HUNK, OLIG1 and CBR1 (Bergholdt et al. 2007a), Figure 9 .

Evaluation of T1D related gene expression in human pancreatic islets is of extreme importance. Testing the target tissue of T1D itself, in which the pathogenetic process takes place, is essential. Human pancreatic islets are, however, very limited as a resource for research, and we believe our observations in the eight pancreatic islet preparations are unique and valuable.

By use of this Low Density Array platform, we further compared expression levels of the same 32 human chromosome 21 genes in mRNA from human lymphocytes (the effector cells in T1D) obtained from 10 newly diagnosed T1D patients (sampled less than 20 weeks after diagnosis) and 10 HLA-matched controls. When comparing the means of expression levels in the control group with the T1D group, none of the comparisons, however, reached statistical significance. Genes most differentially expressed in the two groups were ITSN1, CBR1, CLDN8, OLIG1 and CHAF1B , Figure 10 , (Unpublished data by Regine Bergholdt).

Furthermore, a rat beta cell line INS -1 αβ over expressing pdx-1 was used to evaluate gene expression of genes from this region. In this system we compared expression levels of rat homologues (if existing) of the human chromosome 21 genes in un-stimulated conditions with IL-1 β stimulated conditions (Bergholdt et al. 2007a). This analysis was part of a larger project evaluating differences in rat gene transcript expression between ± pdx-1 expression and ± IL-1 β stimulation (Karlsen et al. Unpublished data). Extensive microarray analyses were performed using an Affymetrix rat oligonucleotide array, containing 30,000 rat gene transcripts. We focused exclusively on gene transcripts homologous to human chromosome 21 genes and looked exclusively at differences in gene expression, with and without IL-1 β stimulation for 2 and 24h (Bergholdt et al. 2007a).

Use of other methods than methods controlling for mass significance are increasingly being accepted for microarray studies, despite the many hypotheses tested. We have chosen to use the FDR method to search for the most likely candidate genes, since our aim was to generate hypotheses regarding the potentially most likely genes, which then has to be examined further and validated.

By use of this approach, we identified several "functional" candidate genes. Five rat genes (corresponding to human chromosome 21 genes) significantly changed expression levels by IL-1 β stimulation, CBR3, DSCR1 and CRYZL1 were down regulated, whereas ATP5O and IFNGR2 were up regulated by IL-1 β (Bergholdt et al. 2007a). Findings observed in this model system needs confirmation in additional functional studies using other systems, but may be valuable in generating testable hypotheses about genes of presumed functional significance. In the present study we have used model systems primarily addressing beta cell function and response.

Although it is not always straightforward to extrapolate from one species to the other, high concordance was observed between expression changes in the rat beta cell line and human pancreatic islets (Bergholdt et al. 2007a). Genes showing differential expression in the approaches used were quite similar, supporting the identified specific gene products as functionally interesting. Selected genes were involved in e.g. the NF κ B-pathway, generally in cytokine signaling, oxidative phosphorylation, defense responses and apoptosis. Transcriptional profiling was used to prioritize among genes and select the ones with putative functional significance in T1D, Figure 11 .

Similar findings in human islets and the rat beta cell line strengthens the validity of the methods used, and supports this as a useful approach for gene mapping. We used this as follow-up on our fine mapping of the linkage region on chromosome 21, in which 32 positional candidate genes had been identified. Most important genes we hypothesize are genes identified by several different approaches, genetic as well as functional, Figure 11.

The analyses furthermore revealed several transcription factor binding-sites shared by several of the differently expressed genes. However, none of the binding sites demonstrated statistically significant over-representation (Bergholdt et al. 2007a). This suggests common regulation and functional interaction of this chromosome 21 region with regions on other chromosomes (e.g. via transcription factors), rather than among the chromosome 21 genes themselves. We think this is a valuable novel approach in selecting yet unidentified genes with a functional significance presumed to underlie multi-factorial genetic diseases. We have used the approach to prioritize chromosome 21 genes for further studies, and have especially focused on the CBR1 (carbonyl reductase 1) gene, which we are in the process of evaluating further in genetic as well as functional studies, Figure 12 (Bergholdt et al. 2005b; Bang-Berthelsen et al. 2006).

For many diseases access to the relevant target tissue, for evaluating expression of genes, is a problem. This is obviously true for T1D, since access to human pancreatic islets is limited. It could be speculated that in many diseases, human lymphocytes could act as a kind of easily accessible RNA source, potentially useful as a surrogate measure for gene expression in other (target) tissues. To evaluate this hypothesis, we used the expression data on chromosome 21 genes in human lymphocytes (Unpublished data by Regine Bergholdt) to calculate whether these could be identified as being representative for the expression levels we have obtained in human pancreatic islets. Such a comparison is not a simple procedure, the biological variation is extensive and the islet preparations with and without cytokine stimulation is not independent. We therefore only performed a simple visual analysis in which we compared the average increase or decrease in expression level for each gene, among control lymphocytes versus T1D lymphocytes as opposed to unstimulated islets versus cytokine stimulated islets.

As seen in Figure 12, these analyses do not support a simple correlation between gene expression in lymphocytes and pancreatic beta cells under the experimental procedures applied here. Thus, when studying human beta cell gene expression, human lymphocytes might not be an obvious surrogate marker system. However, this clearly needs further exploration. We have therefore focused on the beta cell data, and the eight human islet preparations used, do represent one of the largest studies of human beta cell gene expression ever.

Whether the gene responsible for the chromosome 21 linkage peak is expressed in islets is unknown. T1D is an immune-mediated disease and the few T1D genes validated to date, are immune system genes. However, the immune system may destroy beta cells because of features in beta cells themselves, e.g. the way they react to the immunological challenge. This is the reason we chose to use an established model for T1D pathogenesis, the cytokine stimulation model, mimicking what is happening in the beta cell in T1D pathogenesis for these studies. We can of course not rule out that important genes are expressed in immunological tissue, but to elucidate this was not the scope of the present study.

8.2. Functional interaction networks

Both genetic and physical data are essential to understand biological systems. For a complete picture they must be integrated and since genetic interactions define functional relationships between the genes, such integration is expected to provide insight into the translation of interactions to phenotype. Individual interactions can be integrated into larger network structures representing protein complexes or pathways. Whereas the genome is rather static, interaction networks are believed to be more dynamic and dependent on the context. They might be active only under certain conditions, in certain cell types or stages of development. Ideally, all conditions and cell types should be tested to capture this presumed variability.

For prioritization of positional candidate genes in broad linkage intervals the use of functional interaction networks (interactomes) may be a valuable method. If linkage intervals obtained for a disease are queried for functional interactions with each other and related to phenotype information for the disease, this holds promise for selection of putative disease genes for further investigation (Gandhi et al. 2006; Lage et al. 2007a). This is supposed to also be applicable to disease associated intervals identified from genome wide association studies, in which often little is known a priori about the functions of the significant loci. Studies like this might generate novel hypotheses regarding pathogenesis of human multi-factorial diseases. They have the potential of identifying new, previously unrecognized components of disease mechanisms, as well as of pin-pointing the most important protein complexes involved.

Furthermore, many diseases have overlapping clinical manifestations/sub-phenotypes and it could be speculated that this may be represented by genetic variation in the same functional pathways. Recently, the existence of so called disease sub-networks was suggested. It was demonstrated that proteins encoded by genes mutated in one inherited genetic disorder, were likely to interact with proteins known to cause similar disorders, presumably by sharing common underlying biochemical mechanisms (Gandhi et al. 2006).

A recent report demonstrated the feasibility of constructing such functional human gene networks and applying these to positional candidate gene identification (Franke et al. 2006). It was shown that obvious candidate genes are not always involved, and that taking an unbiased approach in finding candidate genes, e.g. by using functional networks may result in new testable hypotheses (Franke et al. 2006). In this report a functional human gene network was constructed by integrating information of genes, proteins, protein-protein interactions, microarray co-expressions etc. from a number of databases. Genetic diseases with at least three known disease genes were tested, 17 of the diseases were of complex inheritance. Artificial susceptibility loci with 100 genes around the known disease genes were constructed, and the functional networks then ranked the genes based on known functional interactions. In 54% of cases the known disease gene was ranked among the top five genes for each locus, corresponding to 2.8 fold better than random selection (Franke et al. 2006). This is not too impressive and part of the explanation may be that this report, as well as another recent one (Aerts et al. 2006) incorporate unscored human protein interaction data, and that none of them take advantage of cross-species integration of interaction data. Yet another recent report used only unscored binary interaction pairs to identify candidates (Oti et al. 2006). They identify new candidates and estimate that this corresponds to a 10-fold enrichment compared with positional information alone (Oti et al. 2006), the candidates and thereby the estimated enrichment, however, remains to be validated.

Approaches like these may be valuable in prioritizing candidate genes in linkage regions or from disease associated regions, in which the disease gene(s) are not known. And information of whether genes from the different loci observed, do interact at a functional level are potentially interesting.

Obviously, the input information is crucial for the success of such an approach. Studies will be biased by absence of complete functional information in databases of the majority of genes, and also interaction databases are far from complete. However, hypotheses generated with existing knowledge may be of value, and genes, that would otherwise not have been predicted to be involved in the disease in question, might be identified this way. Data amounts in databases are rapidly increasing. This include increased knowledge regarding genes, proteins, interactions among them, methods integrating high throughput genomic and proteomic approaches, as well as text mining methods extracting functional relationships from the literature. All information should be quality controlled, and preferably integrate information from other species.

In our own studies we searched for functionality for rules identified by data mining studies (Bergholdt et al. 2007b), in which several rules of genetic interaction were predicted. We used protein interaction networks to examine whether the gene products from the interacting genetic regions could also be shown to interact in biological pathways.

We used an analysis tool, recently developed at the Center for Biological Sequence Analysis, Danish Technical University, for integrating protein-protein interactions in generating protein interaction networks. At the Center for Biological Sequence Analysis a data warehouse has been developed, holding information from 17 different species and data pooled from all the largest interaction databases, Figure 13 .

High standards of quality control and the highest confidence score level were set as a threshold and applied to all interactions included in this data warehouse (Lage et al. 2007a), as opposed to in the other reports mentioned (Aerts et al. 2006; Franke et al. 2006; Oti et al. 2006). This combined human protein network can then be queried and functional protein interaction networks, corresponding to the chromosomal regions in question, be extracted.

We queried this network for interactions corresponding to regions 5 Mb on each side of the two microsatellites in every rule of predicted genetic interaction. For the functional networks extracted, a further stringent criterion (network interaction score) for number of interactions allowed for each non-input protein, was applied, in order to reduce noise from highly interacting proteins (Bergholdt et al. 2007b).

We attempted generation of protein-protein interaction networks for all predicted interactions, using the highly stringent confidence score as well as a stringent network interaction score. The strategy used is outlined in Figure 14 .

We demonstrated support for physical interactions at the protein level for all the predicted genetic interactions, demonstrating that genetic interaction is probably important in T1D susceptibility (Bergholdt et al. 2007b). The high degree of compatibility between genetic rules of interaction derived from data mining analysis, as well as the networks obtained, support that the genetic rules have predictive value for T1D and allow for identification of specific candidate genes/proteins of the involved genomic regions. The method developed and the analyses made represent a completely novel exploration of integrative genomics. We integrate genetic and physical protein interactions to identify protein networks that link genomic regions showing genetic interactions. The resulting networks point directly to novel candidates visualized in context of their interaction network, potentially providing even further biological insight. Furthermore, we integrated information from more than just the two directly genetically interacting regions and obtained some highly interesting and significant functional modules, Figure 15 . By integrating e.g. all the regions demonstrated to in themselves genetically interact with the HLA-region, we can directly point at candidate genes for the underlying interaction. We applied a statistical test devised specifically for this work to evaluate the significance of the protein networks linking several genetic loci that have shown genetic interactions. The networks and functional modules looked promising, supported by the highly significant nature of many of them.

Findings are in line with current knowledge regarding mechanisms involved in T1D pathogenesis, since central proteins in the protein-protein interactions were proteins involved in signal transduction, kinase activity, regulation of transcription and apoptosis, reviewed in (Bergholdt et al. 2004b).

In several of the networks including TNFA as a marker for the MHC region, the same four MHC-region genes are present, TUBB, RPS18, BAT1 and ITPR3 , and responsible for the interaction observed, Figure 15. Of these genes, recent focus has been on the ITPR3 gene, which demonstrated strong T1D association in two Swedish case-control cohorts (Roach et al. 2006), as described in chapter 4. It is of potential interest that we were able to pin-point the same gene as important also in putative gene-gene interactions. This may help putting this gene into a functional context of other interacting proteins, and we see the central position of this gene and other known HLA-region genes as an indication of the value of this method.

We examined whether additional information could be obtained by the use of a less stringent interaction score, allowing for more interactions for each non input-protein, but keeping the same confidence score. Interestingly, we did see the original networks being extended (Bergholdt et al. 2007b), and we hypothesized that this may add information of possible underlying biological pathways. As an example the protein-protein network between TNFA = 1 and D11S910 are shown with the stringent, as well as the less stringent interaction score, Figure 16 .

The extended network for this interaction included genes and pathways of possible relevance to T1D, i.e. classical HLA-genes, histone genes, the ITPR3 gene, a calcium-calmodulin associated protein kinase and genes encoding transcription factors. The combination of TNFA and D11S910 was the only rule, which seemed to confer protection against T1D (Bergholdt et al. 2007b), and the networks constructed may provide insight into mechanisms underlying this.

We have identified a number of completely novel T1D candidate genes. Many of the networks are highly significant, showing that the identified candidates are strongly backed up by a combination of the human protein interaction data and the genetic interaction data in patients with T1D. We therefore think this may be a strong and valuable approach for identification of functional candidate genes.

With 30% or more of human genes lacking functional annotation, existing protein interaction databases and maps are still far from complete, and many of the protein interactions in databases have not been rigorously tested and validated. Our data are the most extended suggestions of genetic epistasis in a multi-factorial disease backed up by high-confidence protein network analyses with implications for functionality. Our data present a new approach to the analysis of complex genetic data, as well as a new comprehensive way of surveying and extracting data from protein-interaction databases, representing a methodological advance in genomics and bioinformatics. A paper by several of the co-authors of our paper was very recently published, in which the new method for elucidating protein-protein interaction networks was validated and thoroughly described (Lage et al. 2007a). The combination of these two approaches does reveal significant and interesting observations, and we propose this type of integrative analysis as a general method for elucidation of genes and networks involved in T1D and other complex diseases. Future studies and validation, of the interaction complexes we observed, are necessary to confirm their biological significance.

In summary, we used an interaction network method and confidence scoring system previously developed (Lage et al. 2007b) to make interaction networks linking regions shown to interact genetically (identified by the method described previously (Pociot et al. 2004)), using data that has not previously been investigated by this method. The dataset was furthermore much larger than used before. Hereby we integrated genetic and physical protein interactions to identify protein networks that link genomic regions showing genetic interactions, and point directly to novel candidates. We included a statistical test devised specifically for this work in order to evaluate the significance of protein networks linking several genetic loci. All the identified networks were highly significant. The analyses made and developed do represent a completely novel exploration of integrative genomics. Integrative genomics is an emerging and promising field in order to tackle complex disease.

Other applications for use of knowledge regarding functional human protein-protein interaction networks, could be to look for interactions among all known T1D linked regions or among regions identified in a genome-wide association study. A strategy for this is exemplified in Figure 17 .