Contribution of Known and Unknown Susceptibility Genes to Early-Onset Diabetes in Scandinavia: Evidence for Heterogeneity

Pedigrees suitable for these studies were selected from our family collections in Finland and Sweden. All subjects gave their consent to the study, which was approved by the local ethic committees. Extensive phenotyping was performed including oral glucose tolerance test with measurements of blood glucose, serum insulin, and C-peptide after ingestion of 75 g glucose (21). Serum total cholesterol, HDL cholesterol, and triglyceride concentrations were measured on a Cobas Mira analyzer (Hoffmann-La Roche, Basel, Switzerland). Body weight and height as well as waist and hip circumference were measured, and BMI was calculated (7,22). GADA were determined by a radiobinding assay using ³⁵S-labeled recombinant human GAD65. GADA levels exceeding five reference index units were considered positive (23).

Subjects with fasting blood glucose >6.7 mmol/l or 2-h blood glucose ≥8.5 mmol/l or a previous diagnosis of diabetes documented by records and/or chronic treatment with oral agents and/or insulin were considered affected (22). In the Botnia study, individuals with 2-h blood glucose levels ≥8.5 mmol/l have a very high risk of developing diabetes (24,25). During a 3-year follow-up of our subjects with IGT, 25% with a 2-h blood glucose >8.5 mmol/l developed manifest type 2 diabetes compared with 3% of those with a blood glucose value <8.5 mmol/l (P < 0.0001).

Among our collection of 1,379 families (8,648 individuals, 3,376 of whom have diabetes at present), 857 families had at least two affected subjects, and 227 pedigrees contained at least two subjects with an age of onset of diabetes ≤45 years. From these 227 families, we selected 29 families (222 individuals, 119 affected) suitable for mapping. The average family size was 7.7 individuals, and the average number of affected individuals in each family was 4.1. All nongenotyped individuals and individuals unavailable for phenotyping were considered to have an unknown affectation status. Table 1 shows the clinical characteristics of the affected subjects included in the study.

Altogether, 41 individuals, including at least one from each sibship diagnosed with diabetes at an age ≤45 years, were screened for mutations in the five MODY genes (HNF-1α, GCK, HNF-4α, IPF-1, and HNF-1β genes), the PPARγ gene, and the A3243G mitochondrial DNA mutation. All known exons and promoters, including a newly discovered promoter region and alternate exon 1 in the HNF4α gene, were screened using single-strand conformation polymorphism and direct sequencing as described earlier (20,26–29).

The genome of each subject was scanned using 433 polymorphic microsatellite markers with an average heterozygosity of 0.75 (estimated from the data). On the basis of published maps, the mean sex-averaged distances between adjacent markers were ∼7 centimorgans (cM) with two intervals >20 cM. Genomic DNA (25 ng) prepared from peripheral blood lymphocytes was amplified in 15 μl of PCR reactions using 40 nmol of dNTP; 2.5, 3.75, 5.0, or 6.25 pmol of each primer; 0.25 units AmpliTaq (Perkin Elmer, Foster City, CA); 67 mmol/l TrisHCl (pH 8.8); 16 mmol/l (NH₄)₂SO₄; 0.01% Tween 20; and either 1.5 or 2.5 mmol/l MgCl₂. Each forward primer was labeled with one of three fluorescent dyes (HEX, FAM, or TET). PCR reactions were assembled using a Beckman robotic pipetting station and overlaid with 12 μl mineral oil and run on 96-microtiter format plates. Thirty cycles of PCR were performed (94°C for 30s, 55°C for 30 s, and 72°C for 30s), and 0.5–1.0 μl of each reaction was loaded onto a denaturing 4.25% polyacrylamide gel and electrophoresed for up to 4 h in 1× TBE using ABI377 or ABI373A Sequencers (Perkin Elmer) and thereafter processed using the ABI software GENESCAN/GENOTYPER (Perkin Elmer). Two independent scorers read all gels, and the data were checked for Mendelian segregation using the PEDMANAGER software (M.P. Reeve and M.J. Daly, personal communication). All discrepancies were subjected to an extensive error checking process, including repeating PCR for the entire family in question for the relevant locus. GENATLAS (http://bisance.citi2.fr/GENATLAS/genan.html) and NCBI LocusLink (http://www.ncbi.nlm.nih.gov/LocusLink/) were used for determining the approximate locations of markers and genes. In cases in which an approximate integration of different marker maps was required to localize them in relation to each other, the Marshfield Center for Medical Genetics maps were used (http://research.marshfieldclinic.org/genetics/MapMarkers/maps/IndexMapFrames.html).

The second exon of the DQB1 gene was genotyped by PCR and hybridization with sequence-specific oligonucleotide probes (23). Nine DQB1 probes were used to distinguish DQB1 alleles, 0201 or 0202 (02), 0301, 0302, 0303, 0501, either 0602 or 0603 (0602/3), and 0401 or 0402 (0401/2). Subjects with DQB*02 or *0302 alleles were defined as having type 1 diabetes risk alleles. According to this definition, 32.6% of nondiabetic population controls had risk alleles (23).

As the extent of genetic homogeneity and mode of inheritance of diabetes are largely unknown, we analyzed the data in a number of different ways. First, evidence of linkage was assessed with a nonparametric method using GENEHUNTER software (version 2.0) (30), which performs complete multipoint analysis of the statistical significance of allele sharing identical by descent among all affected family members at each location in the genome, in addition to estimating the information content. Second, we performed parametric multipoint analyses using the GENEHUNTER software (version 2.0). The parametric logarithm of odds (LOD) score and heterogeneity LOD (HLOD) score calculations were performed using an inheritance model proved successful in previous studies of early onset diabetes (MODY3) (13). Allele frequencies used in all of the analyses were those observed in the founders of the families in the study (30).

Interfamilial differences in phenotypes associated with diabetes, such as age at diagnosis (AAD), may reflect distinct subsets of the disease. To identify a more homogeneous group of families, we used an ordered-subset analysis (31) and ranked families according to their mean family value for AAD. Starting with the family with the lowest AAD, the family-specific nonparametric linkage (NPL) scores were added, one family at a time, until all families were included. The NPL score was determined after each family was added and the overall NPL obtained for any subset of the families was finally reported. This analysis was then repeated adding in families in the reverse order, starting with the family with the oldest age at onset. This analysis was simulated 10,000 times through adding families in random order over each chromosome. We report corrected P values for the results presented from this simulation.

To evaluate the contribution of autoimmunity to our linkage results, we reanalyzed our data excluding GADA-positive subjects. The significance of the results was simulated by randomly excluding the same number of subjects from the analysis. Both nominal and corrected P values are presented (Web Appendix 1, http://diabetes.diabetesjournals.org).

To identify possible genetic interactions with the HLA genotypes, we also conducted a conditional analysis whereby the linkage at the analysis locus was evaluated on the basis of the genotypes (DQB*02 and/or 0302) at the HLA locus (the conditioning locus). Family-specific weights for HLA risk genotypes were assigned on that basis of the proportion of affected individuals in each pedigree having risk genotypes. Families received weights on a scale of 0–1, whereby 1 corresponded to the highest risk (all affected individuals had DQB*02 and/or 0302). Multipoint NPL analyses, as implemented in Allegro (32), were then performed in subsets of families defined by the HLA risk. The proportions of families with different weight classes were kept constant in the simulations. A similar approach to assess interaction between genetic loci was described recently (33). This exploratory test does not correct for multiple testing or the different proportions of families in the differently weighted subsets. Therefore, we performed 200 simulations with random assignment of genotypes to match our data for the loci with indications of interactions (P_nominal < 0.05), which were used to assess the significance of any observed NPL score. Both nominal and corrected P values are presented in Table 2.

By definition, significant linkage is considered present if the P value in a genome-wide scan is <2 × 10⁻⁵ and NPL_all >4.1, whereas suggestive linkage requires a P value <1 × 10⁻³ and NPL_all >3.0 (34). To allow for incomplete informativeness in the data set, we estimated the thresholds for suggestive and significant genome-wide linkage in our family collection by simulations using the GENSIM computer program designed by Mark J. Daly (unpublished data). This program created a data set matching ours with respect to marker informativeness, number of genotyped individuals, affection status, and fraction of missing data (3%). Specifically, 200 whole-genome screens (4,400 chromosomes) were generated at the genome-scan marker density. In this manner, the overall genome-wide threshold for suggestive linkage was determined to be at an NPL score of 2.8, whereas significant linkage required NPL scores of 3.5 (P ≤ 0.05) and 5.0 (P ≤ 0.01), respectively.

To evaluate the power of our family set to identify a disease locus, we generated 100 sets of genotype data with identical family structures with the GENSIM program. These simulations matched our data sets with respect to marker informativeness, individuals genotyped, affection status, and fraction of missing data (3%). Each of these data sets was generated assuming a dominant disease model with a single gene segregating with disease, using two different α values (the proportion of families showing linkage to a certain locus) because the extent of heterogeneity is unknown. First, using an α value of 80%, suggestive linkage (NPL score 2.8) was detected with a power of 99% and significant linkage (NPL 3.5) with a power of 96% in our families. When α was set to 50%, suggestive linkage was detected with a power of 75% and significant linkage with a power of 47% in our families.

Furthermore, we used a resampling simulation strategy to evaluate our subanalyses: 1) In the first simulation, results from 26 families were randomly drawn from the total 29 families; 1,000 simulation runs established suggestive linkage at an NPL score of 2.5 and significant linkage at NPL score of 3.1 (P ≤ 0.05) and 3.3 (P ≤ 0.01), respectively. 2) In the second simulation, results from five families were randomly drawn from the 26 families that did not segregate with MODY/MIDD; 1,000 simulation runs yielded suggestive linkage corresponding to an NPL score of 3.0 and significant linkage to NPL scores of 4.1 (P ≤ 0.05) and 4.7 (P ≤ 0.01), respectively. All of our P_corrected values refer to the simulations described above. We also report P_corrected values with an additional Bonferroni correction to correct for the fact that we had performed five subanalyses.

Accession numbers and URLs for data in this article are as follows: Online Mendelian Inheritance in Man (OMIM), http://www.ncbi.nlm.nih.gov/omim/ for MODY1 (MIM#601283), MODY2 (MIM#601407), MODY3 (MIM#600496), MODY4 (MIM*151690), MODY5 (MIM#100640), MIDD (MIM*229300), PIK3C2B (MIM*602838), SLC2A1 (GLUT1; MIM*138140), and PRKAA2 (MIM*600497). The GENEHUNTER software is found at Whitehead Institute for Biomedical research/MIT Center for Genome Research (http://www.genome.wi.mit.edu/ftp/distribution/software/gh2/).

Approximately 26% of all of our families fulfilled the criteria for EOD (two individuals with age of onset before 45 years). Of the 29 screened families, 7 (24%) were found to have either MODY or MIDD mutations. In two families, mutations (R131Q and L107I) in the HNF-1α gene (MODY3) segregated completely with diabetes, and both variants have been shown to be functional (20). Mutations in the IPF-1 gene (MODY4) were found in patients from three families, two of which had the P239Q and one of which had the G212R variant (28). One family had members with an A241T variant in the HNF1-β gene (MODY5), which also may be functional (27). None of the observed MODY4 or MODY5 variants co-segregated completely with diabetes. The A3243G mitochondrial DNA mutation was found to segregate with diabetes and hearing loss in one family (20). In the three families in which MODY/MIDD mutations segregated with disease, four subjects (31%) had an initial diagnosis of type 1 diabetes, eight (62%) subjects had a diagnosis of type 2 diabetes, and one (7%) subject had a diagnosis of IGT (Table 1).

Among the subjects with EOD and available GADA measurements, 20% (20 of 102) were GADA-positive. The prevalence of GADA was approximately the same in subjects from families with MODY/MIDD (18%) as in other subjects with EOD (20%) (Table 1). HLA risk alleles (DQB*02 and/or 0302) were found in 68% (80 of 118) of affected individuals. The alleles were present in 23% of the MODY/MIDD subjects compared with 73% in the other subjects (Table 1).

To screen for novel susceptibility genes that predispose to EOD, we first genotyped all individuals in the 29 families. The highest observed NPL score was 2.7 (P_nominal < 0.01; a suggestive NPL score would have required an NPL score of 2.8 based on our simulations; see research design and methods) between markers D1S473/D1S438 on chromosome 1p. Six other regions on chromosomes 3p, 7q, 11q, 18q, 20q, and 21q showed P_nominal < 0.05 (Fig. 1).

We then repeated the analysis after excluding the three families in which MODY or MIDD mutations segregated with diabetes. The highest NPL scores in the 26 remaining families were found on chromosomes 1p (D1S438; NPL 2.6; P_nominal < 0.01), 3p (D3S1620; NPL 2.2; P_nominal < 0.03), 5q (D5S1465; NPL 2.1; P_nominal < 0.03), 7q (D7S820; NPL 2.0; P_nominal < 0.03), 18q (D18S535; NPL 1.9; P < 0.04), 20q (D20S195; NPL 2.5; P < 0.02), and 21q (D21S1446; NPL 2.2; P_nominal < 0.03; Fig. 1). Notably, the linkage on chromosome 1p31.11 was suggestive according to our resampling simulations (P_corrected < 0.09). When the GADA-positive individuals were excluded from the analysis, this linkage was no longer suggestive (Web Appendix 1).

Furthermore, we performed an ordered subset analysis based on the mean family AAD. The significance of the obtained NPL score for each chromosome was determined in 10,000 simulations. One region on chromosome 22q11–q13 (D22S689–D22S685; NPL score 4.7; P_corrected < 0.01) reached statistical significance in one family; after Bonferroni correction, the P value was <0.05 (Fig. 2). To examine whether this was supported by the other families, we repeated the analysis after exclusion of this particular family. The maximum NPL score of 1.9 in this analysis was above the threshold for suggestive linkage (NPL score ≥1.2). Three regions on chromosomes 3p, 18q, and 21q reached similar NPL scores, but the P values were no longer suggestive after the Bonferroni correction. When analyzing the families added in reverse order, starting with the family with oldest mean age of onset, no significant or suggestive loci were found.

To identify possible genetic interaction with the HLA locus, we also conducted a conditional analysis whereby linkage at one locus was evaluated on the basis of risk genotypes (DQB*02 and/or 0302) at the HLA locus. In families with several affected subjects with risk genotypes (DQB*02 and/or 0302), eight regions on chromosome 1p, 5q, 6p, 7q, 11q, 20q, and 21q displayed P_nominal < 0.05 (Table 2). When the analysis was restricted to the families with the highest (1) weight for risk genotypes (DQB*02 and/or 0302) at the HLA locus, five regions on chromosomes 1p, 2q, 11q, 13q, and 15q showed P_nominal < 0.05 (Table 2). After simulation of our data and correcting for the exploratory analysis, only chromosomes 20q and 21q showed suggestive linkage (P_corrected < 0.05), but this was no longer so after the Bonferroni correction.

To increase homogeneity, we also analyzed a subsample of five families (Fig. 1) with autosomal dominant inheritance of EOD (at least two subjects with AAD ≤35 years). Of the 23 diabetic subjects, 2 (9%) in these families were GADA-positive, and 11 (65%) had HLA risk alleles (Fig. 2). Of the patients, 7 (30%) initially had a diagnosis of type 1 diabetes, 15 (65%) of type 2 diabetes, and 1 (4%) of IGT. The highest NPL score in these families, suggestive by simulations, was seen on chromosome 1p (D1S438–D1S1665; NPL score 3.0; P_nominal < 0.01). Chromosome 16q fell just short of being suggestive (D16S419; NPL score 2.9; P_nominal < 0.01), and four additional regions on chromosomes 4, 5, 8, and 12 showed P_nominal < 0.05. None of these regions reached statistical significance or was suggestive after correcting for the exploratory analysis. We also analyzed linkage to diabetes in these families using a parametric multipoint approach. The highest HLOD scores from this analysis were observed on chromosomes 3, 4, 5, 8, and 12. The position of the loci on chromosomes 4, 5, and 8 overlapped with the results from the NPL analysis (Web Appendix 2, http://diabetes.diabetesjournals.org).

Approximately 26% of familial diabetes in our population showed an early onset (at least two family members with onset of diabetes ≤45 years). Of them, 13% had mutations in MODY or mitochondrial genes (20), indicating that there are underlying genetic causes left to be found in families with EOD. To unravel some of the unknown genetic background of EOD, we performed a genome-wide scan in 29 pedigrees including 119 affected subjects. In addition, we screened them for mutations in known MODY genes, the A3243G mitochondrial DNA mutation, and presence of GADA and type 1 diabetes HLA risk alleles DQB*02 and/or 0302. The frequency of MODY/MIDD mutations in this study is higher than we have previously published (20). This is not an unexpected finding, as we selected large families with diabetes in several generations.

Subjects with MODY/MIDD mutations were not excluded from the initial analysis, as we have previously seen both MODY and mitochondrial mutations segregating in the same families (20,28). To exclude the possible confounding effect of MODY or mitochondrial genes, we repeated the analysis after excluding the three families in which MODY or MIDD segregated completely with the disease, but the results were virtually unchanged. It could be argued that our sample size (29/26 families) is too small to detect linkage, especially if the disease is heterogeneous. However, it was possible to detect linkage to MODY1 and MODY3 using a small number of families with EOD (10,13). Families with EOD have proved successful for dissecting the genetics of other diseases such as breast cancer (35). Our linkage analysis suggests three chromosomal regions as susceptibility regions for EOD: chromosomes 1p31.11, 20q11, and 22q11–13.

The best evidence for linkage in all families was seen on chromosome 1p31.11 (Fig. 1), and this region also showed some evidence of linkage when conditioned for the HLA locus (Table 2). It is interesting that exclusion of the GADA-positive individuals from the analysis abolished the linkage, suggesting that autoimmunity may play a role in the subjects studied. In Pima Indians, modest evidence for linkage was observed between the 1p31.11 region and type 2 diabetes with retinopathy or nephropathy. The presence of advanced microvascular complications could point at an earlier onset of the disease (36). Ehm et al. (37) also reported modest linkage between this region and type 2 diabetes in two of the four GENNID populations.

The 1p31.11 region contains several interesting candidate genes for diabetes, including phosphatidylinositol 3-kinase class 2β (PIK3C2B; MIM*602838), which is involved in receptor-mediated signal transduction (38), and glucose transporter 1, SLC2A1 (GLUT1; MIM*138140), which is expressed in most organs and responsible for basal glucose transport (39). A third candidate, the catalytic subunit of protein kinase α-2 (PRKAA2; MIM*600497), plays a key role in the regulation of fatty acid and cholesterol metabolism (40,41).

It is interesting that linkage was observed between chromosome 20q11 and all diabetic phenotypes, i.e., diabetes, per se, diabetes plus high-risk HLA genotypes, and GADA-negative diabetes. This region has been linked to diabetes, particularly EOD (42,43,44,45,46), in several studies. Zouali et al. (45) reported suggestive linkage to this region in French patients with type 2 diabetes with age at onset ≤45 years. Furthermore, Ghosh et al. (43) reported putative linkage to this region in a sample of patients with late onset type 2 diabetes in the Fusion study from Finland. Similarly, Ji et al. (44) reported some evidence of linkage to the same region in a sample of U.S. white families with an age at onset of 47 ± 12 years. Notably, suggestive linkage to 20q was also seen in a U.S. white sample of patients with type 2 diabetes with nephropathy (42). This could be a corollary of EOD, as nephropathy is more likely to develop in patients with early onset disease. In fact, the mean age of onset in this sample was 45.6 years (42). Permutt et al. (46) recently published the fifth study supporting this region (Z score 2.05) as an interesting candidate region for diabetes.

Impaired insulin secretion seems to be the trigger for EOD. Therefore, the MODY1 gene, HNF-4α, could be a prime candidate on chromosome 20. However, in at least three studies, HNF-4α has been screened and excluded as the cause of linkage to this region (45,47,48). We did not find mutations in the promoter or coding regions of the HNF-4α gene in our subjects.

Chromosome 22q11–13 was linked (P_corrected < 0.05; Fig. 1A) to EOD in the family with the youngest age of diagnosis. This family would not fit our MODYX definition because it lacks clear-cut autosomal dominant inheritance (Fig. 1B). This region has shown a modest LOD score (1.3) for linkage to fasting glucose in nondiabetic Pima Indians (49). Furthermore, this region has shown suggestive evidence of linkage (Z score 2.5; P < 0.01) to type 2 diabetes in a small group of Canadian Oji-Cree sibships (50). In the Canadian study, the most frequent allele of the reported microsatellite (D22S683) was also associated with the disease (P_nominal < 0.02) (50). It is interesting that a microsatellite ∼10 cM away showed strong nominal association (P < 0.0001) with type 2 diabetes in a large sibship sample of Finnish origin (43).

Only two obvious candidate genes are found in the 22q11–13 region: the intestinal sodium/glucose co-transporter (SLC5A1; MIM*182380), which is responsible for “active” transepithelial glucose absorption (51), and the cytochrome c oxidase 7b (COX7b; MIM*603792), which is a subunit of the terminal component of the respiratory chain complex (52). The respiratory chain complex has been implicated in regulation of insulin secretion in a mouse model for mitochondrial diabetes (53).

At least 20% of MODY patients are unaccounted for by known MODY genes (3,8,17–20), suggesting that other MODY genes could still be operative (MODYX). To increase homogeneity and to search for such genes, we analyzed five EOD families (Fig. 3) with an autosomal mode of inheritance in at least two generations and with at least two members with age at onset <35 years. The age limit of 25 years in the classical MODY criteria (3) was replaced with 35 years because ∼25–50% of MODY patients show onset of the disease after 25 years (18,19,20) and regular type 2 diabetes generally occurs later than at 35 years (54).

In this nonparametric subanalysis, the chromosome 1p region showed suggestive linkage after simulations (however, not after correcting for the exploratory analysis; Web Appendix 2). Although the parametric linkage analysis in general yielded lower LOD scores, some of the regions overlapped with those from the nonparametric analysis (Web Appendix 2). This could, of course, indicate that the NPL scores represent random fluctuations, but it must also be taken into consideration that the parametric model might not be optimal because many of the parameters (mode of inheritance, penetrance, etc.) of MODYX are not known and might be expected to differ from MODY3 for which the model was initially specified (13). Another possible explanation is that EOD is polygenic rather than monogenic. That MODY3 and MODY4 mutations and MODY3 and MIDD mutations have been seen to co-segregate with diabetes in families would support such a view.

In conclusion, considerable heterogeneity was observed in our sample of patients with familial EOD from Scandinavia; 24% showed mutations in MODY or MIDD genes, whereas ∼60% had genetic and/or immunologic markers of type 1 diabetes. Given that the majority of the genetic causes of MODY have been unraveled and the remaining part seems to be heterogeneous, collaborative efforts may be needed to obtain enough families for the dissection of the remaining genetic causes of familial EOD.

This work was supported by grants from the Sigrid Juselius Foundation, the Juvenile Diabetes Foundation, Wallenberg Foundation, the Finnish Diabetes Research Foundation, The Swedish Medical Research Council, the Novo-Nordisk Foundation, and a European commission (EC) grant Genome Integrated Force in Type 2 Diabetes (grant no. QLG-CT-1999-00546 to L.C.G), the Medical Faculty of Lund University, Royal Physiographic Society, Anna-Lisa and Sven Lundgrens Foundation and Dir. Albert Påhlssons Foundation (C.M.L., M.L., O.M., and T.T.), and the Swedish Society for Medical Research (M.L., T.T.). C.M.L. is supported by the Foundation for Strategic Research through National Network for Cardiovascular Disease (NNCR).

We thank the families for participation in the study. Anna Berglund, Anita Nilsson, Margareta Svensson, and Malin Åberg are acknowledged for technical assistance during the project, and Mark J. Daly is acknowledged for the permission to use the GENSIM program. The simulation analyses were performed using the computers at the Center for SuperComputers, Finland.