Assessing the phenotypic effects in the general population of rare variants in genes for a dominant mendelian form of diabetes

Abstract

Genome sequencing can identify individuals in the general population who harbor rare coding variants in genes for Mendelian disorders^1–7 – and who consequently may have increased disease risk. However, previous studies of rare variants in phenotypically extreme individuals have ascertainment bias and may demonstrate inflated effect size estimates^8–12. We sequenced seven genes for maturity-onset diabetes of the young (MODY)¹³ in well-phenotyped population samples^14,15 (n=4,003). Rare variants were filtered according to prediction criteria used to identify disease-causing mutations: i) previously-reported in MODY, and ii) stringent de novo thresholds satisfied (rare, conserved, protein damaging). Approximately 1.5% and 0.5% of randomly selected Framingham and Jackson Heart Study individuals carried variants from these two classes, respectively. However, the vast majority of carriers remained euglycemic through middle age. Accurate estimates of variant effect sizes from population-based sequencing are needed to avoid falsely predicting a significant fraction of individuals as at risk for MODY or other Mendelian diseases.

For personal genome sequencing to help identify at-risk individuals for preventative care^1–7, the ascertainment bias typically employed in human genetic research presents a double-edged sword: studying individuals with extreme phenotype has efficiently identified disease variants but has also introduced an upwards bias in reported effect sizes^10–12. Examples include initial inflation of reported BRCA1/BRCA2 (breast cancer; OMIM#114480)^8,16 and HFE (hereditary hemochromatosis; OMIM#235200)⁹ mutation penetrance, remedied once adequately-sized population-based control groups were studied. Absent similar studies for other diseases, widespread personal genomic testing might exaggerate risk estimates, instigating needless intervention in low-risk individuals^1,2,6,17.

Maturity-onset diabetes of the young (MODY, OMIM#606391)^18,19 is a good candidate for personal genomic screening: (a) MODY is caused by dominant Mendelian mutations, such that heterozygous carriers develop disease²⁰; (b) clinical presentation occurs early in life (<25 years) with non-ketotic hyperglycemia^18,19; (c) MODY frequency is 0.1%-0.2% in European populations^21,22, with the majority un- or misdiagnosed²³; (d) MODY diagnosis can significantly impact on diabetes prognosis and treatment^24,25 of the individual or affected family members²⁶; (e) mutations in the MODY genes also influence late onset phenotypes, as common variants near many of these genes are associated with type 2 diabetes (T2D) risk in the general population²⁷; and (f) this risk can be reduced via lifestyle intervention^28,29. Thus, individuals identified through personal genomics to carry variants in MODY genes might be anticipated to exhibit elevated glycemic parameters by young adulthood, or at minimum by middle age, and could benefit from early intervention.

We aimed to characterize the spectrum of low frequency variation in MODY genes in the general population. Specifically: i) how many individuals carry rare variants in MODY genes of the sort that might be bioinformatically flagged in a personal genome sequencing context, and ii) what percentage of these carriers demonstrate an abnormal glycemic phenotype by middle age. In addition to randomly-ascertained individuals from population cohorts, we validated our methods using individuals selected for MODY or T2D diagnosis.

We focused on seven genes: four most-frequently (HNF1A³⁰, GCK^31,32, HNF4A³³, HNF1B³⁴), and three less-frequently (PDX1³⁵, INS³⁶, NEUROD1³⁷), mutated in European MODY patients. Six other MODY genes were not studied: two because loss-of-function causes hypoglycemia rather than diabetes (ABCC8, KCNJ11)³⁸, and four because sequence data were unavailable (PAX4³⁹, BLK⁴⁰, KLF11⁴¹, CEL⁴²; see Methods).

Primary analysis focused on 4,003 individuals drawn from three population-based cohorts (Table 1). First, we randomly ascertained 1,541 individuals of European ancestry from the Framingham Heart Study Offspring Cohort¹⁴ (FHS cohort) and 1,691 of African-American ancestry from the Jackson Heart Study (JHS cohort)¹⁵; these individuals are referred to as ‘unselected’ (Supplementary Figs. 1–3). Second, from Finnish and Swedish cohorts (>27,500 individuals), we ascertained 771 individuals from the extremes of T2D genetic risk (T2D cohort; 362 young lean T2D cases; 409 elderly obese euglycemic controls¹¹); these individuals are referred to as ‘extreme’ (Supplementary Figs. 1,3–5; Methods). Target capture and DNA sequencing⁴³ of the seven analyzed genes (Methods) were used to identify sequence variants: >93% of bases were covered with at least 20 reads in all genes except INS (79.1%) and PDX1 (37.7%; Supplementary Table 1). Genotyping of select variants, including those observed in single individuals, suggested very few (<0.05%) genotype calls were incorrect (Methods, Supplementary Data 1).

Table 1.

Description of studied individuals.

	Unselected cohorts						Extreme T2D cohort		MODY patients
	FHS cohort			JHS cohort
	No DM	DM/No Med	DM/Med	No DM	DM/No Med	DM/Med	No DM	DM
Number	1435	54	52	1345	54	292	409	362	250
Male (%)	48.0	61.1	65.4	38.7	33.3	31.2	50.6	45.6	46.5
Age (yr)	55.1 ± 9.5	59.0 ± 7.8	60.4 ± 8.1	55.5 ± 11.5	57.7 ± 10.0	60.5 ± 10.1	67.4 ± 7.3	53.0 ± 9.2	22.3 ± 12.4
BMI (kg/m2)	25.1 ± 5.1	27.7 ± 5.5	28.6 ± 5.4	31.3 ± 7.1	33.3 ± 5.6	34.2 ± 7.2	33.6 ± 3.6	23.6 ± 2.0	23.5 ± 4.9
FPG (mg/dL)	95.5 ± 9.6	171.9 ± 55.3	183.3 ± 71.6	91.5 ± 9.3	183.8 ± 66.7	141.9 ± 56.7	92.2 ± 6.0	167.9 ± 51.1	134.2 ± 39.2

In total, 121 non-synonymous variants were identified across the seven genes (Supplementary Fig. 6; Supplementary Tables 2–3). Variants with minor allele frequency (MAF) >1% were omitted from analyses as these have been well-studied in larger epidemiological cohorts^44,45. Following the model increasingly used in sequence interpretation^7,17, we categorized the remaining 108 variants into four non-exclusive classes (Fig. 1, Table 2; Methods):

1. low frequency non-synonymous: resulting in altered protein sequence

2. possibly pathogenic: (a) evolutionarily conserved site; (b) private to one study individual, not observed in 1000 genomes project⁴⁶; and (c) computationally predicted as protein-damaging by mutation analysis tools SIFT⁴⁷ and PolyPhen-2⁴⁸

3. HGMD MODY: previously reported as causal for MODY in the Human Gene Mutation Database Professional v2012.1⁴⁹

4. putative pathogenic: meeting criteria for possibly pathogenic and HGMD MODY.

Figure 1. Description of low frequency non-synonymous variants.

Shown is a summary of all low frequency (MAF<1%) non-synonymous variants identified in the three cohorts (Supplementary Tables 1–3 show lists of all variants, not simply those of low frequency). MAF is calculated as the maximum frequency across the three cohorts. Shown also is the number of variants fitting each annotation (HGMD, conserved, rare, and damaging) and each variant class: low frequency non-synonymous (black, magenta, orange, or green), possibly pathogenic (orange or green), HGMD MODY (magenta or green), and putative pathogenic (green). Twelve low frequency variants fit no annotations. Low frequency non-synonymous is abbreviated to ‘Nonsyn <1%’.

Table 2.

Variant counts in each of the seven sequenced genes.

Gene	Total	HGMD	Conserved	Rare	Damaging
GCK	13	3	9	9	4
HNF1A	27	12	17	19	11
HNF1B	17	4	12	9	6
HNF4A	23	6	12	20	6
INS	6	0	0	2	0
NEUROD1	11	0	8	9	3
PDX1	11	0	5	3	5
Total	108	25	63	71	35

Despite the recognized limitations of bioinformatics criteria alone for disease risk prediction^7,50,51, variants similar to those in the HGMD MODY and putative pathogenic classes are likely to be reported from personal genome sequencing^52–55. Possibly pathogenic variants are also relevant in personal genomics as they fit criteria used to ascribe pathogenicity to variants identified via next-generation sequencing in Mendelian disease studies^56,57.

For validation, we applied these bioinformatic criteria to DNA sequence data obtained from 250 Norwegian patients fitting MODY diagnostic criteria (Table 1, Supplementary Fig. 7). In total, 48% of these patients carried variants meeting the low frequency non-synonymous criteria (Supplementary Tables 4–5), 32% carried variants meeting the possibly pathogenic or HGMD MODY criteria, and 19% carried variants in the putative pathogenic class. As expected, a substantial fraction of patients fitting classic MODY diagnostic criteria carry variants that meet these annotation classes.

We used the same bioinformatics approach to analyze variants found in unselected individuals. For FHS and JHS: 4.4% and 5.7% of individuals carried a variant in the low frequency non-synonymous class (Supplementary Table 5), 1.5% carried variants in the HGMD MODY class, and 0.5% carried variants in the possibly pathogenic class. While these frequencies are a tenth the rate observed in MODY patients, they are at least an order of magnitude greater than estimates of MODY prevalence in the general population^22,58.

We asked whether these variant carriers possessed clinical features associated with MODY (Methods): diabetes diagnosis ≤25 years (proband or family members), lean (BMI<25), and family history of diabetes (≥two generations; typically early-onset)^13,20,59. One variant carrier in the JHS cohort, and none in the FHS cohort, fit these criteria (although several non-carriers did; Methods). These results suggest that only a small minority of individuals carrying such mutations demonstrate clinical characteristics consistent with MODY.

Despite lacking signs of MODY, these variant carriers might nonetheless be at elevated risk for T2D or hyperglycemia. We calculated among variant carriers the prevalence of: i) diabetes (‘receiving medication for diabetes’ or fasting plasma glucose (FPG) levels>126mg/dL), or ii) impaired fasting glucose⁶⁰ (IFG; FPG 100–126mg/dL).

For carriers in the unselected cohorts, point estimates of risk for diabetes were not elevated relative to non-carriers (Fig. 2a, Supplementary Table 6). Carriers of low frequency non-synonymous variants showed no trend toward increased risk (FHS OR=1.1, JHS OR=0.9, p>.1), with HGMD MODY or putative pathogenic variants having similar effect size estimates. The majority of variant carriers in these two classes (>90% of the FHS and near 80% of the JHS) did not develop diabetes, and only two of four carriers (50%) of variants in the most stringent putative pathogenic class developed diabetes (p=0.049).

Figure 2. Phenotypic impact of variants in unselected cohorts.

Shown is the fraction of variant carriers in the two unselected cohorts with (a) diabetes and (b) IFG or diabetes. Separate fractions are given for each of the four defined variant classes, with the fraction of non-carriers with each phenotype shown as reference. Error bars reflect 68% confidence intervals in the estimated fractions and are computed via the Clopper-Pearson method. The number of analyzed samples is given in Supplementary Table 6. Fewer individuals had FPG measurements than diabetes measurements; thus the number of analyzed individuals for these two phenotypes differs. ‘Nonsyn <1%’ is low frequency non-synonymous, ‘Pos. Pathogenic’ is possibly pathogenic, and ‘Put. Pathogenic’ is putative pathogenic.

Furthermore, we did not observe a trend towards IFG in variant carriers (Fig. 2b, Supplementary Table 6), despite many individuals being middle aged (Supplementary Fig. 8) and subjected to long-term follow-up¹⁴. Only 32% of HGMD MODY variant carriers, and 31% of proposed pathogenic variant carriers, had IFG or diabetes (compared to ~35% of non-carriers). These results are consistent with previous reports showing lack of association between PDX1 variation and diabetes⁶¹, although they apply more broadly across all seven analyzed genes.

We performed several additional analyses to investigate the observed low penetrance of variants in the unselected cohorts. To further validate our bioinformatics criteria, we computed carrier frequencies in the extreme T2D cohort (Supplementary Tables 5,7). Relative to old obese controls, young lean T2D cases carried a three-fold excess of low frequency non-synonymous variants (4.7% vs. 1.5%, OR=3.2, p=0.011) and an apparent excess of possibly pathogenic (four observations exclusive to cases, p=0.04) or HGMD MODY variants (four case observations, one control observation). Absolute variant frequencies in the extreme cohort are not directly comparable to those in the unselected cohorts due to differences in ethnic composition. The relative frequencies between cases and controls, however, validate the ability of bioinformatics criteria to identify an enrichment of rare mutations not only in individuals with MODY but also in those with T2D – provided that these individuals are pre-selected for phenotype rather than drawn at random from the population.

We next investigated whether analyzing multiple genes at once might influence our results, as MODY clinical presentation varies with gene affected and observed mutation^13,20. While 64 individuals carried possibly pathogenic or HGMD MODY variants in the four most commonly-mutated MODY genes, only four carried variants in INS, PDX1, or NEUROD1. Thus, inclusion of these three additional genes had a minimal impact on our results (Supplementary Tables 8–9).

Next, we analyzed variants specific to GCK, where mutations cause only mild and typically stable elevations in FPG (99–144 versus 72–108mg/dL in normoglycemia)^20,62. While only two carriers of low frequency non-synonymous GCK mutations met diabetes criteria, 67% (8/12) had FPG levels ≥99mg/dL (compared to 35% of non-carriers, combined p=0.054; Fig. 3a; Supplementary Table 10). Thus, individuals who carry GCK variants may display mild hyperglycemia. This result would need to be precisely communicated in the clinic as GCK-MODY patients are typically treated by diet modification alone⁶² and do not show accelerated decline in beta-cell function⁶³.

Figure 3. Phenotypes of GCK and HNF1A variant carriers.

(a) Shown are FPG values for GCK-variant carriers in the FHS cohort (left) and JHS cohort (right). Three dashed lines correspond to defined FPG thresholds: top (solid) line represents diabetes (126 mg/dL), middle (dashed) line IFG (100 mg/dL), bottom (dotted) line GCK-MODY (99mg/dL). A histogram and box plot representing FPG levels in the non-diabetic population (computed separately for each cohort) is shown for comparison. Two GCK variant carriers were on medication for diabetes and are thus excluded from the plot. Tabular form of these results (including the two carriers with diabetes) is shown in Supplementary Table 10. (b) The scatter plot shows fasting-2hr post-OGTT plasma glucose increment (y-axis) and FPG (x-axis) for each HNF1A-variant carrier in the FHS cohort (OGTT information was unavailable for the JHS cohort). Histograms showing FPG and fasting-2hr post-OGTT plasma glucose increments for individuals in the FHS cohort without diabetes are shown on the left and below the scatter plot respectively. FPG values for individuals receiving treatment for diabetes were omitted from the plot. The vertical and horizontal dashed lines represent the IFG threshold (100mg/dL) and a plasma glucose increment consistent with HNF1A-MODY/beta-cell dysfunction (≥90mg/dL) respectively. Points are colored corresponding to the annotation class of the variant; for variants that belong to multiple classes, colors are chosen according to the following precedence: putative pathogenic, HGMD MODY, possibly pathogenic, low frequency non-synonymous.

Finally, we expanded the set of clinical read-outs used in our analysis of specific genes. For HNF1A variant carriers, we asked whether glucose tolerance is affected even if FPG remains within euglycemic range – a possibility suggested as i) HNF1A-MODY patients can demonstrate elevated fasting-2hr blood glucose increments (≥90mg/dL) following an oral glucose tolerance test (OGTT)^63,64, and ii) the common HNF1A c.293C>T (p.Ala98Val) variant is associated with altered beta-cell function⁶⁵. Only one (of 17) HNF1A-variant carriers had an elevated fasting-2hr glucose increment (Fig. 3b; Supplementary Fig. 9). Likewise, HNF1B-MODY can be accompanied by renal dysfunction³⁴; none of the HNF1B variant carriers had abnormal creatinine values⁶⁶ (Supplementary Fig. 10).

In summary, a substantial proportion of individuals in the general population carry low frequency non-synonymous variants in one of seven MODY genes. Two classes of bioinformatics criteria (either previously reported as causing MODY or rare, conserved, and computationally predicted damaging) are each sufficient to identify a substantial enrichment of variants in individuals diagnosed with MODY or selected for an extreme diabetic phenotype. However, for each class, the majority of variant carriers observed in the general population remain euglycemic through middle age. These results highlight the limitations of disease variant databases, as well as objective and stringent bioinformatics criteria, in ascribing pathogenicity to rare variants.

Our study has multiple limitations. As the individuals are drawn from different genetic and environmental backgrounds, frequencies and effect sizes are not directly comparable across cohorts. This reflects one of the potential challenges of personal genomics: that the personal genome analyzed may not match the datasets from which effect size estimates are drawn. Also, because rare coding variants have low counts even in a study of thousands of participants, power is limited for statistical assessment of heterogeneity across genes or variants; it is possible that a subset of the variants identified in fact have very large effects. Expert interpretation of each gene and variant, information about other family members, and functional characterization of variants may identify a subset with large and robust effects – but these are not yet practical, let alone standardized⁷, in the automated analysis of personal genome sequences.

MODY is a useful model for studying the application of personal genome sequencing to disease risk prediction^13,26: a number of causal genes are well established and a dominant pattern of inheritance predicts that heterozygous mutation carriers will have a phenotype. However, even for this one disease, extrapolation of effect size estimates from extreme individuals to unselected individuals (in the FHS and JHS cohorts) might falsely predict a significant fraction as at higher risk for diabetes: 3 in 200 individuals carry a variant previously reported to cause MODY and yet exhibit no trend toward even late-onset T2D or impaired fasting glucose. Even objective bioinformatics criteria might incorrectly classify a significant, although three-fold lower, fraction of individuals as at-risk. The view that rare variants have deterministic effects, whereas common variants have modest effects, reflects in part the ascertainment bias of study designs used in Mendelian genetic research as well as the true penetrance of rare mutations.

Methods

Sample selection

To obtain individuals for the unselected cohorts, we drew from the Framingham Heart Study (FHS) Offspring cohort¹⁴ and from the Jackson Heart Study (JHS)¹⁵ cohort. As previously described, the FHS is a three generation prospective, community-based, family study which began in 1948 and was designed to identify factors that contributing to cardiovascular disease⁷⁴; the Offspring cohort consists of 5,124 of the adult children and spouses (enrolled in 1971) of the original participants¹⁴. The JHS is a large, community-based, observational study whose participants were recruited from urban and rural areas of the Jackson Mississippi metropolitan statistical area (MSA)¹⁵. These studies were performed using protocols approved by FHS, JHS, and institutional ethics committees, and with informed consent from all participants.

To select individuals for the T2D cohort, we drew from 27,500 individuals in three prospective cohorts: the Malmö Preventive Project⁷⁵ and Scania Diabetes Registry⁷⁶ (from Sweden), and the Botnia Study⁷⁷ (from Finland). Individuals were ranked according to a liability model that measured risk for T2D as previously described¹¹. Briefly, liability scores were computed as the difference between diabetes status and the predicted risk based on age, BMI, and gender; extreme cases were selected to have the highest liability scores (with diabetes but with low predicted risk for diabetes), and extreme controls were selected to have the lowest liability scores (without diabetes but with high predicted risk for diabetes). Individuals with age of diabetes diagnosis below 35 were excluded in an attempt to avoid consideration of patients with type 1 diabetes or MODY. The participants gave their written informed consent and the study protocol was approved by the Ethics Committees of Helsinki University Hospital, Finland, and Lund University.

Clinical and phenotypic parameters in unselected cohorts

Phenotype information for the unselected cohorts was contained in the NHLBI Framingham Cohort, dbGAP dataset: phs000007.v16.p6 (FHS) and the NHLBI Jackson Heart Study Candidate Gene Association Resource dbGAP dataset: phs000286.v2.p1 (JHS). We used phenotypic data from exam 5 for individuals from the FHS and phenotypic data from exam 1 for individuals from the JHS.

Individuals were classified as having diabetes if they were i) documented as such in the FHS or JHS exams, or ii) had FPG levels ≥126mg/dL. Individuals with FPG values 100-126mg/dL were said to display IFG. The euglycemic FPG range was consequently defined as 72–99mg/dL⁶⁰. As GCK-MODY patients exhibit only mildly-elevated FPG versus other MODY subtypes, the FPG range fitting this specific phenotype was defined as 99–144mg/dL^20,64. In the case of HNF1A variant carriers, aberrant fasting-2hr post-OGTT plasma glucose increments were defined as those ≥90mg/dL (this indicative of beta-cell dysfunction)^64,78. HNF1B variant carriers were assessed for signs of renal dysfunction by evaluating serum creatinine levels, the normal range defined as 0.7–1.3mg/dL⁶⁶.

Sequencing, quality control, and variant annotation

Although individuals in the T2D, FHS, and JHS cohorts spanned multiple cohorts, all analysis (target capture, sequencing, variant calling, quality control, annotation, and association analysis) was performed in an identical fashion using the same statistical pipeline. MODY patients were sequenced separately, but annotation and association was performed identically to the other three cohorts.

To sequence individuals in the T2D, FHS, and JHS cohorts, we designed two custom hybrid capture arrays, each using the same previously described technology^43,79, to sequence two sets of genes as part of two larger studies. Individuals from the FHS and JHS cohorts were sequenced for 181 genes previously associated with cardiovascular disease risk factors, including 37 genes associated with diabetes. Individuals from the T2D cohort were sequenced for 257 genes previously associated with diabetes or heart attack (either identified through genome-wide associations or reported to cause Mendelian disorders). These arrays had in common nine genes reported to harbor variants that cause MODY: the seven genes analyzed for this study as well as ABCC8 and KCNJ11.

DNA libraries were barcoded using the Illumina index read strategy and sequenced with an Illumina HiSeq2000. Reads were mapped to the human genome hg19 with the BWA algorithm^80,81 and processed with the Genome Analysis Toolkit (GATK) to recalibrate base quality-scores and perform local realignment around known indels⁸². Target coverage or each sample was also computed with the GATK. Single nucleotide polymorphisms (SNPs) and small insertions and deletions (indels) were called with the Unified Genotyper module of the GATK and filtered to remove SNPs with annotations indicative of technical artifacts (such as strand-bias, low variant call quality, or homopolymer runs)⁸². SNPs with differential call rates (p<1e-3 as computed by the PLINK software package⁸³) were excluded from association analysis. Variant calls from sequence data were deposited in dbGAP (FHS: dataset phs000307.v4.p7; JHS: dataset phs000498.v2.p1).

Samples were also genotyped on one of three genome-wide SNP arrays: the Affymetrix 500k GeneChip Mapping Set (FHS cohort), the Affymetrix Genome-Wide Human SNP Array 6.0 (JHS cohort), and the Metabochip (T2D cohort). We computed concordance of sequence genotypes with these SNP array genotypes using the PLINK software package⁸³. Principal component analysis was performed on a set of SNPs common to all three platforms using PLINK and EIGENSTRAT⁸⁴. These analyses verified that all individuals were unrelated (<25% of their genomes identical by descent) and confirmed the distinct genetic background of each cohort (Supplementary Fig. 3). Quality control showed high (>96%) concordance between sequence and SNP array genotypes for the same individuals.

To ensure that all analyzed variants had genotypes strongly supported by sequence data, we used three strategies. First, we ignored all genotypes supported by fewer than 10 reads (e.g. set as “missing”). Second, we examined the raw read data of every called variant and excluded any from analysis that had visual signatures of sequencing artifacts, such as reads of poor mapping quality, evidence for variation supported by only reads on one strand of the genome, or additional called variants nearby; screenshots of all variants are available as Supplementary Data 1. Third, to evaluate the extent to which erroneous sequence genotypes might impact our analysis, we genotyped 143 SNPs called in the T2D cohort (the first cohort sequenced and with the lowest depth of coverage) using a Sequenom iPLEX Assay. This included nine SNPs observed in only a single individual (singletons) and 27 SNPs observed in two individuals (doubletons). Of the 1573 individuals identified as carrying one of these 143 SNPs based on sequence genotypes, only nine had different Sequenom genotypes (>99.4% non-reference concordance). All (100% of) the 63 identified carriers of singleton or doubleton variants were confirmed based on Sequenom genotypes. These analyses collectively suggested that the analyzed genotypes were of high quality and that any errors had at most a minimal impact on our results.

We annotated variants previously reported to cause MODY using a list of mutations given in the Human Gene Mutation Database Professional v2012.1⁴⁹. Evolutionarily conserved variants were defined as those at sites conserved across 46 vertebrates, based on PhyloP scores (LOD >3) downloaded from the UCSC genome browser database⁸⁵. Rare variants were defined as those observed in only a single study individual across the T2D, FHS, and JHS cohorts, and furthermore were not identified as part of the 1000 genomes project⁴⁶. Predicted protein-damaging variants were defined as those either (a) deleterious according to SIFT⁴⁷ and possibly or probably damaging according to Polyphen-2⁴⁸, or (b) nonsense, frame-shift, or essential splice site mutations according to the Variant Effect Predictor⁶⁷. Protein and nucleotide changes, and SIFT and Polyphen-2 scores, were produced for each variant with the Variant Effect Predictor, using the biologically-relevant transcripts for each gene (as listed in the HGMD and shown in the manuscript ‘Accession Numbers’ section).

For analyses that required variants below a given frequency, variants that exceeded the threshold in any of the cohorts were removed from analysis.

Analysis of MODY patients

As a technical validation of the annotation protocol used for the population-based cohorts, variants in 250 subjects with MODY were analyzed using the same methodology. As previously described^{21,36,86–93}, subjects with MODY were recruited from the Norwegian MODY Registry, a national, population-based registry established in 1997. Patients are referred to the registry if they fit the following criteria: i) a first-degree relative with diabetes; ii) onset of diabetes before age 25 years in at least one family member; iii) a low dose insulin requirement; iv) unusual type 1 diabetes (insulin requirement below 0.5 U/kg/day, no antibodies, atypical history). The four most common MODY genes are routinely sequenced in the first instance: HNF1A, HNF4A, GCK and INS. If fasting glucose is >100 mg/dL and HbA1c is less than 7.5 %, GCK is investigated first followed by the other genes from the list above. If renal cysts or renal failure presents before diabetes, HNF1B is investigated. NEUROD1 and PDX1 were investigated as part of a screening program and are not routinely tested. Informed consent was obtained from all participants and the study was approved by the Regional Committee for Research Ethics and the Norwegian Data Inspectorate.

To screen patients, genomic DNA was first extracted from peripheral leukocytes using standard procedures. The coding exons and intron/exon boundaries of HNF1A, HNF4A, GCK, HNF1B, PDX1, NEUROD1 and/or INS were then PCR-amplified and sequenced using an Applied Biosystems 3730 capillary sequencer (Applied Biosystems, Foster City, CA). We imported the samples into the SeqScape Software (Applied Biosystems) and analysed for sequence variations by comparing with published reference sequences (NM_000545.5, NM_175914.3, NM_000162.3, NM_000458.2, NM_000209.3, NM_002500.3 and NM_000207.2, respectively). Variants were annotated and classified using the same procedure as for the three population-based cohorts.

Assessment of MODY-relevant clinical criteria in unselected cohorts

We examined variant carriers in the FHS and JHS cohorts to see if any fit the following phenotypic criteria associated with MODY: diabetes diagnosis ≤25 years (for proband or family members) lean (BMI<25), and family history of diabetes (at least two generations; typically early-onset). We did not have access to age of diabetes diagnosis for all family members and subsequently could only examine diagnosis age of variant carriers.

In the FHS cohort, no variant carriers simultaneously fit criteria for early diabetes diagnosis and low BMI (4/116 variant carriers had diabetes with BMI below 25, but each had age of onset above 55, and only one variant carrier had BMI below 30 and age of onset below 40). The frequency of carriers fitting these criteria was thus comparable to that for non-carriers: one of the 1349 non-carriers had BMI below 25 and age of onset below 30, while eight had BMI below 30 and age of onset below 40.

In the JHS cohort, 1/235 variant carriers had a family history of diabetes, age of onset 25, and BMI of 29.8; the rest had BMI above 30 or age of onset above 42. Similar to the FHS cohort, the frequency of carriers fitting these criteria was similar to that for non-carriers: two of the 1405 non-carriers had a family history of diabetes, age of onset below 25, and BMI below 25, with four more having a family history, age of onset below 40, and BMI below 30.

Statistical analysis

We tested for association between phenotype (whether extreme T2D status, diabetes, IFG, or a gene-specific phenotype) and variant carrier status using the Cochran-Mantel-Haenszel test as implemented in the PLINK software package⁸³. Randomized permutation of phenotypes was used to obtain one-sided p-values. For the T2D cohort, stratified permutation of phenotypes (separately for individuals of Finnish and Swedish descent) was used to control for ethnic differences across individuals; due to the small number of variant counts in this cohort, further investigation of population stratification was statistically challenging. For the unselected cohorts, tests were first run separately for individuals from each of the FHS and JHS cohorts and then jointly for individuals from both cohorts (with stratified permutation of phenotypes within cohort) to obtain combined p-values in all figures or tables (including Figure 2). For all figures, confidence intervals in were obtained via the Clopper-Pearson method as implemented in the R software package.