Study population

Subjects for the discovery GWAS were recruited from an Australian Aboriginal community of Martu ancestry [17, 18] at the edge of the Western Desert in Western Australia. A family-based study was proposed to take account of recent population history and shared ancestry. A memorandum of understanding (MoU) was established with the community, which included permission for access to clinical records held in a Communicare database at the local Aboriginal Health Service. An individual was classified as having T2D if the subject was: (1) diagnosed with T2D by a qualified physician; (2) on a prescribed drug treatment regimen for T2D; and (3) returned biochemical test results of a fasting plasma glucose level of at least 7 mmol/l in SI units based on criteria laid down by the World Health Organization (WHO) consultation group report [19]. Multiple height and weight measures per individual were recorded in Communicare through time and converted within the database to BMI according to the standard formula BMI = weight (kg)/height (m²). DNA was prepared from saliva samples collected into Oragene tubes (DNA Genotek, Ontario, Canada) from 405 consenting family members who were available at the time of visits by the study team during the two-year collection period of the study.

Ethical approvals

Ethical approval for the study was obtained from the Western Australian Aboriginal Health Ethics Committee (WAAHEC; Reference 227 12/12), who reviewed and approved forms for informed consent. Each individual (or the parent or guardian of individuals less than 18 years of age) signed separate informed consent forms to participate in the study and to provide a DNA sample. Following feedback of results to the community, permission to publish was provided by the Board of the local Aboriginal Health Service, which comprised elders representing the extended families residing in the area. Permission to lodge de-indentified genotype and basic demographic data (broad geographical location, age, sex and phenotype information) in the European Genome-phenome Archive (accession number EGAS00001001004) was also obtained from the Board of the local Aboriginal Health Service.

Genotyping, quality control, and analysis of population structure

DNA from the 405 consenting individuals were genotyped on the Illumina Omni2.5 BeadChip (outsourced to the Centre for Applied Genomics, Toronto, Ontario, Canada). Quality control (QC) data from the service provider indicated a call rate of (mean [SD]) 99.680% [0.005%]. Further in-house QC procedures at the level of individuals resulted in one individual being dropped due to a missing data rate >5%, 2 exclusions due to unintentional duplication, and no exclusions due to outlying heterozygosity (using 3 standard deviation limits), discordant sex, or divergent ancestry (since mixed models were employed to take account of ancestry, cf. below). This provided a post-QC dataset of 402 family members, including 361 with repeated body mass index measurements and 391 (89 cases, 302 family members unaffected at the time of collection) with information on doctor diagnosed T2D as per criteria outlined above. SNPs with minor allele frequency (MAF) <0.05, or with more than 5% missing data, were removed prior to association analysis (cf. below). As family data were employed in our study, complicated by a degree of over-relatedness in the pedigrees, a check of Hardy Weinberg Equilibrium (HWE) was not used as a SNP QC check since it was not clear that we should expect all SNPs to be in HWE. A subset of 70,420 genotyped SNPs with pairwise linkage disequilibrium (LD; r²) ≤0.3 and MAF >0.01 was used in principal component analysis (PCA; SMARTPCA within EIGENSOFT [20, 21]) to look at population substructure across the 402 family members.

Imputation procedures

Pre-phasing of genotyped data was performed using SHAPEIT [22], and imputation conducted using IMPUTE2 [23] with 1000 Genomes (1000G) haplotypes [Phase I integrated variant set release (v3)]. The reference panel includes haplotypes of 1,092 individuals from Africa, Asia, Europe and the Americas. For association analysis, probabilistic imputed SNP genotypes were converted to hard genotype calls provided the posterior probability of the most likely genotype was >0.9, and imputed SNPs were retained only if they had `info’ score >0.5, MAF>0.01 and <10% missing data. To assess imputation accuracy, IMPUTE2 provides measures of concordance and squared Pearson correlation (r²) for each genotyped SNP. Directly genotyped SNPs were masked then imputed and concordance between the true genotype and the most likely imputed genotype [where probability prediction threshold > 0.9] for each SNP was calculated [24–26], as well as r² between the true (discrete) allele dosage (coded as 0, 1 and 2 corresponding to the number of minor alleles) and the imputed (continuous) allelic dosages. For low frequency SNPs, r² is the preferred accuracy measure because, unlike concordance, allele frequency is taken into account. We also compared these imputation accuracies for the 195 individuals determined by PCA to represent pure Martu ancestry against the 207 deemed to be of mixed ethnicity by PCA, using the data generated from imputation of the full dataset for the 402 individuals.

Heritability and association analyses

Heritability for mean BMI was determined from self-reported pedigree structures using the QTDT software package [27], and separately on the basis of kinship and IBD sharing from SNP-chip data calculated using the R package GenABEL v1.7–6 [28] and using the genome-wide complex trait association (GCTA) tool [29]. To provide a visual display of relatedness across the 402 genotyped individuals, the genomic kinship matrix was first calculated in GenABEL v1.7–6 [28], and then converted to a distance matrix followed by hierarchical cluster analysis using single linkage on the dissimilarities. The ape package [30] was used to produce a radial tree plot for the 402 genotyped individuals. For association analyses, the FAmily-based Score Test for Association (FASTA) [31] in GenABELv1.7–6 [28] was employed that uses a linear mixed model (LMM) approximation to model the trait outcome, with whole genome data used to estimate kinship (in order to account for relatedness) and to take account of population substructure. Analyses were carried out using quantitative trait data (BMI), and separately to compare all T2D cases with all unaffected individuals. Association results for genotyped data obtained using GenABEL were compared with results using FaST-LMM [32]. Both GenABEL (FASTA) and FaST-LMM model disease status (control/case for T2D, coded 0/1) as if it were a normally distributed quantitative variable, which has been shown [33] to produce a valid test with respect to testing the null hypothesis of no association. Power calculations using QUANTO [34] show that 361 individuals with BMI measures provide a maximum of 80% power to detect associations at genome wide significance levels (P<5×10^-8), or a maximum of 97% power at suggestive significance levels (P<1×10^-5), for SNPs with MAF>0.3 conferring allelic effects (betas) of magnitude half a unit of standard deviation. For effects (betas) of magnitude 0.25 units of standard deviation, the maximum powers achievable are lowered to 1% and 9% respectively. The 89 T2D cases and 302 unaffected family members provide a maximum of 32% power to detect associations at genome wide significance levels (P<5×10^-8), or 71% power at suggestive significance levels (P<1×10^-5), for SNPs with MAF>0.3 conferring allelic odds ratios >2.5. For odds ratios of 1.5, the powers are lowered to 1.4% and 0.06% respectively.

Manhattan plots were generated using the mhtplot() function of ‘gap’, a genetic analysis package for use in R (see URLs). Quantile-quantile (Q-Q) plots were generated, and inflation factors (denoted λ) calculated in R version 2.15.0 by dividing the median of the observed chi-squared statistics by the median of the theoretical chi-squared distribution. Regional plots of association were created using LocusZoom [35] in which—log₁₀ P-values were graphed against their chromosomal location. Pairwise LD patterns between all regional SNPs and the top SNP were calculated specifically for this Australian Aboriginal study data using 146 unrelated individuals from the total sample of 402 genotyped individuals. The 146 unrelated individuals were selected by iterative removal of the individuals with the greatest number of estimated relationships with IBD>0.1875.

Bioinformatic analysis

Global alignment of genomic sequence for the region SLC28A3 to NTRK2 from human, cow and mouse was undertaken to locate evolutionarily conserved non-coding sequences (CNS) that might contain regulatory elements and transcription factor binding sites (TFBS). Genomic sequences for the three organisms were exported in FASTA format from ENSEMBL (Genome Reference Consortium Release 37, Ensembl Release 67) and associated annotation exported in the form of a General Feature File (GFF) file. The global alignment tool Multi-LAGAN [36, 37] was used to align genomic sequences using the guide tree (((human) cow) mouse). SYNPLOT [38] was used to visualize the annotated alignment. CNS, defined here as regions with a nucleotide sequence conservation level of ≥0.7 (i.e. ≥ to the least conserved exon sequence in the genes flanking the region of interest), were analysed for promoter and enhancer elements using PROMO [39], AliBaba v2.1 [40], and MatInspector v8.0.5 [41] with a matrix similarity parameter >0.75. We also used the UCSC genome-browser (see URLs) with custom tracks to assess where selected top-hit SNPs are located in relation to elements such as CpG islands or repeat elements (LINES; SINES). To look de novo for CpG island-like elements, we firstly masked repeat regions using RepeatMasker (see URLs) following which putative CpG islands were searched for using CpGIsland Searcher [42] with parameters set at 50%GC, 0.60 observed CpG/expected CpG ratio, and length 200bp (rather than the default settings of 55%, 0.65, and 500bp).

Characteristics of the study population

The 402 post-QC genotyped individuals used in the GWAS belonged to a small number of inter-related extended pedigrees, as depicted in the radial plot which shows hierarchical clustering of estimated pairwise identity-by-descent allele-sharing (S1 Fig.). Principal component analysis (S2 Fig.) demonstrated a degree of introgression of predominantly Caucasian origin, with a tight cluster of 195 individuals of Martu Aboriginal ancestry across all age groups. Linear mixed models were used in the genetic analysis to take account of both family relationships and this genetic substructure. S1 Table provides basic demographic data (age, sex; at the time of diagnosis and/or collection) for the 391 individuals used in the genetic analyses. De-identified BMI data was also available for 1020 self-reporting Aboriginal individuals in the population-specific Communicare database.

BMI in the study population

A BMI >22kg/m² is a significant risk factor for T2D in Australian Aboriginal populations [16]. Fig. 1 part A provides a plot of all BMI measurements for self-reported Aboriginal individuals in the study population (N = 1020). BMI by age was plotted using the R package SITAR [43] for all records in Communicare. Separate lines trace multiple measurements over time per individual; females (pink lines) and males (blue lines). The heavy lines show the polynomial quintic (power of 5) curves for females (pink heavy line) and males (blue heavy line) that best fit the data. Fitting separate (by gender) curves to the data (Fig. 1 part A) provided a significantly better fit (P = 10^-9) than fitting a single curve with a sex-specific displacement. We constructed standardized residuals from these curves by subtracting the observed BMI value from the fitted curve and dividing by the estimated standard deviation (calculated within 1-year age brackets). All GWAS analyses of BMI (cf. below) were carried out using these standardized residuals. The extreme female outlier was not used in fitting the female polynomial curve or in any of the GWAS analyses. This analysis of BMI shows (a) that the majority of the adult population have BMI >22 and are therefore at increased risk of diabetes; (b) a proportion of 15–20 year olds also fall into this category; and (c) there are many individuals in the community who can be classified as overweight (BMI 25.00–29.99) or obese (including class I obesity BMI 30.00–34.99; severe or class II obesity BMI 35–39.99; and morbid or class III obesity BMI≥40) according to Center for Disease Control (CDC) [44] and World Health Organization (WHO) [45] international criteria. By estimating familial correlations based on self-reported pedigree structures using QTDT [27], we estimated heritability for mean BMI in this population to be 55%, broadly in line with other populations internationally [46–48]. Estimation of heritability on the basis of kinship and IBD sharing from the SNP-chip data using the R package GenABEL [28] gave a lower estimate of 38%, consistent with the estimate of 39% (95% confidence interval [17%-61%]) provided by GCTA [29].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. BMI by age plotted using the R package SITAR.

(A) plot of all records for self—reporting Aboriginals (N = 1020) in the Aboriginal Health Service’s Communicare database; and (B) plot of all records for the 391 genotyped individuals contributing to association analyses for BMI and T2D. Separate lines trace multiple measurements over time per individual; females (pink) and males (blue). The heavy lines in (A) show the polynomial quintic (power of 5) curves for females (pink heavy line) and males (blue heavy line) that best fit the data. Fitting separate (by gender) curves to the data provided a significantly better fit (P = 10^-9) than fitting a single curve. The extreme outlier was not used in fitting the female polynomial curve. In (B), individuals with T2D are shown in heavy lines.

https://doi.org/10.1371/journal.pone.0119333.g001

T2D in the study population

Our dataset for the study population (Fig. 1 part B) showed that >75% of adults (≥20 years of age) fall above the BMI cut-off of >22 for increased risk of T2D in Aboriginal Australians [16]. Fig. 1 part B also highlights BMI curves for individuals in the study sample with T2D (heavy pink lines for females; heavy blue lines for males). This demonstrates that our study population is characterized by T2D predominantly associated with high risk BMI measurements, including a number of cases <20 years of age. S2 Fig. part D highlights on a PCA plot the individuals with T2D used in the GWAS. Of the 391 individuals used in the T2D GWAS (cf. below), there were 65 with T2D from 191 individuals (34%) that fell within the pure Martu ancestry cluster (S2 Fig. parts E and G), and 24 with T2D from 200 individuals (12%) of mixed ethnicity. There was insufficient power to analyse GWAS data separately for T2D associated with low to normal (<25; N = 10) BMI or with low (<20 years of age; N = 5) age at onset. There will be some loss of power in the analysis with the inclusion of unaffected individuals less than 20 years of age (since ~10–20% of them may go on to get T2D[16]). This had to be balanced against the gain in power with the increased sample size and greater contribution of individuals in controlling for population substructure and relatedness in the analysis. S1 Table shows the age range and BMI for unaffected individuals <20 and >20 years of age separately. Importantly, there will be no false positive associations generated by including all unaffected individuals in the GWAS analysis for T2D (cf. below).

GWAS analyses for BMI based on genotyped SNPs

Longitudinal BMI data available for the 361 post-QC individuals with both BMI measurements and genotyped SNP-chip data were representative of the full range of BMI values in the population (Fig. 1 part B). To identify loci associated with BMI we undertook a GWAS using 1,075,436 post quality control genotyped SNPs in these 361 individuals. For longitudinal BMI data (designated BMI-longitudinal), the analyses in GenABEL and FaST-LMM used each individual BMI reading as a separate observation, using the standardized residual from the fitted curve as the trait of interest and modelling the correlation between readings via the estimated kinship. This does not fully account for the correlation between readings as readings from the same person are likely to be more correlated, especially since they are often measured at frequent (close together) time points. The Genomic Control deflation factor [49] was therefore employed in GenABEL to avoid inflation of the overall distribution of test statistics; this correction was not found to be necessary in FaST-LMM, as previously noted [50]. Association analyses were also undertaken in GenABEL and FaST-LMM using the mean of all BMI readings for an individual (designated BMI-mean) as the trait. Although this raises issues of differential variation in BMI-mean due to uneven numbers of readings between individuals, results from these two different analyses were well correlated. The plots presented at S3 Fig. demonstrate that association results for BMI-mean and BMI-longitudinal obtained using FASTA in GenABEL (S3 Fig. part A) or using FaST-LMM (S3 Fig. part B) were highly correlated. Therefore detailed results are given only for BMI-longitudinal (hereinafter referred to as BMI). Similarly, results obtained for each phenotype from GenABEL and FaST-LMM were strongly concordant (S4 Fig.), consistent with our recent evaluation of a range of software implementations for linear mixed models [50]. Q-Q plots (S5 Fig.) for GenABEL (λ BMI-longitudinal = 1.0; λ BMI-mean = 1.02) and FaST-LMM (λ BMI-longitudinal = 1.00; λ BMI-mean = 1.00) analyses also showed equivalent inflation factors. Therefore detailed results presented hereafter are given only for analyses undertaken using FASTA in GenABEL.

A Manhattan plot showing the genome-wide results for BMI based on the genotyped data is presented in Fig. 2 part A. The top hits did not achieve genome-wide significance, commonly accepted as P<5×10^-8 [51] and concordant with the number of post-QC SNPs (P = 0.05/1,075,436 or 4.65×10^-8) used for this analysis of genotyped data. No specific SNPs reported to achieve genome-wide significance for association with BMI in other populations achieved P<10^-2 in our study, as determined by interrogation of the NIH NHGRI Catalogue of GWAS studies [52]. Nevertheless, since we cannot assume similar patterns of linkage disequilibrium or directionality of associations in our Australian Aboriginal population, we have provided information in S2 Table on SNPs with P<10^-2 in our study population for the genes previously reported to achieve genome-wide significance (i.e. P<5×10^-8) for association with BMI in other populations. Notably (cf. below), MC4R (top SNP rs129959775; P_genotyped = 4.49×10^-4) was present in this list. Extending the table to include genes previously reported in GWAS for BMI at P<10^-5 in other populations provided evidence for association at CNTNAP2 (top SNP rs6960319; P_genotyped = 4.65×10^-5). SNPs at CNTNAP2 were also among the top 50 SNPs from the BMI analysis in our study population (S3 Table). Data for CNTNAP2 and other top SNPs in genes of potential functional relevance, i.e. in pathways previously identified to be associated with metabolic diseases, are shown in Table 1. Effect sizes (betas 0.35 to 0.72) for these associations are concordant with our power to detect allelic effects of magnitude half a unit of standard deviation of the trait. The top hit (rs10868204; P_genotyped = 2.73×10^-6) for BMI in our study population lies in the intergenic region (Fig. 3; cf. below) 5’ of NTRK2 encoding the type 2 neurotrophic tyrosine kinase receptor for brain-derived neurotrophic factor (BDNF) that regulates energy balance downstream of melanocortin-4 receptor (MC4R). Both BDNF and MC4R have previously been shown to be associated with obesity in other populations (see S2 Table). Other top hits in our population that are of functional interest (Table 1) occurred in RBM7 (rs6848632 P_genotyped = 1.43×10^-5) which has been related to pancreatic function [53], and at PIK3C2G (rs12816270 P_genotyped = 8.06×10^-6) which has previously been associated with T2D [54].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Manhattan plots of genome-wide association results for BMI undertaken using FASTA in GenABEL.

(A) results for genotyped data; and (B) results for 1000G imputed data. SNPs in red show the region of the top association in this discovery GWAS.

https://doi.org/10.1371/journal.pone.0119333.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Regional association plots (LocusZoom [35]) of the signal for BMI association in the region SLC28A3 to NTRK2 on chromosome 9.

(A) is the plot for genotyped data; and (B) is the plot for 1000G imputed data. The −log₁₀ P-values are shown on the upper part of each plot. SNPs are colored (see key) based on their r² with the labeled hit SNP (purple), calculated in the 146 unrelated genotyped individuals. The bottom section of each plot shows the genes marked as horizontal lines. The second Y axis is for recombination rate, as shown in blue on the plot.

https://doi.org/10.1371/journal.pone.0119333.g003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Top GWAS SNP hits in genes of functional relevance for BMI, organized by chromosome.

https://doi.org/10.1371/journal.pone.0119333.t001

GWAS analyses for T2D based on genotyped SNPs

A Manhattan plot showing the GWAS results for T2D for genotyped SNPs is presented in S6 Fig. part A. Again, the top hits based on analysis of genotyped SNPs did not achieve genome-wide significance for T2D, which was less well-powered than the BMI analysis. In addition, no specific SNPs reported to achieve genome-wide significance for association with T2D in other populations achieved P<10^-2 in our study, as determined by interrogation of the NIH NHGRI Catalogue of GWAS studies [52]. Nevertheless, as for BMI we cannot assume similar patterns of linkage disequilibrium or directionality of associations in our Australian Aboriginal population. Therefore, we have provided information in S4 Table for genotyped SNPs with P<10^-2 in our study population for the genes previously reported to achieve genome-wide significance (i.e. P<5×10^-8) for association with T2D in other populations. Five genes (ANK1, TSPAN8, PROX1, GLIS3 and UBE2E2) had genotyped SNPs with P<10^-3. Of note, no SNPs at P<10^-2 were observed for TCF7L2 [55–59] or KCNJ11 [60], genes that have previously been associated with T2D across multiple ethnicities. However, 4 genes (Table 2) represented in the top 50 SNPs (S5 Table) are supported by strong biological candidacy in related gene pathways. The top hit (rs11240074 P_genotyped = 5.59×10^-6) for T2D in our study population lies 5’ of BCL9 encoding a protein that, along with TCF7L2, promotes beta-catenin’s transcriptional activity in the WNT signaling pathway. Additional hits (1.07×10^-4≤ P_genotyped ≤4.55×10^-5) of functional interest occurred in genes involved in pancreatic (KCNJ6 [61], KCNA1 [62]) and/or GABA (GABRR1 [63], KCNA1 [64]) functions.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Top GWAS SNP hits in genes of functional interest for T2D, organized by chromosome.

https://doi.org/10.1371/journal.pone.0119333.t002

GWAS analyses based on imputed SNP data

Recent reports have focused on improving tools for imputation of SNPs based on both the 1000 genomes (1000G) and HapMap project data [23, 65]. It was of particular interest to determine how well 1000G imputation would work for this Australian Aboriginal population, given that most estimates suggest the arrival of Aboriginal people in Australia more than 45,000 years ago [66, 67]. We therefore examined the efficiency of imputing genotypes across the genomes of our study population using 1000G data. High average concordance (92.8%-97.9%) was observed across all chromosomes independently of MAF. Fig. 4 compares the efficiency of imputation across all chromosomes, as measured by imputation r², for SNPs at different MAF. As expected, r² is lower for SNPs with MAF 0.01–0.03 (mean r² 77.2–85.4) compared to SNPs with MAF 0.03–0.05 (mean r² 81.5–88.5%) and MAF 0.05–0.5 (mean r² 85.9–90.7%). When examined across individuals, mean imputation accuracy for the 195 individuals of pure Martu ancestry (concordance 94.8; r² 88.8) were not different to measures for the 207 individuals of mixed ethnicity (concordance 95.6; r² 90.0) or to measures across all 402 individuals (concordance 95.2; r² 89.5).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Imputation accuracy for 402 genotyped individuals imputed against the 1000G reference panel.

Imputation accuracy is measured as average r² across all autosomes for SNPs of different MAFs (see key).

https://doi.org/10.1371/journal.pone.0119333.g004

In relation to the GWAS analyses for BMI (Fig. 2 part B; Fig. 3 part B; Table 1; S6 Table), 1000G imputation improved significance for associations at NTRK2 (Fig. 3 part B; Table 1: P_genotyped = 1.50×10^-6; P_{imputed_1000G} = 2.90×10^-7) and at PIK3C2G (Table 1: P_genotyped = 8.06×10^-6; P_{imputed_1000G} = 6.28×10^-7) but not at other genes of functional interest (Table 1). Similarly, for the T2D imputed GWAS analysis (S6 Fig. part B; S7 Table), top SNPs for 1000G imputed data in genes of functional interest (Table 2) were generally of the same order of magnitude as top genotyped SNPs, including support for the top hit at BCL9. Novel findings in the top 100 SNPs for 1000G imputed data included a hit at the previously identified gene for T2D IGF2BP2 (S4 Table: rs138306797, P_{imputed_1000G} = 2.55×10^-6) that had not achieved P<10^-2 in the genotyped data, and hits at P_{imputed_1000G} <10^-6 in genes not previously associated genetically or functionally with T2D (S7 Table: SATB2/TYW5 on chromosome 2, MTHFD1L/AKAP12 on chromosome 6; KLF5/KLF12 on chromosome 13).

Genes for obesity or for T2D?

Analysis of T2D and the standardized BMI residual (BMI-longitudinal) was initially performed without any covariate adjustment. Since T2D and BMI are highly correlated and have been shown to have shared genetic influences [68], the question arises as to which phenotype is primarily regulated by the genes of interest identified in this study. We therefore repeated the analyses allowing for BMI and T2D as covariates (S1 Text). Results demonstrated that (A)—log₁₀P-values for T2D adjusted for the standardized BMI residual were highly correlated with the original unadjusted analysis for both genotyped and imputed data, and (B) conversely,—log₁₀P-values for unadjusted BMI and BMI adjusted for T2D were likewise highly correlated. This suggests that genetic effects on T2D as originally calculated are largely independent of any effects attributable to the mean standardized BMI residual. Specifically, the significance levels for genes of functional interest are of the same order of magnitude in both the adjusted (S1 Text) and the original analysis (Table 2). Similarly, genetic effects on BMI as originally calculated are largely independent of any effects attributable to T2D, consistent with the regional association plots (Fig. 3, S7 Fig.) where no evidence for association is seen in plots comparing T2D results across regions that contained the top hits for BMI. This includes the regions containing RMB7, which previously has been related to pancreatic function [53], and PIK3C2G which has been associated with T2D in another population [54].

Interrogating the SLC28A3 to NTRK2 intergenic region

As with most GWAS for common complex diseases [69], top SNP hits identified here were generally located within introns or in potential regulatory regions upstream or downstream of the gene of interest. In the case of the top hit for BMI (Fig. 3), the peak of association was clearly located upstream of the best functional candidate for BMI, the NTRK2 gene, with little evidence for linkage disequilibrium between SNPs in this peak of association and SNPs within the coding region of the gene. Conditioning on the top genotyped and top imputed SNPs (S8 Fig.) reduced all other SNP signals to P>10^-3, suggesting a single major signal regulating BMI across this region. We therefore interrogated the region of top hits upstream of NTRK2 to see if we could find evidence for potential regulatory elements. Initially we used SYNPLOT to plot the location of the top genotyped SNPs (P<5×10^-6) in relation to CNS upstream of NTRK2 (Fig. 5). These 8 genotyped SNPs were all in strong LD with the top SNP (r²>0.8; Fig. 3). SNP rs1866439 (P_genotyped = 1.59×10^-6) was of equivalent significance to the top SNP rs1347857 (P_genotyped = 1.5×10^-6) and was unique in being the only top SNP located within a CNS peak, with the risk allele causing loss of a NHP1 binding site falling within a half-palindromic estrogen response element (TGAGTagtTG^A/_GCC) [70, 71]. NTRK2 expression has been shown to be regulated by estradiol [72]. Other indications that this might be a regulatory region, as annotated on Fig. 5, included: (i) the presence adjacent to rs1866439 of a CpG-island (50%GC; observed CpG/expected CpG 61%; 212bp in length); (ii) a number of peaks of mono-methylation of lysine 4 of histone H3 (H3K4Me1; frequently found near regulatory regions) as measured by the ENCODE project (see URLs), including around rs1866439; and (iii) the peak of association with the top imputed SNPs (Fig. 3 part B) falling within a non-conserved region (no CNSs) of fragments of long (LINE-1 or L1) interspersed elements or retrotransposons (Fig. 5 and S9 Fig.). S9 Fig. shows the positions of these top SNPs in relation to the L1 elements, with the top imputed SNP (rs11140653) specifically falling within a L1MA3 element (as also annotated on Fig. 5). While full-length L1s are required for retrotransposition [73], regulatory elements that overlap with fragments of L1 sequence have been shown to be involved in gene silencing [74]. Further functional studies will be required to determine which variants and elements are important in regulating NTRK2 in this Aboriginal population.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. SYNPLOT [38] for a section of the intergenic region (NCBI Build 37: 87,140,000bp to 87,190,000bp) ~94kb upstream of NTRK2 on chromosome 9q21.33.

Since regulatory elements are usually found with conserved regions of the genome, we interrogated the region of our top hits using this in silico analysis to look for conserved regions across multiple species. The multiple alignments of (top to bottom) human, cow and mouse sequences were performed across the complete intergenic region SLC28A3 to NTRK2 using LAGAN. The segment of the alignments shown here is the region that contains the top association hits (P<5×10^-6). The central plotted curves show the degree of conservation of sequence across all three species, on a vertical scale 0–1 (= 100%), such that the peaks represent CNS. CNS are defined here as regions with a nucleotide sequence conservation level of ≥0.7, i.e. ≥ to the least conserved exon sequence in the two genes (not shown on this plot) SLC28A3 and NTRK2 flanking the intergenic region of interest. Blue boxes indicate repetitive sequence in all 3 species. The human sequence is also annotated with positions of: (i) LINE-1 elements (yellow); (ii) the top genotyped SNPs (red vertical bars) including the top SNP rs1086204, and the SNP of interest rs1866439; (iii) the top imputed SNPs (blue vertical bars) including the top SNP rs1140653; (iv) the CpG island-like element identified using CpG Island Seacher (green); and (v) positions of peaks of mono-methylation of lysine 4 of histone H3 (H3K4Me1) as measured in NHEK or NHLF cell lines (mauve) or in H1-hESC human embryonic stem cells (orange) by the ENCODE project, as identified using the UCSC browser (see S9 Fig.).

https://doi.org/10.1371/journal.pone.0119333.g005