Farp2 and Stk25 Are Candidate Genes for the HDL Cholesterol Locus on Mouse Chromosome 1
Objective— To identify the gene responsible for the quantitative trait locus (QTL) Hdlq14, a high-density lipoprotein cholesterol (HDL) QTL previously identified in a C57BL/6J×129S1/SvImJ cross.
Methods and Results— Hdlq14 was first confirmed as an independent QTL by detecting it in an intercross between NZB/B1NJ and NZW/LacJ, 2 strains that had identical genotypes at nearby QTL genes on chromosome 1. Using the bioinformatics tools of combined cross data and haplotype analysis, we narrowed this QTL from a 45-Mb 225-gene region to 2 genes, Farp2 and Stk25. Sequencing and expression studies showed that Farp2 had an amino acid polymorphism in an important plekstrin domain and that Stk25 had a significant expression difference between the parental strains. These 2 genes are immediately adjacent to each other and share the same haplotype over 45 inbred strains. The haplotype was associated with a significant difference in HDL levels among these strains.
Conclusion— We confirmed Hdlq14 as a separate independent QTL for HDL and narrowed the region to 2 genes, Farp2 and Stk25, with considerable evidence for both. Additional studies are needed to choose between these 2 genes or to show that both are important in determining HDL levels.
Population studies have consistently shown that high-density lipoprotein (HDL) cholesterol levels are a strong independent inverse predictor of cardiovascular disease (CVD).1 This inverse relationship between HDL levels and CVD has generated interest in the genetic factors contributing to variations in HDL. Quantitative trait locus (QTL) analysis is a means of finding novel genes that regulate complex traits, and QTL detected in mouse models of disease often predict the location of human disease QTL, suggesting that the same genes regulate these traits in both human and mouse.
HDL levels vary among common strains of inbred mice, and numerous QTL for this trait have been mapped.2 The most commonly detected QTL is a broad one on distal chromosome (Chr) 1 found in at least 15 crosses (supplemental Figure I, available online at http://atvb.ahajournals.org).2 In one of the crosses, C57BL/6 (B6)×129S1/SvImJ (129) this broad peak contained at least 2 closely linked QTLs, which we named Hdlq14 and Hdlq15.3 Hdlq15 is caused by Apoa2 located at 173.1 Mb.4 However, the QTL gene for Hdlq14 was difficult to identify, particularly using data from existing crosses that contained Apoa2 polymorphisms, because these made it difficult to identify a nearby QTL gene.
In the present study, we carried out a cross between 2 strains, NZW/LacJ (NZW) and NZB/BINJ (NZB), which could not have a QTL at Hdlq15 because they had identical Apoa2 gene sequences. The cross between these 2 strains confirmed the existence of Hdlq14, and we were able to identify the most likely candidate genes as Farp2 and Stk25 using a variety of bioinformatic tools.
Materials and Methods
Animals and Diet
The (NZB×NZW)F1 were obtained from The Jackson Laboratory (Bar Harbor, Me) and mated to produce 272 male and female F2 progeny. Mice were maintained in a temperature- and humidity-controlled environment with a 14-hour light/10-hour dark cycle and given unrestricted access to food and acidified water. Weanling mice were fed standard chow containing 6% fat (LabDiets, 5K52) until 8 weeks old and then fed an atherogenic diet containing (w/w) 15% dairy fat, 1% cholesterol, and 0.5% cholic acid as described previously.5 Experiments were reviewed and approved by the Institutional Animal Care and Use Committee of The Jackson Laboratory.
Data from 3 other crosses are used in this manuscript: the published 294 female mouse B6×129 intercross3 and 2 unpublished crosses, the 277 female mouse B6×C3H/HeJ intercross and the 146 male mouse (NZO/H1LtJ×NON/LtJ)×NON/LtJ backcross. These latter 2 crosses are described in detail in a manuscript in preparation (Su, DiPetrillo, Ishimori, Leiter, Churchill, Paigen, Stylianou), but the raw data from these crosses are archived in a public database and currently available if the reader is interested at http://phenome.jax. org/pub-cgi/phenome/mpdcgi?rtn=projects/qtlprojlist. The relevant LOD score plots for chromosome 1 from these crosses are included in this manuscript.
Phenotyping and Genotyping
After consuming the atherogenic diet for 8 weeks, mice were fasted for 4 hours in the morning, blood collected from the retro-orbital sinus into tubes with EDTA, centrifuged at 9000 rpm for 5 minutes, and plasma frozen at −20°C until analyzed. Plasma samples were analyzed within a week of being collected for HDL concentrations, measured directly with a Synchron CX Delta System (Beckman Coulter) using the assay called HDLD, which is an enzymatic reagent kit (no. 650207, Beckman Coulter). In this measurement, only the HDL lipoprotein particles were solubilized, and the HDL cholesterol was released to react with cholesterol esterase and cholesterol oxidase. All crosses had HDL measured in exactly the same way. Each HDL assay included standards from the manufacturer, our own laboratory standards prepared from a large plasma pool, which was aliquoted and frozen. Once a month we also included biological standards by measuring 6 C57BL/6 mice. Our original HDL assay was carefully validated by ultracentrifugation6 in both chow and atherogenic-diet fed mice, and this assay was validated against our first assay.
DNA was prepared from tail samples. Tail tips were digested with proteinase K, and DNA was extracted with phenol-chloroform and resuspended in 10 mmol/L Tris.Hcl (PH:8.0). Polymorphic MIT microsatellite markers D1Mit1, D1Mit373, D1Mit212, D1Mit177, D1Mit132, D1Mit336, D1Mit218, D1Mit103, D1Mit14, and D1Mit148 were genotyped using agarose gel electrophoresis (NuSieve 3:1, FMC BioProducts) and SNPs rs3717961, rs3708797, and rs3022854 were genotyped by the Allele-Typing Service at The Jackson Laboratory in conjunction with KBiosciences.
HDL values were ln-transformed to reduce the right skew in the HDL distribution, and more importantly, to produce a constant variance among the crosses. This prevents a cross with greater phenotypic variance from dominating the results when cross data are combined.
QTL mapping was carried out as previously described7,8 with a multiple imputation algorithm9 used to account for missing marker genotypes. Sex was first included as an additive covariate to account for overall differences in HDL between the sexes. A second scan included sex as an interactive covariate to identify sex specific QTL. The difference in LOD score (ΔLOD) between these 2 scans constitutes a test for QTL-by-sex interaction, and a LOD of 2 or more represents a significance difference in QTL between sexes.10 QTL were deemed significant if they either met or exceeded the 95% genome-wide adjusted threshold, which was assessed by permutation analysis.9 Analyses were carried out using Pseudomarker 1.1 software (http://www.jax.org/staff/churchill/labsite). QTL confidence intervals were calculated according to the posterior probability density of QTL locations, as described previously.9
We combined the raw Chr 1 data from the B6×1293 and NZB×NZW crosses by recoding the B6 and NZB genotypes as L for low HDL alleles and the 129 and NZW genotypes as H for high HDL alleles.11 An LOD score was computed at 2-cM intervals across the QTL interval for each cross separately and then for both crosses combined. The combined data were analyzed with “HDL phenotype” as standardized and “cross” as an additive covariate.
Data other than the QTL were analyzed using Graphpad Prism (Windows v4.00, GraphPad Software). Student t test was used to compare the plasma HDL concentrations among the different groups.
SNPs used in the haplotype analysis were obtained from the publicly available mouse SNP databases (http://jax.org/phenome/snps). The missing SNP data for some strains were imputed as previously described.12 We identified the regions where the mouse genome displays a haplotype block structure, assigned individual strains to local dominant haplotypes, and inferred the genotypes of missing SNP alleles using a hidden Markov model. These imputed genotypes are publicly available at http://cgd.jax.org/imputedSNPdata/v1.1/.12
To sequence candidate genes in the parental strains NZW, 129, NZB, and B6, we used the genomic sequence of B6, obtained from the UCSC mouse genome assembly (http://genome.ucsc.edu/) to design primers that amplified each exon and at least 50 nucleotides of the adjacent introns (supplemental Table I). Purified polymerase chain reaction (PCR) products were subjected to thermocycle sequencing, and resulting fragments analyzed on capillary-based machines by the Jackson Laboratory DNA Sequence Laboratory. Sequences were analyzed by aligning to the genomic B6 sequence using Sequencher software (version 4.1.4, GeneCodes Technology). These sequences have been submitted to dbSNP (rs13475988 and rs13467752).
Gene Expression Examined by Real-Time PCR
Liver RNA was extracted from 5 female atherogenic diet-fed mice of each strain of B6, 129, NZB, and NZW by using TRIzol (Invitrogen) following the manufacturer’s instructions. Gene expression was examined by quantitative real-time PCR as described previously.8 Briefly, 2-μg RNA was primed with random hexamers to synthesize cDNA, cDNA samples were mixed with SYBR green master mix (Applied Biosystems) and gene-specific primers (supplemental Table II) in a total volume of 25 μL. Quantitative real-time PCR was performed on an ABI PRISM 7500 Sequence Detection System (Applied Biosystems). PCR reactions were carried out in triplicate. Gene expression was normalized to the expression of β-actin, and fold differences between strains were calculated by the ΔΔCt method.13
An Intercross Between Strains NZB and NZW Confirms Hdlq14
To confirm the existence of Hdlq14, we performed a cross that eliminated the effects of Apoa2. Using a dense SNP map from the Broad Institute http://www.broad.mit.edu/mouse/hapmap/, we choose strains to intercross that had identical Apoa2 genes, but different haplotypes over the Hdlq14 region. One pair of readily available strains fit these criteria, NZB and NZW. The Hdlq14 haplotype of NZW was identical to strain 129, and the haplotype of NZB was identical to strain B6. Thus, a cross between NZB×NZW should confirm the Hdlq14 QTL found between B6×129.
Plasma HDL concentrations were measured in both sexes for the parental (NZB×NZW) F1 mice, and 264 F2 progeny after animals had been fed an atherogenic diet for 8 weeks. Compared with sex-matched NZW, NZB mice displayed significantly increased HDL concentrations (P<0.01; Figure 1A). F1 female mice displayed HDL concentrations intermediate between and significantly different (P<0.01) from those of the female parental strains, F1 male mice displayed HDL concentrations comparable to those of NZB males and significantly higher than those of NZW males (P<0.01; Figure 1A). The distribution of log-transformed HDL concentrations in F2 progeny was approximately normal (Figure 1B).
Using sex as additive or interactive covariate, we identified a locus influencing plasma HDL with a peak near D1Mit336 (58.7 cM, 98.44 Mb in NCBI Build 36) and a significant LOD score of 3.3 or 4.4 depending on sex as additive or interactive covariate (Figure 1C). The difference in LOD score (ΔLOD) between the scans with sex as an additive or interactive covariate is 1.1, which is lower than the threshold (ΔLOD is 2 or more) for significant difference in QTL between the sexes, so the QTL did not differ between sexes. This QTL explained 5.5% of the total variance of HDL concentration. The 95% confidence interval was from 80 to 125 Mb, which overlapped with the 82 to 162 Mb Hdlq14 region identified in the 129×B6 intercross. The NZW allele for high HDL levels was recessive (Figure 1D). These data confirm the existence of Hdlq14 separate from the nearby QTL peaks caused by Hdlq15.
Bioinformatic Tools Narrowed Hdlq14 to 2 Genes
The confidence intervals for Hdlq14 were quite large: 80 Mb containing 487 genes for the B6×129 cross (Figure 2B, step a) and 45 Mb containing 225 genes for the NZB×NZW cross (Figure 2B, step b). To narrow Hdlq14, we used a series of bioinformatic and statistical tools,14 including analyzing combined cross data and haplotype analysis.
Analyzing Combined Cross Data Narrowed Hdlq14 to a 27.3-Mb 125-Gene Region
The large confidence interval of a typical QTL results from the limited number of recombinations in a cross. By combining and analyzing data from multiple crosses, the number of recombinations is increased, and the QTL interval is reduced.11 This depends on the assumption that the QTL in the 2 crosses are caused by the same gene. If they are not, then the confidence interval will not be decreased in the combined cross analysis and the QTL may actually divide into 2 peaks. To combine the raw data, we used log transformed values so that the crosses would contribute equally. The genotype information is recoded from a strain-specific code to a phenotype-specific code; B6 and NZB genotypes were recoded as L for the low HDL allele, and 129 and NZW genotypes were recoded as H for the high HDL allele. Our analysis increased the Hdlq14 LOD score to 5.1 (Figure 2A, dashed line), narrowed the 95% confidence interval to 27.3 Mb (from Mb 89.8 to 117.1 Mb; Figure 2B, step c), and reduced the number of genes in the region to 125.
Haplotype Analysis Narrowed Hdlq14 to a Region Containing Only 2 Genes
Haplotype analysis reduces the QTL regions by eliminating those regions that are identical by descent between the 2 strains as inferred by a shared SNP pattern. This is particularly effective for the related strains NZW and NZB because they shared 63.2% of their genomes.8 However, haplotype analysis depends on the assumptions that the mutation causing the QTL is ancestral and that the SNPs are sufficiently dense so that the DNA regions identical by descent can be correctly inferred. These are fairly safe assumptions in this case. Not only are 97% of mutations ancestral, but this QTL was found in 2 different crosses, thus making it very likely that the mutation is indeed ancestral. Furthermore, 3 of the strains, B6, NZW and 129, are among the strains resequenced by Perlegen so that the density of SNPs is very high.
We carried out the haplotype analysis in two steps using 4 crosses (Figure 2B, step d and e). Two of these crosses detected the QTL, B6×129 and NZW×NZB, and 2 crosses failed to detect the QTL, B6×C3H and NZO×NON (Figure 3A). In the first step we used the 2 crosses that detected the QTL to identify genomic regions shared between strains NZW and 129 but different from strains NZB and B6. We compared the haplotypes throughout the reduced Hdlq14 interval, defining a common haplotype block to be 3 or more consecutive shared alleles. We found those regions where the B6 SNPs were identical to the NZB SNPs (both strains have alleles for low HDL), where the NZW SNPs were identical to 129 SNPs (both stains have alleles for high HDL), and where the SNPs for the strains with high and low HDL alleles differed. This haplotype analysis further narrowed Hdlq14 to a 2.3 Mb, 19-gene region (Figure 2B, step d, Figure 3B).
In the second step, we used the haplotypes of the parental strains in the 2 crosses that failed to detect a QTL, the NZO×NON and the B6×C3H crosses (Figure 3A). Because the manuscript for these 2 crosses is still in preparation (Su, DiPetrillo, Ishimori, Leiter, Paigen, and Stylianou), the raw genotype and phenotype data has been archived in http://phenome.jax.org/pub-cgi/phenome/mpdcgi?rtn=projects/qtlprojlist. For these 2 crosses, we searched for a haplotype that is the same for all 4 strains because no QTL was detected. These crosses had identical haplotypes for the SNPs in the Farp2 and Stk25 genes but different haplotypes for the other 17 genes (Figure 3B). Thus, these 2 crosses reduced the QTL region to just 2 genes and eliminated the other 17 genes (Figure 2B, step e).
Expression Studies and Sequencing of Farp2 and Stk25 Provide Additional Evidence for Both Genes
Because the QTL gene must have either an expression difference between the parental strains or a coding region polymorphism that changes function, we sequenced both candidate genes in B6, 129, NZW, and NZB. The SNPs were classified by whether or not they changed an amino acid (Cn for coding-nonsynonymous, Cs for coding region-synonymous).
Sequencing Stk25 revealed only 1 synonymous change (Table 1A), but recent evidence15,16 has shown that a synonymous change can affect a protein’s expression level and function if it changes a codon’s frequency. Using codon frequency tables (http://www.kazusa.or.jp/codon), we found the Cs in Stk25 had a major change in codon frequency (23% for ACA to 5% for ACG). Sequencing Farp2 revealed 1 Cn and 1 Cs (Table 1A). The Cn in Farp2 changed the amino acid from leucine to proline at amino acid 821, which could affect the folding and thus the function of its encoded protein, and this variant was in a highly conserved region of the protein known as a pleckstrin domain17 (Figure 4A). We sequenced Farp2 in 12 strains (129, A/J, B6, C3H, CAST/EiJ, DBA/2J, FVB/NJ, NOD/LtJ, NZB, NZW, RIIIS/J, SM/J, and SJL/J) and determined the amino acid changing SNP (rs13475988) in an additional 4 strains (PERA/EiJ, I/LnJ, NZO/H1J, and NON/LtJ). Combining our data with SNP data in the Mouse Phenome Database, we show the distribution of the key amino acid changing SNP across the 45 inbred strains (Table 1B).
We also compared expression in the livers of atherogenic diet-fed mice for the parental strains and their F1 progeny for both genes using quantitative real-time PCR. Farp2 showed no difference in expression, but Stk25 expression was significantly higher in the low HDL allele strains, 4.6-fold higher in B6 compared to 129, and 4.8-fold higher in NZB compared to NZW (Figure 4B). Stk25 expression in (B6×129) F1 or (NZB×NZW) F1 mice was comparable to the B6 or NZB mice, respectively (Figure 4B). This is consistent with the recessive high HDL allele observed in each cross (Figure 1D).
These studies showed that Farp2 had an amino acid change that might affect function and that Stk25 had an expression difference. Thus, both genes remain viable candidates for the QTL.
The Farp2/Stk25 Haplotype Was Associated With Altered HDL Concentrations
Farp2 and Stk25 are immediately adjacent to each other and partly overlapping, although transcribed in opposite directions (http://www.ensembl.org/). There has been no recombination between these genes among the 45 inbred strains that we examined (supplemental Figure II). To determine whether the key amino acid changing polymorphism in Farp2 was associated with variation in HDL, we analyzed plasma HDL concentrations of the 45 inbred mouse strains (www.jax.org/phenome), avoiding the effect of Hdlq15 by dividing the strains into groups based on their polymorphisms at Apoa2 (the key amino acid change Ala61-to-Val61)4 (supplementary Figure II). In Apoa2 Ala61 group (32 strains; Figure 5), HDL levels of strains having the haplotype associated with the Farp2 821Leu allele (11 strains) were significantly higher than those of strains having the haplotype associated with the 821Pro allele (21 strains) in both female and male mice (65±7 versus 52±5 mg/dL and 83±8 versus 62±6 mg/dL, P<0.01, respectively). Similar trends were found in strains having Apoa2 Val61 (13 strains; 104±12 versus 81±9 mg/dL in females, P<0.01, and 135±13 versus 101±8 mg/dL in males, P<0.01; Figure 5). Thus, the haplotype at Farp2/Stk25 affects HDL concentration in multiple strains.
In the present study, we confirmed another chromosome 1 HDL QTL, previously identified in a B6×129 intercross,3 by crossing 2 strains that did not differ in the genotype for 2 nearby QTL genes. The peak of this locus was at D1Mit336 (58.7 cM, 98.4Mb), and the 95% confidence interval was from 80 Mb through 125 Mb. This study illustrates a useful method of separating closely linked QTLs. Usually such separation is achieved by constructing a set of overlapping congenic strains in the region, a process that requires a minimum of 5 to 6 generations of breeding. This report presents an alternative method of selecting strains that are closely related and that are identical for the gene (or haplotype) of one QTL region but differ in haplotype for the second region thought to possess the second QTL gene. Although this method is faster than constructing congenics, it does require the use of data from multiple F2 crosses, but in this particular case, those crosses were already done. In a 2-generation intercross, we were able to show that this region did indeed carry another QTL gene. Then we narrowed the QTL region by a set of bioinfomatic methods including combining the two crosses and haplotyping with 2 crosses that detected the QTL. We also haplotyped using 2 crosses, B6×C3H and NON×NZO, that failed to detect the QTL. It is always problematic to reason from negative data, but in the case of these 2 crosses we had the raw data from our own work, and we could ensure from the flat LOD score plots (Figure 3A) that these QTLs were truly absent and not just merely present but not significant enough to be reported. Adding these 2 crosses narrowed Hdlq14 from a 45-Mb 271-gene to 2 genes.
We evaluated these candidate genes by sequence and expression analysis. Stk25 had a large expression difference, and Farp2 had an amino acid change. The variant Pro821Leu of Farp2 was in a highly conserved region of the protein (Figure 4A) in the pleckstrin homology (PH) domain,17,18 a region characteristic of proteins with diverse enzymatic or regulatory functions (such as phospholipases, GTPase-regulating proteins, and protein kinases) and lipid-binding proteins.19 The pleckstrin domain itself is thought to bind lipids.18 Stk25 encodes serine/threonine kinase 25, a crucial regulator of AMPK kinase in muscle and liver cells:20 its activity is potentially important in lipid metabolism.
Because rodent and human QTLs for the same trait often map to the homologous genomic regions, these homologous QTLs in mice and humans may be caused by the same underlying gene, which is correct in many instances.21–23 We often use the concordance of mouse and human QTL in narrowing a QTL region.7,8,24 In fact, Farp2 and Stk25 are located on human 2q37.2 to 37.3, where a human QTL for HDL has been identified.25 We did not use this human QTL to narrow the mouse QTL region because it had a low LOD score in the original report, which was apparent only when applied to diabetics and disappeared when applied to the whole population. However, now that we have narrowed the mouse QTL, the existence of the human QTL provides some additional support. It has been also reported that the polymorphisms in Farp2 gene are associated with the variations of human HDL levels in a more recent genome-wide association study.26 These two lines of evidence suggest that Farp2 and Stk25 identified from mice may be likely HDL genes for humans as well.
We were unable to further decide between these 2 candidates; they are immediately adjacent to each other and no historical recombination occurred between them in the 45 inbred strains that we examined. Additional studies are needed to choose between these 2 genes.
The authors thank Harry Whitmore for his invaluable help in mouse husbandry.
Sources of Funding
This work was funded by the U.S. National Institutes of Health grants to B.P. (HL81162, HL74086, and HL77796), the American Heart Association grant 0725905T to Z.S., and by the Novartis Institutes for Biomedical Research.
Received May 20, 2008; revision accepted October 19, 2008.
Ishimori N, Li R, Kelmenson PM, Korstanje R, Walsh KA, Churchill GA, Forsman-Semb K, Paigen B. Quantitative trait loci analysis for plasma HDL-cholesterol concentrations and atherosclerosis susceptibility between inbred mouse strains C57BL/6J and 129S1/SvImJ. Arterioscler Thromb Vasc Biol. 2004; 24: 161–166.
Wang X, Korstanje R, Higgins D, Paigen B. Haplotype analysis in multiple crosses to identify a QTL gene. Genome Res. 2004; 14: 1767–1772.
Nishina PM, Verstuyft J, Paigen B. Synthetic low and high fat diets for the study of atherosclerosis in the mouse. J Lipid Res. 1990; 31: 859–869.
Ishida BY, Blanche PJ, Nichols AV, Yashar M, Paigen B. Effects of atherogenic diet consumption on lipoproteins in mouse strains C57BL/6 and C3H. J Lipid Res. 1991; 32: 559–568.
Su Z, Tsaih SW, Szatkiewicz J, Shen Y, Paigen B. Candidate genes for plasma triglyceride, FFA, and glucose revealed from an intercross between inbred mouse strains NZB/B1NJ and NZW/LacJ. J Lipid Res. 2008; 49: 1500–1510.
Sen S, Churchill GA. A statistical framework for quantitative trait mapping. Genetics. 2001; 159: 371–387.
Li R, Lyons MA, Wittenburg H, Paigen B, Churchill GA. Combining data from multiple inbred line crosses improves the power and resolution of quantitative trait loci mapping. Genetics. 2005; 169: 1699–1709.
Kimchi-Sarfaty C, Oh JM, Kim IW, Sauna ZE, Calcagno AM, Ambudkar SV, Gottesman MM. A “silent” polymorphism in the MDR1 gene changes substrate specificity. Science. 2007; 315: 525–528.
Lemmon MA. Pleckstrin homology (PH) domains and phosphoinositides. Biochem Soc Symp. 2007: 81–93.
Wang X, Ria M, Kelmenson PM, Eriksson P, Higgins DC, Samnegard A, Petros C, Rollins J, Bennet AM, Wiman B, de Faire U, Wennberg C, Olsson PG, Ishii N, Sugamura K, Hamsten A, Forsman-Semb K, Lagercrantz J, Paigen B. Positional identification of TNFSF4, encoding OX40 ligand, as a gene that influences atherosclerosis susceptibility. Nat Genet. 2005; 37: 365–372.
Su Z, Wang X, Tsaih SW, Zhang A, Cox A, Sheehan S, Paigen B. Identifying the genetic basis of HDL variation between 129/SvImJ and C57BL/6J mice: critical importance for testing HDL candidate genes in targeted mutant mice. J Lipid Res. In press.
Elbein SC, Hasstedt SJ. Quantitative trait linkage analysis of lipid-related traits in familial type 2 diabetes: evidence for linkage of triglyceride levels to chromosome 19q. Diabetes. 2002; 51: 528–535.
Saxena R, Voight BF, Lyssenko V, Burtt NP, de Bakker PI, Chen H, Roix JJ, Kathiresan S, Hirschhorn JN, Daly MJ, Hughes TE, Groop L, Altshuler D, Almgren P, Florez JC, Meyer J, Ardlie K, Bengtsson Bostrom K, Isomaa B, Lettre G, Lindblad U, Lyon HN, Melander O, Newton-Cheh C, Nilsson P, Orho-Melander M, Rastam L, Speliotes EK, Taskinen MR, Tuomi T, Guiducci C, Berglund A, Carlson J, Gianniny L, Hackett R, Hall L, Holmkvist J, Laurila E, Sjogren M, Sterner M, Surti A, Svensson M, Tewhey R, Blumenstiel B, Parkin M, Defelice M, Barry R, Brodeur W, Camarata J, Chia N, Fava M, Gibbons J, Handsaker B, Healy C, Nguyen K, Gates C, Sougnez C, Gage D, Nizzari M, Gabriel SB, Chirn GW, Ma Q, Parikh H, Richardson D, Ricke D, Purcell S. Genome-wide association analysis identifies loci for type 2 diabetes and triglyceride levels. Science. 2007; 316: 1331–1336.