In Silico Quantitative Trait Locus Map for Atherosclerosis Susceptibility in Apolipoprotein E–Deficient Mice
Objective— Atherosclerosis susceptibility is a genetic trait that varies between mouse strains. The goal of this study was to use a public mouse single nucleotide polymorphism (SNP) database to define the genetic loci that are associated with this trait, without the need to perform strain intercrosses that are normally required to obtain these loci.
Methods and Results— Apolipoprotein E (apoE)–deficient mice on 6 inbred genetic backgrounds were compared for atherosclerosis lesion size in the aortic root in 2 independent studies. After normalization to the C57BL/6 strain that was used in both studies, lesion areas were found in the following rank order: DBA/2J>C57BL/6>129/SV-ter>AKR/J≈BALB/cByJ≈C3H/HeJ. The log lesion difference in phenotypes between each of the 15 heterologous strain pairs was determined. A mouse SNP database was then used to calculate the genetic differences between the 15 strain pairs in partially overlapping 30-cM bins across the mouse genome. Correlation analyses were preformed to analyze the genetic and phenotypic differences among the strain pairs for each genetic region. The genetic regions with the highest correlations define the in silico quantitative trait loci (QTL) associated with the atherosclerosis phenotype. Five in silico atherosclerosis QTL were identified on chromosomes 1, 10, 14, 15, and 18. The loci on chromosomes 1, 10, 14, and 18 overlap with suggestive atherosclerosis QTL identified through analyses of an F2 cohort derived from apoE-deficient mice on the C57BL/6 and FVB/N strains.
Conclusions— The 5 identified in silico QTL are candidates for further study to confirm the presence and identity of atherosclerosis susceptibility genes within these loci.
Coronary heart disease is the most common cause of death in the United States. Most cases of coronary heart disease can be attributed to or are associated with atherosclerosis, a complex disease often initiated by hypercholesterolemia involving both environmental and genetic factors. Large epidemiological studies have shown that approximately half of those who develop coronary heart disease have moderate total plasma cholesterol levels of <250 mg/dL.1 Thus, the elucidation of genetic risk factors involved in atherosclerosis susceptibility would provide a means to better direct therapeutic intervention to those most likely to develop atherosclerosis who do not have obvious plasma lipid risk factors.
Mouse models have emerged as 1 of the most useful tools for experimental atherosclerosis research. Mice are small, relatively inexpensive, and easy to maintain; they have a short generation time and can multiply rapidly. Mice made deficient in apolipoprotein E (apoE) by gene targeting have elevated plasma cholesterol levels and spontaneously develop severe atherosclerosis on low-fat chow diets.2,3⇓ These apoE-deficient mice develop lesions that are similar in many aspects to human lesions and are thus a valuable paradigm for studying atherosclerosis in humans.4 We and others have bred apoE-deficient mice onto a number of distinct inbred genetic background strains, and lesion area has been scored in these strains.5,6⇓ ApoE-deficient mice on the C57BL/6 background have lesions ≈8-fold larger than do apoE-deficient mice on the FVB/N background, with intermediate levels in F1 hybrids, and a broad distribution overlapping both parental strain values in the F2 generation.5 A traditional quantitative trait locus (QTL) assay has been performed on 2 independent F2 cohorts from these parental strains and has identified highly significant and suggestive loci on chromosomes 10, 14, and 19.7 We have modified and used the recently described method of in silico QTL mapping8 to use a computational approach to accelerate the mapping of atherosclerosis susceptibility genes. Genetic differences along the mouse genome between strains based on a single nucleotide polymorphism (SNP) database were correlated to phenotypic differences, allowing us to identify chromosomal regions that may contain atherosclerosis susceptibility genes.
Methods and Results
ApoE-deficient mice created at Rockefeller University (RU) and using J1 129/SV-ter embryonic stem cells2 were bred back 10 generations onto the DBA/2J and AKR/J backgrounds at RU. ApoE-deficient mice bred 10 generations onto the C57BL/6 background were also obtained from The Jackson Laboratory (Bar Harbor, Me). At Millennium Pharmaceuticals Inc (MPI), the RU apoE-deficient mice were maintained on an inbred 129/SV-ter background and subsequently bred onto the C57BL/6, BALB/cByJ, and C3H/HeJ backgrounds by using the speed congenic method, as previously described.6 All strains of mice were assayed with polymorphic markers at ≈10-cM intervals and found to be >99% inbred. Two atherosclerosis studies were performed, 1 at RU and 1 at MPI. Lesions in the aortic root were quantified in both male and female mice from each of the strains (Table 1). The RU atherosclerosis study was performed on apoE-deficient mice on the C57BL/6, DBA/2J, and AKR/J strains and included 16-week-old mice that had been maintained on a semisynthetic diet containing 4.5% fat and 0.02% cholesterol (wt/wt). The MPI atherosclerosis study was performed on apoE-deficient mice on the C57BL/6, 129/SV-ter, BALB/cByJ, and C3H/HeJ strains and included 20-week-old mice maintained on a breeder chow diet containing 9% fat. There was a wide distribution of lesion areas among the strains, implying that there are significant genetic differences among the strains that contribute to atherosclerosis susceptibility (Table 1). ANOVA analyses showed that the strain effect on lesion area was highly significant in separate analyses of the male and female mice in the RU study (P<0.0001), whereas in the MPI study with smaller sample sizes, the strain effect was significant in females mice (P=0.0002) but only borderline significant in male mice (P=0.077). In the RU study, lesions in female were larger than in male mice for the DBA/2J and C57BL/6 strains (P<0.01), as has been previously noted for mouse aortic root lesions,9,10⇓ whereas there was a nonsignificant trend for larger lesions in male than in female mice on the AKR/J strain (Table 1). In the MPI study, a trend toward larger lesions in C57BL/6 female versus male mice was observed, but the sample size was insufficient to obtain statistical significance. To combine lesion data from all 6 strains of the 2 studies, mean lesion areas from each study were normalized to the mean lesion area of the C57BL/6 strain, which was used in both of the studies (Table 1). The normalized data revealed an ≈15-fold and 23-fold variation in lesion area between the most susceptible and resistant male and female strains, respectively, with the following lesion area strain order: DBA/2J>C57BL/6>129/SV-ter>AKR/J, BALB/cByJ, and C3H/HeJ, which all have smaller lesions with approximately equal lesion areas. To compare the fold differences rather than the absolute differences in lesion areas between the strains, the logarithm of the normalized lesion area for each of the 6 strains was used to calculate the phenotypic differences (ΔP) between the 15 strain pairs (Table 2). The smaller sample sizes in the MPI study did not impair our analysis of the differences between mean lesions areas in the various mouse strains, as it is unlikely that larger sample sizes would change the means significantly, and only the mean values, rather than the sample distribution, are used in the subsequent statistical analyses.
The 19 autosomes and X chromosome of the mouse genome were fractionated into 113 partially overlapping genetic bins. Each bin spans 30 cM (or less, for some bins at the distal end of each chromosome) and starts at 10-cM intervals. For example, chromosome 1 spans 114 cM and was divided into 10 bins, 1a (0 to 30 cM), 1b (10 to 40 cM), etc, through 1j (90 to 114 cM). The in silico QTL method requires calculation of the genotypic difference (ΔG) for each genetic bin for each of the 15 strain pairs. Using the Roche SNP database (http://mousesnp.roche.com), we retrieved data for 2728 SNPs; however, 29 of these were not definitively mapped to a specific chromosomal position and were discarded. In the published article describing the in silico QTL method, it appears that ΔG was calculated by simply counting allelic differences between each strain pair for each bin.8 However, this can lead to 2 potential artifacts: (1) because some SNPs have not been determined for all of the strains under study, the counted ΔG would likely be higher for strain pairs with the most genotype information and (2) because the density of SNP markers is not uniform among the genetic bins, bins with more markers were likely to have greater absolute ΔG values than bins with fewer markers. We were able to overcome these shortcomings and use a set of SNP markers in which information for at least 2 of the strains was present, for a total of 2495 markers. These markers were distributed into the overlapping bins, yielding, on average, 49±23 (mean±SD) markers per bin, with a range between 7 and 118 markers. Aside from the X chromosome, which had a low marker density, all of the bins with <20 markers were on the distal ends of a chromosome, in bins that were smaller than 30 cM. In our modified method, instead of defining ΔG as the absolute number of allele differences between each strain pair in a genetic bin, we defined ΔG as the proportion of genetic variation in each bin by dividing the number of strain-pair allele differences by the total number of markers with allele assignments for the specific strain pair. Thus, strain-pair ΔG values for each bin can fall between 0 and 1 for strains that share all and no SNP alleles, respectively. One bin with 9 markers at the distal end of chromosome 13 had no genetic variation and was eliminated from further analysis. The average ΔG value across the remaining 112 bins was 0.172±0.065 (mean±SD), meaning that, on average, the strains differed from each other at 17% of the SNP markers. However, genetic variation was not uniformly distributed among the 112 bins, with a range of strain-pair mean ΔG values between 0.032 and 0.376.
To find loci in which genetic variance was correlated with phenotypic variance among the strain pairs, the in silico QTL method relies on correlation analysis between an array of strain-pair Δ probability values and an array of strain-pair ΔG values for each genetic bin.8 To perform the in silico QTL analysis, data were entered into Microsoft Excel spreadsheets. Phenotypic differences (log Δ probability values) for the 15 strain pairs as displayed in Table 2 were entered and served as the 15 x values for the subsequent correlation analysis. For each of the 112 partially overlapping 30-cM genetic bins, the spreadsheet contained values for the total number of SNP markers with information for each of the 15 strain-pair combinations and the number of SNPs that differed for each of these 15 strain-pair combinations. The proportional genetic differences (ΔG values) for each bin and each strain pair were then calculated as the number of SNP differences divided by the total number of SNPs, yielding an array of 15×112 ΔG values. This array of ΔG values served as the y values for correlation analysis. Excel was used to calculate the correlation coefficient for each genetic bin by using the 15 shared x values and the 112 sets of 15 unique y values. The raw correlation coefficients could theoretically vary between 1 and −1 for genetic bins in which the Δ probability values were either perfectly or inversely correlated to the ΔG values, respectively. To scale these values, the mean of all of the 112 correlation coefficients was subtracted from each correlation coefficient, and this difference was divided by the SD of all of the correlation coefficients. Thus, a scaled R of 0 represents the mean R value, and positive and negative scaled R values indicate how many SDs the R value is above or below the mean R, respectively. These scaled values are plotted in Figure 1. We modified the original method by assessing proportional rather than absolute ΔG values and by excluding self-strain comparisons that led to values of 0 for both ΔP and ΔG, because these values are purely structural and do not add to the power of the correlation analysis.
The in silico QTL analysis yielded 6 loci for female mice, with scaled R thresholds >2.0, more stringent than the threshold scaled R values of ≥1.5 that were used in the original in silico QTL study.8 We chose to use this stringent threshold to select the loci with the most likelihood of being authentic QTL. These 6 in silico QTLs were in chromosome bins 1a, 8f, 10a, 14c, 15c, and 18a (Figure 1a). The results were quite similar in male mice, with in silico QTLs found in chromosome bins 1a, 8f, 10a, and 14c, whereas bins 15c and 18a just failed to reach the scaled R threshold of 2.0 (Figure 1b). If the less stringent R threshold of 1.5 was used, additional chromosomal bins 3f, 9a, 10b-e, 11a, b, and d, 15c, and 18a in the male mice and bins 9a, 10b–e, and 11d in the female mice would meet this criterion. The chromosome 8f locus was unique among the 6 loci detected by the in silico analysis with R>2.0, in that the ΔG for this bin was quite small, only 0.060, compared with the ΔG values ranging from 0.139 to 0.197 for the other 5 QTL bins. This bin only contained 10 markers, of which only 1 was polymorphic (DBA/2J and C57BL/6 shared an allele that was distinct from the allele shared by the other 4 strains). The other 5 QTL bins contained between 31 and 88 markers. Because of the limited genetic variation and markers in the chromosome 8f bin, we have less confidence in this suggestive atherosclerosis susceptibility locus.
The in silico QTL method as originally published was verified by its ability to predict loci that had been associated with 10 various phenotypes from traditional mouse genetic methods, albeit with some loci missed (potential in silico false-negatives) and other loci not found by the traditional methods (potential in silico false-positives).8 The clear advantage of this method is that it does not require the labor and time-intensive generation of F2 or N2 cohorts and the individual phenotyping and genotyping of each of the hundreds of progeny from these crosses. In addition, traditional QTL methods can only compare 2 strains at a time, whereas the power of the in silico method increases as more strains are studied. The use of more than 2 strains leads to the inclusion of more genetic diversity and thus potentially allows more loci to be uncovered. The minimum number of strains compared in the original description of this method was 4. Analysis of the MPI data set containing 4 strains yielded only 6 strain-pair comparisons and thus 6 ΔP, ΔG data points for the correlation analysis (compared with 15 data points used in the shared data set). In silico QTL analysis performed with only the MPI data set did not yield any loci with a scaled R value ≥2.0 (supplemental Figure 1 available online), and the pattern was considerably different from that observed in the combined data set (Figure 1). We believe that data based on the shared data set derived from 6 strains are superior to the data derived from only 4 strains because the correlation analysis is based on 15 data points per genetic bin rather than 6, despite the fact that these 6 strains were analyzed in 2 separate studies and normalized to the phenotype observed in the C57BL/6 strain that was analyzed in both studies.
The in silico method is subject to limitations based on the unequal distribution of marker information, but as the number and density of mouse SNPs in the database increase, this caveat may be less important. Additionally, the in silico method cannot be used to find epistatic genetic interactions. The in silico QTL method, as it is used here and in its original description, does not pay attention to blocks of linkage disequilibrium between the strains. As the SNP marker density increases, it may well be that analyses that take linkage disequilibrium into account may increase the sensitivity and specificity of this method. Furthermore, because the in silico method relies on evolutionary derived genetic differences between the strains and not authentic linkage through breeding and recombination, this method may not be able to detect phenotype-altering mutations that occurred after the strains diverged from each other. The in silico QTL method, as used here and in the original publication, provides only a crude map in 30-cM partially overlapping bins, which is comparable in scale to QTL intervals that we observed in a traditional QTL analysis that used markers with an ≈10-cM spacing.7 The in silico QTL method requires bins of this size to include sufficient SNP information to obtain a reasonable indication of the genetic variance between any given strain pair. As the SNP marker density increases, it may be possible to perform a somewhat finer map. However, this method is probably not capable of producing a 1-cM fine map for 2 reasons: (1) as the genetic bin size decreases, each SNP becomes too powerful in the analysis, and random SNP divergence and concordance between the strains would make it difficult to observe the authentic genetic similarities between the strain pairs and (2) the regions of linkage disequilibrium between the strains may be so large as to preclude the generation of a fine map.
The information content of each SNP within each bin is the same whether there is no, low, or a high level of strain-pair variation, because each SNP is scored as either being the same or different for each strain pair, and either result is weighted equally in the calculation of ΔG for each strain pair. However, in bins with a small number of SNPs, each SNP is weighted more highly than in bins with a large number of SNPs, and in bins where there is little genetic variation between the strains, even 1 SNP can have a large effect on the ΔG value. For example, there were only 10 SNPs in bin 8f, of which only 1 was polymorphic, and although the analysis of this bin generated a significant R value, we discounted this result because of the low level of the mean strain-pair ΔG value in this bin (0.06). The other 5 bins with the strongest insilico QTLs (R>2.0) each had between 31 and 86 markers, with mean strain-pair ΔG values between 0.14 and 0.20, and in this regard, these bins were of average size and genetic diversity (all parameters were within 1 SD of the mean marker number and ΔG values, with the exception of bin 18a with 86 markers, which was 1.6 SD above the mean marker number). Compared with the original publication of this method, we suggest 2 potentially important modifications: (1) use of a proportional ΔG value to decrease bias due to incomplete marker information and uneven marker distribution and (2) elimination of the self-strain comparisons that do not add to the correlation analysis.
Atherosclerosis susceptibility loci have been previously described in wild-type mice by using high-cholesterol, cholic acid–containing diets to induce lesions. Ath1, initially described from 2 sets of recombinant inbred strains derived from the atherosclerosis-susceptible C57BL/6 and atherosclerosis-resistant C3H strains and the C57BL/6 and BALB/c strains, maps to chromosome 1,11 but, it is quite distal to the chromosome bin 1a locus discovered by in silico QTL mapping in the current study. The veracity of the Ath1 locus has been disputed, because it was not confirmed in a subsequent study that used 1 of the same set of recombinant inbred strains.12 Ath3 has been mapped to chromosome 7 by the use of recombinant inbred strains derived from the susceptible C57BL/6 and resistant A/J strains.13 Ath6 was mapped from an F2 cross between the susceptible C57BL/6 strain and the more susceptible C57BLKS strain to a region on chromosome 12 near the Apob gene, and the locus effect on atherosclerosis was confirmed through the creation of a secondary congenic strain in which the chromosome 12 locus from the C57BLKS strain was bred onto the C57BL/6 background.14 This locus was more finely mapped through additional crosses to a region near the proximal end of chromosome 12, thus ruling out Apob as a candidate gene.15 The best example of an atherosclerosis susceptibility allele derived from a study of wild-type mice fed an atherogenic diet has come from an F2 cohort descended from the susceptible C57BL/6 and resistant CAST/Ei strains.16 In that study, a locus with a peak log of the odds ratio (LOD) score of ≈6.7 was found at 63 cM on chromosome 6.16 The effect of this locus, called Artles, on lesion size was confirmed in interval-specific secondary congenic strains made on both the wild-type and LDL receptor–deficient background.16 In the LDL receptor–deficient background, an N2 cross between mice on the C57BL/6 and MOLF/Ei strains revealed 2 atherosclerosis susceptibility loci, called Athsq1 and Athsq2, near the distal ends of chromosome 4 and 6, respectively,17 with the chromosome 6 locus matching the location of the Artles locus.16 None of these loci were confirmed in the current analysis. It is quite possible that the failure of our in silico QTL to detect these loci is due to biological reasons, in that different genes may affect lesion development when apoE-deficient mice fed a low-cholesterol diet are compared with wild-type or LDL receptor–deficient mice fed high-cholesterol diets. In addition, some of these previously identified loci may be specific for certain strain pairs and thus, not found consistently within the 15 strain-pair comparisons of the in silico QTL analysis.
The current analysis yielded 5 suggestive loci of which we are most confidant in bins 1a, 10a, 14c, 15c, and 18a. Previously, we analyzed atherosclerosis lesion variation in apoE-deficient mice on the C57BL/6 and FVB/NJ backgrounds,5 the latter strain not represented in the mouse SNP database. Along with MPI, we performed a traditional QTL analysis to identify loci associated with lesion variance in 2 separate F2 cohorts.7 Two of the 3 strongest atherosclerosis susceptibility loci to emerge from this traditional QTL study reside on chromosomes 10 (Ath11, highly significant peak LOD score of 7.8 at 10 cM) and 14 (Ath13, suggestive peak LOD score of 3.2 at 25 cM),7 and these 2 loci fall into bins 10a and 14c, 2 of the 5 in silico QTL that we have identified in the current analysis. Another suggestive atherosclerosis QTL on the proximal end of chromosome 1 (peak LOD score 2.3) was also found in this traditional QTL study,7 which falls into the in silico QTL in bin 1a. In addition, an epistatic atherosclerosis susceptibility locus on chromosome 18 has been identified from the same F2 cross by the use of an independent, nonparametric 2-locus pairwise statistical analysis,18 and this locus falls within bin 18a, also identified by the in silico QTL method. Thus, 4 of the 5 in silico QTL for atherosclerosis have been previously implicated in atherosclerosis from the FVB/N×C57BL/6 apoE-deficient F2 cross, verifying the ability of our in silico QTL analysis to detect previously identified atherosclerosis QTL in the apoE-deficient background. However, there were additional loci identified by the traditional QTL that were not observed in the in silico analysis, including a locus on chromosome 19 (Ath16, peak LOD score of 3.8 at ≈35 cM) and a weaker locus on chromosome 16 (peak LOD score of 2.5).7 In addition, an in silico QTL was detected in bin 15c that was not previously identified from the F2 cohort studies. Despite the high degree of concordance between the traditional and in silico QTL for atherosclerosis in the apoE-deficient background, all of the QTL identified by either method must be independently verified by the demonstration of an effect on lesion size through the creation and analysis of interval-specific secondary congenic mice. In the current analysis, in which mice on different diets from 2 studies were combined, it is possible that dietary variation could have played a role in the phenotypic variation observed between the strains. However, we attempted to minimize this variation by normalizing lesion areas from both studies to the same strain, and the concordance between the QTL identified by this analysis and the traditional QTL analysis argues that the dietary variation probably did not play a significant role in our findings.
In conclusion, the in silico QTL method has suggested 5 loci for further study that may contain polymorphic atherosclerosis susceptibility genes among common mouse strains. The elucidation of these genes may provide insight into novel pathways regulating atherogenesis and will allow for the determination of whether common variation in these genes exists among humans, and if so, whether these variations are associated with the incidence of coronary heart disease. The in silico QTL method appears robust in that 4 of the 5 loci identified have been previously implicated in lesion size through analysis of an interstrain F2 cohort on the apoE-deficient background. In the current era of publicly available phenotype and genotype databases (for example, the mouse phenome database is available from a JAX website http://aretha.jax.org/pub-cgi/phenome/mpdcgi?rtn=docs/home), the in silico QTL method can readily be applied to discover loci that may be related to a specific trait, saving several years of work that are needed to discover these loci by traditional methods. The full justification of this method awaits the confirmation of these in silico QTL as authentic phenotype-varying loci and the determination of the responsible gene or genes within them.
This work was supported by National Institutes of Health grant HL54591. We thank Millennium Pharmaceuticals, Inc, for sharing their atherosclerosis data with us.
Received September 29, 2002; revision accepted November 4, 2002.
- ↵Zhang SH, Reddick RL, Piedrahita JA, Maeda N. Spontaneous hypercholesterolemia and arterial lesions in mice lacking apolipoprotein E. Science. 1992; 258: 468–471.
- ↵Nakashima Y, Plump AS, Raines EW, Breslow JL, Ross R. ApoE-deficient mice develop lesions of all phases of atherosclerosis throughout the arterial tree. Arterioscler Thromb. 1994; 14: 133–140.
- ↵Dansky HM, Charlton SA, Sikes J, Heath SC, Simantov R, Levin LF, Shu P, Moore KJ, Breslow JL, Smith JD. Genetic background determines the extent of atherosclerosis in apoE-deficient mice. Arterioscler Thromb Vasc Biol. 1999; 19: 1960–1968.
- ↵Dansky HM, Shu P, Donavan M, Montagno J, Nagle DL, Smutko JS, Roy N, Whiteing S, Barrios J, McBride TJ, Smith JD, Duyk G, Breslow JL, Moore KJ. A phenotype-sensitizing ApoE-deficient genetic background reveals novel atherosclerosis predisposition loci in the mouse. Genetics. 2002; 160: 1599–1608.
- ↵Grupe A, Germer S, Usuka J, Aud D, Belknap JK, Klein RF, Ahluwalia MK, Higuchi R, Peltz G. In silico mapping of complex disease-related traits in mice. Science. 2001; 292: 1915–1918.
- ↵Tangirala RK, Rubin EM, Palinski W. Quantitation of atherosclerosis in murine models: correlation between lesions in the aortic origin and in the entire aorta, and differences in the extent of lesions between sexes in LDL receptor-deficient and apolipoprotein E-deficient mice. J Lipid Res. 1995; 36: 2320–2328.
- ↵Smith JD, Trogan E, Ginsberg M, Grigaux C, Tian J, Miyata M. Decreased atherosclerosis in mice deficient in both macrophage colony-stimulating factor (op) and apolipoprotein E. Proc Natl Acad Sci U S A. 1995; 92: 8264–8268.
- ↵Paigen B, Mitchell D, Reue K, Morrow A, Lusis AJ, LeBoeuf RC. Ath-1, a gene determining atherosclerosis susceptibility and high density lipoprotein levels in mice. Proc Natl Acad Sci U S A. 1987; 84: 3763–3767.
- ↵Mu JL, Naggert JK, Svenson KL, Collin GB, Kim JH, McFarland C, Nishina PM, Levine DM, Williams KJ, Paigen B. Quantitative trait loci analysis for the differences in susceptibility to atherosclerosis and diabetes between inbred mouse strains C57BL/6J and C57BLKS/J. J Lipid Res. 1999; 40: 1328–1335.
- ↵Mehrabian M, Wong J, Wang XP, Jiang ZM, Shi WB, Fogelman AM, Lusis AJ. Genetic locus in mice that blocks development of atherosclerosis despite extreme hyperlipidemia. Circ Res. 2001; 89: 125–130.
- ↵Welch CL, Bretschger S, Latib N, Bezouevski M, Guo Y, Pleskac N, Liang CP, Barlow C, Dansky H, Breslow JL, Tall AR. Localization of atherosclerosis susceptibility loci to chromosomes 4 and 6 using the Ldlr knockout mouse model. Proc Natl Acad Sci U S A. 2001; 98: 7946–7951.
- ↵Dansky HM, Ono JG, Dansky AH, Wittkowski KM, Breslow JL. A novel method to detect gene interactions determining atherosclerosis susceptibility in apoE-deficient mice. Circulation. 2001; 104: II–222.Abstract.