Association of Genetically Enhanced Lipoprotein Lipase–Mediated Lipolysis and Low-Density Lipoprotein Cholesterol–Lowering Alleles With Risk of Coronary Disease and Type 2 Diabetes

Key Points Question Are genetically determined differences in lipoprotein lipase (LPL)–mediated lipolysis and low-density lipoprotein cholesterol (LDL-C)–lowering pathways independently associated with risk of coronary disease and diabetes? Findings In this genetic association study including 392 220 people, triglyceride-lowering alleles in LPL or its inhibitor ANGPTL4 were associated with lower risk of coronary artery disease and type 2 diabetes in a consistent fashion across quantiles of the population distribution of LDL-C–lowering alleles. For a given genetic difference in LDL-C, the association with lower risk of coronary disease conveyed by rare loss-of-function variants in ANGPTL3, which are associated with lower LDL-C levels and enhanced LPL lipolysis, was greater than that conveyed by other LDL-C–lowering genetic mechanisms. Meaning LPL-mediated lipolysis and LDL-C–lowering mechanisms independently contribute to the risk of coronary disease and diabetes, which supports the development of LPL-enhancing agents for use in the context of LDL-C–lowering therapy.


Factorial genetic analysis
Factorial genetic analyses (eFigure 1B) were conducted in each of the UK Biobank, EPIC-InterAct or EPIC-Norfolk studies separately and then results were combined using inverse variance-weighted fixed-effect meta-analysis.
We constructed two independent LPL and LDL-C weighted genetic risk scores with two distinct goals: (1) to overcome the weak individual associations of genetic variants with lipid levels and disease risk and (2) to "naturally-randomize" participants into approximately equally sized groups, which ensures the greatest statistical power for these individual-level analyses and is akin to a factorial randomized controlled trial design.
We constructed these genetic scores to estimate the combined and independent association of triglyceride-lowering LPL-alleles and of LDL-C lowering polymorphisms at 58 genetic loci with (a) circulating lipid levels and (b) the risk of coronary artery disease and type 2 diabetes (eFigure 1B). We selected six genetic variants for inclusion in the LPL genetic score that were previously reported to be independently and strongly associated with triglyceride levels in analyses of the Global Lipids Genetics Consortium. 17 All genetic variants satisfied these criteria: (1) were in the LPL gene or within 10 kb of the gene; (2) were independently and strongly associated with triglyceride levels in conditional analyses of the Global Lipids Genetics Consortium with p < 5 x 10 -8 . In parallel, we built a LDL-C lowering genetic risk score using 58 genetic variants at 58 independent genetic loci reported by the Global Lipids Genetics Consortium 12 to be strongly and independently associated with LDL-C levels. All genetic variants satisfied these criteria: (1) were over 500 kb away from each other and had no or negligible linkage disequilibrium (R 2 < 0.01); (2) the genetic regions were associated with LDL-C levels (p < 5 x 10 -8 ) in the Global Lipids Genetics Consortium analysis of up to 188,577 individuals.
For each participant and each genetic variant, we weighted the number of effect alleles (i.e. the triglyceride-lowering allele for LPL variants or the LDL-C lowering allele for the 58 LDL-C associated variants) for the effect on the respective lipid trait expressed in standardised units. We then dichotomised each score by dividing people in a group below or equal to the median and above the median value of the weighted score. Because polymorphisms included in genetic scores are inherited approximately randomly at the time of conception in a process known as "Mendelian randomisation", 18 and inherited approximately independently of the other polymorphisms included in the genetic score, the number of lipid lowering alleles that a person inherits for each genetic score should also be random. Therefore, partitioning the population into two groups should "naturally randomise" the population into two approximately equal groups with different genetically-determined lipid levels.
The dichotomised LPL and LDL-C genetic risk scores were used to naturally randomise participants into 4 groups: (1) reference, (2) genetically-lower triglycerides via LPL-alleles, (3) genetically-lower LDL-C via alleles at 58 independent genetic loci, or (4) both genetically-lower triglycerides via LPL-alleles and genetically-lower LDL-C via the 58 genetic loci (referred to as the group "naturally-randomised to both genetic exposures" for simplicity). The reference group included people below or equal to the median of both lipidlowering genetic scores. The group "genetically-lower triglycerides via LPL-alleles" included people above the median for the triglyceride-lowering LPL score, but below or equal to the median for the LDL-C lowering score. The group "genetically-lower LDL-C" included people below or equal to the median for the triglyceride-lowering LPL score, but above the median for the LDL-C lowering score. The group "naturally-randomised to both genetic exposures" included people above the median for both scores.
Using the four "naturally randomised" groups constructed as described above, the effects of each group relative to the reference group were estimated using linear regression for LDL-C and triglyceride levels, while the association with coronary artery disease and type 2 diabetes was estimated using logistic regression (for combined prevalent and incident outcomes, i.e. in UK Biobank and EPIC-Norfolk) or Cox proportional hazards models (for incident events, i.e. in the EPIC-InterAct study). All analyses were adjusted for age, sex and the first four genetic principal components.

Stratified genetic analyses
In stratified genetic analyses (eFigure 1C), we investigated the association of LPLgenetic variants with type 2 diabetes and coronary artery disease in strata of the population distribution of LDL-C lowering genetic variants. These included variants at HMGCR (encoding the target of statins), NPC1L1 (encoding the target of ezetimibe) and PCSK9 (encoding the target of PCSK9 inhibitors), the 58-variant genetic score and the 22-variant genetic score (after excluding variants associated with triglyceride levels). For each of these genes, we used sets of previously published LDL-C lowering genetic variants which were shown by Ference et al. to be strongly associated with lower LDL-C levels and lower coronary disease risk in previous genetic analyses. 19,20 We used six approximately independent genetic variants at the HMGCR locus, five approximately independent genetic variants at the NPC1L1 locus and seven approximately independent genetic variants at the PCSK9 locus. 19,20 We used these genetic variants to partition the population in two groups below or above the median of LDL-C lowering alleles (weighted for their association with LDL-C) at each locus or at the 58 or 22 loci. Additional analyses were conducted in quintiles of the 58-variant LDL-C lowering genetic score. People above the median (or in higher quinitiles) can be thought of as a group of individuals naturally randomised to lower LDL-C levels due to genetic variants at HMGCR, NPC1L1 or PCKS9 or the 58 loci, respectively, serving as a proxy for treatment with the corresponding LDL-C lowering drug or general reduction of LDL-C levels via multiple mechanisms. Within each of these resulting groups, we then estimated the associations of the six triglyceride lowering alleles at LPL with type 2 diabetes and coronary artery disease. We combined individual LPL genetic variant estimates using a weighted generalised linear regression method that accounts for the correlation between genetic variants. 21 eMethods 2. Checks of the quality of genetic data A number of quality control procedures were used to ensure the quality of genetic data and genetic analyses presented here.
In UK Biobank, EPIC-Norfolk and the Illumina Core-Exome-genotyped subset of EPIC-InterAct, the six LPL genetic were directly genotyped with high-quality using genome-wide genotyping arrays. In the Illumina 660w quad genotyped subset of EPIC-InterAct, rs10096633 was directly genotyped and the other five genetic variants were imputed with minimum imputation accuracy info score of 0.91 (with a score of 1 indicating direct genotyping or perfect imputation). Genotyping in these studies underwent a number of quality control procedures including (a) routine quality checks carried out during the process of sample retrieval, DNA extraction, and genotype calling; (b) checks for genotype batch effects, plate effects, departures from Hardy-Weinberg equilibrium, sex effects, array effects, and discordance across control replicates; (c) individual and genetic variant call rate filters.
Given that UK Biobank was the largest study included in the analysis and that genetic data on close to 500,000 individuals have been recently released, we performed additional checks of the quality of data in addition to those implemented by the UK Biobank team (described in details by Bycroft and colleagues 7 ). Firstly, 58 out of 82 genetic variants included in the analysis were directly genotyped in UK Biobank and cleared all pre-release quality control filters 7 as well as a further filter for >95% call rate. The remaining genetic variants were all imputed using the Haplotype Reference Consortium. Among the 89 genetic variants, the median imputation accuracy info score was 1 (with a score of 1 indicating direct genotyping or perfect imputation) and the median among 24 imputed genetic variants was 1 (minimum score 0.93), indicating excellent imputation. Because the 58 directly-genotyped genetic variants were also imputed using the Haplotype Reference Consortium, we compared the minor allele frequency in the genotyped and imputed data, finding near identical frequencies (correlation coefficient = 1.00).
For all genetic variants included in the analysis, we automatically aligned the effect allele to be coded as the lipid lowering allele (eTable 2), using automated scripts. The frequency of the coded effect allele was near identical in the UK Biobank, EPIC-Norfolk and EPIC-InterAct studies (correlation coefficients > 0.9987 for each pairwise comparison) and corresponded to what reported in external reference data.
In the meta-analysis of EPIC-Norfolk, EPIC-InterAct and UK Biobank results used for factorial analyses, we observed a high degree of consistency of estimates from the different studies (see Inset Table).
Inset Table. Consistency of estimates of associations of genetic exposures with lipid traits and cardio-metabolic outcomes in factorial genetic analyses presented in this study. The I 2 and p-value for heterogeneity were used to estimate possible heterogeneity.

UK Biobank
UK Biobank is a population-based cohort of over 500,000 people aged between 40-69 years who were recruited in 2006-2010 from several centres across the United Kingdom. 3 Data from UK Biobank contributed to the analyses of the associations with cardio-metabolic risk factors, type 2 diabetes and coronary artery disease. Waist and hip circumference were measured from participants using a Seca 200cm tape measure, height was measured using a Seca 240cm measure, while weight for the measurement of body mass index (BMI) was collected using a Tanita BC418MA body composition analyser. Type 2 diabetes was defined on the basis of self-reported physician diagnosis at nurse interview or digital questionnaire, age at diagnosis > 36 years, use of oral anti-diabetic medications and electronic health records. 4 Coronary artery disease was defined as either myocardial infarction or coronary disease documented in the participant's medical history at the time of enrolment by a trained nurse or hospitalisation or death involving acute myocardial infarction or its complications (i.e. International Statistical Classification of Diseases and Related Health Problems codes I21, I22 or I23), similar to what previously described. 5,6 Participant characteristics 3 and genotyping methods 7 have been reported in detail elsewhere. We describe the details of the quality checks of genetic data in eNote 3. Participant characteristics are summarised in Table  1.

EPIC-Norfolk cohort study
EPIC-Norfolk is a prospective cohort study of over 20,000 individuals aged between 40 and 79 and living in the Norfolk county in the United Kingdom at recruitment. 8 EPIC-Norfolk is a constituent cohort of the European Prospective Investigation of Cancer (EPIC). 2 Data from EPIC-Norfolk contributed to factorial and stratified genetic analyses. Coronary artery disease was defined as either self-reported myocardial infarction at baseline or incident ischemic heart disease defined by International Statistical Classification of Diseases and Related Health Problems codes 410-414 (ICD9), or I20-I25 (ICD10). Diabetes was defined as either self-reported diabetes at baseline or incident diabetes defined by codes 250 (ICD9), or E10-E14 (ICD10). Participant characteristics and genotyping methods have been previously reported 9 and are summarised in Table 1.

DIAbetes Genetics Replication And Meta-analysis (DIAGRAM) Consortium
Data on type 2 diabetes has been contributed by the DIAGRAM 10  Rare loss-of-function alleles in the LPL-inhibitor ANGPTL3 are associated with lower LDL-C and triglyceride levels, [22][23][24] offering a unique genetic model for the combined reduction of LDL-C levels and enhancement of LPL-mediated lipolysis. Genetic studies and clinical trials show that different LDL-C-lowering mechanisms protect against coronary disease with a mechanism-independent log-linear relationship (i.e. the "LDL-C paradigm"). 19,25,26 If the protective effect of ANGPTL3 variants is only via LDL-C reduction, one would expect their association to be the same as that of LDL-C lowering variants in other genes, for a given genetic difference in LDL-C levels. We investigated this hypothesis by meta-analyzing and modelling data from previously published genetic studies about the association of rare loss-of-function variants of ANGPTL3 with LDL-C and coronary disease risk. 23,24 First, we used results from Dewey and colleagues as estimates of the association of rare loss-of-function variants in ANGPTL3 with LDL-C, i.e. 0.23 SD lower LDL-C (~0.23 mmol/L or 9 mg/dL). 23 Second, we estimated the association with coronary artery disease, for a 0.23 SD genetically-lower LDL-C, of LDL-C lowering variants at HMGCR, NPC1L1, PCSK9 or the 58 LDL-C associated loci using data from UK Biobank and CARDIoGRAMplusC4D. Third, we estimated the association with coronary artery disease, for a 0.23 SD genetically-lower LDL-C, of rare loss-of-function variants in ANGPTL3 by meta-analyzing genetic association studies including up to 58,399 cases and 305,796 controls (eFigure 8). 23,24 Fourth, we tested for heterogeneity between the estimate of the 58 LDL-C lowering alleles and that of ANGPTL3 variants, showing evidence of heterogeneity (eFigure 7). We conducted a number of sensitivity analyses using different estimates for the LDL-C lowering alleles and ANGPTL3 variants (eTable 8). For comparison, we show the consistency of estimates for variants at HMGCR, NPC1L1, PCSK9 with those for the 58 variant LDL-C score (eFigure 7). eAppendix 3. Association of a rare loss-of-function variant in APOC3 with cardiometabolic disease outcomes in UK Biobank Drugs that inhibit APOC3, an inhibitor of LPL-mediated lipolysis, are in early clinical development for the treatment of dyslipidemia. 27,28 Rare loss-of-function variants in the encoding gene have been used as genetic model to study the likely consequences of pharmacological APOC3 inhibition. 29,30 These rare variants are imperfectly captured by array genotyping, such that only one of the four variants driving the reported associations was captured by direct genotyping in UK Biobank (rs147210663, p.Ala43Thr, an experimentallyvalidated loss-of-function variant 31 ), but was not available in InterAct or EPIC-Norfolk. Nonetheless, we sought to estimate the associations with cardio-metabolic disease outcomes of this variant in UK Biobank. In 351,285 people with available genotypes, 279 carried the variant (carrier frequency 0.08%). While the carriers had lower risk of type 2 diabetes (odds ratio per copy of the rare variant rs147210663-A allele, 0.78; 95% confidence interval, 0.44-1.36; p=0.38) and coronary disease (odds ratio per copy of the rs147210663-A allele, 0.90; 95% confidence interval, 0.52-1.55; p=0.70) compared to non-carriers, the difference was not statistically significant. Therefore, it was not possible to meaningfully estimate the association of rare loss-of-function variants of APOC3 in strata of the population distribution of LDL-C lowering alleles. Large-scale sequencing studies of the APOC3 gene will be required to estimate this association. Studies investigating the genetic relationship between triglyceride levels and risk of type 2 diabetes have yielded conflicting results. [32][33][34][35] In a comprehensive Mendelian randomization study, White et al. 35 have estimated the genetic association between triglyceride and diabetes, using 140 triglyceride-lowering genetic variants at multiple loci accounting for possible pleiotropic effects by using different methods, including univariate, multivariate and Egger-MR Mendelian randomization analyses. They found inconsistent results between methods, with Egger-MR (a method that is robust to directional pleiotropy) estimates being consistent with a risk-increasing association for triglyceride-lowering alleles, while the two other methods showed no associations. 35 In this study, we observed associations in a protective direction between triglyceride-lowering alleles at LPL and diabetes risk. We asked whether this association was consistent with estimates of the general genetic relationship between triglycerides and diabetes and tested for heterogeneity between our estimates and those from White and colleagues (eTable 6). We found evidence of heterogeneity, suggesting that the protective association at LPL is specific to this gene/pathway.   The figure shows associations with lipid traits expressed in standardized units for each group compared to the reference group. Data on lipid traits were from the EPIC-Norfolk study and the EPIC-InterAct study subcohort. Median values and interquartile ranges for lipid levels are from the EPIC-Norfolk study. Abbreviations: N, number of participants; CI, confidence interval; LDL-C, low-density lipoprotein cholesterol; LPL, lipoprotein lipase; SD, standard deviation; IQR, interquartile range.    Abbreviations: OR, odds ratio; CI, confidence interval; LPL, lipoprotein lipase. a PubMed manuscript ID. b Inverse variance weighted Mendelian randomisation is a primary analysis method in Mendelian randomisation analyses; multi-variable Mendelian randomisation is an analysis method that adjusts for estimates on other traits (i.e. HDL and LDL cholesterol in this case); Egger Mendelian randomisation is a sensitivity analysis method that is robust to directional pleiotropy. c Comparison between the estimate for LPL alleles from this study (reference group) and each of the two estimates from White and colleagues using 140 triglyceride-lowering alleles from multiple genetic loci. eTable 7. Sensitivity analysis of the association between triglyceride-lowering LPL alleles and risk of coronary artery disease and type 2 diabetes in people above or below the median of the population distribution of 22 LDL-C-lowering variants associated with LDL-C but not triglyceride levels Estimates were nearly identical to those obtained with all 58 LDL-C genetic variants ( Figure  2A). Abbreviations: OR, odds ratio; CI, confidence interval; LDL-C, low-density lipoprotein cholesterol; IVW, inverse variance weighted method; MR, Mendelian randomization. a Odds ratio for coronary artery disease per 0.23 SD genetically-lower LDL-C. b Heterogeneity p-value for comparison of effect estimates between ANGPTL3 variants and LDL-C lowering score analysis. c Sensitivity analysis excluding estimates from PennCath study 24 to account for any possible overlap with the Penn Medicine Biobank. 23 d Estimate from UK Biobank and CARDIoGRAMplusC4D of the association of 22 variants associated with LDL-C (p<5×10 -08 ) but not triglyceride (p>0.05) levels in GLGC. 12 e Estimate from White and colleagues of the association with coronary disease of 130 single-nucleotide polymorphisms associated with LDL-C -inverse variance weighted method. 35 f Estimate from White and colleagues of the association with coronary disease of 130 single-nucleotide polymorphisms associated with LDL-C -multivariable Mendelian randomization method (adjusted for HDL-C and triglycerides). 35 g Estimate from White and colleagues of the association with coronary disease of 130 single-nucleotide polymorphisms associated with LDL-C -Egger Mendelian randomization method. 35