Six new loci associated with blood low-density lipoprotein cholesterol, high-density lipoprotein cholesterol or triglycerides in humans

Abstract

Blood concentrations of lipoproteins and lipids are heritable¹ risk factors for cardiovascular disease^2,3. Using genome-wide association data from three studies (n = 8,816 that included 2,758 individuals from the Diabetes Genetics Initiative specific to the current paper as well as 1,874 individuals from the FUSION study of type 2 diabetes and 4,184 individuals from the SardiNIA study of aging-associated variables reported in a companion paper in this issue⁴) and targeted replication association analyses in up to 18,554 independent participants, we show that common SNPs at 18 loci are reproducibly associated with concentrations of low-density lipoprotein (LDL) cholesterol, high-density lipoprotein (HDL) cholesterol, and/or triglycerides. Six of these loci are new (P < 5 × 10⁻⁸ for each new locus). Of the six newly identified chromosomal regions, two were associated with LDL cholesterol (1p13 near CELSR2, PSRC1 and SORT1 and 19p13 near CILP2 and PBX4), one with HDL cholesterol (1q42 in GALNT2) and five with triglycerides (7q11 near TBL2 and MLXIPL, 8q24 near TRIB1, 1q42 in GALNT2, 19p13 near CILP2 and PBX4 and 1p31 near ANGPTL3). At 1p13, the LDL-associated SNP was also strongly correlated with CELSR2, PSRC1, and SORT1 transcript levels in human liver, and a proxy for this SNP was recently shown to affect risk for coronary artery disease⁵. Understanding the molecular, cellular and clinical consequences of the newly identified loci may inform therapy and clinical care.

We recently conducted the Diabetes Genetics Initiative (DGI) genome-wide association study for type 2 diabetes and 18 other traits, including blood lipoprotein and lipid concentrations⁶. Here, we focus on replication analyses related to three traits—concentrations of LDL cholesterol, HDL cholesterol and triglycerides. In DGI, we analyzed the association of 389,878 markers with blood lipoproteins and lipids in 2,758 individuals. From these results, we selected an initial 196 SNPs for replication on the basis of the strength of statistical evidence. We then combined the DGI results with those from two other genome-wide association studies—the Finland–United States Investigation of NIDDM Genetics (FUSION) and the SardiNIA Study of Aging (see companion manuscript for meta-analytic methods⁴)—and selected an additional 30 SNPs for replication on the basis of the combined evidence (see Supplementary Fig. 1 online for study design). The 226 SNPs selected for replication were tested in up to 18,554 separate participants from three studies, (Table 1). Statistical evidence from the DGI genome-wide association study and the three replication studies was summarized using a variance-weighted meta-analysis⁷. We pre-specified P < 5 × 10⁻⁸ as the statistical significance threshold for each new locus.

Table 1. Sample characteristics.

Characteristic	GWAS: DGI n = 2,758	Replication cohort 1: MDC-CC n = 5,519	Replication cohort 2: FINRISK97 n = 7,940	Replication cohort 3: NORDIL n = 5,095
Ascertainment scheme	Cases with type 2 diabetes, controls free of diabetes	Community-based, prospective cohort	Population-based, prospective cohort	Randomized clinical trial of antihypertensive therapy
Mean age (y)	61.5 ± 10.5	57.6 ± 5.9	50.0 ± 13.4	60.3 ± 6.6
Female gender (%)	50.8	58.6	50.3	50.2
Body mass index (kg/m2)	27.6 ± 4.2	25.8 ± 3.9	26.7 ± 4.5	28.1 ± 4.4
Total cholesterol (mg/dl)	227 ± 44	239 ± 42	214 ± 21	244 ± 45
Low-density lipoprotein cholesterol (mg/dl)	151 ± 40	161 ± 38	135 ± 36	160 ± 43
High-density lipoprotein cholesterol (mg/dl)	50 ± 13	53 ± 14	54 ± 14	53 ± 21
Triglycerides (mg/dl)	146 ± 104	122 ± 71	133 ± 93	159 ± 108
Systolic blood pressure (mmHg)	143 ± 22	142 ± 19	138 ± 41	173 ± 19
Individuals with diabetes mellitus (%)	1,359 (49.3)	464 (8.4)	280 (3.5)	442 (8.7)

Values with ‘±’ are means ± s.d. To convert values for cholesterol to millimoles per liter, multiply by 0.02586. To convert the values for triglycerides to millimoles per liter, multiply by 0.01129. The body-mass index is the weight in kilograms divided by the square of the height in meters. GWAS, genome-wide association study; DGI, Diabetes Genetics Initiative; MDC-CC, Malmö Diet and Cancer Study—Cardiovascular Cohort; NORDIL, Nordic Diltiazem Study.

At 18 independent genomic loci, common DNA sequence variants were reproducibly related to at least one blood lipoprotein or lipid trait (Table 2). Previous studies have identified nine loci with compelling evidence for association between common variants and lipoprotein or lipid concentrations (those of APOB, APOE-APOC1-APOC4-APOC2, PCSK9, ABCA1, APOA1-APOC3-APOA4-APOA5, CETP, LIPC, LPL and ANGPTL4)^8-16, and we confirmed eight of these loci (Table 2). Additionally, we recently identified GCKR⁶ as a locus associated with triglyceride concentration.

Table 2. Genetic loci where common SNPs are associated with blood lipoproteins or lipids.

Trait	Unique loci	SNP	Locus	SNP type	Nearest gene(s)	Allele (frequency)a	Combined analysis: DGI GWAS & ≤3 replications Pb	Effect size (s.e.m.)c
Newly identified loci
LDL cholesterol	1	rs646776 or rs599838d	1p13	Intergenic	CELSR2, PSRC1,SORT1	C (0.24) G (0.24)	5 × 10−42	−0.16 (0.01)
LDL cholesterol	2	rs16996148	19p13	Intergenic	CILP2,PBX4	T (0.10)	3 × 10−8	−0.10 (0.02)
Triglycerides		rs16996148	19p13	Intergenic	CILP2,PBX4	T (0.10)	4 × 10−9	−0.10 (0.02)
HDL cholesterol	3	rs4846914	1q42	Intronic	GALNT2	G (0.40)	2 × 10−13	−0.07 (0.01)
Triglycerides		rs4846914	1q42	Intronic	GALNT2	G (0.40)	7 × 10−15	0.08 (0.01)
Triglycerides	4	rs17145738	7q11	Intergenic	BCL7B,TBL2,MLXIPL	T (0.13)	7 × 10−22	−0.14 (0.02)
Triglycerides	5	rs17321515	8q24	3′ downstream	TRIB1	G (0.49)	4 × 10−17	−0.08 (0.01)
Triglycerides	6	rs12130333	1p31	Intergenic	ANGPTL3, DOCK7,ATG4C	T (0.22)	2 × 10−8	−0.11 (0.02)
Loci with prior evidence
LDL cholesterol	7	rs693	2p24	Coding	APOB	A (0.48)	1 × 10−21	0.12 (0.01)
Triglycerides		rs693	2p24	Coding	APOB	A (0.48)	2 × 10−7	0.08 (0.02)
LDL cholesterol	8	rs4420638	19q13	5′ upstream	APOE-C1-C4-C2	G (0.20)	1 × 10−60	0.19 (0.02)
LDL cholesterol	9	rs12654264	5q13	Intronic	HMGCR	T (0.39)	1 × 10−20	0.10 (0.01)
LDL cholesterol	10	rs6511720	19p13	Intronic	LDLR	T (0.10)	2 × 10−51	−0.26 (0.02)
LDL cholesterol	11	rs11591147e	1p32	Coding	PCSK9	T (0.01)	2 × 10−44	−0.47 (0.03)
HDL cholesterol	12	rs3890182	9q31	Intronic	ABCA1	A (0.13)	3 × 10−10	−0.10 (0.02)
HDL cholesterol	13	rs28927680f	11q23	5′ upstream	APOA1-C3-A4-A5, ZNF259, BUD13	G (0.07)	2 ×10−5	−0.13 (0.03)
Triglycerides		rs28927680	11q23	5′ upstream	APOA1-C3-A4-A5, ZNF259, BUD13	G (0.07)	2 × 10−17	0.26 (0.03)
HDL cholesterol	14	rs1800775	16q13	5′ upstream	CETP	C (0.51)	1 × 10−73	−0.18 (0.01)
HDL cholesterol	15	rs1800588	15q21	5′ upstream	LIPC	T (0.21)	2 × 10−32	0.14 (0.01)
HDL cholesterol	16	rs2156552	18q21	Intergenic	LIPG, ACAA2	A (0.18)	2 × 10−7	−0.07 (0.01)
HDL cholesterol	17	rs328	8p21	Coding	LPL	G (0.09)	9 × 10−23	0.17 (0.02)
Triglycerides		rs328	8p21	Coding	LPL	G (0.09)	2 × 10−28	−0.19 (0.02)
Triglycerides	18	rs780094g	2p23	Intronic	GCKR	T (0.34)	3 × 10−14	0.13 (0.02)

^aAlleles for the SNP on the forward strand of human genome reference sequence (National Center for Biotechnology Information Build 35) were modeled. Allele frequency in the MDC-CC replication sample is presented, except in the case of rs599838, where the frequency in NORDIL is shown.

^bVariance-weighted meta-analysis performed using data from up to four samples (DGI, MDC-CC, FINRISK and NORDIL) as described in Methods.

^cEffect size shown is beta-coefficient (β), which represents the proportion of 1 s.d. change (in standardized blood lipid residual with mean = 0 and s.d. = 1 after adjustment for age, age², gender and diabetes status) per copy of the allele modeled.

^dData shown are for a combination of rs646776 results from MDC-CC and FINRISK97 and results for a proxy (rs599839, r² = 0.89 between two SNPs in HapMap population of European ancestry) from DGI and NORDIL. Data for each SNP are separately shown in Table 3.

^ePCSK9 SNP rs11206510 (17% minor allele frequency) was associated with LDL cholesterol in the combined analysis of DGI/FUSION/SardiNIA (P = 7.5 × 10⁻⁶); however, to confirm this locus, we selected rs11591147 for replication because of prior reports suggesting a very strong effect on LDL cholesterol²⁶.

^frs28927680 is highly correlated with APOA1-APOC3-APOA4-APOA5 coding SNP rs3133506 (r² = 0.98 in MDC-CC); SNP rs3133506 (also known as 56C>G) has been reproducibly related to HDL cholesterol and triglyceride concentrations.

^gData in this row on GCKR has been previously reported by our group⁶ and is presented here for completeness.

Prior work has provided suggestive but not definitive evidence for common variants at three loci (LDLR, HMGCR and LIPG)^17,18. We found compelling evidence for common variants at each of the three loci (Table 2).

Six of the 18 loci have not been previously reported to relate to lipoprotein or lipid traits in humans. For these six newly identified loci, the statistical evidence for association was robust, ranging from P = 3 × 10⁻⁸ to P = 5 × 10⁻⁴² in the combined analysis of the DGI genome-wide association study and three replication cohorts (Table 2).

For LDL cholesterol, we identified two new loci and confirmed five loci with prior evidence (Table 3). The first new locus for LDL cholesterol is located on chromosome 1p13. SNPs rs599839 and rs646776 were robustly associated with LDL cholesterol (combined P = 3 × 10⁻²¹ and 3 × 10⁻²⁹, respectively; Table 3). Both SNPs are located in a 97-kb region of linkage disequilibrium containing four genes—CELSR2, PSRC1, MYBPHL and SORT1. The two SNPs are highly correlated with one another (r² = 0.89 in the HapMap population of European ancestry). Each copy of the minor allele at either SNP (24% frequency) decreased LDL cholesterol concentrations by ~5-8 mg/dl (Supplementary Table 1 online).

Table 3. Genetic loci where common SNPs are associated with blood low-density lipoprotein cholesterol.

					Combined analysis: DGI GWAS & ≤3 replications	GWAS: DGI	Replication 1: MDC-CC	Replication 2: FINRISK 97	Replication 3: NORDIL
					P	P	P	P	P
					β	β	β	β	β
SNP(s)	Locus	SNP type	Nearest gene(s)	Allele (frequency)a	(s.e.m.)b	(s.e.m.)	(s.e.m.)	(s.e.m.)	(s.e.m.)
Newly identified loci
rs646776c	1p13	Intergenic	CELSR2, PSRC1, SORT1	C (0.24)	3 × 10−29	8 × 10−8d	2 × 10−10	1 × 10−13	-
					−0.16 (0.01)	−0.18 (.0.03)	−0.15 (0.02)	0.15 (0.02)
rs599839c				G (0.24)	3 × 10−21	9 × 10−8	-	-	1 × 10−14
					−0.18 (0.02)	−0.18 (0.03)			0.18 (0.02)
rs16996148	19p13	Intergenic	CILP2, PBX4	T (0.10)	3 × 10−8	0.04	0.01	0.12	3 × 10−6
					−0.10 (0.02)	−0.10 (0.05)	−0.09 (0.03)	−0.05 (0.03)	−0.05 (0.03)
Loci with prior evidence
rs693	2p24	Coding	APOB	A (0.48)	1 × 10−21	7 × 10−7	2 × 10−11	-	1 × 10−6
					0.12 (0.01)	0.14 (0.03)	0.13 (0.02)		0.09 (0.02)
rs4420638	19q13	5′ upstream	APOE-C1-C4-C2	G (0.20)	1 × 10−60	3 × 10−13	3 × 10−21	8 × 10−19	6 × 10−13
					0.19 (0.02)	0.25 (0.03)	0.23 (0.02)	0.16 (0.02)	0.18 (0.03)
rs12654264	5q13	Intronic	HMGCR	T (0.39)	1 × 10−20	0.0004	0.002	9 × 10−11	5 × 10−7
					0.10 (0.01)	0.10 (0.03)	0.06 (0.02)	0.11 (0.02)	0.11 (0.02)
rs6511720	19p13	Intronic	LDLR	T (0.10)	2 × 10−51	9 × 10−7d	8 × 10−14	9 × 10−22	3 × 10−13
					−0.26 (0.02)	−0.27 (0.05)	−0.25 (0.03)	−0.26 (0.03)	−0.25 (0.03)
rs11591147e	1p32	Coding	PCSK9	T (0.01)	2 × 10−44	-	7 × 10−7	2 × 10−31	8 × 10−9
					−0.47 (0.03)		−0.45 (0.09)	−0.48 (0.04)	−0.47 (0.08)

A dash (-) indicates that we did not genotype that SNP in the sample. Beta-coefficient (β) represents the proportion of 1 s.d. change in standardized LDL cholesterol residual (mean = 0, s.d. = 1 after adjustment for age, age², gender, and diabetes status) per copy of the allele modeled.

^aAlleles for the SNP on the forward strand of human genome reference sequence (National Center for Biotechnology Information Build 35) were modeled. Allele frequency in the MDC-CC sample is presented, except for rs599839, where allele frequency in DGI is presented.

^bVariance-weighted meta-analysis performed using data from up to four samples (DGI, MDC-CC, FINRISK and NORDIL), as described in Methods.

^crs646776 and rs599839 are in high linkage disequilibrium with r² = 0.89 in the HapMap population of European ancestry.

^dAssociation data from imputed SNP genotypes.

^ePCSK9 SNP rs11206510 (17% minor allele frequency) was associated with LDL cholesterol in the combined analysis of DGI/FUSION/SardiNIA (P = 7.5 × 10⁻⁶); however, to confirm this locus, we elected to take forward the previously reported SNP rs11591147 for replication because of its reported strong effect on LDL cholesterol²⁶.

The second new locus for LDL cholesterol is located on chromosome 19p13 in an intergenic region between CILP2 and PBX4. SNP rs16996148 replicated for association with LDL cholesterol (combined P = 3 × 10⁻⁸, Table 3). Two copies of the minor allele at SNP rs16996148 decreased LDL cholesterol concentrations by ~16 mg/dl (Supplementary Table 1).

Besides the 1p13 and 19p13 loci, we confirmed five loci with prior evidence for association with LDL cholesterol concentrations (APOB, APOE-APOC1-APOC4-APOC2, LDLR, HMGCR and PCSK9, Table 3). We found that an intronic LDLR SNP strongly related to LDL cholesterol. In the cardiovascular cohort of the Malmö Diet and Cancer Study, LDL cholesterol values varied by ~7 mg/dl per copy of the minor allele at the LDLR SNP (combined P = 2 × 10⁻⁵¹, Table 3 and Supplementary Table 1). Similarly, an intronic SNP at HMGCR was associated with LDL cholesterol (combined P = 1 × 10⁻²⁰, Table 3).

For HDL cholesterol, we identified one new locus and confirmed six loci for which there was prior evidence of association (Table 4). The new locus for HDL cholesterol is located at 1q42 in an intron of GALNT2 (SNP rs4846914, combined P = 2 × 10⁻¹³ for association, Table 4). Each copy of the minor allele decreased HDL cholesterol concentrations by ~1.5 mg/dl (Supplementary Table 1). In addition, we confirmed six loci with prior evidence (ABCA1, APOA1-APOC3-APOA4-APOA5, CETP, LIPC, LIPG and LPL).

Table 4. Genetic loci where common SNPs are associated with blood high-density lipoprotein cholesterol.

					Combined analysis; GWAS & ≤3 Replications	GWAS: DGI	Replication 1: MDC-CC	Replication 2: FINRISK97	Replication 3: NORDIL
					P	P	P	P	P
					β	β	β	β	β
SNP	Locus	SNP type	Nearest gene(s)	Allele (frequency)a	(s.e.m.)b	(s.e.m.)	(s.e.m.)	(s.e.m.)	(s.e.m.)
Newly identified locus
rs4846914	1q42	Intronic	GALNT2	G (0.40)	2 × 10−13	3 × 10−4	1 × 10−6	1 × 10−4	0.01
					−0.07 (0.01)	−0.10 (0.03)	−0.10 (0.02)	−0.06 (0.02)	−0.05 (0.02)
Loci with prior evidence
rs3890182	9q31	Intronic	ABCA1	A (0.13)	3 × 10−10	3 × 10−5	0.003	0.001	0.02
					−0.10 (0.02)	−0.17 (0.04)	−0.09 (0.03)	−0.09 (0.03)	−0.08 (0.03)
rs28927680c	11q23	5′ upstream	APOA1-C3-A4-A5, ZNF259, BUD13	G (0.07)	2 × 10−5	0.31	6 × 10−6	-	-
					−0.13 (0.03)	−0.06 (0.05)	−017 (0.04)
rs1800775	16q13	5′ upstream	CETP	C (0.51)	1 × 10−73	3 × 10−13	2 × 10−29	1 × 10−23	2 × 10−12
					−0.18 (0.01)	−0.20 (0.03)	−0.22 (0.02)	−0.17 (0.02)	−0.14 (0.02)
rs1800588	15q21	5′ upstream	LIPC	T (0.21)	2 × 10−32	3 × 10−5d	4 × 10−10	8 × 10−17	3 × 10−5
					0.14 (0.01)	0.15 (0.03)	0.15 (0.02)	0.16 (0.02)	0.10 (0.02)
rs2156552	18q21	Intergenic	LIPG, ACAA2	A (0.18)	2 × 10−7	0.02	0.02	0.001	0.03
					−0.07 (0.01)	−0.09 (0.03)	−0.06 (0.03)	−0.07 (0.02)	−0.06 (0.03)
rs328	8p21	Coding	LPL	G (0.09)	9 × 10−23	3 × 10−4d	3 × 10−12	1 × 10−6	6 × 10−5
					0.17 (0.02)	0.17 (0.05)	0.23 (0.03)	0.14 (0.03)	0.14 (0.04)

A dash (-) indicates that we did not genotype that SNP in sample, Beta-coefficient (β) represents the proportion of 1 s.d. change in standardized HDL cholesterol residual (mean = 0, s.d. = 1 after adjustment for age, age², gender and diabetes status) per copy of the allele modeled.

^aAlleles for the SNP on the forward strand of human genome reference sequence (National Center for Biotechnology Information Build 35) were modeled. Allele frequency in the MDC-CC sample is presented.

^bVariance-weighted meta-analysis performed using data from up to four samples (DGI, MDC-CC, FINRISK and NORDIL), as described in Methods.

^crs28927680 is highly correlated with APOA1-APOC3-APOA4-APOA5 coding SNP rs3133506 (r² = 0.98 in MDC-CC); SNP rs3133506 (also known as 56C>G) has been reproducibly related to HDL cholesterol and triglyceride concentrations.

^dAssociation data from imputed SNP genotypes.

For triglycerides, we identified five new loci (Table 5). The five replicated SNPs are located at 7q11 near TBL2 and MLXIPL, 8q24 near TRIB1, 1q42 in GALNT2, 19p13 near CILP2-PBX4 and 1p31 near ANGPTL3 (combined P < 5 × 10⁻⁸ for each SNP, Table 5). Of these, the SNP at 7q11 near TBL2 and MLXIPL had the strongest effect size, with each copy of the minor allele increasing triglyceride concentrations by ~8 mg/dl (Supplementary Table 1). In addition, we confirmed four loci with prior evidence (APOA1-APOC3-APOA4-APOA5, APOB, GCKR and LPL).

Table 5. Genetic loci where common SNPs are associated with blood triglycerides.

					Combined analysis: GWAS & ≤3 Replications	GWAS: DGI	Replication 1: MDC-CC	Replication 2: FINRISK97	Replication 3: NORDIL
					P	P	P	P	P
					β	β	β	β	β
SNP	Locus	SNP type	Nearest gene(s)	Allele (frequency)a	(s.e.m.)b	(s.e.m.)	(s.e.m.)	(s.e.m.)	(s.e.m.)
Newly identified loci
rs17145738	7q11	Intergenic	BCL7B, TBL2, MLXIPL	T (0.13)	7 × 10−22	0.003	3 × 10−8	2 × 10−7	3 × 10−7
					−0.14 (0.02)	−0.12 (0.04)	−0.17 (0.03)	−0.13 (0.02)	−0.16 (0.03)
rs17321515	8q24	3′ downstream	TRIB1	G (0.49)	4 × 10−17	7 × 10−4	1 × 10−5	7 × 10−5	8 × 10−7
					−0.08 (0.01)	−0.10 (0.03)	−0.09 (0.02)	−0.07 (0.02)	−0.10 (0.02)
rs4846914	1q42	Intronic	GALNT2	G (0.40)	7 × 10−15	9 × 10−5	0.001	2 × 10−8	0.01
					0.08 (0.01)	0.11 (0.03)	0.07 (0.02)	0.09 (0.02)	0.05 (0.02)
rs16996148	19p13	Intergenic	CILP2, PBX4	T (0.10)	4 × 10−9	0.05	0.23	2 × 10−5	3 × 10−5
					−0.10 (0.02)	−0.09 (0.05)	−0.04 (0.03)	−0.14 (0.03)	−0.13 (0.03)
rs12130333	1p31	Intergenic	ANGPTL3, DOCK7, ATG4C	T (0.22)	2 × 10−8	0.0006	2 × 10−5	-	-
					−0.11 (0.02)	−0.12 (0.03)	−0.10 (0.02)
Loci with prior evidence
rs28927680c	11q23	5′ upstream	APOA1-C3-A4-A5, ZNF259, BUD13	G (0.07)	2 × 10−17	6 × 10−5	7 × 10−14	-	-
					0.26 (0.03)	0.22 (0.05)	0.29 (0.04)
rs693	2p24	Coding	APOB	A (0.48)	2 × 10−7	7 × 10−4	7 × 10−5	-	-
					0.08 (0.02)	0.09 (0.03)	0.08 (0.02)
rs780094d	2p23	Intronic	GCKR	T (0.34)	3 × 10−14	4 × 10−8	5 × 10−8	-	-
					0.13 (0.02)	0.16 (0.03)	0.11 (0.02)
rs328	8p21	Coding	LPL	G (0.09)	2 × 10−28	4 × 10−7e	9 × 10−9	2 × 10−10	2 × 10−6
					−0.19 (0.02)	−0.24 (0.05)	−0.20 (0.03)	−0.18 (0.03)	−0.16 (0.03)

A dash (-) indicates that we did not genotype that SNP in sample. Beta-coefficient (β) represents the proportion of 1 s.d. change in standardized log triglycerides residual (mean = 0, s.d. = 1 after adjustment for age, age², gender and diabetes status) per copy of the allele modeled.

^aAlleles for the SNP on the forward strand of human genome reference sequence (National Center for Biotechnology Information Build 35) were modeled. Allele frequency in the MDC-CC sample is presented.

^bVariance-weighted meta-analysis performed using data from up to four samples (DGI, MDC-CC, FINRISK and NORDIL), as described in Methods.

^crs28927680 is highly correlated with APOA1-APOC3-APOA4-APOA5 coding SNP rs3133506 (r² = 0.98 in MDC-CC); SNP rs3133506 (also known as 56C>G) has been reproducibly related to triglyceride and HDL cholesterol concentrations.

^dData in this row on GCKR has been previously published⁶ and is presented here for completeness.

^eAssociation data from imputed SNP genotypes.

We observed that SNPs at four of the newly identified loci—19p13 near CILP2 and PBX4, 1q42 in GALNT2, 7q11 near TBL2 and MLXIPL and 8q24 near TRIB1—were associated with multiple lipoprotein or lipid traits (Table 6 and Supplementary Table 2 online). We did not require the associations with the second and/or third trait to meet a genome-wide association threshold of P < 5 × 10⁻⁸. We find these secondary associations to be of interest, as the patterns of association may provide clues to how the locus affects lipoprotein metabolism.

Table 6. Effects of loci on multiple lipoprotein or lipid traits.

SNP	Locus	Nearest gene(s)	Allele (frequency)a	LDL cholesterolb	HDL cholesterol	Triglycerides
Newly identified loci
rs646776	1p13	CELSR2, PSRC1, SORT1	C (0.24)	↑
rs16996148	19p13	CILP2, PBX4, GALNT2	T (0.10)	↓		↓
rs4846914	1q42		G (0.40)		↓	↑
rs17145738	7q11	BCL7B, TBL2, MLXIPL	T (0.13)		↑	↓
rs17321515	8q24	TRIB1	G (0.49)	↓	↑	↓
rs12130333	1p31	ANGPTL3, DOCK7, ATG4C	T (0.22)			↓
Selected loci with prior evidence
rs693	2p24	APOB	A (0.48)	↑		↑
rs28927680	11q23	AP0A1-C3-A4-A5, ZNF259, BUD13	G (0.07)		↓	↑
rs328	8p21	LPL	G (0.09)		↑	↓

We required each new locus to be associated with at least one lipoprotein or lipid trait at a genome-wide association threshold (P ≤ 5 × 10⁻⁸). We did not require the associations with the second and/or third trait to be at this threshold.

^aAlleles for the SNP on the forward strand of human genome reference sequence (National Center for Biotechnology Information Build 35) were modeled. Allele frequency in the MDC-CC replication sample is presented.

^b↑ indicates that the modeled allele increases trait value;↓, lowers trait value. Allele shown in column 4 is modeled. Data derived from a combined analysis of DGI and three replication studies, as described in Table 2.

The minor allele at SNP rs16996148 on 19p13 near CILP2 and PBX4 was associated with lower concentrations of both LDL cholesterol and triglycerides (Table 6). This pattern of association is similar to that of APOB coding SNP rs693, in which a variant allele is associated with both LDL cholesterol and triglycerides in the same direction.

The minor alleles of GALNT2 SNP rs4846914 as well as SNP rs17145738 on 7q11 near TBL2 and MLXIPL were associated with both triglyceride and HDL cholesterol concentrations: rs4846914 is associated with lower HDL concentrations and higher triglyceride concentrations, and rs17145738 is associated with higher HDL concentrations and lower triglyceride concentrations (Table 6). These patterns of association are similar to that of the common LPL nonsense mutation S447X (rs328).

SNP rs17321515 at 8q24 near TRIB1 was strongly associated with triglycerides and was also associated with LDL cholesterol and HDL cholesterol (Table 6). The minor G allele at this SNP was associated with lower triglycerides, lower LDL cholesterol and higher HDL cholesterol. This pattern of association has not been previously described for any lipid-modulating SNP.

Of note, for the 23 associations from 18 common alleles in this study, we found that the effect size of an allele varied inversely with allele frequency (r = −0.49, P = 0.01). For example, lower-frequency alleles, such as the 1% frequency allele at PCSK9, affected LDL cholesterol concentrations by ~0.5 s.d. units, whereas a 48% frequency allele at APOB affected LDL cholesterol by ~0.1 s.d. units (Table 2). Such an inverse relationship is predicted if alleles perturbing physiology are deleterious during evolution, as such alleles would not rise to a high frequency in the population.

Having observed that common variants at 18 loci are convincingly associated with lipoprotein- or lipid-related traits, we next addressed the extent to which these alleles explain inter-individual variability in lipoprotein or lipid concentrations. In the cardiovascular cohort of the Malmö Diet and Cancer Study, after accounting for age, age², gender and diabetes status, we found that, in sum, seven SNPs explained an additional 5.7% of the residual LDL cholesterol variance (Table 3). Meanwhile, seven SNPs explained an additional 5.2% of the residual HDL cholesterol variance (Table 4) and nine SNPs explained an additional 4.5% of the residual triglyceride level variance (Table 5).

Though these common alleles explain an appreciable fraction of variance, it is likely that these values are underestimates of the impact of each validated locus. As nine of the loci with common variants (ABCA1, APOA1-APOC3-APOA4-APOA5, APOB, APOE, CETP, LDLR, LIPC, LPL and PCSK9) have also been shown to cause mendelian syndromes or harbor multiple rare alleles that contribute to trait variation¹⁹, sequencing of each validated locus will be required to discover all common and rare variants and determine the full impact of each locus.

It is not yet clear what the causal variants or even the causal genes are at the new loci. Each of the six associated SNPs is noncoding. The genes nearest to the associated SNPs are annotated in Table 2.

However, the linkage disequilibrium pattern and the genes in the associated intervals suggest functional hypotheses. At 19p13, the variant associated with LDL cholesterol (located between CILP2 and PBX4) is in high linkage disequilibrium with a nonsynonymous coding SNP in the CSPG3 gene encoding neurocan (rs2228603, 329 kb upstream, r² = 0.85 in HapMap population of European ancestry), suggesting that CSPG3 may be the causal gene at the locus. At the 1q42 locus for HDL cholesterol and triglycerides, GALNT2 encodes polypeptide N-acetylgalactosaminyltransferase 2, an enzyme involved in O-linked glycosylation and transfer of N-acetylgalactosamine to the serine or threonine residues on proteins. O-linked glycosylation has a regulatory role for many proteins²⁰. This suggests the hypothesis that enzymatic glycosylation of any of a number of proteins involved in HDL cholesterol and triglyceride metabolism may lead to the observed pattern of association. At the 7q11 locus for triglycerides, the associated interval includes MLXIPL, encoding a transcription factor recently described to connect carbohydrate flux with fatty-acid synthesis in the liver (also called carbohydrate response element binding protein or ChREBP)²¹. Finally, inactivating mutations in ANGPTL3 (encoding angiopoietin-like 3) have already been demonstrated to lead to low triglycerides in mice²².

We next considered one mechanism by which SNPs (and particularly noncoding SNPs) may relate to traits, namely, the regulation of local gene expression. We analyzed the correlation of lipid-associated SNPs with mRNA transcript levels of nearby genes in 60 human liver samples. At five of the six newly identified loci, lipid-associated SNPs showed no effect on expression of local genes (P > 0.05).

However, SNP rs646776 at the 1p13 locus was strongly associated with transcript concentrations of not only a single gene, but three neighboring genes: SORT1 (P = 3 × 10⁻²⁶), CELSR2 (P 2 × 10⁻¹²) and PSRC1 (P = 3 × 10⁻¹²) (Supplementary Fig. 2 online). SNP rs646776 explained 86%, 58%, and 58% of the inter-individual variability in SORT1, CELSR2 and PSRC1 transcript concentrations, respectively. In analyses conditioning on either the CELSR2 or PSRC1 transcript levels, rs646776 remained associated with SORT1 transcript concentration (P = 1 × 10⁻⁵ and 1 × 10⁻⁵, respectively). Conversely, after SORT1 transcript level was accounted for, rs646776 was weakly or not associated with PSRC1 or CELSR2 (P = 0.04 and 0.81, respectively). Overall, our results suggest that variation at the 1p13 interval may have a regional effect on gene expression.

SORT1, or sortilin, functions both as a sorting protein and as a cell-surface receptor, and it is abundant in skeletal muscle and adipocytes^23,24. As a sorting protein, sortilin enables insulin-mediated glucose uptake by catalyzing the biogenesis of insulin-sensitive vesicles that transport the glucose transporter GLUT4 to the plasma membrane. In addition, as a multiligand receptor, sortilin can bind several proteins, including lipoprotein lipase, and potentially facilitate lipoprotein uptake. Overall, these observations suggest a mechanism by which increased sortilin expression seen with the C allele (at SNP rs646776) could lead to lower circulating LDL cholesterol concentrations.

Notably, a proxy for SNP rs646776 at the 1p13 locus, SNP rs599839, was recently reported to affect risk of coronary artery disease⁵. SNP rs599839 was also related to LDL cholesterol (Table 3) in our study, and the allele associated with lower LDL cholesterol (G allele, 24% frequency) was the same as that correlated with lower risk of coronary artery disease (odds ratio 0.78; P = 4.0 × 10⁻⁹)⁵.

As participants in the initial and replication studies were of European ancestry, it remains to be shown whether the new loci will be associated with lipid-related traits in individuals of other ancestries. In a pilot study, we tested whether the six SNPs from the six new loci identified in those of European ancestry would be associated with lipoprotein or lipid traits in a multiethnic sample. We studied 4,259 participants from the Singapore National Health Survey 98 comprising ethnic Chinese, Indians and Malays²⁵. SNPs at two of the six loci (1p13 near CELSR2-PSRC1-SORT1 associated with LDL cholesterol; 7q11 near TBL2-MLXIPL associated with triglycerides) replicated for association in each of the three ethnic groups (Supplementary Table 3 online). Because of well-known differences in linkage disequilibrium structure and allele frequencies across populations of different ancestries, a comprehensive testing of genetic variation at each new locus is needed for each ethnic group.

We have obtained definitive evidence for six new independent loci at which common genetic variation influences one or more lipoprotein or lipid traits. By establishing these loci as relevant to lipoprotein metabolism in humans, we nominate these as high-priority targets for further investigation. Before considering these loci as targets for pharmacological therapy, it will be critical to assess whether causal alleles at each locus affect risk for cardiovascular disease. If alleles are convincingly associated with risk of cardiovascular disease (as has been shown for PCSK9 (ref. ²⁶)), this would give in vivo human proof for the locus as a valid target and support a path forward.

METHODS

Study samples

DGI is a genome-wide association study of type 2 diabetes mellitus and 18 other traits, including blood lipoproteins and lipids. Details of the study design and ascertainment scheme have been recently described⁶. Of the 2,931 participants, 2,758 had DNA samples and data available for at least one lipoprotein or lipid phenotype.

Replication cohort 1 consisted of the Malmö Diet and Cancer Study, a community-based prospective epidemiologic cohort of 28,449 persons recruited for a baseline examination between 1991 and 1996 (ref. ²⁷). From this cohort, 6,103 persons were randomly selected to participate in the cardiovascular cohort, which sought to investigate risk factors for cardiovascular disease. All participants underwent a medical history assessment and a physical examination. Of the participants in the cardiovascular cohort, 5,519 had DNA samples and data available for at least one lipoprotein or lipid phenotype.

Replication cohort 2 consisted of FINRISK97, a population-based cross-sectional survey designed to study the prevalence of cardiovascular risk factors in Finland²⁸. Surveys are conducted every 5 years, and the 1997 survey included 8,389 Finnish men and women aged 25-74. Participants underwent a physical examination and completed a questionnaire regarding cardiovascular risk factors. Of these FINRISK97 participants, 7,940 had DNA samples and data available for at least one lipoprotein or lipid phenotype.

Replication cohort 3 consisted of NORDIL, a randomized controlled trial that compared calcium antagonists with diuretics and beta-blockers on cardiovascular morbidity and mortality in hypertension²⁹. Participants from Norway and Sweden were enrolled between 1992 and 1999. Of the 10,881 participants in this clinical trial, the 5,152 participants from Sweden form the sample for this report. Of these participants, 5,095 had DNA samples and data available from the baseline examination for at least one lipoprotein or lipid phenotype.

All participants in the DGI genome-wide association study and the three replication studies were of self-reported European ancestry.

We attempted to extend our replicated findings to a multiethnic sample from Singapore, the Singapore National Health Survey 98 (NHS98). Singapore NHS98 was an initiative to determine the risk factors for the major noncommunicable diseases in Singapore²⁵. The sampling scheme captured the three major Singaporean ethnic groups—Chinese, Malays and Asian Indians—and has been previously described²⁵. Of the 4,723 individuals who participated, we studied a sample of 4,259 subjects who had DNA samples and data available for lipoprotein or lipid phenotypes.

All participants in the studies provided written informed consent. The individual study protocols were approved by local ethical committees at each participating institution. The conduct of the genome-wide association study was approved by the Massachusetts Institute of Technology Review Board.

Determination of blood lipid concentration

In DGI, the Malmö Diet and Cancer Study, NORDIL and Singapore NHS98, we measured total cholesterol, HDL cholesterol and triglyceride concentrations in fasting blood samples drawn at the baseline examination for each study. In FINRISK97, individuals were instructed to fast for at least 4 h with a mean fasting time of 6 ± 4 h. We measured lipid concentrations according to standard enzymatic methods. LDL cholesterol concentrations were calculated according to Friedewald’s formula, and missing values were assigned to individuals with triglycerides >400 mg/dl.

SNP selection and genotyping

The overall SNP selection procedure is summarized in Supplementary Figure 1. First, we attempted to validate 196 SNPs from a single genome-wide association study for blood lipid traits—the DGI—for which the full association results are publicly available (see URLs section below)⁶. From these association results, we selected 196 SNPs representing the top 40 independent loci from an interim and a final analysis of the DGI scan. These 196 SNPs were genotyped in a single replication cohort, the cardiovascular cohort of the Malmö Diet and Cancer Study.

Second, during the replication effort described above, we were contacted by two groups—FUSION and SardiNIA—who had performed genome-wide association studies for lipid traits⁴. We agreed to share data and collaboratively carried out a meta-analysis of all three studies, as described in the companion manuscript⁴. For each of three traits—LDL cholesterol, HDL cholesterol and triglycerides—the final meta-analysis involved the association testing of ~2.2 million SNPs in 8,816 individuals.

From this meta-analysis, we selected an additional 30 SNPs for replication testing. These 30 SNPs represented independent loci with P < 10⁻⁴ in the meta-analysis. SNPs at several loci that met this criteria (for example, GRIN3A, LCAT and FARS2 (ref. ⁴)) were not taken forward into replication because of failure to design primers or probes within one Sequenom genotyping assay. At PCSK9, SNP rs11206510 (17% minor allele frequency) was associated with LDL cholesterol in the combined analysis of DGI, FUSION and SardiNIA (P = 7.5 × 10⁻⁶); however, to confirm this locus, we substituted rs11591147 for replication because of prior reports suggesting a very strong effect of rs11591147 on LDL cholesterol²⁶. Replication genotyping for 30 SNPs was conducted in up to 18,554 participants from all three replication cohorts: cardiovascular cohort of the Malmö Diet and Cancer Study, FINRISK97 and NORDIL. Genotyping was conducted for six SNPs at six new loci in Singapore NHS98.

We carried out genotyping using matrix-assisted laser desorption/ionization time-of-flight mass spectrometry on the Sequenom MassARRAY platform or using allelic discrimination method on the ABI 7900 instrument (Applied Biosystems). We excluded SNPs with a Hardy-Weinberg equilibrium P < 0.001. SNPs with a genotyping call rate <90% were also excluded from analysis. The average genotyping call rate exceeded 97% in each of the replication samples. We estimated a genotyping error rate for the Sequenom platform using 15 samples placed in quadruplicate and found this rate to be <0.7%.

mRNA expression and genotyping in human liver samples

Total RNA and DNA were extracted from 60 human liver tissue samples from the University of Washington School of Pharmacy Human Liver Bank. Full details of transcript level measurement, SNP genotyping and association analyses of SNPs with transcript concentrations are described in the Supplementary Methods online.

Statistical analysis

For each participant in the three replication cohorts and the one multiethnic cohort, we log transformed triglyceride values. We adjusted LDL cholesterol, HDL cholesterol and log triglyceride values for the same variables as in DGI (age, age², gender, diabetes status and enrolling center, if applicable). A participant’s multivariable-adjusted residual lipid concentration served as the phenotype in genotype-phenotype association analyses. A small proportion of individuals were on lipid lowering therapy (4.9% in DGI, 1.8% in the cardiovascular cohort of the Malmo Diet and Cancer Study and 3.3% in FINRISK97), and these subjects were excluded.

For genotype-lipid association analyses, we assumed an additive model of inheritance. We conducted multivariable linear regression analyses to test the null hypothesis that LDL cholesterol, HDL cholesterol or triglyceride residuals did not differ by increasing minor allele copy number. We were able to standardize the analyses across the discovery and replication cohorts in the manner described above with two exceptions: in FINRISK97, age² was not considered and outlier individuals in the top 0.5% and the lower 0.5% of the lipid distributions were excluded, and in NORDIL and Singapore NHS98, information on lipid-lowering therapy was not available and thus not considered. Association analyses were conducted in either SAS, SPSS or PLINK³⁰.

To summarize the statistical evidence across the discovery and three replication cohorts, we conducted a fixed-effects variance-weighted meta-analysis⁷, as described in Supplementary Methods.