Genome-wide association study of iron traits and relation to diabetes in the Hispanic Community Health Study/Study of Latinos (HCHS/SOL): potential genomic intersection of iron and glucose regulation?

Abstract

Genetic variants contribute to normal variation of iron-related traits and may also cause clinical syndromes of iron deficiency or excess. Iron overload and deficiency can adversely affect human health. For example, elevated iron storage is associated with increased diabetes risk, although mechanisms are still being investigated. We conducted the first genome-wide association study of serum iron, total iron binding capacity (TIBC), transferrin saturation, and ferritin in a Hispanic/Latino cohort, the Hispanic Community Health Study/Study of Latinos (>12 000 participants) and also assessed the generalization of previously known loci to this population. We then evaluated whether iron-associated variants were associated with diabetes and glycemic traits. We found evidence for a novel association between TIBC and a variant near the gene for protein phosphatase 1, regulatory subunit 3B (PPP1R3B; rs4841132, β = −0.116, P = 7.44 × 10⁻⁸). The effect strengthened when iron deficient individuals were excluded (β = −0.121, P = 4.78 × 10⁻⁹). Ten of sixteen variants previously associated with iron traits generalized to HCHS/SOL, including variants at the transferrin (TF), hemochromatosis (HFE), fatty acid desaturase 2 (FADS2)/myelin regulatory factor (MYRF), transmembrane protease, serine 6 (TMPRSS6), transferrin receptor (TFR2), N-acetyltransferase 2 (arylamine N-acetyltransferase) (NAT2), ABO blood group (ABO), and GRB2 associated binding protein 3 (GAB3) loci. In examining iron variant associations with glucose homeostasis, an iron-raising variant of TMPRSS6 was associated with lower HbA1c levels (P = 8.66 × 10⁻¹⁰). This association was attenuated upon adjustment for iron measures. In contrast, the iron-raising allele of PPP1R3B was associated with higher levels of fasting glucose (P = 7.70 × 10⁻⁷) and fasting insulin (P = 4.79 × 10⁻⁶), but these associations were not attenuated upon adjustment for TIBC—so iron is not likely a mediator. These results provide new genetic information on iron traits and their connection with glucose homeostasis.

Introduction

Iron is necessary for various metabolic processes, including oxygen transport and storage, redox reactions, cell signaling and microbial defense. Absorption, transport and storage of iron are carefully regulated (reviewed in 1,2), presumably to avert potential toxic effects of free iron (such as formation of superoxide radicals and cellular apoptosis). Both iron overload and iron deficiency can be detrimental to health (1), so iron homeostasis is essential. Although many factors that take part in iron homeostasis are known (1,2), mechanisms by which the body regulates iron stores are still being elucidated (3).

In plasma, ferric iron is normally carried by transferrin (TF), en route to the TF receptor in organs such as the bone marrow, liver and spleen. Levels of serum ferritin, a key iron storage protein in the cytosol, nucleus and mitochondria, correlate with iron stores under most physiological conditions. Biomarkers of iron homeostasis, including serum iron, total iron binding capacity (TIBC), transferrin saturation (SAT) and ferritin, are used clinically to assess iron overload or deficiency. Genetic factors play a role in inter-individual variation in levels of these iron biomarkers. For example, hemochromatosis (HFE) variants are associated with hereditary hemochromatosis, which can lead to iron overload and damage to the liver and other organs (4).

Iron biomarkers show significant heritability, apart from hereditary hemochromatosis (5). Single nucleotide polymorphisms (SNPs) associated with iron trait variation have been reported in European populations, where 11 loci associated with iron traits were detected in a recent meta-analysis (6). Other genome-wide association study (GWAS) analyses have been conducted in African Americans (7) and East Asian men (8). Analysis of candidate variants in 233 Hispanic/Latino individuals replicated several associations for the TF locus and TIBC found in other populations (9). However, there has been no published genome-wide analysis of iron traits in Hispanic/Latino populations. Analysis of iron trait genetics in diverse populations is important, as there are known differences between ancestry groups. For example, hereditary hemochromatosis is more common in individuals of European versus African or Hispanic ancestry (10). There may also be differences in iron levels between population groups, such as lower SAT but higher serum ferritin among African Americans (compared with Europeans) (11). These differences could be due in part to genetic factors.

A connection between metabolic disorders—such as diabetes, insulin resistance and fatty liver disease—and iron overload due to hereditary hemochromatosis (12), thalassemia with hypertransfusion (13) or other conditions (14) is well-established. Moreover, excess iron is epidemiologically associated with a higher incidence of type 2 diabetes (T2D) (reviewed in 15,16). For example, elevated serum ferritin levels were associated with roughly a 4-fold increase in the odds of newly diagnosed diabetes in the National Health and Nutrition Education Survey study (17). A few studies have examined genetic factors that influence both iron stores and diabetes risk (18,19), yielding various results. Further work is needed. Better understanding of the genetic underpinnings of iron homeostasis may help elucidate new pathways and connections between iron metabolism and chronic diseases, including diabetes.

Here we report a GWAS analysis of iron traits in the Hispanic Community Health Study/Study of Latinos (HCHS/SOL), to identify potential novel factors involved in iron regulation. We also aimed to assess the role of previously identified iron trait variants in a Hispanic/Latino population. A third goal was to examine the overlap between iron-related loci and risk factors for diabetes or impaired glycemic control. Our results confirm known associations of iron traits with eight loci and potentially identify a novel iron locus.

Results

Outcome measures

Demographic characteristics and levels of iron measures are in Table 1, for the entire sample and for females (n = 7241) and males (n = 5122) separately. Histograms for the distribution of iron measures in the cohort are in Supplementary Material, Fig. S1. Average age in the cohort was 46.1 years (SD 13.9). The average body mass index (BMI) was 29.8 (SD 6.1, median 29), indicating that a high percentage of the cohort was overweight or obese. Prevalence of diabetes was 19.6%, which is higher than the estimated prevalence in the entire HCHS/SOL cohort (16.9%, n = 16 385) (20). Serum iron levels correlated positively with ferritin (r = 0.420) and SAT (r = 0.939) and negatively with TIBC (r = −0.125) (Supplementary Material, Table S1).

Table 1

Demographic characteristics, glycemic traits and iron related measures in the HCHS/SOL, for the entire cohort and females and males

Variable	Entire cohort (n = 12 363)			Females (n = 7241)			Males (n = 5122)
	Mean (SD) or %	Median (Min, Max)	N	Mean (SD) or %	Median (Min, Max)	N	Mean (SD) or %	Median (Min, Max)	N
Age (years)	46.12 (13.86)	48 (18,76)	12 363	46.75 (13.64)	48 (18,76)	7241	45.23 (14.13)	47 (18,75)	5122
BMI (kg/m2)	29.82 (6.07)	29 (14.3,70.3)	12 330	30.36 (6.5)	29.4 (14.3,70.3)	7224	29.06 (5.32)	28.5 (14.9,58.1)	5106
Waist/hip ratio	0.92 (0.08)	0.93 (0.52,1.42)	12 326	0.9 (0.07)	0.9 (0.52,1.42)	7221	0.95 (0.07)	0.95 (0.55,1.25)	5105
Central American heritage	10.85%	–	12 363	10.94%	–	7241	10.72%	–	5122
Cuban heritage	17.73%	–	12 363	15.96%	–	7241	20.23%	–	5122
Dominican heritage	9.29%	–	12 363	10.33%	–	7241	7.81%	–	5122
Mexican heritage	37.30%	–	12 363	38.17%	–	7241	36.08%	–	5122
Puerto Rican heritage	17.55%	–	12 363	17.22%	–	7241	18.02%	–	5122
South American heritage	7.28%	–	12 363	7.37%	–	7241	7.15%	–	5122
Former cigarette use	20.12%	–	12 363	15.51%	–	7241	26.65%	–	5122
Current cigarette use	20.13%	–	12 363	16.20%	–	7241	25.69%	–	5122
Diabetes	19.61%	–	12 363	20.04%	–	7241	19.00%	–	5122
Fasting insulin (µU/ml)	13.35 (12.93)	10.61 (0.33,726)	12 325	13.47 (11.54)	10.83 (0.33,296)	7225	13.18 (14.69)	10.33 (0.56,726)	5100
Fasting glucose (mg/dl)	104.83 (36.81)	95 (52,642)	12 360	102.52 (34.61)	94 (52,642)	7238	108.08 (39.5)	97 (53,449)	5122
2 h Post-load glucose (mg/dl)	123.79 (43.01)	115 (33,398)	10 168	127.78 (42.67)	119 (38,398)	5909	118.27 (42.86)	110 (33,342)	4259
HbA1c (%)	5.9 (1.29)	5.6 (3.3,19.1)	12 317	5.88 (1.21)	5.6 (3.7,19.1)	7213	5.92 (1.38)	5.6 (3.3,15.7)	5104
HOMA-B	141.02 (186.54)	113 (1.1,12600)	12 314	150.05 (217.02)	120 (1.1,12600)	7216	128.24 (130.83)	99.9 (1.9,5333.9)	5098
HOMA-IR	3.6 (4.3)	2.58 (0.07,200.77)	12 325	3.58 (4.17)	2.57 (0.07,179.06)	7225	3.62 (4.47)	2.6 (0.11,200.77)	5100
C-reactive protein (mg/l)	4.02 (6.18)	2.12 (0.11, 112.4)	12 357	4.66 (6.45)	2.69 (0.12, 112.4)	7237	3.12 (5.67)	1.64 (0.11, 104.2)	5120
Ferritin (µg/l)	126.99 (136.5)	89 (7,3061)	12 363	79.02 (74.27)	58 (7,1083)	7241	194.79 (171.26)	152 (7,3061)	5122
Serum iron (µg/dl)	89.82 (33.46)	87 (8,256)	12 363	81.76 (30.95)	80 (8,239)	7241	101.22 (33.56)	98 (13,256)	5122
TIBC (µg/dl)	319.85 (46.86)	316 (140,564)	12 363	326.9 (48.92)	322 (140,564)	7241	309.88 (41.8)	308 (157,500)	5122
SAT (%)	28.81 (11.62)	28 (2,99)	12 363	25.74 (10.62)	25 (2,97)	7241	33.16 (11.59)	32 (4,99)	5122

Abbreviations: HbA1c, glycated hemoglobin; HOMA-B, homeostasis model assessment of β-cell function; HOMA-IR, homeostasis model assessment of insulin resistance; TIBC, total iron-binding capacity; SAT, transferrin saturation.

Primary GWAS analysis

SNPs representing the previously reported TF, HFE, fatty acid desaturase 2 (FADS2)/ myelin regulatory factor (MYRF) and transmembrane protease, serine 6 (TMPRSS6) loci were associated with one or more iron traits at genome-wide significance (P < 5 × 10⁻⁸) (Table 2). An additional novel variant at the protein phosphatase 1, regulatory subunit 3B (PPP1R3B) locus was nearly genome-wide significant for TIBC (P = 7.44 × 10⁻⁸). All genome-wide significant variants listed in Table 2 were directly genotyped or had an imputation r² > 0.95. Manhattan and quantile–quantile (Q–Q) plots and results for all SNPs with P < 1 × 10⁻⁶ for each iron trait are in the supplement (serum iron: Supplementary Material, Fig. S2, Table S2; TIBC: Supplementary Material, Fig. S3, Table S3; SAT: Supplementary Material, Figure S4, Table S4; ferritin: Supplementary Material, Fig. S5, Table S5). Results for each locus are detailed below.

Table 2

SNPs significantly associated with one or more iron traits

Chr	SNP	Position	Nearest gene	Annotation	Effect allele	EAF	Alternate allele	Iron traita	β-value	P-value
3	rs4637289	133458497	TF	intronic	A	0.874	G	Ferritin	0.003	0.625
								Serum iron	0.036	0.296
								TIBC	0.255	1.8 × 10−24
								SAT	−0.058	0.005
3	rs2692666	133478354	TF	intronic	G	0.969	A	Ferritin	0.000	0.978
								Serum iron	0.306	2.57 × 10−6
								TIBC	1.376	8.06 × 10−190
								SAT	−0.240	1.72 × 10−9
3	rs6762719	133480817	TF	intronic	A	0.602	G	Ferritin	0.010	0.038
								Serum iron	−0.081	5.55 × 10−4
								TIBC	−0.676	0 (<1.0 × 10−306)
								SAT	0.149	4.05 × 10−25
3	rs1405023	133481128	TF	intronic	T	0.614	C	Ferritin	−0.014	0.003
								Serum iron	0.003	0.900
								TIBC	0.152	4.50 × 10−18
								SAT	−0.038	0.009
6	rs1799945	26091179	HFEb	His63Asp	C	0.883	G	Ferritin	−0.043	1.13 × 10−9
								Serum iron	−0.365	3.65 × 10−26
								TIBC	0.218	7.37 × 10−18
								SAT	−0.274	2.21 × 10−38
6	rs1800562	26093141	HFEb,c	Cys282Tyr	G	0.982	A	Ferritin	−0.027	0.109
								Serum iron	−0.349	2.26 × 10−5
								TIBC	0.522	3.81 × 10−18
								SAT	−0.371	2.10 × 10−13
8	rs4841132	9183596	PPP1R3B	intergenic	A	0.182	G	Ferritin	0.007	0.220
								Serum iron	0.040	0.177
								TIBC	−0.116	7.44 × 10−8
								SAT	0.053	0.003
11	rs174529	61543961	FADS2/MYRF	intronic (MYRF)	T	0.465	C	Ferritin	0.009	0.073
								Serum iron	0.026	0.283
								TIBC	−0.099	4.51 ×10−8
								SAT	0.045	0.002
22	rs855791	37462936	TMPRSS6d,e	Val736Ala	A	0.439	G	Ferritin	−0.023	5.83 × 10−7
								Serum iron	−0.322	4.61 × 10−46
								TIBC	0.056	9.92 × 10−4
								SAT	−0.202	2.73 × 10−48

a

A log base 10 transformation was used for ferritin, and square root transformation was used for serum iron, TIBC, and transferrin saturation.

b

The minor alleles for both of these HFE variants are associated with higher iron levels [represented by a G nucleotide on the + strand at rs1799945 (and aspartate at residue 63) and an A nucleotide on the + strand at rs1800562 (and tyrosine at residue 282)].

c

Transferrin saturation HFE variant rs79220007 is an LD proxy for coding variant rs1800562 p.Cys282Tyr (r² = 0.98). Variant rs79220007 was thus not included.

d

Ferritin lead TMPRSS6 variant rs760719 is in high LD (r² =  0.79) with coding variant rs855791, which is known to be associated with iron traits (6), and its effect was non-significant when conditioned on rs855791. Variant rs760719 was thus not included.

e

The valine allele is listed as the reference allele in NCBI databases (2321A > G; Val736Ala, with an A nucleotide on the – strand at rs855791), but alanine more likely represents the ancestral allele, because it is conserved across multiple species (gorilla, mouse, rat, chicken) (21).

Gene symbols and abbreviations: TF, transferrin; HFE, hereditary hemochromatosis; PPP1R3B, protein phosphatase 1, regulatory subunit 3B; MYRF, myelin regulatory factor; FADS2, fatty acid desaturase 2; TMPRSS6, transmembrane protease, serine 6; EAF, effect allele frequency; TIBC, total iron-inding capacity; SAT, transferrin saturation.

Table 4

Generalization of iron related variants identified in African American populations by Li et al. (7) to the HCHS/SOL

						Li et al.			HCHS/SOL
rsID	Chr	Position	Nearest gene	Effect allele	Trait	EAF	β-value	P-value	EAF	β-value	r-valuea
rs8177253	3	133480192	TF	T	Total iron-binding capacity	0.24	19.86	1.80 × 10−47	0.398	24.07	1.63 × 10−50
rs115923437	6	22678302	HDGFL1	C	Total iron-binding capacity	0.06	14.84	1.10 × 10−8	0.008	−1.21	1
rs16951289	16	79790621	MAF-DYNLRB2	T	Total iron-binding capacity	0.07	13.38	2.00 × 10−8	0.012	0.75	1
rs141555380	X	153906012	GAB3	T	Log_Ferritin	0.14	0.17	1.10 × 10−8	0.023	0.10	3.20 × 10−4

Reported β-values are from the generalization analysis, which used the same phenotype transformations as in Li et al. (7).

Gene symbols and abbreviations: TF, transferrin; HDGFL1, hepatoma derived growth factor-like 1; MAF-DYNLRB2, MAF bZIP transcription factor—dynein, light chain, roadblock-type 2; GAB3, GRB2 associated binding protein 3; EAF, effect allele frequency.

a

Significant r-values are bolded (r < 0.05).

The TF locus. Variants in the TF locus reached genome-wide significance for TIBC and SAT (Supplementary Material, Fig. S6). Conditional analysis of the TF locus showed five distinct genome-wide significant signals for TIBC, represented by non-coding variants rs6762719, rs2692666, rs1130459, rs1405023 and rs4637289. Two of these variants are in modest LD with rs6762719 in HCHS/SOL (rs1405023, r² = 0.43; rs1130459, r² = 0.22) and two are not (rs2692666 and rs4637289, r² < 0.1). Variant rs1130459 was not associated with TIBC in the primary analysis (β-value = −0.0006, S.E. = 0.018, P= 0.971), but became significant when conditioned on variants rs6762719 and rs2692666 (β-value = 0.269; S.E. = 0.018, P = 8.68 × 10⁻⁵¹ after conditioning). Our top TF SNP (rs6762719) is in very high LD with the lead variant at TF reported in Europeans (rs8177240; r² = 0.98 in the 1000 Genomes European reference panel) (6). Variant rs1405023 shows modest LD with rs8177179 (r² = 0.42 in 1000 Genomes Europeans), which was identified by conditional analysis at the TF locus in Europeans (6). Conditional analysis showed a single independent signal at TF for SAT (rs6762719).

The HFE locus. The HFE locus reached genome-wide significance for all tested iron traits (Supplementary Material, Fig. 7). For serum iron and ferritin, a single signal was revealed by conditional analysis (rs1799945, p.His63Asp). For TIBC and SAT, conditional analysis showed two independent signals at the HFE locus, represented by rs1799945 (p.His63Asp) and rs1800562 (p.Cys282Tyr) (which are in very low LD, r² = 0.002). These coding HFE variants are associated with hereditary hemochromatosis in homozygotes and compound heterozygotes (hemochromatosis, type 1, OMIM 235200). The second independent signal for SAT was led by rs79220007, which is in very strong linkage disequilibrium (LD) with rs1800562 (r² = 0.98).

The FADS2/MYRF locus. A single intron variant in MYRF was associated with TIBC (P = 4.51 × 10⁻⁸; Supplementary Material, Fig. S8). The variant is in strong LD with an intron variant in FADS2 (rs174577, r² = 0.76 in 1000 Genomes Europeans), which was associated with TF among Europeans (6).

The TMPRSS6 locus. There was a single independent signal at TMPRSS6 (Supplementary Material, Fig. S9), which was confirmed by conditional analysis. The signal was represented by lead variant rs855791 (p.Val736Ala) for serum iron and SAT, and by lead variant rs760719 for ferritin. Variants rs760719 and rs855791 are in strong LD (r² = 0.79).

The PPP1R3B locus. A novel association involving a variant near the PPP1R3B locus and TIBC showed near genome-wide significance (rs4841132, β  = −0.116, P = 7.44 × 10⁻⁸; Supplementary Material, Fig. S10).

Additional signals. A low-frequency intergenic variant on chromosome 1 was associated with TIBC (rs78455250, β = −0.621, P = 3.83 × 10⁻⁸; Supplementary Material, Fig. S11). A low-frequency intronic variant on chromosome 5 was associated with SAT (rs2442120, P = 4.92 × 10⁻⁸; Supplementary Material, Fig. S12). However, the association became non-significant when conditioned on the TF, HFE and TMPRSS6 loci, and replication was not pursued.

Replication of novel signals

The newly identified PPP1R3B variant was nominally associated with TF levels (which are linearly related to TIBC values) in the European meta-analysis (β = −0.0577, P = 0.0014; personal communication with Dr Benyamin) (6). A significant association was not observed in the Jackson Heart Study, although the estimated association was in the same direction (β = −2.162, P = 0.246) (7). Variant rs4841132 was most common in the admixed American populations in 1000 Genomes Phase 1 (19% minor allele frequency, compared with 10% in African, 8% in European and 1% in Asian populations).

As mentioned earlier, a variant on chromosome 1 was also associated with TIBC (rs78455250). The variant had a low minor allele frequency in HCHS/SOL (0.2%). The minor allele frequency was higher in the Jackson Heart Study (2.3%), but association with TIBC was in the opposite direction and non-significant (β = 1.451, P = 0.762). This variant was not observed in European 1000 Genomes participants, so replication in Europeans was not attempted. The variant is not considered further here.

Sensitivity analysis excluding individuals with iron deficiency

Excluding individuals with iron deficiency reduced effect sizes for TF and SAT loci in the European meta-analysis (6). Therefore, we excluded individuals with iron deficiency, as a sensitivity analysis for our HCHS/SOL GWAS. Most associations (28/36; Supplementary Material, Table S6) became more significant, despite the reduced sample size. For example, the association of rs4841132 (near PPP1R3B) with TIBC became stronger and genome-wide significant (original model, β = −0.116, P= 7.44 × 10⁻⁸, n = 12 585; excluding iron deficient individuals, β = −0.121, P= 4.78 × 10⁻⁹, n = 11 614).

Moreover, an additional 213 SNPs had genome-wide significant effects after excluding iron deficient individuals (1397 genome-wide significant SNPs, Supplementary Material, Table S7). All but 14 of these variants were in previously recognized loci in HCHS/SOL. For example, associations of SNPs in TF with serum iron strengthened to genome-wide significance levels (lead variant rs2692696, P= 5.11 × 10⁻⁹; Supplementary Material, Fig. S13.a). However, a new genome-wide significant association of SNPs in ABO blood group (ABO) with ferritin was observed, led by intron variant rs657152 (β = 0.025, P= 1.94 × 10⁻⁸; Supplementary Material, Fig. S13.b). Variant rs657152 is in modest LD (r² = 0.4 in the 1000 Genomes Europeans) with rs651007, a variant downstream of ABO that was associated with ferritin levels in European populations (6) and generalized to HCHS/SOL (r-value = 1.84 × 10⁻³; Table 3). Two imputed variants not at previously identified loci were also associated with serum iron (rs144628729 and rs181143083).

Table 3

Generalization of iron related variants identified in European populations by Benyamin et al. (6) in the HCHS/SOL

				Effect allele		Benyamin et al.			HCHS/SOL
rsID	Chr	Position	Nearest gene		Trait	EAF	β-value	P-value	EAF	β-value	r-valuea
rs744653	2	190378750	WDR75–SLC40A1	T	Transferrinb	0.854	0.068	1.35 × 10−11	0.887	0.016	0.16
					Log_Ferritin	0.854	–0.089	8.37 × 10−19	0.887	–0.002	1
rs8177240	3	133477701	TF	T	Serum iron	0.669	–0.066	6.65 × 10−20	0.613	–0.022	4.33 × 10−4
					Transferrin	0.669	–0.38	8.43 × 10−610	0.613	–0.241	4.20 × 10−95
					Saturation	0.669	0.1	7.24 × 10−38	0.613	0.061	3.48 × 10−23
rs9990333	3	195827205	TFRC	T	Transferrin	0.46	–0.051	1.95 × 10−13	0.331	–0.009	0.24
rs1800562	6	26093141	HFE (Cys282Tyr)	A	Serum iron	0.067	0.328	2.72 × 10−97	0.018	0.096	1.20 × 10−5
					Transferrin	0.067	–0.479	8.90 × 10−196	0.018	–0.189	1.97 × 10−16
					Saturation	0.067	0.577	2.19 × 10−270	0.018	0.172	7.81 × 10−16
					Log_Ferritin	0.067	0.204	1.54 × 10−38	0.018	0.033	0.33
rs1799945	6	26091179	HFE (His63Asp)	C	Serum iron	0.85	–0.189	1.10 × 10−81	0.883	–0.094	4.03 × 10−27
					Transferrin	0.85	0.114	9.36 × 10−30	0.883	0.081	6.33 × 10−17
					Saturation	0.85	–0.231	5.13 × 10−109	0.883	–0.120	1.51 × 10−41
					Log_Ferritin	0.85	–0.065	1.71 × 10−10	0.883	–0.053	3.59 × 10−5
rs7385804	7	100235970	TFR2	A	Serum iron	0.621	0.064	1.36 × 10−18	0.704	0.025	1.14 × 10−4
					Saturation	0.621	0.054	6.07 × 10−12	0.704	0.018	7.72 × 10−3
rs4921915	8	18272466	NAT2	A	Transferrin	0.782	0.079	7.05 × 10−19	0.669	0.027	9.17 × 10−5
rs651007	9	136153875	ABO	T	Log_Ferritin	0.202	–0.05	1.31 × 10−8	0.164	–0.032	1.84 × 10−3
rs6486121	11	13355770	ARNTL	T	Transferrin	0.631	–0.046	3.89 × 10−10	0.515	–0.012	0.09
rs174577	11	61604814	FADS2	A	Transferrin	0.33	0.062	2.28 × 10−17	0.551	0.035	6.76 × 10−7
rs411988	17	56709034	TEX14	A	Log_Ferritin	0.564	–0.044	1.59 × 10−10	0.407	–0.009	1
rs855791	22	37462936	TMPRSS6	A	Serum iron	0.446	–0.181	1.32 × 10−139	0.439	–0.082	9.41 × 10−47
			(Val736Ala)		Transferrin	0.446	0.044	1.98 × 10−9	0.439	0.020	2.63 × 10−3
					Saturation	0.446	–0.19	6.41 × 10−137	0.439	–0.087	1.30 × 10−50
					Log_Ferritin	0.446	–0.055	1.38 × 10−14	0.439	–0.029	6.99 × 10−6

Reported β-values are from the generalization analysis, which used the same phenotype transformations as in Benyamin et al. (6)

a

Significant r-values are bolded (r < 0.05).

b

TIBC values were used to generalize transferrin results.

Gene symbols and abbreviations: WDR75–SLC40A1, WD repeat domain 75—solute carrier family 40 (iron-regulated transporter), member 1 (also known as ferroportin); TF, transferrin; TFRC, transferrin receptor; HFE, hereditary hemochromatosis; TFR2, transferrin receptor 2; NAT2, N-acetyltransferase 2 (arylamine N-acetyltransferase); ABO, ABO blood group; ARNTL, aryl hydrocarbon receptor nuclear translocator like; FADS2, fatty acid desaturase 2; TEX14, testis expressed 14; TMPRSS6, transmembrane protease, serine 6; EAF, effect allele frequency; TIBC, total iron-binding capacity.

Generalization of previously identified GWAS signals in HCHS/SOL

In the largest European meta-analysis, Benyamin et al. detected associations between iron traits and 11 loci [HFE (two independent variants), solute carrier family 40 (iron-regulated transporter), member 1 (SLC40A1), TF, transferrin receptor 2 (TFR2), transferrin receptor (TFRC), TMPRSS6, ABO, aryl hydrocarbon receptor nuclear translocator like (ARNTL), FADS2, N-acetyltransferase 2 (arylamine N-acetyltransferase) (NAT2), testis expressed 14 (TEX14)] (6). We investigated all variants for all iron traits, because several were associated with multiple iron traits. Eight of the twelve variants from Benyamin et al. generalized to HCHS/SOL, with 5 generalizing for multiple iron traits (Table 3; Supplementary Material, Fig. S14.a–d).

There were a number of loci with different associations with iron traits in HCHS/SOL, compared with the European meta-analysis. For example, rs1800562 in HFE was highly significant for ferritin in the discovery cohort, but it did not generalize and had a much lower estimated effect in HCHS/SOL (Supplementary Material, Fig. S14.d). This variant had strong associations in HCHS/SOL with SAT, TIBC and serum iron, so lack of association with ferritin may not be simply due to lack of power. Additionally, rs744653 near the WDR75–SLC40A1 locus was associated with both ferritin and TF in Benyamin et al., but it was not associated with any iron trait in HCHS/SOL. Variants rs9990333 (near TFRC), rs6486121 (in ARNTL) and rs411988 (in TEX14) also showed no evidence of generalization in HCHS/SOL. Finally, some variants that did generalize had confidence intervals that did not overlap with previously reported β-values. For example, HFE variant rs1800562 and TMPRSS6 variant rs85579 consistently had more modest estimated β-values in HCHS/SOL, suggesting differences in effect sizes between populations.

The largest analysis of iron-related traits in African Americans found four loci associated with ferritin and TIBC (7). Two of these generalized to HCHS/SOL (Table 4; Supplementary Material, Fig. S15.a and b). The estimated confidence intervals for the non-generalized loci are consistent with no association (rather than lack of power; Supplementary Material, Fig. 15.b). One variant was identified in an East Asian GWAS of ferritin, but the variant did not generalize to HCHS/SOL (rs5742933 on chromosome 2 (8), P = 0.122; not included in Tables 3 and 4).

Associations of iron loci with glycemic traits

We tested the lead variants from each iron-associated locus for associations with glycemic traits: rs6762719 for TF, rs1799945 for HFE, rs4841132 for PPP1R3B, rs174529 for FADS2/MYRF and rs855791 for TMPRSS6 (Table 5). All effect alleles in Table 5 correspond to the iron raising allele, for easier interpretation of the β-values.

Table 5

Associations of iron-related loci with glycemic traits

	rsID	rs6762719	rs1799945	rs4841132	rs174529	rs855791
	Chromosome	3	6	8	11	22
	Position	133480817	26091179	9183596	61543961	37462936	GRSc
	Gene	TF	HFE	PPP1R3B	FADS2/MYRF	TMPRSS6	5-variant	10-variant
	Iron raising allele	A	G	A	T	G
Diabetes	Odds ratio	.95	0.97	1.00	1.03	0.92	0.99	0.99
	P-valuea	0.225	0.622	0.983	0.554	0.060	0.075	1.43 × 10−3
	n	7746	7746	7745	7746	7746	7745	7745
2-h postload glucose	β-valueb	0.063	−1.367	−1.652	0.537	−0.691	−0.458	−0.388
	P-value	0.885	0.035	0.002	0.229	0.103	0.035	0.021
	n	9622	9622	9621	9622	9622	9621	9621
Fasting glucose	β-value	0.132	0.115	0.706	0.266	−0.209	0.167	0.092
	P-value	0.257	0.499	7.70 × 10−7	0.027	0.063	0.004	0.040
	n	10065	10065	10064	10065	10065	10064	10064
Fasting insulin	β-value	−0.001	0.006	0.020	−0.004	−0.001	0.003	0.001
	P-value	0.831	0.268	4.79 × 10−6	0.323	0.824	0.129	0.371
	n	10053	10053	10052	10053	10053	10052	10052
HbA1c	β-value	−0.006	−0.031	−0.009	0.013	−0.030	−0.012	−0.016
	P value	0.235	2.44 × 10−5	0.173	0.013	8.66 × 10−10	5.64 × 10−6	1.11 × 10−15
	n	9636	9636	9635	9636	9636	9635	9635
HOMA-B	β-value	−0 .002	0.005	0.012	−0.008	0.001	0.0002	−0.00003
	P-value	0.510	0.365	0.009	0.026	0.742	0.893	0.982
	n	10045	10045	10044	10045	10045	10044	10044
HOMA-IR	β-value	0.000	0.006	0.023	−0.002	−0.002	0.004	0.002
	P-value	0.977	0.256	4.98 × 10−7	0.543	0.649	0.062	0.248
	n	10052	10052	10051	10052	10052	10051	10051

a

Significant P-values are bolded (P < 0.001 for single variant tests, P < 0.004 for the GRS).

b

All β-values are calculated with respect to the indicated iron raising alleles.

c

The GRSs are unweighted and based on either the five sentinel SNPs (5-variant) or the five sentinel SNPs plus 5 generalized SNPs from reported genome-wide association studies (10-variant). The GRSs are calculated with respect to the iron raising alleles.

d

A log base 10 transformation was used for fasting insulin, HOMA-IR and HOMA-B. Other glycemic trait measures were not transformed.

Gene symbols and abbreviations: TF, transferrin; HFE, hereditary hemochromatosis; PPP1R3B, protein phosphatase 1, regulatory subunit 3B; FADS2, fatty acid desaturase 2; MYRF, myelin regulatory factor; TMPRSS6, transmembrane protease, serine 6; HbA1c, glycated hemoglobin; HOMA-B, homeostasis model assessment of β-cell function; HOMA-IR, homeostasis model assessment of insulin resistance.

The PPP1R3B locus. The PPP1R3B variant was associated with higher fasting glucose (P = 7.70 × 10⁻⁷), insulin levels (P = 4.79 × 10⁻⁶) and homeostasis model assessment of insulin resistance (HOMA-IR), a correlated, calculated measure of insulin resistance (P = 4.98 × 10⁻⁷). Adjustment for TIBC did not attenuate associations of PPP1R3B with glycemic traits, with little change in the P-values (insulin resistance, P = 1.02 × 10⁻⁸; fasting glucose, P = 5.05 × 10⁻⁷; fasting insulin, P = 1.3 × 10⁻⁷) or β-values (Supplementary Material, Table S8.a).

Other loci. The iron-raising alleles at TMPRSS6 (rs855791, P = 8.66 × 10⁻¹⁰) and HFE (rs1799945, P = 2.44 × 10⁻⁵) were associated with lower HbA1c levels in HCHS/SOL. Adjustment for iron measures attenuated, but did not eliminate, the associations with HbA1c (Supplementary Material, Table S8.b). Adjustment for SAT most strongly attenuated the HFE association (P= 0.0166) and the TMPRSS6 association (P= 2.40 × 10⁻⁵). Iron is therefore at least one of the likely mediators of the HbA1c associations.

An unweighted genetic risk score (GRS) based on the five sentinel SNPs was associated with lower HbA1c levels (P = 5.64 × 10⁻⁶) and was more nominally associated with higher levels of fasting glucose (P = 0.004) (multiple comparisons threshold, P < 0.004; Table 5). These five sentinel SNPs together explain only a small percentage of the residual variance in iron traits (9.6% of TIBC, 1.8% of SAT, 1.6% of serum iron, <0.1% of ferritin), after adjusting for the potential confounders in our SNP association models (such as age, sex, heritage group and kinship). We also derived a 10-variant, unweighted GRS, based on the 5 sentinel SNPs plus 5 variants that generalized to HCHS/SOL. Single variant associations for the five generalized SNPs with glycemic traits are shown in Supplementary Material, Table S9. The 10-variant GRS was associated with lower odds for diabetes (P = 1.43 × 10⁻³) and lower HbA1c (P= 1.11 × 10⁻¹⁵; Table 5). For both the 5- and 10-variant scores, the association with HbA1c became non-significant when signals at HFE and TMPRSS6 (and GRB2 associated binding protein 3 (GAB3) for the 10-variant GRS) were removed (P > 0.1). A more nominal association (P = 0.004) of the 5-variant score with higher fasting glucose was attenuated when the signal at PPP1R3B was removed (P > 0.32).

Replication in other studies. We examined associations of our iron-related loci with T2D and related traits in published GWAS meta-analyses. No variant was associated with T2D status. The iron-raising allele of the PPP1R3B variant was associated with higher fasting glucose (P = 3.35 × 10⁻⁹) and higher fasting insulin (P = 1.71 × 10⁻⁹), as well as more nominally with higher insulin resistance (P = 0.001), as in HCHS/SOL (Supplementary Material, Table S10). SNPs at the loci for TMPRSS6 (rs855791, P = 2.74 × 10⁻¹⁴) and HFE (rs1799945, P = 1.43 × 10⁻⁴) were associated with lower HbA1c levels. MYRF/FADS2 proxy variant rs174577 was strongly associated with higher fasting glucose (P = 1.54 × 10⁻¹⁰) and more modestly with lower β-cell function (P = 0.0001). The closely linked FADS1 gene is a recognized genome-wide significant locus for fasting glucose and homeostasis model assessment of β-cell function (HOMA-B) (22,23).

Discussion

Associations of variants with iron traits

In the HCHS/SOL cohort, we found genome-wide significant associations of iron traits with four previously known loci (TF, HFE, FADS2/MYRF, TMPRSS6) and identified a potential novel iron locus (PPP1R3B).

PPP1R3B is a subunit of the serine/threonine phosphatase, protein phosphatase-1 (PP1). PPP1R3B (originally called G_L) targets PP1 to glycogen synthase in the liver, regulating glycogen storage. PPP1R3B is also highly expressed in human skeletal muscle (24). To our knowledge, there is no direct link between PPP1R3B and iron metabolism, and this variant has not been previously associated with iron traits. The observed association was replicated for TF in results from a European meta-analysis (6; Benyamin, personal communication). However, the PPP1R3B association did not replicate in African Americans from the Jackson Heart Study—though the estimated association was in the same direction (7). The lack of replication could be due to differences in effect allele frequencies between populations, because the identified variant has a higher effect allele frequency in admixed American populations in the 1000 Genomes reference panel (compared with European, African, or Asian populations). This higher frequency may have facilitated detection of the novel association with TIBC.

The SNP associated with TIBC (rs4841132) lies within a noncoding RNA locus (LOC157273)—roughly 200 kilobases away from the PPP1R3B gene. Recent siRNA knockdown experiments suggest that LOC157273, a hepatocyte specific transcript, regulates PPP1R3B mRNA levels (25). Moreover, LOC157273/PPP1R3B locus variants have been associated with a variety of cardiometabolic phenotypes, including levels of LDL, HDL, total cholesterol (26), C-reactive protein (27), blood lactate (28), fat accumulation in the liver measured by computed tomography (29), fasting blood glucose (23), including levels during pregnancy (30), fasting blood insulin (23) and liver enzymes (31). The potential pleiotropic effects of PPP1R3B or LOC157273 on these multiple cardiometabolic phenotypes warrant further study. The impact of the locus on iron traits in particular should be further assessed in diverse populations.

Generalization of previously observed variants

We attempted to generalize variant-phenotype associations observed in the largest GWAS analyses of European and African American populations (7). Eight of the twelve variants from the European meta-analysis generalized to HCHS/SOL for at least one iron trait. Two of the four variants from the African American study generalized. Several variants, including a variant near GAB3 in African Americans and variants near ABO, NAT2 and FADS2 in European Americans, have not been previously confirmed in an independent study. Therefore, generalization to HCHS/SOL provides important validation of these results.

Other variants did not generalize to HCHS/SOL, including SNPs at WDR75–SLC40A1, TFRC, ARNTL and TEX14 from the European meta-analysis and hepatoma derived growth factor-like 1 (HDGFL1) and MAF bZIP transcription factor – dynein, light chain, roadblock-type 2 (MAF-DYNLRB2) from the African American meta-analysis. Lack of generalization may be due to different effect sizes in the original study population versus HCHS/SOL. Alternatively, lack of power may have led to reduced precision in effect size estimation. For the variants from the European meta-analysis that did not generalize, the estimated directions of effects were consistent with those in HCHS/SOL (Supplementary Material, Figs S14.b and d) (6). For the two variants from the African American meta-analysis that did not generalize, the estimated direction of effect in HCHS/SOL was the same as in the original study for rs16951289, but not for rs115923437 (Supplementary Material, Fig. S15.b). Both variants had a much smaller effect allele frequency in HCHS/SOL than in the original analysis by Li et al. (7). Differences in allele frequencies are less extreme for the non-generalized variants from the European GWAS. Specifically, no non-generalized variant had a minor allele frequency <10% in either cohort. However, several variants do have lower minor allele frequencies in HCHS/SOL than in the Benyamin et al. meta-analysis, which may modestly impact power (e.g. rs9990333). Other possible reasons for lack of generalization include differences in causal variants or haplotype structure between populations, gene–environment interactions and heterogeneity in environmental exposures between populations (such as iron in the diet), sampling error in either the original study or HCHS/SOL (such that the tested subpopulation does not represent the larger population), and unaccounted for sources of variance.

Apart from variants that did not generalize, some variants showed evidence of different effect sizes in HCHS/SOL. For example, HFE variant rs1800562 (p.Cys282Tyr) consistently showed a smaller effect size in HCHS/SOL than in the European meta-analysis – likely due to the higher prevalence of hereditary hemochromatosis and p.Cys282Tyr homozygotes in Europeans (10). The TMPRSS6 (rs855791) and TF (rs8177240) variants also had significantly lower estimated effect sizes in HCHS/SOL. The reasons for these population differences in effect size are unclear. Differences in non-genetic factors between studies, such as socioeconomic status, demographic traits and diet, may play a role. However, our results generally support the role of known iron trait loci in a Hispanic/Latino population, suggesting that these alleles are derived from a common founder population.

Associations of variants with iron and diabetes related traits

Excess iron is epidemiologically associated with an increased incidence of T2D (15,16). Many mechanisms for this link have been proposed. Although iron is needed for β-cell function and survival, free iron gives rise to reactive oxygen species in these metabolically active cells. Excess iron may also influence glucose utilization in hepatic and adipose tissue. As a possible mechanism, high iron (like high oxygen) decreases levels of hypoxia inducible factors, which in turn reduces levels of glucose transporters and insulin secretion. Iron also may impact circadian rhythms in the liver (altering hepatic gluconeogenesis) or mediate epigenetic effects, such as histone demethylation (15,16). Therefore, associations of iron loci with diabetes and glycemic traits are of interest. Five sentinel SNPs were selected from among the iron related loci found in HCHS/SOL (TF, HFE, PPP1R3B, FADS2/MYRF and TMPRSS6). We hypothesized that alleles associated with raised iron levels may also be associated with increased diabetes risk.

PPP1R3B. The TIBC lowering and SAT raising allele for PPP1R3B variant rs4841132 was associated with higher levels of fasting glucose, fasting insulin and insulin resistance in HCHS/SOL. In earlier GWAS analyses, rs4841132 was associated with higher fasting insulin and fasting glucose levels (23), including levels during pregnancy (30). As discussed earlier, PPP1R3B regulates glycogen storage in the liver, providing a potential mechanistic link to glucose homeostasis (32). Adjustment for TIBC did not attenuate associations of PPP1R3B with fasting glucose, fasting insulin, or insulin resistance. Therefore, iron is not a likely mediator of the observed diabetes-related effects of rs4841132. PPP1R3B or LOC157273 may instead have pleiotropic effects on both iron and glycemic phenotypes. An effect of a PPP1R3B variant (rs9987289) on CRP levels in a previous study was similarly not attenuated by adjustment for HDL, LDL or total cholesterol (all traits associated with rs9987289). Lack of attenuation is compatible with pleiotropic effects on these traits (33). A number of iron loci identified here or in previous studies have been associated with lipid traits by the Global Lipids Genetics Consortium (PPP1R3B, FADS2, ABO, HFE, NAT2) (26), suggesting there may be shared determinants of lipids and iron that could have relevance to diabetes.

TMPRSS6 and HFE. Iron-raising alleles for TMPRSS6 (rs855791, P = 8.66 × 10⁻¹⁰) and HFE (rs1799945, P = 2.44 × 10⁻⁵) were associated with lower HbA1c levels. Iron attenuated these associations in HCHS/SOL. In support of our results, TMPRSS6 (rs855791) and HFE (rs1800562) variants were associated with lower HbA1c levels in a meta-analysis of nondiabetic adults of European descent (34). Red blood cell production involves iron, so we performed look-ups for our five sentinel variants with eight hematological traits in a previous HCHS/SOL GWAS (Supplementary Material, Table S11). Iron-raising variants of HFE and TMPRSS6 were associated with higher hemoglobin (P≤ 8.20 × 10⁻⁸), mean corpuscular volume (MCV) (P≤ 8.49 × 10⁻⁶), and mean corpuscular hemoglobin (MCH) (P≤ 2.39 × 10⁻¹¹), along with other red blood cell traits—similar to results from a European meta-analysis (35). Iron-raising alleles of TMPRSS6 (at rs855791) and HFE (at rs1799945 and rs1800562) may therefore contribute to increased hemoglobin levels (35)—resulting in a lower calculated percentage of HbA1c. Alternatively, iron may impact the rate of hemoglobin glycation. Iron-deficiency anemia has been associated with spuriously high HbA1c, compared with other markers of glycemia. Moreover, iron replacement therapy lowers HbA1c and raises total hemoglobin levels in diabetic and non-diabetic individuals (36–38). Malondialdehyde, which is elevated in patients with iron deficiency anemia (38), can increase hemoglobin glycation in vitro, providing a possible mechanism (39). Further assessment of the role of genetic factors that influence iron traits on HbA1c as a screening tool for diabetes is warranted.

Genetic risk scores. Overall, we found few compelling associations with either a 5- or 10-variant GRS, except for an association with lower HbA1c—likely due to effects on red cells, instead of glucose.

No association with increased risk for diabetes was observed. We did see nominal associations with quantitative glycemic traits indicative of higher diabetes risk, e.g. higher fasting glucose.

Our results are consistent with the limited evidence for a link between iron variants and diabetes risk in previous studies. A GRS based on six iron variants was modestly associated with increased odds for diabetes among men in the Health Professionals Follow-Up Study (18). However, the association was not observed in women and was attenuated by BMI adjustment. In a Chinese cohort, the iron lowering TMPRSS6 rs855791 A variant (with valine at position 736) was associated with lower plasma ferritin levels (β = −0.061, P = 0.002) and with modestly lower risk for T2D (OR 0.801, 95% CI 0.65–0.98; P = 0.031) (19).

Strengths and limitations

Our study represents the first GWAS of iron indices in a Hispanic/Latino cohort. Additional strengths are analyses of the impact of iron deficiency on SNP-iron trait associations (which were generally stronger when iron deficient individuals were excluded) and a formal generalization analysis of GWAS loci identified in European and African American populations. We hypothesize that excluding individuals with iron deficiency reduces heterogeneity due to non-genetic factors, such as dietary iron, blood loss (among premenopausal women), regional background and socioeconomic status. Limitations include lack of a sample size large enough for a more comprehensive evaluation of variants and inability to further explore reasons for lack of generalization. We were also unable to further clarify potential reasons for significant differences in effect sizes between populations in the generalization analysis. The low variance in iron traits explained by our five sentinel SNPs limits our power to detect modest associations between the genetic determinants of iron levels (as assessed by a GRS) and diabetes or related traits. Identification of more variants associated with iron traits in larger Hispanic/Latino cohorts would improve power. We thus cannot exclude a modest effect of iron related variants on diabetes related traits. Lack of differentiation between type 1 and type 2 diabetes (in a small percentage of cases) in HCHS/SOL may also impact our ability to detect associations with diabetes case/control status.

Future directions

Important future studies of iron genetics may include trans-ethnic meta-analyses and evaluations of iron biomarkers in additional Hispanic/Latino cohorts, which may help identify other iron-associated variants. Assessment of iron variants in multi-ethnic cohorts may also be useful for identifying reasons for lack of generalization between populations or different variant effect sizes. Large scale sex-stratified analyses may also be an important future direction, as women generally have lower body iron stores (particularly premenopausal women). However, we observed no heterogeneity in effects between men and women for genome-wide significant loci in HCHS/SOL (by Cochran Q testing; results not shown). The lack of compelling associations of iron related SNPs with glycemic traits and diabetes in HCHS/SOL should also be explored in additional ancestry groups, because a relationship between iron traits and diabetes may vary between populations.

Summary

Our results show the generalization of previously reported associations between iron traits and TF, HFE, FADS2/MYRF, TMPRSS6, TFR2, NAT2, ABO and GAB3 to a Hispanic/Latino population. We identified a potential novel iron locus, PPP1R3B, where lead variant rs4841132 was associated with lower TIBC levels. The association was replicated in summary statistics from a previous European meta-analysis and warrants further exploration. The PPP1R3B variant was associated with higher fasting glucose and insulin levels, and higher insulin resistance, but these effects were unrelated to TIBC. None of the other iron variants was associated with increased risk for diabetes, either individually or as part of a GRS. Further research is needed to explore potential links between the genetic determinants of iron and glucose homeostasis.

Materials and Methods

HCHS/SOL

The HCHS/SOL cohort began in 2006 as a prospective study of Hispanic/Latino populations in the USA. (40–42). From 2008 to 2011, 16 415 adults were recruited from a random sample of households in four communities (the Bronx, Chicago, Miami, and San Diego). Each Field Center recruited >4000 participants from diverse socioeconomic groups. Most participants self-identified as having Cuban, Dominican, Puerto Rican, Mexican, Central American or South American heritage. A total of 12 803 individuals consented for genotyping. The study protocol was approved by the institutional review boards of all collaborating institutions, and written informed consent was obtained for all participants. Here we use data collected from study participants at their first clinic visit.

Outcome measures

Tests used to characterize body iron stores included serum iron, TIBC and ferritin. Serum iron was measured on a Roche Modular P chemistry analyzer using a ferrozine reagent (Roche Diagnostics, Indianapolis, IN). Unsaturated iron binding capacity (UIBC) was assayed on the same sample, and TIBC was calculated by the formula: TIBC = serum iron + UIBC. Ferritin was measured in serum with Roche reagents on a Cobas 6000 Analyzer (Roche Diagnostics), using a particle enhanced immunoturbidimetric assay. SAT levels were calculated by the formula: SAT = serum iron/TIBC × 100.

Diabetes mellitus status was defined using American Diabetes Association criteria (43). Individuals with a level of fasting (>8 h) glucose ≥126 mg/dl, non-fasting (≤8 h) glucose or 2-h postload glucose from an oral glucose tolerance test (OGTT) ≥200 mg/dl, HbA1c ≥6.5%, or who used a diabetes medication were defined as diabetes cases. Individuals with levels of fasting glucose <100 mg/dl, 2-h post-load glucose from an OGTT <140 mg/dl, and HbA1c <5.6% were defined as diabetes controls. Individuals with intermediate values, which may indicate an increased risk for diabetes, were not included as controls. There were 2499 cases and 5247 controls defined by these criteria. The prevalence of diabetes among 18–29-year olds was only 2.6% in the entire HCHS/SOL cohort (20). We estimate that the prevalence of T2D among diabetic individuals in our analysis is roughly >95% (with a small percentage of type 1 diabetes cases likely included), based on the low diabetes prevalence in younger HCHS/SOL participants. Diabetes cases were excluded from analyses of glycemic traits, including 2-h postload glucose from an OGTT, fasting glucose and insulin, HbA1c and the homeostasis model assessments of insulin resistance and β-cell function (HOMA-IR and HOMA-B).

Genotyping and imputation

Genotyping and imputation protocols have been described previously in (44). Genotyping was performed by Illumina Microarray Services with a custom array [>2.5 million SNPs from the HumanOmni2.5-8v1-1 array and ∼150 000 custom SNPs; Illumina (San Diego, CA)]. Genotype data were cleaned and quality checked centrally at Illumina Microarray Services, LA Biomed, and the University of Washington, which heads the HCHS/SOL Genetic Analysis Center. SHAPEIT2 was used for prephasing (45), and imputation to the 1000 Genomes full, cosmopolitan phase 1 version 1 reference panel was completed using IMPUTE2 (46). There were >25.5 million imputed genotypes, for a total of 27.7 million tested SNPs. We excluded variants with low minor allele counts (<30) and low imputation quality (<0.3) from all results shown.

Primary GWAS analysis

GWAS analyses of iron traits were performed in all individuals with genomic and iron trait data (12 375 participants for ferritin; 12 580 for serum iron; 12 586 for TIBC; and 12 589 for SAT). Extreme outliers were excluded (ferritin > 10 000 ng/ml; serum iron >270 µg/dl; TIBC > 580 µg/dl). Participants with malignant tumors of the blood or lymph system or bone cancer (n = 28), high immature granulocyte counts (n = 2), end-stage renal disease (n = 46) or pregnancy at the time of iron trait measurement (n = 8) were excluded, as were participants on chemotherapy (n = 54).

To approximate a normal distribution, a log base 10 transformation was used for ferritin, and square root transformation was used for serum iron, TIBC and SAT. We used a mixed model approach implemented in the GENESIS package (GENetic EStimation and Inference in Structured samples) in R (version 2.0.1), which can account for population structure/ancestry in related individuals (47), to assess SNP associations with iron-related traits. A detailed description of the GWAS analysis methods for HCHS/SOL has been previously published in (48). SNP associations were adjusted for sex, age, cigarette use, heritage group (Mexican, Central American, South American, Cuban, Dominican, Puerto Rican), sampling design, kinship, household, census block, recruitment center and the first five ancestry principal components. Individuals who were outliers for ancestry principal components were excluded (n = 56). Chromosome X was analyzed separately, with adjustment for chromosome X specific eigenvectors. The HCHS/SOL Genetic Analysis Center tracks all genome-wide analyses performed (see Supplementary Material, Table S12 for analysis identification numbers). Genomic inflation was assessed with Q–Q plots. Variants with P < 10⁻⁶ were annotated with ANNOVAR (49) and included in the online Supplementary Tables. We defined associations with P < 5 × 10⁻⁸ as genome-wide significant. LocusZoom plots (50) were used to visualize genome-wide significant loci, with LD for these plots calculated from HCHS/SOL genotype data. HaploReg v4.1 was used to explore the potential functional significance of identified lead variants (51).

To study the possibility that an associated locus may contain more than one distinct association signal, conditional analyses were done for each genome-wide significant locus using GENESIS. The lead variant was included as a covariate, with sequential inclusion of the next lead variant until no genome-wide significant signals remained within 1 Mb.

Replication of novel signals

We attempted to replicate the associations for two novel genome-wide significant signals (for TIBC) by accessing summary statistics from previous GWAS analyses for these variants. We contacted the lead authors of a meta-analysis of European populations (48 972 subjects from 11 discovery and eight replication cohorts) (6) and a GWAS of African Americans in the Jackson Heart Study (2347 subjects from a population-based study from the Jackson, MS, metropolitan area) (7) to obtain these results. In Europeans, results for TF were used to replicate TIBC associations, because TF and TIBC levels are linearly related (52,53). An association was considered replicated if the variant had the same direction of association and P < 0.05.

Sensitivity analysis excluding individuals with iron deficiency

We also repeated our genome-wide analyses for ferritin, TIBC, serum iron and SAT after excluding individuals with iron deficiency (defined by ferritin values below the reference range: <25 µg/l for men and <15 µg/l for women).

Generalization analysis

We tested loci identified in GWAS analyses in African American and European populations for generalization in HCHS/SOL. Following Sofer et al. (54), the generalization null hypothesis takes into account both the discovery study and the generalizing study (here HCHS/SOL). The generalization null hypothesis is rejected if evidence for SNP-phenotype association exists in both studies, with the same direction of association. As a first step, we estimated SNP-phenotype associations in HCHS/SOL using the same phenotype transformations that were used in the original study. Next, we tested the generalization null hypothesis, while controlling for the false discovery rate (FDR) due to multiple testing, by calculating FDR-controlling directional r-values (54) for all SNP associations. We concluded that an association was generalized if the SNP-phenotype association had an r-value < 0.05. All 12 variants from the European meta-analysis (6) were tested for generalization for all four iron traits in HCHS/SOL. However, the four variants from the African American GWAS were tested on only the significant trait from the original report, because association results were not available for all iron traits (7).

Analysis of iron variants with glycemic traits

We chose a single lead variant from each identified iron locus to test for associations with diabetes and six glycemic traits [2-h postload glucose from an OGTT, fasting glucose and insulin, HbA1c, HOMA-IR (insulin resistance) and HOMA-B (β-cell function)]. Five sentinel SNPs were assessed in this way, representing the most significant SNP at all genome-wide significant loci and the near genome-wide significant locus at PPP1R3B. Associations with P < 0.001 were accepted as significant, based on a correction factor of 1/35 (for testing five SNPs and seven traits).

To increase power, we derived an unweighted GRS based on the five sentinel SNPs (calculated with respect to iron raising alleles). Sample sizes for GRS analyses depended on the number of participants with no missing genotypes for any of the variants. We also derived a 10-variant, unweighted GRS, based on the 5 sentinel SNPs plus 5 previously reported variants (6,7) that generalized to HCHS/SOL (rs1800562 in HFE, rs7385804 in TFR2, rs4921915 near NAT2, rs651007 near ABO, rs141555380 in GAB3). The five-variant GRS focused on loci that were genome-wide significant in HCHS/SOL. The 10-variant GRS more comprehensively examined all independent loci that showed evidence for influencing iron traits in HCHS/SOL. None of the generalized variants was in significant LD with the five sentinel SNPs (r² < 0.1). Associations with P < 0.004 were considered significant for GRS analyses, based on a correction factor of 1/14 (for 2 GRS values and 7 traits).

We additionally assessed effects of our five iron-related loci on T2D status in a published trans-ethnic GWAS meta-analysis of European, East Asian, South Asian, Mexican and Mexican American ancestry individuals (55). Examination of previous European ancestry GWAS results was also performed for glycemic traits, including 2-hour post-load glucose from OGTT data (56), fasting glucose and insulin (unadjusted for BMI) (23), HbA1c (34), HOMA-IR and HOMA-B (22) in nondiabetic individuals. Data on glycemic traits were contributed by MAGIC investigators (downloaded from www.magicinvestigators.org). Data were not available from these GWAS meta-analyses for rs174529 (FADS2/MYRF) or rs6762719 (TF), so proxy SNPs were selected. We chose to analyze the lead European variants at these loci (rs8177240 for TF and rs174577 for FADS2), because both generalized to HCHS/SOL and were strongly associated with our iron traits of interest.

Look-up of HCHS/SOL hematological trait results

We performed look-ups for our five sentinel iron variants in a previous HCHS/SOL GWAS on hematological traits (red blood cell count, hemoglobin, hematocrit, MCV, MCH, mean corpuscular hemoglobin concentration, red blood cell distribution width, platelet count). All complete blood counts were assessed at the central laboratory at the University of Minnesota Medical Center, Fairview, in Minneapolis. We used a multiple testing corrected significance threshold of P < 0.001.

Supplementary Material

Supplementary Material is available at HMG online.

Acknowledgements

We thank the staff and participants of HCHS/SOL for their important contributions. We also thank the authors of reported iron-related analyses for providing look-up of novel associations (Benyamin et al. and Li et al.). The HCHS/SOL is a collaborative study supported by contracts from the National Heart, Lung, and Blood Institute (NHLBI) to the University of North Carolina (N01- HC65233), University of Miami (N01-HC65234), Albert Einstein College of Medicine (N01-HC65235), Northwestern University (N01-HC65236) and San Diego State University (N01-HC65237).

Conflict of Interest statement. None declared.

Funding

The Genetic Analysis Center at the University of Washington was supported by NHLBI and NIDCR contracts (HHSN268201300005C AM03 and MOD03). The following Institutes/Centers/Offices contribute to the HCHS/SOL through a transfer of funds to the NHLBI: National Institute on Minority Health and Health Disparities, National Institute on Deafness and Other Communication Disorders, National Institute of Dental and Craniofacial Research, National Institute of Diabetes and Digestive and Kidney Diseases, National Institute of Neurological Disorders and Stroke, NIH Institution-Office of Dietary Supplements. Dr Raffield is supported by a training grant (T32HL129982) from the NHLBI, and Dr Qi is supported by a Scientist Development Award (K01HL129892) from the NHLBI. The views expressed in this article are those of the authors and do not necessarily represent the views of the National Heart, Lung, and Blood Institute; the National Institutes of Health; or the U.S. Department of Health and Human Services.