β-Carotene 15,15′-monooxygenase 1 single nucleotide polymorphisms in relation to plasma carotenoid and retinol concentrations in women of European descent

Abstract

Background: Carotenoids have been hypothesized to reduce the risk of many diseases, but associations with intakes or blood concentrations may arise from other constituents of fruit and vegetables. Use of genetic variation in β-carotene 15,15′-monooxygenase 1 (BCMO1), a key enzyme in provitamin A carotenoid metabolism, as a surrogate for carotenoid exposure may aid in determining the role of carotenoids unconfounded by other carotenoid-containing food constituents, but important variants must be identified.

Objective: Our goal was to select BCMO1 single nucleotide polymorphisms (SNPs) that predict plasma carotenoid concentrations for use in future epidemiologic studies.

Design: We assessed the associations between 224 SNPs in BCMO1 ± 20 kb imputed from the 1000 Genomes Project EUR reference panel with plasma carotenoid and retinol concentrations by using 7 case-control data sets (n = 2344) within the Nurses’ Health Study, randomly divided into training (n = 1563) and testing (n = 781) data sets. SNPs were chosen in the training data set through stepwise selection in multivariate linear regression models; β-coefficients were used as weights in weighted gene scores.

Results: Two or 3 SNPs were selected as predictors of β-carotene, α-carotene, β-cryptoxanthin, and lutein/zeaxanthin. In the testing data set, the weighted gene scores were significantly associated with plasma concentrations of the corresponding carotenoid (P = 6.4 × 10⁻¹², 3.3 × 10⁻³, 0.02, and 1.8 × 10⁻¹⁷, respectively), and concentrations differed by 48%, 15%, 15%, and 36%, respectively, across extreme score quintiles.

Conclusions: SNPs in BCMO1 are associated with plasma carotenoid concentrations. Given adequate sample size, the gene scores may be useful surrogates for carotenoid exposure in future studies.

INTRODUCTION

Carotenoids are pigments obtained in the diet primarily from fruit and vegetables. Certain carotenoids can be converted to vitamin A, but the health effects of carotenoids may instead or additionally result from their antioxidant activity (1) or other mechanisms, such as selective induction of apoptosis (2). β-Carotene 15,15′-monooxygenase 1 (BCMO1)⁴ centrally cleaves dietary provitamin A carotenoids, including β-carotene, α-carotene, and β-cryptoxanthin, to form retinal, which is the first step in converting provitamin A carotenoids to vitamin A (3–6). Leung et al (7) identified 2 nonsynonymous single nucleotide polymorphisms (SNPs; A379V, rs7501331 and R267S, rs12934922) in the coding region of BCMO1 that were associated with higher fasting plasma β-carotene and reduced conversion of a pharmacologic dose of β-carotene to retinyl palmitate. Additionally, through a genome-wide association study (GWAS), Ferrucci et al (8) identified 4 other SNPs in modest (r² = 0.22–0.55) linkage disequilibrium (LD) with each other near BCMO1 that were related to circulating carotenoid concentrations. Specifically, rs6564851 was associated with increased β- and α-carotene and decreased lutein, zeaxanthin, and lycopene (8); rs6564851 was recently reported to reduce BCMO1 activity by 48% in females (9).

Other research suggests that carotenoids may reduce the risk of several diseases, including breast cancer. A meta-analysis suggests that dietary intake of α- and β-carotene is associated with a 9% and 6% lower breast cancer risk, respectively, comparing highest versus lowest intakes (10). These modest associations may result from attenuation of a stronger causal association due to misclassification of an individual's internal carotenoid dose as measured by his or her dietary intake. Alternatively, the apparent inverse associations may be due to residual confounding by personal characteristics and/or other beneficial components of fruit and vegetables. In accordance with the principles of Mendelian randomization (11, 12), genetic variation in carotenoid metabolism is likely to be uncorrelated with dietary and lifestyle factors, thus providing an unconfounded surrogate for carotenoid exposure that may also improve prediction of internal carotenoid dose when used in combination with dietary intake. Thus, the aim of this analysis was to identify important genetic variants related to plasma carotenoid concentrations and develop gene scores for use in future studies as unconfounded measures of carotenoid exposure. Retinol was included in these analyses because of its relation with the provitamin A carotenoids. BCMO1 was selected as the candidate locus of interest given its known role in provitamin A carotenoid metabolism (3–6) and previous identification of polymorphisms in the region associated with circulating carotenoid concentrations (7, 8).

SUBJECTS AND METHODS

Subjects and data sources

The Nurses’ Health Study (NHS) was established in 1976 when 121,700 female registered nurses in the United States aged 30–55 y completed a mailed questionnaire regarding medical history and other health-related exposures. The women have since been followed biennially by mailed self-administered questionnaires. In 1989–1990, 32,826 NHS participants provided blood samples. Details of the blood collection were published previously (13). In brief, women arranged to have their blood drawn and shipped, via overnight courier with an icepack, to our laboratory where samples were immediately centrifuged, portioned into aliquots, and stored in liquid nitrogen freezers. Ninety-seven percent of samples were received within 26 h of collection.

This study includes participants from case-control studies, nested within the blood subcohort, of breast cancer, cataract, colorectal cancer, colorectal adenoma, and myocardial infarction for whom plasma carotenoid and retinol concentrations were assayed. Because the subclinical presence of a major chronic disease could alter plasma biomarker concentrations, women with breast cancer, colon cancer, myocardial infarction, or diabetes diagnosed within 2 y after blood collection were excluded (n = 592).

Some individuals in the biomarker-assayed groups were included in GWASs of breast cancer [National Cancer Institute's Cancer Genetic Markers of Susceptibility (14)], ischemic heart disease, type 2 diabetes, or kidney stones. SNPs were imputed for these 2344 women, hereafter referred to as women with genome-wide scans. rs12934922, rs7501331, and rs6564851, which were previously identified (7, 8), and rs28380273, rs4889286, and rs4448930, which were identified during our analysis, were genotyped by the TaqMan assay in 3492, 3317, 872, 2205, 2495, and 2535 women, respectively (see Supplemental Table 1 under “Supplemental data” in the online issue). rs7501331 and rs6564851 were genotyped by GWAS in 661 and 2340 women, respectively (see Supplemental Table 1 under “Supplemental data” in the online issue). When both genotyped and imputed data were available, genotyped data were used; TaqMan genotyped data were prioritized over GWAS genotyped data when both were available. To eliminate concern over population stratification, the analyses were limited to women of European descent. The final sample consisted of 4135 women, of whom 2149 were controls. Of the 2344 women with genome-wide scans, 1197 were controls. The study protocol was approved by the Institutional Review Board of Brigham and Women's Hospital.

All participants in this analysis completed a semiquantitative food-frequency questionnaire (FFQ) in 1986 and/or 1990 (3632 completed both FFQs, 383 completed only the 1990 FFQ, and 120 completed only the 1986 FFQ). FFQs were considered complete if a woman reported a plausible total energy intake (600–3500 kcal/d) and left ≤70 food items blank. To reduce within-person variation, average nutrient intakes from the 1986 and 1990 FFQs were used when possible. To determine nutrient intakes, reported consumption frequencies of commonly used units or portion sizes of food items were multiplied by the nutrient contents of the specified unit and summed over all foods. Food nutrient values were derived primarily from USDA sources. Use of multivitamins and other supplements, as well as dose and duration of use, also were assessed and incorporated into retinol and β-carotene intake calculations. The reproducibility and validity of the FFQs were reported previously (15–17). The correlation between vitamin A intake from food and supplements estimated from the FFQ compared with four 1-wk dietary records was 0.49; the validity for specific carotenoids was not calculated (16). In another analysis, the correlations between 2 averaged FFQs and plasma carotenoid concentrations in nonsmoking women were 0.48 for α-carotene, 0.27 for β-carotene, 0.32 for β-cryptoxanthin, 0.21 for lycopene, and 0.27 for lutein; retinol was not calculated (18). Except for alcohol, nutrient intakes in this analysis were adjusted for total energy intake by the residual method (19).

Nondietary variables were obtained from questionnaires completed by participants. Menopausal status and postmenopausal hormone (PMH) use were determined from a questionnaire completed at the time of blood collection. Date of birth and height were determined from the baseline NHS questionnaire. BMI was calculated by using height reported in 1976 and weight reported on the blood collection questionnaire. If weight was missing, weight reported on the 1990 NHS questionnaire was used (n = 109). For the 3 women with both weights missing, weight reported on the 1988 NHS questionnaire was used. For the 9 women (4 with genome-wide scans) missing all weights or height, the median BMI value for all women was used. Smoking status was determined from the 1990 NHS questionnaire. Given the low percentage of smokers among women with available data (13.7%), women missing smoking status were assumed to be never/former smokers (n = 10 for all women, 4 for women with genome-wide scans).

Genotyping and imputation

Genotypes for rs12934922, rs7501331, rs28380273, rs4889286, and rs4448930 in breast cancer cases and controls were determined by measuring endpoint fluorescence by using the 5′ nuclease assay (TaqMan) on the ABI PRISM 7900HT Sequence Detection System (20). Genotypes for rs12934922, rs7501331, and rs6564851 in the colorectal cancer and adenoma cases and controls were determined according to the manufacturer's instructions by using the TaqMan OpenArray SNPGenotyping Platform (Applied Biosystems).

For women with genome-wide scans, SNP genotypes were imputed by GWAS data set in 2 steps. First, genotypes were phased by using MaCH (v1.0.16) (21). Then, Minimac (v.beta.2010) was used to impute (22) ∼11 million SNPs by using 1000 Genomes Project EUR data [release August 2010/National Center for Biotechnology Information (NCBI) 37] as the reference panel. Imputation results for each SNP were expressed as “allele dosages” (fractional values from 0 to 2) and were extracted for the BCMO1 gene region plus 20 kilobases up- and downstream (NCBI 37 chr16:81252296–81344747). For women in the Cancer Genetic Markers of Susceptibility and kidney stones GWAS groups (genotyped by using Illumina 550k and 610Q, respectively), 17 SNPs in the region were GWAS genotyped. Thirty SNPs were GWAS genotyped for women in the ischemic heart disease and type 2 diabetes GWAS groups (genotyped by using Affymetrix 6.0). SNPs with imputation r² ≥ 0.50 in all 4 GWAS data sets were initially considered for further analysis (n = 225 of 548). Later, rs28380273 was removed because of poor genotyping results, which left 224 SNPs. Reported minor allele frequencies and imputation r² are weighted averages based on sample size across the 4 individual GWAS data sets.

Plasma biomarker assays data set

Plasma carotenoid and retinol concentrations were measured by the Micronutrient Analysis Laboratory in the Department of Nutrition at the Harvard School of Public Health by using reversed-phase HPLC by the methods described by El-Sohemy et al (23). Lycopene was assayed as total lycopene. Blinded quality-control samples (10%) were randomly placed throughout sample boxes. Laboratory technicians also were blinded to case-control status. CVs were calculated for each biomarker within each laboratory batch. Across the 12 batches for each biomarker, CVs were generally <15%, except for 1 batch for β-carotene (CV = 20.7%), 2 batches for β-cryptoxanthin (CV = 19.4% and 20.6%), 1 batch for lycopene (CV = 17.5%), and 3 batches for retinol (CV = 16.2%, 17.8%, and 25.8%). Values were missing for some biomarkers because of laboratory technical difficulties (9 α-carotene, 10 β-carotene, 10 β-cryptoxanthin, 9 lutein/zeaxanthin, 9 lycopene, and 13 retinol).

To account for laboratory variation that was evident across batches, plasma concentrations were standardized to an average batch. β-Coefficients from linear regression models adjusted for the covariates listed below with batch indicators as the independent variables and natural log–transformed biomarker concentrations as the dependent variable were averaged. The differences between coefficients for each biomarker measurement's batch and the average across batches were subtracted from each biomarker value. A similar method was used previously to account for study effects among 11 blood pressure studies (24).

Total plasma cholesterol was assayed in 15 batches by using the enzymatic methods described by Allain et al (25). CVs were 2.1–15.6%. Plasma cholesterol data were not available for 482 women (355 with genome-wide scans), and 1 outlier identified by the generalized extreme Studentized deviate many-outlier detection method (26) was set to missing.

Statistical analysis

All plasma biomarker values were natural log transformed to improve normality, and all SNP dosages or genotypes were coded as additive variables (ie, allele counts). Outlying biomarker concentrations among all women were identified within batch by a generalized extreme Studentized deviate many-outlier detection method (26) and set to missing (6 of 4126 α-carotene, 9 of 4125 β-carotene, 6 of 4125 β-cryptoxanthin, 3 of 4126 lutein/zeaxanthin, 24 of 4126 lycopene, and 7 of 4122 retinol).

Biomarker residuals were derived from multivariate linear regression of each plasma biomarker on covariates. Correlations bewteen back-transformed plasma biomarker residuals were evaluated by Spearman correlation coefficients. For reduction of extraneous variation in biomarker concentrations, multivariate linear regression models included the following covariates: age at blood collection (y), case-control status (as 7 sets of indicator variables coded as not in data set, control, or case for breast cancer, colon cancer, colorectal adenoma, cataract, myocardial infarction, kidney stones, and diabetes), menopausal status and PMH (premenopausal; postmenopausal, no PMH; postmenopausal, current PMH; or dubious/unknown), BMI (kg/m²), smoking status (never/former, current <15 cigarettes/d, or current ≥15 cigarettes/d or an unknown amount/d), plasma cholesterol (mg/dL or missing indicator), alcohol intake (g/d), energy-adjusted fat intake (g/d), and energy-adjusted intake of the nutrient of interest (μg/d for carotenoids; IU/d for retinol). Plasma β-carotene, α-carotene, and β-cryptoxanthin models also were adjusted for dietary retinol intake (IU/d) to account for possible reduced conversion of provitamin A carotenoids in the presence of high preformed retinol intake. Results from multivariate linear regressions of each plasma biomarker on each SNP among women with genome-wide scans were used to create regional association plots based on NCBI Build 36 positions by using SNAP version 2.2 (27).

To select SNPs associated with the plasma biomarkers and provide unbiased estimates of model performance after SNP selection, women with genome-wide scans were randomly divided into 2 groups: a two-thirds training subset (n = 1563) and a one-third testing subset (n = 781). In the training subset, SNPs were chosen by stepwise selection by using SAS PROC GLMSELECT and the Schwartz Bayesian information selection criterion with the previously listed covariates forced into the model. In the testing subset, unweighted gene scores were calculated by adding dosages of biomarker-increasing alleles across selected SNPs. Weighted gene scores were calculated by multiplying each selected SNP's biomarker-increasing allele dosage by its β-coefficient in the final selected model and adding across SNPs.

Composite LD among SNPs was evaluated by Pearson correlation coefficients. Percentage variation explained by the genetic terms (gene scores or SNPs) was determined by a multiple partial R², which was calculated by subtracting the sum of squared errors (SSE) for the model including both the genetic term and covariates from the SSE for the covariate-only model and then dividing by the SSE for the covariate-only model. This method provides the percentage of variation in the residual biomarker concentration explained by the genetic term after accounting for variation explained by the covariates. Covariate-adjusted geometric mean plasma biomarker concentrations were determined by back-transforming the least-squares means from SAS PROC GLM. Gene-diet interactions were assessed among all women with genome-wide scans by including cross-product terms for dietary intake (represented as quintile medians of the nutrient of interest) and the continuous gene score in multivariate models as well as the main effects of dietary intake (quintile medians) and gene score. All reported P values are 2-sided and unadjusted for multiple testing. All statistical analyses were performed by using SAS version 9 (SAS Institute).

RESULTS

Participant characteristics are shown in Table 1. Spearman correlations among back-transformed plasma biomarker concentration residuals after regressing out covariates are shown elsewhere (see Supplemental Table 2 under “Supplemental data” in the online issue). The β-carotene and α-carotene residuals were the most correlated (r = 0.74), and the β-carotene and retinol residuals were the least correlated (r = 0.02).

TABLE 1.

Participant characteristics at the time of blood collection

	All women		Women with genome-wide scans		Women without genome-wide scans 1
Characteristic	n	Value	n	Value	n	Value
Age (y)	4135	57.8 ± 6.82	2344	58.8 ± 6.4	1791	56.5 ± 7.2
BMI (kg/m2)	4135	25.3 ± 4.4	2344	25.4 ± 4.6	1791	25.1 ± 4.2
Plasma cholesterol (mg/dL)	3653	219 ± 40	1989	222 ± 40	1664	217 ± 40
Postmenopausal or unknown status (%)	3411	82.5	2059	87.8	1352	75.5
Current postmenopausal hormone use (%)3	1482	48.8	951	50.7	531	45.6
Current smoking <15 cigarettes/d (%)	230	5.6	148	6.3	82	4.6
Current smoking ≥15 cigarettes/d (%)	336	8.1	213	9.1	123	6.9
Daily dietary intake4
β-Carotene (μg)	4135	4306 ± 2209	2344	4394 ± 2322	1791	4191 ± 2046
α-Carotene (μg)	4135	762 ± 505	2344	775 ± 520	1791	745 ± 485
β-Cryptoxanthin (μg)	4135	177 ± 94	2344	179 ± 94	1791	174 ± 94
Lutein/zeaxanthin (μg)	4135	2881 ± 1608	2344	2914 ± 1673	1791	2836 ± 1518
Lycopene (μg)	4135	6357 ± 3301	2344	6413 ± 3382	1791	6285 ± 3192
Retinol (IU)	4135	4184 ± 3957	2344	4256 ± 4137	1791	4090 ± 3707
Total fat (g)	4135	57 ± 9	2344	56 ± 9	1791	57 ± 9
Alcohol (g)	4135	6 ± 10	2344	6 ± 10	1791	6 ± 9
Plasma biomarker concentration (μg/L)
β-Carotene	4116	301 ± 250	2331	303 ± 258	1785	300 ± 239
α-Carotene	4120	74 ± 54	2330	74 ± 55	1790	75 ± 53
β-Cryptoxanthin	4119	84 ± 50	2332	85 ± 50	1787	83 ± 51
Lutein/zeaxanthin	4123	187 ± 78	2333	187 ± 78	1790	187 ± 78
Lycopene	4102	432 ± 194	2323	429 ± 195	1779	436 ± 193
Retinol	4115	568 ± 138	2340	578 ± 136	1775	554 ± 139

¹Differences between women with and without genome-wide scans for the variables presented in the table were tested by Wilcoxon's 2-sample and chi-square tests by using a Bonferroni-corrected α = 0.0025. Age (P < 0.0001), plasma cholesterol concentration (P = 0.0002), menopausal status (P < 0.0001), smoking status (P = 0.0007), and plasma retinol concentration (P < 0.0001) were significantly different between groups.

²Mean ± SD (all such values).

³Among postmenopausal women not missing data.

⁴Energy-adjusted, except for alcohol.

Associations between the 3 previously identified SNPs (7, 8) and each plasma biomarker among all women, as determined by multivariate linear regression β-coefficients, are shown elsewhere (see Supplemental Table 3 under “Supplemental data” in the online issue). The rs6564851 G allele was significantly associated with higher β-carotene, α-carotene, and β-cryptoxanthin and with lower lutein/zeaxanthin (P = 5.2 × 10⁻²⁶, 1.5 × 10⁻⁴, 2.2 × 10⁻⁴, and < 1.0 × 10⁻³⁰, respectively); no association was observed with lycopene or retinol. The rs12934922 T allele was significantly associated with higher β-carotene, α-carotene, β-cryptoxanthin, and lutein/zeaxanthin (P = 1.1 × 10⁻¹², 3.1 × 10⁻⁷, 2.4 × 10⁻³, and 5.9 × 10⁻⁶, respectively) and was selected for the β-carotene score. The SNP was not associated with lycopene or retinol. The rs7501331 T allele was significantly associated with lower β-carotene and α-carotene (P = 1.6 × 10⁻⁵ and 0.03, respectively); no associations were observed with the other biomarkers. Results for rs12934922 and rs7501331 were similar when restricted to genotyped data (see Supplemental Table 3 under “Supplemental data” in the online issue); results are shown only for rs6564851 genotyped data because the SNP was only imputed in 4 women with available GWAS data. When we limited the rs7501331 analysis to genotyped data among currently nonsmoking fasting controls with a BMI (in kg/m²) <30 who did not report use of multivitamins, β-carotene supplements, or vitamins A, C, or E to best approximate the inclusion criteria of the previous study (7), the T allele was nonsignificantly inversely associated with plasma β-carotene (n = 427, P = 0. 26; data not shown). A gene score composed of counts of T alleles in rs12934922 and rs7501331 was borderline-significantly positively associated with plasma β-carotene concentration in this group (n = 369, P = 0.07; data not shown).

Regional association plots for each biomarker and 224 SNPs for women with genome-wide scans indicate that the most significant SNPs for β-carotene (rs12925563), β-cryptoxanthin (rs12918164), and lutein/zeaxanthin (rs56389890) were slightly upstream of BCMO1 near PKD1L2 (polycystic kidney disease 1-like 2) (Figure 1). The most significant SNPs for α-carotene (rs12923433) and retinol (rs12934922) were in the center of BCMO1.

FIGURE 1.

Regional (NCBI 36 chr16:79829797–79882248) association plots between SNPs in BCMO1 ± 20 kb and plasma carotenoid and retinol concentrations among women with genome-wide scans (n = 2323–2340). Plotted P values are from multivariate linear regression of each natural log–transformed plasma biomarker concentration on additively modeled SNPs adjusted for the following covariates: age, case-control status, cholesterol (mg/dL or missing indicator), BMI (kg/m²), smoking (never/former, current <15 cigarettes/d, current ≥15 cigarettes/d), menopausal status and PMH use (premenopausal; postmenopausal, no PMH; postmenopausal, PMH; unknown), alcohol intake (g/d), total fat intake (g/d), and dietary intake of the outcome of interest (μg/d for carotenoids or IU/d for retinol). Dietary retinol intake (IU/d) was also included in the models for α-carotene, β-carotene, and β-cryptoxanthin. A: β-carotene; B, α-carotene; C, β-cryptoxanthin; D, lutein/zeaxanthin; E, lycopene; F, retinol. Left y axis = observed (-logP); right y axis = recombination rate (centiMorgans/megabase); x axis = chromosome 16 position (human genome 18) (kb). chr, chromosome; NCBI, National Center for Biotechnology Information; PMH, postmenopausal hormone; SNPs, single nucleotide polymorphisms.

In the original stepwise selection from 225 SNPs for an association with plasma biomarker concentrations in the training data set, zero to 3 SNPs were selected for each biomarker (3 for β-carotene, α-carotene, and β-cryptoxanthin; 1 for lutein/zeaxanthin; and 0 for lycopene and retinol). rs28380273 was selected for lutein/zeaxanthin and also was the individual SNP most significantly associated with β-carotene (P = 6.3 × 10⁻²¹), despite not being selected by stepwise selection for the β-carotene gene score. We thus sought confirmation of these associations via genotyping. The SNP was subsequently removed from further analyses because of the 85% genotyping success rate and lack of Hardy-Weinberg equilibrium (P < 1 × 10⁻³⁰). After repeating stepwise selection from the 224 remaining SNPs, SNP selection changed only for lutein/zeaxanthin, for which 2 other SNPs were selected.

Associations between the selected SNPs and respective biomarkers are shown in Table 2 for the training and testing subsets. Because no SNPs were selected for lycopene or retinol, these biomarkers are not included in Table 2. rs4889293 was selected for both α-carotene and β-cryptoxanthin with the G allele being associated with higher concentrations of both carotenoids. rs4889286 was selected for both β- and α-carotene, with the T allele being associated with higher concentrations of both carotenoids. rs4448930 was selected for β-carotene, α-carotene, and β-cryptoxanthin, with the C allele being associated with higher concentrations of all 3 carotenoids. Because of their selection in multiple gene scores, we genotyped rs4448930 and rs4889286 for confirmation. After the gene scores derived in the training subset from imputed data were applied to genotyped data in the testing subset, associations for rs4448930 were substantially attenuated. (Table 2). Because of this attenuation, and the moderate-to-high composite LD between the remaining selected SNPs and previously identified SNPs (see Supplemental Table 4 under “Supplemental data” in the online issue), we elected to exclude rs4448930 from the 3 gene scores for which it was selected.

TABLE 2.

Associations between selected SNPs in BCMO1 and plasma carotenoid concentrations among women with genome-wide scans by subset¹

			Training subset (n = 1554–1556)				Testing subset (n = 775–778)
Biomarker and SNP	MAF (allele)2	Imputation3 r2	Imputation4	Multiple partial5 R2	β (SE)6	P value	Imputation	Multiple partial R2	β (SE)	P value
			%				%
β-Carotene
rs4889286	0.50 (T)	0.98	100	0.061	0.1762 (0.022)	8.5 × 10−16	29	0.061	0.1748 (0.030)	9.0 × 10−9
rs12934922	0.44 (T)	0.90	25		0.1377 (0.022)	5.9 × 10−10	26		0.1412 (0.031)	5.9 × 10−6
rs4448930	0.15 (C)	0.74	100		0.09747 (0.034)	3.8 × 10−3	28		0.01753 (0.042)	0.68
α-Carotene
rs4889293	0.42 (G)	0.85	73	0.027	0.1183 (0.022)	8.7 × 10−8	71	0.012	0.06480 (0.032)	0.04
rs4448930	0.15 (C)	0.74	100		0.09750 (0.032)	2.3 × 10−3	28		−0.006095 (0.041)	0.88
rs4889286	0.50 (T)	0.98	100		0.06581 (0.020)	1.3 × 10−3	29		0.06844 (0.029)	0.02
β-Cryptoxanthin
rs4448930	0.15 (C)	0.74	100	0.019	0.09872 (0.027)	2.2 × 10−4	28	0.008	0.02774 (0.035)	0.43
rs12918164	0.17 (A)	0.83	100		0.07583 (0.024)	1.9 × 10−3	100		0.05948 (0.035)	0.09
rs4889293	0.42 (G)	0.85	73		0.06043 (0.018)	1.0 × 10−3	71		0.04819 (0.027)	0.07
Lutein/zeaxanthin
rs56389940	0.36 (A)	0.94	100	0.074	0.1636 (0.015)	2.6 × 10−27	100	0.097	0.1792 (0.021)	2.0 × 10−17
rs10048138	0.14 (A)	0.92	73		0.06580 (0.020)	1.3 × 10−3	71		0.1334 (0.030)	1.1 × 10−5

¹Multivariate linear regression of each natural log–transformed plasma biomarker concentration (μg/L) on additively modeled SNPs selected by stepwise selection by using the Schwartz Bayesian information criterion and the following covariates were forced into the model: age, case-control status, cholesterol (mg/dL, or missing indicator), BMI (kg/m²), smoking (never/former, current <15 cigarettes/d, or current ≥15 cigarettes/d), menopausal status and PMH use (premenopausal; postmenopausal, no PMH; postmenopausal, PMH; or unknown), alcohol intake (g/d), total fat intake (g/d), and dietary intake of the outcome of interest (μg/d for carotenoids or IU/d for retinol). Dietary retinol intake (IU/d) was also forced into the models for β-carotene, α-carotene, and β-cryptoxanthin. Results are shown for rs4889286 and rs4448930 original imputed values in the training subset on which selection was performed. Results in the testing subset include later genotyped values when available. MAF, minor allele frequency; PMH, postmenopausal hormone; SNP, single nucleotide polymorphism; SSE, sum of squared errors.

²Weighted average MAF among all women with genome-wide scans.

³Minimac imputation r² among all women with genome-wide scans.

⁴Percentage of women with imputed data.

⁵Multiple partial R²: SSE for the covariate-only model − SSE for the model with SNPs and covariates divided by SSE for the covariate-only model.

⁶β-Coefficient from the final selected model described above.

After finalizing the gene scores, we determined geometric mean plasma concentrations for each weighted gene score quintile among the testing subset (Table 3). The β-carotene–weighted gene score represented the largest difference across extreme quintiles (48%), and the α-carotene- and β-cryptoxanthin-weighted gene scores represented the smallest difference (15%). Results for the unweighted gene scores were similar to the results for the weighted gene scores (data not shown).

TABLE 3.

BCMO1-weighted gene score associations with plasma carotenoid concentrations among the testing subset of women with genome-wide scans (n = 775–778)¹

	Geometric mean concentration by gene score quintile (95% CI) 2
Biomarker	1	2	3	4	5	β (SE)3	P value	Multiple partial4R2
	μg/L
β-Carotene5	194 (176, 214)	197 (178, 218)	218 (190, 250)	244 (227, 262)	288 (264, 314)	1.076 (0.154)	6.4 × 10−12	0.061
α-Carotene6	57 (52, 63)	57 (52, 62)	54 (49, 60)	61 (56, 66)	66 (60, 71)	0.7362 (0.250)	3.3 × 10−3	0.011
β-Cryptoxanthin7	68 (63, 73)	69 (64, 75)	74 (69, 80)	70 (64, 75)	78 (72, 84)	0.8191 (0.354)	0.02	0.007
Lutein/zeaxanthin8	140 (133, 147)	167 (157, 178)	176 (166, 186)	190 (180, 202)	191 (181, 203)	1.099 (0.126)	1.8 × 10−17	0.092

¹PMH, postmenopausal hormone; SSE, sum of squared errors.

²Covariate-adjusted geometric mean plasma biomarker concentrations determined by back-transforming the least-squares means from the linear regression model described below with gene scores modeled as quintile indicator variables.

³β-Coefficient and SE from multivariate linear regression of each natural log–transformed plasma biomarker concentration (μg/L) on corresponding continuous weighted gene score were adjusted for age, case-control status, cholesterol (mg/dL, or missing indicator), BMI (kg/m²), smoking (never/former, current <15 cigarettes/d, or current ≥15 cigarettes/d), menopausal status and PMH use (premenopausal; postmenopausal, no PMH; postmenopausal, PMH; or unknown), alcohol intake (g/d), total fat intake (g/d), and dietary intake of the outcome of interest (μg/d for carotenoids or IU/d for retinol). β-Carotene, α-carotene, and β-cryptoxanthin were also adjusted for dietary retinol intake (IU/d).

⁴Multiple partial R²: SSE for the covariate-only model − SSE for the model with gene score and covariates divided by SSE for the covariate-only model.

⁵Weighted gene score = (0.1637 × rs4889286 T) + (0.1274 × rs12934922 T); weights derived from testing data set were 0.1744 and 0.1397, respectively.

⁶Weighted gene score = (0.06530 × rs4889286 T) + (0.1075 × rs4889293 G); weights derived from testing data set were 0.06860 and 0.06538, respectively.

⁷Weighted gene score = (0.07709 × rs12918164 A) + (0.04916 × rs4889293 G); weights derived from testing data set were 0.05722 and 0.04560, respectively.

⁸Weighted gene score = (0.1636 × rs56389940 A) + (0.06580 × rs10048138 A); weights derived from testing data set were 0.1792 and 0.1334, respectively.

Spearman correlations between each weighted gene score and dietary intake of the corresponding carotenoid among women with genome-wide scans (n = 2344) ranged from −0.03 to 0.03 (P ≥ 0.17). We may thus assume the weighted gene scores are unassociated with dietary intake of other constituents of carotenoid-containing foods as well.

To compare the ability of dietary carotenoid intake compared with BCMO1 SNPs to explain variation in plasma carotenoid concentrations, we determined the multiple partial R² for the dietary carotenoid variables adjusted for all covariates, including the relevant weighted gene score. The percentages of residual variation explained by dietary intake of the specific carotenoids in the testing data set were 9% for β-carotene, 8% for α-carotene, 8% for β-cryptoxanthin, and 6% for lutein/zeaxanthin.

Each weighted gene score for one carotenoid was significantly associated with plasma concentrations of other carotenoids among the testing subset (n = 775–780, data not shown). Whereas the β-carotene–weighted gene score was associated with 48% higher β-carotene across extreme quintiles (Table 3), it was also associated with 24% higher α-carotene (P-continuous score = 1.1 × 10⁻³) and 18% lower lutein/zeaxanthin (P = 6.2 × 10⁻⁷), explaining 1.4% and 3.3% of the residual variation, respectively. The score was borderline-significantly associated with 10% higher β-cryptoxanthin (P = 0.07) and was not associated with lycopene or retinol. The α-carotene–weighted gene score, associated with 15% higher α-carotene across extreme quintiles (Table 3), was also associated with 40% higher β-carotene (P = 8.2 × 10⁻⁹), 10% higher β-cryptoxanthin (P = 0.04), and 10% lower lutein/zeaxanthin (P = 0.01), explaining 4.3%, 0.6%, and 0.9% of the residual variation, respectively. The score was not associated with lycopene or retinol. Whereas the β-cryptoxanthin–weighted gene score was associated with 15% higher β-cryptoxanthin across extreme quintiles (Table 3), the score also was associated with 25% higher β-carotene (P = 3.3 × 10⁻⁵) and 6% lower retinol (P = 4.1 × 10⁻³), explaining 2.0% and 1.1% of the residual variation, respectively. The score was borderline-significantly associated with 3% higher α-carotene (P = 0.09) and was not associated with lutein/zeaxanthin or lycopene. The lutein/zeaxanthin-weighted gene score, associated with 36% higher lutein/zeaxanthin across extreme quintiles (Table 3), was also associated with 14% lower α-carotene (P = 7.7 × 10⁻³) and 28% lower β-carotene (P = 1.1 × 10⁻⁸), explaining 0.9% and 4.2% of the residual variation, respectively. The score was borderline-significantly associated with 7% lower β-cryptoxanthin (P = 0.06) and not associated with lycopene or retinol. Because the SNPs in the other gene scores included or were in composite LD with at least one of the SNPs in the β-carotene gene score, we examined the relation between rs4889286 and rs12934922 and plasma biomarker concentrations among all women with available genotyped data (see Supplemental Table 5 under “Supplemental data” in the online issue). Gene scores composed of these SNPs were significantly associated with all biomarkers other than lycopene (P = 0.20 and 0.18 for unweighted and weighted scores, respectively; data not shown). An increasing number of T alleles was associated with higher β-carotene, α-carotene, and β-cryptoxanthin and lower retinol. In comparison with rs12934922, lutein/zeaxanthin was strongly influenced by rs4889286, in which an increasing number of T alleles was associated with lower lutein/zeaxanthin.

The β-carotene–, α-carotene–, and β-cryptoxanthin–weighted gene scores did not appear to interact with dietary intake of their respective nutrients (Table 4). A significant (P = 0.04) weighted gene score × diet interaction was found for lutein/zeaxanthin, wherein the association between the weighted gene score and plasma lutein/zeaxanthin was weakest in women with the lowest dietary lutein/zeaxanthin intake. P values for the unweighted gene score × diet cross-product terms were 0.69 for β-carotene, 0.05 for α-carotene, 0.68 for β-cryptoxanthin, and 0.13 for lutein/zeaxanthin (data not shown).

TABLE 4.

BCMO1-weighted gene score associations with plasma carotenoid concentration stratified by quintile of dietary intake of the nutrient of interest among women with genome-wide scans (n = 465–468 per quintile)¹

	Diet quintile 1		Diet quintile 2		Diet quintile 3		Diet quintile 4		Diet quintile 5
Biomarker	β (SE)2	P value	β (SE)2	P value	β (SE)2	P value	β (SE)2	P value	β (SE)2	P value	P-interaction3
β-Carotene4	1.275 (0.193)	1.2 × 10−10	0.9422 (0.186)	6.2 × 10−7	0.9096 (0.208)	1.5 × 10−5	0.8386 (0.212)	8.7 × 10−5	1.283 (0.219)	8.8 × 10−9	0.77
α-Carotene5	1.617 (0.316)	4.7 × 10−7	0.7327 (0.299)	0.01	0.8791 (0.316)	5.7 × 10−3	0.1871 (0.324)	0.56	0.9488 (0.360)	8.7 × 10−3	0.26
β-Cryptoxanthin6	1.327 (0.510)	9.6 × 10−3	1.115 (0.477)	0.02	0.8857 (0.443)	0.05	0.0724 (0.451)	0.87	1.324 (0.446)	3.2 × 10−3	0.62
Lutein/zeaxanthin7	0.6766 (0.156)	1.7 × 10−5	1.287 (0.161)	7.4 × 10−14	0.9160 (0.166)	5.9 × 10−8	1.106 (0.166)	8.9 × 10−11	1.188 (0.177)	6.1 × 10−11	0.04

¹PMH, postmenopausal hormone.

²β-Coefficient and SE from multivariate linear regression of each natural log–transformed plasma biomarker concentration (μg/L) on corresponding continuous weighted gene score adjusted for age, case-control status, cholesterol (mg/dL, or missing indicator), BMI (kg/m²), smoking (never/former, current <15 cigarettes/d, or current ≥15 cigarettes/d), menopausal status and PMH use (premenopausal; postmenopausal, no PMH; postmenopausal, PMH; or unknown), alcohol intake (g/d), total fat intake (g/d), and dietary intake of the outcome of interest (μg/d for carotenoids or IU/d for retinol). β-Carotene, α-carotene, and β-cryptoxanthin were also adjusted for dietary retinol intake (IU/d).

³value for gene score × diet interaction cross-product term included in the multivariate linear regression models described above by using diet quintile medians for the cross-product and diet main effect.

⁴Weighted gene score = (0.1637 × rs4889286 T) + (0.1274 × rs12934922 T).

⁵Weighted gene score = (0.06530 × rs4889286 T) + (0.1075 × rs4889293 G).

⁶Weighted gene score = (0.07709 × rs12918164 A) + (0.04916 × rs4889293 G).

⁷Weighted gene score = (0.1636 × rs56389940 A) + (0.06580 × rs10048138 A).

To assess the influence on our results of subsequent disease diagnosed after blood collection, we ran multivariate linear regression models with weighted gene scores in the testing subset, excluding cases (n = 404–406; data not shown). The associations between the weighted gene scores and the respective biomarker concentrations were not materially different for β-carotene, β-cryptoxanthin, and lutein/zeaxanthin, although the association was no longer significant for β-cryptoxanthin. The association between the α-carotene–weighted gene score and plasma α-carotene was both attenuated and no longer significant.

DISCUSSION

Here, BCMO1 SNPs were associated with plasma carotenoid concentrations. Similar to the previous GWAS (8), the rs6564851 G allele was associated with higher β- and α-carotene and lower lutein/zeaxanthin. Rs4889286 or rs56389940, in composite LD with rs6564851 and in the same upstream area, was selected for the β-carotene, α-carotene, and lutein/zeaxanthin gene scores. SNPs upstream of BCMO1, rs7501331 + rs12934922, and rs7501331 alone were previously reported to reduce BCMO1 activity and were associated with fasting plasma β-carotene (7, 9). Here, the rs12934922 T allele was positively associated with plasma β-carotene, and either rs12934922 or rs4889293, in composite LD with rs12934922, was selected for the provitamin A carotene gene scores. The inverse association between the rs7501331 T allele and plasma β-carotene, however, was opposite from that expected based on the positive association reported previously (7). The magnitude of our inverse association was similar, although nonsignificant, when we limited our data to an approximation of the previous study's data, but other differences between study groups, such as age, remained.

The provitamin A carotenoid-weighted gene scores were significantly associated with their corresponding plasma concentrations in the testing subset. Human or chicken BCMO1 catalyzes central cleavage of carotenoids with ≥1 unsubstituted β-ionone ring, explaining its activity toward β-carotene, α-carotene, and β-cryptoxanthin (4, 6). Chicken BCMO1, the gene for which has 67% sequence identity with human BCMO1 (4), is catalytically more efficient with a higher retinoid conversion yield with β-carotene as substrate compared with α-carotene and β-cryptoxanthin (6). Thus, as observed, we expected a stronger association between BCMO1 variants and plasma β-carotene than α-carotene or β-cryptoxanthin. The lack of a major association with plasma retinol and the low correlation between the plasma carotenoids and retinol also was expected. A recent GWAS of circulating retinol did not detect any genome-wide significantly associated SNPs in or near BCMO1 (28). Circulating retinol concentrations are maintained homeostatically in the absence of critically low liver stores (29). All plasma retinol concentrations were >200 μg/L, implying sufficient vitamin A status (30) and thus limited interindividual variation. Regardless, the β-cryptoxanthin gene score was significantly inversely associated with plasma retinol in the testing subset, and the β-carotene gene score, when weighted for retinol, was significantly associated with plasma retinol in all women with genotyped data for the relevant SNPs. Liver retinol stores may be more affected by BCMO1 variation, as suggested by the reduced liver vitamin A concentrations in β-carotene–fed BCMO1 knockout mice (31, 32). Overall, our results support the hypothesis that BCMO1 variants affect provitamin A carotenoid conversion, presumably because of altered BCMO1 function (7, 9).

BCMO1 is not thought to catalyze lutein/zeaxanthin or lycopene cleavage because of the lack of unsubstituted β-ionone rings (4, 6). Our null lycopene results thus agree with biological expectation and provide further evidence that BCMO1 does not cleave lycopene. Our observed lutein/zeaxanthin associations, while confirming previous findings (8), remain biologically unsubstantiated. The associations may be due to an indirect effect through provitamin A carotenoids (8), because carotenoids interact at many stages of absorption and metabolism (33). BCMO1 SNPs resulting in higher absorption of intact β-carotene may lead to reduced plasma lutein/zeaxanthin. The area under the serum lutein-versus-time curve with oral lutein plus β-carotene is smaller than the area with lutein alone, which suggests that β-carotene can reduce lutein absorption (34). β-Carotene supplementation in the Alpha-Tocopherol, Beta-Carotene Cancer Prevention Study in male smokers reduced serum lutein concentrations by 11% (35), but plasma lutein concentrations were not significantly different after β-carotene supplementation in mostly nonsmoking men in the Physicians Health Study (36). Alternatively, β,β-carotene-9′,10′-oxygenase (BCDO2), which catalyzes asymmetric cleavage of provitamin A and nonprovitamin A carotenoids, may be involved (37–39). Both BCMO1 and BCDO2 are expressed in several human tissues, including the small intestine epithelium and liver hepatocytes (40). BCMO1-deficient mice have elevated liver BCDO2 mRNA (41), so SNPs reducing BCMO1 function may indirectly alter BCDO2 expression. BCDO2’s apparent preferential cleavage of lutein and zeaxanthin over β-cryptoxanthin (37) may explain the stronger effect of this indirect pathway on lutein/zeaxanthin compared with the other carotenoids. Finally, BCMO1 may have as yet unknown direct activity on lutein/zeaxanthin.

The gene scores may be useful markers of carotenoid exposure in future studies. In an analysis of plasma carotenoids and breast cancer risk in the NHS, the multivariate-adjusted ORs for the comparison of the top with the bottom quintiles were 0.64 (0.47–0.88) for a 113-μg/L difference in α-carotene, 0.73 (0.53–1.02) for a 521-μg/L difference in β-carotene, and 0.74 (0.55–1.01) for a 148-μg/L difference in lutein/zeaxanthin (42). From these ORs and our observed genetic associations with the plasma carotenoids, the expected ORs of breast cancer for comparisons of extreme weighted gene score quintiles were 0.97 (α-carotene), 0.94 (β-carotene), and 0.90 (lutein/zeaxanthin). Although the expected ORs might be larger for the weighted gene scores because they presumably represent lifetime exposure rather than a single blood measurement, a large sample may be required when these scores are used in future studies. Some SNPs were selected as predictors of multiple carotenoids and/or were in composite LD with other selected SNPs, causing significant associations between gene scores for one carotenoid with concentrations of other carotenoids. Gene scores composed of rs4889286 and rs12934922 predicted plasma concentrations of all biomarkers other than lycopene. This pleiotropy may not only limit attribution of genetic associations to specific carotenoids but may also attenuate associations if both provitamin A carotenoids and lutein/zeaxanthin are similarly associated with disease risk, given their opposite associations with some BCMO1 SNPs. Whereas no SNPs were selected for retinol, pleiotropy also may exist for gene scores modeling differential converter phenotypes because they may be associated with site-specific retinol concentrations that are less regulated than plasma concentrations.

Exclusion of other genes was a limitation. We also did not adjust P values for multiple testing. Some of our nominally statistically significant results may be false positive; thus, results from the more exploratory analyses (eg, gene-diet interaction tests) should be interpreted with special care. However, our main findings (replication of the β-carotene and lutein/zeaxanthin gene scores) are quite robust, passing a conservative Bonferroni correction for the number of marker-biomarker tests conducted and surpassing a widely used genome-wide significance threshold (P < 5 × 10⁻⁸). The failure of the TaqMan genotyping assay for rs28380273 exemplifies a limitation of our reliance on mostly imputed genetic data. This failure may have been due to the apparently triallelic nature of rs28380273 that the other selected SNPs do not seem to share (43). The high LD between particular SNPs in this region combined with random variation also means that SNP selection is sensitive to the data set; other analysts may select different SNPs. Differences also may occur in non-European study populations given the variation in allele frequencies noted for some BCMO1 SNPs across ethnicities (7, 9), which limits this study's generalizability. Despite limitations, the gene scores may still be useful in epidemiologic analyses because they should be largely unconfounded and may be more readily available in large data sets than dietary intakes or plasma concentrations. The SNPs were selected by using a large data set with thorough covariates and were comprehensive across the entire BCMO1 region ± 20 kb.

Our results support an important role for BCMO1 variation on plasma carotenoid concentrations, particularly β-carotene and lutein/zeaxanthin. Genetic surrogates for 4 plasma carotenoids are now available for future studies of disease risk, which provides an additional avenue for research into the health effects of carotenoids.