Genetic predisposition to macronutrient preference and workplace food choices

Abstract

Prior research identified genetic variants influencing macronutrient preference, but whether genetic differences underlying nutrient preference affect long-term food choices is unknown. Here we examined the associations of polygenic scores for carbohydrate, fat, and protein preference with 12 months’ workplace food purchases among 397 hospital employees from the ChooseWell 365 study. Food purchases were obtained retrospectively from the hospital’s cafeteria sales data for the 12 months before participants were enrolled in the ChooseWell 365 study. Traffic light labels, visible to employees when making purchases, measured the quality of workplace purchases. During the 12-month study period, there were 215,692 cafeteria purchases. Each SD increase in the polygenic score for carbohydrate preference was associated with 2.3 additional purchases/month (95%CI, 0.2 to 4.3; p=0.03) and a higher number of green-labeled purchases (β=1.9, 95%CI, 0.5 to 3.3; p=0.01). These associations were consistent in subgroup and sensitivity analyses accounting for additional sources of bias. There was no evidence of associations between fat and protein polygenic scores and cafeteria purchases. Findings from this study suggest that genetic differences in carbohydrate preference could influence long-term workplace food purchases and may inform follow-up experiments to enhance our understanding of the molecular mechanisms underlying food choice behavior.

Experimental studies suggest that most species have evolved dedicated circuits and pathways to detect and motivate sugar and fat intake [1–5]. These studies have identified gut-brain axis circuits involving the hypothalamus, brainstem, and mesolimbic system that play a pivotal role in regulating eating behavior. In humans, activating these selective circuits is considered an important contributor to the overconsumption of energy-dense and highly palatable foods and the concomitant increase in obesity rates [6–8], but there has been limited human research to identify molecular profiles underlying nutrient preference and food choice behavior.

A clinical study including 14 adults with heterozygous loss of function MC4R variants and 20 lean and weight-matched controls showed that MC4R deficient individuals exhibited a markedly increased preference for high-fat foods [9]. However, loss of function variation at the MC4R gene is rare in the population. In separate pediatric laboratory meal studies, children harboring genetic mutations in FTO had an increased preference for calorie-dense and high-fat foods [10–12], suggesting that the mechanisms by which FTO is associated with adiposity are in part mediated by alterations in nutrient-sensing circuits and pathways. While these findings are important as they provide early evidence of genetic profiles and molecular mechanisms underlying human food choice behavior, they rely on a relatively short dietary exposure, sometimes a single meal, and partial characterization of genetic susceptibility to macronutrient preference.

Genome-wide association studies (GWAS) have successfully identified genomic regions associated with variation in macronutrient preference [13–17]. The most recent GWAS, which identified 26 distinct genome-wide significant loci associated with macronutrient preference, showed that genetic variants for macronutrient preference map exclusively to specific brain regions [14]. Integration with brain single-cell RNA sequencing data identified neuronal subpopulations differently enriched for genes associated with carbohydrate, fat, or protein preference, including, for example, a subtype of GABAergic somatostatin expressing neuron in the lateral hypothalamic area that was enriched for genes associated with increased carbohydrate preference. Because distinct pathways sense specific nutrients [4], genetic susceptibility underlying nutrient preference could be used to insight into molecular mechanisms driving human food intake and behavior. In this study, we obtained cafeteria sales data from 12 months for 397 Massachusetts General Hospital employees to test the hypothesis that genetic susceptibility underlying macronutrient preference is associated with long-term workplace food choices.

Methods

Study design and participants

This study is a secondary analysis of participants in the “ChooseWell 365” study [18]. In brief, ChooseWell 365 was a randomized controlled trial that enrolled 602 Massachusetts General Hospital employees between September 2016 and February 2018 to investigate if providing employees with objective feedback and personalized nudges (health/lifestyle tips, social norms, incentives) in the context of an environmental strategy (traffic-light labeling) would increase healthy food choices and prevent weight gain. Employees were eligible for the trial if they were between 20 and 75 years of age and used their employee badges to purchase cafeteria items at least four times per week for at least six weeks during 12 weeks before recruitment. The analysis reported here included 397 European-ancestry participants who provided written informed consent for genetic studies. The study protocol was approved by the Mass General Brigham Institutional Review Board (protocol #2015P000135).

Cafeteria purchases collection procedures

Participant’s purchases were obtained retrospectively from the hospital’s cafeteria sales data for the 12 months before their enrollment in the ChooseWell 365 study. More than 1,200 food items were available, including meals and entrees (e.g., hot prepared meals, prepared sandwiches and salads, and pizza), a large salad bar, snacks, and desserts as hot and cold beverages. All food and drinks were labeled with traffic-light labels, as described elsewhere [18, 19]. The traffic-light labeling system, implemented hospital-wide in 2015, was based on the USDA Dietary Guidelines. A green rating connoted the highest food quality level, and a red rating indicated the lowest level. Overall, food and beverage purchases were evenly distributed between color labels: 34% were green-labeled, 37% were yellow-labeled, and 29% were red-labeled. The average costs of red, yellow, and green items were comparable, and items across different prices were available in each color category [20]. The quality of workplace purchases was measured by the total number of green, yellow, and red items purchased each month.

Polygenic scores for macronutrient preference

Genotyping was conducted using the Infinium Global Screening Array-24 v2.0. Imputation was performed using the Michigan Imputation server with the Haplotype Reference Consortium (HRC, Version r1.1 2016) reference panel for imputation [21]. Haplotype phasing was performed using Eagle v2.3 [22]. Low-quality genetic markers in Hardy-Weinberg disequilibrium (P < 10⁻⁶), low minor allele frequency (<0.01), and low call rate (<98%) were excluded. Samples were tested for low-quality genetic samples with a low sample call rate (<95%), high heterozygosity rate, or relatedness.

Polygenic scores for macronutrient preference were generated for each participant based on previously identified genetic variants and effect sizes for carbohydrate, fat, and protein [14]. Details on the genetic variants and weights to generate each polygenic score are provided in Supplementary Table 1. Genome-wide single-nucleotide polymorphism (SNP)-based heritability was estimated at 4.1% (SE = 0.01), 3.5% (SE = 0.01), and 4.4% (SE = 0.02) for carbohydrate, fat, and protein, respectively. To calculate individual scores in our study population, we used directly genotyped or imputed variants with an imputation quality (minimal rsq) ≥0.50. Each variant was coded with the expected number of associated alleles and weighted by its relative effect size on each phenotype. Each polygenic score was then standardized to a mean of 0 and a standard deviation of 1.

Assessment of covariates

Clinical research nursing staff measured participants’ weight and height at the baseline visit, and body mass index (BMI) was calculated (kg/m2). Participants also completed an online survey that provided self-reported age, sex, education, and health behaviors, including physical activity, diet, and smoking. The physical activity level at baseline was measured with the International Physical Activity Questionnaire[23] and categorized into low, intermediate, and high physical activity levels. Education was ascertained as less than high school, high school, some college, college, or graduate school. Job type was extracted from human resources data and categorized as administrative/service, craft/technicians, management/professionals, and MDs/PhDs. We computed principal components of ancestry based on genetic data using TRACE [24].

Statistical analysis

We summarized continuous measurements using medians and interquartile range and presented categorical observations using frequencies and percentages. Of the 602 participants included in the ChooseWell 365 study, 499 consented to have genotyping, and 397 were of European ancestry and included in the current analysis. The rationale for restricting the study to European ancestry participants was that the discovery of genetic variants for macronutrient preference was conducted in this population. There were no missing values for the main covariates and outcome data. We described associations of demographic and clinical characteristics with monthly food purchases using linear mixed-effects models to account for repeated and clustered observations [25].

Linear mixed-effects models with a random intercept and random slope were used to test the association of polygenic scores with the total number of cafeteria purchases. We modeled polygenic scores as a continuous variable and provided adjusted estimates per 1SD increase in each polygenic score. Linear mixed-effects models were adjusted for age (continuous, years), sex (categorical, male/female), and the first three principal components of ancestry (model 1). We further adjusted for education level (high school/some college, college degree, graduate degree) and job type (administrative/service, craft/technicians, management/professionals, MDs/PhDs) (model 2). Multivariable model 3 was further adjusted for BMI (18.5 to 24.9 kg/m2, 25.0 to 29.9 kg/m2, and ≥30 kg/m2), current smoking status (yes, no), and physical activity (categorized into low, intermediate, and high physical activity level).

In subsequent analyses, we used linear mixed-effects models with a random intercept and random slope to test the association of polygenic scores with the quality of cafeteria purchases measured by the total number of green, yellow, and red items. We adjusted for the same covariates used in previous models. For color-labeled items with evidence of genetic associations, we further assessed the association of polygenic scores with specific types of purchases (i.e., beverages, sides, entrees, and condiments) after adjusting for the same confounders. To investigate whether the associations between polygenic scores and cafeteria purchases were influenced by reverse causation and/or pleiotropy, we conducted a secondary analysis excluding genetic variants in or near genes previously associated with food intake and other relevant phenotypes, including APOE and ADH1B [26].

In subgroup analyses, we assessed the association between polygenic scores and cafeteria purchases according to age, sex, education, job type, and BMI categories. We conducted sensitivity analyses to account for other behaviors and confounders associated with food choice behavior, including the frequency of consumption of meals outside of work and seasonality. For analyses of consumption of meals outside of work, we used baseline survey data about participants’ self-reported frequency of breakfast, lunch, and dinner prepared at home. Breakfast, lunch, and dinner prepared at home variables were categorized into “Never or less than once per week,” “sometimes,” and “every day or almost every day.” We included the same covariates for analyses of meal consumption outside the home as model 3. For seasonality analyses, sine and cosine functions of the enrollment date were used to adjust for seasonality in the participants’ 12-month purchasing periods [27].

Because residual confounding due to obesity might persist even though our analyses accounted for BMI [28], we conducted additional analyses to investigate the effect of BMI on genetic predisposition to carbohydrate preference. We examined whether the polygenic score for carbohydrate preference varied across BMI categories using the Kruskal-Wallis test. Further, we used generalized linear models to investigate whether BMI was an independent predictor of the polygenic score for carbohydrate preference. Finally, structural equation models were implemented to conduct a mediation analysis of BMI. For the mediation analysis, BMI and the polygenic score for carbohydrate preference were modeled as continuous variables. We estimated the relative contribution of BMI to the association between the polygenic score for carbohydrate and cafeteria purchases. We computed the proportion of the total effect explained by the indirect effects of BMI. Indirect effects were estimated by taking the product of the effect of the exposure on the mediator and the effect of the mediator on the outcome [29]. To calculate the proportion of the mediated effect, we divided the indirect effect by the total effect. The direct effect, which is the association of polygenic score for carbohydrates on cafeteria purchases through mechanisms independent of mediation, was estimated from regressing cafeteria purchases on the polygenic score for carbohydrate preference.

All statistical analyses were performed using R software, version 4.0.3 (R Foundation). Plots were created using the packages ggplot2 and sjPlot.

Results

Characteristics of the 397 European-descent participants included in this study are shown in Table 1. Participants were predominantly female (n=321, 80.9%), with a median age of 45 (P_25–75, 33 to 56) and a median BMI of 26.5 (P_25–75, 23.5 to 30.8 kg/m²). The study sample was representative of the overall ChooseWell 365 study population with no major differences in clinical, health behavior, and demographic characteristics other than race/ethnicity (Supplementary Table 2).

Table 1:

Characteristics of the study participants.

	Study participants (n=397)
Demographic characteristics
Age, years	45 (33–56)
≥20–39	161 (40.5)
40–59	167 (42.1)
≥60	69 (17.4)
Sex, No. (%)
Male	76 (19.1)
Female	321 (80.9)
Education, No. (%),
High school/some college	32 (8.1)
College degree	172 (43.1)
Graduate degree	193 (48.4)
Job type, No. (%)
Administrative & Service	36 (9.1)
Craft & Technicians	37 (9.3)
Management & Professionals	283 (71.3)
MD/Phd	41 (10.3)
Lifestyle characteristics
Smoking status, No (%)
Never	279 (70.3)
Former	108 (27.2)
Current	10 (2.5)
Physical activity, No (%)
Low activity	9 (2.3)
Moderate activity	111 (28.0)
High activity	277 (69.8)
Body mass index, Kg/m2	26.5 (23.5–30.8)
18.0–24.9	154 (38.8)
25–29.9	134 (33.8)
≥30	109 (27.5)
Foods and beverages purchases
Total Purchases over 12 months,	215,692
Green Items, No (%)	116,528 (54)
Yellow Items, No (%)	68,370 (31.7)
Red items, No (%)	30,794 (14.3)

Values are median (P₂₅-P₇₅) for continuous variables; numbers and (percentages) for categorical variables. Physical activity was measured with the International Physical Activity Questionnaire [23] and categorized into low, intermediate, and high physical activity levels.

During the 12-month study period, there were 215,692 cafeteria purchases. Among these purchases, 54% (n=116,528), 31.7% (n=68,370), and 14.3% (n=30,794) corresponded to green, yellow, and red-labeled items, respectively (Table 1). Older participants, female participants, participants with obesity, or those with lower education were more likely to buy more food and beverage items than younger participants, male participants, participants without obesity, or those with higher education (P<0.05 for all, Supplementary Figure 1).

The polygenic scores for macronutrient preference were normally distributed (Supplementary Figure 2). The polygenic score for carbohydrate preference was associated with a higher number of cafeteria purchases, with an estimated 2.4 additional purchases/month per SD increase in the carbohydrate polygenic score (95% CI, 0.4 to 4.5; Table 2) after adjusting for age, sex, and PCAs. This association persisted after adjusting for education, job type, BMI, smoking status, and physical activity levels (Table 2). There was no evidence of significant associations between the polygenic scores for fat or protein preference and total cafeteria purchases (Table 2).

Table 2:

Association between polygenic scores for macronutrient preference and workplace purchases.

	Total purchases/month
	β (95% CI)	P
Polygenic score for carbohydrate preference
Age and sex-adjusted model	2.4 (0.4 to 4.5)	0.023
Multivariable model 2	2.7 (0.6 to 4.7)	0.011
Multivariable model 3	2.3 (0.2 to 4.3)	0.029
Polygenic score for fat preference
Age and sex-adjusted model	−1.0 (−3.1 to 1.1)	0.37
Multivariable model 2	−0.7 (−2.8 to 1.4)	0.50
Multivariable model 3	−0.1 (−2.0 to 2.1)	0.94
Polygenic score for protein preference
Age and sex-adjusted model	−1.4 (−3.5 to 0.6)	0.18
Multivariable model 2	−2.0 (−4.0 to 0.1)	0.07
Multivariable model 3	−1.9 (−3.9 to 0.3)	0.09

Linear mixed-effects models with a random intercept and random slope were used to test the association of polygenic scores with the total number of cafeteria purchases. Data represent beta coefficients and 95% CI for cafeteria purchases per 1SD increase in polygenic scores.

Mixed effects models were adjusted for age, sex, and PCAs (Age and sex-adjusted model).

Multivariable model 2 was further adjusted for education level (high school/some college, college degree, graduate degree) and job type (administrative/service, craft/technicians, management/professionals, MDs/PhDs).

Model 3 was further adjusted for body mass index (18.5 to 24.9 kg/m2, 25.0 to 29.9 kg/m2, and ≥30 kg/m2), current smoking status (yes, no), and physical activity (measured with the International Physical Activity Questionnaire at the baseline visit and categorized into low, intermediate, and high physical activity level)

Next, we examined the association of polygenic scores for macronutrient preference with the quality of cafeteria purchases. The polygenic score for carbohydrate preference was positively associated with the number of green-labeled items purchased (1.9 purchases/month per SD increase in the carbohydrate polygenic score (95% CI, 0.5 to 3.3; Figure 1). This association was mainly driven by a higher number of green-labeled beverages, with an adjusted estimate of 1.1 purchases/month (95% CI, 0.3 to 1.9; Supplementary Table 3). The polygenic scores for fat and protein preference were not associated with the quality of cafeteria purchases. To enhance the interpretation of our findings, we investigated the nutritional composition of the most frequently purchased color-coded foods. This analysis showed that, compared to yellow or red items, green items were more likely to be higher in carbohydrates and lower in fat (Supplementary Table 4). Further, we conducted a secondary analysis to investigate the influence of highly pleiotropic genetic variants included in the polygenic scores. This analysis showed that the associations between polygenic scores and cafeteria purchases were consistent after excluding genetic variants in or near APOE and ADH1B (Supplementary Tables 5–7).

Figure 1. Association between polygenic scores for macronutrient preference and types of cafeteria purchases.

Beta coefficients and 95% CI for types of cafeteria purchases per 1SD increase in polygenic scores. Linear mixed-effects models with a random intercept and random slope were adjusted for the same covariates as the previous model 3.

In stratified analyses, there was no evidence that the association between the polygenic score for carbohydrate preference and the number of cafeteria purchases differed by age, sex, education level, job type, or BMI categories (p>0.05 for all; Table 3). Further, we conducted a sensitivity analysis to account for meals prepared at home, as consuming these meals could bias the association between polygenic scores and cafeteria purchases. The mean difference in the polygenic score for carbohydrate preference between participants who consumed meals prepared at home almost every day (n=33) or less than once a week/never (n=41) was 0.37 SD units (95% CI, −0.23 to 0.98; p=0.32). In a multivariable model accounting for the frequency of consumption of meals prepared at home, we showed that the polygenic score for carbohydrate preference was still associated with a higher number of cafeteria purchases, with an adjusted estimate of 2.0 purchases/month per SD increase in the carbohydrate polygenic score (95% CI, 0.1 to 4.1; Supplementary Table 8). In a separate sensitivity analysis, we further adjusted our models for seasonality. This analysis showed that the carbohydrate preference polygenic score was associated with a higher number of cafeteria purchases regardless of the recruitment month (Supplementary Table 8).

Table 3:

Adjusted estimates of workplace purchases according to polygenic scores stratified by sociodemographic and clinical characteristics.

Factor	No. of participants/ Total purchases	β (95% CI) per SD increase	P
Age,
≥20–39	161 / 79,137	1.4 (−1.4 to 4.3)
40–59	167 / 95,098	2.9 (−0.3 to 6.1)
≥60	69 / 41,421	−0.7 (6.9 to 5.5)	0.49
Sex,
Male	76 / 47,121	4.6 (−1.4 to 10.7)
Female	321 / 168,571	1.3 (−0.9 to 3.4)	0.05
Education,
High school/some college	32 / 21,296	1.8 (−10.5 to 14.1)
College degree	171 / 94,818	2.8 (−0.2 to 5.8)
Graduate degree	192 / 98,418	2.1 (−0.7 to 4.8)	0.97
Job type,
Administrative & Service	36 / 22,233	−6.9 (−15.8 to 2.1)
Craft & Technicians	37 / 25,759	11.9 (1.9 to 22.1)
Management & Professionals	283 / 146,306	3.0 (0.7 to 5.2)
MD/Phd	41 / 21,394	−3.6 (−9.8 to 2.6)	0.46
Body mass index, Kg/m2
<25	154 / 69,784	1.2 (−1.8 to 4.2)
25–30	134 / 77,646	1.8 (−2.0 to 5.5)
≥30	109 / 68,262	3.8 (−0.5 to 8.1)	0.71

Association between the carbohydrate preference polygenic score and total cafeteria purchases according to sociodemographic and clinical characteristics.

Linear mixed-effects models were adjusted for the same covariates as the previous model 3.

P-values obtained from linear mixed-effects models include an interaction term between the polygenic score and each factor.

Further, we examined the role of obesity on genetic predisposition to carbohydrate preference. We showed that the polygenic score for carbohydrate preference was similar across BMI categories (p=0.79; Supplementary Figure 3). In addition, we found no evidence that BMI was associated with a higher genetic preference for carbohydrate intake (β = 0.01, 95% CI, −0.04 to 0.04; per 1 kg/m² increase; Supplementary Figure 4). Finally, we conducted a mediation analysis by BMI. We showed evidence of a direct effect of the polygenic score on total purchases and that BMI mediated 8.3% (95% CI, 2.1 to 14.9) of this effect (Supplementary Table 9).

Discussion

Results from this study investigating the association of genetic predisposition to macronutrient preference with worksite cafeteria purchases suggest that a higher genetic predisposition to carbohydrate preference was associated with a higher number of total and more healthy purchases. These associations were consistent in subgroup and sensitivity analyses accounting for additional sources of bias, including obesity, potential pleiotropic effects of included genetic variants, and the frequency of meals consumed outside work. Although preliminary, these findings illustrate how distinct genetic susceptibility underlying nutrient preference influences long-term food choices.

Our observation that genetic variation in macronutrient preference is associated with a higher number of total and healthy purchases expands previous studies showing the modest but significant contribution of genetic variation on food preference [7, 9, 10, 30]. For example, these studies have documented that individuals harboring loss-of-function mutations in MC4R have a greater preference for high-fat foods but a lower preference for high-sucrose [9]. One potential mechanism explaining these associations is due to a preserved pattern of activation of the dopaminergic reward system among carriers of the MC4R variants [29]. In contrast, genes that reside in genomic regions associated with carbohydrate preference are predominantly expressed in specific subtypes of GABAergic neurons, a broad spectrum of inhibitory neurons [14, 30, 31]. While our study was not specifically designed to elucidate molecular mechanisms, the evidence that carbohydrate preference is related to inhibitory control and healthy food choices could support functional follow-up experiments to characterize these initial observations.

Our study suggests that the association between genetic susceptibility to carbohydrate preference and food purchases is unlikely to be confounded by obesity. In a previous study, genetic predisposition to increased BMI was associated with larger and more unhealthy food purchases [33]. Because individuals with increased genetic susceptibility to BMI are more likely to be obese, this study could not discern whether genetic risk for obesity induces unhealthy food choices or whether unhealthy food choices are a consequence of obesity. We showed no difference in the carbohydrate preference polygenic score distribution according to BMI categories and that BMI was not a significant predictor of carbohydrate preference. While we cannot completely exclude potential pleiotropic effects of genetic variants for carbohydrate intake on other phenotypes [34, 35], the corroboration that our results were consistency after excluding highly pleiotropic genetic variants from our polygenic scores provides evidence of a molecular profile associated with a phenotype crucial for the development of obesity and related metabolic complications. However, our results regarding estimated effect sizes are modest and need to be interpreted in the context of relatively healthy employees from a large urban hospital. Further research in independent samples is required to validate our initial findings.

In the U.S., poor dietary habits are estimated to account for >650,000 deaths per year, and 14% of all disability-adjusted life-years lost [36]. Improving our modern food environment is essential to mitigate the deleterious consequences of unhealthy food choices [36, 37]. This is particularly relevant in the workplace as there is often limited access to healthier options. Several studies have shown that changing the workplace food environment, such as traffic-light labeling systems or offering more healthy foods in cafeterias, improves diet quality [20, 38, 39]. Our findings could aid in previous studies by illustrating how individuals with a different genetic predisposition to macronutrient preference benefit from workplace healthy eating efforts.

We acknowledge several limitations. First, as an observational study, we cannot completely rule out the potential for residual confounding as individuals who eat healthier diets tend to engage in other healthful behaviors, such as engaging in physical activities or being less likely to smoke. However, it is reassuring that our findings were consistent despite controlling for smoking and physical activity, two of the most relevant factors associated with healthy food purchases. Second, our study population comprised relatively healthy employees of European ancestry at a large urban hospital in the US. Thus, our findings may not be generalizable to other types of employees, non-employed people, or individuals of non-European ancestry. Third, the relatively small sample size, especially in the stratified analysis, could have limited our ability to find interactions between polygenic scores and clinical or demographic characteristics on food choice behavior. Four, while cafeteria purchasing data provided an opportunity to objectively assess employees’ cafeteria purchases, our results were limited to available cafeteria food and beverages. They may not have reflected actual food consumption that work shift schedules and cafeteria hours could influence. However, our sensitivity analyses, in which we adjusted for foods from outside the hospital, provided consistent findings.

In conclusion, our findings on employees of a large hospital provide evidence for the association of genetic predisposition to macronutrient preference and food choice behavior. The observed associations between genetic predisposition for carbohydrate preference and the higher number of total and healthier cafeteria purchases illustrate how molecular processes underlying nutrient preference influence worksite food choices. Our preliminary results could inform follow-up experiments to enhance our understanding of food choice behavior.

Data availability

Data described in the manuscript will be made available upon request, pending application and approval.

Code availability

Code to reproduce analyses for this manuscript will be made available upon publication on GitHub.