Exposure to food additive mixtures in 106,000 French adults from the NutriNet-Santé cohort

Food additives (e.g. artificial sweeteners, emulsifiers, dyes, etc.) are ingested by billions of individuals daily. Some concerning results, mainly derived from animal and/or cell-based experimental studies, have recently emerged suggesting potential detrimental effects of several widely consumed additives. Profiles of additive exposure as well as the potential long-term impact of multiple exposure on human health are poorly documented. This work aimed to estimate the usual intake of food additives among participants of the French NutriNet-Santé cohort and to identify and describe profiles of exposure (single substances and mixtures). Overall, 106,489 adults from the French NutriNet-Santé cohort study (2009-ongoing) were included. Consumption of 90 main food additives was evaluated using repeated 24 h dietary records including information on brands of commercial products. Qualitative information (as presence/absence) of each additive in food products was determined using 3 large-scale composition databases (OQALI, Open Food Facts, GNPD), accounting for the date of consumption of the product. Quantitative ingested doses were estimated using a combination of laboratory assays on food matrixes (n = 2677) and data from EFSA and JECFA. Exposure was estimated in mg per kg of body weight per day. Profiles of exposure to food additive mixtures were extracted using Non-negative Matrix Factorization (NMF) followed by k-means clustering as well as Graphical Lasso. Sociodemographic and dietary comparison of clusters of participants was performed by Chi-square tests or linear regressions. Data were weighted according to the national census. Forty-eight additives were consumed by more than 10% of the participants, with modified starches and citric acid consumed by more than 90%. The top 50 also included several food additives for which potential adverse health effects have been suggested by recent experimental studies: lecithins (86.6% consumers), mono- and diglycerides of fatty acids (78.1%), carrageenan (77.5%), sodium nitrite (73.9%), di-, tri- and polyphosphates (70.1%), potassium sorbate (65.8%), potassium metabisulphite (44.8%), acesulfame K (34.0%), cochineal (33.9%), potassium nitrate (31.6%), sulfite ammonia caramel (28.8%), bixin (19.5%), monosodium glutamate (15.1%) and sucralose (13.5%). We identified and described five clusters of participants more specifically exposed to five distinct additive mixtures and one additional cluster gathering participants with overall low additive exposure. Food additives, including several for which health concerns are currently debated, were widely consumed in this population-based study. Furthermore, main mixtures of additives were identified. Their health impact and potential cocktail effects should be explored in future epidemiological and experimental studies.

Statistical analyses. Descriptive analyses were weighted according to the 2016 French national census report data by using the CALMAR macro run by sex and based on categories for age, socio-professional status and housing area 68 . Intakes of each food additive were described in mg per day as well as in mg/kg bodyweight per day (% of consumers, mean, SD, median, 95 th percentile). Toxicity of each food additive is assessed by EFSA to determine its Acceptable Daily Intake (ADI), which is then used to set maximum authorized levels in the different food groups. However, additives without a specified ADI can be used quantum satis, i.e., with 'no limitations other than current good manufacturing practice' . Proportion of participants exceeding the ADI 69 (when available) were calculated for each food additive. To evaluate the variation that may have been caused by reformulations across a 10y period, the top 50 most consumed food additives was compared between 3 different periods in a sensitivity analysis (2009-2013/2013-2017/2017-2020).
Nonnegative Matrix Factorization (NMF) was used to determine food additive profiles of exposure. This size reduction technique is specifically adapted to sparse matrixes containing positive values 70 . It is described in detail in Appendix 2. Choice of algorithm was carried out according to the measure of residuals and of sparseness 71 , and the number of ranks r was determined according to the method proposed by Brunet et al 72 , using the smallest value of r for which this coefficient starts decreasing. The NMF was performed using the R package NMF 73 . Then, the scores arising from the components were scaled and introduced to a k-means clustering process, and the number of clusters of participants was determined using the elbow method, which examines the percentage of variance explained depending on the number of clusters.
Clusters of participants were described in terms of socio-demographic characteristics, food additive intake and food consumption. Regarding food and food additive consumption, means adjusted for energy intake and number of dietary records were used for description. Comparisons between clusters were performed by Chisquare tests or linear regression, as appropriate.
Partial correlation corresponds to the degree of association between two variables, controlling for other variables. To visualize the partial correlations between food additives, a network was generated using the glasso package 74 which computes a sparse gaussian graphical model with graphical lasso 75 . It can be interpreted as follows: when two food additives are connected by a blue line, it means that they are more consumed by the same participants, when they are connected by a red line, it means that they are rarely consumed by the same participants. Bolder is the line, higher is the correlation. The network was generated for the 90 analyzed food additives. R version 3.6.2 (R Foundation, Vienna, Austria) was used for the analyses. www.nature.com/scientificreports/
Food additive mixtures derived by NMF and clusters of participants according to additive intake. The NMF procedure resulted in 5 components that discriminated food additive exposure profiles (Appendix 6a). Figure 2 displays the network of partial correlation of food additives generated with graphical lasso. This method, complementary to NMF, yielded overall consistent results in terms of mixtures of additives identified.
After scaling NMF components of food additive mixtures, k-means clustering was performed and 6 clusters of participants were extracted using the elbow method. Appendix 6b shows the mean of each scaled NMF component, by cluster of participants. Each of the first 5 clusters mostly corresponded to one of the 5 specific food additive mixture NMF component, while cluster 6 corresponded to the participants with an overall low additive exposure. Table 2 displays sociodemographic and lifestyle characteristics according to clusters of participants. Table 3 and 4 display the mean (SD) consumption of food additives and food groups (respectively) by cluster, adjusted for energy intake and number of dietary records. Figure 3 displays a synthesis of cluster's consumptions of food additives and food groups. The clusters were described as follows: Cluster 1: Consumers of additives found in cookies and sweet cakes. This cluster constituted 9.8% of the study sample. Participants from this cluster presented notably the highest intakes of e322 lecithins, e471 mono-and diglycerides of fatty acids, e500 sodium carbonates, e450 diphosphates, e503 ammonium carbonates, e422 glycerol and e420 sorbitol. Participants from this cluster displayed the highest proportion of postgraduate and nonsmoker individuals and had the lowest mean BMI. They had the highest caloric and lipid intakes, but the lowest protein, alcohol and sodium intakes. They were the highest consumers of fatty and sweet cakes and cookies (consistent with their higher intakes of lecithins, mono-and diglycerides of fatty acids, sodium carbonates and glycerol) and fatty and salty products.
Cluster 2: Consumers of additives found in broths, meal substitutes, butter and bread. This cluster constituted 14.7% of the sample. They were the highest consumers of e14xx modified starches, e621 monosodium glutamate, e304 fatty acid esters of ascorbic acid and e320 butylated hydroxyanisole (BHA). Participants from this cluster were notably the oldest and the most physically active, with the lowest proportion of current smokers. They had the highest intake of sodium, the lowest lipid intake and were among the highest consumers of butter and margarines, meal substitutes and broth (consistent with the high intakes of monosodium glutamate and BHA).
Cluster 3: Consumers of additives found in dairy desserts, breakfast cereals and pastries. This cluster constituted 8.4% of the sample. They were notably the highest consumers of e407 carrageenan, e270 lactic acid, e282 calcium propionate, e452 polyphosphates, e160b annatto, e1442 hydroxy propyl distarch phosphate. These participants had relatively high carbohydrate intakes. They were the highest consumers of dairy desserts, which could explain the high intakes of carrageenans, lactic acid and hydroxy propyl distarch phosphate. They were also high consumers of pastries (consistent with higher intakes of calcium propionate), sweetened breakfast cereals and cereal bars.
Cluster 4: Consumers of additives found in industrial sauces and processed meat. This cluster constituted 12.3% of the sample. They notably had the highest intakes of e250 sodium nitrite, e316 sodium erythorbate, e451 triphosphates and e120 cochineal which are particularly used in processed meat. They also had the highest intakes of e330 citric acid, e415 xanthan gum, e202 potassium sorbate, e412 guar gum, e224 potassium metabisulphite and e150a plain caramel. This cluster included the highest proportion of men and had an overall lower level of education. They had the lowest carbohydrate intake. They were the highest consumers of bread, fish, rice, semolina, dressings and sauces (the latest being consistent with higher intakes of e415 xanthan gum, e202 www.nature.com/scientificreports/ potassium sorbate, e412 guar gum, e224 potassium metabisulphite). They were also high consumers of processed meat and pork and poultry hams (consistent with higher intakes of sodium nitrite, sodium erythorbate, triphosphates and cochineal).
Cluster 5: Consumers of additives found in sugary and artificially sweetened beverages. This cluster constituted 2.6% of the sample. They were notably the highest consumers of the 4 main sweeteners (e950 acesulfame K, e951 aspartame, e955 sucralose, e960 steviol glycosides), and of e440 pectins, e160a carotenes, e331 sodium citrates, e301 sodium ascorbate, e160c paprika extract, e150d sulphite ammonia caramel, e100 curcumin, e252 potassium nitrate, e338 phosphoric acid, e161b lutein, e211 sodium benzoate, e472 esters of mono-and diglycerides and e212 potassium benzoate. These participants had higher BMI, were the youngest, had the lowest physical activity and were more likely to be smokers. They had an intermediate caloric intake, the highest protein and UPF intakes, and the lowest proportion of organic food in their diet. They were notably the highest consumers of non-alcoholic sweetened and unsweetened drinks (including sugary and artificially sweetened sodas, in line with higher intakes of sweeteners, sodium and potassium benzoates, sodium citrates, phosphoric acid and sulphite ammonia caramel), processed meat, pork and poultry hams (consistent with higher intakes of potassium nitrate), and table-top sweeteners in powder.
Cluster 6: Consumers of various staple foods with low additive content. This cluster constituted 52.1% of the study sample, with the highest proportion of women (74.3%). It presented the lowest mean intakes for all food additives. It was characterized by its lower caloric intake, higher proportion of organic food and lower proportion of UPF in the diet, and higher alcohol intake. Participants of this cluster were high consumers of "staple foods": whole-grain products, pulses, breakfast cereals with little or no added sugar, vegetable juice, oleaginous fruits, vegetable oils, and cheese.

Discussion
To our knowledge, this large population-based study was the first to estimate chronic exposure to food additive mixtures based on detailed consumption and composition data for a wide range of substances. Forty-eight additives were consumed by more than 10% of the participants, with modified starches and citric acid consumed by more than 90%. The top 50 also included several food additives for which potential adverse health effects have been suggested by recent experimental studies. We identified and described five clusters of participants more www.nature.com/scientificreports/ specifically exposed to five additive mixtures and one additional cluster gathering participants with overall low additive exposure. Since 2012, EFSA has started the re-evaluation of all food additives authorized before January 2009. The agency's opinions on an additive are subject to change as evidenced by the update on TiO 2 , which is no longer considered as safe. EFSA has carried out simulations of exposure, combining average food consumption data from European member states with doses of additives reported by manufacturers, for additives and countries for which such data were available. Overall, when comparing exposure estimates with EFSA's, intakes of the NutriNet-Santé population were relatively lower. For instance, for modified starches, we estimated a mean intake of 24.33 mg/kg bodyweight per day (95th percentile: 66.4 mg/kg), versus 112.0 mg/kg bodyweight/day (95th percentile: 235.5 mg/kg) in EFSA's non-brand-loyal scenario (French population group) 76 . Similarly for lecithins: 0.83 mg/kg bodyweight per day (95th percentile: 2.6 mg/kg), versus 6.0 mg/kg bodyweight/day (95th percentile: 13.0 mg/kg) 77 . This overall lower exposure may in part be due to the more health-conscious profiles of NutriNet-Santé participants. However, this may also be due to methodological differences: in the present study, presence/absence of food additives was precisely determined based on the commercial brand and the precise list of ingredients, whereas EFSA stimulations use an average information by product category. Some other studies performed intake estimations for several specific food additives, in particular nitrites/nitrates, colors, monosodium glutamate and sulfites 33,[78][79][80][81][82][83][84][85][86][87][88][89][90][91][92] . Although comparisons of different populations are not straightforward, some similarities in the exposure estimates were observed. For instance, in China, similar mean intakes of monosodium glutamate were found: mean (SD): 2.2 (1.6) g/d 33 versus 2.4 (20.2) in the present study. However, as for EFSA simulations and except in rare cases, these studies were based on generic food data (not accounting for the specific brand consumed and thus the precise ingredient list). Besides, these studies focused on one specific additive or a very limited number of additives, which did not permit the investigation of mixtures. A recent study by the French food observatory "Observatoire de l'Alimentation" (Oqali) evaluated the occurrence of certain food additives in a selection of food products of the French market 93 . The additives most frequently found where consistent with the most consumed in our study (e.g. citric acid, modified starches and lecithins in the top 3; acesulfame K as the most used/consumed sweetener).
For several food additives widely consumed in this study, potential adverse health effects have been suggested by recent in-vivo/in-vitro, and-rarely-epidemiological studies. For instance, an experiment in humans demonstrated that phosphatidylcholine found in lecithin is converted by bacteria in the gut into trimethylamine-N-oxide, which may potentially contribute to hardening of the arteries or atherosclerosis and heart attack 94 . A www.nature.com/scientificreports/ potential role in the development of Chron's disease has also been suggested for lecithins [95][96][97] , and an experimental study among humans suggests a link between lecithins and coronary artery disease through the production of a proatherosclerotic metabolite, trimethylamine-N-oxide (TMAO) 94 . In a study on ex-vivo models of human microbiota, 20 emulsifiers were tested and a large majority (including carboxymethylcellulose, polysorbate 80, carrageenans, guar/xanthan gums, lecithins), were able to directly modify the gut microbiota in a way that could promote gut inflammation 23 . Carrageenans have been linked to fasting hyperglycemia and exacerbated glucose intolerance and hyperlipidemia without effect on weight in mice 31 . Also, sodium nitrite and potassium nitrate intakes have been associated in prospective cohorts with all-cause mortality (nitrates/nitrites from preserved/processed meat) 24 , and colorectal, gastric and pancreatic cancers [25][26][27]98 , although their impact remains debated. Phosphates have been associated with vascular effects (e.g. endothelial dysfunction and vascular calcification) in experimental studies among humans 41,42 . Sulfites have been associated with alteration of the gut and mouth microbiome in-vitro at concentrations close to those found in foods 99 . The effects of non-nutritive sweeteners such as acesulfame K, sucralose and aspartame on human cardiometabolic health and adiposity are controversial 37 , and these additives have been linked with hematopoietic neoplasia and gut microbiota alteration in experimental studies on rodents 21,[38][39][40] . Sulfite ammonia caramel, present in almost every cola sodas, might carry 4-methylimidazole (4-MEI) defined as possibly carcinogenic to humans by the International Agency for Research on Cancer (IARC). Monosodium glutamate might have patho-physiological and toxicological effects on human health 32,34 and was associated with overweight in a prospective cohort 33 . Carboxymethylcellulose has been associated with changes in microbiota composition, intestinal inflammation and metabolic syndrome (in-vivo) 43,100-102 , pro-inflammation (in-vivo, ex-vivo) 46,103-106 and promotion of tumor development (in-vivo) 45 .
In mouse models, it was recently shown that in the presence of intestinal inflammation, the food additive www.nature.com/scientificreports/ ethylenediaminetetraacetate (EDTA) was capable of exacerbating inflammation and inducing colorectal carcinogenesis at doses presumed to be safe 107 . The NMF procedure followed by k-means clustering allowed us to describe profiles of exposure to mixtures of food additives, which corresponded to specific socio-demographic profiles and dietary behaviors. Although about half of the population study pertained to cluster 6 and tended to have a relatively limited exposure to food additives overall. The other half of the study population was exposed to different additive mixtures (5 Table 2. Sociodemographic and lifestyle characteristics of clusters of food additive consumers, NutriNet-Santé cohort, France, 2009-2020 (N = 106,489). Weighted according to the French national census report data by using the CALMAR macro run by sex and based on age, socio-professional category and housing area 69 . Values are percentages unless stated otherwise. IPAQ was available for 91,675 participants, education for 99,725, smoking status for 106,242 and proportion of organic food for 28,075 participants. *Comparisons between clusters using Chi-square tests or linear regressions, as appropriate. **Adjusted for energy intake and number of 24 h dietary records.  www.nature.com/scientificreports/ main mixtures identified). In a previous work consisting in the exploration of the Open Food Facts database, we identified clusters of additives found in food products of the French market 30 . The mixtures of food additives identified in the present work resulted 1) from the co-occurrence of several additives in a same industrial product (as shown previously 30 ) and 2) from the co-consumption of various food products within usual dietary patterns. For instance, participants of cluster 1 were notably the highest consumers of sweet cakes and cookies, thus, they were particularly exposed to food additives of a specific cluster in our previous work ("stabilizers and emulsifiers mostly used in biscuits and cakes"). So far, detailed information on potential cocktail effects of food additives is lacking. However, several studies started to suggest potential interactions and synergies. For instance, mixture of colorings with sodium benzoate were associated with increased hyperactivity in children 108 . Neurotoxic effects were also observed between combinations of brilliant blue with L-glutamic acid and quinoline yellow with aspartame in-vitro 109 and a mixture of food coloring additives increased oxidative stress in rats 110 . Future prospective studies and experimental research should investigate the health effects of chronic exposure to these mixtures of food additives, as they are consumed in real life.

Consumers of various staple foods with low additive content
Strengths of this study included the large sample size and the accuracy of dietary intake data used to estimate additive exposure at the individual level, which is necessary for future etiological studies (population-based   www.nature.com/scientificreports/  www.nature.com/scientificreports/ simulations are not appropriate for this purpose). Indeed, repeated 24 h records allowed us to collect detailed information on > 3500 generic foods/beverages, each declined in dozens of commercial brands, which is a strength compared to previous nutritional studies. Three complementary databases were used to determine qualitative additive composition and thousands of assays were performed and complemented by EFSA and GSFA data to retrieve information on quantitative doses. However, some limitations should be acknowledged. First, not all food additives could be covered due to a lack of quantitative data for some additives. However, the latter were not the most relevant in terms of potential public health impact since they were mostly consumed by less than 10% of the population. Second, as it is generally the case for cohorts with a primary etiological focus, the recruitment method was based on a voluntary participation and the study population was not intended to be representative of the French population. Thus, the individuals included in the cohort were more often women, with "healthier" behaviors, a higher socioeconomic status and a higher level of education than the general French population 111,112 . However, even if lowest socioeconomic statuses were under-represented, the cohort still included about 6% of unemployed citizens or state aid recipients, which is lower than the national ≈10%, but higher than in other health studies that are not Internet-based. Moreover, the geographical distribution of the cohort was close to that of metropolitan France 113 . Also, the proportion of energy intake brought by ultra-processed foods (i.e. the main sources of food additives) among the participants of the cohort was 30-35%, consistent with the 31% assessed in two French nationally representative surveys 114,115 . Besides, a potential selection bias has been minimized since all analyses were weighted according to the characteristics of the French population (INSEE 2016 census). Last, industrial products may be reformulated across time by choice of manufacturers or regulation requirements, thereby complicating exposure assessment. However, bias linked to this aspect was limited in the present study since 1) the composition and consumption data were matched taking into account the year (dynamic matching), accounting for different compositions for a same product/brand consumed several years apart; and 2) the top 50 of most consumed food additives computed for 3 different periods of time in the 2009-2020 time-frame marginally changed, which illustrates the relative stability of additive exposure. In future etiological studies, it will be possible to study food additive exposure as time-dependent variables. This large population-based study provided for the first time a comprehensive overview of intakes for a wide range of additives, highlighting a widespread consumption of food additives for which health concerns are currently debated, and identified mixtures of food additives that were associated to consumer and food consumption profiles. Their health impact and potential cocktail effects should be explored in future epidemiological and experimental studies. In the meantime, and following the precautionary principle, several public health authorities worldwide recently started to recommend limiting the consumption of ultra-processed foods and, in practice, choosing food products of better nutritional quality (according to the Nutri-Score 116 ) and without or with as few additives as possible 117,118 .

Data availability
Data described in the manuscript, code book, and analytic code will be made available upon request pending application and approval. Researchers from public institutions can submit a collaboration request including information on the institution and a brief description of the project to collaboration@etude-nutrinet-sante.fr. All requests will be reviewed by the steering committee of the NutriNet-Santé study. A financial contribution may be requested. If the collaboration is accepted, a data access agreement will be necessary and appropriate authorizations from the competent administrative authorities may be needed. In accordance with existing regulations, no personal data will be accessible.