Participants and study design

The Tsuruoka Metabolome Cohort Study (TMCS) is a population-based, prospective cohort study conducted in Tsuruoka city, Yamagata Prefecture, Japan, and is designed particularly to discover metabolomics biomarkers related to environmental and genetic factors and those for common diseases and disorders. Detailed information on the cohort study methods has been published elsewhere [22–24]. Briefly, the participants of the TMCS were 11,002 residents or workers in Tsuruoka aged 34–74 years at the time of the baseline survey conducted from 2012–2015. All participants provided written informed consent for the study and its protocol was approved by the Medical Ethics Committee of the Keio University School of Medicine, Tokyo, Japan (approval no. 20110264). Firstly, 7,303 participants aged 40–69 years without a medical history of stroke, coronary heart disease, or cancer were chosen for this cross-sectional study. Among those, the following participants were excluded from this analysis: those who did not respond to the FFQ, those who had missing data regarding staple food frequency (n = 40), those who had missing data regarding drinking status (n = 10), those with an unassessed metabolome (n = 39), outliers of biochemical test values (n = 4), outliers of estimated food frequency (n = 26), those who were not fasting before blood sample collection (n = 143), and those with missing data regarding fasting status (n = 32). Finally, 7,012 participants were included in the analysis, comprising 3,198 males and 3,814 females. A flowchart of participant inclusion and exclusion in the analysis is shown in S1 Fig.

All data and blood samples were obtained at the time of the baseline survey. The participants responded to a self-reported questionnaire that included information on demographics, physical activity, alcohol consumption, smoking habits, personal medical history, and other lifestyle factors. Energy intake and daily food consumption were assessed based on a validated FFQ (see later for details). Omissions or inconsistencies in participants’ responses were addressed by the trained survey staff through face-to-face interviews. The medical history was evaluated based on both, the self-reported questionnaire and medical checkup results. Biochemical test results were obtained from medical checkup institutions with the consent of the patients, including the height and weight, to calculate the body mass index (BMI).

The fasting plasma samples were analyzed to obtain non-targeted metabolomics data, using capillary electrophoresis time-of-flight mass spectrometry (CE-TOF-MS), which predominantly measures charged low molecular compounds, such as amino acids and their analogs. A 16 ml blood sample was collected from each participant between 8:30 and 10:30 in the morning after 12 hours of fasting from the previous night to avoid short-term metabolic fluctuations. The sample was divided into 0.5–1 ml portions, then metabolites were extracted from plasma within 6 hours of collection to further minimize the effects of metabolic changes and stored frozen until used for analysis. Sample preparation methods and analysis protocols for CE-TOF-MS have been described in detail previously [22–24]. We quantified the absolute concentrations of 115 metabolites that were expected to be stably observed in most human plasma samples and were compatible for comparison with standard compounds. Raw data were analyzed with our proprietary software, MasterHands [25] (see the summary of instruments and analytical conditions in S2 Table).

Dietary assessment

The questionnaire on dietary habits administered as part of the cohort study was created based on the Semi-Quantitative Food Frequency Questionnaire (SQFFQ) developed by the Department of Health Promotion and Preventive Medicine, Graduate School of Medicine, Nagoya City University [26, 27]. The validity and reproductivity of the SQFFQ had been assessed for energy, selected macro and micronutrients, and food consumption [28, 29]. A total of 76 questions were asked concerning the frequency of intake of 47 food items by the self-administered reminder method to assess eating habits in the past year. Responses to the questions on food intake were categorized at eight levels (never or seldom, 1 to 3 times per month, 1 to 2 times per week, 3 to 4 times per week, 5 to 6 times per week, once per day, twice per day, and three times or more per day) [27]. For staple foods such as rice, bread, and noodles, we asked about the intake frequency at breakfast, lunch, and dinner, as well as the number of portions (cups/pieces) per serving. For alcohol, we inquired about different kinds of alcohol, the number of drinking days per month/week, and the number of drinks per occasion in a questionnaire on lifestyle (see more details of questionnaires in the S1 Table).

In the present study, the 47 food items and different kinds of alcohol assessed via questionnaires were classified into four main food categories consisting of 17 food groups [26]: energy-giving foods (rice, other carbohydrates, confectionary, and oily food), protein-rich foods (meat, fish/seafood, eggs, dairy products, and soy products), fruits and vegetables (carotenoid-rich vegetables, leafy/other vegetables, seaweed, seeds, and fruits), as well as beverages (green tea, coffee, and alcohol). The daily intake of each food group (g/d) was calculated by summing the intake of included food items. The intake of each food item was calculated by multiplying the food intake frequency (per day) by the standard portion size (in grams) set for the SQFFQ nutrition calculation. If the intake frequency was less than once per day, a conversion weight was assigned (never or seldom: 0.05, 1 to 3 times per month: 0.1, 1 to 2 times per week: 0.2, 3 to 4 times per week: 0.5, and 5 to 6 times per week: 0.8). Alcohol intake was calculated based on the reported frequency and quantity consumed per occasion. The total consumption of different kinds of alcohol was calculated according to the percentage of ethanol and shown in comparison to Japanese sake. We finally focused on the three categories (protein-rich foods, fruits and vegetables, and beverages) that were suitable for identifying food biomarkers using CE-MS. Seaweed and seeds, which had a very low intake among the target population, were excluded from the analysis.

Statistical analysis

First, we examined the characteristics of the study population by total and sex-specific data. Data with normal distributions are reported as means and standard deviations (SDs), while skewed data are reported as medians and interquartile ranges (IQRs). For the population intake status of the 17 food groups, we calculated means and 10^th-90^th percentile ranges for both total and sex-specific data.

For metabolome data, we excluded metabolites which had plasma concentrations below the assay limit of detection (LOD) in more than 60% of the entire study population, and 94 substances (54 anions and 40 cations) were assessed in the final analysis (the list of metabolites is shown in S3 Table). For samples with undetectable levels below the LOD, values were imputed using half of the LOD values.

Firstly, a principal component analysis was performed to detect outliers, and two samples were excluded from the analysis beforehand (see details of outlier detection in S2 Fig). Since plasma metabolite concentrations are multivariate data with relatively strong correlations between substances which might change simultaneously due to biochemical interactions in vivo, we used the PLS-R model to select metabolites that greatly contributed to responses to food intake. Then, PLS-R was performed using the Nonlinear Iterative Partial Least Squares algorithm [30]. In this procedure, the intake of each food item was treated as a continuous response variable X, and the 94 metabolites were dealt with as continuous predictor variables Y. All response and predictor variables were log-transformed and standardized. For each food group, observations without responses to food intake were treated as missing values. For alcohol, only data obtained from habitual male drinkers (n = 2,449) were used.

To determine the optimal number of factors required to avoid model over-fitting, leave-one-out cross-validation (LOO-CV) was performed. Using the Van der Voet test, the optimal factor number for each food group was provided with the critical value of p > 0.1 for Hotelling’s T² statistic. For cases in which the optimal factor was less than two, the factor number was set to two. The explained variation in the X matrix (R²X), the explained variation in the Y matrix (R²Y), the predicted variation in the Y matrix (Q²) and their cumulative values were calculated to confirm the goodness-of-fit of the models. In a PLS-R model, R²Y is the proportion of variance in the dependent factors that is predictable from the independent factors, while Q² is the R² when the model built on the training set is applied to the test set. Adding a factor always raises R²Y, whereas Q² does not raise in case of over-fitting. Therefore, the closer the cumulative Q² is to 1, the better the predictive performance of the model. The contribution of individual metabolites in the metabolic signature for each food group were evaluated from variable importance in projection (VIP) scores and positive PLS coefficients. To further complement this, the associations between food intake and plasma metabolite levels were assessed with partial rank-order Spearman correlation coefficients, controlling for sex, total physical activity levels, and smoking status as potential confounders. Statistical analyses were performed using SAS version 9.4 (SAS Institute Inc., Cary, NC, USA), and JMP version 15 (SAS Institute Inc., Cary, NC, USA) was used for outlier detection to visualize the results.

Participants’ characteristics

Table 1 shows the characteristics of the study population. The mean age was 57.8 ± 8.2 years, and the mean BMI was 23.3 ± 3.3 kg/m², which was within the Japanese standard range (18.5–25 kg/m²). The population showed general sex differences among Japanese people. That is, while males were more likely to have a BMI that fell within the overweight range, females were more likely to have a BMI that fell within the underweight range. Moreover, males were likely to have much higher current smoking and habitual drinking rates than females. Concerning nutrition status, males were more likely to have high energy intake and carbohydrate ratio and women were more likely to have a high lipid ratio.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Characteristics of the target population.

https://doi.org/10.1371/journal.pone.0246456.t001

Table 2 shows the distribution of food classification and the mean daily intake for each food group among the population. All grouped food items are common foodstuffs that are usually eaten in a typical Japanese diet. Overall, the population had a high rice intake as staple food compared with bread and noodles, and the participants obtained more protein from fish and soy products than from meat. There were some sex differences in food intake; while males were more likely to consume higher amounts of rice and alcohol, females were more likely to consume more fruits, vegetables, and dairy products. Detailed information is shown in the S6 Table.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Food classification and population intake status.

https://doi.org/10.1371/journal.pone.0246456.t002

Identification of food intake biomarkers

To avoid model over-fitting, we performed the LOO-CV and Van der Voet test, which was proposed as a statistical test with the T² statistic for comparing the predicted residual sum of squares from different models. The PLS-R analyses resulted in final models with a cumulative R²X range of 0.11–0.24, a cumulative R²Y range of 0.05–0.29, and a cumulative Q² range of 0.01–0.53. The food groups could be classified into three predictive performance levels: the lower level for eggs (Q²_cum 0.02 for the final model), green tea (Q²_cum 0.05), and meat (Q²_cum 0.07); the middle level for fish/seafood (Q²_cum 0.21), soy products (Q²_cum 0.23), carotenoid-rich vegetables (Q²_cum 0.28), other vegetables (Q²_cum 0.31), and dairy (Q²_cum 0.33); and the higher level for fruit (Q²_cum 0.47), alcohol (Q²_cum 0.53), and coffee (Q²_cum 0.55). Most validation sets were fitted reasonably for studying fasting concentrations by self-reported FFQs, which are commonly reported with lower validation sets than under conditions of rapid intake. Details of the CV analyses of the goodness-of-fit are shown in S3 Fig and S4 Table.

Compounds that contributed to each food intake with high VIP are shown in Table 3. While the VIP score is generally used for screening variables with PLS modeling, the score is a relative value and has a large variation due to the variable preprocessing method. Therefore, metabolites that are shown to have predominance by univariate and/or multivariate analyses are more likely to be reliable [31]. Important metabolites were selected by referring to VIP scores and PLS coefficients as well as supplementary Spearman’s correlation coefficients. Relationships among them for the characteristic food groups are illustrated in Fig 1. The correlation matrix diagram is shown in Fig 2.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Overview of food biomarker candidates assessed by PLS-R (n = 7,012).

Relationships between the VIP score and the PLS coefficient are described using Spearman’s correlation coefficients. The vertical axis corresponds to the VIP score, the horizontal axis corresponds to the PLS coefficient, and the area of each circle corresponds to the correlation coefficient. Notable metabolites with VIP scores ≥ 1.5 and PLS coefficients ≥ 0.03 are highlighted. (A) meat, (B) fish/seafood, (C) eggs, (D) dairy, (E) soy products, (F) carotenoid-rich vegetables, (G) fruits, (H) coffee, and (I) alcohol. PLS-R, partial least square regression; VIP, variable importance in projection; TMAO, trimethylamine-N-oxide.

https://doi.org/10.1371/journal.pone.0246456.g001

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Correlation matrix diagram of the relationship between food groups and metabolites (n = 7,012).

This heatmap was generated with partial rank-order Spearman correlation coefficients, controlling for sex, smoking, and total physical activity levels. (A) cation, (B) anion. ^a Data of male drinkers (n = 2,449) were used in the analysis for alcohol.

https://doi.org/10.1371/journal.pone.0246456.g002

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Promising food biomarker candidates (n = 7,012).

https://doi.org/10.1371/journal.pone.0246456.t003

Metabolites with specificities of a VIP score ≥ 1.5 and PLS coefficient ≥ 0.3 were more secure than the others as candidate compounds, considering the multiple evaluations (Fig 1). Thus, we identified a total of 21 metabolites with the criteria as candidate habitual food intake markers for nine food groups in three categories, including protein-rich food, fruit/vegetables, and beverages. Major possible markers for protein-rich food intake were hydroxyproline (VIP score 2.66) and 3-methylhistidine (3-MH; VIP score 2.11) for meat, trimethylamine-N-oxide (TMAO; VIP score 2.63) for fish/seafood, and choline (VIP score 2.88) for eggs. In common with these sources of animal protein, 2-aminobutyrate (2-AB) and creatine were substances that were related to changes in intake. Galactarate (VIP score 2.14), threonate (VIP score 1.97), and phenylalanine (VIP score 1.95) were marker candidates related to dairy intake. On the other hand, cystine (VIP score 1.73) and betaine (VIP score 1.53) were metabolites related to the intake of soybean and soy products. Notable metabolites common to the intake of carotenoid-rich vegetables and other vegetables were threonate (VIP scores 2.23 and 1.85, respectively) and galactarate (VIP scores 2.06 and 1.51, respectively). Also, proline betaine (VIP score 3.80) was a prominent candidate marker for the intake of fruits. For beverages, metabolites such as quinate (VIP score 4.59), trigonelline (VIP score 3.13), and hippurate (VIP score 1.88) showed a relation with coffee consumption. Pipecolate (VIP score 2.78), and 2-AB (VIP score 1.92) were closely related to alcohol metabolite concentrations (see more details of the results in the S5 Table).

Protein-rich foods

Fruits and vegetables

Metabolites that were common to fruits and vegetables were threonate and galactarate, consistent with the results of previous studies [14, 32, 34]. As mentioned above, threonate, a sugar acid of threose, is also disassembled from ascorbic acid. It is generally known that fruits and vegetables are rich in ascorbic acid [63]; thus, the component was likely to influence the changes in threonate concentrations. Besides, the plasma concentrations of galactarate (mucate), a sugar acid of galactose, also increased with fruit and vegetable consumption. The metabolite is found in many foods that contain mucins, such as vegetables, potatoes, and root vegetables, as well as ripe fruits that are high in pectin (such as pear, peach, and pomes). Pectin is a structural acidic heteropolysaccharide present in the primary cell walls of plants, in which mucate exists in the form of the polymer polygalacturonate [68].

A metabolite closely related to fruit intake was proline betaine, which is known to be one of the most secure food biomarkers in plasma and urine, particularly for the consumption of citrus fruits such as oranges and grapefruits [8, 35, 69]. The metabolite is a rich component of citrus fruits and so may serve as an indicator of a healthy diet as an intake marker. Mandarin oranges are one of the most commonly consumed fruits in Japan, and the FFQ results also confirmed that citrus fruits including mandarin were consumed in large amounts in the survey population. Therefore, proline betaine was shown to be a potential biomarker for citrus fruit consumption in the Japanese population as well.

Beverages

Three substances (quinate, trigonelline, and hippurate) were prominent metabolites related to coffee consumption, consistent with previous reports [36, 70, 71]. Esters of quinate and caffeate are abundant in coffee beans in the form of chlorogenate, which is the most commonly known coffee polyphenol and easily pyrolytically decomposed into these two compounds by heating [72]. Besides, trigonelline is a methyl betaine of nicotinate (niacin), which is also found in high levels in coffee beans [36]. Hippurate is known to exist in the urine of herbivores, and it has also been reported that it is biosynthesized from quinate by the gut microbiota [73].

Endogenous metabolites such as pipecolate and 2-AB, which are likely to indicate chronic metabolic changes, were shown to be markers of alcohol consumption. Plasma amino acid abnormalities have been frequently reported in alcoholics. Pipecolate is a metabolite of the essential amino acid lysine, generally found in urine and plasma. Studies have shown that plasma concentrations of pipecolate are elevated in patients with chronic liver diseases [74]. Besides, as mentioned earlier, the higher plasma concentrations of 2-AB with habitual alcohol intake may reflect altered glutathione metabolism and lipid peroxidation due to alcoholic liver dysfunction [66]. Indeed, comparative studies of healthy populations have suggested that active drinkers without the liver disease have higher 2-AB concentrations than non-drinkers [22, 75].

We avoided identifying potential markers for green tea consumption. The intake of green tea, the most common beverage consumed while eating in Japan, tends to increase with the frequency of meals in Japan. Therefore, metabolites with high VIP scores such as threonate, galactarate, proline betaine, cystine, and TMAO might reflect the relations with Japanese dietary patterns that are rich in fish and vegetables, rather than specific for green tea itself. Also, although tea catechins were likely to be characteristic components of green tea [76], they were not suitable for measurement via CE-MS because they are non-polar high-molecular-weight polyphenols.

Strengths and limitations

The present study had several strengths that should be noted. Being part of a large cohort study, there were enough data to draw a statistically supported and meaningful conclusion on the behavior of a large population. The study was also carefully designed for both epidemiological and metabolomics analyses. This indicates that there was an advantage in identifying circulating blood metabolites in long-term dietary habits in free life, which was carried out with minimal metabolic variations under a strict protocol by a non-targeted approach. Moreover, the study was carefully executed under overnight fasting conditions to limit the short-term effects of food intake.

Also, our study was pioneer research that aimed to clarify candidate food biomarkers in the Asian population, particularly habitual Japanese characteristics, although its findings may have limited generalizability. The study area was a rural town chosen for its steady population that remained less affected by rapid Westernization, unlike other urban areas. Therefore, it was ideal for research on long-term eating habits. Our findings will also help researchers to consider the influence of dietary factors in exploring biomarkers in various fields.

Despite these strengths, the study had some limitations. While large-scale epidemiological studies such as the cohort study on which our cross-sectional study was based have the great advantage of deriving meaningful knowledge from a large amount of data, it was necessary to adopt practical methods for collecting such data. Dietary assessment by FFQ, which represents participants’ habitual dietary intake over a longer time, is generally regarded as advantageous in terms of cost and time. However, the method tends to have inferior accuracy randomly as well as lower estimates of intake systematically. Having said this, in the statistical analysis of the present study, a relative log-normalized amount of each food was used instead of the absolute amount. Thus, it may have been affected by random errors rather than systematic errors.

An assessment categorized by food groups is likely to make it more challenging to distinguish the effects of specific food items that may differentially associate with metabolites, but it may also improve the interpretability of dietary status based on complex intake of various food items. Thus, we can say it is suitable for practical applications to assess long-term dietary habits. Indeed, the approach by food groups has been adopted to identify habitual food intake biomarkers in typical epidemiological studies [14, 64, 71].

Although CE-MS is an optimal metabolomics measurement technology with high resolution for capturing intracellular metabolites, it was unsuitable for measuring the levels of low polar molecules like lipids and high-molecular-weight polyphenols, which are associated with the intake of various foods. It is difficult to completely cover the metabolite profile with one platform; thus, dietary biomarker identification consisting of a wider range of chemical classes may need to be assessed with an integrated approach.

Another important question is whether the metabolite signatures detected statistically here could be used as dietary biomarkers. Among various types of dietary biomarkers defined previously [8, 9], it is arguable whether the metabolite signatures we observed could be defined as biomarkers of intake (i.e., concentration of replacement dietary biomarker). Follow-up surveys are ongoing now to estimate diet-disease risk association as well as to examine the function, mechanism of action, and the validity and reproducibility of the objective quantities. Effects of endogenous and extrinsic factors on metabolite concentrations are also to be evaluated.

Finally, dietary biomarkers are influenced not only by the intake of individual foods but also by the interaction of sex-specific and/or personal dietary behaviors and preferences. Therefore, it is important to consider such other factors in future studies of dietary biomarkers. The effects of gut microbiota and genetic factors on a diet also cannot be overlooked.