Study design
A total of 91 participants were recruited from January 2013 to March 2014 as part of the Sakura Diet Study. This study is part of the Japan Multi-Institutional Collaborative Cohort (J-MICC) Study in the Shizuoka-Sakuragaoka area, and has been described previously [18, 19]. This investigation was conducted according to the Declaration of Helsinki human subject principles, and the study protocol was approved by the ethics committee of the University of Shizuoka (No. 24–24). In brief, a representative sample of citizens living in Shizuoka city, Japan was recruited and asked by to supply both a 3-day WDR and a blood sample in each of the four seasons. Written informed consent was obtained from each participant by trained medical staff. At the beginning of the study (i.e., the first season, winter), subjects filled out a FFQ and a lifestyle questionnaire with questions about their medical history. Exclusion criteria for the current study included the following: 1) subjects with dyslipidemia, diabetes, angina pectoris or cerebral stroke; and 2) subjects who did not provide > = 8 h’ fasting blood samples in any seasons.
There were two study designs: 1) a cross-sectional design, using data based on the FFQ, 3-day WDRs and a serum lipid profile collected in the first season; and 2) a seasonal variation design, using data derived from 3-day WDRs and a serum lipid profiles collected in each season. Three datasets were prepared: i) ‘complete cases (CC)’ composed of subjects with no missing values in any of the four seasons; ii) ‘observed subjects (OS)’, which included subjects who provided data in each season, even if some values were missing; and iii) a MI dataset calculated by means of applying the MI procedure to the OS dataset.
Dietary assessments and lifestyle factors
Initially, 26 nutrient intakes were estimated using the FFQ, with 47 food items over the past 1 year [10, 20, 21]. Subsequently, using three-nonconsecutive-day WDRs, subjects were asked to record their eaten foods and drinks on two weekdays and one weekend day in each season. All 12-day WDRs were systematically reviewed by two trained registered dieticians. According to Standard Tables of Food Composition in Japan, 2015 (seventh revised edition) and other items [22,23,24,25], individual representative intake of foods and nutrients was given using an average value of WDRs for 3 days in each season. Food items not found in these sources were replaced with similar items in order to convert nutrients, with the advice of two expert registered dieticians (i.e., nutritional epidemiologists), who were involved with the J-MICC study. Nutrients were highlighted include fat, cholesterol, TDF and ethanol. Energy-adjusted intake was used for all analyses, except ethanol.
Using a self-administered lifestyle questionnaire, alcohol consumption, including consumption of six types of alcoholic beverages (sake, shochu (distilled spirit), chuuhai, beer, whiskey, and wine) was assessed [26]. In brief, ethanol consumption (g/day) of current alcohol consumers was estimated based on the frequency and amount of each type of alcoholic beverage consumed over the past year, with reference to 180 ml of sake (one “go”), 633 ml of beer (one bottle), 108 ml of shochu containing 25% ethanol, etc., as 23 g of ethanol. Smoking status was categorized as either current smoker (which included a small number of ex-smokers) or non-smoker. Physical activity was assessed by self-report and included questions about the number of hours of daily activity and leisure-time exercise which was then converted into metabolic equivalent task hours (METs•hour/day) [27].
Biochemical and anthropometric measurements
Concentrations of LDL-C, total and HDL cholesterol and triglyceride levels were measured by an automatic biochemical analyzer in the SRL clinical laboratory (CAP ISO 15189). LDL-C concentration was calculated by Friedewald equation [28], or applied measurement values of LDL-C in the case of subjects with triglyceride levels > = 4.49 mmol/L. Height and weight were measured within 0.1 cm and 0.1 kg respectively, after which body mass index (BMI) was calculated.
Genotyping and quality control
DNA was extracted by a QIAcube (Tokyo, Japan) with a Qiagen DNA extraction kit (QIAamp DNA Blood Mini Kit). All 659,253 tag SNPs were genotyped with the Japonica array (Toshiba, Tokyo, Japan), which covered 96.9% of common SNPs (minor allele frequency > 5%) and 67.2% of low-frequency SNPs (0.5% < minor allele frequency < =5%) [29]. In addition, this array was reported to provide better imputation performance for Japanese individuals than other SNP arrays with reference to 1KJPN and the International 1000 genomes projects panel. After systematically reviewing the GWAS findings on LDL-C concentration in East Asians (key terms; GWAS, LDL-C and lipids, and eligible populations; Japanese, Korean and Chinese), a total of 39 SNPs related to LDL-C concentration were selected as candidates with reference to their coding genes and minor allele frequencies (Additional file 1) [13, 30,31,32,33,34,35]. Characterized SNPs found in American-European people were excluded because of being higher LDL-C concentrations than in East Asian populations [36, 37]. Of the 39 candidate SNPs, 10 were measured with Japonica array: PSRC1 rs599839; HMGCR rs12654264; TIMD4 rs6882076; TBL2 rs17145738; rs651007; intron between ABO and LCNIP2 rs579459; LDLR rs7258950 and rs2738446; TOMM40 rs1160985; APOC1 rs445925. The remaining SNPs were not covered by the array chip because they are minor allele frequencies. After quality control checks, three additional SNPs were excluded from the analyses for the following reasons; 1) the call rate of rs2738446 was < 100%; 2) rs7258950 was not in Hardy-Weinberg equilibrium (P < 0.05); and 3) rs579459 was in linkage disequilibrium with rs651007 (r = 1.0).
Handling of missing values
Utilizing the ‘mice’ package, MI was conducted using chained equations for missing values [38]. Missing values were assumed to be missing at random, and 50 and 100 imputed datasets (m) and 40 iterations were executed. The assumption was assessed with a density plot [38]. Without collinearity in the models, an imputed dataset was created; subsequently, a number of iterations were counted using the ‘norm2’ package. To investigate the convergence of MI, a trace plot was drawn [38], and results from m = 50 (data not shown) and 100 were compared. Sensitivity analysis was applied to assess the robustness of the results under a missing not at random assumption [38].
Statistical analysis
All analyses were performed using R software, version 3.4 (R Project for Statistical Computing), and a P < 0.05 was considered statistically significant. The continuous and categorical variables were represented as mean ± SD and by adding the numbers and percentages. Non-normally distributed variables such as LDL-C concentration, were log-transformed. To avoid the value of zero, 1 g/day of ethanol consumption was assigned before the log-transformation, and no effect was observed after this assignment. Using the ‘lmerTest’ package, a mixed-effect model was applied to assess seasonal variation, and the two variables, season and subject, were used as a fixed effect (‘1 to 4’ for ‘winter to autumn’) and a random intercept (‘a serial number’), respectively [39]. Dunnett’s test was used for post hoc test with reference to winter. For quality control of genotyping, Hardy-Weinberg equilibrium was examined using chi-square tests, and linkage disequilibrium was evaluated with Pearson’s correlation coefficients.
In the cross-sectional study, a multiple linear regression was applied to LDL-C concentration to detect any independent relationships with nutrient intake and SNPs (homozygous for the major allele/other genotypes = 0/1). Variables used as conventional confounding factors included age, BMI, physical activity, ethanol consumption, and total energy intake (which was excepted in case of SNPs) as continuous variables; and sex (men/women = 0/1) and smoking status (non−/smokers = 0/1) as categorical variables. In the seasonal variation study, using the ‘lmerTest’ package, a mixed-effect model was applied, taking into account each effect of season and subject. Subsequently, interactions between nutrient intake and SNPs were calculated when the P < 0.20 was found in the OS dataset of the seasonal variation study.
