Prospective analysis of circulating metabolites and breast cancer in EPIC

EPIC is an ongoing multi-center cohort study including approximately 520,000 participants recruited between 1992 and 2000 from ten European countries [19]. Female participants (n = 367,903) were aged 35–75 years old at inclusion. At recruitment, detailed information was collected on dietary, lifestyle, reproductive, medical, and anthropometric data [19]. Around 246,000 women from all countries provided a baseline blood sample. Blood was collected according to a standardized protocol in France, Germany, Greece, Italy, the Netherlands, Norway, Spain, and the UK [19]. Serum (except in Norway), plasma, erythrocytes, and buffy coat aliquots were stored in liquid nitrogen (− 196 °C) in a centralized biobank at IARC. In Denmark, blood fractions were stored locally in the vapor phase of liquid nitrogen containers (− 150 °C), and in Sweden, they were stored locally at − 80 °C in standard freezers.

Incident cancer cases were identified through record linkage with cancer registries in most countries and through health insurance records, cancer and pathology registries, and active follow-up of study subjects in France, Germany, and Greece. For each EPIC center, closure dates of the study period were defined as the latest dates of complete follow-up for both cancer incidence and vital status (dates varied between centers, from June 2008 to December 2012).

All participants provided written informed consent to participate in the EPIC study. This study was approved by the ethics committee of the International Agency for Research on Cancer (IARC) and all centers.

Subjects were selected among participants who were cancer-free (other than non-melanoma skin cancer) and had donated blood at recruitment into the cohort. Cancers were coded according to the Third Edition of the International Classification of Diseases for Oncology (code C50). Women diagnosed with first primary invasive breast cancer at least 2 years after blood collection and before December 2012, for whom estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2) statuses of the tumors were available, were selected as cases for the current study.

For each breast cancer case, one control was chosen at random among appropriate risk sets comprising all female cohort members who were alive and without cancer diagnosis (except non-melanoma skin cancer) at the time of diagnosis of the index case. Using incidence density sampling, controls were matched to cases on center of recruitment, age (± 6 months), menopausal status (premenopausal, perimenopausal, postmenopausal, surgically postmenopausal [20]), phase of the menstrual cycle [20], use of exogenous hormone at blood collection, time of the day (± 1 h), and fasting status at blood collection (non-fasting (< 3 h since last meal), in between (3–6 h), fasting (> 6 h), unknown).

Initially, 1626 cases and 1626 controls were eligible for the study, but after the exclusion of pregnant women at blood collection, a final population of 1624 cases and 1624 controls were included in the analysis.

Metabolites were analyzed in samples from 3247 distinct subjects (one subject included in 2 pairs). Completeness of measures and coefficients of variation (median = 5.3%, interquartile range = 1.4%) are shown in Additional file 1: Table S1. Values lower than the lower limit of quantification (LLOQ), or higher than the upper limit of quantification (ULOQ), as well as lower than batch-specific limit of detection (LOD) (for compounds measured with a semi-quantitative method: acylcarnitines, glycerophospholipids, sphingolipids), were considered out of the measurable range. Metabolites were excluded from the statistical analyses if more than 20% of observations were outside the measurable range (n = 20). A total of 127 metabolites (8 acylcarnitines, 20 amino acids, 6 biogenic amines, 78 glycerophospholipids, 14 sphingolipids and hexoses) were finally retained for statistical analyses. Of these 127 metabolites, 113 had all values included in the measurable range. For the remaining 14 metabolites, values outside the quantifiable range (all lower than LLOQ or LOD) were imputed with half the LLOQ or half the batch-specific LOD, respectively.

Characteristics of cases and controls were described using mean and standard deviation (SD) or frequency. Geometric means were used to describe non log-transformed metabolite concentrations among cases and controls. Log-transformed metabolite concentrations were used in all other analyses. Partial Pearson’s correlations between metabolites, adjusted for age at blood collection, were estimated among controls.

We used conditional logistic regression to estimate the risk of breast cancer per standard deviation (SD) increase in metabolite concentration. The analysis was conditioned on the matching variables. Likelihood ratio tests were performed to compare linear models with cubic polynomial models in order to assess departure from linearity. Multiple testing was addressed by controlling for family-wise error rate at α = 0.05 by permutation-based stepdown minP adjustment of P values, as this method better accounts for the dependence of the tests [21, 22]. For comparison with previous studies, we also adjusted the raw P values using Bonferroni correction (P < 0.05/127) and controlling for the false discovery rate (FDR) at α = 0.05 [23]. All statistical tests were two-sided.

Metabolites showing a statistically significant association with risk of breast cancer after correcting for multiple testing were categorized into quintiles based on the distribution of the concentrations among controls, and odds ratios (OR) for risk of breast cancer were estimated in each category. For tests of linear trend, participants were assigned the median value in each quintile and we modeled the corresponding variable as a continuous term. To identify potential confounders, models of the metabolites of interest (continuous and quintiles) were adjusted separately for each potential confounder and estimates obtained were compared with estimates from models with matching variables only. Only variables that changed parameter estimates by more than 10% were retained in the multivariable model. Variables tested were as follows: age at first menstrual period (continuous), number of full-term pregnancies (0/1/2/≥ 3), age at first full-term pregnancy (never pregnant/quartiles), breastfeeding (ever/never/never pregnant/missing; duration in quintiles), ever use of oral contraceptive (yes/no), ever use of MHT (yes/no/missing), smoking status (never/former/current), level of physical activity (Cambridge index [24]: inactive/moderately inactive/moderately active/active), alcohol consumption (nondrinkers/> 0–3/3–12/12–24 g/day), education level (no schooling or primary/technical, professional or secondary/longer education), energy intake (continuous, quintiles), height (continuous, quintiles), sitting height (missing/quartiles), weight (continuous, quintiles), body mass index (continuous, quintiles), waist circumference (continuous, quintiles), hip circumference (continuous, quintiles), and hypertension (yes/no). For these variables, missing values were assigned the median (continuous variables) or mode (categorical variables) if they represented less than 5% of the population, or were otherwise classified in a “missing” category (breastfeeding, ever use of MHT, sitting height). Only waist circumference (continuous), hip circumference (continuous), and weight (continuous) were included in the final models. Given the correlations between these variables (> 0.77), these variables were included separately in three different models.

For those metabolites showing a significant association with breast cancer risk after controlling for multiple testing, heterogeneity was investigated by menopausal status at blood collection, use of exogenous hormones at blood collection, fasting status at blood collection, age at diagnosis (age 50 or older/younger than age 50), breast cancer subtype (ER+PR+/−HER2+, ER+PR+/−HER2−, ER−PR−HER2+, ER−PR−HER2−), time between blood collection diagnosis (2–8.6 years/more than 8.6 years), and at recruitment waist circumference (WC) (< 80 cm/≥80 cm), BMI (< 25 kg/m²/≥25 kg/m²), and country, by introducing interaction terms in the models. Subgroup analyses were conducted on the raw models. For WC, unconditional logistic regression adjusted for each matching factor was used. P values were not corrected for multiple tests since heterogeneity was investigated only for metabolites showing statistically significant associations with risk overall, after correction for multiple testing.

A sensitivity analysis of all 127 metabolites was performed on hormone non-users (1124 cases and 1124 controls) and by cancer subtype.

Analyses were conducted using SAS software for Windows (version 9.4, Copyright© 2017, SAS Institute Inc.) and R software (packages Epi and NPC) [25, 26].

Prior to correction for multiple testing, 29 metabolites were significantly associated with the risk of breast cancer with a raw P value lower than 0.05 (Fig. 1a and Table 2), mainly amino acids, PCs (inversely associated), and acylcarnitines (directly associated). However, after adjusting for multiple testing (Fig. 1b), only C2 (OR for 1 SD increment = 1.15, 95% CI = 1.06–1.24, corrected P value = 0.031) and phosphatidylcholine PC ae C36:3 (OR for 1 SD increment = 0.88, 95% CI = 0.82–0.95, corrected P value = 0.044) remained significantly associated with risk of breast cancer (Table 2). Adjustment for multiple testing using FDR procedure identified similar significant metabolites, while with Bonferroni correction, only C2 remained associated with risk of breast cancer with a borderline significant P value (Bonferroni P value = 0.051) (Table 2). Departure from linearity was suggested for glutamate, C0, kynurerine, and SDMA. However, when non-linear models were examined, and after controlling for multiple tests, no non-linear association remained significant (results not shown).

When C2 and PC ae C36:3 were further analyzed as categorical variables, results similar to those of the linear analysis were obtained; logistic regression conditioned on the matching variables showed a linear trend across quintiles of C2 (OR quintile 5 versus quintile 1 = 1.54, 95% CI = 1.21–1.95, P trend = 0.0002) and of PC ae C36:3 (OR quintile 5 versus quintile 1 = 0.73, 95% CI = 0.58–0.91, P trend = 0.0003) (Table 3). Adjusting for anthropometric variables in separate models had little effect on the risk estimates (Table 3).

Statistically significant heterogeneity was observed by use of hormones at blood collection for the associations of C2 (P homogeneity = 0.035) and PC ae C36:3 (P homogeneity = 0.017) with breast cancer, with statistically significant associations restricted to hormone non-users (C2: OR per SD = 1.23, 95% CI = 1.11–1.35; PC ae C36:3: OR per SD = 0.83, 95% CI = 0.76–0.90) and no associations observed in users (C2: OR per SD = 1.03, 95% CI = 0.91–1.17; PC ae C36:3: OR per SD = 1.00, 95% CI = 0.88–1.13; Fig. 2).

In an analysis of the 127 metabolites restricted to hormone non-users (n = 2248) (Fig. 3), we identified additional metabolites showing statistically significant inverse associations with risk of breast cancer after adjustment of P values for multiple testing, for which heterogeneity was also investigated. These metabolites were as follows: arginine (OR per SD = 0.79, 95% CI = 0.70–0.90; P homogeneity = 0.002), asparagine (OR per SD = 0.83, 95% CI = 0.74–0.92; P homogeneity = 0.12), PC aa C36:3 (OR per SD = 0.84, 95% CI = 0.77–0.93; P homogeneity = 0.12), PC ae C34:2 (OR per SD = 0.85, 95% CI = 0.78–0.94; P homogeneity = 0.04), PC ae C36:2 (OR per SD = 0.85, 95% CI = 0.78–0.88; P homogeneity = 0.04), and PC ae C38:2 (OR per SD = 0.84, 95% CI = 0.0.76–0.93; P homogeneity = 0.10).

No significant heterogeneity was observed for the association of C2 and PC ae C36:3 with breast cancer by menopausal status, fasting status at blood collection, breast cancer subtype, age at diagnosis, WC (P homogeneity all > 0.12, Fig. 2), country (P homogeneity of 0.50 for C2 and 0.12 for PC ae C36:3) or by time between blood collection and diagnosis (2–8.6/≥8.6 years (median); P homogeneity of 0.17 for C2 and 0.98 for PC ae C36:3) (data not shown).

Stratification by breast cancer subtypes for all metabolites (see Additional file 1: Figure S2) showed that no metabolite reached statistical significance after correction for multiple testing in each subtype, although for ER+PR+/−HER2− cases (n = 1084 cases), PC ae C36:3 and PC aa C36:3 had adjusted P values close to statistical significance (0.066 and 0.074, respectively).