Metabolic profiling detects early effects of environmental and lifestyle exposure to cadmium in a human population

One hundred and eighty adult participants were recruited to the current study as described in Thomas et al. (2009) [37], and 178 samples in total were analyzed by ¹H NMR spectroscopy. Briefly, each individual was asked to provide one spot morning urine sample; each sample was frozen on the day of collection. Each volunteer completed a lifestyle questionnaire that provided information on a number of parameters including smoking status (current, past, never), age and sex. Data were anonymized for the present study. U-Cd and U-NAG were previously determined [37] and were both log-normally distributed. Ethical approval for the study was granted by Bristol (UK) South and Central Ethics Committee and all participants provided their written informed consent before data collection.

Samples were thawed on ice and prepared in a single batch. For each sample, an aliquot (0.4 mL) was mixed with buffer (0.2 mL, 0.2 M sodium phosphate, pH 7.4) containing ²H₂O (20% v/v) and sodium 3-trimethylsilyl-1-[2, 2, 3, 3,-²H₄] propionate (TSP, 0.3 mM) and centrifuged (16,000 g, ten minutes) to remove suspended particulate matter.

High-resolution ¹H NMR spectroscopy experiments were conducted using a Bruker AVANCE 600 spectrometer (Bruker Biospin, Rheinstetten, Germany) fitted with a flow-injection mode probehead (5 mm FI TXB ¹H-¹³C/¹⁵N-²H Z-GRD H8432/K0201 Z8432/0201, Bruker) at a field strength of 14.1 T (600.13 MHz ¹H frequency). The probe temperature was set to 300 K. Samples were delivered to the probehead in 96-well plates using an automated flow injection system (Bruker). Following introduction to the probe, samples were left to equilibrate (three minutes) prior to gradient shimming using the ¹H channel to ensure good magnetic field homogeneity. All liquid sample handling, automation and acquisition were controlled using XWIN-NMR 3.1 software (Bruker) running on a UNIX workstation (Silicon Graphics, USA).

One-dimensional NMR spectra were acquired using a standard pulse sequence using excitation sculpting with gradients to suppress the water resonance [38]. Following eight dummy scans, each spectrum was acquired into 64 k datapoints over a spectral width of 20 ppm as the sum of 64 transients. The acquisition time was 2.73 seconds, giving a native free induction decay (FID) resolution of 0.183 Hz. The relaxation delay was set at two seconds. The total acquisition time was approximately six minutes per sample. An apodization function equivalent to a line-broadening of 0.3 Hz was applied to each FID prior to Fourier transformation. Assignment of peaks to specific metabolites was based on the addition of known standards to the biological samples, together with published literature [39], on-line metabolomics databases and statistical total correlation spectroscopy (STOCSY) [40]. Additional NMR spectroscopy experiments were conducted to assign a doublet at a chemical shift of δ 1.11. A one-dimensional ¹H selective TOCSY experiment (See additional file 1, Figure S1) observed three related ¹H resonances in the same molecule. Further two-dimensional experiments [See additional file 1, Figure S2] allowed the full ¹H and ¹³C characterization [See additional file 1, Table S1] of the two diastereoisomers of 2,3-dihydroxybutanoic acid [See additional file 1, Figure S3]: 4-deoxy-erythronic acid and 4-deoxy-threonic acid, both of which have previously been identified in human urine [41].

Data were imported and manipulated in Matlab (Mathworks, Natick, MA, USA) using in-house code for automatic phasing, baseline correction, and referencing chemical shifts to the TSP resonance at δ 0. To account for variance in dilution of the urine samples each spectrum was subsequently normalized by the median-fold change to a reference spectrum generated by calculation of the median of all spectra at each spectral point [42]. The spectra were analyzed at two resolutions: (1) reduced resolution of 0.01 ppm width (1,127 data points), for partial least squares (PLS) regression analysis; and (2) high resolution (32,697 data points), for covariance/correlation analysis.

For multivariate analysis the reduced resolution data were exported to SIMCA-P+ (Umetrics) with the TSP (δ < 0.15) and residual water (δ 4.675 to 4.895) resonances removed. PLS regression analysis was applied to these data to model optimally the association between the metabolic profile and the lifestyle and biological factors (U-Cd, age, sex, U-NAG and smoking status) and to identify individual metabolites significantly contributing to any associations observed. Permutation analysis was carried out for each PLS model to test for validity using 1,000 permutations per test [See additional file 1, Figure S4].

Additionally, we used a variable-wise covariance and correlation based analysis of the high resolution data to assist in the identification of individual metabolites associated with lifestyle/biological factors. The results were visualized by color-scale projection onto the plots to indicate the correlation of metabolites with the individual factors, with red indicating a high correlation and dark blue no correlation. The direction and magnitude of the signals represent the co-variation of the metabolites with the specific lifestyle factor.

We measured the area under selected spectral peaks for six specific metabolites (3-hydroxyisovalerate (3-HV), dimethylglycine (DMG), citrate, 4-deoxy-erythronic acid (4-DEA), creatinine and creatine) identified from the above analyses. These data were then analyzed by multiple linear regression (MLR) using SPSS (v19, IBM, Armonk, NY, USA) controlling for major confounding factors (age, sex and smoking status). We developed three separate MLR models for citrate. Model 1 excluded individuals who did not answer the question regarding smoking status (n = 3); model 2 additionally excluded all individuals classed as current smokers (n = 32); and model 3 additionally excluded current (n = 32) and past smokers (n = 35). Additionally, we used non-parametric tests (Kruskal-Wallis/Mann-Whitney or Spearman's ρ) to test for a significant (P < 0.05) difference or trend, respectively, in the concentration of citrate between the three classes of smoker (current-, past-, and never smoked). A normal distribution of values in each group in the dataset was observed (P < 0.05, Shapiro-Wilk test).

Urinary-8-oxodG (U-8-oxodG) was quantified in samples from the upper and lower 15^th percentile according to U-Cd concentration, as described previously [43]. The U-8-oxodG concentrations were normalized to the creatinine concentration (taken from [37]) and a non-parametric Mann-Whitney test used to test for a significant (P < 0.05) difference between the groups. A non-normal distribution of values in the upper 15^th percentile group was observed (P < 0.05, Shapiro-Wilk test).

Our first objective was to determine if there were any associations between the global metabolic profile and U-Cd. Initial inspection of the spectral data indicated that a substantial number of individuals had detectable levels of analgesics (that is, paracetamol and ibuprofen) and/or ethanol detectable in their samples (51/178, approximately 29%, see additional file 1, Table S2). In order to avoid subsequent models being negatively affected by the global diuretic or pharmacological effects of these compounds on the spectra, these individuals were excluded from initial pattern recognition analyses. From the reduced resolution data-set (see Methods), PLS regression could predict U-Cd based on the metabolic profile data (Q² = 0.237, P < 0.01, Table 1). Inspection of model regression weights and bivariate analyses at both low and high resolution (Figure 2A) identified a number of urinary metabolites responsible for this association, namely: citrate, 3-HV, DMG, creatinine, creatine and 4-DEA. DMG, 3-HV, creatinine and 4-DEA were negatively correlated to U-Cd concentration and citrate was positively correlated.

Since smoking is a major source of cadmium exposure we excluded smokers from PLS modelling to test if the link between metabolite profiles and U-Cd was independent of smoking status. PLS models with significant prediction of U-Cd could be generated either with current smokers excluded (Q² = 0.330, P < 0.001, Table 1) or with both current smokers and past smokers excluded (Q² = 0.142, P < 0.001, Table 1), suggesting that exposure to Cd from environmental and dietary sources alone may be sufficient to produce detectable metabolic effects.

It has already been shown in this population that U-Cd is significantly associated with age, sex, U-NAG and smoking status [37]. In order to obtain a better understanding of possible causal links and the specificity of the responses to Cd we conducted further pattern recognition analyses to attempt to derive models for these potentially confounding factors. We could predict a statistically significant degree of variation by PLS regression in each of: age (Q² = 0.224, P < 0.001), smoking status (Q² = 0.185, P < 0.01) and U-NAG levels (Q² = 0.162, P < 0.001) from the urinary NMR data (Table 1). Inspection of regression weights for each of these models indicated that several metabolites correlated with U-Cd levels (that is, 3-HV, DMG, 4-DEA and creatinine) were also associated with these other physiological and lifestyle factors. Further specific metabolite correlates were defined, including cotinine to smoking status and creatine to sex (Figures 2B-2E).

Next we measured the area under selected spectral peaks (the integral) for the six metabolites (3-HV, DMG, citrate, creatinine, creatine and 4-DEA) identified by pattern recognition analysis as associated with U-Cd, and analyzed these in more detail by MLR. Samples previously excluded (n = 51) from pattern recognition analysis due to the presence of ethanol or analgesics were evenly distributed across age, smoking status, gender and U-Cd, reducing the likelihood that including these samples would confound the analysis [See additional file 1, Table S2]. Hence we felt it was justified to include these data for individual metabolite analyses to increase power wherever the relevant resonances could be integrated without interference. The integral of each metabolite was regressed as a function of U-Cd, age, sex and smoking status (Table 2). There was a significant association between citrate levels and both U-Cd (beta = 0.194, P < 0.05) and smoking status (beta = -0.298, P < 0.001) controlling for age and sex. Interestingly, citrate was positively correlated to U-Cd but negatively correlated to smoking status. This suggests that exposure to cigarette smoke via inhalation, although known to significantly increase Cd body burden, may have a different effect than Cd exposure alone via ingestion and/or inhalation (as assessed by U-Cd). Other metabolites showed significant associations to the potential confounding factors but not to U-Cd. 3-HV was negatively associated with age (beta = -0.326, P < 0.001) and sex (beta = -0.233, P < 0.01), and 4-DEA to age only (beta = -0.370, P < 0.001). Creatinine was negatively associated with age (beta = -0.427, P < 0.001) and sex (beta = -0.395, P < 0.001). Creatine was positively associated with sex only (beta = 0.321, P < 0.001). Again, we excluded smokers from the regression models, in order to establish if environmental exposure alone could account for the link between Cd body burden and perturbed endogenous metabolism. Citrate levels remained significantly associated with U-Cd while controlling for all other factors with either current smokers, or current and past smokers, removed from the regression model (Table 2; models citrate(2) and citrate(3) respectively).

Interestingly, there was a significant association between smoking status and citrate, even when current smokers were removed, suggesting that there was a metabolic difference between those individuals who had never smoked and those who had previously smoked. Normalized citrate levels were significantly different between groups (P < 0.05, Kruskal-Wallis test), specifically between current smokers and those who had never smoked (P < 0.01, Mann-Whitney test). There was also a clear negative trend in citrate levels with smoking status (Spearman's ρ = -0.225, P < 0.01), such that ex-smokers had intermediate levels of citrate on average compared to the other two groups (Figure 3A). This suggests that smoking could have induced a long-term change in citrate metabolism or transport, possibly reflecting an effect on mitochondria, which was manifested in decreased urinary citrate levels.

Alterations in citrate levels could imply changes in the consumption of citrate via the Krebs cycle in the mitochondrion. A perturbation to oxidative metabolism would be expected to lead to a change in ROS production and accompanying oxidative stress. Cd has been shown to increase ROS production in vivo [31, 44] and indirectly damages DNA in vitro via oxidation by hydrogen peroxide production [45]. We attempted to define the association of heavy metal exposure (using U-Cd as an indicator) and the potential for oxidative DNA damage by measuring U-8-oxodG, which is widely used as a biomarker of oxidative stress [46, 47]. We measured U-8-oxodG in a subset of 40 individuals (the upper and lower 15% of volunteers according to U-Cd concentration (Figure 3B)). The mean U-Cd concentration in the lower group was 0.115 ± 0.006 (nmol/mmol creatinine ± standard error of the mean (SEM)) and 0.766 ± 0.052 in the higher group. There was a significant (P < 0.05, Mann-Whitney test) increase in U-8-oxodG concentration (median ± SD) in the individuals in the upper 15^th percentile group (1.77 ± 0.82 pmol/mmol creatinine), compared to those in the lower group (1.39 ± 0.45 pmol/mmol creatinine). This suggests a positive association between U-8-oxodG and U-Cd, which could be representative of mitochondrial dysfunction and increased ROS production, consistent with our hypothesis.