Oliver
Robinson
abcd,
Mireille B.
Toledano
a,
Caroline
Sands
e,
Olaf
Beckonert
e,
Elizabeth
J. Want
e,
Rob
Goldin
f,
Michael L.
Hauser
gh,
Alan
Fenwick
h,
Mark R.
Thursz
f and
Muireann
Coen
*e
aMRC-PHE Centre for Environment and Health, School of Public Health, Imperial College London, UK
bISGlobal, Centre for Research in Environmental Epidemiology (CREAL), Spain
cHospital del Mar Medical Research Institute (IMIM), Barcelona, Spain
dCIBER Epidemiología y Salud Pública (CIBERESP), Spain
eComputational and Systems Medicine, Department of Surgery & Cancer, Faculty of Medicine, Imperial College London, UK. E-mail: m.coen@imperial.ac.uk
fDepartment of Medicine, Imperial College London, UK
gOne Health Foundation, Switzerland
hSchistosomiasis Control Initiative, School of Public Health, Imperial College, London, UK
First published on 12th August 2016
Several hundred cases of Hirmi Valley Liver Disease (HVLD), an often fatal liver injury, occurred from 2001 to 2011 in a cluster of rural villages in Tigray, Ethiopia. HVLD is principally caused by contamination of the food supply with plant derived pyrrolizidine alkaloids (PAs), with high exposure to the pesticide DDT among villagers increasing their susceptibility. In an untargeted global approach we aimed to identify metabolic changes induced by PA exposure through 1H NMR spectroscopic based metabolic profiling. We analysed spectra acquired from urine collected from HVLD cases and controls and a murine model of PA exposure and PA/DDT co-exposure, using multivariate partial least squares discriminant analysis. In the human models we identified changes in urinary concentrations of tyrosine, pyruvate, bile acids, N-acetylglycoproteins, N-methylnicotinamide and formate, hippurate, p-cresol sulphate, p-hydroxybenzoate and 3-(3-hydroxyphenyl) propionic acid. Tyrosine and p-cresol sulphate were associated with both exposure and disease. Similar changes to tyrosine, one-carbon intermediates and microbial associated metabolites were observed in the mouse model, with tyrosine correlated with the extent of liver damage. These results provide mechanistic insight and implicate the gut microflora in the human response to challenge with toxins. Pathways identified here may be useful in translational research and as “exposome” signals.
During the period 2001–2011, an outbreak of several hundred cases of an often-fatal toxic hepatitis, referred to here as Hirmi Valley Liver Disease (HVLD), occurred in a cluster of rural villages in the North-Western zone of Tigray, Ethiopia.5,6 It is characterised by epigastric pain and abdominal swelling and appears to be principally caused by exposure to the plant hepatotoxins, pyrrolizidine alkaloids (PAs),7 including acetyllycopsamine (AL).8 The invasive weed Ageratum conyzoides which is highly prevalent in fields of staple crops such as millet, has been identified as the likely source. PAs have caused several outbreaks of severe liver disease worldwide following food contamination and are metabolised in the liver to produce toxic pyrrole metabolites.9 The primary liver injury among acute, recent-onset HVLD cases is centrilobular necrosis while among chronic, long-term cases the pathological features include cytomegaly, bile ductular reaction and various stages of fibrosis. Toxicological testing has demonstrated that AL induces similar pathology in mice including centrilobular necrosis and cytomegaly. The residents of the affected villages were also highly exposed to the pesticide DDT (dichlorodiphenyltrichloroethane), due to its use on food grain to protect against storage pests. In the mouse model DDT increased susceptibility to the hepatotoxic effects of AL through induction of the cytochrome P450 enzymes, primarily CYP3A.8
Here we have applied 1H NMR spectroscopic based metabolic profiling to urine samples collected from HVLD cases and controls. We aimed to identify markers of pyrrolizidine alkaloid induced liver disease that would provide mechanistic insight into the pathogenesis of the disease and that could potentially aid in differential diagnosis. Furthermore, we have explored the utility of this approach as an ‘exposomics’ platform to detect urinary markers of known and unknown toxin exposure. Finally, we aimed to validate and translate our results in humans through a comparative analysis of urine samples collected from mice exposed to both AL alone and a combination of AL and DDT under controlled conditions.
Urine samples (n = 90) were collected over two visits to the Tigray region. Collections were made during morning visits to clinics at the Kiburto kabelle health post in 2008 and the Kelakil kabelle healthpost in 2009, in a newly-built village for displaced residents of Tseada Amba, the village originally affected by HVLD. 41 subjects that met the pre-defined case definition (abdominal distension, hepatomegaly or splenomegaly on clinical examination and either abdominal pain for at least two weeks or another household member with similar symptoms8) and 41 subjects that met the control definition (individuals with no signs of liver disease) were included in the analysis. Furthermore, in the 2009 collection, two cases were hospitalised at the time of collection enabling overnight longitudinal sampling. From one of these patients, four samples were collected and their spectra were included in the analysis, to account for temporal variation. Controls were further classified as either a ‘household control’ (a healthy individual who shares a household with at least one HVLD case), or a ‘village control’ (a healthy individual who lives in a household free from HVLD).
Samples were collected in sterile containers and stored on ice until they were frozen at −20 °C and then transported on ice packs to Imperial College London and stored at −80 °C prior to analysis. The 2009 collected samples were collected into containers pre-filled with the preservative boric acid, while the 2008 collected samples were collected without the addition of preservative. Relative levels of AL in each urine sample were measured by ultra performance liquid chromatography – mass spectrometry (UPLC-MS) as previously described.8
Hepatotoxicity was assessed by alanine transaminase (ALT) activity and histopathological examination of plasma and liver samples collected on sacrifice. To grade the magnitude of hepatotoxic effects in the DDT and AL co-dosed group, a histopathological severity score was constructed as follows: 1 = normal, 2 = mild vacuolation/congestion, 3 = hepatocyte swelling, 4 = zone 3 necrosis, 5 = zone 2 and 3 necrosis. Urine samples were collected before dosing in the single dose experiment and on sacrifice in all experiments by free-catch or by extraction directly from the bladder and immediately frozen at −80 °C following the addition of 1% sodium azide as a preservative.
One-dimensional (1D) 1H NMR spectra were acquired using a Bruker Avance II NMR spectrometer (Bruker Biospin, Rheinstetten, Germany) operating at a 1H frequency of 600 MHz. A standard 1D solvent suppression pulse sequence was used to acquire the free induction decay (FID; relaxation delay – 90° pulse – 4 μs delay – 90° pulse – mixing time – 90° pulse – acquire FID). The D2O present in the buffer provided a field frequency lock, whilst the TSP served as the chemical shift reference compound (δ1H = 0.00). For acquisition of the human samples 256 scans and 8 dummy scans were collected into 65000 data points with a spectral width of 12 ppm, relaxation delay of 4 seconds, mixing time of 100 ms and an acquisition time of 4.56 seconds. For acquisition of the mouse samples 128 scans and 8 dummy scans were collected into 65000 data points with a spectral width of 20 ppm, relaxation delay of 2 seconds, mixing time of 100 ms and an acquisition time of 2.72 seconds. A line-broadening factor of 0.3 Hz was applied prior to Fourier transformation. Spectra were manually phased and baseline corrected using TOPSPIN (version 2.1, Bruker BioSpin).
Partial Least Squares Discriminant Analysis (PLS-DA) was applied to model the data and identify discriminatory features based on class membership. For the human samples orthogonal-PLS-DA (O-PLS-DA) was used which incorporates an orthogonal signal correction filter to remove variance not correlated to class membership. All PLS models were first internally validated using 7-fold cross-validation, where portions of the data are subsequently left out and their class membership then predicted to give an indicator of predictive ability, Q2. Models were then externally validated through random permutation of class membership 1000 times. If the Q2 of the actual model was in the top 5% of possible Q2 scores, the model was accepted as valid (i.e. p < 0.05 significance level). Similar permutation of the spectral data (10000 times) was used to assess the significance of metabolites identified as most strongly loading onto the discriminatory component of the PLS-DA models; metabolites with significance level p < 0.05 were reported.
Differences in age, urinary AL levels (log-transformed) and duration of disease were assessed by t-tests and gender by χ2 tests. Univariate testing of selected metabolites in the co-treatment mouse model was conducted using Mann–Whitney tests and Spearman's correlations with liver injury.
Collection | HVLD status | N | % Male | Mean age, years (range) | Geometric mean urinary AL, a.u. (95% C.I.) | % Ill less than 1 year | Mean duration of illness, months (range) |
---|---|---|---|---|---|---|---|
a Significantly different compared to 2009 collection cases (p < 0.001) based on two-tailed t-test. b Significantly different compared to village controls (p = 0.004) based on χ2 test. c Significantly different compared to village controls (p = 0.001) based on one-tailed t-test. A.U. = arbitrary units. Duration of illness refers to time since reported onset of symptoms. | |||||||
Kiburto Health Post 2008 | Cases | 10 | 80 | 23.4 (4–63) | 89.9 (45–179.5) | 70 | 13 (2–48)a |
Controls | 7 | 57 | 37.6 (18–69) | 39.7 (13.5–116.6) | — | — | |
Kelakil Health Post 2009 | Cases | 31 | 60b | 28.8 (4–48) | 126.4 (91.9–173.8)c | 6 | 37 (6–60) |
Household controls | 18 | 11 | 31.2 (5–70) | 84.2 (44.5–159.2) | — | — | |
Village controls | 17 | 28 | 23.7 (4–55) | 50.7 (29.1–88.6) | — | — |
Three separate cross-validated O-PLS-DA models were constructed with the model statistics presented in Table 2 showing the validity of all models. Model 1 compared 10 cases and 7 controls from the 2008 collection with the scores plot given in Fig. 1a, showing clear separation of cases and control by cross-validated predictive scores (Tcv). The loadings coefficient plot for this model showed that concentrations of bile acids, pyruvate, tyrosine, formate and N-methylnicotinamide were elevated among cases while concentrations of microbial-associated metabolites p-cresol sulphate, hippurate, 3-(3-hydroxyphenyl) propionic acid (3-HPPA) were reduced (Fig. 1b, Table 2).
Model 1 | Model 2 | Model 3 | |
---|---|---|---|
Collection | 2008 | 2009 | 2009 |
Class comparison (N. of samples) | Case (10) vs. control (7) | Case (34a) vs. village control (17) | Household control (18) vs. village control (17) |
Orthogonal components | 2 | 3 | 2 |
Q 2 Y | 0.48 | 0.35 | 0.20 |
R 2 Y | 0.98 | 0.97 | 0.96 |
R 2 X | 0.33 | 0.29 | 0.22 |
Permutation p value | 0.035 | 0.001 | 0.041 |
Metabolite | Loading correlation coefficient (r) | ||
---|---|---|---|
Bold = Spectral permutation p < 0.05; * = p < 0.01; ** = p < 0.005. ‘– ‘ = metabolite not detectable or obscured by overlap with other spectral peaks.a Four longitudinal samples from one case included. 3-HPPA = 3-(3-hydroxyphenyl) propionic acid. | |||
Bile acids | 0.67** | 0.50** | 0.12 |
p-Cresol sulphate | −0.57** | −0.39** | −0.31 |
Pyruvate | 0.62** | — | −0.06 |
Hippurate | −0.75** | −0.33** | −0.24 |
3-HPPA | −0.69** | −0.07 | 0.04 |
Tyrosine | 0.62** | 0.22 | 0.46** |
Formate | 0.55* | 0.29 | 0.16 |
N-Acetylglycoprotein 1 | 0.14 | 0.41** | 0.14 |
N-Acetylglycoprotein 2 | — | — | 0.45* |
p-Hydroxybenzoate | — | −0.42** | −0.45 |
N-Methylnicotinamide | 0.42 | 0.30** | 0.03 |
Unknown 1 | −0.75** | — | — |
Unknown 2 | 0.76** | — | — |
Unknown 3 | 0.75** | — | — |
Unknown 5 | 0.01 | 0.40** | 0.19 |
Unknown 6 | 0.14 | 0.44** | 0.15 |
Model 2 compared spectra from 31 cases (including four longitudinal spectra from one individual) and 17 village controls from the 2009 collection (Fig. 1c). Since household controls shared the same food supply as cases, and were therefore likely to have higher long-term PA exposure than village controls, we excluded household controls from the analysis in model 2 to improve contrast between classes. The loadings coefficient plot of model 2 (Fig. 1d) showed that the relative levels of bile acids, N-acetylglycoproteins and the methyl acceptor N-methylnicotinamide were found to be higher in samples from cases than in samples from village controls, while the concentration of p-cresol sulphate, hippurate and another microbial associated metabolite p-hydroxybenzoate were relatively lower in case samples (Table 2).
The effects of differential exposure levels among controls was explored in model 3, which compared 17 village to 18 household control samples collected in 2009 (Fig. 1e). The loading coefficient plot of model 3 (Fig. 1f) showed that the relative concentration of tyrosine and N-acetylglycoprotein (although a different glycoprotein resonance than that identified in model 2) was greater and the relative concentration of p-cresol sulphate and p-hydroxybenzoate was lower among household control samples (Table 2).
Similar trends in the relative concentration of the discriminatory metabolites bile acids, N-acetylglycoproteins (identified in model 2), hippurate, p-cresol sulphate, formate and N-methylnicotinamide were observed across the sample classes in both collections (Fig. 2). NMR spectra and loadings coefficient plots showed the presence of singlets from bile acid C-18 methyl groups in models 1 and 2, which were likely to arise from taurine-conjugated or unconjugated species.16 A number of metabolites that remain unidentified were also perturbed in each model (Table 2).
Due to the uneven gender distributions between cases and controls, a further O-PLS-DA model was constructed to assess the metabolome relationship with gender. Of the metabolites identified above, only p-cresol sulphate levels were found to be associated with gender with lower levels in samples from male subjects (loading coefficient r = −0.15) However the association of this metabolite with case samples was robust to adjustment for gender in multiple logistic regression models (see ESI†).
In the single AL dose study, 6/8 mice had plasma ALT levels above 1000 U L−1 (upper limit of assay) 24 hours post-dose. Hepatocellular necrosis of the zone 3 region, sometimes extending into zone 2, of the liver lobule was observed in 7/8 mice. A PLS-DA model (R2Y = 0.75, R2X = 0.33, Q2Y = 0.61, validation p = 0.001) was applied to distinguish urinary metabolic profiles from six pre-dose samples and five post-dose samples available in this study. The metabolites significantly contributing to the discriminatory component included amino acids creatine (loading r = 0.93), taurine (loading r = 0.84) and tyrosine (loading r = 0.72) which were raised 24 h post dose. The microbial associated metabolites hippurate (loading r = −0.80) and trimethylamine (loading r = −0.69) were decreased, while phenyacetylglycine (loading r = 0.79) and p-cresol glucuronide (loading r = 0.79) were elevated post dose.
In the AL and DDT co-dosing study, mean plasma ALT levels were 85 U L−1 (Standard Error (SE): 16) in the AL only dosed group and 470 U L−1 (SE: 205) in the AT + DDT co-dosed group. In the AL only dosed group histopathological damage was limited to swollen hepatocytes in the zone 1 region in one mouse. In the AL + DDT co-dosed group histopathological features included no observable injury (1/10 mice), mild vacuolation (2/10 mice), zone 3 congestion (1/10 mice), swollen hepatocytes in the zone 3 region (3/10 mice), extensive zone 3 necrosis (1/10 mice), extensive zone 2 and 3 necrosis (2/10 mice).
Since multivariate methods could not clearly distinguish samples from the two dosing groups in the DDT and AL co-treatment study (data not shown), the relative urinary concentrations of endogenous metabolites identified as being perturbed following acute AL dosing were compared between mice in the AL-only and DDT + AL dosed groups (ESI Fig. s4†). No significant differences were observed between the groups, although the differences in median concentrations were greater for metabolites associated with gut microfloral co-metabolism than for metabolites associated with liver injury. The largest difference was observed for hippurate, which had a lower median concentration among DDT + AL treated mice, but this was not statistically significant (p = 0.075). Despite similar median levels in both dosing groups, tyrosine was correlated with the extent of liver injury (Fig. 3) in both the AL only dosed group (r = 0.72, p = 0.02, with alanine transaminase (ALT) score) and the DDT + AL dosed group (r = 0.70, p = 0.03, with histopathology score). Interestingly, p-cresol glucuronide was positively correlated with liver injury in the AL only dosed group (r = 0.58, p = 0.09, with ALT score) but negatively correlated in the DDT + AL dosed group (r = −0.57, p = 0.09, with histopathology score) although this was not significant at the 5% level.
We sought to validate our results through similar analysis of urine samples collected from mice following an acute single dose of AL. Again metabolites falling into similar metabolic pathways as observed in the human comparisons; liver-associated (tyrosine), one-carbon metabolism (taurine, creatine,) or gut microbial (hippurate, p-cresol glucuronide, phenylacetylglycine and trimethylamine), were perturbed in the mice models. Furthermore, levels of specific metabolites, tyrosine and hippurate, were perturbed in both the human and mouse analyses reflecting similar metabolic responses to PA exposure. Since the population affected by HVLD were also highly exposed to DDT, we also tested the effect of prior dosing with DDT on the metabolic response to AL dosing. Despite greater hepatotoxicity in the DDT pre-dosed group, there were only small differences in metabolic response, suggesting a ‘saturation effect’ for markers such as taurine and creatine, with only little change past a certain level of liver damage. However, for tyrosine and p-cresol glucuronide, which were more closely related to the extent of hepatocellular damage, there was evidence for modulation of the metabolic response following DDT pre-dosing.
Fig. 4 summarises the novel panel of discriminatory metabolites identified across the human and mouse studies and relates them to the known effects of AL exposure. Tyrosine elevations were observed in both mouse and human models with the more acute hepatocellular injury. Hypertyrosemia is often observed following chronic liver diseases21 and it is normally accompanied by rises in other aromatic amino acids such as phenylalanine reflecting reduced hepatic metabolism. However, in this study no changes were observed in phenyalanine indicating the changes in tyrosine appear to be a toxin specific response rather than a result of general liver insufficiency. It is likely that AL interrupts the ability to re-synthesise the short lived enzyme tyrosine aminotransferase (TAT) involved in tyrosine catabolism.22 The dampening effect on the tyrosine response, considering the increased hepatocellular injury, when DDT was co-administered would support this interpretation since phenobarbital-type inducers are known to up-regulate TAT production.23
One-carbon metabolism intermediaries have been suggested as useful probes of liver specific processes.24 Rises in the methyl-acceptor N-methylnicotinamide, the one carbon hub metabolite formate, and also taurine, an end-product of sulphur metabolism, may reflect a fall in one-carbon cycling due to reduced cysteine demand following an overall drop in protein synthesis. When elevation of taurine is accompanied by a rise in creatine, as observed following acute AL dosing, this is thought to relate to active synthesis due to the cytoprotective properties of taurine.25
Increases in resonances from the N-acetyl group of the acute phase proteins α1-acid glycoproteins, were detected among both disease cases and highly exposed controls. These NMR spectral resonances are becoming established as indicators of systemic inflammation in many disease contexts.26 The rises observed in bile acids appear to result from specific molecular forms rather than resulting from general cholestasis, which was a minor component of HVLD. Specific changes to bile acid profiles have been detected by targeted UPLC-MS and have been associated with specific forms of liver injury,27 and may reflect the observed bile ductular reaction in cases or host – gut microfloral interaction.
Lower levels of gut-microbial associated metabolites, including hippurate and p-cresol sulphate, were observed in case samples. Reductions in p-cresol sulphate have not been reported in studies of liver disease in humans previously, indicating the importance of unique microbial profiles in HVLD. Hippurate is produced by the conversion of dietary aromatic compounds by bacteria such as Clostridia species to benzoic acid, followed by glycine conjugation in the kidney or liver.28 The mouse model supports that, for hippurate, lower levels among HVLD cases are a result of toxin exposure. However, rises observed in AL-only dosed mice of phenylacetylglycine and p-cresol glucuronide, the rodent conjugated form of p-cresol, suggests that some microbial species in mice are resistant to AL exposure or that conjugation capacity is altered by AL exposure. Modulation of host xenobiotic metabolic machinery by the microbiome is now well established either through, for example, direct enzyme induction29 or competition for conjugation capacity.30 It is conceivable therefore, that the higher levels of p-cresol sulphate observed in controls compared to cases in the human studies may, in part, precede and modulate the effects of AL exposure. The trend for reduced levels of p-cresol glucuronide with increased hepatotoxicity in mice co-dosed with both DDT and AL suggests additional complexity in the toxin-host-microbiome relationship.
Limitations of this study were a lack of accompanying traditional clinical chemistry measures in the patient samples to assess the severity of disease within classes and strengthen the disease classification. While the majority of cases in the first collection were confirmed by liver function test (i.e. raised serum γ-glutamyl transpeptidase activity8) or biopsy, the confirmation by liver function test in the second collection was not logistically possible. This may have led to misclassification of some cases (and controls) in the analysis of samples from the second collection, although the same specific case definition was used in both collections. Urinary AL measurements only reflect recent exposure, and the use of other markers such as pyrrole-protein adducts would have proved an assessment of long-term, although non-specific, PA exposure. It was also not feasible to completely age and gender match cases and controls, although potential confounding has been adjusted for. Specific information on diet was not obtained, and diet is known to affect urinary metabolic profiles. However, the relative homogeneity of the diet of the participants, relative to western populations, negates the importance of this to some extent. Furthermore, our sample size was relatively small. A larger sample size may have allowed the use of other statistical approaches such as multiple univariate testing with correction for multiple testing. Here we have used a single multivariate modelling approach, partially to increase statistical power and avoid the multiple testing burden. While the study was sufficiently powered to detect differential metabolic phenotypes, future studies may consider increasing sample size to detect additional associations. The NMR platform used has limited sensitivity, and may therefore have been unable to detect metabolites of very low concentration that had aetiological or diagnostic relevance. However, the reproducibility of the platform and the structural information it provides makes it particularly suitable for ‘untargeted’ analyses. Finally, the cross-sectional nature of the human study provided only a ‘snapshot’ without information regarding the temporal sequence of metabolic changes and clinical disease onset. However, this was addressed to some extent by the analysis of differentially exposed healthy controls and the integration of mouse models, which provided both validation and further information on the causal direction of detected associations. Future work may include adopting a ‘One Health approach’ and performing similar analyses in livestock of the affected villages, that were similarly affected by a parallel outbreak of liver disease.
Footnote |
† Electronic supplementary information (ESI) available. See DOI: 10.1039/c6tx00221h |
This journal is © The Royal Society of Chemistry 2016 |