Yubo
Li
a,
Haoyue
Deng
a,
Liang
Ju
a,
Xiuxiu
Zhang
a,
Zhenzhu
Zhang
a,
Zhen
Yang
a,
Lei
Wang
a,
Zhiguo
Hou
a and
Yanjun
Zhang
*b
aTianjin State Key Laboratory of Modern Chinese Medicine, School of Traditional Chinese Materia Medica, Tianjin University of Traditional Chinese Medicine, 312 Anshan west Road, Tianjin 300193, China
bTianjin State Key Laboratory of Modern Chinese Medicine, Tianjin University of Traditional Chinese Medicine, 312 Anshan west Road, Tianjin 300193, China. E-mail: tianjin_tcm001@sina.com; Fax: +86-22-59596221; Tel: +86-22-59596221
First published on 5th November 2015
Currently, drug-induced nephrotoxicity is widespread and seriously affects human health. However, the conventional indexes of renal function lack sensitivity, leading to a delay in the detection of nephrotoxicity. Therefore, we need to identify more sensitive indexes for evaluating nephrotoxicity. In this study, we used gentamicin (100 mg kg−1), etimicin (100 mg kg−1) and amphotericin B (4 mg kg−1) to establish renal injury models in rats, and we collected information using ultra-performance liquid chromatography quadrupole time-of-flight mass spectrometry in the screening stage. Thirteen nephrotoxicity metabolites were selected after multivariate statistical and integration analyses. Then, we conducted trend analysis to select 5 nephrotoxicity biomarkers [thymidine, LysoPC(16:1), LysoPC(18:4), LysoPC(20:5), and LysoPC(22:5)] whose content changed consistently at different timepoints after drug administration. To verify the sensitivity and specificity of these biomarkers for nephrotoxicity, receiver operating characteristic (ROC) and support vector machine (SVM) analyses were applied. The area under the curve of the 5 biomarkers were 0.806–0.901 at the 95% confidence interval according to the ROC analysis. We used the SVM classified model to verify these biomarkers, and the prediction rate was 95.83%. Therefore, the 5 biomarkers have strong sensitivity and high accuracy; these biomarkers are more sensitive indexes for evaluating renal function to identify nephrotoxicity and initiate prompt treatment.
Metabolomics, which is an important part of systems biology, is used to investigate changes in endogenous substances when the biological system is affected by external disturbances.6–8 With research developments, metabolomics technology has been used extensively to evaluate drug toxicity. Particularly, it has promoted the study of drug-induced nephrotoxicity to gain new insights into the associated pathophysiological mechanisms.9–11 Plasma metabolomics is broadly used in human health care and drug safety evaluations because it provides a large amount of information on endogenous substances.12,13 Given its high sensitivity, extensive dynamic range and good separation ability, ultra-performance liquid chromatography-mass spectrometry (UPLC-MS) has become one of the most versatile techniques, and is being gradually applied to various fields such as metabolomics, proteomics and traditional Chinese medicine. UPLC-MS-based metabolomics has great potential for identifying useful biomarkers for disease diagnosis (such as hepatocarcinoma and liver cirrhosis, lung cancer and pneumonia, and Alzheimer's disease and schizophrenia, etc.) and drug-induced toxicity assessment (such as cardiotoxicity, hepatotoxicity, nephrotoxicity, etc.).14–17 Metabolomics biomarkers can reveal the metabolic differences in the physiological and pathological states of organisms in a dynamic and sensitive manner.18–20
The support vector machine (SVM) is an intelligent pattern recognition technology that has been extensively used in different fields.21–23 It effectively solves the binary classification problem because it generates the optimal linear interface of two categories of substances.24 SVM provides a new direction in metabolomics and genomics data processing because of its robustness, and deals well with high-dimensional data and small sample sizes.25,26 Therefore, we utilized SVM to predict and classify the related biomarkers by feature selection and classification prediction.
In this study, we used receiver operating characteristic (ROC) and SVM to analyse plasma metabolomics data to identify biomarkers for evaluating nephrotoxicity. We used gentamicin, etimicin and amphotericin B to establish rat models of renal injury. Information on the plasma samples was collected using an UPLC quadrupole time-of-flight MS (UPLC-Q-TOF/MS) platform. After the multivariate statistical analysis, integration analysis and content analysis, we obtained nephrotoxicity biomarkers whose content changed consistently at different timepoints after drug administration. Next, we used ROC to evaluate the sensitivity and specificity of the nephrotoxicity biomarkers. Then, we predicted nephrotoxicity using these biomarkers after combining with cardiotoxicity and hepatotoxicity data by SVM. The method can provide a systematic tool for screening and validating other toxic biomarkers using metabolomics and can promote the development of metabolomics.
105 rats were divided into ten groups to identify nephrotoxicity biomarkers: the NS, GM-1d, GM-3d, GM-7d, ETI-1d, ETI-2d, ETI-3d, AMB-1d, AMB-3d and AMB-7d groups. 70 rats were divided into seven groups to verify the nephrotoxicity biomarkers: the NS group, two nephrotoxicity groups (TAA and DDP), two cardiotoxicity groups (CP and 5FU), and two hepatotoxicity groups (CCl4 and TC). The groups, doses, administration routes and sampling times are shown in Table 1.36,45
Drug | Grouping | Number | Dose | Mode of administration | Sampling time | |
---|---|---|---|---|---|---|
a The screening stage for nephrotoxicity biomarkers. b The validation stage for nephrotoxicity biomarkers. c Intraperitoneal injection. d Intragastric administration. e Subcutaneous injection. | ||||||
Stage Ia | NS | NS | 15 | 5 ml kg−1 | i.p.c, single-dose | 1 day |
GM | GM-1d | 10 | 100 mg kg−1 | i.p.c, single-dose | 1 day | |
GM-3d | 10 | 100 mg kg−1 | i.p.c, successive administration | 3 days | ||
GM-7d | 10 | 100 mg kg−1 | i.p.c, successive administration | 7 days | ||
ETI | ETI-1d | 10 | 100 mg kg−1 | i.p.c, single-dose | 1 day | |
ETI-2d | 10 | 100 mg kg−1 | i.p.c, successive administration | 2 days | ||
ETI-3d | 10 | 100 mg kg−1 | i.p.c, successive administration | 3 days | ||
AMB | AMB-1d | 10 | 4 mg kg−1 | i.p.c, single-dose | 1 day | |
AMB-3d | 10 | 4 mg kg−1 | i.p.c, successive administration | 3 days | ||
AMB-7d | 10 | 4 mg kg−1 | i.p.c, successive administration | 7 days | ||
Stage IIb | NS | NS | 10 | 5 ml kg−1 | i.p.c, single-dose | 1 day |
TAA | TAA | 10 | 200 mg kg−1 | i.p.c, successive administration | 6 days | |
DDP | DDP | 10 | 6 mg kg−1 | i.p.c, successive administration | 3 days | |
CP | CP | 10 | 200 mg kg−1 | i.p.c,successive administration | 5 days | |
5FU | 5FU | 10 | 125 mg kg−1 | i.g.d, single-dose | 1 day | |
CCl4 | CCl4 | 10 | 5 mL kg−1 | i.s.e, successive administration | 2 days | |
TC | TC | 10 | 1500 mg kg−1 | i.g.d, successive administration | 5 days |
For H&E staining, the fixed tissues were embedded in paraffin wax. Then, 5 μm thick slices were cut and fixed on glass slides. The slices were deparaffinized with xylene, hydrated, stained with haematoxylin, differentiated with hydrochloric alcohol, stained with eosin and dehydrated in a graded alcohol series. Then, the slides were cleaned with xylene, and the histopathological changes were observed by light microscopy at 100× magnification.36,37
In the screening stage, the data was processed by multivariate statistical analysis using SIMCA-P+11.5 software (Umetrics, Sweden). In our study, principal component analysis (PCA) was used to identify the outliers in the samples, and partial least squares-discriminate analysis (PLS-DA) was used to distinguish the variables with a high contribution between the NS group and the drug treatment group at different times. The model was visualized with a score plot. The variables with a variable-importance plot (VIP) greater than 1 (VIP > 1) at different administration times were analysed using Student's t-test in SPSS 17.0 (SPSS, USA), and the variables with p < 0.05 represented potential nephrotoxicity metabolites. The potential nephrotoxicity metabolites from the three drugs at different times were processed by integration analysis to identify nephrotoxicity metabolites using Venn diagrams (http://bioinfogp.cnb.csic.es/tools/venny/index.html). The heat map was generated using Cluster software based on the relative content of each nephrotoxicity metabolite in each treatment group. Next, the change in content of nephrotoxicity metabolites at different timepoints was analysed to identify those metabolites whose content changed consistently. These metabolites were identified by MS/MS information and confirmed with HMDB (http://www.hmdb.ca/) and KEGG (http://www.genome.jp/kegg/) as nephrotoxicity biomarkers. The ROC curves of nephrotoxicity biomarkers based on the nephrotoxic drug groups were determined using the binary logistic regression model in SPSS 17.0 (SPSS, USA).
Then, we combined the data of nephrotoxic drug groups with non-nephrotoxicity (cardiotoxicity and hepatotoxicity) data to validate the nephrotoxicity biomarkers using an SVM model in MATLAB R2010a (MathWorks, USA). The peak areas of the nephrotoxicity biomarkers were the input variables, and the training set was used to build an SVM classification model with the optimal penalty parameter (c) and kernel function (g). The factor c is used to determine the characteristics of subspace-regulated learning, and g is a function for mapping from low-dimensional space to high-dimensional space.36,39 We obtained the accuracy rate of the model using the test set. Cross-validation was used to determine the confidence and experience risk ratio ranges of the model.36,40
Kidney damage is indicated by increased Scr and BUN contents. In our study, the levels of Scr and BUN in the drug-treated groups were compared with those in the NS group by Student's t-test (Fig. 2). The Scr and BUN levels were significantly increased (p < 0.05) only in the GM-7d, ETI-3d, and AMB-7d groups compared with the NS group. In the other groups, the two indexes were not increased significantly at the same timepoints. However, the content of the two indexes showed a temporal correlation.
Fig. 2 BUN and Scr levels in serum samples. (A) Changes in BUN levels. (B) Changes in Scr levels. Data are presented as the mean ± SD (*p < 0.05, **p < 0.01, compared with the NS group). |
We obtained the PCA and PLS-DA score plots using multivariate statistical analysis (Fig. S3 and Table S2†). Some stray samples were removed according to the PCA. We selected variables with VIP > 1 based on the PLS-DA for analysis by Student's t-test. The variables with p < 0.05 represented potential metabolites associated with each drug at different times. Then, they were processed by integration analysis to identify potential nephrotoxicity-associated metabolites of each drug (Fig. S4†). Finally, we obtained 13 nephrotoxicity metabolites, and the Venn diagram is shown in Fig. 3. The heat map of the relative content of each nephrotoxicity-associated metabolite is shown in Fig. 4. We retained 5 biomarkers whose content changed consistently in the GM, ETI and AMB groups at different times (Fig. 5). These biomarkers were identified by mass spectrometry (Fig. S5†). Detailed information regarding the 5 nephrotoxicity biomarkers is provided in Table 2.
Fig. 4 Heat map of the relative content of each nephrotoxicity-associated metabolite in all the drug-treated groups. |
No. | t R (min) | Metabolites | Obsd [M + H]+ | Calcd [M + H]+ | Error (ppm) | Formula | MS/MS |
---|---|---|---|---|---|---|---|
1 | 2.51 | Thymidine | 243.0987 | 243.0975 | 4.80 | C10H14N2O5 | 243.1 [M + H]+ |
127.1 [M + H − C5H8O3]+ | |||||||
2 | 5.16 | LysoPC(16:1) | 494.3244 | 494.3241 | 0.61 | C24H48NO7P | 494.3 [M + H]+ |
476.3 [M + H − H2O]+ | |||||||
184.0 [M + H − C19H34O3]+ | |||||||
125.0 [M + H − C22H43NO3]+ | |||||||
104.1 [M + H − C19H35O6P]+ | |||||||
3 | 5.15 | LysoPC(18:4) | 516.3062 | 516.3085 | −4.45 | C26H46NO7P | 516.3 [M + H]+ |
498.3 [M + H − H2O]+ | |||||||
184.0 [M + H − C21H32O3]+ | |||||||
104.1 [M + H − C21H33O6P]+ | |||||||
4 | 5.55 | LysoPC(20:5) | 542.3219 | 542.3241 | −4.06 | C28H48NO7P | 542.3 [M + H]+ |
524.3 [M + H − H2O]+ | |||||||
259.1 [M + H − C17H33NO2]+ | |||||||
184.0 [M + H − C23H34O3]+ | |||||||
125.0 [M + H − C26H43NO3]+ | |||||||
104.1 [M + H − C23H35O6P]+ | |||||||
5 | 6.13 | LysoPC(22:5) | 570.3549 | 570.3554 | −0.88 | C30H52NO7P | 570.4 [M + H]+ |
552.3 [M + H − H2O]+ | |||||||
184.0 [M + H − C25H38O3]+ | |||||||
125.0 [M + H–C28H47NO3]+ | |||||||
104.1 [M + H–C25H38O6P]+ |
We used ROC to evaluate the diagnostic potential of the 5 biomarkers for nephrotoxicity in the screening stage. The ROC analysis showed that the 5 nephrotoxicity biomarkers had a high accuracy for evaluating nephrotoxicity based on the area under the curve, and the sensitivity and specificity at the best cutoff points (Table 3 and Fig. 6).
Fig. 6 ROC curves of the 5 nephrotoxicity biomarkers [thymidine, LysoPC(16:1), LysoPC(18:4), LysoPC(20:5) and LysoPC(22:5)] at the screening stage. |
Biomarkers | AUCa | Sensitivity (%) | Specificity (%) | Standard errorb | 95% CI | |
---|---|---|---|---|---|---|
Lower | Upper | |||||
a The area under the curve. b Under the nonparametric assumption. | ||||||
Thymidine | 0.901 | 0.956 | 0.714 | 0.044 | 0.815 | 0.988 |
LysoPC(16:1) | 0.817 | 0.813 | 0.786 | 0.069 | 0.682 | 0.952 |
LysoPC(18:4) | 0.806 | 0.846 | 0.714 | 0.064 | 0.680 | 0.933 |
LysoPC(20:5) | 0.886 | 0.791 | 0.857 | 0.033 | 0.821 | 0.951 |
LysoPC(22:5) | 0.830 | 0.758 | 0.786 | 0.062 | 0.710 | 0.951 |
Fig. 7 3D view of the SVM classified model of the 5 nephrotoxicity biomarkers (parameters: best c = 0.76, best g = 6.96, CV accuracy = 100%). |
To a certain extent, biochemical indicators reflect organ damage. However, their sensitivity and specificity are poor because they are often affected by other factors. However, histopathological analysis can reveal organ damage directly. When Scr and BUN levels are significantly increased, it is likely that the kidney has been injured by the pathological condition. Hence, we used histopathological examinations to evaluate the extent of kidney damage. From the serum biochemistry results, we ascertained that the kidneys were injured only in the GM-7d, ETI-3d, and AMB-7d groups. However, the histopathological examination revealed that the kidneys were injured in all the drug-treated groups. This finding indicates that the existing methods do not detect nephrotoxicity with adequate accuracy or sensitivity. Compared with serum biochemistry, nephrotoxicity biomarkers underwent significant changes at different timepoints after drug administration, which reveals sensitive metabolic differences in organisms. Additionally, these biomarkers can help explain the biological mechanism of drug-induced nephrotoxicity.
Currently, some studies have used single nephrotoxic drug to identify nephrotoxicity biomarkers by metabolomics. Similar metabolic processes in the body may be affected by different toxic drugs.43 Therefore, the biomarkers in the reported studies were not exclusive to nephrotoxicity. Additionally, the application of these biomarkers has not been promoted.44,45 In our study, we established nephrotoxicity models based on three nephrotoxic drugs at three different administration times to identify nephrotoxicity biomarkers. To exclude the impact of other forms of toxicity, we combined the nephrotoxicity with other toxic drugs (cardiotoxicity and hepatotoxicity) to validate our biomarkers using an SVM model. The SVM model was used to verify the accuracy of the biomarkers in predicting nephrotoxicity. The results of the SVM model showed that the best combination of biomarkers [thymidine, LysoPC(16:1), LysoPC(18:4), LysoPC(20:5), and LysoPC(22:5)] had higher sensitivity and accuracy compared with the reported studies.44,45 Therefore, this combination has potential for broad application in drug safety evaluations and drug development as well as in the clinical evaluation and prediction of drug-induced nephrotoxicity.
In our study, we established a comprehensive and systematic method for identifying nephrotoxicity biomarkers. This represents a new tool for discovering and verifying biomarkers in other areas related to metabolomics, such as drug-induced toxicity, clinical diagnostics and plant metabolomics. Furthermore, it is conducive to the development of metabolomics. To obtain specific and exclusive nephrotoxicity biomarkers, we controlled the experimental animals (male Wistar rats, 6 weeks old, weighing 200 ± 20 g) to investigate the differences in metabolism in response to different toxic drugs in this study. Considering the universal applicability of our nephrotoxicity biomarkers, we should combine them with the factors such as gender and age for verification purposes in future studies.
Footnote |
† Electronic supplementary information (ESI) available: Histopathological examination results (Fig. S1). BPI chromatograms of QC samples in the positive ion mode of UPLC-Q-TOF/MS (Fig. S2). PCA and PLS-DA score plots (Fig. S3). Potential metabolites of each drug at different times were processed by integration analysis using Venn diagrams (Fig. S4). The substance identification with mass spectrometry information (Fig. S5). The results of methodology experiments (Table S1). The parameters of PLS-DA score plots (Table S2). See DOI: 10.1039/c5tx00171d |
This journal is © The Royal Society of Chemistry 2016 |