Discovery of Hepatotoxic Equivalent Combinatorial Markers from Dioscorea bulbifera tuber by Fingerprint-Toxicity Relationship Modeling

Due to extremely chemical complexity, identification of potential toxicity-related constituents from an herbal medicine (HM) still remains challenging. Traditional toxicity-guided separation procedure suffers from time- and labor-consumption and neglects the additive effect of multi-components. In this study, we proposed a screening strategy called “hepatotoxic equivalent combinatorial markers (HECMs)” for a hepatotoxic HM, Dioscorea bulbifera tuber (DBT). Firstly, the chemical constituents in DBT extract were globally characterized. Secondly, the fingerprints of DBT extracts were established and their in vivo hepatotoxicities were tested. Thirdly, three chemometric tools including partial least squares regression (PLSR), back propagation-artificial neural network (BP-ANN) and cluster analysis were applied to model the fingerprint-hepatotoxicity relationship and to screen hepatotoxicity-related markers. Finally, the chemical combination of markers was subjected to hepatotoxic equivalence evaluation. A total of 40 compounds were detected or tentatively characterized. Two diterpenoid lactones, 8-epidiosbulbin E acetate (EEA) and diosbulbin B (DIOB), were discovered as the most hepatotoxicity-related markers. The chemical combination of EEA and DIOB, reflecting the whole hepatotoxicity of original DBT extract with considerable confidential interval, was verified as HECMs for DBT. The present study is expected not only to efficiently discover hepatotoxicity-related markers of HMs, but also to rationally evaluate/predict the hepatotoxicity of HMs.

The similarity results of 21 batches of DBT samples are presented in Table S1, and the similarity values fell within the range of 0.919-0.996. The results of common peak areas are listed in Table S2. As is evidenced from chemical profile data and similarity values among DBT extracts, it is difficult to discriminate the randomly selected samples only by fingerprint analysis.
Hepatotoxicity of DBT extracts. The levels of serum ALT and AST activities are used as biomarkers for the early acute liver injury, and obvious elevation of these enzymes reflects the damage of hepatic function 33,34 . To assess hepatotoxicity of DBT extract, serum ALT and AST in mice treated with DBT extracts were detected. It was noted that a dose of 2 g/kg DBT extracts resulted in serum ALT/AST levels significantly increased (p < 0.01), comparing with 32 ± 6 U/L (ALT) and 83 ± 15 U/L (AST) in mice treated with vehicle ( Fig. 4(A,B)). In addition, severe liver damage appeared as hepatic cell necrosis, inflammatory cell infiltration and local spotty necrosis by histopathologic analysis (Fig. 4(C)). These findings indicated the mice administered DBT extracts at 2 g/kg caused potential hepatotoxicity.

Discovery of principal hepatotoxic markers by PLSR, BP-ANN and cluster analysis.
In the present study, we tried to establish the fingerprint-hepatotoxicity relationship model, and three chemometric methods, including PLSR, BP-ANN and cluster analysis were applied to screen the possible hepatotoxic markers from DBT extracts.
PLSR is a particular type of multivariate analysis using the two-block predictive PLS model to find the relationship between two matrices, X and Y 35 . Here, a PLSR model to correlate hepatotoxicity and chromatographic data was constructed with the 21 batches of DBT extracts. The parameter R-squared and adjusted R-squared of the model were 0.902 and 0.576 respectively, which indicated that PLSR was appropriate in modeling fingerprint-hepatotoxicity correlation. The importance of the X-variables for a model could be evaluated by variable importance for the projection (VIP) values (usually with a threshold >1.0). The VIP values of ten peaks are given in Table 2. As observed, the VIP values of peaks 9 and 10 were higher than 1.0 (1.025 and 1.798, respectively), indicating that DIOB and EEA might induce liver injury.
Furthermore, a nonlinear BP-ANN mathematical model was applied to clarify the fingerprint-hepatotoxicity relationship. The model contained three layers and three knots, and the iteration times were 5000. All data sets were from the 21 batches of DBT extracts samples. Mean impact value (MIV) is one of the best indexes reflecting changes of the weights matrix and evaluating the correlation of variables in the neural network 36,37 . In general, the MIV can be described as: where n represents the number of observations. After training on set data, MIVs of the ten common peaks are list in Table 2. The mean squared error (MSE) of training, the magnitude of the gradient and the correlation Figure 2. The typical total ion chromatogram of DBT extract by UHPLC-QTOF MS analysis in negative ion mode (S01). The peak numbers were in accordance with the compound numbers in Table 1. *HEMCs.
coefficient (R) are depicted in Fig. 5. In line with PLSR modeling results, compounds 9 and 10 had highest coefficient values (0.255 and 0.342, respectively) among all common peaks. Finally, heat map was performed to obtain more penetrating understanding of this relationship. The relative concentration trend between the common peaks and serum ALT/AST levels in all test samples is illustrated in Fig. 6. The cluster analyses also suggested that peaks 9 and 10 were more responsible for the hepatotoxicity.
Summing up the results from the extended PLSR, BP-ANN and cluster analysis, DIOB and EEA were finally discovered as the main hepatotoxic compounds in DBT extract. Accordingly, these two compounds were tentatively assigned as candidate HECMs accounting for the whole hepatotoxicity of original DBT extracts. Assessment of hepatotoxic equivalence between candidate HECMs and original DBT extracts. The contents of DIOB and EEA in DBT extracts of S22-S24 are summarized in Table 3, and the hepatotoxicities of original DBT extracts and candidate HECMs are shown in Fig. 7 In the hepatotoxic equivalence assessment, both original DBT extracts and candidate HECMs caused significant elevations of serum ALT and AST (p < 0.01) (Fig. 7(A)); histopathologic analysis (n = 5 per group) revealed local spotty necrosis and inflammatory cell infiltration in the liver of mice given original DBT extracts and candidate HECMs (Fig. 7(B)). The 90% confidence interval (CI) for the hepatotoxicity of candidate HECMs in ALT and AST tests were 99.9-125.6%, 83.4-143.3%, 83.7-121.4% and 93.0-136.9%, 87.4-135.6%, 101.5-117.7%, respectively.

Discussion
Chemical profiling of HMs is always the prerequisite task for discovery of bioactive compounds 38,39 . Flavonoids and diterpenoid lactones are reported as two main types of ingredients in DBT. Flavonoids are phenolic substances abundantly presented in the plant kingdom, numerous studies have demonstrated their health-promoting properties 40 . In contrast, diterpenoid lactones in DBT are found to be hepatotoxic. Actually, the parent diterpenoid lactones do not appear to be hepatotoxic, while the metabolic activation of the furan ring by cytochromes P450 (CYP450) is the key procedure of acute liver injury 25,[41][42][43] . Considering low CYP expression in cell lines 44 , the in vivo animal experiments were therefore conducted to evaluate the potential hepatotoxicity of the DBT extracts.
Similarity analysis of chromatographic fingerprint is well-recognized as a useful measure to evaluate the batch-to-batch chemical consistency of HMs [45][46][47][48] . The chemical similarities among 21 batches of DBT extracts were measured by the fingerprint analysis. It was noted that all the similarity values were higher than 0.90, which suggested very little chemical fluctuation existed among those samples. Conversely, the hepatotoxic effects of DBT extracts on mice highly varied. The inconsistence between chemical composition and hepatotoxicity indicated that it is urgent to establish appropriate strategy to screen hepatotoxic markers in DBT extracts. In this study, three chemometric methods including PLSR, BP-ANN and cluster analysis were applied to discover the hepatotoxic markers from DBT extracts. Compared with the traditional toxicity-guided isolation process, the fingerprint-hepatotoxicity modeling approach showed its distinct advantages in efficiency, cost and general compatibility with holistic mode of HMs. The two principal discriminatory compounds, EEA and DIOB, were discovered as the hepatotoxic markers of DBT. Consequently, the combination of these two compounds was assigned as the candidate HECMs accounting for the whole hepatotoxicity of original DBT extracts. As shown in Fig. 4(A), the sum of DIOB and EEA correlated well with hepatotoxicity. In samples S01, 02, 04, 07, 08, 10, 11, 16 and 21, low amount of DIOB and EEA associated with low serum ALT/AST levels. While in samples S06, 09, 13, 14,  15, 17, 18 and 20, a similar trend was observed, viz. relatively higher contents of DIOB and EEA corresponded to more potent toxicity (Fig. 4(B)). The content-toxicity correlation implied that the combination of DIOB and EEA might be the candidate HECMs for the hepatotoxicity evaluation of DBT extract. Additionally, as listed in Table 2, the relatively higher VIP value and MIV of EEA indicated it was more potent than DIOB (1.798 vs 1.025, and 0.342 vs 0.255, respectively), which was consistent with a previous report 25 . Thus, compared with DIOB, EEA might play a major role in the DBT-induced liver injury.
Seen from Fig. 7(A), it was found that the mixtures of DIOB and EEA at identical amounts with those in S22-S24 showed nearly equivalent hepatotoxicities with corresponding DBT extracts, since all the 90% CI fell within the range of 70-143% 49 , demonstrating that candidate HECMs hepatotoxically equaled to original DBT extracts. Therefore, the chemical combination of EEA and DIOB was confirmed as the real HECMs for DBT.
In conclusion, a strategy based on the fingerprint-toxicity relationship modeling and hepatotoxic equivalence assessment was initially proposed for discovery and verification of HECMs from DBT. The chemical constituents in DBT extract was characterized by UHPLC-QTOF MS and a total of 40 compounds were identified or tentatively assigned. Based on the fingerprint-toxicity relationship modeling, EEA and DIOB were discovered as the main hepatotoxic markers. Furthermore, the chemical combination of those two markers was confirmed as HECMs which could account for the whole hepatotoxicity of original DBT extracts with 90% CI. Using DBT as a case study, this work provides not only a promising strategy for efficient discovery of potential hepatotoxic constituents from HMs, but also a rational terminology-HECMs for hepatotoxicity evaluation/ prediction of HMs.   (2), succinic acid (4), gallic acid (6), epigallocatechin (7), protocatechuic acid (9)  Hepatotoxicity of DBT extracts. Mice were orally administered DBT extracts (2 g/kg, suspended in 0.5% CMC-Na, n = 10) for 36 h, and 0.5% sodium carboxymethyl cellulose (CMC-Na) was used as a vehicle control (n = 10) 25,50 . They were fasted from food, but no water 12 h prior to the administration of the test suspension. Blood was collected from the eyeball for measurement of ALT, AST. Serum ALT and AST activities were measured on Cobas 8000 modular analyzer (Basel, Switzerland). Liver tissues were fixed in 10% neutral buffered formalin, paraffin processed, and sectioned at 3 μm. For histological evaluation, the tissue sections were stained with hematoxylin and eosin (H&E).   Establishing of UHPLC fingerprint. All DBT samples were chemically profiled under the above mentioned chromatographic and mass spectrometric conditions. The fingerprints of 21 batches of samples (S01-S21) were matched automatically using Similarity Evaluation System for Chromatographic Fingerprint of Traditional Chinese Medicine (2012 Version, Committee of Chinese Pharmacopeia). The simulative mean chromatogram as a representative standard for those fingerprints was calculated and generated automatically by median method. Based on careful comparison of UV, MS/MS spectra and relative retention time, peaks detected in all fingerprints were defined as "common characteristic peaks" and structurally elucidated. Similarity values between each two chromatographic fingerprints were then determined using the above mentioned officially recommended software.
Quantitative analysis of the major diterpenoid lactones. The contents of DIOB and EEA were quantified by means of the external standard method. Prior to quantification, the developed UHPLC method was fully validated in terms of specificity, linearity, limit of detection (LOD), limit of quantitation (LOQ), precision (i.e. repeatability, intra-day and inter-day variability), stability, repeatability and accuracy.
Statistical analysis. Results were expressed as the mean ± standard deviation (SD) for continuous variables and as the number (percent) for categorical variables. To maximize identification the fingerprint-toxicity relationship between groups, PLSR model was applied using SIMCA version 14.0.1 (Umetrics AB, Umea, Sweden). In addition to the multivariate statistical method, the BP-ANN model was also employed to correlate fingerprints with hepatotoxicity using Matlab R2016a (Mathworks, Natick, USA). Heat maps and hierarchical cluster analyses were conducted using MeV version 4.6.0. Statistical analyses were performed using SPSS software version 19.0 (IBM Corp., Armonk, USA). An adjusted p value < 0.05 was considered statistically significant. Hepatotoxicity evaluation between candidate HECMs and original DBT extracts. Hepatotoxic equivalence was evaluated by calculating 90% CI of the ratio between the toxicities of candidate HECMs and original DBT extracts (two one-sided t test), the equivalent relationship of the 90% CI was calculated by the following equation: Where Y B and Y H are the least squares means of the candidate HECMs and original DBT extracts treatment, σ W 2 is the mean square, from Analysis of Variance (ANOVA) after logarithmic transformation, and . + − t n n 0 95, 2 1 2 is the 0.95 quantile of the central t-distribution with n 1 + n 2 -2 degrees of freedom 26 . If the 90% CI of relative hepatotoxicity compared to original DBT fell within the range of 70-143% 49 , the candidate HECMs were considered to be hepatotoxic equivalent with original DBT extracts. Three batches of additional DBT samples from Sichuan (S22), Anhui (S23) and Hubei (S24) were chosen to evaluate the hepatotoxic equivalence between candidate HECMs and original DBT extracts. The mixtures of DIOB and EEA were dissolved in 0.5% CMC-Na solution at a dose equivalent to that of the two tested compounds found in the DBT extracts. The mice were sacrificed 36 h after the administration, and the serum ALT and AST activities were used to assess the hepatotoxicity. Meanwhile, the histological samples were also examined.