Metabolomics Study of Stepwise Hepatocarcinogenesis From the Model Rats to Patients: Potential Biomarkers Effective for Small Hepatocellular Carcinoma Diagnosis*

The aim of this study is to find the potential biomarkers from the rat hepatocellular carcinoma (HCC) disease model by using a non-target metabolomics method, and test their usefulness in early human HCC diagnosis. The serum metabolic profiling of the diethylnitrosamine-induced rat HCC model, which presents a stepwise histopathological progression that is similar to human HCC, was performed using liquid chromatography-mass spectrometry. Multivariate data analysis methods were utilized to identify the potential biomarkers. Three metabolites, taurocholic acid, lysophosphoethanolamine 16:0, and lysophosphatidylcholine 22:5, were defined as “marker metabolites,” which can be used to distinguish the different stages of chemical hepatocarcinogenesis. These metabolites represented the abnormal metabolism during the progress of hepatocarcinogenesis, which could also be found in patients. To test their diagnosis potential 412 sera from 262 patients with HCC, 76 patients with cirrhosis and 74 patients with chronic hepatitis B were collected and studied, it was found that 3 marker metabolites were effective for the discrimination of small liver tumor (solitary nodules of less than 2 cm in diameter) patients, achieved a sensitivity of 80.5% and a specificity of 80.1%,which is better than those of α-fetoprotein (53 and 64%, respectively). Moreover, they were also effective for the discrimination of all HCCs and chronic liver disease patients, which could achieve a sensitivity of 87.5% and a specificity of 72.3%, better than those of α-fetoprotein (61.2 and 64%). These results indicate metabolomics method has the potential of finding biomarkers for the early diagnosis of HCC.

Hepatocellular carcinoma (HCC) 1 is a type of malignancy with a high mortality rate worldwide, especially in the East Asian countries (1). In China, chronic hepatitis B virus (HBV) infection and subsequent liver cirrhosis are major precancerous lesions in the majority of HCC cases. An early diagnosis of small HCCs in these precancerous cases may greatly improve the outcome of HCC treatment (2). The current screening methods for the high risk population, such as ultrasound or the serum surveillance of tumor markers (mainly ␣-fetoprotein (AFP)), although effective, are far from ideal (3,4). For example, a subset of patients with chronic hepatitis and/or cirrhosis exhibit modest elevations (10 -500 ng/ml) of serum AFP, which may lead to an incorrect diagnosis (5). Hence, new biomarkers for monitoring hepatocarcinogenesis would be of great clinical importance.
Emerging platforms in the biomedical arena provide a new methodology to identify novel biomarkers. Within the framework of systems biology, metabolomics focuses on the quantitative measurement of holistic endogenous metabolites and is increasingly used in clinical fields that focus on the pathophysiological and diagnostic study of diseases (6). Metabolic fingerprinting and metabolite biomarkers have been studied for use in the discrimination or diagnosis of carcinoma (7,8), diabetes mellitus (9) and inborn errors (10). Nontarget metabolomics approaches have also been used to search for new biomarkers and to explore the mechanism of carcinogenesis in hepatic diseases (11)(12)(13)(14)(15)(16)(17)(18)(19)(20). According to these stud-ies, several circulation metabolites have been reported to be predictors for HCC, such as hexanal, 1-octen-3-ol, octane (18), and a combination of multimetabolites (21). Soga et al. found that serum ␥-glutamyl dipeptides can be used as biomarkers for discriminating different liver diseases (22). In our previous metabolomics studies on HCC based on the metabolic profiling of serum and urine (14,19), we found that the levels of several metabolites, such as sphingosines in blood and carnitines in urine, were statistically significantly different between the patients and controls. These studies examined a range of metabolites that represented the metabolic deregulation of HCC patients and illustrated the ability of metabolomics to identify the potential biomarkers of HCC. However, no individual metabolites or their combination were evaluated for their value as biomarkers to monitor the hepatocarcinogenesis process and HCC early diagnosis.
The metabolome in the clinical samples fluctuates frequently depending on various genetic and environmental factors, and discovery of biomarkers using these clinical samples faces a great technical challenge because of the enormous complexity of the serum metabolome. In addition, it is not easy to obtain serial clinical samples of individual liver diseases that represent the stages of chronic hepatitis, cirrhosis, and HCC, respectively. In reverse, the different stages of liver disease development in the rat model of HCC can be well anticipated. It has been found that histological and genetic signatures of diethylnitrosamine (DEN)-induced hepatocarcinogenesis are similar to those of human HCC (23,24). This rat model has also been used to investigate the progression of cirrhosis to HCC (25)(26)(27). Therefore, in the present study, a metabolomics approach was applied to the DEN-induced rat model to study the metabolic features of hepatocarcinogenesis using liquid chromatography-mass spectrometry (LC-MS). The "marker metabolites," which were defined from the rat model, were applied to distinguish patients with chronic hepatitis, cirrhosis, or small HCC to evaluate their capacity for diagnosing small HCC.

EXPERIMENTAL PROCEDURES
Animal Experiments-The present study conformed to the Guide for the Care and Use of Laboratory Animals from the Second Military Medical University. The animal experiments were briefly described in our previous paper (28). A total of 80 male Sprague-Dawley (S.D.) rats were obtained from the Shanghai Experimental Animal Centre and were enrolled in the present study at the age of 42 days. The starting time for the animal experiment was defined as week 0. The model group included 52 rats, and 6 of these rats died of liver failure during the experiment. The control group consisted of 28 rats. The DEN administrations (70 mg/kg body weight) were given to the model rats using intraperitoneal injections once a week after 14 days of inhabitation. In addition, saline injections of equivalent volumes were administered to the control animals. Ten injections were administered to each animal between week 2 and week 11.
To verify the histological progression of HCC, 4 rats from the model group and 2 rats from the control group were sacrificed under pentobarbital anesthesia every 2 weeks from week 2 until all of the surviving animals (n ϭ 20) were sacrificed (20 weeks). The liver tissues were fixed in 10% buffered formalin and embedded in paraffin. The gross examinations and histopathological studies (hematoxylin-eosin staining) were used to monitor the progress of the carcinogenesis.
The sera were sequentially collected after 8 h of overnight fasting. This collection was conducted from week 6 to week 20 once every 2 weeks. The sera were stored at Ϫ80°C until analysis. The analyses were performed in a month after all samples were collected.
The detailed serum and tissue collection times are shown in Fig. 1.
Collection of the Human Sera-Human samples were obtained from National Liver Tissue Bank in the Second Military Medical University, Shanghai, China. All samples were collected and stored under the same standard operation protocol. Written informed consent was given by all participants. The present study was approved by the ethics committee of the Second Military Medical University, Shanghai, China. All of the serum specimens for the present study were collected with permission from 262 patients with HCC, 76 patients with cirrhosis, and 74 patients with hepatitis B. HCC was histopathologically diagnosed after the tumor excision. The age of HCC patients was 30 to 78 years including 39 female and 233 male cases. A total of 77 HCC patients were in early stages of HCC (Edmonson I, II) and 95 patients had a solitary nodule smaller than 2 cm. The diagnosis of chronic hepatitis B or cirrhosis was made using clinical, imaging, and laboratory evidence of hepatic decompensation or portal hypertension. The age and gender distributions of the participants in each group were matched, and the collecting time point and thereby the storage time of samples from different groups were similar. Detailed information about the participants is shown in Table I. The sera were stored at Ϫ80°C until analysis. Sera were thawed at room temperature before analysis. Acetonitrile (400 l) was added to the sera (100 l) to quench the enzyme activity and precipitate the proteins. After 10 min of centrifugation at 10,000 ϫ g, the supernatant (350 l) from each sample was stored in an autosampler vial.
Metabolic Profiling and the Target Metabolite Analysis-The rat serum metabolic profiling analysis was performed using a LC-MS method on a 1200 Rapid Resolution Liquid Chromatography (RRLC) system that was coupled to a 6510 quadrupole time-of-flight (Q TOF) MS (Agilent, Santa Clara, CA). The used column was ZORBAX TM SB-AQ (10 cm ϫ 2.1 mm 1.8 m, Agilent, USA) running with the buffer consisting of water containing 0.1% formic acid (v/v) (A) and acetonitrile (B). The detailed LC and MS methods were described in our previous papers (14,28) and the supplementary materials. The injection volume was 4 l and the flow rate of the LC system was 0.3 ml/min. The total running time was 35 min including 11 min rinsing with acetonitrile and column equilibration.
The human serum metabolic profilings were acquired using a Thermo Fisher Accela LC system that was coupled to an LTQ Orbitrap XL system. The LC column and the elution gradient for this system were the same as those in the Agilent 1200 system. A high resolution MS of LTQ Orbitrap was used as the detector. The MS capillary temperature was held at 300°C, and the spray voltage was 4.5 kV. The flow rate of the sheath gas was 35 arbitrary units, and the flow rate of the aux gas was 5 arbitrary units. The MS system was operated using the positive ion mode, and the mass range was set at 100 -1000 m/z. The resolution of the Orbitrap was set at 60000, with a scan rate of 1 s/spectrum. At the defined operational conditions, the mass errors of the three "marker metabolites" were less than 2 ppm.
To obtain the contents of the three metabolites, all of the metabolites in a total ion chromatogram were extracted and aligned using the SIEVE software (V1.2, Thermo Fisher). The frame settings were 0.5 min (retention time) and 0.005 m/z (mass error). Next, the total peak area from one sample was assigned to a constant of 100000; the peak areas of all of the metabolites were normalized to this total area. After the confirmation of the three target metabolites, their normalized peak areas were exported to an Excel table (Microsoft, Redmond, WA) for the statistical analysis.
To ensure the stability and repeatability of the LC-MS systems, pooled quality control (QC) samples were used as in the literature (29 -31). The QC samples were prepared from 10 l of each sample and analyzed together with the other samples. A total of five runs of the QC samples were performed on the system before the sample sequence. The QC samples were also inserted and analyzed in every 10 samples.
Data Analysis-SIMCA-P 11.0 (Umetrics AB, Umeå, Sweden) was used for the chemometrics analysis. The multivariate pattern recognition technique of the partial least squares discriminant analysis (PLS-DA) with the pareto scaling was performed. The parameters of PLS-DA (R 2 Y, Q 2 Y) were used for the evaluation of the models. R 2 Y and Q 2 Y represent the goodness of the fit and the prediction ability of the models (32). Response permutation test was used to assess whether the model established exhibited overfitting because of the chance correlations. SPSS 13.0 for Windows was used for the statistical analysis. The data were analyzed using the Wilcoxon Mann-Whitney Test, with p Ͻ 0.05 set as the level of statistical significance. A binary logistic regression was also performed using this software (33). After the regression, the values of the prediction probability were applied to the discrimination of the samples. Receiver operating characteristic curve (ROC) was made by using the SPSS software. The cutoff values were calculated based on the results of the ROC. The optimized cutoff values in this study were those corresponding with the highest accuracy (maximum sensitivity and specificity). Because different ROC curves were performed for animals, small HCC and HCC, 3 optimized cutoff values were defined. The Multi Experiment Viewer software (Version 4.5.1, http://www.tm4.org) was used for the hierarchical clustering analysis (HCA) and significance analysis for microarrays (SAM). The SAM method was performed according to Tusher et al. (34).

Metabolic
Profiling of the Rat HCC Model-During DEN administration of 52 rats, six of these rats died of liver function failure or a hemorrhage of the tumor during the experiment. A total of 36 rats from the model group and 18 rats from the control group were sacrificed for histological observations (Fig. 1). According to the histological findings, all of the DEN treated rats that were alive at the end of the study (week 20) exhibited incidences of the liver tumor, as shown in the pathological analysis. The serial progression of hepatocarcinogenesis in this animal model was divided into three stages: the inflammation stage (week 4 -8), the cirrhosis stage (weeks 10 -14), and the HCC stage (week 16 -20). The time points at "week 6," "week 14, " and "week 20" were the characteristic histological changes of the inflammation, cirrhosis, and HCC stages, respectively ( Fig. 1). In addition, "week 10" and "week 16" were time points that were between two stages with mixed features, depending on the individual animals (data not shown here). Finally, the sera from each stage of the 10 DEN treated rats and the 10 matched control rats were collected for a metabolic profile analysis (Fig. 1). Fig. 2A shows the rat serum total ion chromatograms (TIC) of the control group at the normal growth process and the diseased rats at the different pathological stages, including inflammation (week 6), cirrhosis (week 14), and carcinoma (week 20). These chromatograms show the metabolic alterations among the different groups.
After the peak detection and alignment of all TICs, a total of 1459 metabolite ions were enrolled in the final data set for the statistical analysis. To model and evaluate the systemic changes in the rat metabolome, a partial least squares discriminate analysis (PLS-DA) of the data for each typical stage of the lesions was performed. Five components were calculated, and the cumulative R 2 Y and Q 2 were 0.6 and 0.26, respectively. No overfitting of the data was observed based As an ideal biomarker, it should be related to only the disease state, and have very few or no interference from the nondisease factors. To reduce the influence of the animal age, a Wilcoxon Mann-Whitney test was performed between the control and model groups with the same ages. It was found that 382, 427, and 445 variables significantly changed (p Ͻ 0.05) between the model and the control at the inflammation stage (week 6, 8), the cirrhosis (week 12, 14) and carcinoma (week 18, 20) stages, respectively. A total of 706 metabolites exhibited statistically significant differences (p Ͻ 0.05) in at least one stage of liver disease (Fig. 2C). These metabolites could better reflect the metabolic trends of the tumorigenesis and were also able to classify the three stages of the liver diseases (inflammation, cirrhosis, and carcinoma) using a PLS-DA model (R 2 Y was 0.76, and Q 2 was 0.4, Fig. 2D). The PLS-DA model achieved a better classification between cirrhosis and carcinoma, which was overlapped when all the variables were used in Fig. 2B.
Selection and Identification of Important Differential Metabolites-Using the above PLS-DA model that was based on 706 ions with significant differences, the statistically important metabolites were studied. According to the variable importance in the projection (VIP), a total of 52 variables (ions) with a VIP Ͼ2 were selected, and the chemical structures of 44 ions (supplementary Table S1) were identified based on the metabolite identification strategy (35,36). It was observed that these metabolites could also be found in the data set from the human metabolic profiling. Therefore, the results that were obtained from the rat models may be possibly extended to patients.
To further select the potential biomarkers from the 52 variables (ions) with a VIP Ͼ2, a heat map was constructed, providing the relative average contents of the selected ions in the model animals compared with the corresponding contents in the age-matched control animals (Fig. 3A). Several metabolites exhibited a characteristic trend of alterations that indicated the stage of the progression of hepatocarcinogenesis. To narrow down the scope of the biomarker pool, HCA was performed to understand the potential relationships among the metabolites. These metabolites were clustered according to their Pearson correlation coefficients, which were shown on the plot at different colors (Fig. 3B). The closely related metabolites were clustered, six major clusters were observed (Fig. 3B, I-VI). Cluster I included two poly-unsaturated fatty acids (PUFAs) (clupanodonic acid and docosahexaenoic acid) and LPC 22:5. Cluster II was mainly fatty acids and carnitines. Cluster III was all LPCs with different carbon chains. Cluster IV consisted of LPE, and cluster V was mainly LPC. Cluster VI was mainly bile acids and related ions.
Based on the above results and our previous work (28), the representative characteristic metabolites were selected from each cluster including LPC 22:5, palmityl-L-carnitine, LPC 22:6, LPE 16:0, LPC O-16:0 and TCA, which correspond to clusters I to VI, respectively. The metabolites in the same cluster also had similar changing trends. The metabolic trajectories of these six characteristic metabolites during carcinogenesis are shown in Fig. 4. The level of LPC 22:5 appeared to increase significantly (Mann-Whitney test, p Ͻ 0.05) during the early stage of HCC (week 16) and greatly increased as the tumor progressed compared with the control animals (Fig.  4A). The level of palmityl-L-carnitine decreased with the age of the animals in both groups (Fig. 4B), however, it increased significantly (p Ͻ 0.05) in week 8 between two groups. The level of LPE 16:0 increased as the cirrhosis developed (weeks 8 -14), and reached its peak during the advanced stage of HCC (weeks 20, Fig. 4C), and during this period, it increased significantly between the models and the controls (p Ͻ 0.05). The level of LPC 22:6 ( Fig. 4D) decreased during the late stage of inflammation (weeks 8 -10, p Ͻ 0.01), while increased in the late stage of HCC (weeks 20, p Ͻ 0.05). LPC O-16:0 (Fig. 4E) increased during the cirrhotic stage (weeks 12, p Ͻ 0.05) and in the advanced tumor stage (weeks 20, p Ͻ 0.001). The level of TCA significantly increased compared with that in the corresponding control (p Ͻ 0.01), especially at week 10 (Fig. 4F).
The "significance analysis for microarrays, (SAM)" method (34) was used to select the most significant metabolites. The Discrimination of the Hepatocarcinogenesis Stages-According to the relative contents of the metabolites (Fig. 4) and the SAM analysis (supplemental Fig. S1A, S1B), a discrimination flowchart is given in Fig. 5A.
Samples were randomly divided into a training data set (80% samples) to build the logistic model and a test data set (20% samples) for validation. LPE 16:0 and TCA were first applied to the classification of the control and diseased animals; 95.2% of the controls and 85.9% of the diseased animals were correctly discriminated in the training group, and 93.8% of the controls and 87.5% of the diseased animals in the test group were also correctly discriminated using the same model (cutoff value: 0.5, Fig. 5B). Next, the HCC animals were discriminated from other non-HCC rats using LPC 22:5, 81.2% of the HCC (weeks 18,20), and 97.5% of the non-HCC (weeks 6 -15) rats were correctly discriminated, and the accuracy was 80 and 75% in the test group, respectively (cutoff value: 0.5, Fig. 5C). Similarly, 93.8% of the inflammation (weeks 6, 8) and 93.8% of the cirrhosis (weeks 12, 14) animals were correctly identified using all three metabolites, and only 1 animal with hepatitis was wrongly classified (cutoff value: 0.5, Fig. 5D). Furthermore, Results from another DEN-induced HCC rat model cohort also provided similar ability of classification using the above logistic model (supplemental Fig. S2). The animals that were in the border stages ("week 10" and "week 16") were also predicted using the above binary logistic FIG. 5. Diagnostic potential of the three marker metabolites using binary logistic regression methods with the data from the rat models. A, Flowchart of the discriminating liver diseases based on the three marker metabolites. B, Liver disease animal models (weeks 6 -20) versus control animals (weeks 6 -20). C, HCC (weeks 18,20) versus CLD (weeks 6 -14). The state in the border stage (week 16) was predicted using the above HCC and the CLD binary logistic regressions. D, Hepatitis (Hep, weeks 6 -8) versus cirrhosis (Cir, weeks [12][13][14]. The state in the border stage (10W, week 10) was predicted using the above hepatitis versus cirrhosis binary logistic regressions. T: training dataset, S: test data set. regression, four of the 10 rats at week 10 were identified as having cirrhosis, and three of the 10 rats at "week 16" were predicted to be tumor-bearing.
Preclinical Validation of the Three Marker Metabolites-To test the usefulness of the three marker metabolites for human HCC diagnosis, especially discriminating small HCC from precancer cihhrosis and chronic hepatitis, 412 serum samples from 262 patients with HCC, 76 patients with cirrhosis and 74 patients with chronic hepatitis B were analyzed using the Orbitrap MS, and the normalized contents of the three marker metabolites were calculated. For the 95 patients with small HCC (solitary nodules with a diameter of less than 2 cm), the three marker metabolites reached a sensitivity of 80.5% and a specificity of 80.1% (the cutoff value: 0.45, Fig. 6A), whereas the results of AFP (the cutoff value: 20 ng/ml) for these patients were 53 and 64%, respectively. Furthermore, the ROC curve analysis of the three marker metabolites yielded an AUC of 0.882, which was greater than that of AFP (0.648, Fig. 6B). However, the combination of metabolic markers and AFP achieved an AUC of 0.879. When all HCC samples were enrolled, the sensitivity of the HCC diagnosis with three marker metabolites was 87.5%, and the specificity was 72.3% (the cutoff value: 0.65, Fig. 6C). When AFP was used as a biomarker for the same cohorts, 61.2% of the HCC group and 64% of the non-HCC group were correctly diagnosed (the cutoff value: 20 ng/ml). The ROC curve analysis revealed that the AUC of the three selected marker metabolites was 0.821, which was greater than that of AFP (0.678) (Fig. 6B).
The above results were calculated according to the optimized cut off values. Interestingly, there were also natural cut offs which were represented on Fig. 6. The natural cut offs are 0.35 for small HCC and 0.4 for all the patients. When these cutoffs were applied, a sensitivity of 92.7% and specificity of 59.7% could be achieved for the diagnosis of small HCC. And a sensitivity of 97.8% and specificity of 42.7% could be achieved for discrimination of all HCC patients.
When a cutoff value of 20 ng/ml AFP was used, 100 of 262 HCC patients were misdiagnosed. However, 91 of these 100 patients had a level of the metabolite markers higher than the FIG. 6. Diagnostic potential of the three marker metabolites. Using a binary logistic regression method, the samples from liver patients with chronic hepatitis, cirrhosis or HCC were classified. A, Discrimination of the non-HCC and small HCC patients using the three marker metabolites. B, ROC curves for the differential diagnosis by AFP and the metabolic markers using data from non-HCC and small HCC patients. C, Discrimination of the non-HCC and all HCC patients using the three marker metabolites. D, ROC curves for the differential diagnosis by AFP and the metabolic markers using data from non-HCC and all HCC patients. cutoff value (supplemental Fig. S3), implying the occurrence of HCC. On the other hand, for the 54 CLD patients with AFP above 20 ng/ml, only 16 cases showed positive for the "metabolite markers". In addition, the use of the 3 marker metabolites plus AFP significantly increased the specificity and the diagnostic performance of the metabolic markers (AUC: 0.895, Fig. 6B).

DISCUSSION
Early or subclinical diagnosis of HCC would be helpful for the prevention and treatment. Because HCC patients with chronic liver diseases usually take a long time before HCC occurs, there should exist an opportunity to discover the early biomarkers of the occurrence and development of HCC. The nontarget metabolomics provides a global view of the organism and can be used to monitor the dynamic metabolic alterations that occur in different pathological processes (37).
The hypothesis of the present study was that a metabolomics study of hepatocarcinogenesis would provide early or even subclinical metabolic markers for HCC, the serum metabolic profiling analysis of DEN-induced rat models of HCC was the first step of this work. The tissue pathological results showed that the rat HCC model was successfully constructed. The serial progression of hepatocarcinogenesis could be found, including the inflammation stage, the cirrhosis stage, and the carcinoma stage.
The study was first focused on the systemic metabolic changes in the rat metabolome using a PLS-DA model. As shown in Fig. 1B, aging and carcinogenesis are major influence factors in model rats. Moreover, these two factors are parallel to the time axis. So the age-related metabolites were first removed using a statistical analysis (Mann-Whitney test) and the remaining 706 ions were also able to classify the stages of liver diseases (Fig. 2D), which provide possibility for the exploration of marker candidates. By using PLS-DA to 706 ions, 52 statistically important variables with VIPϾ2 were defined. The identified differential metabolites (supplemental Table S1) showed that many metabolic pathways including fatty acids, bile acids, amino acids, phospholipids etc. were influenced during carcinogenesis in the model rats. Among these metabolic alterations, the six clusters of metabolites covered most of these differential metabolites and represented the major metabolic disturbance.
Our results indicated that the deregulation of lipids metabolism is of great importance during the carcinogenesis procedure. Fatty acids and carnitines were included in the list of "most differential metabolites" which reflected the abnormal metabolism of lipids. Poly unsaturated fatty acids (PUFA) such as clupanodonic acid and docosahexaenoic acid, could play important roles in the modulation of cell proliferation through cellular lipids peroxidation (38). The deregulation of PUFA has been considered as the early events in the liver carcinogenesis (39). L-carnitine acts as a protector from liver carcinogenesis by decreasing free fatty acid in blood and lower the oxidative damage from lipids peroxidation (40,41). Phospholipids take part in the metabolism of lipids. The balance of phospholipids constituting cellular membrane were kept by enzyme such as PE N-methyltransferase 2 (PEMT2), which catalyzes the conversion of PE to PC in the liver when dietary choline supply is inadequate. However, the expression of PEMT2 mRNA is reduced or absent in HCC and may cause the accumulation of PE or LPE, partly accounting for the high level of circulation PE or LPE (42). As evidence, the ratio of PE to PC in both blood and tissue elevated significantly, which indicated the low expression or activity in the model rats (data not shown). When clustered with the commonly used clinical liver injury indicator of aminotransferases (ALT, AST) and bilirubin, the reason for the increase of PE could partly be attributable to hepatocellular damage (supplemental Fig. S4). Bile acids are synthesized from cholesterol in the liver, and cycle in gallbladder and intestine. The high levels of bile acids in the circulation would be useful references for hepatocellular damages. Liver injury in HCC and chronic liver diseases could also be reflected on the increase of blood bile acids (43). The deregulation of lipids metabolism, lipids peroxidation, and cellular damage were the major metabolic events in the model rats during carcinogenesis. The possibly relationship of differential metabolites and the hepatocarcinogenesis procedure was also given in supplemental Fig. S5 of the supplementary materials.
In the exploration of the disease mechanism, we hope to identify as many differential metabolites from a variety of pathways as possible. As a result, the strategy of a combination of several feature selection methods is required (28). However, for the diagnosis of diseases, redundant features are not necessary. In this study, although 52 ions were found to be of great value for the discrimination of hepatocarcinogenesis, additional methods were still necessary to narrow down the scope of the biomarker pool. With the HCA analysis 52 ions were clustered into six types. The results of SAM indicated that TCA and LPE 16:0 are the most significant metabolic features which are positively related to the model, and LPC 22:5 is the most significant metabolic feature of the carcinoma stage. These three metabolites, LPE 16:0, TCA, and LPC 22:5, referred to as the marker metabolites, were finally defined for their potential applications. Using the binary logistic regression method (33), the capacity of the above three marker metabolites to discriminate liver diseases in rats was evaluated. The results provided preliminary evidence for these metabolites to be used as markers for the classification of HCC and chronic liver diseases.
The model rats in the different disease stages were usually discriminated according to their age, which is not precise but is convenient to estimate the progression of HCC. However, the rats in the border stages exhibited variable pathological differences, depending on the individuals. The rats at week 10 suffered from inflammation or cirrhosis, and several rats at week 16 were tumor-bearing. Thus, the discrimination of the rats at week 10 or week 16 was suggestive of the "early diagnosis" of cirrhosis or carcinoma, respectively. As shown in Figs. 5C and 5D, four rats at week 10 were classified as the cirrhosis, and three rats at week 16 were classified as the HCC. Although there is no direct pathological proof for the resulting classifications, the above results are valuable for guiding subsequent work aimed at discovering markers for the warning or subclinical diagnosis of HCC.
To determine whether the three metabolic markers that were identified using the rat models could be extended for clinical HCC diagnosis, we performed a preliminary validation using the sera from 412 patients with liver diseases, including chronic hepatitis B, cirrhosis, and HCC. Because the diagnosis of small HCC is a challenge, more attention was taken to it. Therefore, 95 cases of small HCC were first studied. Compared with the traditional HCC biomarker AFP, the combinational metabolic markers showed a better sensitivity. Similarly, data from all 412 patients show three marker metabolites have a better AUC value than the AFP (Fig. 6). It should be pointed out that metabolites are regulated by diverse intrinsic or extrinsic factors, the specificity of the metabolic markers is still to be improved, the one of ways is the combinational use of the AFP and the metabolic markers (Figs. 6B and 6D).
In summary, the results of the present metabolomics study revealed the dynamic changes in hepatocarcinogenesis using the DEN-induced rat model. However, to apply these results for clinical use, the differences between rats and humans and between chemical-and HBV-induced carcinogenesis must be considered. Therefore, the validation using large scaled serum samples from patients is very important. In the present work, after acquiring the rat serum metabolic profile using LC-MS, multivariate data analysis methods were utilized to screen potential biomarkers. Of 52 differential metabolites, three marker metabolites that provided the effective classification of the disease stages of tumorigenesis were defined. More importantly, these metabolites were shown to be effective in distinguishing patients with chronic hepatitis, cirrhosis or HCC, especially those small HCCs. Further work is needed for confirmation of these metabolites as early or subclinical biomarkers. In addition, these metabolic features indicated the deregulation of lipid metabolism during hepatocarcinogenesis, which provides useful clues for future mechanism exploration and identification of therapeutic targets of HCC. The present study also highlights the ability of nontarget metabolomics approaches to investigate the dynamic metabolic alterations that occur during the complex biological processes in carcinogenesis. *