The prediction of chemosensitivity is a challenging problem in the management of cancer. In the present study, a metabonomic approach was proposed to assess the feasibility of chemosensitivity prediction in a human xenograft model of gastric cancer. BALB/c-nu/nu mice were transplanted with MKN-45 cell line to establish the xenograft model. The mice were then randomized into treatment group (cisplatin and 5-fluorouracil) and control group (0.9% sodium chloride), and their plasma were collected before treatment. Metabolic profiles of all plasma samples were acquired by using high-performance liquid chromatography coupled with a quadrupole time-of-flight mass spectrometer (HPLC/Q-TOF-MS). Based on the data of metabolic profiles and k-Nearest Neighbor algorithm, a prediction model for chemosensitivity was developed and an average accuracy of 90.4% was achieved. In addition, a series of endogenous metabolites, including 1-acyl-lysophosphatidycholines, polyunsaturated fatty acids and their derivatives, were determined as potential indicators of chemosensitivity. In conclusion, our results suggest that the proposed metabonomic approach allows effective chemosensitivity prediction in human xenograft model of gastric cancer. The approach presents a new concept in the chemosensitivtiy prediction of cancer and is expected to be developed as a powerful tool in the personalized cancer therapy.
If you can't find a tool you're looking for, please click the link at the top of the page to go "Back to old version". We'll be adding more features regularly and your feedback is important to us, so please let us know if you have comments or ideas for improvement.
Chemotherapy with cytotoxic drugs plays an essential role in the management of advanced gastric cancer.1 However, not all patients with gastric cancer respond well to the cytotoxic drugs.2 Although these patients might have similar histopathological type or TNM stage when they receive similar chemotherapy regimens, individual difference in chemosensitivity still could be observed. In particular, to a portion of patients, chemotherapy not only yields little positive response on the tumor but brings severe toxicity and functional impairment.3, 4 Therefore, how to pick out the individuals with high chemosensitivity from gastric cancer patients is crucial to the personalized therapy for gastric cancer.5
In early years, there are quite a few ways to predict chemosensitivity of cancer, such as clone formation, proliferation, cell metabolic activity assays in vitro, tumor growth and survival assays in vivo.6 Unfortunately, these methods usually suffer from less sensitivity, specificity and accuracy, and lack of feasibility and reproducibility in clinical practice.7 Recently, a few of specific genes or proteins have been demonstrated as useful biomarkers for chemotherapy of gastric cancer based on pharmacogenetics.8–15 However, just by these limited biomolecules, the chemosensitivity could not be effectively determined. Thereby, high-throughput “omic” methods, such as genomics and proteomics, have been employed as powerful tools for better stratification of cancer patients to facilitate the individualization of chemotherapy.16–20
Metabonomics, as another high-throughput “omic” approach, can provide information distinctive from genomics and proteomics.21 In the last few years, metabonomics has been successfully applied in several fields of cancer research including diagnosis,22, 23 prognosis24 and drug evaluation,25–27 and it is also believed to be an alternative strategy for individualized therapy of cancer.28–29 In the present study, to determine the potential value of metabonomics in chemosensitivity prediction, a high-performance liquid chromatography/quadrupole time-of-flight mass spectrometer (HPLC/Q-TOF-MS) based metabonomic approach was introduced to acquire initial metabolic profiles of plasma and develop a prediction model of chemosensitivity in a human xenograft model of gastric cancer. The proposed metabonomic approach might bring new insights into chemosensitivity prediction and individualized chemotherapy for cancer patients.
Material and Methods
Cisplatin (CDDP) and 5-fluorouracil (5-FU) were obtained from Qilu Pharmaceuticals and Shanghai Xudong Haipu Pharmaceuticals, China, respectively. Before their use, CDDP was immediately dissolved in 0.9% sodium chloride injection to a working concentration of 0.4 mg/ml and 5-FU was dissolved in the same solvent to a working concentration of 2 mg/ml individually. Three commercial standards for MS/MS spectra, 1-stearoylglycerophosphocholine, 1-palmitoyllysophosphatidylcholine, and glycerophosphorylcholine were purchased from Sigma Chemicals (USA).
MKN-45, a cell line of poor differentiated human gastric adenocarcinoma, was kindly provided by Dr. Ming Yao (Shanghai Cancer Institute, China). The cell line was cultured in RPMI 1640 medium (GIBCO) supplemented with 10% fetal bovine serum and antibiotics in a humidified atmosphere containing 5% CO2 at 37°C.
Specific pathogen-free BALB/c-nu/nu mice were obtained from Shanghai Cancer Institute at 4–6 weeks of age and quarantined for 2 weeks prior to being eligible for entry into the study. Mice were fed under aseptic conditions which strictly followed the National Institutes of Health Guide for the Care and Use of Laboratory Animals. All animals were given free access to normal diet and drinking all the time. The protocol was approved by the Medical Ethics Committee, Second Military Medical University (Shanghai, China).
Human xenograft model of gastric cancer
The efficacy of CDDP plus 5-FU (PF regimen) was investigated in a human xenograft model of gastric cancer. The xenograft tumors were established in 6–8-week-old mice through subcutaneous injection of 1 × 106 MKN-45 cells into the dorsal flank. After a fifth passage culture in vivo, the xenograft tumors were harvested, divided and inoculated for a large-scale preparation of xenograft models. When most of xenograft tumors reached a volume of ∼100 mm3, mice were randomized into treatment group and control group. Every mouse was labeled by ear-cutting and their blood (150–200 μl) was collected from the orbital venous plexus for HPLC/MS analysis on the day before treatment (Day 0). The treatment group received CDDP (4 mg/kg) on Day 1, followed by 5-FU (20 mg/kg/day) on Day 2–6. The control group received equal volumes of 0.9% sodium chloride injection (10 ml/kg/day) on Day 1–6. Drug solutions were delivered by intraperitoneal injection.
Evaluation of chemotherapy response
The tumor diameter of each mouse was measured daily. The tumor volume (TV) and relative tumor volume (RTV) were calculated by the following formulas:
where a is the length and b is the width of tumor.
where V0 is the TV at the Day 0 and Vt is the TVs of succedent measurements.
RTV was considered as an index of chemosensitivity in our study. The smaller the RTV, the higher the chemosensitivity. Based on the hypothesis that different individual mice would have different response to CDDP plus 5-FU, mice in the treatment group was divided into 3 subgroups according to their RTVs: sensitive group (mice with smaller RTVs), intermediate group (mice with intermediate RTVs) and resistant group (mice with larger RTVs).
The overall antitumor activity of CDDP plus 5-FU was evaluated by both the relative tumor growth rate and the tumor growth inhibitory rate, which were calculated by the following formulas:
Relative tumor growth rate = RTVT/RTVC × 100% (RTVT: RTV of treatment group; RTVC: RTV of control group)
Tumor growth inhibitory rate = (1 − WT/WC) × 100% (WT: mean tumor weight of treatment group; WC: mean tumor weight of control group)
The blood plasma was separated by centrifugation at 3,000 g for 20 min after anticoagulation, and the aliquots were stored at −80°C. Before HPLC/MS analysis, all samples were thawed at room temperature. A double volume of methanol was added into the supernatant and shaken vigorously for 30 sec. The mixture was tranquillized for 10 min, and then centrifuged at 10,000 g for 5 min. The supernatant was filtered through a 0.2-μm syringe filter for HPLC/MS analysis. Additionally, when the treatment was over, the mice were sacrificed and the subcutaneous tumors were excised, weighed, and prepared for histopathological examination with hematoxylin-eosin staining.
An Agilent-1200 RRLC/6510 Q-TOF system (Agilent, MA) and an Eclipse XDB-C18 column (3.0 × 100 mm, 1.8 μm, Agilent) were used for HPLC/MS analysis. The mobile phase was composed of water (A) and acetonitrile (B). The gradient program was as follows: 0–2 min, 98% A, 2% B; 5 min, 35% A, 65% B; 13 min, 25% A, 75% B; 16–17 min, 5% A, 95% B. Elution was performed at a solvent flow rate of 0.4 ml/min, and a portion of the column effluent (0.2 ml/min) was delivered into the ion source of mass spectrometry. The column compartment was kept at 25°C, and the sample injection volume was 6 μl. The conditions of mass analysis were as follows: drying gas N2 8 l/min, temperature 330°C, pressure of nebulizer 35 psi, capillary voltage 4,000 V and scan range 50–1000 m/z.
LC/MS data were deconvoluted and aligned with mass and retention time tolerances using GeneSpring software (Version 1.1, Agilent, USA) to generate a matrix containing information on mass, retention time, and intensities for all peaks. Prior to multivariate statistical analysis, data of each chromatogram was normalized to a constant integrated intensity of the number of peaks to partially compensate for the concentration bias of each sample, and then, the original variables were standardized mathematically by:
where p is the number of variables, and is the average of variable j.
To develop prediction model of chemosensitivity, the dataset of treatment group was randomly divided into a training set of 40 samples and a test set of 20 samples. The training set was used to build prediction model while the test set was used to evaluate the model performance. To reduce the risk of overfitting, the test set was not used to monitor the training process. The random division of training set and test set were performed 40 times to control the possible bias and variance, and the means and standard deviations of prediction accuracy were indicated.
The dataset produced by HPLC/MS profiling involved a large degree of redundant information, and partial least squares (PLS) was used for information extraction. After PLS, principal components which denoted the most variance of the whole dataset were applied to build a chemosensitivity prediction model for gastric cancer. In our study, both PLS and hierarchical PLS (HPLS) were applied.30, 31 PLS was carried out in only 1 step while HPLS involved 2 steps as follows: first, PLS was directly conducted on the merged data of all 3 kinds of samples; then, data of samples in sensitive group were removed, and the rest data (samples in intermediate and resistant groups) were fed for the second PLS. The flowchart of the general strategy for HPLS is given in Figure 1.
K-Nearest Neighbor (k-NN) was applied to build prediction model. K-NN methodology is based on a simple similarity learning approach, whereby an unknown sample in the test set is classified into the class to which the majority of its k most similar neighbors in the training set belongs.32 K-NN algorithm is implemented simply as follows: (i) calculate the Euclidean distance between an unknown sample and each of training samples; (ii) select k samples from the training set, which are nearest to the unknown sample according to the calculated distance; (iii) classify the unknown sample with the class to which the majority of the k samples belongs. Parameter k was optimized via leave-one-out cross-validation method. Programs of HPLS/PLS and kNN were coded in MATLAB 7.0 (Mathworks, USA; see Supporting Information 1).
Chemosensitivity-related metabolites were determined by the loading plot of PLS. The chemical structures of these metabolites were identified as follows: first, the Agilent METLIN Personal Metabolite Database (Version B.01.00) was searched by mass weight and a list of candidates was obtained; then tandem mass analysis was carried out, and according to the possible fragment mechanisms, items without characteristic mass fragment information were removed from the list, with the most probable metabolic indicators survived; finally, by comparing the retention times and mass spectra to the commercial standards, part of the chemosensitivity-related metabolites were structurally confirmed.
Quantitative data was presented as mean ± SD. Means of a single factor were compared by t-test and analysis of variance (ANOVA). p < 0.05 was considered significant. The statistical analysis was conducted by SPSS 10.0 software package (SPSS, Chicago, IL).
One hundred and eight xenograft models were successfully established. The rate of tumor formation after 1 week reached 100%. The TVs of most mice reached ∼100 mm3 after 2 weeks. Twenty-eight mice models with irregular tumors were excluded and the remaining 80 mice were included in the study. After random allocation, 60 mice were classified as treatment group, and the other 20 mice were regarded as control group. No difference was found in view of TV, RTV and body weight between the 2 groups (p values > 0.05) before treatment. From the 3rd day of administration, the tumors of mice in treatment group grew significantly slower than those in control group (Figure 2). The relative tumor growth rate was 59.6% and the tumor growth inhibitory rate was 41.9%, which suggested that CDDP plus 5-FU were active agents for gastric cancer (p values < 0.001).
Remarkable disparity existed in individual mice concerning their chemosensitivity to CDDP plus 5-FU. According to their RTVs, the treatment group was rough equally divided by 2 cut-points (3 and 4) into 3 subgroups: sensitive group (RTVs ≤ 3, n = 19), intermediate group (3 < RTVs ≤ 4, n = 20) and resistant group (RTVs > 4, n = 21). As shown in Figure 3, the RTVs of 3 subgroups in treatment group were all less than that of control group. Furthermore, RTVs decreased gradually when their chemosensitivity increased (p < 0.001). But there was no significant difference between the RTVs of resistant group and control group (p = 0.392), which suggested that CDDP plus 5-FU had little antitumor activity for the mice in resistant group. In addition, the histopathological examination of tumor xenografts among the 4 groups also demonstrated that the difference of antitumor effect was significant. Representative histopathological features of the tumor xenografts are shown in Figure 4, which were consistent with the variation trend of their RTVs.
HPLC/MS assay was performed and its operating conditions were optimized to acquire features as many as possible in a single injection. Both positive and negative ion mode analyses were experimented, and the HPLC/MS profile in the positive mode was finally adopted since it provided much more metabolite information. In our study, totally 8,873 features (defined by a pair of mass-charge ratio and retention time) were resolved from LC/MS profile of each sample (Raw data of HPLC/MS analysis were attached as Supporting Information 2). Figure 5 lists total ion current (TIC) profiles of typical samples of each group. Prior to PLS and modeling, the data matrix containing intensities of all peaks resolved from each sample was standardized according to the equation in Data preprocessing section.
Both HPLS and PLS were performed on the training set, and k-NN method was employed to construct predictive models of chemosensitivity for experimental gastric cancer. In 40 times of k-NN modeling, the parameter k was optimized via leave-one-out cross validation method in each procedure (validation results were indicated in Supporting Information Table I). The test set was used to evaluate the performance of k-NN prediction models. As a result, HPLS-kNN method achieved an average prediction accuracy of 90.4 ± 6.4%, and comparatively, the average prediction accuracy was 87.1 ± 6.8% for direct PLS-kNN method (detailed prediction accuracy of each random split was enclosed in Supporting Information Table II). HPLS-kNN permitted the metabolic characteristics of chemosensitivity for gastric cancer to be more efficiently extracted hierarchically, and higher prediction accuracy could be obtained (t = 2.181, p = 0.032). Overall, the metabonomic approach combined with multivariate data analysis and k-NN was able to accurately predict the possible chemotherapy response of experimental gastric cancer.
Chemosensitivity-related metabolites were determined by the loading plot of PLS, and different data sets used for PLS analysis would bring about distinct metabolic indicators. In our study, PLS analysis was carried out on 2 different data sets from the treatment group. One data set was based on the whole 60 mice samples, and the other was on a reduced data set composed of 30 representative mice samples only. The 30 mice in the latter data set were defined as 10 mice in each 3 group with their RTVs away from the cut-points. Figures 6a and 6b present the score plots of PLS on the data matrix of all 60 samples and representative 30 mice samples, respectively. Comparatively, the latter shows a satisfactory clustering according to 3 groups, which would accordingly produce potential metabolic indicators with stronger discriminatory power. For this reason, the chemosensitivity-related metabolites were discovered based on PLS analysis on the reduced data set composing of representative 30 mice samples.
Several potential metabolic indicators were suggested in the loading plot of PLS (Figure 7), and 18 chemosensitivity-related metabolites were structurally postulated according to possible fragment mechanisms (the MS2 fragments analysis for these metabolites was attached as Supporting Information 3). As shown in Table 1, nearly a half of these metabolites were a series of 1-acyl-lysophosphatidylcholines such as linolenoyl lysolecithin and arachidonyl lysolecithin, and another half of these metabolite were made up of a cluster of polyunsaturated fatty acids and their active derivatives, including docosahexaenoic acid, arachidonic acid, 2,3-dinor-PGE1, 2,3-dinor-TXB2 and so on. In addition, 3 of these metabolites were further confirmed by comparing their retention time and mass spectra with commercial standards (see MS spectra in Supporting Information Fig. 1).
Table 1. Data of MS fragments of postulated metabolites
The prediction of chemosensitivity is a challenging problem in the clinical management of cancer. Despite many predictive methods developed, most of them fail to achieve an efficient prediction because chemotherapy response reflect not only properties intrinsic to the cancer cell, but also host metabolic properties.33 Moreover, these methods usually suffer from time-consuming, invasive, and lack of high predictive accuracy.34 These shortcomings disable their extensive application and an integral approach with high specificity and accuracy is urgently needed in clinical practice. Metabonomics, which is defined as a systems approach to investigating the metabolic consequences of physiopathologic modification in a multivariate and dynamic manner,35 is believed to be a possible alternative strategy to prediction of chemotherapy response for cancer patients.28–29 Theoretically, alterations of metabolome can describe the activities of varied genes, drug enzymes, different functional proteins and detoxification pathways which may influence the activity of the anticancer agents,36 and therefore associations between metabolic profile and chemotherapy response may contribute to the chemosensitivity prediction and the tailoring of chemotherapy. In the present study, a HPLC/MS based metabonomic approach was successfully applied and a prediction model with 90.4% accuracy for chemosensitivity was built using HPLS combined with k-NN method. The results suggested that the metabonomic approach to the chemosensitivity prediction was feasible and was expected to be developed as a potential tool for personalized cancer therapy.
PLS has been widely used as a feature extraction technique in metabolic profiling and can be performed directly or hierarchically.37, 38 In the case of direct PLS, all samples of training set is merged into one matrix, on which PLS is directly performed. In the case of HPLS, it needs multiple steps, and in each step the data matrix contains only data that have not been properly classified in previous PLS steps. Generally, direct PLS is sufficient to fulfill feature extraction of gigantic dimensional data in most cases. However, especially in case of the multi-catalogs of data, direct PLS might not reach satisfactory results for each catalog of data simultaneously, and HPLS should be considered as an alterative solution. In our study, direct PLS and HPLS were comparatively applied and the results indicated that HPLS was superior to direct PLS for feature extraction. In addition, the model performance is also highly associated with the modeling technique applied. It is generally agreed that diseases and metabolites behave in a very complicated nonlinear relationships and thus linear algorithm often shows its limitations in application. Therefore, a nonlinear modeling approach of k-NN was used to build the prediction model in our study.
PLS is also a widely accepted approach to detect potential biomarkers in metabonomic studies.39 PLS score plot represents the relative position of the objects in the space (two-dimensional or three-dimensional) of the principal components and clusters of various classes of objects may be easily separated. PLS loading plot will directly identify variables with high discriminating power for various class pairs. The outliers in loading plot are the most contributors to the separation of classes and thus are often considered as potential biomarkers. Generally speaking, the discrimination ability of these outliers directly rely on the separation of classes displayed in PLS score plot. In the present study, 2 different data sets were comparatively fed for PLS analysis. When data matrix of all 60 samples used, the score plot of PLS shows a poor clustering of 3 groups and could not be well separated from each other (Fig. 6a). The main reason might lie in the nature of RTVs (indicating anticancer efficacy), which were continuous variables ranged from 0.53 to 9.04 with no objective cut-points for categorization. The splitting that the treatment group was divided into 3 groups according to their RTVs just was based on our subjective experiences. For mice with similar RTVs around the cut-points, there may be little individual difference between them even though they were divided into 2 different groups. Under this condition, the comparison of these similar individuals between 2 near groups would lead to poor clustering and potential biomarkers with poor discriminatory power. So, we removed half of mice whose RTVs are near to cut-points in each 3 groups to make the 3 groups with different features clearly. As we assumed, based on the reduced data set composing 30 representative samples, the score plot (Fig. 6b) shows a better clustering and could be clearly separated from each other, which means the indicator candidates might show higher specificity.
An electrospray ionization (ESI) based LC/MS system was employed in our study. Compared to conventional ESI, microspray and nanospray techniques can provide more efficient ionization and much higher sensitivity, and thus have a higher potential to be used for the profiling of complicated analyte such as serum and urine.40–41 In our assays, although the limitation of sensitivity might lead to a miss of biomarkers with less concentration but more specificity, abundance of metabolic information (8,873 features) still could be resolved from TIC profiles using conventional ESI. Our prediction model achieved over 90% accuracy, which suggested that ESI-MS based metabolic profiling could properly reveal different characteristics of chemosensitivity from sensitive, intermediate, and resistant groups.
In the present study, a series of metabolites were proposed as potential indicators of chemosensitivity for gastric cancer, which might be meaningful for us to understand the mechanism of chemotherapy response and had value in the prediction of chemosensitivity in clinical settings. The indicators mainly conclude 1-acyl-lysophosphatidylcholines and polyunsaturated fatty acids, which are simultaneous hydrolysis products of phosphatidylcholine by phospholipase A2. Other metabolites, such as diacylglycerol and glycerophosphocholine, can be considered as the products of hydrolysis of phosphatidylcholine by phospholipase C and 1-acyl-lysophosphatidylcholine by phospholipase B1 individually. As a result, the biosynthetic and degradation pathways of 1-acyl-lysophosphatidylcholines might correlate notably with the chemosensitivity of gastric cancer. From the viewpoint of biochemistry, the biosynthetic and degradation pathways of 1-acyl-lysophosphatidylcholines are composed of enzymes such as phospholipase A2 and B1, lecithin-cholesterol acyltransferase (LCAT) and lysophosphatidylcholine acyltransferases (LPCATs),42–45 which might be key switch modulating the chemosensitivity of gastric cancer to cytotoxic drugs. However, our present findings about these metabolites and enzymes are as yet still at the initial stage, and their role as indicators for chemosensitivity prediction remains to be validated in clinical trials.
In conclusion, our results suggest that the proposed metabonomic approach allows effective chemosensitivity prediction in a human xenograft model of gastric cancer. The approach presents a new concept in the chemosensitivtiy prediction of cancer and is expected to be developed as a powerful tool in the personalized cancer chemotherapy.
We sincerely thank Mr. Ming Yao, Ms. Ming-Xia Yan, and Mr. De-Shui Jia, Shanghai Cancer Institute, China, for their kind assistance in the animal experiments.