• <tr id="yyy80"></tr>
  • <sup id="yyy80"></sup>
  • <tfoot id="yyy80"><noscript id="yyy80"></noscript></tfoot>
  • 99热精品在线国产_美女午夜性视频免费_国产精品国产高清国产av_av欧美777_自拍偷自拍亚洲精品老妇_亚洲熟女精品中文字幕_www日本黄色视频网_国产精品野战在线观看 ?

    Plasma-metabolite-based machine learning is a promising diagnostic approach for esophageal squamous cell carcinoma investigation

    2021-09-14 07:46:50ZhongjianChenXiancongHuangYunGaoSuZengWeiminMao
    Journal of Pharmaceutical Analysis 2021年4期

    Zhongjian Chen , Xiancong Huang , Yun Gao , Su Zeng , Weimin Mao ,*

    a Laboratory of Pharmaceutical Analysis and Drug Metabolism, College of Pharmaceutical Sciences, Zhejiang University, Hangzhou, 310058, China

    b The Cancer Research Institute,The Cancer Hospital of the University of Chinese Academy of Sciences(Zhejiang Cancer Hospital),Institute of Basic Medicine and Cancer (IBMC), Chinese Academy of Sciences, Hangzhou, 310022, China

    Keywords:Diagnostic Esophageal squamous cell carcinoma(ESCC)Metabolomics Machine learning Prognostic

    ABSTRACT The aim of this study was to develop a diagnostic strategy for esophageal squamous cell carcinoma(ESCC)that combines plasma metabolomics with machine learning algorithms.Plasma-based untargeted metabolomics analysis was performed with samples derived from 88 ESCC patients and 52 healthy controls. The dataset was split into a training set and a test set. After identification of differential metabolites in training set, single-metabolite-based receiver operating characteristic (ROC) curves and multiple-metabolite-based machine learning models were used to distinguish between ESCC patients and healthy controls. Kaplan-Meier survival analysis and Cox proportional hazards regression analysis were performed to investigate the prognostic significance of the plasma metabolites. Finally, twelve differential plasma metabolites (six up-regulated and six down-regulated) were annotated. The predictive performance of the six most prevalent diagnostic metabolites through the diagnostic models in the test set were as follows: arachidonic acid (accuracy: 0.887), sebacic acid (accuracy: 0.867), indoxyl sulfate (accuracy: 0.850), phosphatidylcholine (PC) (14:0/0:0) (accuracy: 0.825), deoxycholic acid(accuracy: 0.773), and trimethylamine N-oxide (accuracy: 0.653). The prediction accuracies of the machine learning models in the test set were partial least-square (accuracy: 0.947), random forest (accuracy:0.947),gradient boosting machine(accuracy:0.960),and support vector machine(accuracy:0.980).Additionally, survival analysis demonstrated that acetoacetic acid was an unfavorable prognostic factor(hazard ratio (HR): 1.752), while PC (14:0/0:0) (HR: 0.577) was a favorable prognostic factor for ESCC.This study devised an innovative strategy for ESCC diagnosis by combining plasma metabolomics with machine learning algorithms and revealed its potential to become a novel screening test for ESCC.

    1. Introduction

    Esophageal squamous cell carcinoma (ESCC), a predominant histological subtype of esophageal cancer, occurs most frequently in developing countries such as China[1].It is difficult to diagnose ESCC at an early stage due to a lack of typical symptoms as well as specific and sensitive biomarkers of this tumor [2]. Consequently,patients are often diagnosed at a relatively advanced stage,usually accompanied by lymph node metastasis and invasion. Unfortunately,no effective treatment has been reported available for such plight currently[3].Despite significant improvements in diagnostic modalities and treatments including surgery, radiation, chemotherapy,and their combination,the prognosis for ESCC still remains unsatisfactory[4,5].Furthermore,the 5-year overall survival rate is about 20% and only 1% for those with advanced stages [6]. Therefore,early detection is extremely important with an urgent need for a novel and accurate means by which to diagnose ESCC.

    A hallmark of malignancy is metabolic changes [7] through which cancer cells reprogram normal metabolic pathways that support uncontrolled proliferation. Some of the most striking alterations include elevation of glycolysis[8],up-regulation of amino acid [9] and lipid metabolism [10,11], as well as macromolecule biosynthesis [12]. Thus, the investigation of metabolic perturbations in cancer may be a promising means to discover novel cancer biomarkers and therapeutic targets.

    Metabolomics is a powerful and efficient tool for the discovery of metabolic biomarkers and targets within biological specimens.Previously, plasma- and tissue-based ESCC metabolomic studies have identified potential diagnostic and prognostic metabolites(e.g., tryptophan and kynurenine), suggesting metabolic reprogramming to be associated with the initiation and development of ESCC [2,13-17]. Also, liquid biopsy for cancer diagnosis has many advantages, such as relatively low invasiveness [18]. Herein,plasma/serum-based metabolomics is an attractive approach for the discovery of ESCC diagnostic biomarkers.

    For diagnosis, metabolomics data analysis requires statistical and machine learning-based classification methods [19]. Machine learning is a type of computer algorithm focusing on prediction through pattern recognition [20]. Principal component analysis(PCA) and partial least-square discriminant analysis (PLS-DA) are the two most widely used multivariate analysis methods for metabolomic studies [21]. In addition to these two well-known methods, the application of other machine learning algorithms,such as random forest(RF),gradient boosting machine(GBM),and support vector machine (SVM), has emerged in recent years and shown a promising diagnostic potential when combined with metabolomics data [19,22]. However, when reviewing previously published ESCC metabolomics studies,most of them were found to solely focus on the discovery of differential metabolites, and use single-metabolite diagnostic models with a rare evaluation of multiple-metabolite-based machine learning models [15,16,23].Therefore, further studies employing a combination of plasma metabolomics and machine learning algorithms are encouraged,which might have potential clinical usefulness.

    Initially, an untargeted plasma metabolomics study was conducted in a cohort consisting of 88 ESCC patients and 52 healthy controls. Based on the differential metabolites annotated, both single-metabolite-based receiver operating characteristic (ROC)curves and multiple-metabolite-based machine learning models,including PLS, RF, GBM, and SVM, were established in the training set (n =100). In the test set, the predictive performance of the machine learning models had an accuracy range of 0.947-0.980,which was higher than that of single-metabolite models (an accuracy range of 0.653-0.887). These findings indicate that plasmametabolites-based machine learning models are an excellent diagnostic strategy for ESCC.

    2. Materials and methods

    2.1. Reagents

    High-performance liquid chromatography (HPLC) grade acetonitrile and methanol were purchased from Tedia Co.Inc.(Fairfield,OH, USA). HPLC grade formic acid was purchased from Roe Scientific Inc. (Newark, DE, USA). Distilled water was from Wahaha Group Co., Ltd. (Hangzhou, China).

    2.2. Plasma samples

    ESCC plasma samples were collected from 88 patients recruited after histopathologic confirmation of ESCC and radical resection at Zhejiang Cancer Hospital,China,from May 2010 to December 2012.The clinical stages of ESCC patients were determined based on the American Joint Committee on Cancer 8th edition staging system[24]. Participants were followed up until December 2017, and the overall survival(OS) from their surgery to the date of death or the last follow-up visit was evaluated.Healthy controls,recruited from our health examination center, were matched with ESCC patients based on age and sex. Fasting blood samples were collected from preoperative patients and healthy controls at approximately 8 a.m.,with plasma immediately separated from whole blood by centrifugation at 1000 g, 4°C for 10 min. Di-potassium salt of ethylenediaminetetraacetic acid was used as the anticoagulant.All the plasma samples were stored at-80°C until analysis.The basic characteristics of these samples are listed in Table 1.

    The study protocol was performed in accordance with the Declaration of Helsinki, approved by the Research Ethics Committee of Zhejiang Cancer Hospital, with written informed consent obtained from all individuals.

    2.3. Sample preparation

    Sample preparation was conducted according to Huang et al.[25]. Briefly, each plasma sample (50 μL) was thawed on ice and immediately mixed with 200 μL of ice-cold acetonitrile. After vortexing for 1 min,the mixture was centrifuged at 16,000 g, 4°C for 15 min.The supernatant(150 μL)was transferred into a fresh tube and lyophilized till dry. Residues were dissolved by mixing for 1 min with 80 μL of a solution consisting of 25%acetonitrile and 75%water. After centrifuging for 15 min at 16,000 g and 4°C, 60 μL of the supernatant was transferred into a sample vial. An aliquot of 5 μL supernatant was used for liquid chromatography-mass spectrometry (LC-MS) analysis.

    2.4. LC-MS analysis

    An Ultimate 3000 UPLC system (Dionex, Idstein, Germany)linked to a Q Exactive Orbitrap mass spectrometer(Thermo Fisher Scientific, Bremen, Germany) was used for this un-targeted metabolomics study. Separation was performed on an ACQUITY UPLC HSS T3 column (2.1 mm × 100 mm,1.8 μm, Waters, Milford,MA,USA)at 35°C.The mobile phase was water containing 0.1%(V/V)formic acid(phase A)and acetonitrile(phase B),with a flow rate of 0.3 mL/min.The linear gradients of phase B were as follows:2%for 0-1 min,2%-100%for 1-10 min,100%-2%for 10-13 min,and 2%for 13-16 min.Linear gradients of phase A changed accordingly complementary to that of phase B. The electrospray voltages were 3.5 kV in positive mode and 2.5 kV in negative mode. The probeheater temperatures were set at 320°C and 350°C in positive and negative modes, respectively. The sheath gas was set at 35 and 40 arb in positive and negative modes,respectively.For collecting MS/MS spectra,a data-dependent acquisition mode for top 10 ions was conducted with a mass resolution of 17,500 and stepped collision energies of 10, 20, and 40 eV.

    Table 1 Clinical information of the patients.

    Quality control (QC) samples were prepared by pooling redissolved samples in equal amounts (15 μL) and periodically analyzed throughout the entire analytical run to monitor instrument stability.

    2.5. Metabolomics data analysis

    Metabolomics data analysis was performed according to Yang et al. [26]. Briefly, R package XCMS (version 3.8.2) was utilized for processing mass raw data,including peak detection,retention time alignment, peak matching, and correction. R package MetaX package (version 1.4.16) was used for ion filtration based on the following exclusion criteria:(1)ions not detected in over 50%of all QC samples or over 80% of all non-QC samples; (2) ions with relative standard deviation >30% in QC samples. QC-based robust LOESS signal correction was applied to reduce the influence of signal shift.

    The cohort data(n =140,ESCC case(C):88,healthy control(N):52)were randomly divided into a training set(n =100,C:63,N:37)and a test set (n =40, C: 25, N: 15). To discover differential metabolites, an unsupervised PCA was first conducted to investigate the trends for all samples in the training set.Then supervised PLSDA was performed to identify the most discriminating ion features between ESCC plasma and non-cancerous counterparts based on VIP values. Finally, those with VIP > 1, false discovery rate(FDR) <0.05, and |log2(fold change)| >0.584 were defined as differential ion features. Metabolite annotation was performed using Progenesis QI (Waters, Milford, MA, USA)software based on METLIN (http://metlin.scripps.edu) and HMDB (http://www.hmdb.ca/). Metabolism pathway analysis was conducted using the online tool, Metaboanalyst (https://www.metaboanalyst.ca/MetaboAnalyst/home.xhtml).

    2.6. Development of diagnostic models using single-metabolite ROC curves and metabolite-based machine learning models

    For single metabolites, ROC curves were first analyzed for each metabolite in the training set. Youden's index (sum of sensitivity and specificity minus one)was used as a criterion for selecting the optimum cut-off point for each metabolite.With cut-off points for each metabolite, predictive classes were calculated for each unknown sample in the test set.The predictive performance including accuracy,sensitivity,and specificity was then calculated for the test set.

    Fig.1. Principle component analysis (PCA) score plot of plasma metabolic profiles of ESCC patients and healthy controls.(A)Training set and(B)test set.C:ESCC patient;N:healthy control.

    For metabolite-based machine learning modeling, data in the training set were preprocessed with“scaling”and“centering”.The same preprocessing methods with the same parameters were applied to the test set.Algorithms including PLS,RF,GBM,and SVM were investigated for cancerous and non-cancerous classification.R package caret(version 6.0-85)was utilized to train and test PLS,RF,and GBM models,while R package e1071(version 1.7-3)was used to train and test SVM model. Ten repeated and five-fold crossvalidation was performed to train the models PLS,RF,and GBM,and optimization was conducted using R package caret, in which the number of components in PLS, “mtry” in RF, as well as “n.trees”,“interaction.depth”, “shrinkage”, and “n.minobsinnode” in GBM,were tuned.For the SVM model,linear kernel was used and value of“cost” was screened from 1 to 10. To reduce model complexity,models with different amounts of top features,which were ranked in each model, were tested.Predictive accuracy in the test set was used to evaluate the predictive performance of models.

    2.7. Survival analysis for plasma metabolites in ESCC

    Kaplan-Meier curves were used to identify the relationships between metabolite levels in ESCC patients and their OS through log-rank test with a median split. Proportional hazard regression was performed for each metabolite to calculate the hazard ratio(HR) value. Factors with P values <0.05 were considered significantly prognostic.

    2.8. Statistical analysis

    Statistical analysis was performed using R software (version 3.6.2). Normality of the variables was tested by the Shapiro-Wilk normality test in R. Cox proportional hazard regression analysis was conducted using R package survival (version 3.1-8). ROC analysis was performed by R package pROC (version 1.15.3). Student's t-test was used to compare the means between two groups,whereas ANOVA test was used to compare the means among three or more groups. A two-tailed P value <0.05 was considered to be statistically significant.

    3. Results

    3.1. Metabolic shift in plasma of ESCC patients

    Un-targeted metabolomics was performed to investigate differential metabolites within the plasma of ESCC patients and healthy controls.A total of 3090 metabolic features in electrospray ionization positive (ESI+) mode and 3399 metabolic features in electrospray ionization negative (ESI-) mode were obtained from the metabolomics data. PCA analysis demonstrated a significant separation trend in plasma between ESCC patients and healthy controls, indicating a metabolic shift in ESCC plasma (Fig. 1).Furthermore, PLS-DA analysis demonstrated that ESCC patients were markedly separated from healthy controls,suggesting a global metabolic shift between the two groups (Fig. 2A). Volcano plots illustrating differential metabolomic features are shown in Fig.2B.A total of 840 differential metabolic features were obtained based on the criteria VIP >1, |log2fold change| >0.584, and FDR <0.05.Among these,12 features were annotated with specific metabolites(Table 2,Table S1).The heatmap demonstrated that 12 differential metabolites were able to distinguish ESCC patients from healthy controls (Fig. 2C). Pathway analysis of the 12 differential metabolites revealed that the top 3 significant metabolism pathways were synthesis and degradation of ketone bodies,butanoate metabolism,and lysine degradation (Fig. 2D).

    3.2. Predictive performance of single-metabolite models

    For single metabolite-based biomarker development, ROC curve analysis for metabolites in the training set showed that there were six metabolites with AUC values of over 0.85: indoxyl sulfate, phosphatidylcholine (PC) (14:0/0:0), sebacic acid, trimethylamine N-oxide,arachidonic acid,and deoxycholic acid(Fig.3).These six metabolites were further used to develop singlemetabolite-based diagnostic models for ESCC. After calculating the predictive classes for the unknown samples in the test set,confusion matrices were obtained, and the testing predictive performance of each metabolite is listed in Table 3. Arachidonic acid displayed the highest predictive accuracy (0.887, 95%CI:0.732-0.958), followed by sebacic acid (0.867, 95%CI:0.701-0.943), indoxyl sulfate (0.850, 95%CI: 0.701-0.942), PC(14:0/0:0) (0.825, 95%CI: 0.672-0.926), deoxycholic acid (0.773,95%CI: 0.644-0.910), and trimethylamine N-oxide (0.653, 95%CI:0.535-0.834).

    3.3. Predictive performance of multiple-metabolite-based machine learning models

    For the PLS model, the optimized number of components used in the model was 1, and the AUC of the ROC curve (AUCROC) was 0.981 (95%CI: 0.906-1.000) in the training set (Fig. 4A), and 0.973(95%CI:0.924-1.000)in test set(Fig.4E).The predictive accuracies in the training set and the test set were 0.955(95%CI:0.887-0.984)and 0.947 (95%CI: 0.830-0.994), respectively (Table 3 and Table S2).

    For the RF model, the optimized value for entry was 2, and the AUCROCwas 1.000(95%CI:0.906-1.000)in the training set(Fig.4B),and 0.997(95%CI:0.989-1.000) in test set(Fig.4F). The predictive accuracies in the training set and the test set were 1.000 (95%CI:0.964-1.000)and 0.947(95%CI:0.831-0.994),respectively(Table 3 and Table S2).

    Fig.2. Metabolic shift in plasma of ESCC patients compared with healthy controls.(A)PLS-DA score plot derived from partial least-squares discriminant analysis in the training set;(B) differential ion features were defined as VIP >1, |log2 FC| >0.584, and an FDR <0.05; (C) heatmap analysis of 12 plasma differential metabolites revealed a metabolic shift in ESCC patients compared with healthy controls;(D)pathway analysis of 12 differential metabolites.FC:fold change;FDR:Benjamini-Hochberg false discover rate;C:ESCC patient;N:healthy control.

    Table 2 Summary of the 12 differential metabolites found in plasma of ESCC patients and healthy controls a.

    Fig.3. Receiver operating characteristic(ROC)curves of single-metabolite models and boxplots of peak intensity distribution;(A)indoxyl sulfate,(B)PC(14:0/0:0),(C)sebacic acid,(D) trimethylamine N-oxide, (E) arachidonic acid, (F)deoxycholic acid. AUC:area under the curve.Two-tailed student's t-test was used with P value <0.05 considered significant.**** P <0.0001.

    Table 3 Predictive performance of different diagnostic models with the test seta.

    For the GBM model,the final optimized model had the following parameters: n.trees value of 150, interaction depth value of 2,shrinkage value of 0.1,n.minobsinnode value of 10,with the AUCROC1.000 (95%CI: 0.906-1.000) in the training set (Fig. 4C), and 1.000(95%CI:1.000-1.000)in test set(Fig.4G).The predictive accuracies in the training set and the test set were 1.000(95%CI:0.964-1.000)and 0.960 (95%CI: 0.830-0.994), respectively (Table 3 and Table S2).

    For the SVM model, linear SVM was finally selected with “Cclassification”as the type,“cost”value of 3,and 16 support vectors,with the AUCROCof 0.996 (95%CI: 0.866-1.000) in the training set(Fig. 4D), 0.979 (95%CI: 0.935-1.000) in test set (Fig. 4H). The predictive accuracies in the training set and the test set were 0.987(95%CI:0.946-0.999)and 0.980(95%CI:0.868-0.999),respectively(Table 3 and Table S2).

    In comparison to the single-metabolite models, the four metabolite-based machine learning models displayed higher predictive performance (Fig. 4J; Table 3), demonstrating the ascendancy of combined metabolomics data and machine learning approaches. Among the four machine learning models, SVM showed the highest accuracy of 0.980 in the test set among the four models.In terms of computational time(Fig.4I), the fastest model was SVM (0.05 s), followed by PLS (1.43 s), GBM (3.5 s), and RF(4.53 s). Taken together, four machine learning models, especially SVM,were all considered as promising diagnostic models for ESCC.

    3.4. Feature metabolite selection

    By ranking the feature importance of annotated metabolites via different machine learning models,it was reported from all models that several metabolites were of high importance, including indoxyl sulfate, arachidonic acid, and trimethylamine N-oxide. On the other hand, some were of low importance, including acetoacetic acid, pipecolic acid, and carnitine. Moreover, inter-model variations in feature importance of the same metabolite were noted.For instance,deoxycholic acid was reported of relatively high importance in PLS,RF,and SVM,while of relatively low importance in GBM (Fig. 5A).

    Fig. 5. (A) Feature importance of 12 metabolites in different machine learning models, and (B-E) machine learning models with different feature metabolites. The curves of predictive accuracy values increase as the number of feature metabolites grows in the (B) PLS model, (C) RF model, (D) GBM model, and (E) SVM model.

    To reduce model complexity, optimization of machine learning models can be achieved through the use of fewer variables. Thus,models with different amounts of top features were investigated(from top 2 to top 12). The results showed that three models (i.e.,PLS,RF,and GBM)reached accuracies of over 0.900 for the training set and over 0.850 for the test set, for the top 3 variables(Figs. 5B-D). A constant accuracy of 1.000 was reported in the training set of RFs,while in the corresponding test set a downward trend was observed when modeling with top 8-12 (Fig. 5C). The SVM model showed an escalating accuracy and achieved 0.987 for the training set and 0.980 for the test set (Fig. 5E).

    3.5. Prognostic value of plasma metabolites for ESCC

    To assess the prognostic value of the 12 differential metabolites,survival analysis was performed, and the results demonstrated acetoacetic acid to be negatively associated with OS for ESCC,with an HR of 1.752 (95%CI: 1.012-3.033) (Fig. 6A). PC (14:0/0:0) was positively related to OS, with an HR of 0.577(95%CI: 0.333-1.002)(Fig.6B).Two metabolites,acetoacetic acid and trimethylamine Noxide,were significantly increased in ESCC patients with advanced stages(TNM III-IV)compared with early stages(TNM I-II)(Figs.6C and D). Compared with healthy controls, ESCC patients had an evident increase in acetoacetic acid levels, while a significant decrease in trimethylamine N-oxide level.With regard to smoking status, the levels of 7Z, 10Z, 13Z, 16Z, and 19Z-docosapentaenoic acid were lower in smoking ESCC patients than in non-smoking ones, while the level of this metabolite was higher in ESCC patients than in healthy controls independent of their smoking status(Fig. 6E). Based on these results, plasma acetoacetic acid was an unfavorable prognostic factor for ESCC and might be related to the progression of ESCC.

    Fig.6. Prognostic significance of plasma metabolites.Kaplan-Meier survival curves for ESCC patients stratified by plasma metabolites with a median-split:(A)acetoacetic acid;(B)PC (14:0/0:0). Relative plasma concentrations of (C) acetoacetic acid, and (D) trimethylamine N-oxide among healthy controls, ESCC patients with early stages (I, II) and patients with advanced stages(III, IV). Relative plasma concentration of (E) 7Z,10Z,13Z,16Z,19Z-docosapentaenoic acid among healthy controls,smoking ESCC patients and non-smoking ones. Log-rank test was used with P value <0.05 considered significant. Two-tailed student's t-test and ANOVA test were used with P value <0.05 considered significant.

    4. Discussion

    ESCC patients survive longer when diagnosed at an early stage.Therefore, it is urgent to develop accurate and convenient diagnostic methods for early stage ESCC diagnosis. Previous metabolomic studies have demonstrated metabolic reprogramming to be a significant feature of ESCC, with the associated metabolites considered potential diagnostic biomarkers [2,14-17,23,27].Plasma/serum are the most common clinical fluid biopsies. These specimens are an excellent and relatively non-invasive source of precise,rapid,and real-time diagnostic biomarkers[28].Previously,several plasma/serum metabolomic studies[15,23,27]have found a group of metabolites differentially present in ESCC patients compared with healthy controls.For example,Cheng et al.[2]have found an increase in tryptophan metabolites including kynurenine,5-hydroxytryptamine, 5-hydroxytryptophan, and 5-hydroxyindole-3-acetic acid in ESCC serum. Mir et al. [15] have revealed a dysregulation of serum PC in ESCC patients.Liu et al.[23]have demonstrated six plasma phospholipids, phosphatidylserine,phosphatidic acid, phosphocholine, phosphatidylinositol, phosphatidylethanolamine, and sphinganine 1-phosphate, to be significantly up-regulated in ESCC. However, these studies have limitations as follows: diagnostic significance of the metabolites was not clearly elucidated by proper validation design, such as splitting the data set into a training set and a test set; current multivariate analysis methods used for metabolomics data are PCA and PLS-DA (one form of PLS when Y is categorical), which can result in classifications that are over-optimistic or over-fitting. In order to have an in-depth understanding of the diagnostic significance of ESCC plasma metabolites and to enhance the clinical application of metabolomics, the present study developed and assessed metabolite-based machine learning models to discriminate between plasma samples of ESCC patients and healthy controls.

    Four machine learning algorithms, PLS, RF, GBM, and SVM, are all widely used in the healthcare field, particularly in the area of medical diagnosis. However, with the exception of PLS, the other three machine learning algorithms have not been fully investigated for metabolomic data analysis yet. Based on our results, the metabolite-based machine learning models used in this study showed satisfactory predictive performance (accuracy range of 0.947-0.980), which was significantly higher than that of singlemetabolite-based models (accuracy range of 0.653-0.887). The SVM exhibited highest predictive performance among the four models, both in the training set(0.987) and in the test set (0.980).Meanwhile, it had the lowest computational cost, altogether indicating SVM may be most suitable for analysis of large metabolomics data sets. Taken together, this study demonstrated machine learning methods other than PLS to be useful for clinical metabolomics studies, encouraging the use of combined metabolomics and machine learning approaches for the development of diagnostic cancer tools.

    In the present study, a significant relationship was observed between TNM stage and acetoacetic acid, which was herein evidenced as the most prominent metabolite, possessing both diagnostic and prognostic value. Acetoacetic acid was originally considered as a ketone body, mainly produced in the liver during periods of nutrient deprivation,that served as high-energy fuel for extrahepatic tissues like the brain,heart, and skeletal muscle[29].Consistent with our results, other studies have reported upregulated ketone bodies in ESCC cancerous tissues [13]. These were characterized as an accumulation of ketone bodies (acetone and acetoacetic acid) as well as up-regulated ketone transportermonocarboxylate transporter 1 in ESCC [13,30]. Taken together,these results suggested a potential functional role for acetoacetic acid in ESCC. However, contradicting results regarding ketones in cancer also exist.For example,Poff et al.[31]claimed cancer cells to be unable to efficiently utilize ketones, while ketones slow the proliferation of tumor cells. Martinez-Outschoorn et al. [32] illustrated an opposite effect of ketones which increased the stemness of cancer cells,resulting in recurrence,metastasis,and poor clinical outcomes in breast cancer. Therefore, it is essential to clarify whether ESCC cancer cells utilize ketones as an energy resource as well as to determine the biological role of acetoacetic acid in ESCC.

    Lipids are essential to cancer cell structure,signal transduction,and cancer cell proliferation [33-35]. Our results showed a significantly decreased PC(14:0/0:0)level in ESCC,which is favorable to this cancer. Similarly, Mir et al. [15] also detected a group of decreased serum phosphatidylcholines, such as PC (18:2/0:0), PC(18:1/18:2),and PC(20:4/0:0)in ESCC.Meanwhile,Kamphorst et al.[34] proposed the capability of cancer cells in lipid uptake and utilization from the circulation through macro-pinocytosis.Collectively, these findings suggested that alterations in circulating lipids may be associated with enhanced lipid consumption by cancer cells.It is herein evident that lower circulating PC(14:0/0:0) levels, corresponding to higher consumption of PC (14:0/0:0)by cancer cells, are related to poorer survival of ESCC.

    Our results exhibited an indicative decrease in indoxyl sulfate.The highest AUC value in ROC analysis was observed for this metabolite in the training set when performing single-metabolitebased diagnostic model analysis(AUC =0.917).Its AUC value in test set, though not the greatest, was also very high. Moreover, great importance of indoxyl sulfate was observed in all multiplemetabolite-based diagnostic models combined with machine learning models. These results thus suggest that indoxyl sulfate might be a promising diagnostic biomarker for ESCC.In addition,a previous study revealed that indoxyl sulfate is related to microbial tryptophan catabolism [36]. In accordance with these results,Cheng et al.[2]reported disturbed tryptophan metabolism in ESCC.Altogether, tryptophan metabolism, especially microbial tryptophan catabolism, is potentially associated with ESCC initiation or progression.

    Deoxycholic acid is a secondary bile acid, the metabolic byproduct of intestinal bacteria. It is known to increase intracellular production of reactive oxygen as well as reactive nitrogen species,and higher levels are associated with increased frequencies of colon cancer[37-39].Additionally,deoxycholic acid has multifunctional biological activities, such as disrupting the intestinal mucosal barrier [40] and enhancing Wnt signaling [41]. However, there have been no reports about deoxycholic acid in ESCC yet; thus it is worthy to further investigate its biological functions.

    7Z,10Z,13Z,16Z,19Z-docosapentaenoic acid is an Omega-3 polyunsaturated fatty acid with 5 double bonds in 7-,10-,13-,16-,19-positions. The present study detected a significant increase in 7Z,10Z,13Z,16Z,19Z-docosapentaenoic acid levels in ESCC patients,which indicates a diagnostic potential for ESCC. Consistently, Liu et al. [42] have recently reported that people with high docosapentaenoic acid levels are vulnerable to lung cancer, indicating a biological role of docosapentaenoic acid in cancer initiation and development. Therefore, more work is needed to investigate the potential mechanisms of docosapentaenoic acid in ESCC.

    This study also presented conflicting results: down-regulated trimethylamine N-oxide was found in ESCC patients at early stage compared with healthy controls,while an accretion was denoted at advanced stage in comparison to early stage,though the levels were still lower than those of healthy controls. A plausible explanation for such outcome is that levels of trimethylamine-N-oxide are determined by two factors: trimethylamine production from precursor molecules such as choline and L-carnitine by the metabolism of gut microbiota; and dietary intake of trimethylamine-N-oxiderich foods such as high-choline or high-carnitine diet [43].Accordingly, plasma levels of trimethylamine-N-oxide in ESCC patients might be altered by both changes in dietary compositions and intestinal bacteria. Meanwhile, controversial results are suggested in previous studies. Bae et al. [44] postulated a positive correlation between incidence of colorectal cancer (CRC) and plasma levels of trimethylamine N-oxide in US women, while Guertin et al.[45]indicated no correlation between this metabolite and risks of CRC. It remains opaque, thus requiring further researches to investigate, whether an increase in trimethylamine Noxide level is a cause or a consequence of cancer.

    Limitations of this study must be addressed. First, the sample size was relatively small for machine learning algorithms. A larger cohort is needed to validate model performance and finely optimize model parameters. Second, metabolite annotation efficiency was relatively low due to a lack of an in-house database and online MS2spectral data,resulting in many diagnostic metabolite ions not being annotated.Third,the current prediction capacity of machine learning models is limited to two classes of plasma samples, so more diverse samples are needed to improve future performance of machine learning algorithms. Last but not least, the detailed function of the identified metabolites, such as acetoacetic acid, is required to be further clarified.

    In conclusion, this study successfully established plasma metabolite-based machine learning models to distinguish ESCC cancer patients from healthy controls, demonstrating that the combination of metabolomics and machine learning is a novel and efficient diagnostic strategy for ESCC and possibly for other cancers.Although this study was a pilot in nature, due to relatively small sample size and limited diversity within the training set,the results could encourage future applications of machine learning algorithms to clinical metabolomics studies and accordingly aid medical diagnostic development. In addition to the discovery of diagnostic metabolites,this study explored progression-associated metabolites, which may provide potential prognostic biomarkers for ESCC.The findings of this study contribute to an understanding of the molecular pathogenesis of ESCC and provide useful information for individualized cancer therapy. In summary, further studies with larger cohorts should be conducted through a combined application of metabolomics and machine learning. This approach is promising in cancer diagnosis and will contribute significantly to cancer treatment.

    Declaration of competing interest

    The authors declare that there are no conflicts of interest.

    Acknowledgments

    This work was supported by the National Natural Science Foundation of China (Grant Nos. 81672315, 81802276, and 81302840),Key R&D Program Projects in Zhejiang Province(Grant No. 2018C04009), and 1022 Talent Training Program of Zhejiang Cancer Hospital.

    Appendix A. Supplementary data

    Supplementary data to this article can be found online at https://doi.org/10.1016/j.jpha.2020.11.009.

    亚洲avbb在线观看| 亚洲av成人不卡在线观看播放网| 亚洲久久久国产精品| 纵有疾风起免费观看全集完整版| 久久久国产一区二区| 午夜91福利影院| 久久久久久久大尺度免费视频| 日韩精品免费视频一区二区三区| 男女之事视频高清在线观看| 精品国产一区二区久久| 热re99久久国产66热| 丝袜在线中文字幕| 宅男免费午夜| 久久午夜亚洲精品久久| 亚洲色图综合在线观看| www.自偷自拍.com| 国产精品自产拍在线观看55亚洲 | av又黄又爽大尺度在线免费看| 精品卡一卡二卡四卡免费| 在线观看www视频免费| 久久精品国产亚洲av香蕉五月 | 777米奇影视久久| 欧美乱码精品一区二区三区| 丁香欧美五月| 肉色欧美久久久久久久蜜桃| 久久精品国产99精品国产亚洲性色 | 国产无遮挡羞羞视频在线观看| 考比视频在线观看| 日韩大码丰满熟妇| 色视频在线一区二区三区| 一级黄色大片毛片| 免费在线观看完整版高清| 亚洲av国产av综合av卡| 我的亚洲天堂| 国产无遮挡羞羞视频在线观看| 久久av网站| 99riav亚洲国产免费| 免费久久久久久久精品成人欧美视频| 成年人午夜在线观看视频| 国产日韩欧美在线精品| 日韩人妻精品一区2区三区| 午夜激情久久久久久久| 亚洲精品成人av观看孕妇| 日韩欧美一区二区三区在线观看 | 国产黄色免费在线视频| 50天的宝宝边吃奶边哭怎么回事| 久久久久久久国产电影| 极品教师在线免费播放| 午夜激情久久久久久久| 亚洲成人免费电影在线观看| 老汉色av国产亚洲站长工具| 精品一区二区三区四区五区乱码| 成人影院久久| 亚洲 欧美一区二区三区| 高潮久久久久久久久久久不卡| 亚洲欧美一区二区三区黑人| 国产成人免费观看mmmm| 国产精品国产高清国产av | 男女免费视频国产| 法律面前人人平等表现在哪些方面| 美女国产高潮福利片在线看| 他把我摸到了高潮在线观看 | 老熟妇仑乱视频hdxx| 老汉色av国产亚洲站长工具| e午夜精品久久久久久久| 91九色精品人成在线观看| 精品亚洲乱码少妇综合久久| 国产成人精品久久二区二区91| 老司机午夜福利在线观看视频 | 露出奶头的视频| 无人区码免费观看不卡 | cao死你这个sao货| 欧美大码av| 中文亚洲av片在线观看爽 | 欧美精品一区二区免费开放| 男男h啪啪无遮挡| www.自偷自拍.com| 51午夜福利影视在线观看| 操出白浆在线播放| 熟女少妇亚洲综合色aaa.| 咕卡用的链子| 国产精品香港三级国产av潘金莲| 国产精品偷伦视频观看了| 国产人伦9x9x在线观看| 老汉色∧v一级毛片| 国产欧美亚洲国产| 久久久久国产一级毛片高清牌| 99国产精品一区二区三区| 国产不卡一卡二| 欧美精品人与动牲交sv欧美| 午夜免费鲁丝| 免费看a级黄色片| 精品国产乱码久久久久久男人| 在线av久久热| 亚洲欧美日韩另类电影网站| 19禁男女啪啪无遮挡网站| 国产欧美日韩综合在线一区二区| 欧美性长视频在线观看| 久久中文看片网| 日本av免费视频播放| av国产精品久久久久影院| 天天躁夜夜躁狠狠躁躁| 精品久久久久久久毛片微露脸| 成人国语在线视频| 免费高清在线观看日韩| 大码成人一级视频| 建设人人有责人人尽责人人享有的| 国产一卡二卡三卡精品| 99热国产这里只有精品6| 大片免费播放器 马上看| 97在线人人人人妻| 久久亚洲真实| 国产精品电影一区二区三区 | 国产av精品麻豆| 欧美一级毛片孕妇| 动漫黄色视频在线观看| 天天添夜夜摸| 久久久精品94久久精品| 一区二区三区激情视频| 亚洲人成77777在线视频| 成人国产av品久久久| videosex国产| 午夜激情久久久久久久| 午夜视频精品福利| 在线看a的网站| 高清在线国产一区| 嫩草影视91久久| 日韩欧美免费精品| 久久国产精品人妻蜜桃| 精品亚洲成国产av| 嫁个100分男人电影在线观看| 色视频在线一区二区三区| 成人国产一区最新在线观看| 69av精品久久久久久 | videosex国产| 成人18禁在线播放| 韩国精品一区二区三区| 日韩人妻精品一区2区三区| 久久久久久久久免费视频了| 日韩欧美国产一区二区入口| 麻豆av在线久日| 精品少妇久久久久久888优播| 丝瓜视频免费看黄片| 免费观看av网站的网址| √禁漫天堂资源中文www| a级毛片黄视频| www.精华液| 青青草视频在线视频观看| 国产精品久久久久成人av| 超碰97精品在线观看| 久久国产精品男人的天堂亚洲| 成年人免费黄色播放视频| 桃红色精品国产亚洲av| 91麻豆av在线| 成人精品一区二区免费| 国产精品久久久久久精品古装| 考比视频在线观看| 大型av网站在线播放| 亚洲成a人片在线一区二区| 国产精品电影一区二区三区 | 欧美日韩精品网址| 国产淫语在线视频| 狠狠狠狠99中文字幕| 成人三级做爰电影| 欧美日韩亚洲高清精品| cao死你这个sao货| 丝袜美足系列| 蜜桃在线观看..| 日本wwww免费看| 女性被躁到高潮视频| 人人妻人人爽人人添夜夜欢视频| 精品久久久久久电影网| 亚洲成人免费电影在线观看| 一夜夜www| 久久国产精品人妻蜜桃| 啦啦啦免费观看视频1| 汤姆久久久久久久影院中文字幕| 国产成+人综合+亚洲专区| av视频免费观看在线观看| 精品熟女少妇八av免费久了| 欧美激情久久久久久爽电影 | 精品国产乱码久久久久久男人| 精品人妻在线不人妻| 中文字幕av电影在线播放| 麻豆成人av在线观看| 男女床上黄色一级片免费看| 亚洲国产av影院在线观看| 97人妻天天添夜夜摸| 欧美激情极品国产一区二区三区| 正在播放国产对白刺激| 国产一区有黄有色的免费视频| 久久精品成人免费网站| 成人永久免费在线观看视频 | www.精华液| 中文字幕制服av| 最近最新中文字幕大全免费视频| 久久青草综合色| 久久国产精品男人的天堂亚洲| 欧美成狂野欧美在线观看| 91九色精品人成在线观看| 久久天堂一区二区三区四区| 国产成人一区二区三区免费视频网站| 成年版毛片免费区| 极品教师在线免费播放| 丰满迷人的少妇在线观看| 精品人妻1区二区| 久久精品成人免费网站| 色婷婷久久久亚洲欧美| 亚洲人成77777在线视频| 男女高潮啪啪啪动态图| 日本欧美视频一区| 老司机午夜十八禁免费视频| 不卡一级毛片| 亚洲av第一区精品v没综合| 91老司机精品| 欧美黑人精品巨大| 欧美黄色淫秽网站| 男人舔女人的私密视频| 亚洲五月婷婷丁香| 九色亚洲精品在线播放| 性色av乱码一区二区三区2| 9色porny在线观看| 日韩精品免费视频一区二区三区| 欧美黄色淫秽网站| 老熟妇仑乱视频hdxx| 老鸭窝网址在线观看| 在线观看免费高清a一片| 80岁老熟妇乱子伦牲交| 首页视频小说图片口味搜索| 超碰97精品在线观看| 满18在线观看网站| 香蕉久久夜色| 成人免费观看视频高清| 午夜免费鲁丝| a级片在线免费高清观看视频| 日本撒尿小便嘘嘘汇集6| 丰满饥渴人妻一区二区三| 性色av乱码一区二区三区2| 精品少妇久久久久久888优播| av在线播放免费不卡| 国产精品美女特级片免费视频播放器 | 一个人免费在线观看的高清视频| 国产成人av激情在线播放| 啦啦啦免费观看视频1| 欧美日韩福利视频一区二区| 不卡一级毛片| 嫁个100分男人电影在线观看| 我要看黄色一级片免费的| 亚洲成人国产一区在线观看| 亚洲精品国产一区二区精华液| 男人舔女人的私密视频| 久久香蕉激情| 欧美精品亚洲一区二区| 男女免费视频国产| 精品少妇内射三级| 久久久水蜜桃国产精品网| 亚洲avbb在线观看| 久久国产精品男人的天堂亚洲| 一进一出抽搐动态| 丝袜在线中文字幕| 欧美成狂野欧美在线观看| 手机成人av网站| 90打野战视频偷拍视频| 国产精品电影一区二区三区 | 亚洲五月婷婷丁香| 侵犯人妻中文字幕一二三四区| 肉色欧美久久久久久久蜜桃| 在线观看舔阴道视频| 十八禁网站网址无遮挡| 久久人妻av系列| 国产又色又爽无遮挡免费看| 久久久久视频综合| av线在线观看网站| bbb黄色大片| 麻豆乱淫一区二区| 国产精品 国内视频| 91九色精品人成在线观看| 国产在线视频一区二区| 久久人人爽av亚洲精品天堂| 黄片小视频在线播放| 高清av免费在线| 老熟女久久久| 久久久精品国产亚洲av高清涩受| 天天躁夜夜躁狠狠躁躁| 久久精品亚洲av国产电影网| 建设人人有责人人尽责人人享有的| 国产成人精品无人区| 亚洲国产欧美网| 91精品三级在线观看| √禁漫天堂资源中文www| 日韩成人在线观看一区二区三区| 99热网站在线观看| 精品人妻1区二区| 久久香蕉激情| 欧美 日韩 精品 国产| 高清毛片免费观看视频网站 | 丰满人妻熟妇乱又伦精品不卡| tube8黄色片| 99热国产这里只有精品6| 精品国产超薄肉色丝袜足j| 天天操日日干夜夜撸| 精品久久久精品久久久| 天天添夜夜摸| netflix在线观看网站| 在线天堂中文资源库| 国产一区二区 视频在线| 人妻 亚洲 视频| 国产成人欧美在线观看 | 国内毛片毛片毛片毛片毛片| 国产免费视频播放在线视频| 精品少妇黑人巨大在线播放| 亚洲情色 制服丝袜| 男女无遮挡免费网站观看| 欧美黄色淫秽网站| tube8黄色片| aaaaa片日本免费| 欧美 日韩 精品 国产| 久久性视频一级片| 亚洲中文字幕日韩| 久久九九热精品免费| 黄色成人免费大全| 久久久久国内视频| 黄网站色视频无遮挡免费观看| 精品福利永久在线观看| 欧美精品亚洲一区二区| 亚洲一区二区三区欧美精品| 国产视频一区二区在线看| 亚洲av欧美aⅴ国产| 免费日韩欧美在线观看| 久久精品国产亚洲av高清一级| 天天躁日日躁夜夜躁夜夜| 性高湖久久久久久久久免费观看| 久久青草综合色| 亚洲成av片中文字幕在线观看| 91大片在线观看| 日韩欧美一区二区三区在线观看 | 国产精品电影一区二区三区 | 亚洲人成电影免费在线| 日韩免费av在线播放| 久久香蕉激情| 伊人久久大香线蕉亚洲五| 欧美成狂野欧美在线观看| 一边摸一边做爽爽视频免费| 成在线人永久免费视频| 久久久精品94久久精品| 动漫黄色视频在线观看| 国产免费现黄频在线看| 日韩中文字幕视频在线看片| 99热网站在线观看| 欧美精品亚洲一区二区| 国产精品九九99| 熟女少妇亚洲综合色aaa.| 中文字幕制服av| 啪啪无遮挡十八禁网站| 丁香六月欧美| 男女无遮挡免费网站观看| 国产极品粉嫩免费观看在线| 欧美激情久久久久久爽电影 | 亚洲天堂av无毛| 午夜91福利影院| 两个人免费观看高清视频| 操出白浆在线播放| 亚洲av成人不卡在线观看播放网| 国产成人免费观看mmmm| 午夜福利视频在线观看免费| 欧美成人免费av一区二区三区 | 天堂动漫精品| av国产精品久久久久影院| 日本黄色日本黄色录像| 免费一级毛片在线播放高清视频 | 99国产精品一区二区三区| 性色av乱码一区二区三区2| 老司机深夜福利视频在线观看| 狠狠婷婷综合久久久久久88av| 我要看黄色一级片免费的| 国产伦人伦偷精品视频| 女性被躁到高潮视频| 纵有疾风起免费观看全集完整版| 变态另类成人亚洲欧美熟女 | 91字幕亚洲| 在线观看免费午夜福利视频| av线在线观看网站| 可以免费在线观看a视频的电影网站| 少妇精品久久久久久久| 成人三级做爰电影| 他把我摸到了高潮在线观看 | 国产一区二区三区在线臀色熟女 | 久久国产亚洲av麻豆专区| av有码第一页| 一区二区三区乱码不卡18| 免费观看人在逋| 高清欧美精品videossex| 国产免费福利视频在线观看| 午夜激情av网站| 亚洲精品自拍成人| 欧美日韩黄片免| 久久香蕉激情| 国产男靠女视频免费网站| 成人国语在线视频| 欧美黑人精品巨大| 国产精品一区二区免费欧美| 亚洲精品美女久久久久99蜜臀| 精品少妇久久久久久888优播| 亚洲精品粉嫩美女一区| 王馨瑶露胸无遮挡在线观看| 天堂中文最新版在线下载| www日本在线高清视频| 国产成人精品无人区| 黄色成人免费大全| 亚洲熟女精品中文字幕| 国产精品影院久久| 欧美一级毛片孕妇| 蜜桃在线观看..| 午夜免费鲁丝| 一个人免费在线观看的高清视频| 热99国产精品久久久久久7| 精品福利永久在线观看| 国产亚洲一区二区精品| 国产欧美日韩综合在线一区二区| 亚洲美女黄片视频| 熟女少妇亚洲综合色aaa.| 另类亚洲欧美激情| 50天的宝宝边吃奶边哭怎么回事| 人人妻人人澡人人爽人人夜夜| 精品久久久精品久久久| 夜夜爽天天搞| 超碰97精品在线观看| 精品国产乱码久久久久久小说| av天堂在线播放| 少妇的丰满在线观看| 成年动漫av网址| 制服人妻中文乱码| 国产精品.久久久| 欧美日韩国产mv在线观看视频| 亚洲欧美一区二区三区黑人| 国产亚洲av高清不卡| 制服诱惑二区| 极品少妇高潮喷水抽搐| 亚洲精品乱久久久久久| 久久久欧美国产精品| 青青草视频在线视频观看| 大香蕉久久网| 99国产精品一区二区三区| 中亚洲国语对白在线视频| 亚洲成人免费电影在线观看| 亚洲中文av在线| 国产精品电影一区二区三区 | 久久久久视频综合| 国产不卡av网站在线观看| 日本av手机在线免费观看| 亚洲第一青青草原| 18禁观看日本| 99riav亚洲国产免费| 又黄又粗又硬又大视频| 女人久久www免费人成看片| 精品一区二区三区av网在线观看 | 精品久久久久久久毛片微露脸| 亚洲 欧美一区二区三区| 美女高潮喷水抽搐中文字幕| √禁漫天堂资源中文www| 丝瓜视频免费看黄片| 别揉我奶头~嗯~啊~动态视频| 在线观看人妻少妇| 啪啪无遮挡十八禁网站| 午夜福利影视在线免费观看| 欧美日韩中文字幕国产精品一区二区三区 | 另类精品久久| 国产欧美日韩精品亚洲av| 一级片'在线观看视频| 悠悠久久av| 交换朋友夫妻互换小说| 国产成+人综合+亚洲专区| 久久国产精品大桥未久av| 真人做人爱边吃奶动态| 黄色视频不卡| 人人妻人人添人人爽欧美一区卜| 蜜桃在线观看..| 大片电影免费在线观看免费| 久久精品国产a三级三级三级| 欧美成狂野欧美在线观看| 国产精品免费视频内射| 中文字幕另类日韩欧美亚洲嫩草| 王馨瑶露胸无遮挡在线观看| 国产精品.久久久| 在线观看免费高清a一片| 午夜福利免费观看在线| 久久国产亚洲av麻豆专区| 欧美激情极品国产一区二区三区| 热99re8久久精品国产| 夫妻午夜视频| 久久久国产欧美日韩av| 日本一区二区免费在线视频| 日韩一区二区三区影片| 男女边摸边吃奶| 欧美精品亚洲一区二区| 狠狠狠狠99中文字幕| 一区二区三区乱码不卡18| 大陆偷拍与自拍| 亚洲欧洲精品一区二区精品久久久| 黄色a级毛片大全视频| 十八禁高潮呻吟视频| 国产极品粉嫩免费观看在线| 精品国产亚洲在线| 日韩人妻精品一区2区三区| a级毛片黄视频| 国产成人精品无人区| 亚洲欧美一区二区三区久久| 露出奶头的视频| 免费在线观看日本一区| 国产不卡av网站在线观看| 亚洲精品国产一区二区精华液| 人人妻,人人澡人人爽秒播| 日本wwww免费看| 久久婷婷成人综合色麻豆| 国产成人精品无人区| 女性被躁到高潮视频| 国产伦人伦偷精品视频| 91九色精品人成在线观看| 视频区欧美日本亚洲| 亚洲色图综合在线观看| 91精品三级在线观看| 精品少妇久久久久久888优播| 十八禁网站网址无遮挡| 日日爽夜夜爽网站| 90打野战视频偷拍视频| 午夜91福利影院| 国内毛片毛片毛片毛片毛片| 老司机深夜福利视频在线观看| 在线看a的网站| 水蜜桃什么品种好| a在线观看视频网站| 三上悠亚av全集在线观看| 成人国产av品久久久| av网站免费在线观看视频| 建设人人有责人人尽责人人享有的| 久9热在线精品视频| 狠狠精品人妻久久久久久综合| 操出白浆在线播放| 亚洲伊人色综图| 在线av久久热| 欧美在线黄色| 国产片内射在线| www.999成人在线观看| 久久 成人 亚洲| 大陆偷拍与自拍| 99国产精品一区二区三区| 亚洲黑人精品在线| 国产视频一区二区在线看| 欧美日韩视频精品一区| 午夜精品国产一区二区电影| 69精品国产乱码久久久| 欧美+亚洲+日韩+国产| 两人在一起打扑克的视频| 国产精品久久久av美女十八| av网站在线播放免费| 国产欧美日韩一区二区三| 人人澡人人妻人| 国产三级黄色录像| 亚洲午夜理论影院| 在线看a的网站| 又紧又爽又黄一区二区| 欧美乱妇无乱码| 男人舔女人的私密视频| 99热国产这里只有精品6| 嫩草影视91久久| 极品人妻少妇av视频| 亚洲专区中文字幕在线| 国产精品美女特级片免费视频播放器 | 久久人妻熟女aⅴ| 欧美日韩精品网址| 一区二区三区精品91| 精品国产超薄肉色丝袜足j| 操出白浆在线播放| 免费看a级黄色片| 亚洲avbb在线观看| 亚洲av美国av| 一区二区三区国产精品乱码| 欧美激情高清一区二区三区| 欧美日韩精品网址| 久久九九热精品免费| www.自偷自拍.com| 欧美日韩亚洲高清精品| 老司机午夜十八禁免费视频| 日韩视频在线欧美| 国产高清国产精品国产三级| 亚洲精品粉嫩美女一区| 一本久久精品| 天天操日日干夜夜撸| 日本黄色视频三级网站网址 | av又黄又爽大尺度在线免费看| 一区福利在线观看| 亚洲欧洲精品一区二区精品久久久| 一本一本久久a久久精品综合妖精| 亚洲色图综合在线观看| 午夜福利免费观看在线| netflix在线观看网站| 欧美激情高清一区二区三区| 午夜福利影视在线免费观看| 香蕉丝袜av| 欧美激情久久久久久爽电影 | 国产黄频视频在线观看| 午夜视频精品福利| 99热网站在线观看| 午夜福利乱码中文字幕| 久久久久久久久免费视频了| 人人妻人人澡人人爽人人夜夜| 国产老妇伦熟女老妇高清| 国产精品成人在线| 人人妻,人人澡人人爽秒播| 国产男女内射视频| 高清视频免费观看一区二区| 大片免费播放器 马上看| 夜夜爽天天搞| 美女视频免费永久观看网站| 亚洲国产av新网站|