• <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 ?

    Procleave:Predicting Protease-specific Substrate Cleavage Sites by Combining Sequence and Structural Information

    2020-07-29 05:34:42FuyiLiAndreLeierQuanzhongLiuYananWangDongxuXiangTatsuyaAkutsuGeoffreyWebbIanSmithTatianaMarquezLagoJianLiJiangningSong
    Genomics,Proteomics & Bioinformatics 2020年1期

    Fuyi Li ,Andre Leier ,Quanzhong Liu ,Yanan Wang ,Dongxu Xiang 1,,Tatsuya Akutsu ,Geoffrey I.Webb ,A.Ian Smith ,Tatiana Marquez-Lago ,Jian Li ,Jiangning Song ,6,*

    1 Biomedicine Discovery Institute and Department of Biochemistry and Molecular Biology,Monash University,Melbourne,VIC 3800,Australia

    2 Monash Centre for Data Science,Faculty of Information Technology,Monash University,Melbourne,VIC 3800,Australia

    3 School of Medicine,University of Alabama at Birmingham,Birmingham,AL 35233,USA

    4 College of Information Engineering,Northwest A&F University,Yangling 712100,China

    5 Bioinformatics Center,Institute for Chemical Research,Kyoto University,Uji,Kyoto 611-0011,Japan

    6 ARC Centre of Excellence in Advanced Molecular Imaging,Monash University,Melbourne,VIC 3800,Australia

    7 Biomedicine Discovery Institute and Department of Microbiology,Monash University,Melbourne,VIC 3800,Australia

    KEYWORDS Protease;Cleavage site prediction;Machine learning;Conditional random field;Structural determinants

    Abstract Proteases are enzymes that cleave and hydrolyse the peptide bonds between two specific amino acid residues of target substrate proteins.Protease-controlled proteolysis plays a key role in the degradation and recycling of proteins,which is essential for various physiological processes.Thus,solving the substrate identification problem will have important implications for the precise understanding of functions and physiological roles of proteases,as well as for therapeutic target identification and pharmaceutical applicability.Consequently,there is a great demand for bioinformatics methods that can predict novel substrate cleavage events with high accuracy by utilizing both sequence and structural information.In this study,we present Procleave,a novel bioinformatics approach for predicting protease-specific substrates and specific cleavage sites by taking into account both their sequence and 3D structural information.Structural features of known cleavage sites were represented by discrete values using a LOWESS data-smoothing optimization method,which turned out to be critical for the performance of Procleave.The optimal approximations of all structural parameter values were encoded in a conditional random field(CRF)computational framework,alongside sequence and chemical group-based features.Here,we demonstrate the outstanding performance of Procleave through extensive benchmarking and independent tests.Procleave is capable of correctly identifying most cleavage sites in the case study.Importantly,when applied to the human structural proteome encompassing 17,628 protein structures,Procleave suggests a number of potential novel target substrates and their corresponding cleavage sites of different proteases.Procleave is implemented as a webserver and is freely accessible at http://procleave.erc.monash.edu/.

    Introduction

    Protease-specific cleavage is a ubiquitous type of irreversible post-translational modification(PTM)that occurs when proteases specifically cleave the peptide bonds between the P1 and P1′sites of target proteins or peptide substrates[1].Numerous experimental studies indicate that proteolytic cleavage plays a critical role in a variety of developmental and physiological processes,including cell cycle,pathway regulation,and protein degradation.On the other hand,the dysregulation of proteases is associated with numerous diseases[2].Thus,it is very important to identify protease-specific substrate cleavage sites,as such knowledge can provide deeper insights into the mechanisms and biological functions of proteases,which in turn might lead to novel therapeutic targets and pharmaceutical applicability.However,current existing experimental methods for protease substrate cleavage site identification are expensive,labourintensive,and time-consuming.Therefore,the development of cost-effective computational approaches for precise prediction of protease-specific proteolytic events is very important.Such tools can not only provide high-quality predictions of target substrates for a specific protease,but also guide hypothesisdriven experimental efforts to identify substrate specificity and associated biological functions of proteases.

    Due to the importance and the benefits of computational predictions of protease-specific target substrates,over the past two decades,more than 20 computational methods have been proposed[3,4].In our recent review paper,we categorized these methods into two major groups according to the employed methodologies:(i)sequence-scoring function-based methods,such as PoPS[5],SitePrediction[6],and CAT3[7],and(ii)machine learning methods,such as Pripper[8],Cascleave[9],PROSPER[10],LabCaS[11],ScreenCap3[12],Cascleave 2 [13], iProt-Sub [14], and PROSPERous [15]. These publicly available computational tools have successfully guided experiments in finding novel cleavage sites and obtaining a better understanding of protease-substrate interactions.

    A number of encouraging studies have been done regarding the development of computational methods and tools for predicting protease-specific cleavage sites.However,all of these existing prediction methods are developed based on protein sequences and they are only used for predicting the cleavage sites from substrate sequences.Previous studies have shown that protease cleavage sites are primarily distributed in loop regions of the substrate proteins,while cleavage within other structural regions of substrate proteins,such as α-helices and β-sheets,is also possible[16-18].These findings indicate protease substrate cleavage specificity at the secondary structure(SS)level.The majority of existing predictors did not consider the structure-level preference and parameters, which can potentially improve the prediction performance and also help better understand the biological functions of proteases.

    In this study,we introduce Procleave to fill the knowledge gap outlined above and enhance protease substrate cleavage site prediction by incorporating 3D structural features of substrate cleavage segments.More specifically,Procleave uses the data curated from the MEROPS database[19]and maps substrate sequences to PDB structures by performing BLAST search,thereby generating an extensive 3D structural substrate dataset.Multi-faceted sequence and structural features are then extracted,which are further integrated into a novel conditional random field (CRF) algorithm with a datasmoothing framework to train cleavage site prediction models.A comprehensive performance test confirms that smoothed structural features combined with sequence-based features can greatly improve the prediction performance.Subsequently,we implement a webserver for 27 major proteases,taking advantage of the findings in this study,and make it publicly accessible.

    Method

    Overall framework

    Figure 1provides an overview of the Procleave framework.Five major steps are involved in the construction and evaluation of Procleave.At the first step,i.e.,data collection and pre-processing,the benchmark training and independent test datasets were collected from MEROPS[19].At the second step,multi-faceted sequence features and 3D structure features were generated.At the third step,a novel integrative CRF framework was developed for model training and optimization.At the fourth step,the trained CRF models were further evaluated and validated by performing the independent test.A performance comparison with currently existing methods was also conducted.At the final step,the Procleave webserver was implemented to facilitate public use.

    Dataset collection and pre-processing

    Figure 1 The overall framework of Procleave

    The experimentally verified protein substrate cleavage annotations for training and benchmarking Procleave were extracted from the MEROPS database(Release 9.0)[19].MEROPS is a public resource and knowledgebase for experimentally validated protease substrates and cleavage sites,which is accessible via https://www.ebi.ac.uk/merops/.To develop reliable prediction models and objectively evaluation the model performance,we discarded highly homologous sequences from the initial substrate dataset with a sequence identity(SI)threshold of 70%between any two substrate protein sequences.This avoids overestimating the prediction performance in cross-validation tests.It is noticeable that a number of existing studies used SI cut-off values of 70%[9,14,15]or a higher,e.g.,80%[12].The MEROPS database was recently updated(Release 12.0,26-April-2019)and we decided to use all the newly added protease substrates and cleavage sites as the independent test dataset to assess the performance of trained Procleave models and conduct the performance comparison with existing methods.In addition,in order to perform a more fairly independent test,we used a stricter SI threshold(30%)to remove the sequence redundancy in the independent test dataset.CD-HIT[20]was applied to remove the redundant sequences between the independent test datasets and training datasets at the SI threshold of 30%.This ensures that any two substrates in the training and independent test datasets have a SI of<30%.A statistical summary of both benchmark and independent test datasets is provided in Tables S1 and S2,respectively.Subsequently,the remaining sequences were mapped to PDB[21]by performing PSI-BLAST[22]to search against the PDB sequence database(using the‘pdbaa’file)with three iterations,with an e-value of 10-3,and a SI threshold of 95%.We only retained the X-ray crystallography(X-ray)structures,while nuclear magnetic resonance(NMR)and electron microscopy(EM)structures were discarded.After this procedure,all substrate cleavage sites were mapped onto respective 3D structures using our in-house Perl script and all of these cleavage sites were used as positive samples to train the Procleave models.Sites that have been not annotated as cleavage sites in substrate proteins were considered as negative samples.Accordingly the same number of negative sites was randomly selected as that of the positive samples.In this study,a gallery of all mapped respective 3D structures with visualized cleavage sites can be accessed at http://procleave.erc.monash.edu/gallery.html.

    Feature engineering

    The substrate cleavage site prediction task can be regarded as a binary classification problem.Each cleavage site is denoted as an N-dimensional feature vector F={f1,f2,...,fN}.Three major types of features were extracted,namely structural features, sequence features, and chemical group features. A detailed description of each feature type is presented below.

    Structural features

    In this study,several different types of 3D structural descriptors were extracted from the P4-P4′local windows surrounding cleavage sites,which include:

    (1)Protrusion and depth index.We calculated the protrusion(cx)index and the depth index by CX[23]and DPX[24]programs,respectively.

    (2)Solvent accessibility.Naccess[25]was employed to compute the absolute and relative solvent accessibility features using the default settings.There are five types of solvent accessibility features,including all atoms,total side chain,main chain,non-polar side chain,and allpolar side chain solvent accessibility.

    (3)Packing.Packing was calculated using the method proposed previously[26].

    (4)Molecular surface accessibility.Molecular surfaces are either solvent-accessible surfaces (SAS) or solventexcluded surfaces(SES).Both were calculated by the MSMS program[27].

    (5)Secondary structure features.The DSSP program[28]was used to calculate the secondary structure features.These encompass hydrogen bonds,secondary structures(eight classes were transformed to three classes,i.e.,αhelix,β-sheet,and coil),and backbone torsion angles.The HBPLUS v.3.06 program[29]was used to calculate the hydrogen bond.

    (6)Solvent exposure properties.Half-sphere exposure properties were also used as candidate features.They were extracted using the Biopython package [30]. They included contact number(CN),the number of Cα atoms in the upper half-sphere(HSEAU),the number of Cα atoms in the lower half-sphere(HSEAD),the number of Cβ atoms in the upper half-sphere(HSEBU),and the number of Cβ atoms in the lower half-sphere(HSEBD).

    (7)B-factor.The B-factor values of all atoms were extracted from PDB files,with the average values being used as the input feature[31].

    Sequence features

    We employed the binary encoding scheme to extract and encode sequence features. In particular, a sliding window approach(P4-P4′)centred around the potential cleavage sites was used to extract the local sequence features.Each amino acid(AA)residue was encoded by a binary vector with 20 dimensions.Therefore,the total number of dimensions of the obtained vector is 8×20=160.

    Chemical group features

    Apart from structural and sequence features,the chemical/structural groupings of AAs were also used as candidate features. According to the chemical/structural properties, 20 AAs were clustered into eight chemical groups[32].These include sulfur-containing(residues C and M),aliphatic 1(residues A,G,and P),aliphatic 2(residues I,L,and V),acidic(residues D and E),basic(residues H,K,and R),aromatic(residues F,W,and Y),amide(residues N and Q),and small hydroxy(residues S and T)residues.Then,these eight chemical groups were encoded as input features using the one-hot encoding.The total number of dimensions of the chemical group features is 8×8=64(for any 8-AA window).

    Model training and optimization

    CRFs and LOWESS data smoothing

    CRFs are a type of undirected graphical models originally introduced by Lafferty et al.[33]to deal with the segmentation and labelling tasks of text sequences.CRFs have been proven to be effective in a number of applications with structured outputs,such as information extraction,image processing,and parsing.A CRF is an undirected graph,and its nodes can be categorized as two disjoint sets,namely the observed variables X and the output variables Y.Its principle is to define a conditional probability distribution p(Y|X)over label sequences Y ={y1,y2,···,yn}, given the observational sequence X ={x1,x2,···,xn}.Yis a sequence of hidden state variables that needs to be inferred given the observation.y1,···,yi,yi+1,···,ynare structured to form a chain,with an edge between each yiand yi+1.The distribution of the network has the following form:

    Since a CRF does not have the assumption for the distribution of inputs and, instead, finds the decision boundary directly,it may be considered as an extended version of logistic regression to model sequential data.CRFs have been applied to bioinformatics rather recently and have delivered promising results,such as for gene prediction[34]and phosphorylation sites prediction[32].CRFs can capture sophisticated dependencies and combine information from different aspects.The specific advantages of CRFs are well-suited for incorporating structural information into a cleavage site prediction algorithm.Many of the structural parameters are closely related,and structural parameters contain important information for determining the potential cleavage site that might be better captured by CRFs.

    In this study,our input variables X are the structural,sequence,and chemical group features of a given substrate peptide and the output variables are binary labels corresponding to‘‘cleavage site”or‘‘non-cleavage site”.The CRF models were trained by maximizing the likelihood that the positive samples of a training set were cleavage sites,given their structural,sequence,and chemical group features.We used the open source package CRF++(version 0.54)and,as part of the CRF implementation,used Boolean feature functions to train the models.As the Boolean feature functions evaluate one of the two states of being true or false for a feature appearing at an exact position,all structural features are regarded in the form of discrete instead of continuous values during the model training.In addition,considering that the substrate cleavage depends on the overall 3D shape or neighbourhood of multiple AAs,structural features recognized by cleavage sites,e.g.,the overall shape of the P4-P4′segment surrounding the potential cleavage sites,we combined CRF with a LOWESS data-smoothing approach[35]and examined whether cleavage site prediction could be further improved.Specifically,feature optimization first ran the LOWESS smoothing algorithm on the input vectors of each structural feature.Then the resulting vectors were discretized into equally sized bins to group similar values for use by the Boolean feature functions.Algorithm 1 describes the detailed procedures of the LOWESS smoothing algorithm.

    Algorithm 1 LOWESS data-smoothing algorithm Input:

    The input to Algorithm 1 was the smoothing range range and the initial feature vector @iniArry, which needed to be smoothed and tuned.In this study,each type of structural feature was described by an 8-bit vector,where each bit was associated with the feature value of a local sliding window(P4-P4′)surrounding the potential cleavage site.The output of Algorithm 1 was the 8-bit vector smoothedArray.The smoothing procedure was performed in a ‘for’ loop. At step 1,#iniArray was the length of feature vector,which equals to eight.At the second step,four variables,namely avey,avex,norm,and weight,were set to 0.These variables represented the average value of y(i.e.,values of the features),the average value of x(positions of the feature vector),the normalization variable,and the weight of the variable,respectively.At step 3, the if statement has three different expressions ExpressionN,which can be presented as:

    For these three expressions,the range SNof the parameter range in the step 4 and step 12 is different:

    Then,at step 5,the weight of the variable was calculated.The method used for calculating the variable weight is also different:

    At steps 6 and 7,weight was used to calculate the normalized values of x and y.Then,avey and avex were updated at step 10 by dividing the normalization variable calculated at step 8.At step 11,the smoothed value of mtop and the smoothed bottom value of mbot were initialized to 0.At steps 12-16,these two variables were calculated and updated,and at step 17 the final output smoothedArray was generated according to these two values.

    We set the smoothing range range from 1 to 5 and the bin number from 1 to 10,respectively,in this study.The smoothing procedure and the number of bins for each type of structural feature were optimized by maximizing the area under the curve (AUC) of the receiver operating characteristic(ROC)curves on the 5-fold cross validation test using the benchmark dataset.In this way,by optimizing the smoothing range and the number of bins for each of the structural features,the optimal combination of smoothing and discretization that best represented structural features of all samples in the training set could be determined.

    Performance evaluation

    To assess the performance of the Procleave models and benchmark it with other currently available methods,a set of five commonly used performance measures were applied,including sensitivity (Sn), specificity (Sp), precision, accuracy (Acc),Matthew’s correlation coefficient(MCC),and AUC.Sn,Sp,Precision,Acc,and MCC are defined as:

    where TP,TN,FP,and FN represent the numbers of true positives, true negatives, false positives, and false negatives,respectively. Moreover, we plotted the ROC curves and accordingly calculated the AUCs,as a primary measure to assess the prediction performance of Procleave models and all compared methods.

    Results and discussion

    Characterization of structural features in the proximity of cleavage sites

    To better understand the structural determinants surrounding cleavage sites of different proteases,we examined the structural features of protease cleavage sites using the curated PDB structure datasets.Bar graphs for a total of 27 proteases presented in Figure 2(9 proteases)and Figure S1(18 proteases)show the secondary structure preferences of protease-specific substrates across the P4-P4′sites surrounding the cleavage sites.As shown in these figures,different protease cleavage sites generally have distinctly different secondary structure preferences.However,on the other hand,some proteases also share similar secondary structure preferences.For instance,the P4-P4′site surrounding cleavage sites of caspase-3,granzyme B(human)(Figure 2E and H),cathepsin S,caspase-6,meprin α subunit,meprin β subunit,and LAST_MAM peptidase(Figure S1G,J,and L-N)are more likely to be located in loop regions than in helix and strand regions.In addition,the cleavage sites of most proteases can be found in all three types of secondary structures,except for those of necepsin-1,cathepsin L1(Fasciola sp.),falcipain-2,and falcipain-3(Figure S1D,F,H,and I).The cleavage sites of these four proteases are predominately found in helix and loop regions,but not in strands.The results are in good agreement with the findings of existing studies and suggest that proteases prefer to cleave within loop regions of substrate proteins,while cleavage within helix/sheet regions is also possible[16-18].In addition,we plotted the boxplots for other structural features of positive samples(cleavage sites)for all 27 proteases.These results are provided in supplementary figures,including protrusion index(Figure S2),depth index(Figure S3),solvent accessibility calculated by Naccess(Figures S4-S13),packing(Figure S14),solvent exposure properties(Figures S15 and S16),solvent accessibility calculated by DSSP(Figure S17),backbone torsion angles(Figures S18 and S19),solvent exposure properties(Figures S20-24),B-factor(Figure S25),and hydrogen bonds(Figure S26).

    Figure 2 Structural determinants of the substrate specificity of nine proteases across the P4-P4′cleavage sites

    Performance assessment

    To examine how the structural features help to predict the cleavage sites and how our proposed feature smoothing algorithm improves the prediction performance of trained CRF models,we evaluated the performance of different types of feature combinations.The experiments were conducted by performing 10 times of 5-fold cross-validation tests using the benchmark datasets.The evaluated features/feature combinations include Seq only(using sequence features only),Seq+Chem(using sequence features together with chemical features),Seq+Chem+real structure(using sequence,chemical,and original structural features,without any smoothing),Seq+Chem+smooth DSSP (using sequence, chemical,and smoothed DSSP structural features),and Seq+Chem+smooth structure(using sequence,chemical,and smoothed structural features).Performance comparisons of different feature combinations in terms of AUC values(average AUC values of 10 times of 5-fold cross-validation tests)are shown in Figure 3 and Table S3.

    From these results,several important observations can be made. The Seq+Chem+smooth structure models performed the best compared with all other feature combinations in terms of AUC values for 22 of the 27 tested proteases(see Figure 3 and detailed results in Table S3).Also,the Seq+Chem+smooth DSSP models achieved highest AUC values for meprin β and chymotrypsin A(bovine),while the Seq+Chem models achieved highest AUC values for cathepsin B and lysyl peptidase(bacteria).Seq only model performed the best for HIV-1 retropepsin.These results demonstrate that the sequence features and chemical group features are more relevant and important for the three proteases,while the structural features may not be useful for further improving the cleavage site prediction performance for these proteases.Not surprisingly,the Seq+Chem+real structure models performed the worst among all the compared feature combination models,because the Boolean feature functions of the CRF cannot deal properly with continuous values.This not only leads to the loss of some useful feature information,but also affects the model training.

    In addition,to test and verify the statistical significance of AUC improvement by the Seq+Chem+smooth structure models, we conducted a student’s t-test to compare the AUC values of different feature combination models trained with CRF.The P values of the student’s t-test are given in Table 1, indicating that the AUCs of the Seq+Chem+smooth structure models were significantly (P ≤0.01,marked in bold)higher than those of other models according to the pairwise tests.Feature combinations that achieved the best performance during each comparison test are underlined in Table 1. Furthermore, the AUC values of the Seq+Chem+smooth DSSP models were significantly higher than those of the Seq+Chem and the Seq+Chem+real structure models,while inconclusive with the Seq only models.Altogether,both the performance comparisons and pairwise t-test comparisons demonstrate that structural features smoothed by the LOWESS data smoothing algorithm can greatly help to boost the performance of CRF models.A possible explanation is that the LOWESS smoothing takes the structural variables defined over the cleavage segment P4-P4′sites,and flattens the fluctuations of the structural variables over the eight AA residues of the cleavage sites.This makes intuitive sense because the structural variables are defined over the crystal structure of the protein,which represents only one of the many conformations that constitute the equilibrium ensemble of the protein in solution.In particular,the cleavage site is generally located on or near the surface of the protein,where the side chains of residues on the surface are particularly prone to fluctuations due to thermal contact with the water[16].As such,a single value for the structural variables of a given AA residue will not be a fair representation,especially given that in crystal structures,sidechain conformations on the surface are often flush against symmetric repeats of the protein[16].As such,the smoothing of the structural parameters provides a way to reduce these effects and a more appropriate representation of the structural determinants of cleavage sites.

    Moreover,in order to further illustrate the advantage of CRF,we benchmarked the performance of CRF models with that of the other two popular machine learning algorithms,i.e.,support vector machine(SVM)and random forest(RF),on both the training and independent test datasets.The performance results on the 5-fold cross validation and independent tests are provided in Tables S3 and S4,respectively.As a result,the CRF models achieved the best performance across almost all comparative experiments on the training datasets.The only exceptions were the Seq+Chem+real structure feature for matrix metallopeptidase 2(MMP-2)and the Seq+Chem feature for both astacin and meprin α,for which the RF models achieved the best prediction results.For the performance evaluation on the independent test,we applied the SVM and RF models trained using the Seq+Chem+smooth structure feature combinations,as the SVM and RF models trained on this feature combination performed the best compared to all the other feature combinations.The performance results on the independent tests confirm that the CRF models of Procleave achieve overall a better performance than SVM and RF models,for all 27 proteases examined.Taken together,the performance results on both 5-fold cross validation and independent tests demonstrate the superiority of the CRF framework,making it the model of choice for the development of Procleave.

    Therefore,we accordingly built two prediction models for protease cleavage site prediction from both protein sequences and structures.We built the Procleave_sequence based on Seq+Chem feature combination models for protease cleavage site prediction from protein sequences;while the Procleave_smooth based on Seq+Chem+smooth structure feature combination was built for protease cleavage site prediction from protein structures.

    Comparison with existing methods

    Figure 3 Performance comparison of CRF models trained using different feature combinations in terms of AUC values

    We compared the performance of two variant models‘Procleave_sequence’ and ‘Procleave_smooth’ against five existing tools, including PoPS, SitePrediction, PROSPER,PROSPERous,and iProt-Sub,by performing the independent test.In order to avoid any potential bias and objectively assess the performance,we submitted the PDB sequences in the FASTA format in the independent test dataset to each of the webservers of these methods.The detailed performance results are summarized in Table S4.In addition to AUC,MCC,Acc,Sn,Sp,and precision are also provided and listed in Table S4,while ROC curves are presented in Figure 4.

    Table 1 P values for pairwise t-test comparisons of prediction performance using different feature combinations

    Figure 4 Comparison of cleavage site prediction performance of Procleave and other methods in terms of AUC values for 5 different proteases

    Figure 4 displays the ROC curves of PoPS,SitePrediction,PROSPER,PROSPERous,iProt-Sub,Procleave_smooth,and Procleave_sequence on the independent test dataset.As the entries in the independent test dataset were obtained solely from the newly identified protease substrates and cleavage sites from the most-recent version of MEROPS(12.0)as compared to its previous version(release 9.0),the amount of newly added data was relatively small,and there was even fewer data remaining after mapping onto the PDB 3D structures.Therefore,only five proteases were used for the test,including cathepsin E,caspase-3,caspase-6,MMP-2,and granzyme B(human).As can be seen,Procleave_smooth(red line)performed the best and Procleave_sequence(green line)ranked second in terms of AUC for Cathepsin E (Figure 4A),MMP-2(Figure 4D),and granzyme B(human)(Figure 4E).For caspase-3, Procleave_sequence and PROSPERous achieved the best performance (AUC=1) and Procleave_smooth achieved the second highest AUC(0.990)(Figure 4B). While for caspase-6, PROSPERous achieved the highest AUC(0.999)value and Procleave_smooth ranked second(Figure 4C).To summarize,all these results demonstrate that Procleave is a reliable and powerful bioinformatics approach that improves protease cleavage site prediction.In particular,there are three important factors that account for the good performance of Procleave.First,the high quality and comprehensive 3D structural substrate cleavage data provide solid foundation for the training of Procleave.Second,extracting useful and complementary 3D structural features as calculated by multiple software tools provides a better description of the characteristics of substrate cleavage sites.And lastly,processing initial 3D structural features using the LOWESS data-smoothing algorithm is necessary to enable CRF to learn the underlying rules and characteristics of protease-specific cleavage events.

    Webserver implementation

    To facilitate bioinformatics analyses of novel protease target substrates and cleavage sites, we implemented the CRFbased Procleave approach and developed a publicly available webserver for the wider research community.The Procleave webserver was implemented using HTML and Perl.The webserver is freely accessible at http://procleave.erc.monash.edu/.Procleave webserver is operated by Tomcat7 and configured in a Linux server with an eight-core CPU,500-GB hard disk and 16-GB memory.Both the Procleave_smooth and Procleave_sequence variant models are implemented on the web server.The web server requires two steps of inputs in order to make a prediction of the potential cleavage sites for the given protein.First,Procleave_smooth requires users to supply a protein 3D structure file(*.pdb file is preferred),while for Procleave_sequence models,users are required to input the FASTA formatted protein sequences.Second,users need to specify the PDB chain name and protease type in the case of submitting the 3D structure file. Each submission takes approximately 3-4 min to complete.The prediction outcome for the submitted structure file is returned on the result webpage.The prediction results can be exported in the CSV,Excel,and PDF formats.3Dmol.js[36]is also employed for protein 3D structure visualization at the webserver.The predicted potential cleavage sites are labelled at their corresponding positions.

    Structural proteome-wide prediction

    Furthermore,we conducted a structural proteome-wide prediction of novel protease substrate cleavage sites(containing 17,628 human proteins extracted from the PDB database)by applying the Procleave_smooth model.The results are briefly summarized in this section.We applied an Sp threshold of 99%to all predictions[15,31,37,38]to generate a compendium of high-confidence predicted cleavage sites and then performed the statistical analyses.Statistics of the identified cleavage substrates and the predicted cleavage sites for the 27 different proteases are summarized in Table S5.The results of the identified cleavage substrates and their cleavage sites are also accessible at the Procleave webserver,which can be freely downloaded at http://procleave.erc.monash.edu/.

    Case study

    Figure 5 Predicted cleavage sites of four substrate protein structures

    To illustrate the utility and capacity of Procleave,a case study of the protease-specific cleavage site prediction in four substrate proteins was conducted in this section.The four proteins were selected from the independent test dataset.The first protein is human αB crystalline(PDB ID:3L1G,chain A),which functions as a chaperone and oligomeric assembly.It serves as a stability sensor and can recognize and bind to destabilized proteins in eye lens and other tissues[39].The second protein is human interferon β(PDB ID:1AU1,chain A),which is the protein to defend the cells from various viruses[40].The third protein is an ATPase p97 mutant(PDB ID:3HU2,chain A).ATPase p97 is one of the most abundant cytosolic proteins and can interact with different adaptor proteins involved in many cellular activities,including protein degradation,cell cycle regulation,and membrane fusion[41].The fourth protein is human enolase 1(PDB ID:3B97,chain A),which is a glycolytic enzyme expressed in most tissues.A previous study indicates that this protein is involved in many diseases,including metastatic cancer,ischaemia,autoimmune disorders,and bacterial infection[42].Structure scanning results and the predicted cleavage sites are shown in Figure 5 and Table S6. All correctly predicted cleavage sites are highlighted in red.These prediction results of demonstrate that Procleave could correctly identify all the experimentally verified cleavage sites.These results suggest that Procleave is a useful tool and can be used to identify cleavage sites based on the 3D structural information of the substrate proteins.

    Conclusion

    In the present work,we developed Procleave,a new CRF approach, which combines both sequence and structural information to enhance the protease-specific cleavage site prediction.Procleave employs multi-faceted 3D structure-based features,in combination with a LOWESS smoothing optimization algorithm to train and optimize the CRF-based cleavage site prediction models for a protease.We conducted a comprehensive set of empirical benchmarking tests to benchmark the performance of CRF models built based on different combinations of sequence,chemical,and structural features.We also assessed the performance of Procleave with several state-of-the-art approaches.The comparison results demonstrate that Procleave outperforms these methods,and the LOWESS smoothing optimization is critical to the performance of Procleave.The aim of this study is to systematically investigate whether both sequence-derived and real 3D structural information can be integrated in a machine learning framework to improve the substrate cleavage site prediction for 27 major proteases.A user-friendly webserver of Procleave has been made available as an implementation of the proposed approach.All predicted cleavage sites of the human proteome with 3D the structure data available are provided for further protease biology research.We envisage that Procleave will become a useful tool in the future,facilitating community-wide hypothesis-driven experimental design and functional characterization studies. As a generally useful framework,the CRF-based methodology combined with the LOWESS smoothing optimization algorithm can be readily extended and applied to develop useful methods for predicting other important types of PTM sites[43-46]and functional sites that utilize 3D structural information in future work.

    Data availability

    The datasets and proteome-wide prediction results are publicly accessible at http://procleave.erc.monash.edu/gallery.html/.

    Authors’contributions

    JS,JL,and TML conceived the project and supervised the study.JS,FL,and AL designed the algorithm and drafted the manuscript.FL performed the machine learning experiments and analyzed the results.FL and YW analyzed the performance comparison results.FL,QL,and DX implemented the online webserver.TA,GIW,and AIS revised the manuscript critically for important intellectual content.All authors read,revised,and approved the final manuscript.

    Competing interests

    The authors have declared no competing interests.

    Acknowledgments

    JS was financially supported by grants from the Australian Research Council (ARC) (Grant Nos. LP110200333 and DP120104460),National Health and Medical Research Council of Australia (NHMRC) (Grant Nos. APP1127948,APP1144652,and APP490989),the National Institute of Allergy and Infectious Diseases of the National Institutes of Health,USA(Grant No.R01 AI111965),and a Major Inter-Disciplinary Research(IDR)Grant Awarded by Monash University,Australia(Grant Nos.2019-32 and 2018-28).AL and TML was supported in part by Informatics start-up packages through the School of Medicine,University of Alabama at Birmingham,USA.JL is a NHMRC Principal Research Fellow.

    Supplementary material

    Supplementary data to this article can be found online at https://doi.org/10.1016/j.gpb.2019.08.002.

    ORCID

    0000-0001-5216-3213(Li F)

    0000-0002-2647-2693(Leier A)

    0000-0002-0611-0556(Liu Q)

    0000-0002-3820-8443(Wang Y)

    0000-0002-2937-2786(Xiang D)

    0000-0001-9763-797X(Akutsu T)

    0000-0001-9963-5169(Webb GI)

    0000-0002-4143-2892(Smith AI)

    0000-0003-3279-5592(Marquez-Lago T)

    0000-0001-7953-8230(Li J)

    0000-0001-8031-9086(Song J)

    少妇被粗大猛烈的视频| а√天堂www在线а√下载| av天堂中文字幕网| 国产淫片久久久久久久久| 少妇的逼好多水| 欧美日韩一区二区视频在线观看视频在线 | 欧美又色又爽又黄视频| 欧美日韩综合久久久久久| 中文亚洲av片在线观看爽| 日本成人三级电影网站| 黄色一级大片看看| 赤兔流量卡办理| 国产亚洲欧美98| 18禁裸乳无遮挡免费网站照片| www日本黄色视频网| 婷婷精品国产亚洲av| 免费观看精品视频网站| 国产69精品久久久久777片| 国产精品久久久久久av不卡| 亚洲精品色激情综合| 少妇人妻一区二区三区视频| 在线天堂最新版资源| 麻豆一二三区av精品| 国产久久久一区二区三区| 欧美色欧美亚洲另类二区| 免费在线观看成人毛片| 日本欧美国产在线视频| 日韩av在线大香蕉| 国产一区二区三区在线臀色熟女| 国产高清视频在线播放一区| 啦啦啦韩国在线观看视频| 在线观看午夜福利视频| 九九热线精品视视频播放| 日本 av在线| 看十八女毛片水多多多| 久久久a久久爽久久v久久| 亚洲成a人片在线一区二区| 我要看日韩黄色一级片| 亚洲欧美成人综合另类久久久 | ponron亚洲| 中文字幕人妻熟人妻熟丝袜美| 91在线精品国自产拍蜜月| 国产精品久久久久久精品电影| 大香蕉久久网| 伊人久久精品亚洲午夜| 成人二区视频| 熟女人妻精品中文字幕| 精品人妻熟女av久视频| 在现免费观看毛片| 最后的刺客免费高清国语| 亚洲高清免费不卡视频| 最近视频中文字幕2019在线8| 99国产极品粉嫩在线观看| 亚洲成人久久性| 色综合色国产| 特级一级黄色大片| 欧美人与善性xxx| 免费看av在线观看网站| 国产精华一区二区三区| 一个人免费在线观看电影| 搡女人真爽免费视频火全软件 | 日韩国内少妇激情av| 久久精品夜色国产| 国语自产精品视频在线第100页| 桃色一区二区三区在线观看| 观看美女的网站| 99九九线精品视频在线观看视频| 人人妻,人人澡人人爽秒播| 日本黄大片高清| 一区二区三区高清视频在线| 国产午夜福利久久久久久| 国产午夜福利久久久久久| 亚洲电影在线观看av| 久久鲁丝午夜福利片| 在线看三级毛片| 国产精品一区www在线观看| 天天躁日日操中文字幕| 美女xxoo啪啪120秒动态图| 亚洲av免费在线观看| 国产 一区精品| 亚洲第一区二区三区不卡| 国产精品一区二区三区四区免费观看 | 舔av片在线| 深爱激情五月婷婷| 亚洲最大成人av| 综合色丁香网| a级毛片a级免费在线| 国产精品一区二区三区四区久久| 国产一区二区三区av在线 | 国产精品一区二区三区四区免费观看 | 婷婷亚洲欧美| 久久久久久久久久成人| 99热这里只有精品一区| 在线看三级毛片| 热99re8久久精品国产| 综合色av麻豆| 在线免费观看的www视频| 美女大奶头视频| 亚洲乱码一区二区免费版| 国内精品美女久久久久久| 中文亚洲av片在线观看爽| h日本视频在线播放| 色5月婷婷丁香| 看黄色毛片网站| 国产探花在线观看一区二区| 国产69精品久久久久777片| 国产淫片久久久久久久久| 一个人观看的视频www高清免费观看| 精品一区二区免费观看| 18禁黄网站禁片免费观看直播| 亚洲国产精品合色在线| 青春草视频在线免费观看| av黄色大香蕉| 国产成人a区在线观看| 日本一本二区三区精品| 精品一区二区三区av网在线观看| 内地一区二区视频在线| 十八禁国产超污无遮挡网站| 亚洲无线观看免费| 国产精品99久久久久久久久| 亚洲欧美日韩东京热| 免费人成在线观看视频色| 日本黄色片子视频| 国内精品宾馆在线| 日韩av在线大香蕉| 久久久精品欧美日韩精品| 免费不卡的大黄色大毛片视频在线观看 | 成年女人毛片免费观看观看9| 国产人妻一区二区三区在| aaaaa片日本免费| 国产69精品久久久久777片| 午夜精品国产一区二区电影 | 亚洲无线观看免费| 日韩一区二区视频免费看| 99热精品在线国产| 大又大粗又爽又黄少妇毛片口| 秋霞在线观看毛片| 免费看av在线观看网站| 久久草成人影院| 欧美高清性xxxxhd video| 精品一区二区三区av网在线观看| 精品99又大又爽又粗少妇毛片| 免费看日本二区| 久久久久精品国产欧美久久久| 亚洲美女搞黄在线观看 | 亚洲成人av在线免费| 亚洲中文字幕一区二区三区有码在线看| 全区人妻精品视频| 成年版毛片免费区| 老女人水多毛片| 久久久久久久久久久丰满| 亚洲av不卡在线观看| 国产v大片淫在线免费观看| 乱系列少妇在线播放| 久久国内精品自在自线图片| 精品久久久久久久末码| 亚洲精品久久国产高清桃花| 69人妻影院| 毛片女人毛片| 一级av片app| 男人狂女人下面高潮的视频| 麻豆成人午夜福利视频| 给我免费播放毛片高清在线观看| 日本免费a在线| 成年女人永久免费观看视频| 国产亚洲91精品色在线| 丰满乱子伦码专区| 国产成人福利小说| 国产一区二区在线av高清观看| 亚洲美女黄片视频| 国产伦在线观看视频一区| 久久久久久久久中文| 成人精品一区二区免费| 国产精品一区二区三区四区免费观看 | 伦理电影大哥的女人| 高清午夜精品一区二区三区 | 身体一侧抽搐| 12—13女人毛片做爰片一| 悠悠久久av| 午夜福利18| 日日撸夜夜添| 麻豆国产97在线/欧美| 亚洲无线在线观看| 嫩草影院新地址| 男女视频在线观看网站免费| videossex国产| 国产欧美日韩一区二区精品| 免费看日本二区| 欧美国产日韩亚洲一区| 中文字幕久久专区| 伦理电影大哥的女人| 小说图片视频综合网站| 亚洲精品亚洲一区二区| 日韩中字成人| 国产精品爽爽va在线观看网站| 丰满乱子伦码专区| 夜夜爽天天搞| 99热6这里只有精品| 欧美国产日韩亚洲一区| 免费一级毛片在线播放高清视频| 91狼人影院| 成人高潮视频无遮挡免费网站| av福利片在线观看| 成人亚洲欧美一区二区av| 三级男女做爰猛烈吃奶摸视频| 国产成人freesex在线 | 日韩一本色道免费dvd| 午夜福利在线在线| 美女免费视频网站| 午夜福利在线在线| 97人妻精品一区二区三区麻豆| 婷婷精品国产亚洲av在线| 日韩欧美一区二区三区在线观看| 搡女人真爽免费视频火全软件 | 岛国在线免费视频观看| 最近在线观看免费完整版| 精品午夜福利视频在线观看一区| 国产色爽女视频免费观看| 在线国产一区二区在线| 欧美不卡视频在线免费观看| 一进一出抽搐gif免费好疼| av国产免费在线观看| 99热只有精品国产| 综合色丁香网| 丝袜美腿在线中文| 久久国内精品自在自线图片| 成年免费大片在线观看| 国产乱人偷精品视频| 成人欧美大片| 特级一级黄色大片| 亚洲欧美精品自产自拍| 国产精品久久久久久亚洲av鲁大| 午夜影院日韩av| 菩萨蛮人人尽说江南好唐韦庄 | 性插视频无遮挡在线免费观看| 国产美女午夜福利| 精品久久久噜噜| 伊人久久精品亚洲午夜| 97碰自拍视频| 久久精品国产清高在天天线| 亚洲av第一区精品v没综合| 国产欧美日韩一区二区精品| 级片在线观看| 国产熟女欧美一区二区| 99热只有精品国产| 成人无遮挡网站| 最近的中文字幕免费完整| 啦啦啦观看免费观看视频高清| 久久亚洲国产成人精品v| 欧美日韩国产亚洲二区| 日韩欧美在线乱码| 在线天堂最新版资源| 精品国产三级普通话版| 色哟哟·www| 插阴视频在线观看视频| 欧美3d第一页| 看片在线看免费视频| 波野结衣二区三区在线| 午夜福利在线观看免费完整高清在 | 亚洲成人久久爱视频| 亚洲不卡免费看| 亚洲最大成人中文| av中文乱码字幕在线| 日本精品一区二区三区蜜桃| 中国美女看黄片| 色视频www国产| 精品久久久久久久久久免费视频| 啦啦啦韩国在线观看视频| 免费无遮挡裸体视频| 日本免费一区二区三区高清不卡| 日韩欧美精品v在线| 97在线视频观看| 能在线免费观看的黄片| 久久久精品大字幕| 国产熟女欧美一区二区| 少妇被粗大猛烈的视频| 中国国产av一级| 中国美女看黄片| 少妇熟女欧美另类| 成人综合一区亚洲| 亚洲人成网站在线播放欧美日韩| 日本黄大片高清| 人人妻,人人澡人人爽秒播| 色噜噜av男人的天堂激情| 在线观看一区二区三区| 免费高清视频大片| 国产高清有码在线观看视频| 国产真实乱freesex| 国内久久婷婷六月综合欲色啪| 老师上课跳d突然被开到最大视频| 亚洲三级黄色毛片| 国产探花极品一区二区| 亚洲成人精品中文字幕电影| 国产精品精品国产色婷婷| 一本久久中文字幕| 中文在线观看免费www的网站| 中文资源天堂在线| 婷婷精品国产亚洲av在线| 国产精品三级大全| 亚洲人与动物交配视频| 麻豆一二三区av精品| 国产精品一区二区性色av| 我的女老师完整版在线观看| 两个人视频免费观看高清| 日本黄色片子视频| 看黄色毛片网站| 悠悠久久av| 日韩三级伦理在线观看| 欧美成人a在线观看| 国产精品伦人一区二区| 中文字幕av在线有码专区| 亚洲欧美精品自产自拍| 久久精品国产清高在天天线| 欧美在线一区亚洲| 国产v大片淫在线免费观看| 在线播放无遮挡| 身体一侧抽搐| 最新中文字幕久久久久| 午夜a级毛片| 淫秽高清视频在线观看| 美女xxoo啪啪120秒动态图| 国内精品宾馆在线| 色尼玛亚洲综合影院| 黄色配什么色好看| av中文乱码字幕在线| 男女边吃奶边做爰视频| 99久国产av精品国产电影| 免费黄网站久久成人精品| 久久鲁丝午夜福利片| eeuss影院久久| 美女xxoo啪啪120秒动态图| 在线免费观看不下载黄p国产| 亚洲av成人精品一区久久| 亚洲经典国产精华液单| 黄片wwwwww| 狠狠狠狠99中文字幕| 成人鲁丝片一二三区免费| 欧美激情国产日韩精品一区| 日本与韩国留学比较| 校园人妻丝袜中文字幕| 国产精品一二三区在线看| 国产亚洲精品av在线| 日韩av不卡免费在线播放| 97在线视频观看| 欧美xxxx性猛交bbbb| 99精品在免费线老司机午夜| 成人av一区二区三区在线看| 久久久久久久久久黄片| 午夜福利视频1000在线观看| 国产又黄又爽又无遮挡在线| 国产老妇女一区| 少妇的逼好多水| 99热6这里只有精品| 国产午夜精品论理片| 欧美日韩精品成人综合77777| 天美传媒精品一区二区| 久久精品国产99精品国产亚洲性色| 欧美潮喷喷水| 国产爱豆传媒在线观看| 九九热线精品视视频播放| 美女 人体艺术 gogo| 日韩欧美在线乱码| 国产精品一区二区性色av| 少妇的逼好多水| a级毛色黄片| 日韩精品中文字幕看吧| 亚洲av.av天堂| 国产亚洲精品久久久com| 国产 一区 欧美 日韩| 波野结衣二区三区在线| 国产男人的电影天堂91| 日本欧美国产在线视频| 久久久成人免费电影| 日日干狠狠操夜夜爽| 色av中文字幕| 国产精品99久久久久久久久| 91av网一区二区| 色播亚洲综合网| 国产在线精品亚洲第一网站| 国产精品一区二区三区四区免费观看 | 婷婷精品国产亚洲av| 国产一区二区亚洲精品在线观看| 国内精品宾馆在线| 亚洲天堂国产精品一区在线| 全区人妻精品视频| 色综合站精品国产| 国产日本99.免费观看| 黄色一级大片看看| 国产伦精品一区二区三区视频9| 在线观看午夜福利视频| 成人美女网站在线观看视频| 午夜爱爱视频在线播放| 老司机影院成人| 看片在线看免费视频| 又粗又爽又猛毛片免费看| 亚洲av一区综合| 精品久久久久久成人av| 在线天堂最新版资源| АⅤ资源中文在线天堂| 午夜激情欧美在线| 久久久久久久久久成人| 久久精品国产自在天天线| 欧美成人a在线观看| 欧美一级a爱片免费观看看| 日韩人妻高清精品专区| 欧美激情在线99| 亚洲av成人精品一区久久| 欧美一区二区精品小视频在线| 成人国产麻豆网| 欧美成人a在线观看| 最近最新中文字幕大全电影3| 少妇熟女aⅴ在线视频| av中文乱码字幕在线| 亚洲性夜色夜夜综合| 男人狂女人下面高潮的视频| 18+在线观看网站| 91午夜精品亚洲一区二区三区| 欧美日韩在线观看h| 色综合站精品国产| 美女 人体艺术 gogo| 欧美激情国产日韩精品一区| 欧美bdsm另类| 亚洲精品久久国产高清桃花| 国产在线精品亚洲第一网站| 精品人妻偷拍中文字幕| 听说在线观看完整版免费高清| 非洲黑人性xxxx精品又粗又长| 麻豆一二三区av精品| 国产午夜精品久久久久久一区二区三区 | 白带黄色成豆腐渣| 日韩欧美一区二区三区在线观看| 日本一二三区视频观看| 国产在线精品亚洲第一网站| 91久久精品电影网| 亚洲性夜色夜夜综合| 女的被弄到高潮叫床怎么办| 亚洲专区国产一区二区| 天堂√8在线中文| a级毛色黄片| 老熟妇乱子伦视频在线观看| 亚洲综合色惰| 在线免费十八禁| 亚洲无线观看免费| 美女黄网站色视频| 如何舔出高潮| 97碰自拍视频| 一级毛片我不卡| 国产伦在线观看视频一区| 日本精品一区二区三区蜜桃| 美女高潮的动态| 国产成人精品久久久久久| 超碰av人人做人人爽久久| 中文字幕久久专区| 日本 av在线| 99在线视频只有这里精品首页| 欧美高清成人免费视频www| 久久这里只有精品中国| 国产麻豆成人av免费视频| 久久久久性生活片| 看免费成人av毛片| 国产一区二区激情短视频| 免费看av在线观看网站| 三级国产精品欧美在线观看| av在线播放精品| 小蜜桃在线观看免费完整版高清| 精品国产三级普通话版| 精品无人区乱码1区二区| 免费无遮挡裸体视频| 成人特级av手机在线观看| 乱码一卡2卡4卡精品| 国产精品无大码| 别揉我奶头~嗯~啊~动态视频| 看十八女毛片水多多多| 成年女人永久免费观看视频| 国产av一区在线观看免费| 日日啪夜夜撸| 卡戴珊不雅视频在线播放| 日韩大尺度精品在线看网址| 欧美日韩在线观看h| 最近手机中文字幕大全| 亚洲性久久影院| 国产一区二区三区在线臀色熟女| 久久久久久久亚洲中文字幕| avwww免费| 久久草成人影院| 免费在线观看影片大全网站| 久久韩国三级中文字幕| 国产高清三级在线| 人人妻人人澡欧美一区二区| 久久久久九九精品影院| 日本免费a在线| 日韩大尺度精品在线看网址| 日韩欧美三级三区| 精品人妻一区二区三区麻豆 | 国产精品人妻久久久影院| 国产91av在线免费观看| 国产爱豆传媒在线观看| 禁无遮挡网站| 欧美人与善性xxx| 波多野结衣高清作品| 如何舔出高潮| 国产爱豆传媒在线观看| av卡一久久| 中文资源天堂在线| 日韩欧美三级三区| 又爽又黄无遮挡网站| 免费在线观看影片大全网站| 亚洲av熟女| 午夜视频国产福利| 亚洲av免费在线观看| 精品人妻一区二区三区麻豆 | 亚洲18禁久久av| 色尼玛亚洲综合影院| 一级a爱片免费观看的视频| 麻豆成人午夜福利视频| 亚洲人成网站在线播放欧美日韩| 我要看日韩黄色一级片| av专区在线播放| 国产欧美日韩精品亚洲av| 久久久久久久亚洲中文字幕| 精品国产三级普通话版| 国产v大片淫在线免费观看| 亚洲av中文av极速乱| 国产淫片久久久久久久久| 国产精品亚洲一级av第二区| 搡老妇女老女人老熟妇| 赤兔流量卡办理| 日本精品一区二区三区蜜桃| 欧美国产日韩亚洲一区| 国产 一区精品| 国产69精品久久久久777片| 91午夜精品亚洲一区二区三区| 免费搜索国产男女视频| 成人高潮视频无遮挡免费网站| 九九热线精品视视频播放| 啦啦啦韩国在线观看视频| 亚洲av.av天堂| 国产91av在线免费观看| 97人妻精品一区二区三区麻豆| 最近2019中文字幕mv第一页| 身体一侧抽搐| 国产精华一区二区三区| av福利片在线观看| 综合色丁香网| 黄色一级大片看看| 最近中文字幕高清免费大全6| 精品国内亚洲2022精品成人| 丰满的人妻完整版| 晚上一个人看的免费电影| 日日啪夜夜撸| 国产成人aa在线观看| 成人av一区二区三区在线看| 搡老妇女老女人老熟妇| 在线播放无遮挡| 国国产精品蜜臀av免费| 成人美女网站在线观看视频| 丰满的人妻完整版| 精品99又大又爽又粗少妇毛片| 久久精品人妻少妇| 麻豆国产av国片精品| 高清午夜精品一区二区三区 | 色综合亚洲欧美另类图片| 精品午夜福利视频在线观看一区| 成人av一区二区三区在线看| 久久久久国内视频| 看十八女毛片水多多多| 男女之事视频高清在线观看| 少妇的逼水好多| 麻豆国产97在线/欧美| 亚洲人成网站高清观看| 色综合亚洲欧美另类图片| 成年av动漫网址| 久久久成人免费电影| 国产久久久一区二区三区| 亚洲欧美日韩无卡精品| 桃色一区二区三区在线观看| 亚洲精品亚洲一区二区| 可以在线观看的亚洲视频| 1000部很黄的大片| 日韩亚洲欧美综合| 精品一区二区三区视频在线观看免费| 亚洲av第一区精品v没综合| 搡老熟女国产l中国老女人| 91久久精品国产一区二区三区| 精品久久久久久久久久免费视频| 男人舔女人下体高潮全视频| 99久国产av精品| 精品国产三级普通话版| av黄色大香蕉| 国产一区亚洲一区在线观看| 可以在线观看毛片的网站| 成人特级av手机在线观看| 欧美高清成人免费视频www| 深爱激情五月婷婷| 欧美人与善性xxx| 日本爱情动作片www.在线观看 | 久久精品国产清高在天天线| 午夜久久久久精精品| 国产国拍精品亚洲av在线观看| 久久久精品大字幕| 国产黄a三级三级三级人| 99久久精品一区二区三区| 丰满乱子伦码专区| 亚洲国产日韩欧美精品在线观看| 淫秽高清视频在线观看| 亚洲成av人片在线播放无| 在线观看美女被高潮喷水网站| 欧美日韩国产亚洲二区| 亚洲七黄色美女视频| 国产黄片美女视频| 国模一区二区三区四区视频| 亚洲经典国产精华液单| 一级毛片我不卡| 三级国产精品欧美在线观看| 最近2019中文字幕mv第一页|